wehub/huggingface--lerobot
1
0
Fork 0
Code Issues Pull Requests Actions Deployments Models Agent Notes Packages Projects Releases Wiki Activity

chore: import upstream snapshot with attribution

Browse Source
This commit is contained in:
wehub-resource-sync
2026-07-13 12:58:47 +08:00
commit e68525745e
990 changed files with 259649 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/source/_toctree.yml
+192
View File
@@ -0,0 +1,192 @@
- sections:
- local: index
title: LeRobot
- local: installation
title: Installation
- local: cheat-sheet
title: Cheat sheet
title: Get started
- sections:
- local: il_robots
title: Imitation Learning for Robots
- local: lelab
title: LeLab - Lerobot GUI
- local: bring_your_own_policies
title: Adding a Policy
- local: integrate_hardware
title: Bring Your Own Hardware
- local: hilserl
title: Train a Robot with RL
- local: hilserl_sim
title: Train RL in Simulation
- local: multi_gpu_training
title: Multi GPU training
- local: hil_data_collection
title: Human In the Loop Data Collection
- local: peft_training
title: Training with PEFT (e.g., LoRA)
- local: rename_map
title: Using Rename Map and Empty Cameras
title: "Tutorials"
- sections:
- local: hardware_guide
title: Compute Hardware Guide
- local: torch_accelerators
title: PyTorch accelerators
title: "Compute & Hardware"
- sections:
- local: lerobot-dataset-v3
title: Using LeRobotDataset
- local: porting_datasets_v3
title: Porting Large Datasets
- local: using_dataset_tools
title: Using the Dataset Tools
- local: language_and_recipes
title: Language Columns and Recipes
- local: tools
title: Tools
- local: annotation_pipeline
title: Annotation Pipeline
- local: video_encoding_parameters
title: Video encoding parameters
- local: streaming_video_encoding
title: Streaming Video Encoding
title: "Datasets"
- sections:
- local: act
title: ACT
- local: smolvla
title: SmolVLA
- local: pi0
title: π₀ (Pi0)
- local: pi0fast
title: π₀-FAST (Pi0Fast)
- local: pi05
title: π₀.₅ (Pi05)
- local: molmoact2
title: MolmoAct2
- local: vla_jepa
title: VLA-JEPA
- local: eo1
title: EO-1
- local: lingbot_va
title: LingBot-VA
- local: fastwam
title: FastWAM
- local: evo1
title: EVO1
- local: groot
title: NVIDIA GR00T
- local: xvla
title: X-VLA
- local: multi_task_dit
title: Multitask DiT Policy
- local: walloss
title: WALL-OSS
title: "Policies"
- sections:
- local: sarm
title: SARM
- local: robometer
title: ROBOMETER
- local: topreward
title: TOPReward
title: "Reward Models"
- sections:
- local: inference
title: Policy Deployment (lerobot-rollout)
- local: async
title: Use Async Inference
- local: rtc
title: Real-Time Chunking (RTC)
title: "Inference"
- sections:
- local: envhub
title: Environments from the Hub
- local: envhub_leisaac
title: Control & Train Robots in Sim (LeIsaac)
title: "Simulation"
- sections:
- local: adding_benchmarks
title: Adding a New Benchmark
- local: libero
title: LIBERO
- local: libero_plus
title: LIBERO-plus
- local: metaworld
title: Meta-World
- local: robotwin
title: RoboTwin 2.0
- local: robocasa
title: RoboCasa365
- local: robocerebra
title: RoboCerebra
- local: robomme
title: RoboMME
- local: envhub_isaaclab_arena
title: NVIDIA IsaacLab Arena Environments
- local: vlabench
title: VLABench
title: "Benchmarks"
- sections:
- local: introduction_processors
title: Introduction to Robot Processors
- local: debug_processor_pipeline
title: Debug your processor pipeline
- local: implement_your_own_processor
title: Implement your own processor
- local: processors_robots_teleop
title: Processors for Robots and Teleoperators
- local: env_processor
title: Environment Processors
- local: action_representations
title: Action Representations
title: "Robot Processors"
- sections:
- local: so101
title: SO-101
- local: so100
title: SO-100
- local: koch
title: Koch v1.1
- local: lekiwi
title: LeKiwi
- local: hope_jr
title: Hope Jr
- local: reachy2
title: Reachy 2
- local: unitree_g1
title: Unitree G1
- local: earthrover_mini_plus
title: Earth Rover Mini
- local: omx
title: OMX
- local: openarm
title: OpenArm
- local: rebot_b601
title: reBot B601-DM
title: "Robots"
- sections:
- local: phone_teleop
title: Phone
- local: isaac_teleop
title: Isaac Teleop
title: "Teleoperators"
- sections:
- local: cameras
title: Cameras
title: "Sensors"
- sections:
- local: notebooks
title: Notebooks
- local: feetech
title: Updating Feetech Firmware
- local: damiao
title: Damiao Motors and CAN Bus
title: "Resources"
- sections:
- local: contributing
title: Contribute to LeRobot
- local: backwardcomp
title: Backward compatibility
title: "About"
docs/source/act.mdx
+91
View File
@@ -0,0 +1,91 @@
# ACT (Action Chunking with Transformers)
ACT is a **lightweight and efficient policy for imitation learning**, especially well-suited for fine-grained manipulation tasks. It's the **first model we recommend when you're starting out** with LeRobot due to its fast training time, low computational requirements, and strong performance.
<div class="video-container">
<iframe
width="100%"
height="415"
src="https://www.youtube.com/embed/ft73x0LfGpM"
title="LeRobot ACT Tutorial"
frameborder="0"
allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture"
allowfullscreen
></iframe>
</div>
_Watch this tutorial from the LeRobot team to learn how ACT works: [LeRobot ACT Tutorial](https://www.youtube.com/watch?v=ft73x0LfGpM)_
## Model Overview
Action Chunking with Transformers (ACT) was introduced in the paper [Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware](https://arxiv.org/abs/2304.13705) by Zhao et al. The policy was designed to enable precise, contact-rich manipulation tasks using affordable hardware and minimal demonstration data.
### Why ACT is Great for Beginners
ACT stands out as an excellent starting point for several reasons:
- **Fast Training**: Trains in a few hours on a single GPU
- **Lightweight**: Only ~80M parameters, making it efficient and easy to work with
- **Data Efficient**: Often achieves high success rates with just 50 demonstrations
### Architecture
ACT uses a transformer-based architecture with three main components:
1. **Vision Backbone**: ResNet-18 processes images from multiple camera viewpoints
2. **Transformer Encoder**: Synthesizes information from camera features, joint positions, and a learned latent variable
3. **Transformer Decoder**: Generates coherent action sequences using cross-attention
The policy takes as input:
- Multiple RGB images (e.g., from wrist cameras, front/top cameras)
- Current robot joint positions
- A latent style variable `z` (learned during training, set to zero during inference)
And outputs a chunk of `k` future action sequences.
## Installation Requirements
1. Install LeRobot by following our [Installation Guide](./installation).
2. ACT is included in the base LeRobot installation, so no additional dependencies are needed!
## Training ACT
ACT works seamlessly with the standard LeRobot training pipeline. Here's a complete example for training ACT on your dataset:
```bash
lerobot-train \
--dataset.repo_id=${HF_USER}/your_dataset \
--policy.type=act \
--output_dir=outputs/train/act_your_dataset \
--job_name=act_your_dataset \
--policy.device=cuda \
--wandb.enable=true \
--policy.repo_id=${HF_USER}/act_policy
```
### Training Tips
1. **Start with defaults**: ACT's default hyperparameters work well for most tasks
2. **Training duration**: Expect a few hours for 100k training steps on a single GPU
3. **Batch size**: Start with batch size 8 and adjust based on your GPU memory
### Train using Google Colab
If your local computer doesn't have a powerful GPU, you can utilize Google Colab to train your model by following the [ACT training notebook](./notebooks#training-act).
## Evaluating ACT
Once training is complete, you can evaluate your ACT policy using the `lerobot-record` command with your trained policy. This will run inference and record evaluation episodes:
```bash
lerobot-rollout \
--strategy.type=base \
--policy.path=${HF_USER}/act_policy \
--robot.type=so101_follower \
--robot.port=/dev/ttyACM0 \
--robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
--display_data=true \
--task="Your task description" \ # can be skipped for ACT
--duration=60
```
docs/source/action_representations.mdx
+223
View File
@@ -0,0 +1,223 @@
# Action Representations
This guide explains the different ways robot actions can be represented in LeRobot, how they relate to each other, and when to use each one.
## Joint Space vs End-Effector Space
Before discussing action representations, it helps to understand the two coordinate spaces actions can live in.
### Joint Space
Joint-space actions directly specify target positions for each motor. For a 6-DOF arm with a gripper, a joint-space action might look like:
```
action = [shoulder_pan: 45.0, shoulder_lift: -20.0, elbow: -30.0, wrist_pitch: 10.0, wrist_roll: 0.0, wrist_yaw: 5.0, gripper: 0.8]
```
Joint space is the default in LeRobot. It is simple, requires no kinematics model, and maps directly to motor commands. Most beginner setups (SO-100, Koch) use joint-space actions.
### End-Effector (EE) Space
End-effector-space actions specify the desired position and orientation of the robot's tool tip (gripper) in Cartesian coordinates:
```
action = [x: 0.25, y: -0.10, z: 0.15, wx: 0.0, wy: 0.0, wz: 0.1, gripper: 0.8]
```
EE space is more intuitive for tasks like pick-and-place because it directly describes where the gripper should go, but it requires a kinematics model (URDF) to convert between EE poses and joint angles.
### Converting Between Spaces
LeRobot provides processor steps for converting between joint and EE spaces using forward and inverse kinematics. These are built on top of `RobotKinematics`, which loads a URDF model of your robot.
```python
from lerobot.model.kinematics import RobotKinematics
from lerobot.robots.so_follower.robot_kinematic_processor import (
ForwardKinematicsJointsToEE,
InverseKinematicsEEToJoints,
)
kinematics = RobotKinematics(
urdf_path="./SO101/so101_new_calib.urdf",
target_frame_name="gripper_frame_link",
joint_names=["shoulder", "elbow", "wrist_pitch", "wrist_roll", "wrist_yaw"],
)
# Joints → EE (for observations: "where is my gripper?")
fk_step = ForwardKinematicsJointsToEE(kinematics=kinematics, motor_names=[...])
# EE → Joints (for actions: "move my gripper here")
ik_step = InverseKinematicsEEToJoints(kinematics=kinematics, motor_names=[...])
```
See [`examples/so100_to_so100_EE/`](https://github.com/huggingface/lerobot/tree/main/examples/so100_to_so100_EE) for a complete working example of recording, replaying, and evaluating with EE-space actions on an SO-100 arm.
## Absolute, Relative, and Delta Actions
Regardless of whether you work in joint space or EE space, the action values can be expressed in three different ways. The terminology follows [UMI (Chi et al., 2024)](https://arxiv.org/abs/2402.10329).
### Absolute Actions (LeRobot default)
Each action specifies the target position directly.
**Example** (joint space, chunk of 4):
```
current_state = [45.0, -30.0, 10.0]
action_chunk = [
[46.0, -29.0, 11.0], # go to 46, -29, 11
[47.5, -27.0, 12.0], # go to 47.5, -27, 12
[49.0, -25.0, 13.5], # go to 49, -25, 13.5
[50.0, -24.0, 15.0], # go to 50, -24, 15
]
```
Each value is a target position in the robot's coordinate frame. Simple and direct, but requires a consistent global coordinate frame. This is the default in LeRobot.
### Relative Actions (used by OpenPI / pi0)
Each action in the chunk is an offset from the **current state at the moment of prediction**. All actions in the chunk share the same reference point:
```
current_state = [45.0, -30.0, 10.0]
relative_chunk = [
[1.0, 1.0, 1.0], # +1 from current → target 46, -29, 11
[2.5, 3.0, 2.0], # +2.5 from current → target 47.5, -27, 12
[4.0, 5.0, 3.5], # +4 from current → target 49, -25, 13.5
[5.0, 6.0, 5.0], # +5 from current → target 50, -24, 15
]
```
The conversion is straightforward: `relative = absolute - current_state`. To recover absolute: `absolute = relative + current_state`.
**Why use relative actions?** The model learns to predict offsets centered around zero, which is easier to normalize and leads to more stable training. Because every chunk references the same current state, there is no error accumulation across chunks.
### Delta Actions (sequential differences)
Each action is an offset from the **previous action** (or from the current state for the first step):
```
current_state = [45.0, -30.0, 10.0]
delta_chunk = [
[1.0, 1.0, 1.0], # current → 46, -29, 11
[1.5, 2.0, 1.0], # previous action → 47.5, -27, 12
[1.5, 2.0, 1.5], # previous action → 49, -25, 13.5
[1.0, 1.0, 1.5], # previous action → 50, -24, 15
]
```
Here each step is relative to the one before it. To recover absolute positions you must sum all previous deltas, which means errors accumulate over time. UMI explicitly argues against this representation for this reason.
### Visual Comparison
The figure below (based on a figure from [UMI, Chi et al., 2024](https://arxiv.org/abs/2402.10329)) illustrates the key difference. With **relative trajectory**, every action in the chunk points back to the same origin (current state), so a new inference step cleanly resets the reference. With **delta**, each action depends on the previous one, so errors accumulate. **Absolute** actions require a consistent global coordinate frame.
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/action_representations_umi.png"
alt="Relative Trajectory as Action Representation (UMI, Chi et al., 2024)"
width="85%"
/>
## Using Relative Actions in LeRobot
LeRobot provides `RelativeActionsProcessorStep` to convert between absolute and relative actions inside the processor pipeline. This is how pi0, pi0.5, and pi0_fast support relative actions.
> **Note:** All pi models (pi0, pi0.5, pi0*fast) apply relative conversion \_before* normalization (`relative → normalize`), so the normalizer always sees delta (relative) values. This means **relative action stats are required** for all of them when training with `use_relative_actions=true`. In pi0_fast the `RelativeActionsProcessorStep` only modifies the action — the state observation is unchanged — so `NormalizerProcessorStep` still runs before the state tokenizer and the tokenizer continues to receive normalized state as expected.
### How it works
During **training** (preprocessing), actions are converted from absolute to relative before the model sees them:
```
raw absolute action → RelativeActionsProcessorStep → normalize → model
```
During **inference** (postprocessing), model predictions are converted back to absolute before being sent to the robot:
```
model output → unnormalize → AbsoluteActionsProcessorStep → robot
```
The `AbsoluteActionsProcessorStep` reads the cached current state from its paired `RelativeActionsProcessorStep`, so the two must be wired together (handled automatically by the policy factory).
### Enabling relative actions for the pi family (pi0, pi0.5, pi0_fast)
**Step 1**: Precompute relative action statistics for your dataset:
```bash
lerobot-edit-dataset \
--repo_id your_dataset \
--operation.type recompute_stats \
--operation.relative_action true \
--operation.chunk_size 50 \
--operation.relative_exclude_joints "['gripper']"
```
**Step 2**: Train with relative actions enabled:
```bash
lerobot-train \
--dataset.repo_id=your_dataset \
--policy.type=pi0 \
--policy.use_relative_actions=true \
--policy.relative_exclude_joints='["gripper"]'
```
The `relative_exclude_joints` parameter specifies joints that should remain in absolute space. For example, gripper commands are typically binary (open/close) and don't benefit from relative encoding.
### Combining relative actions with RTC
[RTC](https://arxiv.org/abs/2506.07339) runs policy inference at high frequency and sends actions to the robot as they are predicted rather than waiting for a full chunk. Relative actions and RTC are fully compatible: because every chunk in relative mode references the **same** current state (captured at the start of inference), each predicted action in the chunk remains a valid offset even if the robot has already moved. No special handling is needed — `RelativeActionsProcessorStep` caches the state once per inference call and `AbsoluteActionsProcessorStep` applies it to every action in the streamed output.
### Combining relative actions with EE space
Relative actions work in both joint space and EE space. For example, if your dataset stores EE actions, relative encoding converts them to offsets from the current EE pose:
```
current_ee_state = [x: 0.25, y: -0.10, z: 0.15, gripper: 0.8]
absolute_ee_chunk = [
[0.26, -0.09, 0.16, 0.8],
[0.28, -0.07, 0.18, 0.8],
]
relative_ee_chunk = [
[0.01, 0.01, 0.01, 0.0], # offset from current EE pose
[0.03, 0.03, 0.03, 0.0], # offset from current EE pose
]
```
## Processing Pipeline Summary
Here is how the different processors compose. Each arrow is a processor step, and they can be chained in a `RobotProcessorPipeline` or `PolicyProcessorPipeline`:
```
┌─────────────────────────────────────────┐
Action Space │ Joint Space ←──IK──→ EE Space │
│ ForwardKinematicsJointsToEE │
│ InverseKinematicsEEToJoints │
└─────────────────────────────────────────┘
┌─────────────────────────────────────────┐
Representation │ Absolute ←────→ Relative │
│ RelativeActionsProcessorStep (pre) │
│ AbsoluteActionsProcessorStep (post) │
└─────────────────────────────────────────┘
┌─────────────────────────────────────────┐
Normalization │ Raw ←────→ Normalized │
│ NormalizerProcessorStep (pre) │
│ UnnormalizerProcessorStep (post) │
└─────────────────────────────────────────┘
```
A typical training preprocessor might chain: `raw absolute joint actions → relative → normalize`. A typical inference postprocessor: `unnormalize → absolute → (optionally IK to joints)`.
## References
- [Universal Manipulation Interface (UMI)](https://arxiv.org/abs/2402.10329) - Chi et al., 2024. Defines the relative trajectory action representation and compares it with absolute and delta actions.
- [Introduction to Processors](./introduction_processors) - How processor pipelines work in LeRobot.
- [`examples/so100_to_so100_EE/`](https://github.com/huggingface/lerobot/tree/main/examples/so100_to_so100_EE) - Complete example of recording and evaluating with EE-space actions.
docs/source/adding_benchmarks.mdx
+322
View File
@@ -0,0 +1,322 @@
# Adding a New Benchmark
This guide walks you through adding a new simulation benchmark to LeRobot. Follow the steps in order and use the existing benchmarks as templates.
A benchmark in LeRobot is a set of [Gymnasium](https://gymnasium.farama.org/) environments that wrap a third-party simulator (like LIBERO or Meta-World) behind a standard `gym.Env` interface. The `lerobot-eval` CLI then runs evaluation uniformly across all benchmarks.
## Existing benchmarks at a glance
Before diving in, here is what is already integrated:
| Benchmark | Env file | Config class | Tasks | Action dim | Processor |
| -------------- | ------------------- | ------------------ | ------------------- | ------------ | ---------------------------- |
| LIBERO | `envs/libero.py` | `LiberoEnv` | 130 across 5 suites | 7 | `LiberoProcessorStep` |
| Meta-World | `envs/metaworld.py` | `MetaworldEnv` | 50 (MT50) | 4 | None |
| IsaacLab Arena | Hub-hosted | `IsaaclabArenaEnv` | Configurable | Configurable | `IsaaclabArenaProcessorStep` |
Use `src/lerobot/envs/libero.py` and `src/lerobot/envs/metaworld.py` as reference implementations.
## How it all fits together
### Data flow
During evaluation, data moves through four stages:
```
1. gym.Env ──→ raw observations (numpy dicts)
2. Preprocessing ──→ standard LeRobot keys + task description
(preprocess_observation in envs/utils.py, env.call("task_description"))
3. Processors ──→ env-specific then policy-specific transforms
(env_preprocessor, policy_preprocessor)
4. Policy ──→ select_action() ──→ action tensor
then reverse: policy_postprocessor → env_postprocessor → numpy action → env.step()
```
Most benchmarks only need to care about stage 1 (producing observations in the right format) and optionally stage 3 (if env-specific transforms are needed).
### Environment structure
`make_env()` returns a nested dict of vectorized environments:
```python
dict[str, dict[int, gym.vector.VectorEnv]]
# ^suite ^task_id
```
A single-task env (e.g. PushT) looks like `{"pusht": {0: vec_env}}`.
A multi-task benchmark (e.g. LIBERO) looks like `{"libero_spatial": {0: vec0, 1: vec1, ...}, ...}`.
### How evaluation runs
All benchmarks are evaluated the same way by `lerobot-eval`:
1. `make_env()` builds the nested `{suite: {task_id: VectorEnv}}` dict.
2. `eval_policy_all()` iterates over every suite and task.
3. For each task, it runs `n_episodes` rollouts via `rollout()`.
4. Results are aggregated hierarchically: episode, task, suite, overall.
5. Metrics include `pc_success` (success rate), `avg_sum_reward`, and `avg_max_reward`.
The critical piece: your env must return `info["is_success"]` on every `step()` call. This is how the eval loop knows whether a task was completed.
## What your environment must provide
LeRobot does not enforce a strict observation schema. Instead it relies on a set of conventions that all benchmarks follow.
### Env attributes
Your `gym.Env` must set these attributes:
| Attribute | Type | Why |
| -------------------- | ----- | ---------------------------------------------------- |
| `_max_episode_steps` | `int` | `rollout()` uses this to cap episode length |
| `task_description` | `str` | Passed to VLA policies as a language instruction |
| `task` | `str` | Fallback identifier if `task_description` is not set |
### Success reporting
Your `step()` and `reset()` must include `"is_success"` in the `info` dict:
```python
info = {"is_success": True} # or False
return observation, reward, terminated, truncated, info
```
### Observations
The simplest approach is to map your simulator's outputs to the standard keys that `preprocess_observation()` already understands. Do this inside your `gym.Env` (e.g. in a `_format_raw_obs()` helper):
| Your env should output | LeRobot maps it to | What it is |
| ------------------------- | -------------------------- | ------------------------------------- |
| `"pixels"` (single array) | `observation.image` | Single camera image, HWC uint8 |
| `"pixels"` (dict) | `observation.images.<cam>` | Multiple cameras, each HWC uint8 |
| `"agent_pos"` | `observation.state` | Proprioceptive state vector |
| `"environment_state"` | `observation.env_state` | Full environment state (e.g. PushT) |
| `"robot_state"` | `observation.robot_state` | Nested robot state dict (e.g. LIBERO) |
If your simulator uses different key names, you have two options:
1. **Recommended:** Rename them to the standard keys inside your `gym.Env` wrapper.
2. **Alternative:** Write an env processor to transform observations after `preprocess_observation()` runs (see step 4 below).
### Actions
Actions are continuous numpy arrays in a `gym.spaces.Box`. The dimensionality depends on your benchmark (7 for LIBERO, 4 for Meta-World, etc.). Policies adapt to different action dimensions through their `input_features` / `output_features` config.
### Feature declaration
Each `EnvConfig` subclass declares two dicts that tell the policy what to expect:
- `features` — maps feature names to `PolicyFeature(type, shape)` (e.g. action dim, image shape).
- `features_map` — maps raw observation keys to LeRobot convention keys (e.g. `"agent_pos"` to `"observation.state"`).
## Step by step
<Tip>
At minimum, you need two files: a **gym.Env wrapper** and an **EnvConfig
subclass** with a `create_envs()` override. Everything else is optional or
documentation. No changes to `factory.py` are needed.
</Tip>
### Checklist
| File | Required | Why |
| ---------------------------------------- | -------- | ------------------------------------------------------------ |
| `src/lerobot/envs/<benchmark>.py` | Yes | Wraps the simulator as a standard gym.Env |
| `src/lerobot/envs/configs.py` | Yes | Registers your benchmark and its `create_envs()` for the CLI |
| `src/lerobot/processor/env_processor.py` | Optional | Custom observation/action transforms |
| `src/lerobot/envs/utils.py` | Optional | Only if you need new raw observation keys |
| `pyproject.toml` | Yes | Declares benchmark-specific dependencies |
| `docs/source/<benchmark>.mdx` | Yes | User-facing documentation page |
| `docs/source/_toctree.yml` | Yes | Adds your page to the docs sidebar |
### 1. The gym.Env wrapper (`src/lerobot/envs/<benchmark>.py`)
Create a `gym.Env` subclass that wraps the third-party simulator:
```python
class MyBenchmarkEnv(gym.Env):
metadata = {"render_modes": ["rgb_array"], "render_fps": <fps>}
def __init__(self, task_suite, task_id, ...):
super().__init__()
self.task = <task_name_string>
self.task_description = <natural_language_instruction>
self._max_episode_steps = <max_steps>
self.observation_space = spaces.Dict({...})
self.action_space = spaces.Box(low=..., high=..., shape=(...,), dtype=np.float32)
def reset(self, seed=None, **kwargs):
... # return (observation, info) — info must contain {"is_success": False}
def step(self, action: np.ndarray):
... # return (obs, reward, terminated, truncated, info) — info must contain {"is_success": <bool>}
def render(self):
... # return RGB image as numpy array
def close(self):
...
```
**GPU-based simulators (e.g. MuJoCo with EGL rendering):** If your simulator allocates GPU/EGL contexts during `__init__`, defer that allocation to a `_ensure_env()` helper called on first `reset()`/`step()`. This avoids inheriting stale GPU handles when `AsyncVectorEnv` spawns worker processes. See `LiberoEnv._ensure_env()` for the pattern.
Also provide a factory function that returns the nested dict structure:
```python
def create_mybenchmark_envs(
task: str,
n_envs: int,
gym_kwargs: dict | None = None,
env_cls: type | None = None,
) -> dict[str, dict[int, Any]]:
"""Create {suite_name: {task_id: VectorEnv}} for MyBenchmark."""
...
```
See `create_libero_envs()` (multi-suite, multi-task) and `create_metaworld_envs()` (difficulty-grouped tasks) for reference.
### 2. The config (`src/lerobot/envs/configs.py`)
Register a config dataclass so users can select your benchmark with `--env.type=<name>`. Each config owns its environment creation and processor logic via two methods:
- **`create_envs(n_envs, use_async_envs)`** — Returns `{suite: {task_id: VectorEnv}}`. The base class default uses `gym.make()` for single-task envs. Multi-task benchmarks override this.
- **`get_env_processors()`** — Returns `(preprocessor, postprocessor)`. The base class default returns identity (no-op) pipelines. Override if your benchmark needs observation/action transforms.
```python
@EnvConfig.register_subclass("<benchmark_name>")
@dataclass
class MyBenchmarkEnvConfig(EnvConfig):
task: str = "<default_task>"
fps: int = <fps>
obs_type: str = "pixels_agent_pos"
features: dict[str, PolicyFeature] = field(default_factory=lambda: {
ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(<action_dim>,)),
})
features_map: dict[str, str] = field(default_factory=lambda: {
ACTION: ACTION,
"agent_pos": OBS_STATE,
"pixels": OBS_IMAGE,
})
def __post_init__(self):
... # populate features based on obs_type
@property
def gym_kwargs(self) -> dict:
return {"obs_type": self.obs_type, "render_mode": self.render_mode}
def create_envs(self, n_envs: int, use_async_envs: bool = True):
"""Override for multi-task benchmarks or custom env creation."""
from lerobot.envs.<benchmark> import create_<benchmark>_envs
return create_<benchmark>_envs(task=self.task, n_envs=n_envs, ...)
def get_env_processors(self):
"""Override if your benchmark needs observation/action transforms."""
from lerobot.processor import PolicyProcessorPipeline
from lerobot.processor.env_processor import MyBenchmarkProcessorStep
return (
PolicyProcessorPipeline(steps=[MyBenchmarkProcessorStep()]),
PolicyProcessorPipeline(steps=[]),
)
```
Key points:
- The `register_subclass` name is what users pass on the CLI (`--env.type=<name>`).
- `features` tells the policy what the environment produces.
- `features_map` maps raw observation keys to LeRobot convention keys.
- **No changes to `factory.py` needed** — the factory delegates to `cfg.create_envs()` and `cfg.get_env_processors()` automatically.
### 3. Env processor (optional — `src/lerobot/processor/env_processor.py`)
Only needed if your benchmark requires observation transforms beyond what `preprocess_observation()` handles (e.g. image flipping, coordinate conversion). Define the processor step here and return it from `get_env_processors()` in your config (see step 2):
```python
@dataclass
@ProcessorStepRegistry.register(name="<benchmark>_processor")
class MyBenchmarkProcessorStep(ObservationProcessorStep):
def _process_observation(self, observation):
processed = observation.copy()
# your transforms here
return processed
def transform_features(self, features):
return features # update if shapes change
def observation(self, observation):
return self._process_observation(observation)
```
See `LiberoProcessorStep` for a full example (image rotation, quaternion-to-axis-angle conversion).
### 4. Dependencies (`pyproject.toml`)
Add a new optional-dependency group:
```toml
mybenchmark = ["my-benchmark-pkg==1.2.3", "lerobot[scipy-dep]"]
```
Pinning rules:
- **Always pin** benchmark packages to exact versions for reproducibility (e.g. `metaworld==3.0.0`).
- **Add platform markers** when needed (e.g. `; sys_platform == 'linux'`).
- **Pin fragile transitive deps** if known (e.g. `gymnasium==1.1.0` for Meta-World).
- **Document constraints** in your benchmark doc page.
Users install with:
```bash
pip install -e ".[mybenchmark]"
```
### 5. Documentation (`docs/source/<benchmark>.mdx`)
Write a user-facing page following the template in the next section. See `docs/source/libero.mdx` and `docs/source/metaworld.mdx` for full examples.
### 6. Table of contents (`docs/source/_toctree.yml`)
Add your benchmark to the "Benchmarks" section:
```yaml
- sections:
- local: libero
title: LIBERO
- local: metaworld
title: Meta-World
- local: envhub_isaaclab_arena
title: NVIDIA IsaacLab Arena Environments
- local: <your_benchmark>
title: <Your Benchmark Name>
title: "Benchmarks"
```
## Verifying your integration
After completing the steps above, confirm that everything works:
1. **Install** — `pip install -e ".[mybenchmark]"` and verify the dependency group installs cleanly.
2. **Smoke test env creation** — call `make_env()` with your config in Python, check that the returned dict has the expected `{suite: {task_id: VectorEnv}}` shape, and that `reset()` returns observations with the right keys.
3. **Run a full eval** — `lerobot-eval --env.type=<name> --env.task=<task> --eval.n_episodes=1 --policy.path=<any_compatible_policy>` to exercise the full pipeline end-to-end. (`batch_size` defaults to auto-tuning based on CPU cores; pass `--eval.batch_size=1` to force a single environment.)
4. **Check success detection** — verify that `info["is_success"]` flips to `True` when the task is actually completed. This is what the eval loop uses to compute success rates.
## Writing a benchmark doc page
Each benchmark `.mdx` page should include:
- **Title and description** — 1-2 paragraphs on what the benchmark tests and why it matters.
- **Links** — paper, GitHub repo, project website (if available).
- **Overview image or GIF.**
- **Available tasks** — table of task suites with counts and brief descriptions.
- **Installation** — `pip install -e ".[<benchmark>]"` plus any extra steps (env vars, system packages).
- **Evaluation** — recommended `lerobot-eval` command with `n_episodes` for reproducible results. `batch_size` defaults to auto; only specify it if needed. Include single-task and multi-task examples if applicable.
- **Policy inputs and outputs** — observation keys with shapes, action space description.
- **Recommended evaluation episodes** — how many episodes per task is standard.
- **Training** — example `lerobot-train` command.
- **Reproducing published results** — link to pretrained model, eval command, results table (if available).
See `docs/source/libero.mdx` and `docs/source/metaworld.mdx` for complete examples.
docs/source/annotation_pipeline.mdx
+291
View File
@@ -0,0 +1,291 @@
# Annotation Pipeline
`lerobot-annotate` watches each episode's video with a vision-language
model (VLM) and writes natural-language annotations back into your
dataset. It fills the two language columns from the
[Language Columns and Recipes](./language_and_recipes) page —
`language_persistent` and `language_events` — straight into
`data/chunk-*/file-*.parquet`.
In short: point it at a LeRobot dataset, and it adds subtasks, plans,
memory, interjections, speech, and visual Q&A that a policy can be
trained on.
## How it fits together
```text
your dataset lerobot-annotate
(LeRobot v3.1)
│
▼
┌─────────────────────────────────────────────────────┐
│ read episodes │
└──────────────────────────┬──────────────────────────┘
│
┌────────────────────┼────────────────────┐
▼ ▼ ▼
┌──────────┐ ┌───────────────┐ ┌──────────┐ one shared Qwen-VL
│ plan │ │ interjections │ │ vqa │ ◀── server (vLLM, OpenAI
└────┬─────┘ └───────┬───────┘ └────┬─────┘ API) drives all three
└────────────────────┼─────────────────────┘
│ each module stages raw JSONL
▼ into .annotate_staging/
┌─────────────────┐
│ validator │ ◀── checks everything
└────────┬────────┘
▼
┌─────────────────┐
│ writer │
└────────┬────────┘
▼
data/chunk-*/file-*.parquet
(+ meta/info.json tools)
```
Three modules (`plan`, `interjections`, `vqa`) all talk to **one** shared
VLM. Each module stages its output to disk, a validator checks it, and a
single writer rewrites the dataset shards in place.
## What the pipeline produces
Each module emits a few kinds of annotation ("styles"), routed to one of
the two language columns:
| Style / atom | Column | Module |
| ------------------------------------------- | --------------------- | --------------- |
| `subtask` (Pi0.7-style "how, not what") | `language_persistent` | `plan` |
| `plan` (initial + refresh on interjection) | `language_persistent` | `plan` |
| `memory` (MEM-style compression) | `language_persistent` | `plan` |
| `task_aug` (rephrasings of the task) | `language_persistent` | `plan` |
| `interjection` | `language_events` | `interjections` |
| speech tool-call atom (`style=null`, `say`) | `language_events` | `interjections` |
| `vqa` (user / assistant pair) | `language_events` | `vqa` |
### How subtasks are generated
The `plan` module doesn't ask the VLM for subtasks in one shot. Instead
it uses a two-step **describe → segment** flow:
1. **Describe** — the VLM narrates only what it actually sees in the
chosen camera (no guessing about the task).
2. **Segment** — that description is fed back in, and the VLM splits the
episode into consecutive atomic subtasks.
Both passes see the episode as **timestamped contact sheets** — frames
sampled at `frames_per_second` (0.5s by default) and packed into JPEG
grids with each frame's time burned into its corner, so the VLM cites
exact boundary times directly. This is far cheaper in vision tokens than
one image per frame, so the sampling can stay dense; episodes longer than
`max_frames_per_prompt` are split into windows at the same density and
merged. Both prompts also carry a causal **event-boundary** definition (a
new event starts when an object becomes held / is released / reaches a new
location / a lid changes state / contents move) to sharpen where cuts land.
The resulting spans are then stitched into a gap-free, full-episode
cover, so **every frame has exactly one active subtask**. See
[`run_hf_job.py`](https://github.com/huggingface/lerobot/blob/main/examples/annotations/run_hf_job.py)
for the production settings (single camera, timestamped contact sheets,
auto-windowed subtask generation).
### Tools
The writer does **not** add a `tools` column to the parquet. The tool
catalog lives in `meta/info.json["tools"]` instead (see [Tools](./tools)).
After every run, the pipeline makes sure the canonical `say` schema is in
that list, keeping any tools you declared beforehand.
Want to add your own tool? Edit `meta/info.json["tools"]` directly — the
pipeline preserves whatever is already there. That makes the tool visible
to the chat template, so the model can learn to _generate_ the call. The
runtime layer that actually _executes_ a generated call (the `Tool`
protocol / `TOOL_REGISTRY` under `src/lerobot/tools/`) is not part of
this PR — the [Tools](./tools) doc marks those pieces as
not-yet-implemented.
## Running on Hugging Face Jobs
Annotation runs on [Hugging Face Jobs](https://huggingface.co/docs/hub/en/jobs).
The repo ships a launcher script you copy and tweak for your dataset:
```bash
HF_TOKEN=hf_... uv run python examples/annotations/run_hf_job.py
```
[`run_hf_job.py`](https://github.com/huggingface/lerobot/blob/main/examples/annotations/run_hf_job.py)
starts a single-GPU `h200` job (bump it to `h200x4` for big datasets)
that:
1. installs `lerobot` (from `main`) plus the annotation extras,
2. boots one vLLM server per GPU (using the `vllm/vllm-openai` image) and
drives it over the OpenAI-compatible API,
3. runs the `plan` / `interjections` / `vqa` modules across the dataset
with `lerobot-annotate`,
4. with `--push_to_hub=true`, uploads the result to `--new_repo_id` (or
back to `--repo_id` in place if you leave that unset).
To use a different dataset, model, or hub repo, edit the `CMD` block in
the script. Every flag there maps directly to a `lerobot-annotate` flag
(run `lerobot-annotate --help` for the full list).
## Key options
These are the flags you'll reach for most often. Run
`lerobot-annotate --help` for everything else; the defaults are tuned for
short manipulation episodes.
### Dataset in / out
| Flag | Default | What it does |
| ----------------- | ------- | ----------------------------------------------------------------------- |
| `--repo_id` | — | Hub dataset to annotate (downloaded if `--root` unset). |
| `--root` | — | Annotate a local dataset directory instead. |
| `--new_repo_id` | — | Push the result to a new repo (leaves the source repo untouched). |
| `--push_to_hub` | `false` | Upload after annotating (to `--new_repo_id`, else back to `--repo_id`). |
| `--only_episodes` | all | Annotate just these episode indices (handy for a test run). |
| `--seed` | `1729` | Seeds the RNGs that pick interjection timestamps + VQA question types. |
### Which modules run
Every module is on by default and can be toggled independently (set to
`false` to skip it, e.g. to iterate on one module at a time):
| Flag | Default | Turns off |
| ------------------------- | ------- | ----------------------------------- |
| `--plan.enabled` | `true` | subtasks + plan + memory + task_aug |
| `--interjections.enabled` | `true` | interjections + speech atoms |
| `--vqa.enabled` | `true` | the VQA pairs |
### The VLM (`--vlm.*`)
| Flag | Default | What it does |
| -------------------------- | ------------------ | ----------------------------------------------------------------------------------- |
| `--vlm.model_id` | `Qwen/Qwen3.6-27B` | The model to serve and prompt. |
| `--vlm.camera_key` | first `images.*` | Which camera every prompt is grounded on. |
| `--vlm.serve_command` | auto | The exact `vllm serve …` command (set TP size, GPU memory, `--max-model-len` here). |
| `--vlm.parallel_servers` | `1` | Independent servers for round-robin routing (one per GPU). |
| `--vlm.num_gpus` | `0` | GPUs per server (`0` = one each). |
| `--vlm.client_concurrency` | `16` | In-flight requests across all servers. |
| `--vlm.max_new_tokens` | `512` | Generation cap per call. |
| `--vlm.temperature` | `0.2` | Sampling temperature. |
### Subtasks / plan / memory (`--plan.*`)
| Flag | Default | What it does |
| ------------------------------- | ---------- | ------------------------------------------------------------------------------------------------------------------------- |
| `--plan.frames_per_second` | `2.0` | Frame sampling rate for the contact sheets (`2.0` = one frame every 0.5s). |
| `--plan.max_frames_per_prompt` | `60` | Frame budget per VLM call. Episodes whose sampling exceeds this are auto-windowed at the same density, then stitched. |
| `--plan.contact_sheet_columns` | `5` | Columns per contact-sheet grid (`contact_sheet_frames_per_sheet` tiles, time row-major). |
| `--plan.plan_max_steps` | `8` | Upper bound on subtasks per episode. |
| `--plan.subtask_describe_first` | `true` | Run the describe→segment grounding pass (best subtask quality; +1 call/episode). |
| `--plan.emit_plan` | `true` | Emit the numbered `plan` rows (`false` = subtasks + memory only). |
| `--plan.emit_memory` | `true` | Emit the `memory` rows (`false` = subtasks + plan only); symmetric to `emit_plan`. |
| `--plan.n_task_rephrasings` | `10` | How many `task_aug` rephrasings to emit (`0` disables). |
| `--plan.derive_task_from_video` | `if_short` | Use the dataset task as-is (`off`), only when it's missing/short (`if_short`), or always re-derive from video (`always`). |
### Interjections + VQA
| Flag | Default | What it does |
| ----------------------------------------------- | ------- | ---------------------------------------------------------- |
| `--interjections.max_interjections_per_episode` | `3` | Cap on interjection/speech pairs per episode. |
| `--vqa.vqa_emission_hz` | `1.0` | How often VQA pairs are emitted. |
| `--vqa.restrict_to_default_camera` | `false` | Ground VQA only on `--vlm.camera_key` (else every camera). |
| `--executor.episode_parallelism` | `16` | Episodes processed concurrently within each phase. |
## Contributing new modules
The pipeline is built to grow, and **contributions are very welcome** —
a brand-new module (say, trajectory traces or affordances), a new prompt
template, a smarter grounding flow, or quality fixes to the existing
`plan` / `interjections` / `vqa` modules.
Every module lives under
`src/lerobot/annotations/steerable_pipeline/modules/`, shares the VLM
client and the keyframe cache, writes its raw output to the staging
tree, and plugs into the executor as its own phase. Got an idea? Open an
issue or PR on [the repo](https://github.com/huggingface/lerobot).
## How recipes consume the output
The annotations are meant to be read by recipes (see
[Language Columns and Recipes](./language_and_recipes)). Typically:
- low-level / high-level / memory-update branches read
`subtask` / `plan` / `memory` from `language_persistent`.
- an interjection-response branch reads `interjection` events plus the
paired speech atom (merged into one assistant turn via `tool_calls_from`)
and the matching `plan` refresh at the same timestamp.
- a VQA branch reads the `(vqa, user)` and `(vqa, assistant)` pairs from
`language_events`.
## Why state and events are split
Two ideas shape the design:
1. **Persistent state vs. exact events.** Persistent rows (`subtask`,
`plan`, `memory`) apply to the whole episode and answer "what's true
right now?". Event rows (`interjection`, `vqa`, speech) appear only on
the one frame whose timestamp matches. Timestamps are copied straight
from the source parquet — never recomputed in floating point.
2. **One VLM pass.** All three modules share a single VLM client (the
OpenAI-compatible client talking to the job's vLLM server), so you pay
for one model load per dataset, not three.
## Re-running a single module
Each module stages its raw output to
`<root>/.annotate_staging/episode_{N:06d}/<module>.jsonl`. This makes
prompt iteration cheap: re-running one module overwrites only its own
JSONL, then the writer recomposes the final parquet. Disable modules you
don't want with `--plan.enabled=false` (and likewise
`--interjections.enabled` / `--vqa.enabled`) to test one at a time.
## What the validator checks
Before the writer runs, `StagingValidator` confirms:
- every event row lands exactly on a real frame timestamp;
- no speech / interjection pairs are left orphaned;
- `plan` is refreshed at every interjection timestamp;
- `memory` rows fall on subtask boundaries (a warning, not an error);
- each VQA assistant `content` is valid JSON in one of the
bbox / keypoint / count / attribute / spatial shapes;
- every row goes to the column chosen by `column_for_style(style)`.
Any error aborts the writer. Pass `--skip_validation=true` to override
while debugging.
## Where each module's ideas come from
- **`plan` — subtasks.** Hi Robot ([Shi 2025](https://arxiv.org/abs/2502.19417))
for atom granularity ("pick up one piece of lettuce", "place bowl to
box"); Pi0.7 ([Physical Intelligence 2025](https://pi.website/pi07))
for "how, not what" detail.
- **`plan` — memory.** MEM ([Torne 2026](https://arxiv.org/abs/2603.03596)):
keep only the minimal relevant information — preserve outcomes, drop
specific attributes.
- **`interjections`.** Hi Robot's scenario taxonomy: negative task,
situated correction, specific constraint, preference. Speech is a
tool-call-only atom
(`tool_calls=[{type:function, function:{name:"say", arguments:{text:...}}}]`).
- **`vqa`.** ECoT ([Zawalski 2024](https://arxiv.org/abs/2407.08693)) for
grounded features (pixel bounding boxes `[x_min, y_min, x_max, y_max]`,
keypoints) and Steerable VLA Policies
([Zhao 2025](https://arxiv.org/abs/2509.07626)) for multi-abstraction
grounding. Pi0.7 also grounds answers across abstraction levels.
When improving a module, tweak its prompt template in
`src/lerobot/annotations/steerable_pipeline/prompts/` rather than
rewriting from scratch.
## Roughly how much it costs
Per episode, the pipeline makes about `max_steps` plan calls,
`max_interjections_per_episode` interjection calls, and
`vqa_emission_hz × episode_seconds` VQA calls. With the defaults (8
subtasks, 1 interjection, 1 Hz × 3 pairs) on a 30-second episode, that's
~50 VLM calls.
Storage stays small: `language_persistent` is at most tens of KB per
episode (parquet dictionary-encodes the one entry that repeats across
frames), and `language_events` is empty on most frames — its size scales
with the number of emissions, not `num_frames × num_emissions`.
docs/source/async.mdx
+313
View File
@@ -0,0 +1,313 @@
# Asynchronous Inference
With our [SmolVLA](https://huggingface.co/papers/2506.01844) we introduced a new way to run inference on real-world robots, **decoupling action prediction from action execution**.
In this tutorial, we'll show how to use asynchronous inference (_async inference_) using a finetuned version of SmolVLA, and all the policies supported by LeRobot.
**Try async inference with all the policies** supported by LeRobot!
**What you'll learn:**
1. Why asynchronous inference matters and how it compares to, more traditional, sequential inference.
2. How to spin-up a `PolicyServer` and connect a `RobotClient` from the same machine, and even over the network.
3. How to tune key parameters (`actions_per_chunk`, `chunk_size_threshold`) for your robot and policy.
If you get stuck, hop into our [Discord community](https://discord.gg/s3KuuzsPFb)!
In a nutshell: with _async inference_, your robot keeps acting while the policy server is already busy computing the next chunk of actions---eliminating "wait-for-inference" lags and unlocking smoother, more reactive behaviours.
This is fundamentally different from synchronous inference (sync), where the robot stays idle while the policy computes the next chunk of actions.
---
## Getting started with async inference
You can read more information on asynchronous inference in our [blogpost](https://huggingface.co/blog/async-robot-inference). This guide is designed to help you quickly set up and run asynchronous inference in your environment.
First, install `lerobot` with the `async` tag, to install the extra dependencies required to run async inference.
```shell
pip install -e ".[async]"
```
Then, spin up a policy server (in one terminal, or in a separate machine) specifying the host address and port for the client to connect to.
You can spin up a policy server running:
```shell
python -m lerobot.async_inference.policy_server \
--host=127.0.0.1 \
--port=8080
```
This will start a policy server listening on `127.0.0.1:8080` (`localhost`, port 8080). At this stage, the policy server is empty, as all information related to which policy to run and with which parameters are specified during the first handshake with the client. Spin up a client with:
```shell
python -m lerobot.async_inference.robot_client \
--server_address=127.0.0.1:8080 \ # SERVER: the host address and port of the policy server
--robot.type=so100_follower \ # ROBOT: your robot type
--robot.port=/dev/tty.usbmodem585A0076841 \ # ROBOT: your robot port
--robot.id=follower_so100 \ # ROBOT: your robot id, to load calibration file
--robot.cameras="{ laptop: {type: opencv, index_or_path: 0, width: 1920, height: 1080, fps: 30}, phone: {type: opencv, index_or_path: 0, width: 1920, height: 1080, fps: 30}}" \ # POLICY: the cameras used to acquire frames, with keys matching the keys expected by the policy
--task="dummy" \ # POLICY: The task to run the policy on (`Fold my t-shirt`). Not necessarily defined for all policies, such as `act`
--policy_type=your_policy_type \ # POLICY: the type of policy to run (smolvla, act, etc)
--pretrained_name_or_path=user/model \ # POLICY: the model name/path on server to the checkpoint to run (e.g., lerobot/smolvla_base)
--policy_device=mps \ # POLICY: the device to run the policy on, on the server (cuda, mps, xpu, cpu)
--actions_per_chunk=50 \ # POLICY: the number of actions to output at once
--chunk_size_threshold=0.5 \ # CLIENT: the threshold for the chunk size before sending a new observation to the server
--aggregate_fn_name=weighted_average \ # CLIENT: the function to aggregate actions on overlapping portions
--debug_visualize_queue_size=True # CLIENT: whether to visualize the queue size at runtime
```
In summary, you need to specify instructions for:
- `SERVER`: the address and port of the policy server
- `ROBOT`: the type of robot to connect to, the port to connect to, and the local `id` of the robot
- `POLICY`: the type of policy to run, and the model name/path on server to the checkpoint to run. You also need to specify which device should the sever be using, and how many actions to output at once (capped at the policy max actions value).
- `CLIENT`: the threshold for the chunk size before sending a new observation to the server, and the function to aggregate actions on overlapping portions. Optionally, you can also visualize the queue size at runtime, to help you tune the `CLIENT` parameters.
Importantly,
- `actions_per_chunk` and `chunk_size_threshold` are key parameters to tune for your setup.
- `aggregate_fn_name` is the function to aggregate actions on overlapping portions. You can either add a new one to a registry of functions, or add your own in `robot_client.py` (see [here](NOTE:addlinktoLOC))
- `debug_visualize_queue_size` is a useful tool to tune the `CLIENT` parameters.
## Done! You should see your robot moving around by now 😉
## Async vs. synchronous inference
Synchronous inference relies on interleaving action chunk prediction and action execution. This inherently results in _idle frames_, frames where the robot awaits idle the policy's output: a new action chunk.
In turn, inference is plagued by evident real-time lags, where the robot simply stops acting due to the lack of available actions.
With robotics models increasing in size, this problem risks becoming only more severe.
<p align="center">
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/async-inference/sync.png"
width="80%"
></img>
</p>
<p align="center">
<i>Synchronous inference</i> makes the robot idle while the policy is
computing the next chunk of actions.
</p>
To overcome this, we design async inference, a paradigm where action planning and execution are decoupled, resulting in (1) higher adaptability and, most importantly, (2) no idle frames.
Crucially, with async inference, the next action chunk is computed _before_ the current one is exhausted, resulting in no idleness.
Higher adaptability is ensured by aggregating the different action chunks on overlapping portions, obtaining an up-to-date plan and a tighter control loop.
<p align="center">
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/async-inference/async.png"
width="80%"
></img>
</p>
<p align="center">
<i>Asynchronous inference</i> results in no idleness because the next chunk is
computed before the current chunk is exhausted.
</p>
---
## Start the Policy Server
Policy servers are wrappers around a `PreTrainedPolicy` interfacing them with observations coming from a robot client.
Policy servers are initialized as empty containers which are populated with the requested policy specified in the initial handshake between the robot client and the policy server.
As such, spinning up a policy server is as easy as specifying the host address and port. If you're running the policy server on the same machine as the robot client, you can use `localhost` as the host address.
<hfoptions id="start_policy_server">
<hfoption id="Command">
```bash
python -m lerobot.async_inference.policy_server \
--host=127.0.0.1 \
--port=8080
```
</hfoption>
<hfoption id="API example">
<!-- prettier-ignore-start -->
```python
from lerobot.async_inference.configs import PolicyServerConfig
from lerobot.async_inference.policy_server import serve
config = PolicyServerConfig(
host="localhost",
port=8080,
)
serve(config)
```
<!-- prettier-ignore-end -->
</hfoption>
</hfoptions>
This listens on `localhost:8080` for an incoming connection from the associated`RobotClient`, which will communicate which policy to run during the first client-server handshake.
---
## Launch the Robot Client
`RobotClient` is a wrapper around a `Robot` instance, which `RobotClient` connects to the (possibly remote) `PolicyServer`.
The `RobotClient` streams observations to the `PolicyServer`, and receives action chunks obtained running inference on the server (which we assume to have better computational resources than the robot controller).
<hfoptions id="start_robot_client">
<hfoption id="Command">
```bash
python -m lerobot.async_inference.robot_client \
--server_address=127.0.0.1:8080 \ # SERVER: the host address and port of the policy server
--robot.type=so100_follower \ # ROBOT: your robot type
--robot.port=/dev/tty.usbmodem585A0076841 \ # ROBOT: your robot port
--robot.id=follower_so100 \ # ROBOT: your robot id, to load calibration file
--robot.cameras="{ laptop: {type: opencv, index_or_path: 0, width: 1920, height: 1080, fps: 30}, phone: {type: opencv, index_or_path: 0, width: 1920, height: 1080, fps: 30}}" \ # POLICY: the cameras used to acquire frames, with keys matching the keys expected by the policy
--task="dummy" \ # POLICY: The task to run the policy on (`Fold my t-shirt`). Not necessarily defined for all policies, such as `act`
--policy_type=your_policy_type \ # POLICY: the type of policy to run (smolvla, act, etc)
--pretrained_name_or_path=user/model \ # POLICY: the model name/path on server to the checkpoint to run (e.g., lerobot/smolvla_base)
--policy_device=mps \ # POLICY: the device to run the policy on, on the server
--actions_per_chunk=50 \ # POLICY: the number of actions to output at once
--chunk_size_threshold=0.5 \ # CLIENT: the threshold for the chunk size before sending a new observation to the server
--aggregate_fn_name=weighted_average \ # CLIENT: the function to aggregate actions on overlapping portions
--debug_visualize_queue_size=True # CLIENT: whether to visualize the queue size at runtime
```
</hfoption>
<hfoption id="API example">
<!-- prettier-ignore-start -->
```python
import threading
from lerobot.robots.so_follower import SO100FollowerConfig
from lerobot.cameras.opencv import OpenCVCameraConfig
from lerobot.async_inference.configs import RobotClientConfig
from lerobot.async_inference.robot_client import RobotClient
from lerobot.async_inference.helpers import visualize_action_queue_size
# 1. Create the robot instance
"""Check out the cameras available in your setup by running `python lerobot/find_cameras.py`"""
# these cameras must match the ones expected by the policy
# check the config.json on the Hub for the policy you are using
camera_cfg = {
"top": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
"side": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30)
}
robot_cfg = SO100FollowerConfig(
port="/dev/tty.usbmodem585A0076841",
id="follower_so100",
cameras=camera_cfg
)
# 3. Create client configuration
client_cfg = RobotClientConfig(
robot=robot_cfg,
server_address="localhost:8080",
policy_device="mps",
client_device="cpu",
policy_type="smolvla",
pretrained_name_or_path="<user>/smolvla_async",
chunk_size_threshold=0.5,
actions_per_chunk=50, # make sure this is less than the max actions of the policy
)
# 4. Create and start client
client = RobotClient(client_cfg)
# 5. Specify the task
task = "Don't do anything, stay still"
if client.start():
# Start action receiver thread
action_receiver_thread = threading.Thread(target=client.receive_actions, daemon=True)
action_receiver_thread.start()
try:
# Run the control loop
client.control_loop(task)
except KeyboardInterrupt:
client.stop()
action_receiver_thread.join()
# (Optionally) plot the action queue size
visualize_action_queue_size(client.action_queue_size)
```
<!-- prettier-ignore-end -->
</hfoption>
</hfoptions>
The following two parameters are key in every setup:
<table>
<thead>
<tr>
<th>Hyperparameter</th>
<th>Default</th>
<th>What it does</th>
</tr>
</thead>
<tbody>
<tr>
<td>
<code>actions_per_chunk</code>
</td>
<td>50</td>
<td>
How many actions the policy outputs at once. Typical values: 10-50.
</td>
</tr>
<tr>
<td>
<code>chunk_size_threshold</code>
</td>
<td>0.7</td>
<td>
When the queue is ≤ 50% full, the client sends a fresh observation.
Value in [0, 1].
</td>
</tr>
</tbody>
</table>
<Tip>
Different values of `actions_per_chunk` and `chunk_size_threshold` do result
in different behaviours.
</Tip>
On the one hand, increasing the value of `actions_per_chunk` will result in reducing the likelihood of ending up with no actions to execute, as more actions will be available when the new chunk is computed.
However, larger values of `actions_per_chunk` might also result in less precise actions, due to the compounding errors consequent to predicting actions over longer timespans.
On the other hand, increasing the value of `chunk_size_threshold` will result in sending out to the `PolicyServer` observations for inference more often, resulting in a larger number of updates action chunks, overlapping on significant portions. This results in high adaptability, in the limit predicting one action chunk for each observation, which is in turn only marginally consumed while a new one is produced.
This option does also put more pressure on the inference pipeline, as a consequence of the many requests. Conversely, values of `chunk_size_threshold` close to 0.0 collapse to the synchronous edge case, whereby new observations are only sent out whenever the current chunk is exhausted.
We found the default values of `actions_per_chunk` and `chunk_size_threshold` to work well in the experiments we developed for the [SmolVLA paper](https://huggingface.co/papers/2506.01844), but recommend experimenting with different values to find the best fit for your setup.
### Tuning async inference for your setup
1. **Choose your computational resources carefully.** [PI0](https://huggingface.co/lerobot/pi0) occupies 14GB of memory at inference time, while [SmolVLA](https://huggingface.co/lerobot/smolvla_base) requires only ~2GB. You should identify the best computational resource for your use case keeping in mind smaller policies require less computational resources. The combination of policy and device used (CPU-intensive, using MPS, or the number of CUDA cores on a given NVIDIA GPU) directly impacts the average inference latency you should expect.
2. **Adjust your `fps` based on inference latency.** While the server generates a new action chunk, the client is not idle and is stepping through its current action queue. If the two processes happen at fundamentally different speeds, the client might end up with an empty queue. As such, you should reduce your fps if you consistently run out of actions in queue.
3. **Adjust `chunk_size_threshold`**.
- Values closer to `0.0` result in almost sequential behavior. Values closer to `1.0` → send observation every step (more bandwidth, relies on good world-model).
- We found values around 0.5-0.6 to work well. If you want to tweak this, spin up a `RobotClient` setting the `--debug_visualize_queue_size` to `True`. This will plot the action queue size evolution at runtime, and you can use it to find the value of `chunk_size_threshold` that works best for your setup.
<p align="center">
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/async-inference/queues.png"
width="80%"
></img>
</p>
<p align="center">
<i>
The action queue size is plotted at runtime when the
`--debug_visualize_queue_size` flag is passed, for various levels of
`chunk_size_threshold` (`g` in the SmolVLA paper).
</i>
</p>
---
## Conclusion
Asynchronous inference represents a significant advancement in real-time robotics control, addressing the fundamental challenge of inference latency that has long plagued robotics applications. Through this tutorial, you've learned how to implement a complete async inference pipeline that eliminates idle frames and enables smoother, more reactive robot behaviors.
**Key Takeaways:**
- **Paradigm Shift**: Async inference decouples action prediction from execution, allowing robots to continue acting while new action chunks are computed in parallel
- **Performance Benefits**: Eliminates "wait-for-inference" lags that are inherent in synchronous approaches, becoming increasingly important as policy models grow larger
- **Flexible Architecture**: The server-client design enables distributed computing, where inference can run on powerful remote hardware while maintaining real-time robot control
- **Tunable Parameters**: Success depends on properly configuring `actions_per_chunk` and `chunk_size_threshold` for your specific hardware, policy, and task requirements
- **Universal Compatibility**: Works with all LeRobot-supported policies, from lightweight ACT models to vision-language models like SmolVLA
Start experimenting with the default parameters, monitor your action queue sizes, and iteratively refine your setup to achieve optimal performance for your specific use case.
If you want to discuss this further, hop into our [Discord community](https://discord.gg/s3KuuzsPFb), or open an issue on our [GitHub repository](https://github.com/huggingface/lerobot/issues).
docs/source/backwardcomp.mdx
+151
View File
@@ -0,0 +1,151 @@
# Backward compatibility
## Policy Normalization Migration (PR #1452)
**Breaking Change**: LeRobot policies no longer have built-in normalization layers embedded in their weights. Normalization is now handled by external `PolicyProcessorPipeline` components.
### What changed?
| | Before PR #1452 | After PR #1452 |
| -------------------------- | ------------------------------------------------ | ------------------------------------------------------------ |
| **Normalization Location** | Embedded in model weights (`normalize_inputs.*`) | External `PolicyProcessorPipeline` components |
| **Model State Dict** | Contains normalization statistics | **Clean weights only** - no normalization parameters |
| **Usage** | `policy(batch)` handles everything | `preprocessor(batch)` → `policy(...)` → `postprocessor(...)` |
### Impact on existing models
- Models trained **before** PR #1452 have normalization embedded in their weights
- These models need migration to work with the new `PolicyProcessorPipeline` system
- The migration extracts normalization statistics and creates separate processor pipelines
### Migrating old models
Use the migration script to convert models with embedded normalization:
```shell
python src/lerobot/processor/migrate_policy_normalization.py \
--pretrained-path lerobot/act_aloha_sim_transfer_cube_human \
--push-to-hub \
--branch migrated
```
The script:
1. **Extracts** normalization statistics from model weights
2. **Creates** external preprocessor and postprocessor pipelines
3. **Removes** normalization layers from model weights
4. **Saves** clean model + processor pipelines
5. **Pushes** to Hub with automatic PR creation
### Using migrated models
```python
# New usage pattern (after migration)
from lerobot.policies import make_policy, make_pre_post_processors
# Load model and processors separately
policy = make_policy(config, ds_meta=dataset.meta)
preprocessor, postprocessor = make_pre_post_processors(
policy_cfg=config,
dataset_stats=dataset.meta.stats
)
# Process data through pipeline
processed_batch = preprocessor(raw_batch)
action = policy.select_action(processed_batch)
final_action = postprocessor(action)
```
## Hardware API redesign
PR [#777](https://github.com/huggingface/lerobot/pull/777) improves the LeRobot calibration but is **not backward-compatible**. Below is a overview of what changed and how you can continue to work with datasets created before this pull request.
### What changed?
| | Before PR #777 | After PR #777 |
| --------------------------------- | ------------------------------------------------- | ------------------------------------------------------------ |
| **Joint range** | Degrees `-180...180°` | **Normalised range** Joints: `–100...100` Gripper: `0...100` |
| **Zero position (SO100 / SO101)** | Arm fully extended horizontally | **In middle of the range for each joint** |
| **Boundary handling** | Software safeguards to detect ±180 ° wrap-arounds | No wrap-around logic needed due to mid-range zero |
---
### Impact on existing datasets
- Recorded trajectories created **before** PR #777 will replay incorrectly if loaded directly:
- Joint angles are offset and incorrectly normalized.
- Any models directly finetuned or trained on the old data will need their inputs and outputs converted.
### Using datasets made with the previous calibration system
We provide a migration example script for replaying an episode recorded with the previous calibration here: `examples/backward_compatibility/replay.py`.
Below we take you through the modifications that are done in the example script to make the previous calibration datasets work.
```diff
+ key = f"{name.removeprefix('main_')}.pos"
action[key] = action_array[i].item()
+ action["shoulder_lift.pos"] = -(action["shoulder_lift.pos"] - 90)
+ action["elbow_flex.pos"] -= 90
```
Let's break this down.
New codebase uses `.pos` suffix for the position observations and we have removed `main_` prefix:
<!-- prettier-ignore-start -->
```python
key = f"{name.removeprefix('main_')}.pos"
```
<!-- prettier-ignore-end -->
For `"shoulder_lift"` (id = 2), the 0 position is changed by -90 degrees and the direction is reversed compared to old calibration/code.
<!-- prettier-ignore-start -->
```python
action["shoulder_lift.pos"] = -(action["shoulder_lift.pos"] - 90)
```
<!-- prettier-ignore-end -->
For `"elbow_flex"` (id = 3), the 0 position is changed by -90 degrees compared to old calibration/code.
<!-- prettier-ignore-start -->
```python
action["elbow_flex.pos"] -= 90
```
<!-- prettier-ignore-end -->
To use degrees normalization we then set the `--robot.use_degrees` option to `true`.
```diff
python examples/backward_compatibility/replay.py \
--robot.type=so101_follower \
--robot.port=/dev/tty.usbmodem5A460814411 \
--robot.id=blue \
+ --robot.use_degrees=true \
--dataset.repo_id=my_dataset_id \
--dataset.episode=0
```
### Using policies trained with the previous calibration system
Policies output actions in the same format as the datasets (`torch.Tensors`). Therefore, the same transformations should be applied.
To find these transformations, we recommend to first try and and replay an episode of the dataset your policy was trained on using the section above.
Then, add these same transformations on your inference script (shown here in the `record.py` script):
```diff
action_values = predict_action(
observation_frame,
policy,
get_safe_torch_device(policy.config.device),
policy.config.use_amp,
task=single_task,
robot_type=robot.robot_type,
)
action = {key: action_values[i].item() for i, key in enumerate(robot.action_features)}
+ action["shoulder_lift.pos"] = -(action["shoulder_lift.pos"] - 90)
+ action["elbow_flex.pos"] -= 90
robot.send_action(action)
```
If you have questions or run into migration issues, feel free to ask them on [Discord](https://discord.gg/s3KuuzsPFb)
docs/source/bring_your_own_policies.mdx
+389
View File
@@ -0,0 +1,389 @@
# Adding a Policy
This guide walks you through implementing a custom policy and getting it to work with LeRobot's training, evaluation, and deployment tools. There are two paths:
- **Plugin (out-of-tree)** — ship your policy as a standalone `lerobot_policy_*` package. Faster, no PR required, easy to iterate. Right for experimentation, internal use, or when you want to publish independently.
- **In-tree (contributed to LeRobot)** — land your policy directly in `src/lerobot/policies/`. Requires a PR, but makes your policy a first-class citizen of the library.
The plugin route is usually the right starting point — promote to in-tree once the policy has stabilized and there's clear value in shipping it with the library.
Either way, the building blocks are the same: a configuration class, a policy class, and a processor factory. The first half of this guide covers those shared pieces; the second half covers the path-specific scaffolding ([Path A](#path-a-out-of-tree-plugin), [Path B](#path-b-contributing-in-tree)).
A note on tone: robot-learning is an actively evolving field, and "what a policy looks like" can shift with each new architecture. The conventions described here exist because they let `lerobot-train` and `lerobot-eval` work uniformly across very different models. When a new policy genuinely doesn't fit them, raise it (in your PR, or an issue) — the conventions are not sacred.
---
## Anatomy of a policy
Three building blocks make up every policy. The names below use `my_policy` as a placeholder — replace with your policy's name. That name is load-bearing: it must match the string you pass to `@PreTrainedConfig.register_subclass`, the `MyPolicy.name` class attribute, and the `make_<name>_pre_post_processors` factory function (more on each below).
### Configuration class
Inherit from [`PreTrainedConfig`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/configs/policies.py) and register your policy type. Here is a template — customize the parameters and methods as needed for your policy's architecture and training requirements.
```python
# configuration_my_policy.py
from dataclasses import dataclass, field
from lerobot.configs import PreTrainedConfig
from lerobot.optim import AdamWConfig
from lerobot.optim import CosineDecayWithWarmupSchedulerConfig
@PreTrainedConfig.register_subclass("my_policy")
@dataclass
class MyPolicyConfig(PreTrainedConfig):
"""Configuration class for MyPolicy.
Args:
n_obs_steps: Number of observation steps to use as input
horizon: Action prediction horizon
n_action_steps: Number of action steps to execute
hidden_dim: Hidden dimension for the policy network
# Add your policy-specific parameters here
"""
horizon: int = 50
n_action_steps: int = 50
hidden_dim: int = 256
optimizer_lr: float = 1e-4
optimizer_weight_decay: float = 1e-4
def __post_init__(self):
super().__post_init__()
if self.n_action_steps > self.horizon:
raise ValueError("n_action_steps cannot exceed horizon")
def validate_features(self) -> None:
"""Validate input/output feature compatibility.
Call this explicitly from your policy's __init__ — the base class does not.
"""
if not self.image_features:
raise ValueError("MyPolicy requires at least one image feature.")
if self.action_feature is None:
raise ValueError("MyPolicy requires 'action' in output_features.")
def get_optimizer_preset(self) -> AdamWConfig:
return AdamWConfig(lr=self.optimizer_lr, weight_decay=self.optimizer_weight_decay)
def get_scheduler_preset(self):
"""Return a LRSchedulerConfig from lerobot.optim, or None."""
return None
@property
def observation_delta_indices(self) -> list[int] | None:
"""Relative timestep offsets the dataset loader provides per observation.
Return `None` for single-frame policies. For temporal policies that consume
multiple past or future frames, return a list of offsets, e.g. `[-20, -10, 0, 10]` for
3 past frames at stride 10 and 1 future frame at stride 10.
"""
return None
@property
def action_delta_indices(self) -> list[int]:
"""Relative timestep offsets for the action chunk the dataset loader returns."""
return list(range(self.horizon))
@property
def reward_delta_indices(self) -> None:
return None
```
The string you pass to `@register_subclass` must match `MyPolicy.name` (next section) and is what users supply as `--policy.type` on the CLI. Default to `AdamW` from `lerobot.optim` for `get_optimizer_preset` unless you genuinely need otherwise.
### Policy class
Inherit from [`PreTrainedPolicy`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/pretrained.py) and set two class attributes — both are checked by `__init_subclass__`:
```python
# modeling_my_policy.py
import torch
import torch.nn as nn
from typing import Any
from lerobot.policies import PreTrainedPolicy
from lerobot.utils.constants import ACTION
from .configuration_my_policy import MyPolicyConfig
class MyPolicy(PreTrainedPolicy):
config_class = MyPolicyConfig # must match the string in @register_subclass
name = "my_policy"
def __init__(self, config: MyPolicyConfig, dataset_stats: dict[str, Any] = None):
super().__init__(config, dataset_stats)
config.validate_features() # not called automatically by the base class
self.config = config
self.model = ... # your nn.Module here
def reset(self):
"""Reset per-episode state. Called by lerobot-eval at the start of each episode."""
...
def get_optim_params(self) -> dict:
"""Return parameters to pass to the optimizer (e.g. with per-group lr/wd)."""
return {"params": self.parameters()}
def predict_action_chunk(self, batch: dict[str, torch.Tensor], **kwargs) -> torch.Tensor:
"""Return the full action chunk (B, chunk_size, action_dim) for the current observation."""
...
def select_action(self, batch: dict[str, torch.Tensor], **kwargs) -> torch.Tensor:
"""Return a single action for the current timestep (called every step at inference)."""
...
def forward(self, batch: dict[str, torch.Tensor]) -> tuple[torch.Tensor, dict | None]:
"""Compute the training loss.
Returns `(loss, output_dict)`. `output_dict` may be `None`; everything in it must be
logging-friendly Python natives (no tensors with gradients).
`batch["action_is_pad"]` is a bool mask of shape (B, horizon) that marks
timesteps padded because the episode ended before `horizon` steps; you
can exclude those from your loss.
"""
actions = batch[ACTION]
action_is_pad = batch.get("action_is_pad")
...
return loss, {"some_loss_component": some_loss_component.item()}
```
The methods called by the train/eval loops:
| Method | Used by | What it does |
| ----------------------------------------------------------------- | ----------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `reset() -> None` | `lerobot-eval` | Clear per-episode state at the start of each episode. |
| `select_action(batch, **kwargs) -> Tensor` | `lerobot-eval` | Return the next action `(B, action_dim)`. Called every step. |
| `predict_action_chunk(batch, **kwargs) -> Tensor` | the policy itself | Return an action chunk `(B, chunk_size, action_dim)`. Currently abstract on the base class — raise `NotImplementedError` if your policy doesn't chunk. |
| `forward(batch, reduction="mean") -> tuple[Tensor, dict \| None]` | `lerobot-train` | Return `(loss, output_dict)`. Accept `reduction="none"` if you want to support per-sample weighting. |
| `get_optim_params() -> dict` | the optimizer | Return `self.parameters()` for simple policies; return a named parameter dict for [multi-optimizer policies](https://github.com/huggingface/lerobot/blob/ecd38c50d7d15b4184cf42649ff1185ee2e11eeb/src/lerobot/policies/sac/modeling_sac.py#L61-L73). |
| `update() -> None` _(optional)_ | `lerobot-train` | Called after each optimizer step _if defined_. Use for EMA, target nets, replay buffers (TDMPC uses this). |
Batches are flat dictionaries keyed by the constants in [`lerobot.utils.constants`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/utils/constants.py): `OBS_STATE` (`observation.state.<motor>`), `OBS_IMAGES` (`observation.images.<camera>`), `OBS_LANGUAGE`, `ACTION`, etc. Reuse the constants — don't invent new prefixes.
### Processor functions
LeRobot uses `PolicyProcessorPipeline`s to normalize inputs and de-normalize outputs around your policy. For a concrete reference, see [`processor_act.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/act/processor_act.py) or [`processor_diffusion.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/diffusion/processor_diffusion.py).
```python
# processor_my_policy.py
from typing import Any
import torch
from lerobot.processor import PolicyAction, PolicyProcessorPipeline
def make_my_policy_pre_post_processors(
config,
dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
) -> tuple[
PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
PolicyProcessorPipeline[PolicyAction, PolicyAction],
]:
preprocessor = ... # build your PolicyProcessorPipeline for inputs
postprocessor = ... # build your PolicyProcessorPipeline for outputs
return preprocessor, postprocessor
```
**Important — function naming:** LeRobot discovers your processor by name. The function **must** be called `make_{policy_name}_pre_post_processors` (matching the string you passed to `@PreTrainedConfig.register_subclass`).
---
## Path A: Out-of-tree plugin
The fastest way to ship a policy: package it as a standalone Python distribution and install it alongside LeRobot. No PR required, you own the release cycle, and you can publish to PyPI under your own namespace.
### Package structure
Create a package with the prefix `lerobot_policy_` (IMPORTANT!) followed by your policy name:
```bash
lerobot_policy_my_policy/
├── pyproject.toml
└── src/
└── lerobot_policy_my_policy/
├── __init__.py
├── configuration_my_policy.py
├── modeling_my_policy.py
└── processor_my_policy.py
```
### `pyproject.toml`
```toml
[project]
name = "lerobot_policy_my_policy"
version = "0.1.0"
dependencies = [
# your policy-specific dependencies
]
requires-python = ">= 3.12"
[build-system]
build-backend = # your-build-backend
requires = # your-build-system
```
### Package `__init__.py`
Expose your classes in the package's `__init__.py` and guard against missing `lerobot`:
```python
# __init__.py
"""Custom policy package for LeRobot."""
try:
import lerobot # noqa: F401
except ImportError:
raise ImportError(
"lerobot is not installed. Please install lerobot to use this policy package."
)
from .configuration_my_policy import MyPolicyConfig
from .modeling_my_policy import MyPolicy
from .processor_my_policy import make_my_policy_pre_post_processors
__all__ = [
"MyPolicyConfig",
"MyPolicy",
"make_my_policy_pre_post_processors",
]
```
### Install and use
```bash
cd lerobot_policy_my_policy
pip install -e .
# Or install from PyPI if published
pip install lerobot_policy_my_policy
```
Once installed, your policy automatically integrates with LeRobot's training and evaluation tools:
```bash
lerobot-train \
--policy.type my_policy \
--env.type pusht \
--steps 200000
```
---
## Path B: Contributing in-tree
When your policy has stabilized and there's clear value in shipping it with the library, you can land it directly in LeRobot. Read the general [contribution guide](./contributing) and the [PR template](https://github.com/huggingface/lerobot/blob/main/.github/PULL_REQUEST_TEMPLATE.md) first — that's where you'll find the testing/quality expectations every PR has to meet (`pre-commit run -a`, `pytest`, the community-review rule, etc.). What's below is the policy-specific layer on top of that.
### In-tree layout
```
src/lerobot/policies/my_policy/
├── __init__.py # re-exports config + modeling + processor factory
├── configuration_my_policy.py # MyPolicyConfig + @register_subclass
├── modeling_my_policy.py # MyPolicy(PreTrainedPolicy)
├── processor_my_policy.py # make_my_policy_pre_post_processors
└── README.md # symlink → ../../../../docs/source/policy_my_policy_README.md
```
Two notes:
- The `README.md` next to the source is a **symlink** into `docs/source/policy_<name>_README.md` — the actual file lives under `docs/`. Existing policies (act, smolvla, diffusion, …) all do this; copy one of those symlinks. The policy README is conventionally minimal: paper link + BibTeX citation.
- The user-facing tutorial — what to install, how to train, hyperparameters, benchmark numbers — lives separately at `docs/source/<my_policy>.mdx` and is registered in `_toctree.yml` under "Policies".
The file names are load-bearing: the factory does lazy imports by name, and the processor is discovered by the `make_<policy_name>_pre_post_processors` convention.
### Wiring
Four places need to know about your policy. All by name.
1. **`policies/__init__.py`** — re-export `MyPolicyConfig` and add it to `__all__`. **Don't** re-export the modeling class; it loads lazily through the factory (so `import lerobot` stays fast).
2. **`factory.py:get_policy_class`** — add a branch returning `MyPolicy` from a lazy import.
3. **`factory.py:make_policy_config`** and **`factory.py:make_pre_post_processors`** — same idea, two more branches.
4. **`templates/lerobot_modelcard_template.md` and the root `README.md`** — the template is what `push_model_to_hub` renders into the model card of every checkpoint trained with your policy: add a one-line description of your policy in the `model_name` branches, map it in `policy_docs` so cards link to your MDX guide, and optionally add an architecture image to `diagrams`. Then add your policy to the models table in the root `README.md`, under the right category, linking to your doc page.
Mirror an existing policy that's structurally similar to yours; the diff is small.
### Heavy / optional dependencies
Most policies need a heavy backbone (transformers, diffusers, a specific VLM SDK). The convention is **two-step gating**: a `TYPE_CHECKING`-guarded import at module top, and a `require_package` runtime check in the constructor. [`modeling_diffusion.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/diffusion/modeling_diffusion.py) is the canonical reference:
```python
from typing import TYPE_CHECKING
from lerobot.utils.import_utils import _diffusers_available, require_package
if TYPE_CHECKING or _diffusers_available:
from diffusers.schedulers.scheduling_ddim import DDIMScheduler
else:
DDIMScheduler = None # keeps the symbol bindable at import time
class DiffusionPolicy(PreTrainedPolicy):
def __init__(self, config):
require_package("diffusers", extra="diffusion")
super().__init__(config)
...
```
This way:
- `import lerobot.policies` keeps working without the extra installed (the symbol is just bound to `None`).
- Type checkers see the real symbol.
- Instantiating the policy without the extra raises a clear `ImportError` pointing at `pip install 'lerobot[diffusion]'`.
Add a matching extra to [`pyproject.toml`](https://github.com/huggingface/lerobot/blob/main/pyproject.toml) `[project.optional-dependencies]` and include it in the `all` extra so `pip install 'lerobot[all]'` keeps installing everything.
### Benchmarks and a published checkpoint
A new policy is much easier to review — and far more useful — when it ships with a working checkpoint and at least one number you can reproduce.
**Pick at least one in-tree benchmark.** LeRobot ships sim benchmarks with per-benchmark Docker images (LIBERO, LIBERO-plus, Meta-World, RoboTwin 2.0, RoboCasa365, RoboCerebra, RoboMME, VLABench and more). Pick the one that matches your policy's modality — VLAs usually go to LIBERO or VLABench; image-only BC to LIBERO or Meta-World. The full list lives under [Benchmarks](./libero) in the docs sidebar.
**Push the checkpoint & processors** to the Hub under `lerobot/<policy>_<benchmark>` (or your namespace if you don't have write access; a maintainer can mirror it). Use `PreTrainedPolicy.push_model_to_hub` so the repo gets `config.json`, `model.safetensors`, and a model card.
**Report results in your policy's MDX**, with the exact `lerobot-eval` command and hardware so anyone can re-run:
```markdown
## Results
Evaluated on LIBERO with `lerobot/<policy>_libero`:
| Suite | Success rate | n_episodes |
| -------------- | -----------: | ---------: |
| libero_spatial | 87.5% | 50 |
| libero_object | 93.0% | 50 |
| libero_goal | 81.5% | 50 |
| libero_10 | 62.0% | 50 |
| **average** | **81.0%** | 200 |
Reproduce: `lerobot-eval --policy.path=lerobot/<policy>_libero --env.type=libero --env.task=libero_spatial --eval.n_episodes=50` (1× A100 40 GB).
```
Use `n_episodes ≥ 50` per suite for stable success-rate estimates.
If your policy is real-robot-only and no sim benchmark applies, swap the sim eval for: a public training dataset on the Hub, the `lerobot-train` command, the checkpoint, and a real-robot success rate over ≥10 episodes via `lerobot-rollout --policy.path=...`.
### PR checklist
The general expectations are in [`CONTRIBUTING.md`](https://github.com/huggingface/lerobot/blob/main/CONTRIBUTING.md) and the [PR template](https://github.com/huggingface/lerobot/blob/main/.github/PULL_REQUEST_TEMPLATE.md). On top of those, reviewers will look for:
- [ ] `MyPolicy` and `MyPolicyConfig` cover the surface above; `__init_subclass__` accepts the class.
- [ ] `factory.py` and `policies/__init__.py` are wired (lazy imports for modeling).
- [ ] `make_my_policy_pre_post_processors` follows the naming convention.
- [ ] Optional deps live behind a `[project.optional-dependencies]` extra and the `TYPE_CHECKING + require_package` guard.
- [ ] `tests/policies/` updated; backward-compat artifact committed & policy-specific tests.
- [ ] `src/lerobot/policies/<name>/README.md` symlinked into `docs/source/policy_<name>_README.md`; user-facing `docs/source/<name>.mdx` written and added to `_toctree.yml`.
- [ ] `templates/lerobot_modelcard_template.md` has a description entry and a `policy_docs` link for your policy.
- [ ] The models table in the root `README.md` lists your policy in the right category, linking to your doc page.
- [ ] At least one reproducible benchmark eval in the policy MDX with a published checkpoint (sim benchmark, or real-robot dataset + checkpoint).
The fastest way to get a clean PR is to copy the directory of the existing policy closest to yours, rename, and replace contents method by method. Don't wait until everything is polished — open a draft PR early and iterate with us; reviewers would much rather give feedback on a half-finished branch than a fully-merged one.
---
## Examples and community contributions
Check out these example policy implementations:
- [DiTFlow Policy](https://github.com/danielsanjosepro/lerobot_policy_ditflow) — Diffusion Transformer policy with flow-matching objective. Try it out in this example: [DiTFlow Example](https://github.com/danielsanjosepro/test_lerobot_policy_ditflow)
Thanks for taking the time to bring a new policy into LeRobot. Every architecture that lands in `main` — and every plugin published by the community — makes the library a little more useful for the next person, and a little more representative of where robot learning is going. We're looking forward to seeing what you ship. 🤗
docs/source/cameras.mdx
+226
View File
@@ -0,0 +1,226 @@
# Cameras
LeRobot offers multiple options for video capture:
| Class | Supported Cameras |
| ----------------- | ----------------------------------- |
| `OpenCVCamera` | Phone, built-in laptop, USB webcams |
| `ZMQCamera` | Network-connected cameras |
| `RealSenseCamera` | Intel RealSense (with depth) |
| `Reachy2Camera` | Reachy 2 robot cameras |
> [!TIP]
> For `OpenCVCamera` compatibility details, see the [Video I/O with OpenCV Overview](https://docs.opencv.org/4.x/d0/da7/videoio_overview.html).
### Find your camera
Every camera requires a unique identifier to be instantiated, allowing you to distinguish between multiple connected devices.
`OpenCVCamera` and `RealSenseCamera` support auto-discovery. Run the command below to list available devices and their identifiers. Note that these identifiers may change after rebooting your computer or re-plugging the camera, depending on your operating system.
```bash
lerobot-find-cameras opencv # or realsense for Intel Realsense cameras
```
The output will look something like this if you have two cameras connected:
```bash
--- Detected Cameras ---
Camera #0:
Name: OpenCV Camera @ 0
Type: OpenCV
Id: 0
Backend api: AVFOUNDATION
Default stream profile:
Format: 16.0
Width: 1920
Height: 1080
Fps: 15.0
--------------------
(more cameras ...)
```
> [!WARNING]
> When using Intel RealSense cameras in `macOS`, you could get this [error](https://github.com/IntelRealSense/librealsense/issues/12307): `Error finding RealSense cameras: failed to set power state`, this can be solved by running the same command with `sudo` permissions. Note that using RealSense cameras in `macOS` is unstable.
`ZMQCamera` and `Reachy2Camera` do not support auto-discovery. They must be configured manually by providing their network address and port or robot SDK settings.
## Use cameras
### Frame access modes
All camera classes implement three access modes for capturing frames:
| Method | Behavior | Blocks? | Best For |
| ------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------- | ---------------------------------------- |
| `read()` | Waits for the camera hardware to return a frame. May block for a long time depending on the camera and SDK. | Yes | Simple scripts, sequential capture |
| `async_read(timeout_ms)` | Returns the latest unconsumed frame from background thread. Blocks only if buffer is empty, up to `timeout_ms`. Raises `TimeoutError` if no frame arrives. | With a timeout | Control loops synchronized to camera FPS |
| `read_latest(max_age_ms)` | Peeks at the most recent frame in buffer (may be stale). Raises `TimeoutError` if frame is older than `max_age_ms`. | No | UI visualization, logging, monitoring |
### Usage examples
The following examples show how to use the camera API to configure and capture frames from different camera types.
- **Blocking and non-blocking frame capture** using an OpenCV-based camera
- **Color and depth capture** using an Intel RealSense camera
> [!WARNING]
> Failing to cleanly disconnect cameras can cause resource leaks. Use the context manager protocol to ensure automatic cleanup:
>
> ```python
> with OpenCVCamera(config) as camera:
> ...
> ```
>
> You can also call `connect()` and `disconnect()` manually, but always use a `finally` block for the latter.
<hfoptions id="shell_restart">
<hfoption id="Open CV Camera">
<!-- prettier-ignore-start -->
```python
from lerobot.cameras.opencv import OpenCVCamera, OpenCVCameraConfig
from lerobot.cameras import ColorMode, Cv2Rotation
# Construct an `OpenCVCameraConfig` with your desired FPS, resolution, color mode, and rotation.
config = OpenCVCameraConfig(
index_or_path=0,
fps=15,
width=1920,
height=1080,
color_mode=ColorMode.RGB,
rotation=Cv2Rotation.NO_ROTATION
)
# Instantiate and connect an `OpenCVCamera`, performing a warm-up read (default).
with OpenCVCamera(config) as camera:
# Read a frame synchronously — blocks until hardware delivers a new frame
frame = camera.read()
print(f"read() call returned frame with shape:", frame.shape)
# Read a frame asynchronously with a timeout — returns the latest unconsumed frame or waits up to timeout_ms for a new one
try:
for i in range(10):
frame = camera.async_read(timeout_ms=200)
print(f"async_read call returned frame {i} with shape:", frame.shape)
except TimeoutError as e:
print(f"No frame received within timeout: {e}")
# Instantly return a frame - returns the most recent frame captured by the camera
try:
initial_frame = camera.read_latest(max_age_ms=1000)
for i in range(10):
frame = camera.read_latest(max_age_ms=1000)
print(f"read_latest call returned frame {i} with shape:", frame.shape)
print(f"Was a new frame received by the camera? {not (initial_frame == frame).any()}")
except TimeoutError as e:
print(f"Frame too old: {e}")
```
<!-- prettier-ignore-end -->
</hfoption>
<hfoption id="Intel Realsense Camera">
<!-- prettier-ignore-start -->
```python
from lerobot.cameras.realsense import RealSenseCamera, RealSenseCameraConfig
from lerobot.cameras import ColorMode, Cv2Rotation
# Create a `RealSenseCameraConfig` specifying your camera’s serial number and enabling depth.
config = RealSenseCameraConfig(
serial_number_or_name="233522074606",
fps=15,
width=640,
height=480,
color_mode=ColorMode.RGB,
use_depth=True,
rotation=Cv2Rotation.NO_ROTATION
)
# Instantiate and connect a `RealSenseCamera` with warm-up read (default).
camera = RealSenseCamera(config)
camera.connect()
# Capture a color frame via `read()` and a depth map via `read_depth()`.
try:
color_frame = camera.read()
depth_map = camera.read_depth()
print("Color frame shape:", color_frame.shape)
print("Depth map shape:", depth_map.shape)
finally:
camera.disconnect()
```
<!-- prettier-ignore-end -->
</hfoption>
</hfoptions>
### Working with depth
The Intel RealSense and Reachy 2 cameras can capture both color and depth in lockstep. Calling `read()` returns the **color** frame as `(H, W, 3)` `uint8`. Calling `read_depth()` returns the **depth map** as `(H, W, 1)` `uint16`, where each pixel value is the distance from the sensor expressed in **millimetres**. A pixel value of `0` typically means "no measurement available" (out-of-range, occluded, or low-confidence).
During recording, the control loop peeks the freshest buffered frames non-blockingly via `read_latest()` (color) and `read_latest_depth()` (depth), adding the depth map as a sibling feature (e.g. `front_depth` next to `front`).
For how depth streams are stored and encoded when recording a dataset, see the [Depth streams](./video_encoding_parameters#depth-streams) section of the video encoding guide.
## Use your phone's camera
<hfoptions id="use phone">
<hfoption id="iPhone & macOS">
To use your iPhone as a camera on macOS, enable the Continuity Camera feature:
- Ensure your Mac is running macOS 13 or later, and your iPhone is on iOS 16 or later.
- Sign in both devices with the same Apple ID.
- Connect your devices with a USB cable or turn on Wi-Fi and Bluetooth for a wireless connection.
For more details, visit [Apple support](https://support.apple.com/en-gb/guide/mac-help/mchl77879b8a/mac).
</hfoption>
<hfoption id="OBS virtual camera">
If you want to use your phone as a camera using OBS, follow these steps to set up a virtual camera.
1. _(Linux only) Install `v4l2loopback-dkms` and `v4l-utils`_. These packages create virtual camera devices and verify their settings. Install with:
```bash
sudo apt install v4l2loopback-dkms v4l-utils
```
2. _Install the [DroidCam app](https://droidcam.app) on your phone_. This app is available for both iOS and Android.
3. _Download and install [OBS Studio](https://obsproject.com)_.
4. _Download and install the [DroidCam OBS plugin](https://droidcam.app/obs)_.
5. _Start OBS Studio_.
6. _Add your phone as a source_. Follow the instructions [here](https://droidcam.app/obs/usage). Be sure to set the resolution to `640x480` to avoid the watermarks.
7. _Adjust resolution settings_. In OBS Studio, go to `File > Settings > Video` or `OBS > Preferences... > Video`. Change the `Base(Canvas) Resolution` and the `Output(Scaled) Resolution` to `640x480` by manually typing it.
8. _Start virtual camera_. In OBS Studio, follow the instructions [here](https://obsproject.com/kb/virtual-camera-guide).
9. _Verify the virtual camera setup and resolution_.
- **Linux**: Use `v4l2-ctl` to list devices and check resolution:
```bash
v4l2-ctl --list-devices # find VirtualCam and note its /dev/videoX path
v4l2-ctl -d /dev/videoX --get-fmt-video # replace with your VirtualCam path
```
You should see `VirtualCam` listed and resolution `640x480`.
- **macOS**: Open Photo Booth or FaceTime and select "OBS Virtual Camera" as the input.
- **Windows**: The native Camera app doesn't support virtual cameras. Use a video conferencing app (Zoom, Teams) or run `lerobot-find-cameras opencv` directly to verify.
<details>
<summary><strong>Troubleshooting</strong></summary>
> The virtual camera resolution is incorrect.
Delete the virtual camera source and recreate it. The resolution cannot be changed after creation.
> Error reading frame in background thread for OpenCVCamera(X): OpenCVCamera(X) frame width=640 or height=480 do not match configured width=1920 or height=1080.
This error is caused by OBS Virtual Camera advertising a `1920x1080` resolution despite rescaling. The only fix for now is to comment out the width and height check in `_postprocess_image()`.
</details>
</hfoption>
</hfoptions>
If everything is set up correctly, your phone will appear as a standard OpenCV camera and can be used with `OpenCVCamera`.
docs/source/cheat-sheet.mdx
+177
View File
@@ -0,0 +1,177 @@
# Cheat sheet
All of the LeRobot commands in one place. If you forgot how to use a specific command or want to learn about a new one you can do it here.
> [!WARNING]
> For all of the commands listed below remember to change the ports/names/ids to your own values!
> [!TIP]
> Another great way to look at all the commands and get them configured for your specific setup is to use this [Jupyter Notebook](https://github.com/huggingface/lerobot/blob/main/examples/notebooks/quickstart.ipynb).
### Setup and installation
For installation please look at [LeRobot Installation](https://huggingface.co/docs/lerobot/main/en/installation).
### Useful tools
###### Find port
Use this to identify which serial ports your robots are connected to. Follow the instructions in your terminal: you will be asked to unplug the USB cable and press Enter. The script will then detect and print the correct serial port for that robot.
```bash
lerobot-find-port
```
###### Find cameras
Quickly find camera indices and verify their output. This command prints camera information to the terminal and saves test frames from each detected camera to `lerobot/outputs/captured_images`
```bash
lerobot-find-cameras
```
### Calibration
In most cases you will need to perform calibration just once for each robot and teleoperation device. Before performing the calibration make sure that all the joints are roughly in the middle position.
```bash
lerobot-calibrate \
--robot.type=so101_follower \
--robot.port=/dev/ttyACM0 \
--robot.id=my_follower_arm
```
Make sure that you use the same IDs used during calibration later for the other scripts. That's how LeRobot finds the calibration files.
### Teleoperation
Teleoperating with two cameras and displaying the data with Rerun.
```bash
lerobot-teleoperate \
--robot.type=so101_follower \
--robot.port=/dev/ttyACM0 \
--robot.id=my_follower_arm \
--robot.cameras="{ top: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30}, wrist: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30} }" \
--teleop.type=so101_leader \
--teleop.port=/dev/ttyACM1 \
--teleop.id=my_leader_arm \
--display_data=true
```
### Recording a dataset
The dataset is automatically uploaded to the server and saved under repo_id, make sure you are logged in to your HF account with CLI:
`hf auth login`
You can get the token from: [https://huggingface.co/settings/tokens](https://huggingface.co/settings/tokens)
```bash
lerobot-record \
--robot.type=so101_follower \
--robot.port=/dev/ttyACM0 \
--robot.id=my_follower_arm \
--robot.cameras="{ top: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30}, wrist: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30} }" \
--teleop.type=so101_leader \
--teleop.port=/dev/ttyACM1 \
--teleop.id=my_leader_arm \
--dataset.repo_id=${HF_USER}/so101_dataset_test \
--dataset.num_episodes=30 \
--dataset.single_task="put the red brick in a bowl" \
--dataset.streaming_encoding=true \
--display_data=true
```
While collecting the dataset you can control the process with your keyboard:
Control the data recording flow using keyboard shortcuts:
- Press **Right Arrow (`→`)**: Save episode and move to the next.
- Press **Left Arrow (`←`)**: Delete current episode and retry.
- Press **Escape (`ESC`)**: Stop, encode videos, and upload.
### Recording depth
Intel RealSense cameras (`type: intelrealsense`) record a depth stream when you set `use_depth: true`. Depth is quantized to 12-bit codes and stored as its own video.
```bash
lerobot-record \
... \
--robot.cameras="{ head: {type: intelrealsense, serial_number_or_name: \"0123456789\", width: 640, height: 480, fps: 30, use_depth: true} }" \
--dataset.repo_id=${HF_USER}/so101_depth_test \
--dataset.single_task="put the red brick in a bowl" \
--dataset.depth_encoder.depth_min=0.01 \
--dataset.depth_encoder.depth_max=10.0 \
--dataset.depth_encoder.shift=0.0 \
--dataset.depth_encoder.use_log=true
```
### Video encoding parameters
RGB and depth streams are encoded independently via the `--dataset.rgb_encoder.*` and `--dataset.depth_encoder.*` keys.
```bash
lerobot-record \
... \
--dataset.rgb_encoder.vcodec=h264 \
--dataset.rgb_encoder.pix_fmt=yuv420p \
--dataset.rgb_encoder.crf=23 \
--dataset.depth_encoder.vcodec=hevc \
--dataset.depth_encoder.extra_options='{"x265-params": "lossless=1"}'
```
### Training
Depending on your hardware training the policy might take a few hours. That's how you train simple `ACT` policy:
```bash
lerobot-train \
--dataset.repo_id=${HF_USER}/so101_dataset_test \
--policy.type=act \
--output_dir=outputs/train/act_so101_test \
--job_name=act_so101_test \
--policy.device=cuda \
--wandb.enable=true \
--policy.repo_id=${HF_USER}/policy_test \
--steps=20000
```
- Policy Types: `act`, `diffusion`, `smolvla`, `pi05`
- Devices: `cuda` (NVIDIA), `mps` (Apple Silicon), `cpu`
If you want to fine-tune a specific model you can provide the path to the model. In this case path is enough and type can be skipped.
```bash
lerobot-train \
--dataset.repo_id=${HF_USER}/so101_dataset_test \
--policy.path=username/the_policy_to_finetune \
--policy.device=cuda \
--policy.repo_id=${HF_USER}/policy_test \
--output_dir=outputs/train/act_so101_test \
--steps=20000
```
No local GPU? Add `--job.target=<flavor>` (e.g. `a10g-small`) to either command and `lerobot-train` runs it on [Hugging Face Jobs](https://huggingface.co/docs/hub/jobs) instead — it uploads a local-only dataset for you and pushes the trained model. List flavors with `hf jobs hardware`.
To resume, point `--config_path` at a checkpoint and add `--resume=true`. It accepts a local path or a Hub repo id (the latest checkpoint is fetched), and works locally or on a job by adding `--job.target=<flavor>`:
```bash
lerobot-train --config_path=${HF_USER}/policy_test --resume=true --job.target=a10g-small
```
### Inference
Inference means running the trained policy/model on a robot. For that we use `lerobot-rollout`. You will need to provide a path to your policy. It can be a local path or a path to Hugging Face for example "lerobot/folding_latest". Your cameras configuration needs to match what was used when collecting the dataset. Duration is in seconds if unspecified, it will run forever.
> [!TIP]
> If you are using the previous release V0.5.1 instead of `lerobot-rollout` you need to use `lerobot-record`. More information [here](https://huggingface.co/docs/lerobot/v0.5.1/en/il_robots#run-inference-and-evaluate-your-policy).
```bash
lerobot-rollout \
--strategy.type=base \
--policy.path=${HF_USER}/my_policy \
--robot.type=so101_follower \
--robot.port=/dev/ttyACM1 \
--robot.cameras="{ up: {type: opencv, index_or_path: /dev/video1, width: 640, height: 480, fps: 30}, side: {type: opencv, index_or_path: /dev/video5, width: 640, height: 480, fps: 30}}" \
--task="Put lego brick into the transparent box" \
--duration=60
```
docs/source/contributing.md
Symlink
+1
View File
@@ -0,0 +1 @@
../../CONTRIBUTING.md
docs/source/damiao.mdx
+165
View File
@@ -0,0 +1,165 @@
# Damiao Motors and CAN Bus
This guide covers setup and usage of Damiao motors with LeRobot via CAN bus communication.
Currently, only Linux is supported, as the OpenArms CAN adapter only has drivers for Linux.
## Linux CAN Setup
Before using Damiao motors, you need to set up the CAN interface on your Linux system.
### Install CAN Utilities
```bash
sudo apt-get install can-utils
```
### Configure CAN Interface (Manual)
For standard CAN FD (recommended for OpenArms):
```bash
sudo ip link set can0 down
sudo ip link set can0 type can bitrate 1000000 dbitrate 5000000 fd on
sudo ip link set can0 up
```
For standard CAN (without FD):
```bash
sudo ip link set can0 down
sudo ip link set can0 type can bitrate 1000000
sudo ip link set can0 up
```
### Configure CAN Interface (Using LeRobot)
LeRobot provides a utility script to setup and test CAN interfaces:
```bash
# Setup multiple interfaces (e.g., OpenArms Followers with 2 CAN buses)
lerobot-setup-can --mode=setup --interfaces=can0,can1
```
## Debugging CAN Communication
Use the built-in debug tools to test motor communication:
```bash
# Test motors on all interfaces
lerobot-setup-can --mode=test --interfaces=can0,can1
# Run speed/latency test
lerobot-setup-can --mode=speed --interfaces=can0
```
The test mode will scan for motors (IDs 0x01-0x08) and report which ones respond. Example output:
```
can0: UP (CAN FD)
Motor 0x01 (joint_1): ✓ FOUND
→ Response 0x11 [FD]: 00112233...
Motor 0x02 (joint_2): ✓ FOUND
Motor 0x03 (joint_3): ✗ No response
...
Summary: 2/8 motors found
```
## Usage
### Basic Setup
```python
from lerobot.motors import Motor
from lerobot.motors.damiao import DamiaoMotorsBus
# Define your motors with send/receive CAN IDs
motors = {
"joint_1": Motor(id=0x01, motor_type_str="dm8009", recv_id=0x11),
"joint_2": Motor(id=0x02, motor_type_str="dm4340", recv_id=0x12),
"joint_3": Motor(id=0x03, motor_type_str="dm4310", recv_id=0x13),
}
# Create the bus
bus = DamiaoMotorsBus(
port="can0", # Linux socketcan interface
motors=motors,
)
# Connect
bus.connect()
```
### Reading Motor States
```python
# Read single motor position (degrees)
position = bus.read("Present_Position", "joint_1")
# Read from multiple motors
positions = bus.sync_read("Present_Position") # All motors
positions = bus.sync_read("Present_Position", ["joint_1", "joint_2"])
# Read all states at once (position, velocity, torque)
states = bus.sync_read_all_states()
# Returns: {'joint_1': {'position': 45.2, 'velocity': 1.3, 'torque': 0.5}, ...}
```
### Writing Motor Commands
```python
# Enable torque
bus.enable_torque()
# Set goal position (degrees)
bus.write("Goal_Position", "joint_1", 45.0)
# Set positions for multiple motors
bus.sync_write("Goal_Position", {
"joint_1": 45.0,
"joint_2": -30.0,
"joint_3": 90.0,
})
# Disable torque
bus.disable_torque()
```
## Configuration Options
| Parameter | Default | Description |
| -------------- | --------- | ----------------------------------------------------------- |
| `port` | - | CAN interface (`can0`) or serial port (`/dev/cu.usbmodem*`) |
| `use_can_fd` | `True` | Enable CAN FD for higher data rates |
| `bitrate` | `1000000` | Nominal bitrate (1 Mbps) |
| `data_bitrate` | `5000000` | CAN FD data bitrate (5 Mbps) |
## Motor Configuration
Each motor requires:
- `id`: CAN ID for sending commands
- `motor_type`: One of the supported motor types (e.g., `"dm8009"`, `"dm4340"`)
- `recv_id`: CAN ID for receiving responses
OpenArms default IDs follow the pattern: send ID `0x0N`, receive ID `0x1N` where N is the joint number.
## Troubleshooting
### No Response from Motors
1. **Check power**
2. **Verify CAN wiring**: Check CAN-H, CAN-L, and GND connections
3. **Check motor IDs**: Use Damiao Debugging Tools to verify/configure IDs
4. **Test CAN interface**: Run `candump can0` to see if messages are being received
5. **Run diagnostics**: `lerobot-setup-can --mode=test --interfaces=can0`
### Motor Timeout Parameter
If motors were configured with timeout=0, they won't respond to commands. Use Damiao Debugging Tools to set a non-zero timeout value.
### Verify CAN FD Status
```bash
ip -d link show can0 | grep fd
```
docs/source/debug_processor_pipeline.mdx
+299
View File
@@ -0,0 +1,299 @@
# Debug Your Processor Pipeline
Processor pipelines can be complex, especially when chaining multiple transformation steps.
Unlike simple function calls, pipelines lack natural observability, you can't easily see what happens
between each step or where things go wrong.
This guide provides debugging tools and techniques specifically designed to address these challenges
and help you understand data flow through your pipelines.
We'll explore three complementary debugging approaches: **hooks** for runtime monitoring, **step-through debugging** for detailed inspection, and **feature validation** for catching structural mismatches. Each serves a different purpose and together they provide complete visibility into your pipeline's behavior.
## Understanding Hooks
Hooks are functions that get called at specific points during pipeline execution.
They provide a way to inspect, monitor, or modify data without changing your pipeline code.
Think of them as "event listeners" for your pipeline.
### What is a Hook?
A hook is a callback function that gets automatically invoked at specific moments during pipeline execution.
The concept comes from event-driven programming, imagine you could "hook into" the pipeline's execution flow to observe or react to what's happening.
Think of hooks like inserting checkpoints into your pipeline. Every time the pipeline reaches one of these checkpoints, it pauses briefly to call your hook function, giving you a chance to inspect the current state, log information, and validate data.
A hook is simply a function that accepts two parameters:
- `step_idx: int` - The index of the current processing step (0, 1, 2, etc.)
- `transition: EnvTransition` - The data transition at that point in the pipeline
The beauty of hooks is their non-invasive nature: you can add monitoring, validation, or debugging logic without changing a single line of your pipeline code. The pipeline remains clean and focused on its core logic, while hooks handle the cross-cutting concerns like logging, monitoring, and debugging.
### Before vs After Hooks
The pipeline supports two types of hooks:
- **Before hooks** (`register_before_step_hook`) - Called before each step executes
- **After hooks** (`register_after_step_hook`) - Called after each step completes
```python
def before_hook(step_idx: int, transition: EnvTransition):
"""Called before step processes the transition."""
print(f"About to execute step {step_idx}")
# Useful for: logging, validation, setup
def after_hook(step_idx: int, transition: EnvTransition):
"""Called after step has processed the transition."""
print(f"Completed step {step_idx}")
# Useful for: monitoring results, cleanup, debugging
processor.register_before_step_hook(before_hook)
processor.register_after_step_hook(after_hook)
```
### Implementing a NaN Detection Hook
Here's a practical example of a hook that detects NaN values:
```python
def check_nans(step_idx: int, transition: EnvTransition):
"""Check for NaN values in observations."""
obs = transition.get(TransitionKey.OBSERVATION)
if obs:
for key, value in obs.items():
if isinstance(value, torch.Tensor) and torch.isnan(value).any():
print(f"NaN detected in {key} at step {step_idx}")
# Register the hook to run after each step
processor.register_after_step_hook(check_nans)
# Process your data - the hook will be called automatically
output = processor(input_data)
# Remove the hook when done debugging
processor.unregister_after_step_hook(check_nans)
```
### How Hooks Work Internally
Understanding the internal mechanism helps you use hooks more effectively. The pipeline maintains two separate lists: one for before-step hooks and another for after-step hooks. When you register a hook, it's simply appended to the appropriate list.
During execution, the pipeline follows a strict sequence: for each processing step, it first calls all before-hooks in registration order, then executes the actual step transformation, and finally calls all after-hooks in registration order. This creates a predictable, sandwich-like structure around each step.
The key insight is that hooks don't change the core pipeline logic—they're purely additive. The pipeline's `_forward` method orchestrates this dance between hooks and processing steps, ensuring that your debugging or monitoring code runs at exactly the right moments without interfering with the main data flow.
Here's a simplified view of how the pipeline executes hooks:
```python
class DataProcessorPipeline:
def __init__(self):
self.steps = [...]
self.before_step_hooks = [] # List of before hooks
self.after_step_hooks = [] # List of after hooks
def _forward(self, transition):
"""Internal method that processes the transition through all steps."""
for step_idx, processor_step in enumerate(self.steps):
# 1. Call all BEFORE hooks
for hook in self.before_step_hooks:
hook(step_idx, transition)
# 2. Execute the actual processing step
transition = processor_step(transition)
# 3. Call all AFTER hooks
for hook in self.after_step_hooks:
hook(step_idx, transition)
return transition
def register_before_step_hook(self, hook_fn):
self.before_step_hooks.append(hook_fn)
def register_after_step_hook(self, hook_fn):
self.after_step_hooks.append(hook_fn)
```
### Execution Flow
The execution flow looks like this:
```
Input → Before Hook → Step 0 → After Hook → Before Hook → Step 1 → After Hook → ... → Output
```
For example, with 3 steps and both hook types:
```python
def timing_before(step_idx, transition):
print(f"⏱️ Starting step {step_idx}")
def validation_after(step_idx, transition):
print(f"✅ Completed step {step_idx}")
processor.register_before_step_hook(timing_before)
processor.register_after_step_hook(validation_after)
# This will output:
# ⏱️ Starting step 0
# ✅ Completed step 0
# ⏱️ Starting step 1
# ✅ Completed step 1
# ⏱️ Starting step 2
# ✅ Completed step 2
```
### Multiple Hooks
You can register multiple hooks of the same type - they execute in the order registered:
```python
def log_shapes(step_idx: int, transition: EnvTransition):
obs = transition.get(TransitionKey.OBSERVATION)
if obs:
print(f"Step {step_idx} observation shapes:")
for key, value in obs.items():
if isinstance(value, torch.Tensor):
print(f" {key}: {value.shape}")
processor.register_after_step_hook(check_nans) # Executes first
processor.register_after_step_hook(log_shapes) # Executes second
# Both hooks will be called after each step in registration order
output = processor(input_data)
```
While hooks are excellent for monitoring specific issues (like NaN detection) or gathering metrics during normal pipeline execution, sometimes you need to dive deeper. When you want to understand exactly what happens at each step or debug complex transformation logic, step-through debugging provides the detailed inspection you need.
## Step-Through Debugging
Step-through debugging is like having a slow-motion replay for your pipeline. Instead of watching your data get transformed in one quick blur from input to output, you can pause and examine what happens after each individual step.
This approach is particularly valuable when you're trying to understand a complex pipeline, debug unexpected behavior, or verify that each transformation is working as expected. Unlike hooks, which are great for automated monitoring, step-through debugging gives you manual, interactive control over the inspection process.
The `step_through()` method is a generator that yields the transition state after each processing step, allowing you to inspect intermediate results. Think of it as creating a series of snapshots of your data as it flows through the pipeline—each snapshot shows you exactly what your data looks like after one more transformation has been applied.
### How Step-Through Works
The `step_through()` method fundamentally changes how the pipeline executes. Instead of running all steps in sequence and only returning the final result, it transforms the pipeline into an iterator that yields intermediate results.
Here's what happens internally: the method starts by converting your input data into the pipeline's internal transition format, then yields this initial state. Next, it applies the first processing step and yields the result. Then it applies the second step to that result and yields again, and so on. Each `yield` gives you a complete snapshot of the transition at that point.
This generator pattern is powerful because it's lazy—the pipeline only computes the next step when you ask for it. This means you can stop at any point, inspect the current state thoroughly, and decide whether to continue. You're not forced to run the entire pipeline just to debug one problematic step.
Instead of running the entire pipeline and only seeing the final result, `step_through()` pauses after each step and gives you the intermediate transition:
```python
# This creates a generator that yields intermediate states
for i, intermediate_result in enumerate(processor.step_through(input_data)):
print(f"=== After step {i} ===")
# Inspect the observation at this stage
obs = intermediate_result.get(TransitionKey.OBSERVATION)
if obs:
for key, value in obs.items():
if isinstance(value, torch.Tensor):
print(f"{key}: shape={value.shape}, dtype={value.dtype}")
```
### Interactive Debugging with Breakpoints
You can add breakpoints in the step-through loop to interactively debug:
```python
# Step through the pipeline with debugging
for i, intermediate in enumerate(processor.step_through(data)):
print(f"Step {i}: {processor.steps[i].__class__.__name__}")
# Set a breakpoint to inspect the current state
breakpoint() # Debugger will pause here
# You can now inspect 'intermediate' in the debugger:
# - Check tensor shapes and values
# - Verify expected transformations
# - Look for unexpected changes
```
During the debugger session, you can:
- Examine `intermediate[TransitionKey.OBSERVATION]` to see observation data
- Check `intermediate[TransitionKey.ACTION]` for action transformations
- Inspect any part of the transition to understand what each step does
Step-through debugging is perfect for understanding the _data_ transformations, but what about the _structure_ of that data? While hooks and step-through help you debug runtime behavior, you also need to ensure your pipeline produces data in the format expected by downstream components. This is where feature contract validation comes in.
## Validating Feature Contracts
Feature contracts define what data structure your pipeline expects as input and produces as output.
Validating these contracts helps catch mismatches early.
### Understanding Feature Contracts
Each processor step has a `transform_features()` method that describes how it changes the data structure:
```python
# Get the expected output features from your pipeline
initial_features = {
PipelineFeatureType.OBSERVATION: {
"observation.state": PolicyFeature(type=FeatureType.STATE, shape=(7,)),
"observation.image": PolicyFeature(type=FeatureType.IMAGE, shape=(3, 224, 224))
},
PipelineFeatureType.ACTION: {
"action": PolicyFeature(type=FeatureType.ACTION, shape=(4,))
}
}
# Check what your pipeline will output
output_features = processor.transform_features(initial_features)
print("Input features:")
for feature_type, features in initial_features.items():
print(f" {feature_type}:")
for key, feature in features.items():
print(f" {key}: {feature.type.value}, shape={feature.shape}")
print("\nOutput features:")
for feature_type, features in output_features.items():
print(f" {feature_type}:")
for key, feature in features.items():
print(f" {key}: {feature.type.value}, shape={feature.shape}")
```
### Verifying Expected Features
Check that your pipeline produces the features you expect:
```python
# Define what features you expect the pipeline to produce
expected_keys = ["observation.state", "observation.image", "action"]
print("Validating feature contract...")
for expected_key in expected_keys:
found = False
for feature_type, features in output_features.items():
if expected_key in features:
feature = features[expected_key]
print(f"✅ {expected_key}: {feature.type.value}, shape={feature.shape}")
found = True
break
if not found:
print(f"❌ Missing expected feature: {expected_key}")
```
This validation helps ensure your pipeline will work correctly with downstream components that expect specific data structures.
## Summary
Now that you understand the three debugging approaches, you can tackle any pipeline issue systematically:
1. **Hooks** - For runtime monitoring and validation without modifying pipeline code
2. **Step-through** - For inspecting intermediate states and understanding transformations
3. **Feature validation** - For ensuring data structure contracts are met
**When to use each approach:**
- Start with **step-through debugging** when you need to understand what your pipeline does or when something unexpected happens
- Add **hooks** for continuous monitoring during development and production to catch issues automatically
- Use **feature validation** before deployment to ensure your pipeline works with downstream components
These three tools work together to give you the complete observability that complex pipelines naturally lack. With hooks watching for issues, step-through helping you understand behavior, and feature validation ensuring compatibility, you'll be able to debug any pipeline confidently and efficiently.
docs/source/earthrover_mini_plus.mdx
+238
View File
@@ -0,0 +1,238 @@
# EarthRover Mini Plus
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Earth_Rover_Mini_5_240c9adc-4f9e-44b7-982f-5d1dc24af1d8.png.webp"
alt="EarthRover Mini Plus"
width="70%"
/>
The EarthRover Mini Plus is a fully open source mobile robot that connects through the cloud using the Frodobots SDK. This lets you control the robot and record datasets for training AI models.
## What You Need
### Hardware
- EarthRover Mini robot
- Computer with Python 3.12 or newer
- Internet connection
### Setting Up the Frodobots SDK
The robot needs the [Frodobots SDK](https://github.com/frodobots-org/earth-rovers-sdk) running on your computer. Here's how:
1. Download and install the SDK:
```bash
git clone https://github.com/frodobots-org/earth-rovers-sdk.git
cd earth-rovers-sdk
pip install -r requirements.txt
```
2. Save Credentials:
Write your .env variables with the SDK API key and bot name provided by the Frodobots team.
```bash
SDK_API_TOKEN=your_sdk_api_token_here
BOT_SLUG=your_bot_slug_here
CHROME_EXECUTABLE_PATH=/path/to/chrome_or_chromium
# Default value is MAP_ZOOM_LEVEL=18 https://wiki.openstreetmap.org/wiki/Zoom_levels
MAP_ZOOM_LEVEL=18
MISSION_SLUG=your_mission_slug_here
# Image quality between 0.1 and 1.0 (default: 0.8)
# Recommended: 0.8 for better performance
IMAGE_QUALITY=0.8
# Image format: jpeg, png or webp (default: png)
# Recommended: jpeg for better performance and lower bandwidth usage
IMAGE_FORMAT=jpeg
```
3. Start the SDK:
```bash
hypercorn main:app --reload
```
4. Open your web browser and go to `http://localhost:8000`, then click "Join"
The SDK gives you:
- Live video from front and rear cameras
> [!IMPORTANT]
> The SDK must be running before you can use the robot.
## Install LeRobot
Follow our [Installation Guide](./installation) to install LeRobot.
In addition to the base installation, install the EarthRover Mini with hardware dependencies:
```bash
pip install -e ".[hardware]"
```
## How It Works
The robot uses the internet to communicate:
- **Movement commands**: Sent through the SDK
- **Camera video**: Received from the SDK
- **Robot info**: Battery, location, speed from the SDK
You don't need to plug anything in - it all works through the SDK.
## Calibration
No calibration needed! The robot is ready to use as soon as the SDK is running.
## Controlling the Robot
You control the robot using your keyboard - just like playing a video game with WASD keys.
### Keyboard Controls
| Key | Action |
| --- | -------------------------------- |
| W | Move forward |
| S | Move backward |
| A | Turn left (with forward motion) |
| D | Turn right (with forward motion) |
| Q | Rotate left in place |
| E | Rotate right in place |
| X | Stop all movement |
| +/= | Increase speed |
| - | Decrease speed |
| ESC | Disconnect |
### Speed Settings
You can adjust how fast the robot moves:
- **Forward/backward speed**: Default is full speed (1.0)
- **Turning speed**: Default is full speed (1.0)
- **Speed changes**: Use +/- keys to adjust by 0.1 each time
### Try It Out
Test driving the robot before recording data:
```python
from lerobot.robots.earthrover_mini_plus import EarthRoverMiniPlus, EarthRoverMiniPlusConfig
from lerobot.teleoperators.keyboard import KeyboardRoverTeleop, KeyboardRoverTeleopConfig
# Initialize robot
robot_config = EarthRoverMiniPlusConfig()
robot = EarthRoverMiniPlus(robot_config)
# Initialize teleoperator
teleop_config = KeyboardRoverTeleopConfig(
linear_speed=1.0,
angular_speed=1.0,
speed_increment=0.1
)
teleop = KeyboardRoverTeleop(teleop_config)
# Connect
robot.connect()
teleop.connect()
# Teleoperate (use keyboard controls)
try:
while True:
action = teleop.get_action()
robot.send_action(action)
except KeyboardInterrupt:
pass
finally:
robot.disconnect()
teleop.disconnect()
```
> [!TIP]
> If you're using a Mac, you might need to give Terminal permission to access your keyboard for teleoperation. Go to System Preferences > Security & Privacy > Input Monitoring and check the box for Terminal.
## Recording Data
Once you can drive the robot well, you can start recording data to train AI models. The system records:
- **What you do**: How you move the robot (forward, backward, turning)
- **What the robot sees**:
- Videos from both cameras
- Robot speed and direction
- Battery level and location
- GPS position and signal
- Other sensor data
- **When it happened**: Timestamps for everything
### Setting Up Hugging Face
We use Hugging Face to store your data online. First, log in with your token from [Hugging Face settings](https://huggingface.co/settings/tokens):
```bash
hf auth login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
```
Store your Hugging Face username:
```bash
HF_USER=$(hf auth whoami | awk -F': *' 'NR==1 {print $2}')
echo $HF_USER
```
### Start Recording
Use the standard recording command:
```bash
lerobot-record \
--robot.type=earthrover_mini_plus \
--teleop.type=keyboard_rover \
--dataset.repo_id=your_username/dataset_name \
--dataset.num_episodes=2 \
--dataset.fps=10 \
--dataset.single_task="Navigate around obstacles" \
--dataset.streaming_encoding=true \
--dataset.encoder_threads=2 \
# --dataset.rgb_encoder.vcodec=auto \
--display_data=true
```
Replace `your_username/dataset_name` with your Hugging Face username and a name for your dataset.
### What Gets Saved
Your dataset includes:
**Your Actions (2 features)**:
- `linear_velocity`: How much you moved forward/backward
- `angular_velocity`: How much you turned left/right
**Robot Observations (24 features)**:
- Front camera video
- Rear camera video
- Current speed
- Battery level
- Orientation
- GPS (latitude, longitude, signal strength)
- Network signal strength
- Vibration level
- Lamp state (on/off)
- Accelerometer (x, y, z)
- Gyroscope (x, y, z)
- Magnetometer (x, y, z)
- Wheel RPMs (4 wheels)
### Where Your Data Goes
On your computer: `~/.cache/huggingface/lerobot/{repo-id}`
After recording, your data automatically uploads to your Hugging Face page:
```bash
echo https://huggingface.co/datasets/${HF_USER}/earthrover-navigation
```
Your dataset will be tagged with `LeRobot` for community discovery.
docs/source/env_processor.mdx
+437
View File
@@ -0,0 +1,437 @@
# Environment Processors
Environment processors are a critical layer in LeRobot's data processing architecture that handle **environment-specific** transformations, separate from policy-specific processing. This separation of concerns enables cleaner code, better modularity, and easier experimentation with different environments and policies.
## Why Environment Processors?
When working with different robot environments (LIBERO, MetaWorld, Aloha, etc.), each environment often has unique data formats, coordinate systems, and conventions that need standardization **before** policy processing. Without environment processors, these transformations would be:
1. **Hardcoded in environment code** - Making it difficult to experiment with different state representations
2. **Duplicated across policies** - Each policy would need to handle environment-specific quirks
3. **Mixed with policy logic** - Violating separation of concerns and making debugging harder
Environment processors solve this by providing a **dedicated processing layer** between raw environment observations and policy inputs.
## The Processing Pipeline
Here's how data flows through the complete processing pipeline during evaluation:
```python
# In lerobot_eval.py rollout() function:
# 1. Raw environment observation (numpy arrays, various formats)
raw_observation = env.step(action)
# 2. Convert numpy to torch, normalize images [0,1]
observation = preprocess_observation(raw_observation)
# 3. Add task metadata (for multi-task environments)
observation = add_envs_task(env, observation)
# 4. ENVIRONMENT-SPECIFIC preprocessing (NEW!)
# - Flatten robot states
# - Rotate images to match dataset conventions
# - Handle environment-specific coordinate systems
observation = env_preprocessor(observation)
# 5. POLICY-SPECIFIC preprocessing
# - Normalize with dataset statistics
# - Add batch dimensions
# - Move to GPU
# - Tokenize language instructions
observation = preprocessor(observation)
# 6. Policy inference
action = policy.select_action(observation)
# 7. POLICY-SPECIFIC postprocessing
# - Unnormalize actions
# - Remove batch dimensions
action = postprocessor(action)
# 8. ENVIRONMENT-SPECIFIC postprocessing (NEW!)
# - Convert action formats if needed
# - Apply environment-specific constraints
action_transition = {"action": action}
action_transition = env_postprocessor(action_transition)
action = action_transition["action"]
# 9. Execute in environment
env.step(action)
```
## The Benefits
### 1. **Separation of Concerns**
Environment processors handle transformations specific to the **environment's data format**, while policy processors handle transformations specific to the **model's requirements**.
```python
# ❌ Before: Mixed concerns
class LiberoVLAPolicy:
def preprocess(self, obs):
# Environment-specific: Flatten robot state (shouldn't be in policy!)
state = self._flatten_robot_state(obs["robot_state"])
# Policy-specific: Normalize with dataset stats
state = self.normalizer(state)
return state
# ✅ After: Clear separation
# Environment processor: Handles LIBERO's nested robot state
env_preprocessor = LiberoProcessorStep() # Flattens robot_state
# Policy processor: Handles model requirements
policy_preprocessor = NormalizerProcessorStep(stats=dataset_stats)
```
### 2. **Flexibility and Reusability**
The same policy can work with different environment processors, and the same environment processor can work with different policies:
````python
# Use SmolVLA policy with LIBERO environment
# Use SmolVLA policy with LIBERO environment
libero_preprocessor, libero_postprocessor = make_env_pre_post_processors(
env_cfg=libero_cfg,
policy_cfg=smolvla_cfg,
)
smolvla_preprocessor, smolvla_postprocessor = make_pre_post_processors(smolvla_cfg)
# Or use ACT policy with the same LIBERO environment
libero_preprocessor, libero_postprocessor = make_env_pre_post_processors(
env_cfg=libero_cfg,
policy_cfg=act_cfg,
)
act_preprocessor, act_postprocessor = make_pre_post_processors(act_cfg)
```python
# Use SmolVLA policy with LIBERO environment
libero_preprocessor, libero_postprocessor = make_env_pre_post_processors(
env_cfg=libero_cfg,
policy_cfg=smolvla_cfg,
)
smolvla_preprocessor, smolvla_postprocessor = make_pre_post_processors(smolvla_cfg)
# Or use ACT policy with the same LIBERO environment
libero_preprocessor, libero_postprocessor = make_env_pre_post_processors(
env_cfg=libero_cfg,
policy_cfg=act_cfg,
)
act_preprocessor, act_postprocessor = make_pre_post_processors(act_cfg)
### 3. **Easier Experimentation**
Want to try different state representations for LIBERO? Just create a new processor:
```python
# Original: 8D state (pos + quat→axisangle + gripper)
@ProcessorStepRegistry.register("libero_processor")
class LiberoProcessorStep(ObservationProcessorStep):
def _process_observation(self, obs):
eef_pos = robot_state["eef"]["pos"] # 3D
eef_axisangle = quat2axisangle(quat) # 3D
gripper = robot_state["gripper"]["qpos"] # 2D
state = torch.cat([eef_pos, eef_axisangle, gripper], dim=-1) # 8D
return state
# Experiment: Add velocity for better control
@ProcessorStepRegistry.register("libero_velocity_processor")
class LiberoVelocityProcessorStep(ObservationProcessorStep):
def _process_observation(self, obs):
# Include velocities for 14D state
eef_pos = robot_state["eef"]["pos"] # 3D
eef_axisangle = quat2axisangle(quat) # 3D
eef_vel = robot_state["eef"]["vel"] # 3D (NEW)
gripper_pos = robot_state["gripper"]["qpos"] # 2D
gripper_vel = robot_state["gripper"]["qvel"] # 3D (NEW)
state = torch.cat([eef_pos, eef_axisangle, eef_vel,
gripper_pos, gripper_vel], dim=-1) # 14D
return state
````
### 4. **Cleaner Environment Code**
Environments expose **all available data** without needing to know what downstream models will use:
```python
# LIBERO environment exposes full robot state
observation = {
"pixels": {"image": img, "image2": img2},
"robot_state": {
"eef": {"pos": ..., "quat": ..., "vel": ..., "mat": ..., "axisangle": ...},
"gripper": {"qpos": ..., "qvel": ...},
"joints": {"pos": ..., "vel": ...}
}
}
# Environment processor decides what to use
# Policy processor handles model-specific transformations
```
## Using Environment Processors
### Factory Function
The `make_env_pre_post_processors` function follows the same pattern as `make_pre_post_processors` for policies:
```python
from lerobot.envs import make_env_pre_post_processors, PushtEnv
from lerobot.envs.configs import LiberoEnv
# For LIBERO: Returns LiberoProcessorStep in preprocessor
libero_cfg = LiberoEnv(task="libero_spatial", camera_name=["agentview"])
env_preprocessor, env_postprocessor = make_env_pre_post_processors(libero_cfg)
# For other environments: Returns identity processors (no-op)
pusht_cfg = PushtEnv()
env_preprocessor, env_postprocessor = make_env_pre_post_processors(pusht_cfg)
```
### Implementation in `envs/factory.py`
```python
def make_env_pre_post_processors(
env_cfg: EnvConfig,
) -> tuple[
PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
]:
"""
Create preprocessor and postprocessor pipelines for environment observations.
Args:
env_cfg: The configuration of the environment.
Returns:
A tuple containing:
- preprocessor: Pipeline that processes environment observations
- postprocessor: Pipeline that processes environment outputs
"""
# For LIBERO environments, add the LiberoProcessorStep to preprocessor
if isinstance(env_cfg, LiberoEnv) or "libero" in env_cfg.type:
preprocessor = PolicyProcessorPipeline(steps=[LiberoProcessorStep()])
else:
# For all other environments, return an identity preprocessor
preprocessor = PolicyProcessorPipeline(steps=[])
# Postprocessor is currently identity for all environments
# Future: Could add environment-specific action transformations
postprocessor = PolicyProcessorPipeline(steps=[])
return preprocessor, postprocessor
```
### Integration in Evaluation
In `lerobot_eval.py`, the environment processors are created once and used throughout:
```python
def eval_main(cfg: EvalPipelineConfig):
# Create environment
envs = make_env(cfg.env, n_envs=cfg.eval.batch_size)
# Create policy
policy = make_policy(cfg=cfg.policy, env_cfg=cfg.env)
# Create policy processors
preprocessor, postprocessor = make_pre_post_processors(
policy_cfg=cfg.policy,
pretrained_path=cfg.policy.pretrained_path,
)
# Create environment processors (NEW!)
env_preprocessor, env_postprocessor = make_env_pre_post_processors(env_cfg=cfg.env)
# Run evaluation with both processor types
eval_policy_all(
envs=envs,
policy=policy,
env_preprocessor=env_preprocessor, # Environment-specific
env_postprocessor=env_postprocessor, # Environment-specific
preprocessor=preprocessor, # Policy-specific
postprocessor=postprocessor, # Policy-specific
n_episodes=cfg.eval.n_episodes,
)
```
## Example: LIBERO Environment Processor
The `LiberoProcessorStep` demonstrates a real-world environment processor:
```python
from lerobot.processor import ObservationProcessorStep
@dataclass
@ProcessorStepRegistry.register(name="libero_processor")
class LiberoProcessorStep(ObservationProcessorStep):
"""
Processes LIBERO observations into the LeRobot format.
**State Processing:**
- Extracts end-effector position (3D)
- Converts quaternion to axis-angle representation (3D)
- Extracts gripper joint positions (2D)
- Concatenates into 8D state vector
**Image Processing:**
- Rotates images 180° to match HuggingFaceVLA/libero convention
"""
def _process_observation(self, observation):
processed_obs = observation.copy()
# Process images: Flip 180° for camera convention
for key in list(processed_obs.keys()):
if key.startswith("observation.images."):
img = processed_obs[key]
img = torch.flip(img, dims=[2, 3]) # Flip H and W
processed_obs[key] = img
# Process robot_state: Flatten to 8D vector
if "observation.robot_state" in processed_obs:
robot_state = processed_obs.pop("observation.robot_state")
eef_pos = robot_state["eef"]["pos"] # (B, 3)
eef_quat = robot_state["eef"]["quat"] # (B, 4)
gripper_qpos = robot_state["gripper"]["qpos"] # (B, 2)
# Convert quaternion to axis-angle
eef_axisangle = self._quat2axisangle(eef_quat) # (B, 3)
# Concatenate into single state vector
state = torch.cat((eef_pos, eef_axisangle, gripper_qpos), dim=-1)
state = state.float()
processed_obs["observation.state"] = state
return processed_obs
```
### Why These Transformations?
1. **Image Rotation**: The HuggingFaceVLA/libero dataset has images rotated 180° from the raw LIBERO simulator. The processor handles this convention mismatch so policies trained on the dataset work seamlessly.
2. **State Flattening**: The raw LIBERO environment exposes nested dictionaries with all available state information (position, quaternion, velocity, matrix representation, etc.). The processor:
- Selects the relevant components (pos, quat, gripper)
- Converts quaternion to axis-angle (more suitable for learning)
- Flattens to a single 8D vector that policies expect
3. **Flexibility**: The environment still exposes **all** raw data. If you want to try different state representations (e.g., including velocities, using matrix representation instead of axis-angle), you can create a new processor without modifying the environment code.
## Adding Environment Processors for New Environments
To add environment processors for a new environment:
### 1. Create the Processor Step
```python
# In src/lerobot/processor/env_processor.py
@dataclass
@ProcessorStepRegistry.register(name="myenv_processor")
class MyEnvProcessorStep(ObservationProcessorStep):
"""Process observations from MyEnv."""
def _process_observation(self, observation):
processed = observation.copy()
# Your environment-specific transformations
if "myenv.specific.state" in processed:
state = processed.pop("myenv.specific.state")
# Transform to standard format
processed["observation.state"] = self._transform_state(state)
return processed
```
### 2. Update Your `EnvConfig` Subclass
```python
# In src/lerobot/envs/factory.py
def make_env_pre_post_processors(env_cfg: EnvConfig):
if isinstance(env_cfg, LiberoEnv) or "libero" in env_cfg.type:
preprocessor = PolicyProcessorPipeline(steps=[LiberoProcessorStep()])
elif isinstance(env_cfg, MyEnvConfig) or "myenv" in env_cfg.type:
preprocessor = PolicyProcessorPipeline(steps=[MyEnvProcessorStep()])
else:
preprocessor = PolicyProcessorPipeline(steps=[])
postprocessor = PolicyProcessorPipeline(steps=[])
return preprocessor, postprocessor
```
### 3. Use in Evaluation
No changes needed! The evaluation script automatically uses the appropriate processor:
```bash
lerobot-eval \
--policy.path=lerobot/my_policy \
--env.type=myenv \ # Automatically uses MyEnvProcessorStep
--eval.n_episodes=10
```
## Future: Environment Postprocessors
Currently, postprocessors are identity (no-op) for all environments. Future use cases include:
### Action Space Transformations
```python
@dataclass
class MyEnvActionPostprocessor(ProcessorStep):
"""Convert policy actions to environment-specific format."""
def __call__(self, transition: EnvTransition) -> EnvTransition:
action = transition["action"]
# Example: Convert from Cartesian to joint space
if self.action_space == "joint":
action = self.ik_solver(action)
# Example: Apply environment-specific safety limits
action = torch.clamp(action, self.min_action, self.max_action)
transition["action"] = action
return transition
```
### Coordinate System Conversions
```python
@dataclass
class CoordinateTransformPostprocessor(ProcessorStep):
"""Transform actions between coordinate systems."""
def __call__(self, transition: EnvTransition) -> EnvTransition:
action = transition["action"]
# Example: Policy outputs in world frame, env expects base frame
action = self.world_to_base_transform(action)
transition["action"] = action
return transition
```
## Best Practices
1. **Keep environment processors simple**: They should only handle environment-specific data format issues, not complex learning-related transformations.
2. **Use policy processors for model requirements**: Normalization, batching, device placement, and tokenization belong in policy processors.
3. **Expose all data from environments**: Let processors decide what to use rather than hardcoding choices in the environment.
4. **Document conventions**: Clearly document any coordinate system conventions, camera orientations, or data formats that your processor handles.
5. **Test independently**: Environment processors should be testable without loading full policies or environments.
## Summary
Environment processors provide a **clean separation** between environment-specific data transformations and policy-specific model requirements. This architecture:
- ✅ Enables easy experimentation with different state representations
- ✅ Allows policies to work seamlessly across different environments
- ✅ Keeps environment code focused on simulation/hardware interface
- ✅ Makes processor pipelines more maintainable and debuggable
- ✅ Follows the single responsibility principle
The key insight: **Environments define data formats, processors standardize them, policies consume standardized data.** Each layer has a clear, focused responsibility.
docs/source/envhub.mdx
+431
View File
@@ -0,0 +1,431 @@
# Loading Environments from the Hub
The **EnvHub** feature allows you to load simulation environments directly from the Hugging Face Hub with a single line of code. This unlocks a powerful new model for collaboration: instead of environments being locked away inside monolithic libraries, anyone can publish custom environments and share them with the community.
## What is EnvHub?
EnvHub lets you create custom robotics simulation environments with your own robot models and scenarios, and make them easily usable by anyone through the LeRobot framework.
EnvHub packages are stored on the Hugging Face Hub, and can be seamlessly pulled and used in your AI robotics projects through LeRobot with a single line of code.
Thanks to EnvHub, you can:
1. **Create and publish environments** to the Hugging Face Hub as Git repositories, and distribute complex physics simulations without packaging hassles
2. **Load environments** dynamically, without installing them as packages
3. **Version and track** environment changes using Git semantics
4. **Discover** new simulation tasks shared by the community
This design means you can go from discovering an interesting environment on the Hub to running experiments in seconds, or create your own custom robot and environment without worrying about dependency conflicts or complex installation procedures.
When you create an EnvHub package, you can build anything you want inside it and use any simulation tool you like: this is your own space to play with. The only requirement is that the package contains an `env.py` file that defines the environment and allows LeRobot to load and use your EnvHub package.
This `env.py` file needs to expose a small API so LeRobot can load and run it. In particular, you must provide a `make_env(n_envs: int = 1, use_async_envs: bool = False)` or `make_env(n_envs: int = 1, use_async_envs: bool = False, cfg: EnvConfig)` function, which is the main entry point for LeRobot. It should return one of:
- A `gym.vector.VectorEnv` (most common)
- A single `gym.Env` (will be automatically wrapped)
- A dict mapping `{suite_name: {task_id: VectorEnv}}` (for multi-task benchmarks)
You can also pass an `EnvConfig` object to `make_env` to configure the environment (e.g. the number of environments, task, camera name, initial states, control mode, episode length, etc.).
Finally, your environment must implement the standard `gym.vector.VectorEnv` interface so it works with LeRobot, including methods like `reset` and `step`.
## Quick Start
Loading an environment from the Hub is as simple as:
```python
from lerobot.envs import make_env
# Load a hub environment (requires explicit consent to run remote code)
env = make_env("lerobot/cartpole-env", trust_remote_code=True)
```
<Tip warning={true}>
**Security Notice**: Loading environments from the Hub executes Python code
from third-party repositories. Only use `trust_remote_code=True` with
repositories you trust. We strongly recommend pinning to a specific commit
hash for reproducibility and security.
</Tip>
## Repository Structure
To make your environment loadable from the Hub, your repository must contain at minimum:
### Required Files
**`env.py`** (or custom Python file)
- Must expose a `make_env(n_envs: int, use_async_envs: bool)` function
- This function should return one of:
- A `gym.vector.VectorEnv` (most common)
- A single `gym.Env` (will be automatically wrapped)
- A dict mapping `{suite_name: {task_id: VectorEnv}}` (for multi-task benchmarks)
### Optional Files
**`requirements.txt`**
- List any additional dependencies your environment needs
- Users will need to install these manually before loading your environment
**`README.md`**
- Document your environment: what task it implements, observation/action spaces, rewards, etc.
- Include usage examples and any special setup instructions
**`.gitignore`**
- Exclude unnecessary files from your repository
### Example Repository Structure
```
my-environment-repo/
├── env.py # Main environment definition (required)
├── requirements.txt # Dependencies (optional)
├── README.md # Documentation (recommended)
├── assets/ # Images, videos, etc. (optional)
│ └── demo.gif
└── configs/ # Config files if needed (optional)
└── task_config.yaml
```
## Creating Your Environment Repository
### Step 1: Define Your Environment
Create an `env.py` file with a `make_env` function:
```python
# env.py
import gymnasium as gym
def make_env(n_envs: int = 1, use_async_envs: bool = False):
"""
Create vectorized environments for your custom task.
Args:
n_envs: Number of parallel environments
use_async_envs: Whether to use AsyncVectorEnv or SyncVectorEnv
Returns:
gym.vector.VectorEnv or dict mapping suite names to vectorized envs
"""
def _make_single_env():
# Create your custom environment
return gym.make("CartPole-v1")
# Choose vector environment type
env_cls = gym.vector.AsyncVectorEnv if use_async_envs else gym.vector.SyncVectorEnv
# Create vectorized environment
vec_env = env_cls([_make_single_env for _ in range(n_envs)])
return vec_env
```
### Step 2: Test Locally
Before uploading, test your environment locally:
```python
from lerobot.envs.utils import _load_module_from_path, _call_make_env, _normalize_hub_result
# Load your module
module = _load_module_from_path("./env.py")
# Test the make_env function
result = _call_make_env(module, n_envs=2, use_async_envs=False)
normalized = _normalize_hub_result(result)
# Verify it works
suite_name = next(iter(normalized))
env = normalized[suite_name][0]
obs, info = env.reset()
print(f"Observation shape: {obs.shape if hasattr(obs, 'shape') else type(obs)}")
env.close()
```
### Step 3: Upload to the Hub
Upload your repository to Hugging Face:
```bash
# Install huggingface_hub if needed
pip install huggingface_hub
# Login to Hugging Face
hf auth login
# Create a new repository
hf repo create my-org/my-custom-env
# Initialize git and push
git init
git add .
git commit -m "Initial environment implementation"
git remote add origin https://huggingface.co/my-org/my-custom-env
git push -u origin main
```
Alternatively, use the `huggingface_hub` Python API:
```python
from huggingface_hub import HfApi
api = HfApi()
# Create repository
api.create_repo("my-custom-env", repo_type="space")
# Upload files
api.upload_folder(
folder_path="./my-env-folder",
repo_id="username/my-custom-env",
repo_type="space",
)
```
## Loading Environments from the Hub
### Basic Usage
```python
from lerobot.envs import make_env
# Load from the hub
envs_dict = make_env(
"username/my-custom-env",
n_envs=4,
trust_remote_code=True
)
# Access the environment
suite_name = next(iter(envs_dict))
env = envs_dict[suite_name][0]
# Use it like any gym environment
obs, info = env.reset()
action = env.action_space.sample()
obs, reward, terminated, truncated, info = env.step(action)
```
### Advanced: Pinning to Specific Versions
For reproducibility and security, pin to a specific Git revision:
```python
# Pin to a specific branch
env = make_env("username/my-env@main", trust_remote_code=True)
# Pin to a specific commit (recommended for papers/experiments)
env = make_env("username/my-env@abc123def456", trust_remote_code=True)
# Pin to a tag
env = make_env("username/my-env@v1.0.0", trust_remote_code=True)
```
### Custom File Paths
If your environment definition is not in `env.py`:
```python
# Load from a custom file
env = make_env("username/my-env:custom_env.py", trust_remote_code=True)
# Combine with version pinning
env = make_env("username/my-env@v1.0:envs/task_a.py", trust_remote_code=True)
```
### Async Environments
For better performance with multiple environments:
```python
envs_dict = make_env(
"username/my-env",
n_envs=8,
use_async_envs=True, # Use AsyncVectorEnv for parallel execution
trust_remote_code=True
)
```
## URL Format Reference
The hub URL format supports several patterns:
| Pattern | Description | Example |
| -------------------- | ------------------------------ | -------------------------------------- |
| `user/repo` | Load `env.py` from main branch | `make_env("lerobot/pusht-env")` |
| `user/repo@revision` | Load from specific revision | `make_env("lerobot/pusht-env@main")` |
| `user/repo:path` | Load custom file | `make_env("lerobot/envs:pusht.py")` |
| `user/repo@rev:path` | Revision + custom file | `make_env("lerobot/envs@v1:pusht.py")` |
## Multi-Task Environments
For benchmarks with multiple tasks (like LIBERO), return a nested dictionary:
```python
def make_env(n_envs: int = 1, use_async_envs: bool = False):
env_cls = gym.vector.AsyncVectorEnv if use_async_envs else gym.vector.SyncVectorEnv
# Return dict: {suite_name: {task_id: VectorEnv}}
return {
"suite_1": {
0: env_cls([lambda: gym.make("Task1-v0") for _ in range(n_envs)]),
1: env_cls([lambda: gym.make("Task2-v0") for _ in range(n_envs)]),
},
"suite_2": {
0: env_cls([lambda: gym.make("Task3-v0") for _ in range(n_envs)]),
}
}
```
## Security Considerations
<Tip warning={true}>
**Important**: The `trust_remote_code=True` flag is required to execute
environment code from the Hub. This is by design for security.
</Tip>
When loading environments from the Hub:
1. **Review the code first**: Visit the repository and inspect `env.py` before loading
2. **Pin to commits**: Use specific commit hashes for reproducibility
3. **Check dependencies**: Review `requirements.txt` for suspicious packages
4. **Use trusted sources**: Prefer official organizations or well-known researchers
5. **Sandbox if needed**: Run untrusted code in isolated environments (containers, VMs)
Example of safe usage:
```python
# ❌ BAD: Loading without inspection
env = make_env("random-user/untrusted-env", trust_remote_code=True)
# ✅ GOOD: Review code, then pin to specific commit
# 1. Visit https://huggingface.co/trusted-org/verified-env
# 2. Review the env.py file
# 3. Copy the commit hash
env = make_env("trusted-org/verified-env@a1b2c3d4", trust_remote_code=True)
```
## Example: CartPole from the Hub
Here's a complete example using the reference CartPole environment:
```python
from lerobot.envs import make_env
import numpy as np
# Load the environment
envs_dict = make_env("lerobot/cartpole-env", n_envs=4, trust_remote_code=True)
# Get the vectorized environment
suite_name = next(iter(envs_dict))
env = envs_dict[suite_name][0]
# Run a simple episode
obs, info = env.reset()
done = np.zeros(env.num_envs, dtype=bool)
total_reward = np.zeros(env.num_envs)
while not done.all():
# Random policy
action = env.action_space.sample()
obs, reward, terminated, truncated, info = env.step(action)
total_reward += reward
done = terminated | truncated
print(f"Average reward: {total_reward.mean():.2f}")
env.close()
```
## Benefits of EnvHub
### For Environment Authors
- **Easy distribution**: No PyPI packaging required
- **Version control**: Use Git for environment versioning
- **Rapid iteration**: Push updates instantly
- **Documentation**: Hub README renders beautifully
- **Community**: Reach LeRobot users directly
### For Researchers
- **Quick experiments**: Load any environment in one line
- **Reproducibility**: Pin to specific commits
- **Discovery**: Browse environments on the Hub
- **No conflicts**: No need to install conflicting packages
### For the Community
- **Growing ecosystem**: More diverse simulation tasks
- **Standardization**: Common `make_env` API
- **Collaboration**: Fork and improve existing environments
- **Accessibility**: Lower barrier to sharing research
## Troubleshooting
### "Refusing to execute remote code"
You must explicitly pass `trust_remote_code=True`:
```python
env = make_env("user/repo", trust_remote_code=True)
```
### "Module X not found"
The hub environment has dependencies you need to install:
```bash
# Check the repo's requirements.txt and install dependencies
pip install gymnasium numpy
```
### "make_env not found in module"
Your `env.py` must expose a `make_env` function:
```python
def make_env(n_envs: int, use_async_envs: bool):
# Your implementation
pass
```
### Environment returns wrong type
The `make_env` function must return:
- A `gym.vector.VectorEnv`, or
- A single `gym.Env`, or
- A dict `{suite_name: {task_id: VectorEnv}}`
## Best Practices
1. **Document your environment**: Include observation/action space descriptions, reward structure, and termination conditions in your README
2. **Add requirements.txt**: List all dependencies with versions
3. **Test thoroughly**: Verify your environment works locally before pushing
4. **Use semantic versioning**: Tag releases with version numbers
5. **Add examples**: Include usage examples in your README
6. **Keep it simple**: Minimize dependencies when possible
7. **License your work**: Add a LICENSE file to clarify usage terms
## Future Directions
The EnvHub ecosystem enables exciting possibilities:
- **GPU-accelerated physics**: Share Isaac Gym or Brax environments
- **Photorealistic rendering**: Distribute environments with advanced graphics
- **Multi-agent scenarios**: Complex interaction tasks
- **Real-world simulators**: Digital twins of physical setups
- **Procedural generation**: Infinite task variations
- **Domain randomization**: Pre-configured DR pipelines
As more researchers and developers contribute, the diversity and quality of available environments will grow, benefiting the entire robotics learning community.
## See Also
- [Hugging Face Hub Documentation](https://huggingface.co/docs/hub/en/index)
- [Gymnasium Documentation](https://gymnasium.farama.org/index.html)
- [Example Hub Environment](https://huggingface.co/lerobot/cartpole-env)
docs/source/envhub_isaaclab_arena.mdx
+510
View File
@@ -0,0 +1,510 @@
# NVIDIA IsaacLab Arena & LeRobot
LeRobot EnvHub now supports **GPU-accelerated simulation** with IsaacLab Arena for policy evaluation at scale.
Train and evaluate imitation learning policies with high-fidelity simulation — all integrated into the LeRobot ecosystem.
<img
src="https://huggingface.co/nvidia/isaaclab-arena-envs/resolve/main/assets/Gr1OpenMicrowaveEnvironment.png"
alt="IsaacLab Arena - GR1 Microwave Environment"
style={{ maxWidth: "100%", borderRadius: "8px", marginBottom: "1rem" }}
/>
[IsaacLab Arena](https://github.com/isaac-sim/IsaacLab-Arena) integrates with NVIDIA IsaacLab to provide:
- 🤖 **Humanoid embodiments**: GR1, G1, Galileo with various configurations
- 🎯 **Manipulation & loco-manipulation tasks**: Door opening, pick-and-place, button pressing, and more
- ⚡ **GPU-accelerated rollouts**: Parallel environment execution on NVIDIA GPUs
- 🖼️ **RTX Rendering**: Evaluate vision-based policies with realistic rendering, reflections and refractions
- 📦 **LeRobot-compatible datasets**: Ready for training with GR00T N1x, PI0, SmolVLA, ACT, and Diffusion policies
- 🔄 **EnvHub integration**: Load environments from HuggingFace EnvHub with one line
## Installation
### Prerequisites
Hardware requirements are shared with Isaac Sim, and are detailed in [Isaac Sim Requirements](https://docs.isaacsim.omniverse.nvidia.com/5.1.0/installation/requirements.html).
- NVIDIA GPU with CUDA support
- NVIDIA driver compatible with IsaacSim 5.1.0
- Linux (Ubuntu 22.04 / 24.04)
### Setup
```bash
# 1. Create conda environment
conda create -y -n lerobot-arena python=3.11
conda activate lerobot-arena
conda install -y -c conda-forge ffmpeg=7.1.1
# 2. Install Isaac Sim 5.1.0
pip install "isaacsim[all,extscache]==5.1.0" --extra-index-url https://pypi.nvidia.com
# Accept NVIDIA EULA (required)
export ACCEPT_EULA=Y
export PRIVACY_CONSENT=Y
# 3. Install IsaacLab 2.3.0
git clone https://github.com/isaac-sim/IsaacLab.git
cd IsaacLab
git checkout v2.3.0
./isaaclab.sh -i
cd ..
# 4. Install IsaacLab Arena
git clone https://github.com/isaac-sim/IsaacLab-Arena.git
cd IsaacLab-Arena
git checkout release/0.1.1
pip install -e .
cd ..
# 5. Install LeRobot (evaluation extra for env/policy evaluation)
git clone https://github.com/huggingface/lerobot.git
cd lerobot
pip install -e ".[evaluation]"
cd ..
# 6. Install additional dependencies
pip install onnxruntime==1.23.2 lightwheel-sdk==1.0.1 vuer[all]==0.0.70 qpsolvers==4.8.1
pip install numpy==1.26.0 # Isaac Sim 5.1 depends on numpy==1.26.0, this will be fixed in next release
```
## Evaluating Policies
### Pre-trained Policies
The following trained policies are available:
| Policy | Architecture | Task | Link |
| :-------------------------- | :----------- | :------------ | :----------------------------------------------------------------------- |
| pi05-arena-gr1-microwave | PI0.5 | GR1 Microwave | [HuggingFace](https://huggingface.co/nvidia/pi05-arena-gr1-microwave) |
| smolvla-arena-gr1-microwave | SmolVLA | GR1 Microwave | [HuggingFace](https://huggingface.co/nvidia/smolvla-arena-gr1-microwave) |
### Evaluate SmolVLA
```bash
pip install -e ".[smolvla]"
pip install numpy==1.26.0 # revert numpy to version 1.26
```
```bash
lerobot-eval \
--policy.path=nvidia/smolvla-arena-gr1-microwave \
--env.type=isaaclab_arena \
--env.hub_path=nvidia/isaaclab-arena-envs \
--rename_map='{"observation.images.robot_pov_cam_rgb": "observation.images.robot_pov_cam"}' \
--policy.device=cuda \
--env.environment=gr1_microwave \
--env.embodiment=gr1_pink \
--env.object=mustard_bottle \
--env.headless=false \
--env.enable_cameras=true \
--env.video=true \
--env.video_length=10 \
--env.video_interval=15 \
--env.state_keys=robot_joint_pos \
--env.camera_keys=robot_pov_cam_rgb \
--trust_remote_code=True \
--eval.batch_size=1
```
### Evaluate PI0.5
```bash
pip install -e ".[pi]"
pip install numpy==1.26.0 # revert numpy to version 1.26
```
<Tip>PI0.5 requires disabling torch compile for evaluation:</Tip>
```bash
TORCH_COMPILE_DISABLE=1 TORCHINDUCTOR_DISABLE=1 lerobot-eval \
--policy.path=nvidia/pi05-arena-gr1-microwave \
--env.type=isaaclab_arena \
--env.hub_path=nvidia/isaaclab-arena-envs \
--rename_map='{"observation.images.robot_pov_cam_rgb": "observation.images.robot_pov_cam"}' \
--policy.device=cuda \
--env.environment=gr1_microwave \
--env.embodiment=gr1_pink \
--env.object=mustard_bottle \
--env.headless=false \
--env.enable_cameras=true \
--env.video=true \
--env.video_length=15 \
--env.video_interval=15 \
--env.state_keys=robot_joint_pos \
--env.camera_keys=robot_pov_cam_rgb \
--trust_remote_code=True \
--eval.batch_size=1
```
<Tip>
To change the number of parallel environments, use the ```--eval.batch_size```
flag.
</Tip>
### What to Expect
During evaluation, you will see a progress bar showing the running success rate:
```
Stepping through eval batches: 8%|██████▍ | 4/50 [00:45<08:06, 10.58s/it, running_success_rate=25.0%]
```
### Video Recording
To enable video recording during evaluation, add the following flags to your command:
```bash
--env.video=true \
--env.video_length=15 \
--env.video_interval=15
```
For more details on video recording, see the [IsaacLab Recording Documentation](https://isaac-sim.github.io/IsaacLab/main/source/how-to/record_video.html).
<Tip>
When running headless with `--env.headless=true`, you must also enable cameras explicitly for camera enabled environments:
```bash
--env.headless=true --env.enable_cameras=true
```
</Tip>
### Output Directory
Evaluation videos are saved to the output directory with the following structure:
```
outputs/eval/<date>/<timestamp>_<env>_<policy>/videos/<task>_<env_id>/eval_episode_<n>.mp4
```
For example:
```
outputs/eval/2026-01-02/14-38-01_isaaclab_arena_smolvla/videos/gr1_microwave_0/eval_episode_0.mp4
```
## Training Policies
To learn more about training policies with LeRobot, please refer to the training documentation:
- [SmolVLA](./smolvla)
- [Pi0.5](./pi05)
- [GR00T N1.7](./groot)
Sample IsaacLab Arena datasets are available on HuggingFace Hub for experimentation:
| Dataset | Description | Frames |
| :-------------------------------------------------------------------------------------------------------- | :------------------------- | :----- |
| [Arena-GR1-Manipulation-Task](https://huggingface.co/datasets/nvidia/Arena-GR1-Manipulation-Task-v3) | GR1 microwave manipulation | ~4K |
| [Arena-G1-Loco-Manipulation-Task](https://huggingface.co/datasets/nvidia/Arena-G1-Loco-Manipulation-Task) | G1 loco-manipulation | ~4K |
## Environment Configuration
### Full Configuration Options
```python
from lerobot.envs.configs import IsaaclabArenaEnv
config = IsaaclabArenaEnv(
# Environment selection
environment="gr1_microwave", # Task environment
embodiment="gr1_pink", # Robot embodiment
object="power_drill", # Object to manipulate
# Simulation settings
episode_length=300, # Max steps per episode
headless=True, # Run without GUI
device="cuda:0", # GPU device
seed=42, # Random seed
# Observation configuration
state_keys="robot_joint_pos", # State observation keys (comma-separated)
camera_keys="robot_pov_cam_rgb", # Camera observation keys (comma-separated)
state_dim=54, # Expected state dimension
action_dim=36, # Expected action dimension
camera_height=512, # Camera image height
camera_width=512, # Camera image width
enable_cameras=True, # Enable camera observations
# Video recording
video=False, # Enable video recording
video_length=100, # Frames per video
video_interval=200, # Steps between recordings
# Advanced
mimic=False, # Enable mimic mode
teleop_device=None, # Teleoperation device
disable_fabric=False, # Disable fabric optimization
enable_pinocchio=True, # Enable Pinocchio for IK
)
```
### Using Environment Hub directly for advanced usage
Create a file called `test_env_load_arena.py` or [download from the EnvHub](https://huggingface.co/nvidia/isaaclab-arena-envs/blob/main/tests/test_env_load_arena.py):
```python
import logging
from dataclasses import asdict
from pprint import pformat
import torch
import tqdm
from lerobot.configs import parser
from lerobot.configs.eval import EvalPipelineConfig
@parser.wrap()
def main(cfg: EvalPipelineConfig):
"""Run random action rollout for IsaacLab Arena environment."""
logging.info(pformat(asdict(cfg)))
from lerobot.envs import make_env
env_dict = make_env(
cfg.env,
n_envs=cfg.env.num_envs,
trust_remote_code=True,
)
env = next(iter(env_dict.values()))[0]
env.reset()
for _ in tqdm.tqdm(range(cfg.env.episode_length)):
with torch.inference_mode():
actions = env.action_space.sample()
obs, rewards, terminated, truncated, info = env.step(actions)
if terminated.any() or truncated.any():
obs, info = env.reset()
env.close()
if __name__ == "__main__":
main()
```
Run with:
```bash
python test_env_load_arena.py \
--env.environment=g1_locomanip_pnp \
--env.embodiment=gr1_pink \
--env.object=cracker_box \
--env.num_envs=4 \
--env.enable_cameras=true \
--env.seed=1000 \
--env.video=true \
--env.video_length=10 \
--env.video_interval=15 \
--env.headless=false \
--env.hub_path=nvidia/isaaclab-arena-envs \
--env.type=isaaclab_arena
```
## Creating New Environments
First create a new IsaacLab Arena environment by following the [IsaacLab Arena Documentation](https://isaac-sim.github.io/IsaacLab-Arena/release/0.1.1/index.html).
Clone our EnvHub repo:
```bash
git clone https://huggingface.co/nvidia/isaaclab-arena-envs
```
Modify the `example_envs.yaml` file based on your new environment.
[Upload](./envhub#step-3-upload-to-the-hub) your modified repo to HuggingFace EnvHub.
<Tip>
Your IsaacLab Arena environment code must be locally available during
evaluation. Users can clone your environment repository separately, or you can
bundle the environment code and assets directly in your EnvHub repo.
</Tip>
Then, when evaluating, use your new environment:
```bash
lerobot-eval \
--env.hub_path=<your-env-hub-path>/isaaclab-arena-envs \
--env.environment=<your new environment> \
...other flags...
```
We look forward to your contributions!
## Troubleshooting
### CUDA out of memory
Reduce `batch_size` or use a GPU with more VRAM:
```bash
--eval.batch_size=1
```
### EULA not accepted
Set environment variables before running:
```bash
export ACCEPT_EULA=Y
export PRIVACY_CONSENT=Y
```
### Video recording not working
Enable cameras when running headless:
```bash
--env.video=true --env.enable_cameras=true --env.headless=true
```
### Policy output dimension mismatch
Ensure `action_dim` matches your policy:
```bash
--env.action_dim=36
```
### libGLU.so.1 Errors during Isaac Sim initialization
Ensure you have the following dependencies installed, this is likely to happen on headless machines.
```bash
sudo apt update && sudo apt install -y libglu1-mesa libxt6
```
## See Also
- [EnvHub Documentation](./envhub.mdx) - General EnvHub usage
- [IsaacLab Arena GitHub](https://github.com/isaac-sim/IsaacLab-Arena)
- [IsaacLab Documentation](https://isaac-sim.github.io/IsaacLab/)
## Lightwheel LW-BenchHub
[Lightwheel](https://www.lightwheel.ai) is bringing `Lightwheel-Libero-Tasks` and `Lightwheel-RoboCasa-Tasks` with 268 tasks to the LeRobot ecosystem.
LW-BenchHub collects and generates large-scale datasets via teleoperation that comply with the LeRobot specification, enabling out-of-the-box training and evaluation workflows.
With the unified interface provided by EnvHub, developers can quickly build end-to-end experimental pipelines.
### Install
Assuming you followed the [Installation](#installation) steps, you can install LW-BenchHub with:
```bash
conda install pinocchio -c conda-forge -y
pip install numpy==1.26.0 # revert numpy to version 1.26
sudo apt-get install git-lfs && git lfs install
git clone https://github.com/LightwheelAI/lw_benchhub
git lfs pull # Ensure LFS files (e.g., .usd assets) are downloaded
cd lw_benchhub
pip install -e .
```
For more detailed instructions, please refer to the [LW-BenchHub Documentation](https://docs.lightwheel.net/lw_benchhub/usage/Installation).
### Lightwheel Tasks Dataset
LW-BenchHub datasets are available on HuggingFace Hub:
| Dataset | Description | Tasks | Frames |
| :------------------------------------------------------------------------------------------------------------ | :---------------------- | :---- | :----- |
| [Lightwheel-Tasks-X7S](https://huggingface.co/datasets/LightwheelAI/Lightwheel-Tasks-X7S) | X7S LIBERO and RoboCasa | 117 | ~10.3M |
| [Lightwheel-Tasks-Double-Piper](https://huggingface.co/datasets/LightwheelAI/Lightwheel-Tasks-Double-Piper) | Double-Piper LIBERO | 130 | ~6.0M |
| [Lightwheel-Tasks-G1-Controller](https://huggingface.co/datasets/LightwheelAI/Lightwheel-Tasks-G1-Controller) | G1-Controller LIBERO | 62 | ~2.7M |
| [Lightwheel-Tasks-G1-WBC](https://huggingface.co/datasets/LightwheelAI/Lightwheel-Tasks-G1-WBC) | G1-WBC RoboCasa | 32 | ~1.5M |
For training policies, refer to the [Training Policies](#training-policies) section.
### Evaluating Policies
#### Pre-trained Policies
The following trained policies are available:
| Policy | Architecture | Task | Layout | Robot | Link |
| :----------------------- | :----------- | :----------------------------- | :--------- | :-------------- | :------------------------------------------------------------------------------------ |
| smolvla-double-piper-pnp | SmolVLA | L90K1PutTheBlackBowlOnThePlate | libero-1-1 | DoublePiper-Abs | [HuggingFace](https://huggingface.co/LightwheelAI/smolvla-double-piper-pnp/tree/main) |
#### Evaluate SmolVLA
```bash
lerobot-eval \
--policy.path=LightwheelAI/smolvla-double-piper-pnp \
--env.type=isaaclab_arena \
--rename_map='{"observation.images.left_hand_camera_rgb": "observation.images.left_hand", "observation.images.right_hand_camera_rgb": "observation.images.right_hand", "observation.images.first_person_camera_rgb": "observation.images.first_person"}' \
--env.hub_path=LightwheelAI/lw_benchhub_env \
--env.kwargs='{"config_path": "configs/envhub/example.yml"}' \
--trust_remote_code=true \
--env.state_keys=joint_pos \
--env.action_dim=12 \
--env.camera_keys=left_hand_camera_rgb,right_hand_camera_rgb,first_person_camera_rgb \
--policy.device=cuda \
--eval.batch_size=10 \
--eval.n_episodes=100
```
### Environment Configuration
Evaluation can be quickly launched by modifying the `robot`, `task`, and `layout` settings in the configuration file.
#### Full Configuration Options
```yml
# =========================
# Basic Settings
# =========================
disable_fabric: false
device: cuda:0
sensitivity: 1.0
step_hz: 50
enable_cameras: true
execute_mode: eval
episode_length_s: 20.0 # Episode length in seconds, increase if episodes timeout during eval
# =========================
# Robot Settings
# =========================
robot: DoublePiper-Abs # Robot type, DoublePiper-Abs, X7S-Abs, G1-Controller or G1-Controller-DecoupledWBC
robot_scale: 1.0
# =========================
# Task & Scene Settings
# =========================
task: L90K1PutTheBlackBowlOnThePlate # Task name
scene_backend: robocasa
task_backend: robocasa
debug_assets: null
layout: libero-1-1 # Layout and style ID
sources:
- objaverse
- lightwheel
- aigen_objs
object_projects: []
usd_simplify: false
seed: 42
# =========================
# Object Placement Retry Settings
# =========================
max_scene_retry: 4
max_object_placement_retry: 3
resample_objects_placement_on_reset: true
resample_robot_placement_on_reset: true
# =========================
# Replay Configuration Settings
# =========================
replay_cfgs:
add_camera_to_observation: true
render_resolution: [640, 480]
```
### See Also
- [LW-BenchHub GitHub](https://github.com/LightwheelAI/LW-BenchHub)
- [LW-BenchHub Documentation](https://docs.lightwheel.net/lw_benchhub/)
docs/source/envhub_leisaac.mdx
+302
View File
@@ -0,0 +1,302 @@
# LeIsaac × LeRobot EnvHub
LeRobot EnvHub now supports **imitation learning in simulation** with LeIsaac.
Spin up everyday manipulation tasks, teleoperate the robot, collect demos, push them to the Hub, and train policies in LeRobot — all in one loop.
[LeIsaac](https://github.com/LightwheelAI/leisaac) integrates with IsaacLab and the SO101 Leader/Follower setup to provide:
- 🕹️ **Teleoperation-first workflows** for data collection
- 📦 **Built-in data conversion** ready for LeRobot training
- 🤖 **Everyday skills** like picking oranges, lifting cubes, cleaning tables, and folding cloth
- ☁️ **Ongoing upgrades** from [LightWheel](https://lightwheel.ai/): cloud simulation, EnvHub support, Sim2Real tooling, and more
Below you’ll find the currently supported LeIsaac tasks exposed through LeRobot EnvHub.
# Available Environments
The following table lists all available tasks and environments in LeIsaac x LeRobot Envhub. You can also get the latest list of environments by running the following command:
```bash
python scripts/environments/list_envs.py
```
| Task | Environment ID | Task Description | Related Robot |
| :-------------------------------------------------------------------------------------------------------------------------------------------------------------- | :-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | :------------------------------------------------------------------------------------------------------------------------- | :--------------------------------------------------------- |
| <video src="https://github.com/user-attachments/assets/466eddff-f720-4f99-94d5-5e123e4c302c" autoplay loop muted playsinline style="max-width: 300px;"></video> | [LeIsaac-SO101-PickOrange-v0](https://github.com/LightwheelAI/leisaac/blob/main/source/leisaac/leisaac/tasks/pick_orange/pick_orange_env_cfg.py)<br /><br />[LeIsaac-SO101-PickOrange-Direct-v0](https://github.com/LightwheelAI/leisaac/blob/main/source/leisaac/leisaac/tasks/pick_orange/direct/pick_orange_env.py) | Pick three oranges and put them into the plate, then reset the arm to rest state. | Single-Arm SO101 Follower |
| <video src="https://github.com/user-attachments/assets/1e4eb83a-0b38-40fb-a0b2-ddb0fe201e6d" autoplay loop muted playsinline style="max-width: 300px;"></video> | [LeIsaac-SO101-LiftCube-v0](https://github.com/LightwheelAI/leisaac/blob/main/source/leisaac/leisaac/tasks/lift_cube/lift_cube_env_cfg.py)<br /><br />[LeIsaac-SO101-LiftCube-Direct-v0](https://github.com/LightwheelAI/leisaac/blob/main/source/leisaac/leisaac/tasks/lift_cube/direct/lift_cube_env.py) | Lift the red cube up. | Single-Arm SO101 Follower |
| <video src="https://github.com/user-attachments/assets/e49d8f1c-dcc9-412b-a88f-100680d8a45b" autoplay loop muted playsinline style="max-width: 300px;"></video> | [LeIsaac-SO101-CleanToyTable-v0](https://github.com/LightwheelAI/leisaac/blob/main/source/leisaac/leisaac/tasks/clean_toy_table/clean_toy_table_env_cfg.py)<br /><br />[LeIsaac-SO101-CleanToyTable-BiArm-v0](https://github.com/LightwheelAI/leisaac/blob/main/source/leisaac/leisaac/tasks/clean_toy_table/clean_toy_table_bi_arm_env_cfg.py)<br /><br />[LeIsaac-SO101-CleanToyTable-BiArm-Direct-v0](https://github.com/LightwheelAI/leisaac/blob/main/source/leisaac/leisaac/tasks/clean_toy_table/direct/clean_toy_table_bi_arm_env.py) | Pick two letter e objects into the box, and reset the arm to rest state. | Single-Arm SO101 Follower<br /><br />Bi-Arm SO101 Follower |
| <video src="https://github.com/user-attachments/assets/e29a0f8a-9286-4ce6-b45d-342c3d3ba754" autoplay loop muted playsinline style="max-width: 300px;"></video> | [LeIsaac-SO101-FoldCloth-BiArm-v0](https://github.com/LightwheelAI/leisaac/blob/main/source/leisaac/leisaac/tasks/fold_cloth/fold_cloth_bi_arm_env_cfg.py)<br /><br />[LeIsaac-SO101-FoldCloth-BiArm-Direct-v0](https://github.com/LightwheelAI/leisaac/blob/main/source/leisaac/leisaac/tasks/fold_cloth/direct/fold_cloth_bi_arm_env.py) | Fold the cloth, and reset the arm to rest state.<br /><br />_Note: Only the DirectEnv support check_success in this task._ | Bi-Arm SO101 Follower |
# Load LeIsaac directly in LeRobot with one line of code
> EnvHub: Share LeIsaac environments through HuggingFace
[EnvHub](https://huggingface.co/docs/lerobot/envhub) is our reproducible environment hub, spin up a packaged simulation with one line, experiment immediately, and publish your own tasks for the community.
LeIsaac offers EnvHub support so you can consume or share tasks with only a few commands.
<video
controls
src="https://github.com/user-attachments/assets/687666f5-ebe0-421d-84a0-eb86116ac5f8"
style={{ width: "100%", maxWidth: "960px", borderRadius: "8px" }}
/>
## How to get started, environment Setup
Run the following commands to setup your code environments:
```bash
# Refer to Getting Started/Installation to install leisaac firstly
conda create -n leisaac_envhub python=3.11
conda activate leisaac_envhub
conda install -c "nvidia/label/cuda-12.8.1" cuda-toolkit
pip install -U torch==2.7.0 torchvision==0.22.0 --index-url https://download.pytorch.org/whl/cu128
pip install 'leisaac[isaaclab] @ git+https://github.com/LightwheelAI/leisaac.git#subdirectory=source/leisaac' --extra-index-url https://pypi.nvidia.com
# Install lerobot
pip install lerobot==0.4.1
# Fix numpy version
pip install numpy==1.26.0
```
## Usage Example
EnvHub exposes every LeIsaac-supported task in a uniform interface. The examples below load `so101_pick_orange` and demonstrate a random-action rollout and an interactive teleoperation.
### Random Action
<details>
<summary>Click to expand code example</summary>
```python
# envhub_random_action.py
import torch
from lerobot.envs import make_env
# Load from the hub
envs_dict = make_env("LightwheelAI/leisaac_env:envs/so101_pick_orange.py", n_envs=1, trust_remote_code=True)
# Access the environment
suite_name = next(iter(envs_dict))
sync_vector_env = envs_dict[suite_name][0]
# retrieve the isaac environment from the sync vector env
env = sync_vector_env.envs[0].unwrapped
# Use it like any gym environment
obs, info = env.reset()
while True:
action = torch.tensor(env.action_space.sample())
obs, reward, terminated, truncated, info = env.step(action)
if terminated or truncated:
obs, info = env.reset()
env.close()
```
</details>
```bash
python envhub_random_action.py
```
You should see the SO101 arm swinging under purely random commands.
### Teleoperation
LeRobot’s teleoperation stack can drive the simulated arm.
Connect the SO101 Leader controller, run the calibration command below.
```bash
lerobot-calibrate \
--teleop.type=so101_leader \
--teleop.port=/dev/ttyACM0 \
--teleop.id=leader
```
And then launch the teleop script.
<details>
<summary>Click to expand code example</summary>
```python
# envhub_teleop_example.py
import logging
import time
import gymnasium as gym
from dataclasses import asdict, dataclass
from pprint import pformat
from lerobot.teleoperators import ( # noqa: F401
Teleoperator,
TeleoperatorConfig,
make_teleoperator_from_config,
so_leader,
bi_so_leader,
)
from lerobot.utils.robot_utils import precise_sleep
from lerobot.utils.utils import init_logging
from lerobot.envs import make_env
@dataclass
class TeleoperateConfig:
teleop: TeleoperatorConfig
env_name: str = "so101_pick_orange"
fps: int = 60
@dataclass
class EnvWrap:
env: gym.Env
def make_env_from_leisaac(env_name: str = "so101_pick_orange"):
envs_dict = make_env(
f'LightwheelAI/leisaac_env:envs/{env_name}.py',
n_envs=1,
trust_remote_code=True
)
suite_name = next(iter(envs_dict))
sync_vector_env = envs_dict[suite_name][0]
env = sync_vector_env.envs[0].unwrapped
return env
def teleop_loop(teleop: Teleoperator, env: gym.Env, fps: int):
from leisaac.devices.action_process import preprocess_device_action
from leisaac.assets.robots.lerobot import SO101_FOLLOWER_MOTOR_LIMITS
from leisaac.utils.env_utils import dynamic_reset_gripper_effort_limit_sim
env_wrap = EnvWrap(env=env)
obs, info = env.reset()
while True:
loop_start = time.perf_counter()
if env.cfg.dynamic_reset_gripper_effort_limit:
dynamic_reset_gripper_effort_limit_sim(env, 'so101leader')
raw_action = teleop.get_action()
processed_action = preprocess_device_action(
dict(
so101_leader=True,
joint_state={
k.removesuffix(".pos"): v for k, v in raw_action.items()},
motor_limits=SO101_FOLLOWER_MOTOR_LIMITS),
env_wrap
)
obs, reward, terminated, truncated, info = env.step(processed_action)
if terminated or truncated:
obs, info = env.reset()
dt_s = time.perf_counter() - loop_start
precise_sleep(max(1 / fps - dt_s, 0.0))
loop_s = time.perf_counter() - loop_start
print(f"\ntime: {loop_s * 1e3:.2f}ms ({1 / loop_s:.0f} Hz)")
def teleoperate(cfg: TeleoperateConfig):
init_logging()
logging.info(pformat(asdict(cfg)))
teleop = make_teleoperator_from_config(cfg.teleop)
env = make_env_from_leisaac(cfg.env_name)
teleop.connect()
if hasattr(env, 'initialize'):
env.initialize()
try:
teleop_loop(teleop=teleop, env=env, fps=cfg.fps)
except KeyboardInterrupt:
pass
finally:
teleop.disconnect()
env.close()
def main():
teleoperate(TeleoperateConfig(
teleop=so_leader.SO101LeaderConfig(
port="/dev/ttyACM0",
id='leader',
use_degrees=False,
),
env_name="so101_pick_orange",
fps=60,
))
if __name__ == "__main__":
main()
```
</details>
```bash
python envhub_teleop_example.py
```
Running the script lets you operate the simulated arm using the physical Leader device.
## ☁️ Cloud Simulation (No GPU Required)
Don’t have a local GPU or the right drivers? No problem! You can run LeIsaac entirely in the cloud with zero setup.
LeIsaac works out-of-the-box on **NVIDIA Brev**, giving you a fully configured environment directly in your browser.
👉 **Start here:** [https://lightwheelai.github.io/leisaac/docs/cloud_simulation/nvidia_brev](https://lightwheelai.github.io/leisaac/docs/cloud_simulation/nvidia_brev)
Once your instance is deployed, simply open the link for **port 80 (HTTP)** to launch **Visual Studio Code Server** (default password: `password`). From there, you can run simulations, edit code, and visualize IsaacLab environments — all from your web browser.
**No GPU, no drivers, no local installation. Just click and run.**
## Additional Notes
We keep EnvHub coverage aligned with the LeIsaac task. Currently supported:
- `so101_pick_orange`
- `so101_lift_cube`
- `so101_clean_toytable`
- `bi_so101_fold_cloth`
Switch tasks by targeting a different script when calling `make_env`, for example:
```python
envs_dict_pick_orange = make_env("LightwheelAI/leisaac_env:envs/so101_pick_orange.py", n_envs=1, trust_remote_code=True)
envs_dict_lift_cube = make_env("LightwheelAI/leisaac_env:envs/so101_lift_cube.py", n_envs=1, trust_remote_code=True)
envs_dict_clean_toytable = make_env("LightwheelAI/leisaac_env:envs/so101_clean_toytable.py", n_envs=1, trust_remote_code=True)
envs_dict_fold_cloth = make_env("LightwheelAI/leisaac_env:envs/bi_so101_fold_cloth.py", n_envs=1, trust_remote_code=True)
```
Note: when working with `bi_so101_fold_cloth`, call `initialize()` immediately after retrieving the env before performing any other operations:
<details>
<summary>Click to expand code example</summary>
```python
import torch
from lerobot.envs import make_env
# Load from the hub
envs_dict = make_env("LightwheelAI/leisaac_env:envs/bi_so101_fold_cloth.py", n_envs=1, trust_remote_code=True)
# Access the environment
suite_name = next(iter(envs_dict))
sync_vector_env = envs_dict[suite_name][0]
# retrieve the isaac environment from the sync vector env
env = sync_vector_env.envs[0].unwrapped
# NOTE: initialize() first
env.initialize()
# other operation with env...
```
</details>
docs/source/eo1.mdx
+168
View File
@@ -0,0 +1,168 @@
# EO-1
EO-1 is a **Vision-Language-Action policy for robot control**. The LeRobot implementation integrates EO-1 with the standard LeRobot training, evaluation, processor interface.
## Model Overview
EO-1 uses a Qwen2.5-VL backbone for vision-language understanding and adds a continuous flow-matching action head for robot control. The policy formats each robot-control sample as a multimodal conversation: camera images are passed to Qwen2.5-VL, the robot state is represented with EO-1 state tokens, and the future action chunk is represented with EO-1 action tokens.
<img
src="https://huggingface.co/datasets/HaomingSong/lerobot-documentation-images/resolve/main/lerobot/eo_pipeline.png"
alt="An overview of EO-1"
width="85%"
/>
During training, EO-1 learns to denoise continuous action chunks at the action-token positions. During inference, it samples an action chunk, returns continuous actions, and executes `n_action_steps` from the chunk before sampling again.
### What the LeRobot Integration Covers
- Standard `policy.type=eo1` configuration through LeRobot
- Qwen2.5-VL image and text preprocessing through policy processors
- Continuous flow-matching action prediction
- Checkpoint save/load through LeRobot policy APIs
- Training with `lerobot-train` and evaluation with `lerobot-eval`
The broader EO-1 project also includes interleaved vision-text-action pretraining and multimodal reasoning workflows. This page focuses on the LeRobot robot-control policy path.
## Installation Requirements
1. Install LeRobot by following the [Installation Guide](./installation).
2. Install EO-1 dependencies by running:
```bash
pip install -e ".[eo1]"
```
3. If you want to train or evaluate on LIBERO, install the LIBERO dependencies too:
```bash
pip install -e ".[eo1,libero]"
```
EO-1 can use the standard PyTorch scaled-dot-product attention backend through `policy.attn_implementation=sdpa`. If your environment has a compatible `flash_attn` installation, you can request `policy.attn_implementation=flash_attention_2`.
## Data Requirements
EO-1 expects a LeRobot dataset with:
- At least one visual observation, for example `observation.images.image`
- `observation.state`
- `action`
- A language task instruction through the dataset `task` field
If your dataset uses different observation names, use `rename_map` to align them with the names expected by your training or evaluation setup.
## Usage
To use EO-1 in a LeRobot configuration, specify the policy type as:
```python
policy.type=eo1
```
By default, a new EO-1 policy initializes its backbone from:
```python
policy.vlm_base=Qwen/Qwen2.5-VL-3B-Instruct
```
Once a LeRobot-format EO-1 checkpoint is available, load it with:
```python
policy.path=your-org/your-eo1-checkpoint
```
## Training
### Training Command Example
```bash
lerobot-train \
--dataset.repo_id=your_org/your_dataset \
--policy.type=eo1 \
--policy.vlm_base=Qwen/Qwen2.5-VL-3B-Instruct \
--policy.dtype=bfloat16 \
--policy.attn_implementation=sdpa \
--policy.gradient_checkpointing=false \
--output_dir=./outputs/eo1_training \
--job_name=eo1_training \
--steps=300000 \
--batch_size=16 \
--policy.device=cuda
```
### Key Training Parameters
| Parameter | Default | Description |
| -------------------------------------- | ----------------------------- | ----------------------------------------------------------------------- |
| `policy.vlm_base` | `Qwen/Qwen2.5-VL-3B-Instruct` | Qwen2.5-VL checkpoint used to initialize a new policy |
| `policy.dtype` | `auto` | Backbone dtype request: `auto`, `bfloat16`, or `float32` |
| `policy.attn_implementation` | `None` | Optional Qwen attention backend, such as `sdpa` |
| `policy.gradient_checkpointing` | `false` | Reduces memory usage during training |
| `policy.chunk_size` | `8` | Number of future actions predicted per chunk |
| `policy.n_action_steps` | `8` | Number of actions consumed from a sampled chunk |
| `policy.num_denoise_steps` | `10` | Number of flow-matching denoising steps used during sampling |
| `policy.max_state_dim` | `32` | State padding dimension |
| `policy.max_action_dim` | `32` | Action padding dimension |
| `policy.force_fp32_autocast` | `true` | Keeps the flow head in fp32 even when the backbone uses mixed precision |
| `policy.supervise_padding_action_dims` | `true` | Controls whether padded action dimensions are supervised |
| `policy.supervise_padding_actions` | `true` | Controls whether padded future action rows are supervised |
## Evaluation
EO-1 can be evaluated through `lerobot-eval` once you have a LeRobot-format checkpoint:
```bash
lerobot-eval \
--policy.path=your-org/your-eo1-checkpoint \
--env.type=libero \
--env.task=libero_object \
--eval.batch_size=1 \
--eval.n_episodes=20
```
For datasets or environments whose camera names differ from the checkpoint configuration, pass a `rename_map`:
```bash
lerobot-eval \
--policy.path=your-org/your-eo1-checkpoint \
--env.type=libero \
--env.task=libero_object \
--rename_map='{"observation.images.image2":"observation.images.wrist_image"}'
```
## Configuration Notes
### Image Processing
EO-1 uses the Qwen2.5-VL processor. The `policy.image_min_pixels` and `policy.image_max_pixels` settings control the image resizing bounds before the visual tokens are passed into the backbone.
### State and Action Dimensions
The policy pads state and action vectors to `policy.max_state_dim` and `policy.max_action_dim` before the EO-1 flow head. Predictions are cropped back to the original action dimension before being returned by the policy.
### Attention Backend
Use `policy.attn_implementation=sdpa` for a portable setup. Use `flash_attention_2` only when `flash_attn` is installed and compatible with your environment.
## References
- [EO-1 project](https://github.com/EO-Robotics/EO1)
- [EO-1 paper](https://arxiv.org/abs/2508.21112)
- [Qwen2.5-VL-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct)
## Citation
```bibtex
@article{eo1,
title={EO-1: Interleaved Vision-Text-Action Pretraining for General Robot Control},
author={Delin Qu and Haoming Song and Qizhi Chen and Zhaoqing Chen and Xianqiang Gao and Xinyi Ye and Qi Lv and Modi Shi and Guanghui Ren and Cheng Ruan and Maoqing Yao and Haoran Yang and Jiacheng Bao and Bin Zhao and Dong Wang},
journal={arXiv preprint},
year={2025},
url={https://arxiv.org/abs/2508.21112}
}
```
## License
This LeRobot integration follows the **Apache 2.0 License** used by LeRobot. Check the upstream EO-1 model and dataset pages for the licenses of released EO-1 checkpoints and data.
docs/source/evo1.mdx
+191
View File
@@ -0,0 +1,191 @@
# EVO1
EVO1 is a Vision-Language-Action policy for robot control built around an InternVL3 backbone and a continuous flow-matching action head. This LeRobot integration exposes EVO1 as a standard policy type so it can be trained and evaluated with the usual LeRobot dataset, checkpoint, and processor APIs.
## Model Overview
The policy embeds one or more camera images and the language task prompt with InternVL3, pads robot state/action vectors to fixed maximum dimensions, and predicts future action chunks with a flow-matching action head. During inference, the policy samples an action chunk and returns `n_action_steps` actions from that chunk before sampling again.
### What the LeRobot Integration Covers
- Standard `policy.type=evo1` configuration through LeRobot
- InternVL3 image/text embedding with optional FlashAttention fallback
- Stage-based finetuning controls for action-head-only and VLM finetuning runs
- Continuous flow-matching action prediction
- Checkpoint save/load through LeRobot policy APIs
- Training with `lerobot-train` and evaluation with standard policy inference APIs
The broader EVO1 project may include additional training scripts and dataset tooling. This page focuses on the LeRobot robot-control policy path.
## Installation Requirements
1. Install LeRobot by following the [Installation Guide](./installation).
2. Install EVO1 dependencies:
```bash
pip install -e ".[evo1]"
```
For LIBERO evaluation, install the LIBERO extra as well:
```bash
pip install -e ".[evo1,libero]"
```
3. Install a `flash-attn` wheel only if it is compatible with your Python, PyTorch, CUDA, and GPU stack. EVO1 falls back to standard attention when `flash_attn` is not available.
EVO1 uses the native Hugging Face `transformers` InternVL implementation, so `policy.vlm_model_name` must point to a natively converted checkpoint such as `OpenGVLab/InternVL3-1B-hf` (note the `-hf` suffix). The first run may download the configured VLM checkpoint unless `policy.vlm_model_name` points to a local model directory.
## Data Requirements
EVO1 expects a LeRobot dataset with:
- One to `policy.max_views` visual observations, for example `observation.images.image`
- `observation.state`
- `action`
- A language task instruction in the dataset `task` field, or another field configured with `policy.task_field`
State and action vectors are padded to `policy.max_state_dim` and `policy.max_action_dim`. Predictions are cropped back to the dataset action dimension before being returned.
## Usage
To use EVO1 in a LeRobot configuration, specify:
```python
policy.type=evo1
```
By default, a new EVO1 policy initializes its VLM from:
```python
policy.vlm_model_name=OpenGVLab/InternVL3-1B-hf
```
Once a LeRobot-format EVO1 checkpoint is available, load it with:
```python
policy.path=your-org/your-evo1-checkpoint
```
## Training
### Stage 1
Stage 1 freezes the VLM and trains the action head:
```bash
lerobot-train \
--dataset.repo_id=your_org/your_dataset \
--policy.type=evo1 \
--policy.training_stage=stage1 \
--policy.vlm_model_name=OpenGVLab/InternVL3-1B-hf \
--policy.device=cuda \
--policy.chunk_size=50 \
--policy.n_action_steps=50 \
--policy.max_state_dim=24 \
--policy.max_action_dim=24 \
--policy.optimizer_lr=1e-5 \
--batch_size=4 \
--steps=5000 \
--output_dir=./outputs/evo1_stage1
```
### Stage 2
Stage 2 finetunes the VLM branches and action head. A common workflow starts from a Stage 1 checkpoint:
```bash
lerobot-train \
--dataset.repo_id=your_org/your_dataset \
--policy.path=./outputs/evo1_stage1/checkpoints/005000/pretrained_model \
--policy.training_stage=stage2 \
--policy.vlm_model_name=OpenGVLab/InternVL3-1B-hf \
--policy.device=cuda \
--policy.chunk_size=50 \
--policy.n_action_steps=50 \
--policy.max_state_dim=24 \
--policy.max_action_dim=24 \
--policy.optimizer_lr=1e-5 \
--batch_size=4 \
--steps=80000 \
--output_dir=./outputs/evo1_stage2
```
By default, `policy.training_stage` reapplies the finetuning defaults for that stage. This is important when
starting Stage 2 from a Stage 1 checkpoint, because the Stage 1 checkpoint config stores the VLM finetuning
flags as disabled. These stage defaults take precedence over saved or manually supplied `policy.finetune_*`
flags unless `policy.apply_training_stage_defaults=false`, so set that flag only when manually controlling
every finetuning flag.
### Key Training Parameters
| Parameter | Default | Description |
| --------------------------------------------- | --------------------------- | ----------------------------------------------------------------- |
| `policy.vlm_model_name` | `OpenGVLab/InternVL3-1B-hf` | Natively converted InternVL3 checkpoint or local model directory |
| `policy.training_stage` | `stage1` | `stage1` trains the action head; `stage2` finetunes VLM branches |
| `policy.apply_training_stage_defaults` | `true` | Reapplies stage finetuning defaults after loading a checkpoint |
| `policy.vlm_num_layers` | `14` | Number of InternVL3 language layers kept for the policy |
| `policy.vlm_dtype` | `bfloat16` | Requested VLM dtype |
| `policy.use_flash_attn` | `true` | Requests FlashAttention when installed; otherwise falls back |
| `policy.enable_gradient_checkpointing` | `true` | Enables checkpointing on supported InternVL3 modules |
| `policy.gradient_checkpointing_use_reentrant` | `false` | Reentrant setting passed to gradient checkpointing when supported |
| `policy.chunk_size` | `50` | Number of future actions predicted per chunk |
| `policy.n_action_steps` | `50` | Number of actions consumed from a sampled chunk |
| `policy.max_state_dim` | `24` | State padding dimension |
| `policy.max_action_dim` | `24` | Action padding dimension |
| `policy.postprocess_action_dim` | `null` | Optional action dimension returned after EVO1 postprocessing |
| `policy.binarize_gripper` | `false` | Binarizes the postprocessed gripper channel for LIBERO-style eval |
| `policy.task_field` | `task` | Batch field used as the language prompt |
## Inference
Try it out with a trained EVO1 checkpoint:
```bash
lerobot-rollout \
--policy.path=your-org/your-evo1-checkpoint \
--inference.type=rtc \ # optional
...
```
## Results
### LIBERO Evaluation
> [!NOTE]
> Benchmark results for a `lerobot`-hosted LIBERO checkpoint trained with this implementation
> will be added once training completes.
The official EVO1 LIBERO rollout protocol uses the raw LIBERO camera feature names
(`observation.images.agentview_image` and `observation.images.robot0_eye_in_hand_image`), replans every
14 actions, and binarizes the gripper command before stepping the simulator. The EVO1 policy postprocessor
can crop the padded 24D action back to the 7D LIBERO action space and apply that gripper binarization. To
evaluate a LIBERO checkpoint under the same one-episode-per-task setting, keep the raw camera names instead
of the default `image`/`image2` mapping and set the LIBERO action postprocessing flags:
```bash
lerobot-eval \
--policy.path=your-org/your-evo1-libero-checkpoint \
--policy.vlm_model_name=OpenGVLab/InternVL3-1B-hf \
--policy.device=cuda \
--policy.use_flash_attn=true \
--policy.n_action_steps=14 \
--policy.postprocess_action_dim=7 \
--policy.binarize_gripper=true \
--env.type=libero \
--env.task=libero_object \
--env.camera_name_mapping="{agentview_image: agentview_image, robot0_eye_in_hand_image: robot0_eye_in_hand_image}" \
--env.observation_height=448 \
--env.observation_width=448 \
--eval.batch_size=1 \
--eval.n_episodes=1
```
## References
- [EVO1 repository](https://github.com/MINT-SJTU/Evo-1)
- [InternVL3-1B-hf](https://huggingface.co/OpenGVLab/InternVL3-1B-hf)
## License
This LeRobot integration follows the Apache 2.0 License used by LeRobot. Check the upstream EVO1 and InternVL3 model pages for the licenses of released checkpoints and data.
docs/source/fastwam.mdx
+167
View File
@@ -0,0 +1,167 @@
# FastWAM
FastWAM is a World Action Model policy for robot control. The LeRobot integration exposes FastWAM through the standard policy API so it can be configured with `policy.type=fastwam`, trained with `lerobot-train`, and loaded through the LeRobot pretrained policy interface.
## Model Overview
FastWAM keeps video modeling during training, but uses direct action prediction at inference time instead of iteratively generating future observations. This LeRobot policy wraps the FastWAM action model, adapts LeRobot batches to FastWAM training samples, and provides the standard processor pipeline for normalization and action postprocessing.
The implementation initializes the visual world-model components from `Wan-AI/Wan2.2-TI2V-5B` by default and predicts action chunks with shape `[batch, action_horizon, action_dim]`.
### What the LeRobot Integration Covers
- Standard `policy.type=fastwam` configuration through LeRobot
- Image, state, action, and language-task batch adaptation
- Action chunk inference through `select_action` and `predict_action_chunk`
- Checkpoint save/load through the LeRobot policy APIs
- Configurable LIBERO gripper action postprocessing
## Installation Requirements
Install LeRobot from source, then install FastWAM dependencies:
```bash
pip install -e ".[fastwam]"
```
This installs the FastWAM policy extra from `pyproject.toml`: `transformers`,
`diffusers`, `ftfy`, and `regex`, plus LeRobot's base dependencies.
For LIBERO evaluation, install the benchmark dependencies too:
```bash
pip install -e ".[fastwam,libero]"
```
This installs both extras. In addition to the FastWAM dependencies above, the
`libero` extra installs LeRobot dataset dependencies, `hf-libero` on Linux, and
`scipy`.
FastWAM uses the Wan2.2 TI2V backbone. The default model id is:
```python
policy.model_id=Wan-AI/Wan2.2-TI2V-5B
```
## Data Requirements
FastWAM expects a LeRobot dataset with:
- one or more visual observations whose widths concatenate to `policy.image_size[1]`
- `observation.state` when `policy.proprio_dim` is not `None`
- `action`
- a language task instruction through the dataset task field, or precomputed `context` and `context_mask` tensors
The default visual setup is one image feature named `observation.images.image` with shape `(3, 224, 448)`. If the dataset uses two cameras, configure `policy.input_features` so their heights match `224` and their widths sum to `448`.
## Usage
Create a new FastWAM policy with:
```bash
lerobot-train \
--dataset.repo_id=your-org/your-dataset \
--policy.type=fastwam \
--policy.action_dim=7 \
--policy.proprio_dim=8 \
--policy.action_horizon=32 \
--policy.n_action_steps=10 \
--policy.image_size='[224,448]' \
--output_dir=./outputs/fastwam_training \
--job_name=fastwam_training \
--steps=300000 \
--batch_size=8 \
--policy.device=cuda
```
Evaluate an existing LeRobot-format checkpoint on LIBERO-10 with:
```bash
lerobot-eval \
--policy.path=ZibinDong/fastwam_libero_uncond_2cam224 \
--policy.device=cuda \
--policy.torch_dtype=float32 \
--policy.n_action_steps=10 \
--env.type=libero \
--env.task=libero_10 \
--env.observation_height=224 \
--env.observation_width=224 \
--eval.batch_size=1 \
--eval.n_episodes=50 \
--seed=0 \
--env.episode_length=600
```
For `libero_goal`, `libero_spatial`, and `libero_object`, use
`--env.episode_length=300`.
For real-robot rollout, use the same checkpoint path:
```bash
lerobot-rollout \
--robot.type=so101_follower \
--robot.port=/dev/ttyACM0 \
--policy.path=your-org/fastwam-real-robot
```
## Configuration Notes
### Image Features
`policy.image_size` is the size of the concatenated FastWAM image tensor as `(height, width)`. Each configured image feature must have shape `(3, height, camera_width)`, and all camera widths must sum to the configured width.
### Action Chunking
`policy.action_horizon` controls the number of future actions supervised during training and predicted during inference. `policy.n_action_steps` controls how many actions are consumed before the policy predicts a fresh chunk. `policy.n_action_steps` must be less than or equal to `policy.action_horizon`.
### Wan Components
FastWAM loads the Wan VAE, video DiT, text encoder, and tokenizer from the configured Wan model directory or Hugging Face Hub model id. LeRobot-format FastWAM checkpoints saved by `save_pretrained` also copy the local Wan component files needed by `from_pretrained`.
### Attention Backend
FastWAM's DiT uses PyTorch's `scaled_dot_product_attention` (SDPA) for all attention. It does **not** use FlashAttention: its Mixture-of-Transformers (MoT) routing needs arbitrary boolean `[query, key]` attention masks, which the FlashAttention varlen API cannot express. Installing the `flash-attn` package therefore has no effect on the FastWAM path. (Note that SDPA itself may still select PyTorch's own flash / memory-efficient / math kernel internally — this is unrelated to the `flash-attn` package.)
### LIBERO Action Toggle
FastWAM LIBERO checkpoints use `policy.toggle_action_dimensions=[-1]` by
default to match the gripper action convention used by the original FastWAM
evaluation pipeline:
```bash
--policy.toggle_action_dimensions='[-1]'
```
## Results
Evaluated on LIBERO with [`ZibinDong/fastwam_libero_uncond_2cam224`](https://huggingface.co/ZibinDong/fastwam_libero_uncond_2cam224):
| Suite | Success rate | n_episodes |
| -------------- | -----------: | ---------: |
| libero_spatial | 97.6% | 500 |
| libero_object | 99.0% | 500 |
| libero_goal | 95.0% | 500 |
| libero_10 | 94.0% | 500 |
| **average** | **96.4%** | 2000 |
Reproduce: `lerobot-eval --policy.path=ZibinDong/fastwam_libero_uncond_2cam224 --policy.device=cuda --policy.torch_dtype=float32 --policy.n_action_steps=10 --env.type=libero --env.task=libero_spatial --env.observation_height=256 --env.observation_width=256 --eval.batch_size=1 --eval.n_episodes=50 --seed=0 --env.episode_length=300` (1x H20 140 GB).
## References
- [Fast-WAM paper](https://arxiv.org/abs/2603.16666)
- [Fast-WAM project page](https://yuantianyuan01.github.io/FastWAM/)
- [Fast-WAM code](https://github.com/yuantianyuan01/FastWAM)
- [Released upstream checkpoints](https://huggingface.co/yuanty/fastwam)
- [Wan2.2 TI2V 5B](https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B)
## Citation
```bibtex
@article{yuan2026fastwam,
title = {Fast-WAM: Do World Action Models Need Test-time Future Imagination?},
author = {Tianyuan Yuan and Zibin Dong and Yicheng Liu and Hang Zhao},
journal = {arXiv preprint arXiv:2603.16666},
year = {2026},
url = {https://arxiv.org/abs/2603.16666}
}
```
docs/source/feetech.mdx
+71
View File
@@ -0,0 +1,71 @@
# Feetech Motor Firmware Update
This tutorial guides you through updating the firmware of Feetech motors using the official Feetech software.
## Prerequisites
- Windows computer (Feetech software is only available for Windows)
- Feetech motor control board
- USB cable to connect the control board to your computer
- Feetech motors connected to the control board
## Step 1: Download Feetech Software
1. Visit the official Feetech software download page: [https://www.feetechrc.com/software.html](https://www.feetechrc.com/software.html)
2. Download the latest version of the Feetech debugging software (FD)
3. Install the software on your Windows computer
## Step 2: Hardware Setup
1. Connect your Feetech motors to the motor control board
2. Connect the motor control board to your Windows computer via USB cable
3. Ensure power is supplied to the motors
## Step 3: Configure Connection
1. Launch the Feetech debugging software
2. Select the correct COM port from the port dropdown menu
- If unsure which port to use, check Windows Device Manager under "Ports (COM & LPT)"
3. Set the appropriate baud rate (typically 1000000 for most Feetech motors)
4. Click "Open" to establish communication with the control board
## Step 4: Scan for Motors
1. Once connected, click the "Search" button to detect all connected motors
2. The software will automatically discover and list all motors on the bus
3. Each motor will appear with its ID number
## Step 5: Update Firmware
For each motor you want to update:
1. **Select the motor** from the list by clicking on it
2. **Click on Upgrade tab**:
3. **Click on Online button**:
- If an potential firmware update is found, it will be displayed in the box
4. **Click on Upgrade button**:
- The update progress will be displayed
## Step 6: Verify Update
1. After the update completes, the software should automatically refresh the motor information
2. Verify that the firmware version has been updated to the expected version
## Important Notes
⚠️ **Warning**: Do not disconnect power or USB during firmware updates, it will potentially brick the motor.
## Bonus: Motor Debugging on Linux/macOS
For debugging purposes only, you can use the open-source Feetech Debug Tool:
- **Repository**: [FT_SCServo_Debug_Qt](https://github.com/CarolinePascal/FT_SCServo_Debug_Qt/tree/fix/port-search-timer)
### Installation Instructions
Follow the instructions in the repository to install the tool, for Ubuntu you can directly install it, for MacOS you need to build it from source.
**Limitations:**
- This tool is for debugging and parameter adjustment only
- Firmware updates must still be done on Windows with official Feetech software
docs/source/groot.mdx
+227
View File
@@ -0,0 +1,227 @@
# GR00T Policy
GR00T is an NVIDIA foundation model family for generalized humanoid robot reasoning and skills. It is a cross-embodiment policy that accepts multimodal input, including language, images, and proprioception, to perform manipulation tasks in diverse environments.
LeRobot integrates GR00T N1.7 through the `groot` policy type.
> [!WARNING]
> **Breaking change:** GR00T N1.5 support was removed from LeRobot, and current releases support GR00T N1.7 only. N1.5 checkpoints and configs are rejected with a migration note. To keep using an N1.5 checkpoint, pin the last release that supports it: `pip install 'lerobot==0.5.1'`. To use the current release, migrate to GR00T N1.7 (base model [`nvidia/GR00T-N1.7-3B`](https://huggingface.co/nvidia/GR00T-N1.7-3B)).
## Model Overview
GR00T N1.7 uses a Cosmos-Reason2/Qwen3-VL backbone and provides checkpoints for SimplerEnv, DROID, and LIBERO.
Developers and researchers can post-train GR00T with their own real or synthetic data to adapt it for specific humanoid robots or tasks.
GR00T uses pre-trained vision and language encoders with a flow matching action transformer to model a chunk of actions conditioned on vision, language, and proprioception.
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/lerobot-groot-paper1%20(1).png"
alt="An overview of GR00T"
width="80%"
/>
Its strong performance comes from being trained on an expansive and diverse humanoid dataset, which includes:
- Real captured data from robots.
- Synthetic data generated using NVIDIA Isaac GR00T Blueprint.
- Internet-scale video data.
This approach allows the model to be highly adaptable through post-training for specific embodiments, tasks, and environments.
## Installation Requirements
GR00T is intended for NVIDIA GPU-accelerated systems. Install LeRobot with the GR00T extra:
```bash
pip install "lerobot[groot]"
```
For a source checkout:
```bash
pip install -e ".[groot]"
```
## Usage
To use GR00T N1.7:
```bash
--policy.type=groot
```
## Training
### Training Command Example
Here's a complete training command for finetuning the base GR00T model on your own dataset:
This command is using the `new_embodiment` flag, which is used for the SO-101 robot, [read more about how GR00T handles different embodiments.](https://github.com/NVIDIA/Isaac-GR00T/blob/main/getting_started/policy.md#--embodiment-tag).
```bash
# install extra deps for training
pip install "lerobot[training]"
hf auth login
wandb login
export DATASET_NAME=your_data_set
export HF_USER=your_hf_username
export DATASET=$HF_USER/$DATASET_NAME
export REPO_ID="${DATASET}_GR00T17" #this is the model that will be uploaded to huggingface
export OUTPUT_DIR=outputs/train/$REPO_ID
lerobot-train \
--dataset.repo_id=$DATASET \
--dataset.image_transforms.enable=true \
--policy.type=groot \
--policy.device=cuda \
--policy.base_model_path=nvidia/GR00T-N1.7-3B \
--policy.embodiment_tag=new_embodiment \
--policy.chunk_size=16 \
--policy.n_action_steps=16 \
--policy.use_relative_actions=true \
--policy.relative_exclude_joints='["gripper"]' \
--policy.use_bf16=true \
--policy.push_to_hub=true \
--policy.repo_id=$REPO_ID \
--seed=42 \
--batch_size=64 \
--steps=20000 \
--save_checkpoint=true \
--save_freq=5000 \
--use_policy_training_preset=true \
--env_eval_freq=0 \
--eval_steps=0 \
--log_freq=10 \
--output_dir=$OUTPUT_DIR \
--job_name=$DATASET \
--wandb.enable=true \
--wandb.disable_artifact=true
```
## Performance Results
### LIBERO Benchmark Results
> [!NOTE]
> Follow the [LIBERO](./libero) setup instructions before running `lerobot-eval`.
GR00T N1.7 has demonstrated strong performance on the LIBERO benchmark suite. To reproduce LeRobot results, follow the instructions in the [LIBERO](./libero) section.
### Train on LIBERO
Example training command for a LIBERO suite (here `libero_spatial`):
```bash
IMAGE_TRANSFORMS='{
"brightness": {"weight": 1.0, "type": "ColorJitter", "kwargs": {"brightness": [0.7, 1.3]}},
"contrast": {"weight": 1.0, "type": "ColorJitter", "kwargs": {"contrast": [0.6, 1.4]}},
"saturation": {"weight": 1.0, "type": "ColorJitter", "kwargs": {"saturation": [0.5, 1.5]}},
"hue": {"weight": 1.0, "type": "ColorJitter", "kwargs": {"hue": [-0.08, 0.08]}}
}'
lerobot-train \
--dataset.repo_id=IPEC-COMMUNITY/libero_spatial_no_noops_1.0.0_lerobot \
--dataset.root=/datasets/libero_spatial \
--dataset.revision=main \
--dataset.video_backend=pyav \
--dataset.image_transforms.enable=true \
--dataset.image_transforms.max_num_transforms=4 \
--dataset.image_transforms.tfs="$IMAGE_TRANSFORMS" \
--policy.type=groot \
--policy.base_model_path=nvidia/GR00T-N1.7-3B \
--policy.embodiment_tag=libero_sim \
--policy.push_to_hub=false \
--policy.use_relative_actions=false \
--policy.max_steps=20000 \
--batch_size=320 \
--steps=20000 \
--save_freq=2000 \
--env_eval_freq=0 \
--eval_steps=0 \
--log_freq=10 \
--wandb.enable=true \
--wandb.project=lerobot \
--wandb.mode=online \
--wandb.disable_artifact=true \
--num_workers=4 \
--prefetch_factor=2 \
--persistent_workers=true \
--output_dir=$OUTPUT_DIR \
--job_name=$JOB_NAME
```
This will follow the recipe found [here](https://github.com/NVIDIA/Isaac-GR00T/blob/main/examples/LIBERO/README.md).
### GR00T N1.7 LIBERO Results
Preliminary LeRobot integration results (GR00T-LeRobot, `eval.n_episodes >= 50` per suite):
| Suite | Success rate | Checkpoint |
| ---------------- | -----------: | ------------------------------------------------------------------------------------------------------------- |
| LIBERO Spatial | 95% | [nvidia/gr00t17-lerobot-libero_spatial-640](https://huggingface.co/nvidia/gr00t17-lerobot-libero_spatial-640) |
| LIBERO Object | 100% | [nvidia/gr00t17-lerobot-libero_object-640](https://huggingface.co/nvidia/gr00t17-lerobot-libero_object-640) |
| LIBERO Goal | 98% | [nvidia/gr00t17-lerobot-libero_goal-640](https://huggingface.co/nvidia/gr00t17-lerobot-libero_goal-640) |
| LIBERO 10 (Long) | 93% | [nvidia/gr00t17-lerobot-libero_10-640](https://huggingface.co/nvidia/gr00t17-lerobot-libero_10-640) |
| **Average** | **96.5%** | |
```bash
export MODEL_ID=your_trained_model_on_huggingface
lerobot-eval \
--policy.type=groot \
--policy.base_model_path=$MODEL_ID \
--policy.embodiment_tag=libero_sim \
--env.type=libero \
--env.task=libero_spatial \
--eval.n_episodes=50
```
Use `eval.n_episodes >= 50` per suite when reporting success rates.
### Evaluate in your hardware setup
Once you have trained your model using your parameters you can run inference in your downstream task. Follow the instructions in [Policy Deployment (lerobot-rollout)](./inference). For example:
```bash
# install extra deps for roullout and real hardware
pip install "lerobot[feetech,viz]"
export MODEL_ID=your_trained_model_on_huggingface
# make sure that camera index matches your setup!
# find index using `uv run lerobot-find-cameras opencv`
WRIST_CAM='wrist: {type: opencv, index_or_path: 2, width: 640, height: 480, fps: 30, fourcc: "MJPG"}'
FRONT_CAM='front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30, fourcc: "MJPG"}'
export ROBOT_CAMERAS="{ $WRIST_CAM, $FRONT_CAM }"
export ROBOT_ID=follower_robot
export ROBOT_PORT=/dev/ttyACM0
uv run lerobot-rollout \
--strategy.type=base \
--policy.path=$MODEL_ID \
--policy.base_model_path=nvidia/GR00T-N1.7-3B \
--policy.n_action_steps=8 \
--robot.type=so101_follower \
--robot.port=$ROBOT_PORT \
--robot.id=$ROBOT_ID \
--robot.cameras="$ROBOT_CAMERAS" \
--task="place the vial in the rack" \
--duration=60 \
--device=cuda \
--display_data=true \
--inference.type=rtc \
--inference.rtc.enabled=True \ # set to False if it causes inference instability
--inference.rtc.execution_horizon=8 \
--inference.queue_threshold=0
```
> [!NOTE]
> Value of `inference.queue_threshold` should not exceed 5 to ensure stable inference.
## License
GR00T N1.7 is released under the [NVIDIA Open Model License Agreement](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/).
docs/source/hardware_guide.mdx
+99
View File
@@ -0,0 +1,99 @@
# Compute HW Guide for LeRobot Training
Rough sizing for training a LeRobot policy: how much VRAM each policy needs, what training time looks like, and where to run when local hardware isn't enough.
The numbers below are **indicative** — order-of-magnitude figures for picking hardware, not exact predictions. Throughput depends heavily on dataset I/O, image resolution, batch size, and number of GPUs.
## Memory by policy group
Policies cluster by backbone size; the groupings below give a single VRAM envelope per group instead of repeating numbers per policy. Memory scales roughly linearly with batch size; AdamW (the LeRobot default) carries optimizer state that adds ~30–100% over a forward+backward pass alone.
| Group | Policies | Peak VRAM (BS 8, AdamW) | Suitable starter GPUs |
| ---------- | ------------------------------------------- | ----------------------: | --------------------------------- |
| Light BC | `act`, `vqbet`, `tdmpc` | ~2–6GB | Laptop GPU (RTX 3060), L4, A10G |
| Diffusion | `diffusion`, `multi_task_dit` | ~8–14GB | RTX 4070+ / L4 / A10G |
| Small VLA | `smolvla` | ~10–16GB | RTX 4080+ / L4 / A10G |
| Large VLA | `pi0`, `pi0_fast`, `pi05`, `xvla`, `wall_x` | ~24–40GB | A100 40 GB+ (24 GB tight at BS 1) |
| Multimodal | `groot`, `eo1` | ~24–40GB | A100 40 GB+ |
| RL | `sac` | config-dep. | See [HIL-SERL guide](./hilserl) |
Memory-bound? Drop the batch size (~linear), use gradient accumulation to recover effective batch, or for SmolVLA leave `freeze_vision_encoder=True`.
## Training time
Robotics imitation learning typically converges in **5–10 epochs over the dataset**, not hundreds of thousands of raw steps. Once you know your epoch count, wall-clock is essentially:
```text
total_frames = sum of frames over all episodes # 50 ep × 30 fps × 30 s ≈ 45,000
steps_per_epoch = ceil(total_frames / (num_gpus × batch_size))
total_steps = epochs × steps_per_epoch
wall_clock ≈ total_steps × per_step_time
```
Per-step time depends on the policy and the GPU. The numbers in the table below are anchors — pick the row closest to your setup and scale linearly with `total_steps` if you train longer or shorter.
### Common scenarios
Indicative wall-clock for **5 epochs on a ~50-episode dataset (~45k frames at 30 fps × 30 s)**, default optimizer (AdamW), 640×480 images:
| Setup | Policy | Batch | Wall-clock |
| ------------------------------------ | -------------- | ----- | ---------: |
| Single RTX 4090 / RTX 3090 (24 GB) | `act` | 8 | ~30–60min |
| Single RTX 4090 / RTX 3090 (24 GB) | `diffusion` | 8 | ~2–4h |
| Single L4 / A10G (24 GB) | `act` | 8 | ~1–2h |
| Single L4 / A10G (24 GB) | `smolvla` | 4 | ~3–6h |
| Single A100 40 GB | `smolvla` | 16 | ~1–2h |
| Single A100 40 GB | `pi0` / `pi05` | 4 | ~4–8h |
| 4× H100 80 GB cluster (`accelerate`) | `diffusion` | 32 | ~30–60min |
| 4× H100 80 GB cluster (`accelerate`) | `smolvla` | 32 | ~1–2h |
| Apple Silicon M1/M2/M3 Max (MPS) | `act` | 4 | ~6–14h |
These are order-of-magnitude figures. Real runs deviate by ±50% depending on image resolution, dataset I/O, dataloader threading, and exact GPU SKU. They are useful as "is this run going to take an hour or a day?" intuition, not as SLAs.
### Multi-GPU matters a lot
`accelerate launch --num_processes=N` is the easiest way to cut training time. Each optimizer step processes `N × batch_size` samples in roughly the same wall-clock as a single-GPU step, so 4 GPUs ≈ 4× speedup for compute-bound runs. See the [Multi GPU training](./multi_gpu_training) guide for the full setup.
Reference data points on a 4×H100 80 GB cluster (`accelerate launch --num_processes=4`), 5000 steps, batch 32, AdamW, dataset [`imstevenpmwork/super_poulain_draft`](https://huggingface.co/datasets/imstevenpmwork/super_poulain_draft) (~50 episodes, ~640×480 images):
| Policy | Wall-clock | `update_s` | `dataloading_s` | GPU util | Notable flags |
| ----------- | ---------- | ---------: | --------------: | -------- | ------------------------------------------------------------------------------------------------------------------------------ |
| `diffusion` | 16m 17s | 0.167 | 0.015 | ~90% | defaults (training from scratch) |
| `smolvla` | 27m 49s | 0.312 | 0.011 | ~80% | `--policy.path=lerobot/smolvla_base`, `freeze_vision_encoder=false`, `train_expert_only=false` |
| `pi05` | 3h 41m | 2.548 | 0.014 | ~95% | `--policy.pretrained_path=lerobot/pi05_base`, `gradient_checkpointing=true`, `dtype=bfloat16`, vision encoder + expert trained |
The `dataloading_s` vs. `update_s` ratio is the diagnostic that matters: when `dataloading_s` approaches `update_s`, more GPUs stop helping — your dataloader is the bottleneck and you should look at `--num_workers`, image resolution, and disk speed before adding compute.
### Schedule and checkpoints
If you shorten training (e.g. 5k–10k steps on a small dataset), also shorten the LR schedule with `--policy.scheduler_decay_steps≈--steps`. Otherwise the LR stays near its peak and never decays. Same for `--save_freq`.
## Where to run
VRAM is the first filter. Within a tier, pick by budget and availability — the `$`–`$$$$` columns are relative; check current pricing on the provider you actually use.
| Class | VRAM | Tier | Comfortable for |
| -------------------------- | ----- | ------ | ----------------------------------------------------------- |
| RTX 3090 / 4090 (consumer) | 24 GB | `$` | Light BC, Diffusion, SmolVLA. Tight for VLAs at batch 1. |
| L4 / A10G (cloud) | 24 GB | `$–$$` | Same envelope; common on Google Cloud, RunPod, AWS `g5/g6`. |
| A100 40 GB | 40 GB | `$$$` | Any policy at reasonable batch sizes. |
| A100 80 GB / H100 80 GB | 80 GB | `$$$$` | Multi-GPU clusters; large batches for VLAs. |
| **CPU only** | — | — | Don't train. Use Colab or rent a GPU. |
### Hugging Face Jobs
[Hugging Face Jobs](https://huggingface.co/docs/hub/jobs) lets you run training on managed HF infrastructure, billed by the second, without owning a GPU. `lerobot-train` submits and streams the job for you — just add `--job.target=<flavor>` to a normal training command:
```bash
lerobot-train \
--policy.type=act --dataset.repo_id=<USER>/<DATASET> \
--policy.repo_id=<USER>/act_<task> \
--job.target=a10g-large
```
Notes:
- Run `hf auth login` once before submitting, the job runs under your token.
- `--job.target` maps onto the table above: `t4-small`/`t4-medium` (T4, ACT only), `l4x1`/`l4x4` (L4 24 GB), `a10g-small/large/largex2/largex4` (A10G 24 GB scaled out), `a100-large` (A100). List the current catalogue with pricing via `hf jobs hardware`, or see [https://huggingface.co/docs/hub/jobs](https://huggingface.co/docs/hub/jobs).
- The job defaults to a `2d` (48h) timeout. Override it with `--job.timeout=4h` (or any other valid duration string) to shorten or extend the timeout. The job automatically stops when the command completes.
- For the full walkthrough — dataset upload, checkpoint streaming, resuming a run on a job — see the [imitation-learning training guide](./il_robots#train-using-hugging-face-jobs).
docs/source/hil_data_collection.mdx
+267
View File
@@ -0,0 +1,267 @@
# Human-In-the-Loop Data Collection
Human-In-the-Loop (HIL) data collection lets you improve a trained policy by deploying it on a real robot while a human operator monitors and intervenes when needed. The intervention data (recovery movements and corrections) is recorded alongside autonomous segments, producing a richer training dataset that teaches the policy how to handle failures.
---
## Why Human-In-the-Loop?
Standard behavioral cloning trains policies on successful demonstrations only. During deployment, small errors can compound and push the robot into states never seen during training (distribution shift). HIL data collection addresses this by:
- Running the trained policy on the real robot
- Having a human intervene when the robot is about to fail
- Recording the human's recovery and correction as training data
- Fine-tuning the policy on the combined dataset
This produces a policy that not only knows how to perform the task, but also how to recover when things go wrong.
---
## How It Works
During a HIL session, the human operator follows this loop within each episode:
1. **Watch** the policy run autonomously
2. **Pause** when failure is imminent, the robot holds its position
3. **Take control** and teleoperate the robot back to a good state (recovery), then correct the behavior
4. **Return control to the policy**, the policy resumes autonomous execution
5. Repeat steps 2–4 as many times as needed during the episode
6. **End the episode** when the task is complete, save and move on to the next rollout
Both autonomous and human-controlled segments are recorded. The policy and human can alternate control multiple times within a single episode, and the episode continues from the current state after each handoff (no reset required just because intervention happened). This captures autonomous execution, recovery, and correction in one continuous trajectory. After collection, the combined dataset (original demonstrations + HIL data) is used to fine-tune the policy.
This process can be repeated iteratively: deploy, collect, fine-tune, repeat. Each round targets the current policy's failure modes.
```
┌─────────────────────────────────────────────────────────────────────────┐
│ Policy v0 (trained on demos) │
│ ↓ │
│ HIL Collection (target current failure modes) → Fine-tune → Policy v1 │
│ ↓ │
│ HIL Collection (target new failure modes) → Fine-tune → Policy v2 │
│ ↓ │
│ ... (repeat until satisfactory performance) │
└─────────────────────────────────────────────────────────────────────────┘
```
---
## Hardware Requirements
### Teleoperator Requirements
The `lerobot-rollout --strategy.type=dagger` mode requires **teleoperators with active motors** that can:
- Enable/disable torque programmatically
- Move to target positions (to mirror the robot state when pausing)
**Compatible teleoperators:**
- `bi_openarm_mini` - Bimanual OpenArm Mini
- `so_leader` - SO100 / SO101 leader arm
> [!IMPORTANT]
> The provided commands default to `bi_openarm_follower` + `bi_openarm_mini`.
> `so_follower` + `so_leader` configs are also registered and can be used via CLI flags.
---
## Script
Use `lerobot-rollout` with `--strategy.type=dagger` for HIL data collection. Select the inference backend with `--inference.type=sync|rtc`:
| Mode | Flag | Models |
| ------------------------ | ---------------------- | --------------------- |
| Standard (default) | _(no flag needed)_ | ACT, Diffusion Policy |
| Real-Time Chunking (RTC) | `--inference.type=rtc` | Pi0, Pi0.5, SmolVLA |
---
## Step-by-Step Guide
### Step 1: Pre-train a Base Policy
First, train a policy on your demonstration dataset:
```bash
python src/lerobot/scripts/lerobot_train.py \
--dataset.repo_id=your-username/demo-dataset \
--policy.type=pi0 \
--output_dir=outputs/pretrain \
--batch_size=32 \
--steps=50000
```
### Step 2: Collect HIL Data
**Standard inference (ACT, Diffusion Policy):**
```bash
lerobot-rollout --strategy.type=dagger \
--robot.type=bi_openarm_follower \
--robot.left_arm_config.port=can1 \
--robot.left_arm_config.side=left \
--robot.right_arm_config.port=can0 \
--robot.right_arm_config.side=right \
--robot.cameras='{left_wrist: {type: opencv, index_or_path: "/dev/video0", width: 1280, height: 720, fps: 30}, right_wrist: {type: opencv, index_or_path: "/dev/video4", width: 1280, height: 720, fps: 30}, base: {type: opencv, index_or_path: "/dev/video2", width: 640, height: 480, fps: 30}}' \
--teleop.type=bi_openarm_mini \
--teleop.left_arm_config.port=/dev/ttyACM0 \
--teleop.right_arm_config.port=/dev/ttyACM1 \
--policy.path=outputs/pretrain/checkpoints/last/pretrained_model \
--dataset.repo_id=your-username/rollout_hil_dataset \
--dataset.single_task="Fold the T-shirt properly" \
--dataset.fps=30 \
--strategy.num_episodes=50 \
--interpolation_multiplier=2
```
**With RTC for large models (Pi0, Pi0.5, SmolVLA):**
For models with high inference latency, enable RTC for smooth execution:
```bash
lerobot-rollout --strategy.type=dagger \
--inference.type=rtc \
--inference.rtc.execution_horizon=20 \
--inference.rtc.max_guidance_weight=5.0 \
--inference.rtc.prefix_attention_schedule=LINEAR \
--robot.type=bi_openarm_follower \
--robot.left_arm_config.port=can1 \
--robot.left_arm_config.side=left \
--robot.right_arm_config.port=can0 \
--robot.right_arm_config.side=right \
--robot.cameras='{left_wrist: {type: opencv, index_or_path: "/dev/video0", width: 1280, height: 720, fps: 30}, right_wrist: {type: opencv, index_or_path: "/dev/video4", width: 1280, height: 720, fps: 30}, base: {type: opencv, index_or_path: "/dev/video2", width: 640, height: 480, fps: 30}}' \
--teleop.type=bi_openarm_mini \
--teleop.left_arm_config.port=/dev/ttyACM0 \
--teleop.right_arm_config.port=/dev/ttyACM1 \
--policy.path=outputs/pretrain/checkpoints/last/pretrained_model \
--dataset.repo_id=your-username/rollout_hil_rtc_dataset \
--dataset.single_task="Fold the T-shirt properly" \
--dataset.fps=30 \
--strategy.num_episodes=50 \
--interpolation_multiplier=3
```
**Controls (Conceptual):**
The interaction model is:
- **Pause input**: pause autonomous policy execution
- **Takeover input**: transfer control to the human operator and record intervention data
- **Return-to-policy input**: hand control back to the policy and continue the same episode
- **Episode control inputs**: save/re-record/stop/reset as needed
Exact key/pedal bindings can differ across scripts and hardware integrations. Use each script's printed controls as the source of truth for the concrete mapping on your setup.
**The HIL Protocol:**
1. Watch the policy run autonomously (teleop is idle/free)
2. When you see imminent failure, trigger the **pause input**
- Policy stops
- Teleoperator moves to match robot position (torque enabled)
- No frames recorded during pause
3. Trigger the **takeover input** to take control
- Teleoperator torque disabled, free to move
- **Recovery**: Teleoperate the robot back to a good state
- **Correction**: Correct the behavior
- All movements are recorded
4. Trigger the **return-to-policy input**
- Policy resumes autonomous execution from the current state
- You can intervene again at any time (repeat steps 2–4)
5. End and save the episode when the task is complete (or episode time limit is reached)
6. **Reset**: Teleop moves to robot position, you can move the robot to the starting position
7. Start the next episode
**Foot Pedal Setup (Linux):**
If using a USB foot pedal (PCsensor FootSwitch), ensure access:
```bash
sudo setfacl -m u:$USER:rw /dev/input/by-id/usb-PCsensor_FootSwitch-event-kbd
```
### Step 3: Fine-tune the Policy
Fine-tune on the **combined** dataset (`demo-dataset` + `hil-dataset` merged together):
```bash
python src/lerobot/scripts/lerobot_train.py \
--dataset.repo_id=your-username/hil-dataset \
--policy.type=pi0 \
--policy.pretrained_path=outputs/pretrain/checkpoints/last/pretrained_model \
--output_dir=outputs/hil_finetune \
--steps=20000
```
Then deploy the fine-tuned policy and repeat from Step 2 to target its remaining failure modes.
---
## Tips for Effective HIL Collection
### When to Intervene
Intervene when you see:
- Robot about to make an irreversible mistake
- Robot hesitating or showing uncertain behavior
- Robot deviating from the expected trajectory
### Recovery: Teleoperating Back to a Good State
During recovery, teleoperate the robot back to a state where:
- The robot is in a familiar, in-distribution configuration
- The current subtask can still be completed
- The recovery trajectory itself is informative training data
### Quality of Corrections
During correction:
- Provide **confident, clean** trajectories
- Complete the current subtask fully
- Don't overcorrect or add unnecessary movements
---
## Related Work
This HIL data collection approach builds on ideas from interactive imitation learning:
- **DAgger** (Ross et al., 2011) introduced the core idea: instead of only training on expert demonstrations, query the expert for corrections on states the _learner_ visits. This breaks the compounding-error cycle of standard behavioral cloning by iteratively collecting on-policy data.
- **HG-DAgger** (Kelly et al., 2019) made this practical for robotics: a human expert monitors the robot and only intervenes when needed, rather than labeling every state. The gating between autonomous and human control is exactly the pause → takeover → return-to-policy loop used in the scripts here.
- **RaC** (Hu et al., 2025) scales this loop to long-horizon tasks by explicitly decomposing interventions into **recovery** (teleoperating back to a good state) and **correction** (demonstrating the right behavior from there). This decomposition is the protocol followed by the DAgger strategy in `lerobot-rollout`.
- **π0.6/RECAP** (Physical Intelligence, 2025) applies the same iterative collect-and-finetune loop at scale with VLA models, showing that even large pretrained policies benefit substantially from targeted human corrections on their own failure modes. π0.6 is trained using RECAP.
```bibtex
@article{ross2011dagger,
title={A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning},
author={Ross, Stéphane and Gordon, Geoffrey and Bagnell, Drew},
journal={Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics},
year={2011}
}
@article{kelly2019hgdagger,
title={HG-DAgger: Interactive Imitation Learning with Human Experts},
author={Kelly, Michael and Sidrane, Chelsea and Driggs-Campbell, Katherine and Kochenderfer, Mykel J},
journal={arXiv preprint arXiv:1810.02890},
year={2019}
}
@article{hu2025rac,
title={RaC: Robot Learning for Long-Horizon Tasks by Scaling Recovery and Correction},
author={Hu, Zheyuan and Wu, Robyn and Enock, Naveen and Li, Jasmine and Kadakia, Riya and Erickson, Zackory and Kumar, Aviral},
journal={arXiv preprint arXiv:2509.07953},
year={2025}
}
@article{pi2025recap,
title={π0.6: a VLA That Learns From Experience},
author={Physical Intelligence},
year={2025}
}
```
docs/source/hilserl.mdx
+950
View File
@@ -0,0 +1,950 @@
# HIL-SERL Real Robot Training Workflow Guide
In this tutorial you will go through the full Human-in-the-Loop Sample-Efficient Reinforcement Learning (HIL-SERL) workflow using LeRobot. You will master training a policy with RL on a real robot in just a few hours.
HIL-SERL is a sample-efficient reinforcement learning algorithm that combines human demonstrations with online learning and human interventions. The approach starts from a small set of human demonstrations, uses them to train a reward classifier, and then employs an actor-learner architecture where humans can intervene during policy execution to guide exploration and correct unsafe behaviors. In this tutorial, you'll use a gamepad to provide interventions and control the robot during the learning process.
It combines three key ingredients:
1. **Offline demonstrations & reward classifier:** a handful of human-teleop episodes plus a vision-based success detector give the policy a shaped starting point.
2. **On-robot actor / learner loop with human interventions:** a distributed Soft Actor Critic (SAC) learner updates the policy while an actor explores on the physical robot; the human can jump in at any time to correct dangerous or unproductive behaviour.
3. **Safety & efficiency tools:** joint/end-effector (EE) bounds, crop region of interest (ROI) preprocessing and WandB monitoring keep the data useful and the hardware safe.
Together these elements let HIL-SERL reach near-perfect task success and faster cycle times than imitation-only baselines.
<p align="center">
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/hilserl-main-figure.png"
alt="HIL-SERL workflow"
title="HIL-SERL workflow"
width="100%"
></img>
</p>
<p align="center">
<i>HIL-SERL workflow, Luo et al. 2024</i>
</p>
This guide provides step-by-step instructions for training a robot policy using LeRobot's HilSerl implementation to train on a real robot.
## What do I need?
- A gamepad (recommended) or keyboard to control the robot
- A Nvidia GPU
- A real robot with a follower and leader arm (optional if you use the keyboard or the gamepad)
- A URDF file for the robot for the kinematics package (check `lerobot/model/kinematics.py`)
## What kind of tasks can I train?
One can use HIL-SERL to train on a variety of manipulation tasks. Some recommendations:
- Start with a simple task to understand how the system works.
- Push cube to a goal region
- Pick and lift cube with the gripper
- Avoid extremely long horizon tasks. Focus on tasks that can be completed in 5-10 seconds.
- Once you have a good idea of how the system works, you can try more complex tasks and longer horizons.
- Pick and place cube
- Bimanual tasks to pick objects with two arms
- Hand-over tasks to transfer objects from one arm to another
- Go crazy!
## Install LeRobot with HIL-SERL
To install LeRobot with HIL-SERL, you need to install the `hilserl` extra.
```bash
pip install -e ".[hilserl]"
```
## Real Robot Training Workflow
### Understanding Configuration
The training process begins with proper configuration for the HILSERl environment. The main configuration class is `GymManipulatorConfig` in `lerobot/rl/gym_manipulator.py`, which contains nested `HILSerlRobotEnvConfig` (defined in `lerobot/envs/configs.py`) and `DatasetConfig`. The configuration is organized into focused, nested sub-configs:
<!-- prettier-ignore-start -->
```python
class GymManipulatorConfig:
env: HILSerlRobotEnvConfig # Environment configuration (nested)
dataset: DatasetConfig # Dataset recording/replay configuration (nested)
mode: str | None = None # "record", "replay", or None (for training)
device: str = "cpu" # Compute device
class HILSerlRobotEnvConfig(EnvConfig):
robot: RobotConfig | None = None # Main robot agent (defined in `lerobot/robots`)
teleop: TeleoperatorConfig | None = None # Teleoperator agent, e.g., gamepad or leader arm
processor: HILSerlProcessorConfig # Processing pipeline configuration (nested)
name: str = "real_robot" # Environment name
task: str | None = None # Task identifier
fps: int = 10 # Control frequency
# Nested processor configuration
class HILSerlProcessorConfig:
control_mode: str = "gamepad" # Control mode
observation: ObservationConfig | None = None # Observation processing settings
image_preprocessing: ImagePreprocessingConfig | None = None # Image crop/resize settings
gripper: GripperConfig | None = None # Gripper control and penalty settings
reset: ResetConfig | None = None # Environment reset and timing settings
inverse_kinematics: InverseKinematicsConfig | None = None # IK processing settings
reward_classifier: RewardClassifierConfig | None = None # Reward classifier settings
max_gripper_pos: float | None = 100.0 # Maximum gripper position
# Sub-configuration classes
class ObservationConfig:
add_joint_velocity_to_observation: bool = False # Add joint velocities to state
add_current_to_observation: bool = False # Add motor currents to state
add_ee_pose_to_observation: bool = False # Add end-effector pose to state
display_cameras: bool = False # Display camera feeds during execution
class ImagePreprocessingConfig:
crop_params_dict: dict[str, tuple[int, int, int, int]] | None = None # Image cropping parameters
resize_size: tuple[int, int] | None = None # Target image size
class GripperConfig:
use_gripper: bool = True # Enable gripper control
gripper_penalty: float = 0.0 # Penalty for inappropriate gripper usage
class ResetConfig:
fixed_reset_joint_positions: Any | None = None # Joint positions for reset
reset_time_s: float = 5.0 # Time to wait during reset
control_time_s: float = 20.0 # Maximum episode duration
terminate_on_success: bool = True # Whether to terminate episodes on success detection
class InverseKinematicsConfig:
urdf_path: str | None = None # Path to robot URDF file
target_frame_name: str | None = None # End-effector frame name
end_effector_bounds: dict[str, list[float]] | None = None # EE workspace bounds
end_effector_step_sizes: dict[str, float] | None = None # EE step sizes per axis
class RewardClassifierConfig:
pretrained_path: str | None = None # Path to pretrained reward classifier
success_threshold: float = 0.5 # Success detection threshold
success_reward: float = 1.0 # Reward value for successful episodes
# Dataset configuration
class DatasetConfig:
repo_id: str # LeRobot dataset repository ID
task: str # Task identifier
root: str | None = None # Local dataset root directory
num_episodes_to_record: int = 5 # Number of episodes for recording
replay_episode: int | None = None # Episode index for replay
push_to_hub: bool = False # Whether to push datasets to Hub
```
<!-- prettier-ignore-end -->
### Processor Pipeline Architecture
HIL-SERL uses a modular processor pipeline architecture that processes robot observations and actions through a series of composable steps. The pipeline is divided into two main components:
#### Environment Processor Pipeline
The environment processor (`env_processor`) handles incoming observations and environment state:
1. **VanillaObservationProcessorStep**: Converts raw robot observations into standardized format
2. **JointVelocityProcessorStep** (optional): Adds joint velocity information to observations
3. **MotorCurrentProcessorStep** (optional): Adds motor current readings to observations
4. **ForwardKinematicsJointsToEE** (optional): Computes end-effector pose from joint positions
5. **ImageCropResizeProcessorStep** (optional): Crops and resizes camera images
6. **TimeLimitProcessorStep** (optional): Enforces episode time limits
7. **GripperPenaltyProcessorStep** (optional): Applies penalties for inappropriate gripper usage
8. **RewardClassifierProcessorStep** (optional): Automated reward detection using vision models
9. **AddBatchDimensionProcessorStep**: Converts data to batch format for neural network processing
10. **DeviceProcessorStep**: Moves data to the specified compute device (CPU/GPU)
#### Action Processor Pipeline
The action processor (`action_processor`) handles outgoing actions and human interventions:
1. **AddTeleopActionAsComplimentaryDataStep**: Captures teleoperator actions for logging
2. **AddTeleopEventsAsInfoStep**: Records intervention events and episode control signals
3. **InterventionActionProcessorStep**: Handles human interventions and episode termination
4. **Inverse Kinematics Pipeline** (when enabled):
- **MapDeltaActionToRobotActionStep**: Converts delta actions to robot action format
- **EEReferenceAndDelta**: Computes end-effector reference and delta movements
- **EEBoundsAndSafety**: Enforces workspace safety bounds
- **InverseKinematicsEEToJoints**: Converts end-effector actions to joint targets
- **GripperVelocityToJoint**: Handles gripper control commands
#### Configuration Examples
**Basic Observation Processing**:
```json
{
"env": {
"processor": {
"observation": {
"add_joint_velocity_to_observation": true,
"add_current_to_observation": false,
"display_cameras": false
}
}
}
}
```
**Image Processing**:
```json
{
"env": {
"processor": {
"image_preprocessing": {
"crop_params_dict": {
"observation.images.front": [180, 250, 120, 150],
"observation.images.side": [180, 207, 180, 200]
},
"resize_size": [128, 128]
}
}
}
}
```
**Inverse Kinematics Setup**:
```json
{
"env": {
"processor": {
"inverse_kinematics": {
"urdf_path": "path/to/robot.urdf",
"target_frame_name": "end_effector",
"end_effector_bounds": {
"min": [0.16, -0.08, 0.03],
"max": [0.24, 0.2, 0.1]
},
"end_effector_step_sizes": {
"x": 0.02,
"y": 0.02,
"z": 0.02
}
}
}
}
}
```
### Advanced Observation Processing
The HIL-SERL framework supports additional observation processing features that can improve policy learning:
#### Joint Velocity Processing
Enable joint velocity estimation to provide the policy with motion information:
```json
{
"env": {
"processor": {
"observation": {
"add_joint_velocity_to_observation": true
}
}
}
}
```
This processor:
- Estimates joint velocities using finite differences between consecutive joint position readings
- Adds velocity information to the observation state vector
- Useful for policies that need motion awareness for dynamic tasks
#### Motor Current Processing
Monitor motor currents to detect contact forces and load conditions:
```json
{
"env": {
"processor": {
"observation": {
"add_current_to_observation": true
}
}
}
}
```
This processor:
- Reads motor current values from the robot's control system
- Adds current measurements to the observation state vector
- Helps detect contact events, object weights, and mechanical resistance
- Useful for contact-rich manipulation tasks
#### Combined Observation Processing
You can enable multiple observation processing features simultaneously:
```json
{
"env": {
"processor": {
"observation": {
"add_joint_velocity_to_observation": true,
"add_current_to_observation": true,
"display_cameras": false
}
}
}
}
```
**Note**: Enabling additional observation features increases the state space dimensionality, which may require adjusting your policy network architecture and potentially collecting more training data.
### Finding Robot Workspace Bounds
Before collecting demonstrations, you need to determine the appropriate operational bounds for your robot.
This helps simplify the problem of learning on the real robot in two ways: 1) by limiting the robot's operational space to a specific region that solves the task and avoids unnecessary or unsafe exploration, and 2) by allowing training in end-effector space rather than joint space. Empirically, learning in joint space for reinforcement learning in manipulation is often a harder problem - some tasks are nearly impossible to learn in joint space but become learnable when the action space is transformed to end-effector coordinates.
**Using lerobot-find-joint-limits**
This script helps you find the safe operational bounds for your robot's end-effector. Given that you have a follower and leader arm, you can use the script to find the bounds for the follower arm that will be applied during training.
Bounding the action space will reduce the redundant exploration of the agent and guarantees safety.
```bash
lerobot-find-joint-limits \
--robot.type=so100_follower \
--robot.port=/dev/tty.usbmodem58760431541 \
--robot.id=black \
--teleop.type=so100_leader \
--teleop.port=/dev/tty.usbmodem58760431551 \
--teleop.id=blue
```
**Workflow**
1. Run the script and move the robot through the space that solves the task
2. The script will record the minimum and maximum end-effector positions and the joint angles and prints them to the console, for example:
```
Max ee position [0.2417 0.2012 0.1027]
Min ee position [0.1663 -0.0823 0.0336]
Max joint positions [-20.0, -20.0, -20.0, -20.0, -20.0, -20.0]
Min joint positions [50.0, 50.0, 50.0, 50.0, 50.0, 50.0]
```
3. Use these values in your environment configuration under `env.processor.inverse_kinematics.end_effector_bounds` (see `InverseKinematicsConfig` in `lerobot/envs/configs.py`)
**Example Configuration**
```json
{
"env": {
"processor": {
"inverse_kinematics": {
"end_effector_bounds": {
"max": [0.24, 0.2, 0.1],
"min": [0.16, -0.08, 0.03]
}
}
}
}
}
```
### Collecting Demonstrations
With the bounds defined, you can safely collect demonstrations for training. Training RL with off-policy algorithm allows us to use offline datasets collected in order to improve the efficiency of the learning process.
**Setting Up Record Mode**
Create a configuration file for recording demonstrations (or edit an existing one like [env_config.json](https://huggingface.co/datasets/lerobot/config_examples/resolve/main/rl/env_config.json)):
1. Set `mode` to `"record"` at the root level
2. Specify a unique `repo_id` for your dataset in the `dataset` section (e.g., "username/task_name")
3. Set `num_episodes_to_record` in the `dataset` section to the number of demonstrations you want to collect
4. Set `env.processor.image_preprocessing.crop_params_dict` to `{}` initially (we'll determine crops later)
5. Configure `env.robot`, `env.teleop`, and other hardware settings in the `env` section
Example configuration section:
```json
{
"env": {
"type": "gym_manipulator",
"name": "real_robot",
"fps": 10,
"processor": {
"control_mode": "gamepad",
"observation": {
"display_cameras": false
},
"image_preprocessing": {
"crop_params_dict": {},
"resize_size": [128, 128]
},
"gripper": {
"use_gripper": true,
"gripper_penalty": 0.0
},
"reset": {
"reset_time_s": 5.0,
"control_time_s": 20.0
}
},
"robot": {
// ... robot configuration ...
},
"teleop": {
// ... teleoperator configuration ...
}
},
"dataset": {
"repo_id": "username/pick_lift_cube",
"root": null,
"task": "pick_and_lift",
"num_episodes_to_record": 15,
"replay_episode": 0,
"push_to_hub": true
},
"mode": "record",
"device": "cpu"
}
```
### Using a Teleoperation Device
Along with your robot, you will need a teleoperation device to control it in order to collect datasets of your task and perform interventions during the online training.
We support using a gamepad or a keyboard or the leader arm of the robot.
HIL-Serl learns actions in the end-effector space of the robot. Therefore, the teleoperation will control the end-effector's x,y,z displacements.
The end-effector transformation is applied by the processor pipeline (`InverseKinematicsRLStep`, `EEBoundsAndSafety`, `EEReferenceAndDelta`, `GripperVelocityToJoint`) configured under `env.processor.inverse_kinematics` (`InverseKinematicsConfig`) and `env.processor.gripper` / `env.processor.max_gripper_pos`. The defaults related to the end-effector space are:
<!-- prettier-ignore-start -->
```python
class InverseKinematicsConfig:
"""Configuration for inverse kinematics processing."""
urdf_path: str | None = None
target_frame_name: str | None = None
# bounds for the end-effector in x,y,z direction
end_effector_bounds: dict[str, list[float]] | None = None
# maximum step size for the end-effector in x,y,z direction
end_effector_step_sizes: dict[str, float] | None = None
class HILSerlProcessorConfig:
...
# maximum gripper position that the gripper will be open at
max_gripper_pos: float | None = 100.0
```
<!-- prettier-ignore-end -->
The `Teleoperator` defines the teleoperation device. You can check the list of available teleoperators in `lerobot/teleoperators`.
**Setting up the Gamepad**
The gamepad provides a very convenient way to control the robot and the episode state.
To setup the gamepad, you need to set the `control_mode` to `"gamepad"` and define the `teleop` section in the configuration file.
```json
{
"env": {
"teleop": {
"type": "gamepad",
"use_gripper": true
},
"processor": {
"control_mode": "gamepad",
"gripper": {
"use_gripper": true
}
}
}
}
```
<p align="center">
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/gamepad_guide.jpg?raw=true"
alt="Figure shows the control mappings on a Logitech gamepad."
title="Gamepad Control Mapping"
width="100%"
></img>
</p>
<p align="center">
<i>Gamepad button mapping for robot control and episode management</i>
</p>
**Setting up the SO101 leader**
The SO101 leader arm has reduced gears that allows it to move and track the follower arm during exploration. Therefore, taking over is much smoother than the gearless SO100.
To setup the SO101 leader, you need to set the `control_mode` to `"leader"` and define the `teleop` section in the configuration file.
```json
{
"env": {
"teleop": {
"type": "so101_leader",
"port": "/dev/tty.usbmodem585A0077921",
"use_degrees": true
},
"processor": {
"control_mode": "leader",
"gripper": {
"use_gripper": true
}
}
}
}
```
In order to annotate the success/failure of the episode, **you will need** to use a keyboard to press `s` for success, `esc` for failure.
During the online training, press `space` to take over the policy and `space` again to give the control back to the policy.
<details>
<summary><strong>Video: SO101 leader teleoperation</strong></summary>
<div class="video-container">
<video controls width="600">
<source
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/so101_leader_tutorial.mp4"
type="video/mp4"
/>
</video>
</div>
<p align="center"><i>SO101 leader teleoperation example, the leader tracks the follower, press `space` to intervene</i></p>
</details>
**Recording Demonstrations**
Start the recording process, an example of the config file can be found [here](https://huggingface.co/datasets/aractingi/lerobot-example-config-files/blob/main/env_config_so100.json):
```bash
python -m lerobot.rl.gym_manipulator --config_path src/lerobot/configs/env_config_so100.json
```
During recording:
1. The robot will reset to the initial position defined in the configuration file `env.processor.reset.fixed_reset_joint_positions`
2. Complete the task successfully
3. The episode ends with a reward of 1 when you press the "success" button
4. If the time limit is reached, or the fail button is pressed, the episode ends with a reward of 0
5. You can rerecord an episode by pressing the "rerecord" button
6. The process automatically continues to the next episode
7. After recording all episodes, the dataset is pushed to the Hugging Face Hub (optional) and saved locally
### Processing the Dataset
After collecting demonstrations, process them to determine optimal camera crops.
Reinforcement learning is sensitive to background distractions, so it is important to crop the images to the relevant workspace area.
Visual RL algorithms learn directly from pixel inputs, making them vulnerable to irrelevant visual information. Background elements like changing lighting, shadows, people moving, or objects outside the workspace can confuse the learning process. Good ROI selection should:
- Include only the essential workspace where the task happens
- Capture the robot's end-effector and all objects involved in the task
- Exclude unnecessary background elements and distractions
Note: If you already know the crop parameters, you can skip this step and just set the `crop_params_dict` in the configuration file during recording.
**Determining Crop Parameters**
Use the `crop_dataset_roi.py` script to interactively select regions of interest in your camera images:
```bash
python -m lerobot.rl.crop_dataset_roi --repo-id username/pick_lift_cube
```
1. For each camera view, the script will display the first frame
2. Draw a rectangle around the relevant workspace area
3. Press 'c' to confirm the selection
4. Repeat for all camera views
5. The script outputs cropping parameters and creates a new cropped dataset
Example output:
```
Selected Rectangular Regions of Interest (top, left, height, width):
observation.images.side: [180, 207, 180, 200]
observation.images.front: [180, 250, 120, 150]
```
<p align="center">
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/crop_dataset.gif"
width="600"
/>
</p>
<p align="center">
<i>Interactive cropping tool for selecting regions of interest</i>
</p>
**Updating Configuration**
Add these crop parameters to your training configuration:
```json
{
"env": {
"processor": {
"image_preprocessing": {
"crop_params_dict": {
"observation.images.side": [180, 207, 180, 200],
"observation.images.front": [180, 250, 120, 150]
},
"resize_size": [128, 128]
}
}
}
}
```
**Recommended image resolution**
Most vision-based policies have been validated on square inputs of either **128×128** (default) or **64×64** pixels. We therefore advise setting the resize_size parameter to [128, 128] – or [64, 64] if you need to save GPU memory and bandwidth. Other resolutions are possible but have not been extensively tested.
### Training a Reward Classifier
The reward classifier plays an important role in the HIL-SERL workflow by automating reward assignment and automatically detecting episode success. Instead of manually defining reward functions or relying on human feedback for every timestep, the reward classifier learns to predict success/failure from visual observations. This enables the RL algorithm to learn efficiently by providing consistent and automated reward signals based on the robot's camera inputs.
This guide explains how to train a reward classifier for human-in-the-loop reinforcement learning implementation of LeRobot. Reward classifiers learn to predict the reward value given a state which can be used in an RL setup to train a policy.
**Note**: Training a reward classifier is optional. You can start the first round of RL experiments by annotating the success manually with your gamepad or keyboard device.
The reward classifier implementation in `lerobot/rewards/classifier/modeling_classifier.py` uses a pretrained vision model to process the images. It can output either a single value for binary rewards to predict success/fail cases or multiple values for multi-class settings.
**Collecting a Dataset for the reward classifier**
Before training, you need to collect a dataset with labeled examples. Setting `mode: "record"` in your config and running `gym_manipulator.py` enables the process of collecting a dataset of observations, actions, and rewards.
To collect a dataset, you need to modify some parameters in the environment configuration based on HILSerlRobotEnvConfig.
```bash
python -m lerobot.rl.gym_manipulator --config_path src/lerobot/configs/reward_classifier_train_config.json
```
**Key Parameters for Data Collection**
- **mode**: set it to `"record"` to collect a dataset (at root level)
- **dataset.repo_id**: `"hf_username/dataset_name"`, name of the dataset and repo on the hub
- **dataset.num_episodes_to_record**: Number of episodes to record
- **env.processor.reset.terminate_on_success**: Whether to automatically terminate episodes when success is detected (default: `true`)
- **env.fps**: Number of frames per second to record
- **dataset.push_to_hub**: Whether to push the dataset to the hub
The `env.processor.reset.terminate_on_success` parameter allows you to control episode termination behavior. When set to `false`, episodes will continue even after success is detected, allowing you to collect more positive examples with the reward=1 label. This is crucial for training reward classifiers as it provides more success state examples in your dataset. When set to `true` (default), episodes terminate immediately upon success detection.
**Important**: For reward classifier training, set `terminate_on_success: false` to collect sufficient positive examples. For regular HIL-SERL training, keep it as `true` to enable automatic episode termination when the task is completed successfully.
Example configuration section for data collection:
```json
{
"env": {
"type": "gym_manipulator",
"name": "real_robot",
"fps": 10,
"processor": {
"reset": {
"reset_time_s": 5.0,
"control_time_s": 20.0,
"terminate_on_success": false
},
"gripper": {
"use_gripper": true
}
},
"robot": {
// ... robot configuration ...
},
"teleop": {
// ... teleoperator configuration ...
}
},
"dataset": {
"repo_id": "hf_username/dataset_name",
"root": "data/your_dataset",
"task": "reward_classifier_task",
"num_episodes_to_record": 20,
"replay_episode": null,
"push_to_hub": true
},
"mode": "record",
"device": "cpu"
}
```
**Reward Classifier Configuration**
The reward classifier is configured using `lerobot/rewards/classifier/configuration_classifier.py`. Here are the key parameters:
- **model_name**: Base model architecture (e.g., we mainly use `"helper2424/resnet10"`)
- **model_type**: `"cnn"` or `"transformer"`
- **num_cameras**: Number of camera inputs
- **num_classes**: Number of output classes (typically 2 for binary success/failure)
- **hidden_dim**: Size of hidden representation
- **dropout_rate**: Regularization parameter
- **learning_rate**: Learning rate for optimizer
Example configuration for training the [reward classifier](https://huggingface.co/datasets/aractingi/lerobot-example-config-files/blob/main/reward_classifier_train_config.json):
```json
{
"dataset": {
"repo_id": "hf_username/dataset_name",
"root": null
},
"reward_model": {
"type": "reward_classifier",
"model_name": "helper2424/resnet10",
"model_type": "cnn",
"num_cameras": 2,
"num_classes": 2,
"hidden_dim": 256,
"dropout_rate": 0.1,
"learning_rate": 1e-4,
"device": "cuda",
"input_features": {
"observation.images.front": {
"type": "VISUAL",
"shape": [3, 128, 128]
},
"observation.images.side": {
"type": "VISUAL",
"shape": [3, 128, 128]
}
},
"push_to_hub": true,
"repo_id": "hf_username/model_repo"
},
"batch_size": 16,
"num_workers": 4,
"steps": 5000,
"log_freq": 10,
"env_eval_freq": 1000,
"save_freq": 1000,
"save_checkpoint": true,
"seed": 2,
"resume": false,
"optimizer": {
"grad_clip_norm": 10.0
},
"wandb": {
"enable": true,
"project": "reward-classifier",
"disable_artifact": false
},
"job_name": "reward-classifier"
}
```
**Training the Classifier**
To train the classifier, use the `train.py` script with your configuration:
```bash
lerobot-train --config_path path/to/reward_classifier_train_config.json
```
**Deploying and Testing the Model**
To use your trained reward classifier, configure the `HILSerlRobotEnvConfig` to use your model:
<!-- prettier-ignore-start -->
```python
config = GymManipulatorConfig(
env=HILSerlRobotEnvConfig(
processor=HILSerlProcessorConfig(
reward_classifier=RewardClassifierConfig(
pretrained_path="path_to_your_pretrained_trained_model"
)
),
# Other environment parameters
),
dataset=DatasetConfig(...),
mode=None # For training
)
```
<!-- prettier-ignore-end -->
or set the argument in the json config file.
```json
{
"env": {
"processor": {
"reward_classifier": {
"pretrained_path": "path_to_your_pretrained_model",
"success_threshold": 0.7,
"success_reward": 1.0
},
"reset": {
"terminate_on_success": true
}
}
}
}
```
Run `gym_manipulator.py` to test the model.
```bash
python -m lerobot.rl.gym_manipulator --config_path path/to/env_config.json
```
The reward classifier will automatically provide rewards based on the visual input from the robot's cameras.
**Example Workflow for training the reward classifier**
1. **Create the configuration files**:
Create the necessary json configuration files for the reward classifier and the environment. Check the examples [here](https://huggingface.co/datasets/lerobot/config_examples/resolve/main/reward_classifier/config.json).
2. **Collect a dataset**:
```bash
python -m lerobot.rl.gym_manipulator --config_path src/lerobot/configs/env_config.json
```
3. **Train the classifier**:
```bash
lerobot-train --config_path src/lerobot/configs/reward_classifier_train_config.json
```
4. **Test the classifier**:
```bash
python -m lerobot.rl.gym_manipulator --config_path src/lerobot/configs/env_config.json
```
### Training with Actor-Learner
The LeRobot system uses a distributed actor-learner architecture for training. This architecture decouples robot interactions from the learning process, allowing them to run concurrently without blocking each other. The actor server handles robot observations and actions, sending interaction data to the learner server. The learner server performs gradient descent and periodically updates the actor's policy weights. You will need to start two processes: a learner and an actor.
**Configuration Setup**
Create a training configuration file (example available [here](https://huggingface.co/datasets/lerobot/config_examples/resolve/main/rl/train_config.json)). The training config is based on the main `TrainRLServerPipelineConfig` class in `lerobot/rl/train_rl.py`.
1. Configure the policy settings (`type="gaussian_actor"`, `device`, etc.)
2. Configure the algorithm settings under the top-level `algorithm` block (`type="sac"`, learning rates, discount, etc., defined in `lerobot/rl/algorithms/sac/configuration_sac.py`).
3. Set `dataset` to your cropped dataset
4. Configure environment settings with crop parameters
5. Check the other parameters related to the Gaussian Actor in [configuration_gaussian_actor.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/gaussian_actor/configuration_gaussian_actor.py#L79).
6. Verify that the `policy` config is correct with the right `input_features` and `output_features` for your task.
**Starting the Learner**
First, start the learner server process:
```bash
python -m lerobot.rl.learner --config_path src/lerobot/configs/train_config_hilserl_so100.json
```
The learner:
- Initializes the policy network
- Prepares replay buffers
- Opens a `gRPC` server to communicate with actors
- Processes transitions and updates the policy
**Starting the Actor**
In a separate terminal, start the actor process with the same configuration:
```bash
python -m lerobot.rl.actor --config_path src/lerobot/configs/train_config_hilserl_so100.json
```
The actor:
- Connects to the learner via `gRPC`
- Initializes the environment
- Execute rollouts of the policy to collect experience
- Sends transitions to the learner
- Receives updated policy parameters
**Training Flow**
The training proceeds automatically:
1. The actor executes the policy in the environment
2. Transitions are collected and sent to the learner
3. The learner updates the policy based on these transitions
4. Updated policy parameters are sent back to the actor
5. The process continues until the specified step limit is reached
**Human in the Loop**
- The key to learning efficiently is to have human interventions to provide corrective feedback and completing the task to aide the policy learning and exploration.
- To perform human interventions, you can press the upper right trigger button on the gamepad (or the `space` key on the keyboard). This will pause the policy actions and allow you to take over.
- A successful experiment is one where the human has to intervene at the start but then reduces the amount of interventions as the policy improves. You can monitor the intervention rate in the `wandb` dashboard.
<p align="center">
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/hil_effect.png?raw=true"
alt="Figure shows the control mappings on a Logitech gamepad."
title="Gamepad Control Mapping"
width="100%"
></img>
</p>
<p align="center">
<i>
Example showing how human interventions help guide policy learning over time
</i>
</p>
- The figure shows the plot of the episodic reward over interaction step. The figure shows the effect of human interventions on the policy learning.
- The orange curve is an experiment without any human interventions. While the pink and blue curves are experiments with human interventions.
- We can observe that the number of steps where the policy starts achieving the maximum reward is cut by a quarter when human interventions are present.
**Monitoring and Debugging**
If you have `wandb.enable` set to `true` in your configuration, you can monitor training progress in real-time through the [Weights & Biases](https://wandb.ai/site/) dashboard.
### Guide to Human Interventions
The learning process is very sensitive to the intervention strategy. It will takes a few runs to understand how to intervene effectively. Some tips and hints:
- Allow the policy to explore for a few episodes at the start of training.
- Avoid intervening for long periods of time. Try to intervene in situation to correct the robot's behaviour when it goes off track.
- Once the policy starts achieving the task, even if its not perfect, you can limit your interventions to simple quick actions like a simple grasping commands.
The ideal behaviour is that your intervention rate should drop gradually during training as shown in the figure below.
<p align="center">
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/intervention_rate_tutorial_rl.png?raw=true"
alt="Intervention rate"
title="Intervention rate during training"
width="100%"
></img>
</p>
<p align="center">
<i>
Plot of the intervention rate during a training run on a pick and lift cube
task
</i>
</p>
### Key hyperparameters to tune
Some configuration values have a disproportionate impact on training stability and speed:
- **`temperature_init`** (`algorithm.temperature_init`) – initial entropy temperature in SAC. Higher values encourage more exploration; lower values make the policy more deterministic early on. A good starting point is `1e-2`. We observed that setting it too high can make human interventions ineffective and slow down learning.
- **`policy_parameters_push_frequency`** (`policy.actor_learner_config.policy_parameters_push_frequency`) – interval in _seconds_ between two weight pushes from the learner to the actor. The default is `4 s`. Decrease to **1-2 s** to provide fresher weights (at the cost of more network traffic); increase only if your connection is slow, as this will reduce sample efficiency.
- **`storage_device`** (`policy.storage_device`) – device on which the learner keeps the policy parameters. If you have spare GPU memory, set this to `"cuda"` (instead of the default `"cpu"`). Keeping the weights on-GPU removes CPU→GPU transfer overhead and can significantly increase the number of learner updates per second.
Congrats 🎉, you have finished this tutorial!
> [!TIP]
> If you have any questions or need help, please reach out on [Discord](https://discord.com/invite/s3KuuzsPFb).
Paper citation:
```
@article{luo2024precise,
title={Precise and Dexterous Robotic Manipulation via Human-in-the-Loop Reinforcement Learning},
author={Luo, Jianlan and Xu, Charles and Wu, Jeffrey and Levine, Sergey},
journal={arXiv preprint arXiv:2410.21845},
year={2024}
}
```
docs/source/hilserl_sim.mdx
+154
View File
@@ -0,0 +1,154 @@
# Train RL in Simulation
This guide explains how to use the `gym_hil` simulation environments as an alternative to real robots when working with the LeRobot framework for Human-In-the-Loop (HIL) reinforcement learning.
`gym_hil` is a package that provides Gymnasium-compatible simulation environments specifically designed for Human-In-the-Loop reinforcement learning. These environments allow you to:
- Train policies in simulation to test the RL stack before training on real robots
- Collect demonstrations in sim using external devices like gamepads or keyboards
- Perform human interventions during policy learning
Currently, the main environment is a Franka Panda robot simulation based on MuJoCo, with tasks like picking up a cube.
## Installation
First, install the `gym_hil` package within the LeRobot environment:
```bash
pip install -e ".[hilserl]"
```
## What do I need?
- A gamepad or keyboard to control the robot
- A Nvidia GPU
## Configuration
To use `gym_hil` with LeRobot, you need to create a configuration file. An example is provided [here](https://huggingface.co/datasets/lerobot/config_examples/resolve/main/rl/gym_hil/env_config.json). Key configuration sections include:
### Environment Type and Task
```json
{
"env": {
"type": "gym_manipulator",
"name": "gym_hil",
"task": "PandaPickCubeGamepad-v0",
"fps": 10
},
"device": "cuda"
}
```
Available tasks:
- `PandaPickCubeBase-v0`: Basic environment
- `PandaPickCubeGamepad-v0`: With gamepad control
- `PandaPickCubeKeyboard-v0`: With keyboard control
### Processor Configuration
```json
{
"env": {
"processor": {
"control_mode": "gamepad",
"gripper": {
"use_gripper": true,
"gripper_penalty": -0.02
},
"reset": {
"control_time_s": 15.0,
"fixed_reset_joint_positions": [
0.0, 0.195, 0.0, -2.43, 0.0, 2.62, 0.785
]
},
"inverse_kinematics": {
"end_effector_step_sizes": {
"x": 0.025,
"y": 0.025,
"z": 0.025
}
}
}
}
}
```
Important parameters:
- `gripper.gripper_penalty`: Penalty for excessive gripper movement
- `gripper.use_gripper`: Whether to enable gripper control
- `inverse_kinematics.end_effector_step_sizes`: Size of the steps in the x,y,z axes of the end-effector
- `control_mode`: Set to `"gamepad"` to use a gamepad controller
## Running with HIL RL of LeRobot
### Basic Usage
To run the environment, set mode to null:
```bash
python -m lerobot.rl.gym_manipulator --config_path path/to/gym_hil_env.json
```
### Recording a Dataset
To collect a dataset, set the mode to `record` whilst defining the repo_id and number of episodes to record:
```json
{
"env": {
"type": "gym_manipulator",
"name": "gym_hil",
"task": "PandaPickCubeGamepad-v0"
},
"dataset": {
"repo_id": "username/sim_dataset",
"root": null,
"task": "pick_cube",
"num_episodes_to_record": 10,
"replay_episode": null,
"push_to_hub": true
},
"mode": "record"
}
```
```bash
python -m lerobot.rl.gym_manipulator --config_path path/to/gym_hil_env.json
```
### Training a Policy
To train a policy, checkout the configuration example available [here](https://huggingface.co/datasets/lerobot/config_examples/resolve/main/rl/gym_hil/train_config.json) and run the actor and learner servers:
```bash
python -m lerobot.rl.actor --config_path path/to/train_gym_hil_env.json
```
In a different terminal, run the learner server:
```bash
python -m lerobot.rl.learner --config_path path/to/train_gym_hil_env.json
```
The simulation environment provides a safe and repeatable way to develop and test your Human-In-the-Loop reinforcement learning components before deploying to real robots.
Congrats 🎉, you have finished this tutorial!
> [!TIP]
> If you have any questions or need help, please reach out on [Discord](https://discord.com/invite/s3KuuzsPFb).
Paper citation:
```
@article{luo2024precise,
title={Precise and Dexterous Robotic Manipulation via Human-in-the-Loop Reinforcement Learning},
author={Luo, Jianlan and Xu, Charles and Wu, Jeffrey and Levine, Sergey},
journal={arXiv preprint arXiv:2410.21845},
year={2024}
}
```
docs/source/hope_jr.mdx
+283
View File
@@ -0,0 +1,283 @@
# HopeJR
## Prerequisites
- [Hardware Setup](https://github.com/TheRobotStudio/HOPEJr)
## Install LeRobot
Follow the [installation instructions](https://github.com/huggingface/lerobot#installation) to install LeRobot.
Install LeRobot with HopeJR dependencies:
```bash
pip install -e ".[hopejr]"
```
## Device Configuration
Before starting calibration and operation, you need to identify the USB ports for each HopeJR component. Run this script to find the USB ports for the arm, hand, glove, and exoskeleton:
```bash
lerobot-find-port
```
This will display the available USB ports and their associated devices. Make note of the port paths (e.g., `/dev/tty.usbmodem58760433331`, `/dev/tty.usbmodem11301`) as you'll need to specify them in the `--robot.port` and `--teleop.port` parameters when recording data, replaying episodes, or running teleoperation scripts.
## Step 1: Calibration
Before performing teleoperation, HopeJR's limbs need to be calibrated. Calibration files will be saved in `~/.cache/huggingface/lerobot/calibration`
### 1.1 Calibrate Robot Hand
```bash
lerobot-calibrate \
--robot.type=hope_jr_hand \
--robot.port=/dev/tty.usbmodem58760432281 \
--robot.id=blue \
--robot.side=right
```
When running the calibration script, a calibration GUI will pop up. Finger joints are named as follows:
**Thumb**:
- **CMC**: base joint connecting thumb to hand
- **MCP**: knuckle joint
- **PIP**: first finger joint
- **DIP** : fingertip joint
**Index, Middle, Ring, and Pinky fingers**:
- **Radial flexor**: Moves base of finger towards the thumb
- **Ulnar flexor**: Moves base of finger towards the pinky
- **PIP/DIP**: Flexes the distal and proximal phalanx of the finger
Each one of these will need to be calibrated individually via the GUI.
Note that ulnar and radial flexors should have ranges of the same size (but with different offsets) in order to get symmetric movement.
<p align="center">
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/calibration_gui_1.png"
alt="Setting boundaries in the hand calibration GUI"
title="Setting boundaries in the hand calibration GUI"
width="100%"
></img>
</p>
Use the calibration interface to set the range boundaries for each joint as shown above.
<p align="center">
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/calibration_gui_2.png"
alt="Saving calibration values"
title="Saving calibration values"
width="100%"
></img>
</p>
Once you have set the appropriate boundaries for all joints, click "Save" to save the calibration values to the motors.
### 1.2 Calibrate Teleoperator Glove
```bash
lerobot-calibrate \
--teleop.type=homunculus_glove \
--teleop.port=/dev/tty.usbmodem11201 \
--teleop.id=red \
--teleop.side=right
```
Move each finger through its full range of motion, starting from the thumb.
```
Move thumb through its entire range of motion.
Recording positions. Press ENTER to stop...
-------------------------------------------
NAME | MIN | POS | MAX
thumb_cmc | 1790 | 1831 | 1853
thumb_mcp | 1497 | 1514 | 1528
thumb_pip | 1466 | 1496 | 1515
thumb_dip | 1463 | 1484 | 1514
```
Continue with each finger:
```
Move middle through its entire range of motion.
Recording positions. Press ENTER to stop...
-------------------------------------------
NAME | MIN | POS | MAX
middle_mcp_abduction | 1598 | 1718 | 1820
middle_mcp_flexion | 1512 | 1658 | 2136
middle_dip | 1484 | 1500 | 1547
```
Once calibration is complete, the system will save the calibration to `/Users/your_username/.cache/huggingface/lerobot/calibration/teleoperators/homunculus_glove/red.json`
### 1.3 Calibrate Robot Arm
```bash
lerobot-calibrate \
--robot.type=hope_jr_arm \
--robot.port=/dev/tty.usbserial-1110 \
--robot.id=white
```
This will open a calibration GUI where you can set the range limits for each motor. The arm motions are organized as follows:
- **Shoulder**: pitch, yaw, and roll
- **Elbow**: flex
- **Wrist**: pitch, yaw, and roll
<p align="center">
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/calibration_gui_2.png"
alt="Setting boundaries in the arm calibration GUI"
title="Setting boundaries in the arm calibration GUI"
width="100%"
></img>
</p>
Use the calibration interface to set the range boundaries for each joint. Move each joint through its full range of motion and adjust the minimum and maximum values accordingly. Once you have set the appropriate boundaries for all joints, save the calibration.
### 1.4 Calibrate Teleoperator Exoskeleton
```bash
lerobot-calibrate \
--teleop.type=homunculus_arm \
--teleop.port=/dev/tty.usbmodem11201 \
--teleop.id=black
```
The exoskeleton allows one to control the robot arm. During calibration, you'll be prompted to move all joints through their full range of motion:
```
Move all joints through their entire range of motion.
Recording positions. Press ENTER to stop...
-------------------------------------------
-------------------------------------------
NAME | MIN | POS | MAX
shoulder_pitch | 586 | 736 | 895
shoulder_yaw | 1257 | 1374 | 1390
shoulder_roll | 449 | 1034 | 2564
elbow_flex | 3023 | 3117 | 3134
wrist_roll | 3073 | 3096 | 3147
wrist_yaw | 2143 | 2171 | 2185
wrist_pitch | 1975 | 1993 | 2074
Calibration saved to /Users/your_username/.cache/huggingface/lerobot/calibration/teleoperators/homunculus_arm/black.json
```
## Step 2: Teleoperation
Due to global variable conflicts in the Feetech middleware, teleoperation for arm and hand must run in separate shell sessions:
### Hand
```bash
lerobot-teleoperate \
--robot.type=hope_jr_hand \
--robot.port=/dev/tty.usbmodem58760432281 \
--robot.id=blue \
--robot.side=right \
--teleop.type=homunculus_glove \
--teleop.port=/dev/tty.usbmodem11201 \
--teleop.id=red \
--teleop.side=right \
--display_data=true \
--fps=30
```
### Arm
```bash
lerobot-teleoperate \
--robot.type=hope_jr_arm \
--robot.port=/dev/tty.usbserial-1110 \
--robot.id=white \
--teleop.type=homunculus_arm \
--teleop.port=/dev/tty.usbmodem11201 \
--teleop.id=black \
--display_data=true \
--fps=30
```
## Step 3: Record, Replay, Train
Record, Replay and Train with Hope-JR is still experimental.
### Record
This step records the dataset, which can be seen as an example [here](https://huggingface.co/datasets/nepyope/hand_record_test_with_video_data/settings).
```bash
lerobot-record \
--robot.type=hope_jr_hand \
--robot.port=/dev/tty.usbmodem58760432281 \
--robot.id=right \
--robot.side=right \
--robot.cameras='{"main": {"type": "opencv", "index_or_path": 0, "width": 640, "height": 480, "fps": 30}}' \
--teleop.type=homunculus_glove \
--teleop.port=/dev/tty.usbmodem1201 \
--teleop.id=right \
--teleop.side=right \
--dataset.repo_id=<USER>/hand_record_test_with_video_data \
--dataset.single_task="Hand recording test with video data" \
--dataset.num_episodes=1 \
--dataset.episode_time_s=5 \
--dataset.push_to_hub=true \
--dataset.private=true \
--dataset.streaming_encoding=true \
--dataset.encoder_threads=2 \
# --dataset.rgb_encoder.vcodec=auto \
--display_data=true
```
### Replay
```bash
lerobot-replay \
--robot.type=hope_jr_hand \
--robot.port=/dev/tty.usbmodem58760432281 \
--robot.id=right \
--robot.side=right \
--dataset.repo_id=<USER>/hand_record_test_with_camera \
--dataset.episode=0
```
### Train
```bash
lerobot-train \
--dataset.repo_id=<USER>/hand_record_test_with_video_data \
--policy.type=act \
--output_dir=outputs/train/hopejr_hand \
--job_name=hopejr \
--policy.device=mps \
--wandb.enable=true \
--policy.repo_id=<USER>/hand_test_policy
```
### Evaluate
This training run can be viewed as an example [here](https://wandb.ai/tino/lerobot/runs/rp0k8zvw?nw=nwusertino).
```bash
lerobot-record \
--robot.type=hope_jr_hand \
--robot.port=/dev/tty.usbmodem58760432281 \
--robot.id=right \
--robot.side=right \
--robot.cameras='{"main": {"type": "opencv", "index_or_path": 0, "width": 640, "height": 480, "fps": 30}}' \
--display_data=false \
--dataset.repo_id=<USER>/eval_hopejr \
--dataset.single_task="Evaluate hopejr hand policy" \
--dataset.num_episodes=10 \
--dataset.streaming_encoding=true \
--dataset.encoder_threads=2 \
# --dataset.rgb_encoder.vcodec=auto \
--policy.path=outputs/train/hopejr_hand/checkpoints/last/pretrained_model
```
docs/source/il_robots.mdx
+638
View File
@@ -0,0 +1,638 @@
# Imitation Learning on Real-World Robots
This tutorial will explain how to train a neural network to control a real robot autonomously.
**You'll learn:**
1. How to record and visualize your dataset.
2. How to train a policy using your data and prepare it for evaluation.
3. How to evaluate your policy and visualize the results.
By following these steps, you'll be able to replicate tasks, such as picking up a Lego block and placing it in a bin with a high success rate, as shown in the video below.
<details>
<summary><strong>Video: pickup lego block task</strong></summary>
<div class="video-container">
<video controls width="600">
<source
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/lerobot_task.mp4"
type="video/mp4"
/>
</video>
</div>
</details>
This tutorial isn’t tied to a specific robot: we walk you through the commands and API snippets you can adapt for any supported platform.
During data collection, you’ll use a “teloperation” device, such as a leader arm or keyboard to teleoperate the robot and record its motion trajectories.
Once you’ve gathered enough trajectories, you’ll train a neural network to imitate these trajectories and deploy the trained model so your robot can perform the task autonomously.
If you run into any issues at any point, jump into our [Discord community](https://discord.com/invite/s3KuuzsPFb) for support.
<Tip>
Want to quickly get the right commands for your setup? The [quickstart notebook](https://github.com/huggingface/lerobot/blob/main/examples/notebooks/quickstart.ipynb) [![Open in Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/lerobot/blob/main/examples/notebooks/quickstart.ipynb) lets you configure your robot once and generates all the commands below ready to paste.
</Tip>
## Set up and Calibrate
If you haven't yet set up and calibrated your robot and teleop device, please do so by following the robot-specific tutorial.
## Teleoperate
In this example, we’ll demonstrate how to teleoperate the SO101 robot. For each command, we also provide a corresponding API example.
Note that the `id` associated with a robot is used to store the calibration file. It's important to use the same `id` when teleoperating, recording, and evaluating when using the same setup.
<hfoptions id="teleoperate_so101">
<hfoption id="Command">
```bash
lerobot-teleoperate \
--robot.type=so101_follower \
--robot.port=/dev/tty.usbmodem58760431541 \
--robot.id=my_awesome_follower_arm \
--teleop.type=so101_leader \
--teleop.port=/dev/tty.usbmodem58760431551 \
--teleop.id=my_awesome_leader_arm
```
</hfoption>
<hfoption id="API example">
<!-- prettier-ignore-start -->
```python
from lerobot.teleoperators.so_leader import SO101Leader, SO101LeaderConfig
from lerobot.robots.so_follower import SO101Follower, SO101FollowerConfig
robot_config = SO101FollowerConfig(
port="/dev/tty.usbmodem5AB90687491",
id="my_follower_arm",
)
teleop_config = SO101LeaderConfig(
port="/dev/tty.usbmodem5AB90689011",
id="my_leader_arm",
)
robot = SO101Follower(robot_config)
teleop_device = SO101Leader(teleop_config)
robot.connect()
teleop_device.connect()
while True:
action = teleop_device.get_action()
robot.send_action(action)
```
<!-- prettier-ignore-end -->
</hfoption>
</hfoptions>
The teleoperate command will automatically:
1. Identify any missing calibrations and initiate the calibration procedure.
2. Connect the robot and teleop device and start teleoperation.
## Cameras
To add cameras to your setup, follow this [Guide](./cameras#setup-cameras).
## Teleoperate with cameras
With `rerun`, you can teleoperate again while simultaneously visualizing the camera feeds and joint positions. In this example, we’re using the Koch arm.
<hfoptions id="teleoperate_koch_camera">
<hfoption id="Command">
```bash
lerobot-teleoperate \
--robot.type=so101_follower \
--robot.port=/dev/tty.usbmodem5AB90687491 \
--robot.id=my_follower_arm \
--robot.cameras="{front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
--teleop.type=so101_leader \
--teleop.port=/dev/tty.usbmodem5AB90689011 \
--teleop.id=my_leader_arm \
--display_data=true
```
</hfoption>
<hfoption id="API example">
<!-- prettier-ignore-start -->
```python
import time
from lerobot.teleoperators.so_leader import SO101Leader, SO101LeaderConfig
from lerobot.robots.so_follower import SO101Follower, SO101FollowerConfig
from lerobot.cameras.opencv import OpenCVCameraConfig
from lerobot.utils.visualization_utils import init_visualization, log_visualization_data, shutdown_visualization
robot_config = SO101FollowerConfig(
port="/dev/tty.usbmodem5AB90687491",
id="my_follower_arm",
cameras={
"wrist": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
"top": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30)
}
)
teleop_config = SO101LeaderConfig(
port="/dev/tty.usbmodem5AB90689011",
id="my_leader_arm",
)
init_visualization("rerun", session_name="teleoperation") # pass "foxglove" to stream to Foxglove instead
robot = SO101Follower(robot_config)
teleop_device = SO101Leader(teleop_config)
robot.connect()
teleop_device.connect()
TARGET_HZ = 30
TIME_PER_FRAME = 1.0 / TARGET_HZ
while True:
start_time = time.perf_counter()
observation = robot.get_observation()
action = teleop_device.get_action()
robot.send_action(action)
log_visualization_data("rerun", observation=observation, action=action)
elapsed_time = time.perf_counter() - start_time
sleep_time = TIME_PER_FRAME - elapsed_time
if sleep_time > 0:
time.sleep(sleep_time)
```
<!-- prettier-ignore-end -->
</hfoption>
</hfoptions>
## Record a dataset
Once you're familiar with teleoperation, you can record your first dataset.
We use the Hugging Face hub features for uploading your dataset. If you haven't previously used the Hub, make sure you can login via the cli using a write-access token, this token can be generated from the [Hugging Face settings](https://huggingface.co/settings/tokens).
Add your token to the CLI by running this command:
```bash
hf auth login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
```
Then store your Hugging Face repository name in a variable:
```bash
HF_USER=$(NO_COLOR=1 hf auth whoami | awk -F': *' 'NR==1 {print $2}')
echo $HF_USER
```
Now you can record a dataset. To record 5 episodes and upload your dataset to the hub, adapt the code below for your robot and execute the command or API example.
<hfoptions id="record">
<hfoption id="Command">
```bash
lerobot-record \
--robot.type=so101_follower \
--robot.port=/dev/tty.usbmodem585A0076841 \
--robot.id=my_awesome_follower_arm \
--robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 1920, height: 1080, fps: 30}}" \
--teleop.type=so101_leader \
--teleop.port=/dev/tty.usbmodem58760431551 \
--teleop.id=my_awesome_leader_arm \
--display_data=true \
--dataset.repo_id=${HF_USER}/record-test \
--dataset.num_episodes=5 \
--dataset.single_task="Grab the black cube" \
--dataset.streaming_encoding=true \
# --dataset.rgb_encoder.vcodec=auto \
--dataset.encoder_threads=2
```
</hfoption>
<hfoption id="API example">
<!-- prettier-ignore-start -->
```python
from lerobot.cameras.opencv import OpenCVCameraConfig
from lerobot.datasets.lerobot_dataset import LeRobotDataset
from lerobot.utils.feature_utils import hw_to_dataset_features
from lerobot.robots.so_follower import SO101Follower, SO101FollowerConfig
from lerobot.teleoperators.so_leader.config_so_leader import SO101LeaderConfig
from lerobot.teleoperators.so_leader.so_leader import SO101Leader
from lerobot.common.control_utils import init_keyboard_listener
from lerobot.utils.utils import log_say
from lerobot.utils.visualization_utils import init_visualization
from lerobot.scripts.lerobot_record import record_loop
from lerobot.processor import make_default_processors
NUM_EPISODES = 5
FPS = 30
EPISODE_TIME_SEC = 60
RESET_TIME_SEC = 10
TASK_DESCRIPTION = "My task description"
def main():
# Create robot configuration
robot_config = SO101FollowerConfig(
port="/dev/tty.usbmodem5AB90687491",
id="my_follower_arm",
cameras={
"wrist": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
"top": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30)
}
)
teleop_config = SO101LeaderConfig(
port="/dev/tty.usbmodem5AB90689011",
id="my_leader_arm",
)
# Initialize the robot and teleoperator
robot = SO101Follower(robot_config)
teleop = SO101Leader(teleop_config)
# Configure the dataset features
action_features = hw_to_dataset_features(robot.action_features, "action")
obs_features = hw_to_dataset_features(robot.observation_features, "observation")
dataset_features = {**action_features, **obs_features}
# Create the dataset
dataset = LeRobotDataset.create(
repo_id="<hf_username>/<dataset_repo_id>",
fps=FPS,
features=dataset_features,
robot_type=robot.name,
use_videos=True,
image_writer_threads=4,
)
# Initialize the keyboard listener and rerun visualization
_, events = init_keyboard_listener()
init_visualization("rerun", session_name="recording")
# Connect the robot and teleoperator
robot.connect()
teleop.connect()
# Create the required processors
teleop_action_processor, robot_action_processor, robot_observation_processor = make_default_processors()
episode_idx = 0
while episode_idx < NUM_EPISODES and not events["stop_recording"]:
log_say(f"Recording episode {episode_idx + 1} of {NUM_EPISODES}")
record_loop(
robot=robot,
events=events,
fps=FPS,
teleop_action_processor=teleop_action_processor,
robot_action_processor=robot_action_processor,
robot_observation_processor=robot_observation_processor,
teleop=teleop,
dataset=dataset,
control_time_s=EPISODE_TIME_SEC,
single_task=TASK_DESCRIPTION,
display_data=True,
)
# Reset the environment if not stopping or re-recording
if not events["stop_recording"] and (episode_idx < NUM_EPISODES - 1 or events["rerecord_episode"]):
log_say("Reset the environment")
record_loop(
robot=robot,
events=events,
fps=FPS,
teleop_action_processor=teleop_action_processor,
robot_action_processor=robot_action_processor,
robot_observation_processor=robot_observation_processor,
teleop=teleop,
control_time_s=RESET_TIME_SEC,
single_task=TASK_DESCRIPTION,
display_data=True,
)
if events["rerecord_episode"]:
log_say("Re-recording episode")
events["rerecord_episode"] = False
events["exit_early"] = False
dataset.clear_episode_buffer()
continue
dataset.save_episode()
episode_idx += 1
# finalize dataset
log_say("Finalizing dataset...")
dataset.finalize()
# Clean up
log_say("Stop recording")
robot.disconnect()
teleop.disconnect()
dataset.push_to_hub()
if __name__ == "__main__":
main()
```
<!-- prettier-ignore-end -->
</hfoption>
</hfoptions>
#### Dataset upload
Locally, your dataset is stored in this folder: `~/.cache/huggingface/lerobot/{repo-id}`. At the end of data recording, your dataset will be uploaded on your Hugging Face page (e.g. `https://huggingface.co/datasets/${HF_USER}/so101_test`) that you can obtain by running:
```bash
echo https://huggingface.co/datasets/${HF_USER}/so101_test
```
Your dataset will be automatically tagged with `LeRobot` for the community to find it easily, and you can also add custom tags (in this case `tutorial` for example).
You can look for other LeRobot datasets on the hub by searching for `LeRobot` [tags](https://huggingface.co/datasets?other=LeRobot).
You can also push your local dataset to the Hub manually, running:
```bash
hf upload ${HF_USER}/record-test ~/.cache/huggingface/lerobot/{repo-id} --repo-type dataset
```
#### Record function
The `record` function provides a suite of tools for capturing and managing data during robot operation:
##### 1. Data Storage
- Data is stored using the `LeRobotDataset` format and is stored on disk during recording.
- By default, the dataset is pushed to your Hugging Face page after recording.
- To disable uploading, use `--dataset.push_to_hub=False`.
##### 2. Checkpointing and Resuming
- Checkpoints are automatically created during recording.
- If an issue occurs or you want to record additional episodes in the same dataset, you can resume by re-running the same command with `--resume=true`. When resuming a recording, `--dataset.num_episodes` must be set to the **number of additional episodes to be recorded**, and not to the targeted total number of episodes in the dataset! Make sure that you also set `--dataset.root="local_path"`, it's a local path to save the new part of the dataset and is required to resume.
- To start recording from scratch, **manually delete** the dataset directory.
##### 3. Recording Parameters
Set the flow of data recording using command-line arguments:
- `--dataset.episode_time_s=60`
Duration of each data recording episode (default: **60 seconds**).
- `--dataset.reset_time_s=60`
Duration for resetting the environment after each episode (default: **60 seconds**).
- `--dataset.num_episodes=50`
Total number of episodes to record (default: **50**).
##### 4. Keyboard Controls During Recording
Control the data recording flow using keyboard shortcuts:
- Press **Right Arrow (`→`)** or **`n`**: Early stop the current episode or reset time and move to the next.
- Press **Left Arrow (`←`)** or **`r`**: Cancel the current episode and re-record it.
- Press **Escape (`ESC`)** or **`q`**: Immediately stop the session, encode videos, and upload the dataset.
<Tip>
These control-flow shortcuts work on **X11, Wayland, and headless/SSH** sessions. When a global keyboard backend isn't available (Wayland, a headless machine, or macOS without Accessibility permission), `lerobot-record` automatically reads the same keys from the terminal — launch it from an interactive terminal and keep it focused. You can also use the letter equivalents **`n`** (next, same as `→`), **`r`** (re-record, same as `←`) and **`q`** (quit, same as `ESC`). No `$DISPLAY` setup is required.
This applies to the recording control flow only. Keyboard **teleoperation** (driving the robot with the keyboard) still needs a global key backend, so it works only on an X11 session, a Windows desktop, or macOS with Accessibility/Input Monitoring granted — not on Wayland or headless sessions.
</Tip>
#### Tips for gathering data
Once you're comfortable with data recording, you can create a larger dataset for training. A good starting task is grasping an object at different locations and placing it in a bin. We suggest recording at least 50 episodes, with 10 episodes per location. Keep the cameras fixed and maintain consistent grasping behavior throughout the recordings. Also make sure the object you are manipulating is visible on the camera's. A good rule of thumb is you should be able to do the task yourself by only looking at the camera images.
In the following sections, you’ll train your neural network. After achieving reliable grasping performance, you can start introducing more variations during data collection, such as additional grasp locations, different grasping techniques, and altering camera positions.
Avoid adding too much variation too quickly, as it may hinder your results.
If you want to dive deeper into this important topic, you can check out the [blog post](https://huggingface.co/blog/lerobot-datasets#what-makes-a-good-dataset) we wrote on what makes a good dataset.
#### Troubleshooting:
- On Linux, the recording control-flow keys (arrow keys, Escape) work on X11, Wayland, and headless/SSH sessions as long as `lerobot-record` runs in an interactive terminal — no `$DISPLAY` setup is needed. If the keys have no effect, make sure you are in an interactive (TTY) terminal, not a piped/non-TTY session, and that it is focused; the letter equivalents `n` / `r` / `q` also work. Keyboard _teleoperation_ (as opposed to the recording control flow) still requires a global key backend — an X11 session, a Windows desktop, or macOS with Accessibility/Input Monitoring granted — and is unavailable on Wayland or headless machines. See [pynput limitations](https://pynput.readthedocs.io/en/latest/limitations.html#linux).
## Visualize a dataset
If you uploaded your dataset to the hub with `--control.push_to_hub=true`, you can [visualize your dataset online](https://huggingface.co/spaces/lerobot/visualize_dataset) by copy pasting your repo id given by:
```bash
echo ${HF_USER}/so101_test
```
## Replay an episode
A useful feature is the `replay` function, which allows you to replay any episode that you've recorded or episodes from any dataset out there. This function helps you test the repeatability of your robot's actions and assess transferability across robots of the same model.
You can replay the first episode on your robot with either the command below or with the API example:
<hfoptions id="replay">
<hfoption id="Command">
```bash
lerobot-replay \
--robot.type=so101_follower \
--robot.port=/dev/tty.usbmodem58760431541 \
--robot.id=my_awesome_follower_arm \
--dataset.repo_id=${HF_USER}/record-test \
--dataset.episode=0 # choose the episode you want to replay
```
</hfoption>
<hfoption id="API example">
<!-- prettier-ignore-start -->
```python
import time
from lerobot.datasets import LeRobotDataset
from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
from lerobot.utils.robot_utils import precise_sleep
from lerobot.utils.utils import log_say
episode_idx = 0
robot_config = SO100FollowerConfig(port="/dev/tty.usbmodem5AB90687491", id="my_follower_arm")
robot = SO100Follower(robot_config)
robot.connect()
dataset = LeRobotDataset("<hf_username>/<dataset_repo_id>", episodes=[episode_idx])
actions = dataset.select_columns("action")
log_say(f"Replaying episode {episode_idx}")
for idx in range(dataset.num_frames):
t0 = time.perf_counter()
action = {
name: float(actions[idx]["action"][i]) for i, name in enumerate(dataset.features["action"]["names"])
}
robot.send_action(action)
precise_sleep(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
robot.disconnect()
```
<!-- prettier-ignore-end -->
</hfoption>
</hfoptions>
Your robot should replicate movements similar to those you recorded. For example, check out [this video](https://x.com/RemiCadene/status/1793654950905680090) where we use `replay` on a Aloha robot from [Trossen Robotics](https://www.trossenrobotics.com).
## Train a policy
To train a policy to control your robot, use the [`lerobot-train`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/scripts/lerobot_train.py) script. A few arguments are required. Here is an example command:
```bash
lerobot-train \
--dataset.repo_id=${HF_USER}/so101_test \
--policy.type=act \
--output_dir=outputs/train/act_so101_test \
--job_name=act_so101_test \
--policy.device=cuda \
--wandb.enable=true \
--policy.repo_id=${HF_USER}/my_policy
```
Let's explain the command:
1. We provided the dataset as argument with `--dataset.repo_id=${HF_USER}/so101_test`.
2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor states, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
3. We provided `policy.device=cuda` since we are training on a Nvidia GPU, but you could use `policy.device=mps` to train on Apple silicon.
4. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
Training should take several hours. You will find checkpoints in `outputs/train/act_so101_test/checkpoints`.
To resume training from a checkpoint, below is an example command to resume from `last` checkpoint of the `act_so101_test` policy:
```bash
lerobot-train \
--config_path=outputs/train/act_so101_test/checkpoints/last/pretrained_model/train_config.json \
--resume=true
```
`--config_path` also accepts a **Hub repo id**: if a run pushed its checkpoints to the Hub (with `--save_checkpoint_to_hub=true`), you can resume straight from the repo — its latest checkpoint is downloaded and training continues, restoring the optimizer, scheduler, step counter and data order:
```bash
lerobot-train --config_path=${HF_USER}/my_policy --resume=true
```
If you do not want to push your model to the hub after training use `--policy.push_to_hub=false`.
Additionally you can provide extra `tags` or specify a `license` for your model or make the model repo `private` by adding this: `--policy.private=true --policy.tags=\[ppo,rl\] --policy.license=mit`
#### Train using Google Colab
If your local computer doesn't have a powerful GPU you could utilize Google Colab to train your model by following the [ACT training notebook](./notebooks#training-act).
#### Train using Hugging Face Jobs
Hugging Face jobs let's you easily select hardware and run the training in the cloud. So if you don't have a powerful GPU or you need more VRAM or just want to train a model much faster use HF Jobs! It's pay as you go and you simply pay for each second of use, you can see the pricing and additional information [here](https://huggingface.co/docs/hub/jobs).
`lerobot-train` runs locally by default. To run on a HuggingFace GPU, pass `--job.target` with a hardware flavor name:
```bash
lerobot-train \
--dataset.repo_id=${HF_USER}/so101_test \
--policy.type=act \
--policy.repo_id=${HF_USER}/my_policy \
--job.target=a10g-small
```
List available flavors and pricing with `hf jobs hardware`. The run streams its logs to your terminal; press Ctrl-C to detach (the job keeps running in the cloud). Re-attach or cancel with:
```bash
hf jobs logs <job-id>
hf jobs cancel <job-id>
```
If your dataset exists only locally (not yet on the Hub), it is automatically pushed to a **private** Hub repo so the job can download it by `repo_id` (nothing is made public). The trained model is pushed to the model repo at the end of the run. To also push every intermediate checkpoint to the Hub as it is saved (so you can monitor progress mid-run), add `--save_checkpoint_to_hub=true` — this requires a runtime image that includes this feature.
Every job (and any dataset pushed by the run) is tagged `lerobot` so it's easy to find on the Hub. Add your own with `--job.tags '["my-tag"]'`.
By default the job is capped at `2d` (48h) of wall-clock. Override it with an HF Jobs duration string, e.g. `--job.timeout=4h` to fail faster or `--job.timeout=7d` for a longer run.
> **Note:** the model repo is created up front (it holds the staged training config the job runs from). If a run fails before the model is pushed, that repo is left on the Hub so you can inspect it — it is not deleted automatically, so repeated failures can leave empty repos behind. Remove one with `hf repo delete <repo-id>`.
**Prerequisites:** run `hf auth login` before submitting. For Weights & Biases integration, run `wandb login` or set `WANDB_API_KEY` on your machine — the key is forwarded to the job automatically.
**Resuming on a job.** Adding `--job.target` to a resume command runs the resume in the cloud — the same command works locally or remotely. The checkpoint repo is the source of truth, and new checkpoints continue the lineage in the same repo:
```bash
# resume a Hub run on a job (its checkpoints are already on the Hub)
lerobot-train --config_path=${HF_USER}/my_policy --resume=true --job.target=a10g-small
# resume a LOCAL run on a job — the checkpoint is uploaded to a private Hub repo first,
# then the job resumes from it (a local-only dataset is uploaded the same way)
lerobot-train \
--config_path=outputs/train/act_so101_test/checkpoints/last/pretrained_model/train_config.json \
--resume=true \
--job.target=a10g-small
```
Job settings come from the current command, so override `--job.target`, `--job.timeout`, etc. as needed; for the resumed run to itself be resumable later, keep `--save_checkpoint_to_hub=true`.
#### Upload policy checkpoints
Once training is done, upload the latest checkpoint with:
```bash
hf upload ${HF_USER}/act_so101_test \
outputs/train/act_so101_test/checkpoints/last/pretrained_model
```
You can also upload intermediate checkpoints with:
```bash
CKPT=010000
hf upload ${HF_USER}/act_so101_test${CKPT} \
outputs/train/act_so101_test/checkpoints/${CKPT}/pretrained_model
```
## Run inference and evaluate your policy
Use `lerobot-rollout` to deploy a trained policy on your robot. You can choose different strategies depending on your needs:
The examples below load the model from `--policy.path`. To pin a specific pushed version — useful once `--save_checkpoint_to_hub=true` has committed several checkpoints — add `--policy.pretrained_revision` with a commit hash, branch, or tag. Each pushed checkpoint is tagged with its step (e.g. `--policy.pretrained_revision=010000`), so you can recover a checkpoint by step without looking up its commit sha.
<hfoptions id="eval">
<hfoption id="Base mode (no recording)">
```bash
lerobot-rollout \
--strategy.type=base \
--policy.path=${HF_USER}/my_policy \
--robot.type=so100_follower \
--robot.port=/dev/ttyACM1 \
--robot.cameras="{ up: {type: opencv, index_or_path: /dev/video10, width: 640, height: 480, fps: 30}, side: {type: intelrealsense, serial_number_or_name: 233522074606, width: 640, height: 480, fps: 30}}" \
--task="Put lego brick into the transparent box" \
--duration=60
```
</hfoption>
<hfoption id="Sentry mode (with recording)">
```bash
lerobot-rollout \
--strategy.type=sentry \
--strategy.upload_every_n_episodes=5 \
--policy.path=${HF_USER}/my_policy \
--robot.type=so100_follower \
--robot.port=/dev/ttyACM1 \
--robot.cameras="{ up: {type: opencv, index_or_path: /dev/video10, width: 640, height: 480, fps: 30}, side: {type: intelrealsense, serial_number_or_name: 233522074606, width: 640, height: 480, fps: 30}}" \
--dataset.repo_id=${HF_USER}/eval_so100 \
--dataset.single_task="Put lego brick into the transparent box" \
--duration=600
```
</hfoption>
</hfoptions>
The `--strategy.type` flag selects the execution mode:
- `base`: Autonomous rollout with no data recording (useful for quick evaluation)
- `sentry`: Continuous recording with auto-upload (useful for large-scale evaluation)
- `highlight`: Ring buffer recording with keystroke save (useful for capturing interesting events)
- `dagger`: Human-in-the-loop data collection (see [HIL Data Collection](./hil_data_collection))
- `episodic`: Episode-oriented policy recording with reset phases between episodes
All strategies support `--inference.type=rtc` for smooth execution with slow VLA models (Pi0, Pi0.5, SmolVLA).
docs/source/implement_your_own_processor.mdx
+273
View File
@@ -0,0 +1,273 @@
# Implement your own Robot Processor
In this tutorial, you'll learn how to implement your own Robot Processor.
It begins by exploring the need for a custom processor, then uses the `NormalizerProcessorStep` as the running example to explain how to implement, configure, and serialize a processor. Finally, it lists all helper processors that ship with LeRobot.
## Why would you need a custom processor?
In most cases, when reading raw data from sensors or when models output actions, you need to process this data to make it compatible with your target system. For example, a common need is normalizing data ranges to make them suitable for neural networks.
LeRobot's `NormalizerProcessorStep` handles this crucial task:
```python
# Input: raw joint positions in [0, 180] degrees
raw_action = torch.tensor([90.0, 45.0, 135.0])
# After processing: normalized to [-1, 1] range for model training
normalizer = NormalizerProcessorStep(features=features, norm_map=norm_map, stats=dataset_stats)
normalized_result = normalizer(transition)
# ...
```
Other common processing needs include:
- **Device placement**: Moving tensors between CPU/GPU and converting data types
- **Format conversion**: Transforming between different data structures
- **Batching**: Adding/removing batch dimensions for model compatibility
- **Safety constraints**: Applying limits to robot commands
```python
# Example pipeline combining multiple processors
pipeline = PolicyProcessorPipeline([
RenameObservationsProcessorStep(rename_map={}),
AddBatchDimensionProcessorStep(),
NormalizerProcessorStep(features=features, stats=stats),
DeviceProcessorStep(device="cuda"),
# ...
])
```
LeRobot provides a pipeline mechanism to implement sequences of processing steps for both input data and output actions, making it easy to compose these transformations in the right order for optimal performance.
## How to implement your own processor?
We'll use the `NormalizerProcessorStep` as our main example because it demonstrates essential processor patterns including state management, configuration serialization, and tensor handling that you'll commonly need.
Prepare the sequence of processing steps necessary for your problem. A processor step is a class that implements the following methods:
- `__call__`: implements the processing step for the input transition.
- `get_config`: gets the configuration of the processor step.
- `state_dict`: gets the state of the processor step.
- `load_state_dict`: loads the state of the processor step.
- `reset`: resets the state of the processor step.
- `feature_contract`: displays the modification to the feature space during the processor step.
### Implement the `__call__` method
The `__call__` method is the core of your processor step. It takes an `EnvTransition` and returns a modified `EnvTransition`. Here's how the `NormalizerProcessorStep` works:
```python
@dataclass
@ProcessorStepRegistry.register("normalizer_processor")
class NormalizerProcessorStep(ProcessorStep):
"""Normalize observations/actions using dataset statistics."""
features: dict[str, PolicyFeature]
norm_map: dict[FeatureType, NormalizationMode]
stats: dict[str, dict[str, Any]] | None = None
eps: float = 1e-8
_tensor_stats: dict = field(default_factory=dict, init=False, repr=False)
def __post_init__(self):
"""Convert stats to tensors for efficient computation."""
self.stats = self.stats or {}
self._tensor_stats = to_tensor(self.stats, device=self.device, dtype=torch.float32)
def __call__(self, transition: EnvTransition) -> EnvTransition:
new_transition = transition.copy()
# Normalize observations
# ...
# Normalize action
# ...
return new_transition
```
See the full implementation in `src/lerobot/processor/normalize_processor.py` for complete details.
**Key principles:**
- **Always use `transition.copy()`** to avoid side effects
- **Handle both observations and actions** consistently
- **Separate config from state**: `get_config()` returns JSON-serializable params, `state_dict()` returns tensors
- **Convert stats to tensors** in `__post_init__()` for efficient computation
### Configuration and State Management
Processors support serialization through three methods that separate configuration from tensor state. The `NormalizerProcessorStep` demonstrates this perfectly - it carries dataset statistics (tensors) in its state, and hyperparameters in its config:
```python
# Continuing the NormalizerProcessorStep example...
def get_config(self) -> dict[str, Any]:
"""JSON-serializable configuration (no tensors)."""
return {
"eps": self.eps,
"features": {k: {"type": v.type.value, "shape": v.shape} for k, v in self.features.items()},
"norm_map": {ft.value: nm.value for ft, nm in self.norm_map.items()},
# ...
}
def state_dict(self) -> dict[str, torch.Tensor]:
"""Tensor state only (e.g., dataset statistics)."""
flat: dict[str, torch.Tensor] = {}
for key, sub in self._tensor_stats.items():
for stat_name, tensor in sub.items():
flat[f"{key}.{stat_name}"] = tensor.cpu() # Always save to CPU
return flat
def load_state_dict(self, state: dict[str, torch.Tensor]) -> None:
"""Restore tensor state at runtime."""
self._tensor_stats.clear()
for flat_key, tensor in state.items():
key, stat_name = flat_key.rsplit(".", 1)
# Load to processor's configured device
self._tensor_stats.setdefault(key, {})[stat_name] = tensor.to(
dtype=torch.float32, device=self.device
)
# ...
```
**Usage:**
```python
# Save (e.g., inside a policy)
config = normalizer.get_config()
tensors = normalizer.state_dict()
# Restore (e.g., loading a pretrained policy)
new_normalizer = NormalizerProcessorStep(**config)
new_normalizer.load_state_dict(tensors)
# Now new_normalizer has the same stats and configuration
```
### Transform features
The `transform_features` method defines how your processor transforms feature names and shapes. This is crucial for policy configuration and debugging.
For `NormalizerProcessorStep`, features are typically preserved unchanged since normalization doesn't alter keys or shapes:
```python
def transform_features(self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
"""Normalization preserves all feature definitions."""
return features # No changes to feature structure
# ...
```
When your processor renames or reshapes data, implement this method to reflect the mapping for downstream components. For example, a simple rename processor:
```python
def transform_features(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
# Simple renaming
if "pixels" in features:
features["observation.image"] = features.pop("pixels")
# Pattern-based renaming
for key in list(features.keys()):
if key.startswith("env_state."):
suffix = key[len("env_state."):]
features[f"observation.{suffix}"] = features.pop(key)
# ...
return features
```
**Key principles:**
- Use `features.pop(old_key)` to remove and get the old feature
- Use `features[new_key] = old_feature` to add the renamed feature
- Always return the modified features dictionary
- Document transformations clearly in the docstring
### Using overrides
You can override step parameters at load-time using `overrides`. This is handy for non-serializable objects or site-specific settings. It works both in policy factories and with `DataProcessorPipeline.from_pretrained(...)`.
**Foundational model adaptation**: This is particularly useful when working with foundational pretrained policies where you rarely have access to the original training statistics. You can inject your own dataset statistics to adapt the normalizer to your specific robot or environment data.
Example: during policy evaluation on the robot, override the device and rename map.
Use this to run a policy trained on CUDA on a CPU-only robot, or to remap camera keys when the robot uses different names than the dataset.
Direct usage with `from_pretrained`:
```python
from lerobot.processor import RobotProcessorPipeline
# Load a foundational policy trained on diverse robot data
# but adapt normalization to your specific robot/environment
new_stats = LeRobotDataset(repo_id="username/my-dataset").meta.stats
processor = RobotProcessorPipeline.from_pretrained(
"huggingface/foundational-robot-policy", # Pretrained foundation model
overrides={
"normalizer_processor": {"stats": new_stats}, # Inject your robot's statistics
"device_processor": {"device": "cuda:0"}, # registry name for registered steps
"rename_processor": {"rename_map": robot_key_map}, # Map your robot's observation keys
# ...
},
)
```
## Best Practices
Based on analysis of all LeRobot processor implementations, here are the key patterns and practices:
### 1. **Safe Data Handling**
Always create copies of input data to avoid unintended side effects. Use `transition.copy()` and `observation.copy()` rather than modifying data in-place. This prevents your processor from accidentally affecting other components in the pipeline.
Check for required data before processing and handle missing data gracefully. If your processor expects certain keys (like `"pixels"` for image processing), validate their presence first. For optional data, use safe access patterns like `transition.get()` and handle `None` values appropriately.
When data validation fails, provide clear, actionable error messages that help users understand what went wrong and how to fix it.
### 2. **Choose Appropriate Base Classes**
LeRobot provides specialized base classes that reduce boilerplate code and ensure consistency. Use `ObservationProcessorStep` when you only need to modify observations, `ActionProcessorStep` for action-only processing, and `RobotActionProcessorStep` specifically for dictionary-based robot actions.
Only inherit directly from `ProcessorStep` when you need full control over the entire transition or when processing multiple transition components simultaneously. The specialized base classes handle the transition management for you and provide type safety.
### 3. **Registration and Naming**
Register your processors with descriptive, namespaced names using `@ProcessorStepRegistry.register()`. Use organization prefixes like `"robotics_lab/safety_clipper"` or `"acme_corp/vision_enhancer"` to avoid naming conflicts. Avoid generic names like `"processor"` or `"step"` that could clash with other implementations.
Good registration makes your processors discoverable and enables clean serialization/deserialization when saving and loading pipelines.
### 4. **State Management Patterns**
Distinguish between configuration parameters (JSON-serializable values) and internal state (tensors, buffers). Use dataclass fields with `init=False, repr=False` for internal state that shouldn't appear in the constructor or string representation.
Implement the `reset()` method to clear internal state between episodes. This is crucial for stateful processors that accumulate data over time, like moving averages or temporal filters.
Remember that `get_config()` should only return JSON-serializable configuration, while `state_dict()` handles tensor state separately.
### 5. **Input Validation and Error Handling**
Validate input types and shapes before processing. Check tensor properties like `dtype` and dimensions to ensure compatibility with your algorithms. For robot actions, verify that required pose components or joint values are present and within expected ranges.
Use early returns for edge cases where no processing is needed. Provide clear, descriptive error messages that include the expected vs. actual data types or shapes. This makes debugging much easier for users.
### 6. **Device and Dtype Awareness**
Design your processors to automatically adapt to the device and dtype of input tensors. Internal tensors (like normalization statistics) should match the input tensor's device and dtype to ensure compatibility with multi-GPU training, mixed precision, and distributed setups.
Implement a `to()` method that moves your processor's internal state to the specified device. Check device/dtype compatibility at runtime and automatically migrate internal state when needed. This pattern enables seamless operation across different hardware configurations without manual intervention.
## Conclusion
You now have all the tools to implement custom processors in LeRobot! The key steps are:
1. **Define your processor** as a dataclass with the required methods (`__call__`, `get_config`, `state_dict`, `load_state_dict`, `reset`, `transform_features`)
2. **Register it** using `@ProcessorStepRegistry.register("name")` for discoverability
3. **Integrate it** into a `DataProcessorPipeline` with other processing steps
4. **Use base classes** like `ObservationProcessorStep` when possible to reduce boilerplate
5. **Implement device/dtype awareness** to support multi-GPU and mixed precision setups
The processor system is designed to be modular and composable, allowing you to build complex data processing pipelines from simple, focused components. Whether you're preprocessing sensor data for training or post-processing model outputs for robot execution, custom processors give you the flexibility to handle any data transformation your robotics application requires.
Key principles for robust processors:
- **Device/dtype adaptation**: Internal tensors should match input tensors
- **Clear error messages**: Help users understand what went wrong
- **Base class usage**: Leverage specialized base classes to reduce boilerplate
- **Feature contracts**: Declare data structure changes with `transform_features()`
Start simple, test thoroughly, and ensure your processors work seamlessly across different hardware configurations!
docs/source/index.mdx
+23
View File
@@ -0,0 +1,23 @@
<div class="flex justify-center">
<a target="_blank" href="https://huggingface.co/lerobot">
<img
alt="HuggingFace Expert Acceleration Program"
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/lerobot-logo-thumbnail.png"
style="width: 100%"
></img>
</a>
</div>
# LeRobot
**State-of-the-art machine learning for real-world robotics**
🤗 LeRobot aims to provide models, datasets, and tools for real-world robotics in PyTorch. The goal is to lower the barrier for entry to robotics so that everyone can contribute and benefit from sharing datasets and pretrained models.
🤗 LeRobot contains state-of-the-art approaches that have been shown to transfer to the real-world with a focus on imitation learning and reinforcement learning.
🤗 LeRobot already provides a set of pretrained models, datasets with human collected demonstrations, and simulated environments so that everyone can get started.
🤗 LeRobot hosts pretrained models and datasets on the LeRobot HuggingFace page.
Join the LeRobot community on [Discord](https://discord.gg/s3KuuzsPFb)
docs/source/inference.mdx
+299
View File
@@ -0,0 +1,299 @@
# Policy Deployment (lerobot-rollout)
`lerobot-rollout` is the single CLI for deploying trained policies on real robots. It supports multiple execution strategies and inference backends, from quick evaluation to continuous recording and human-in-the-loop data collection.
## Quick Start
No extra dependencies are needed beyond your robot and policy extras.
```bash
lerobot-rollout \
--strategy.type=base \
--policy.path=lerobot/act_koch_real \
--robot.type=koch_follower \
--robot.port=/dev/ttyACM0 \
--task="pick up cube" \
--duration=30
```
This runs the policy for 30 seconds with no recording.
---
## Strategies
Select a strategy with `--strategy.type=<name>`. Each strategy defines a different control loop with its own recording and interaction semantics.
### Base (`--strategy.type=base`)
Autonomous policy execution with no data recording. Use this for quick evaluation, demos, or when you only need to observe the robot.
```bash
lerobot-rollout \
--strategy.type=base \
--policy.path=${HF_USER}/my_policy \
--robot.type=so100_follower \
--robot.port=/dev/ttyACM0 \
--robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
--task="Put lego brick into the box" \
--duration=60
```
| Flag | Description |
| ---------------- | ------------------------------------------------------ |
| `--duration` | Run time in seconds (0 = infinite) |
| `--task` | Task description passed to the policy |
| `--display_data` | Stream observations/actions to Rerun for visualization |
### Sentry (`--strategy.type=sentry`)
Continuous autonomous recording with periodic upload to the Hugging Face Hub. Episode boundaries are auto-computed from camera resolution and FPS so each saved episode produces a complete video file, keeping uploads efficient.
Policy state (hidden state, RTC queue) persists across episode boundaries: the robot does not reset between episodes.
```bash
lerobot-rollout \
--strategy.type=sentry \
--strategy.upload_every_n_episodes=5 \
--policy.path=${HF_USER}/my_policy \
--robot.type=so100_follower \
--robot.port=/dev/ttyACM0 \
--robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
--dataset.repo_id=${HF_USER}/rollout_eval_data \
--dataset.single_task="Put lego brick into the box" \
--duration=3600
```
| Flag | Description |
| -------------------------------------- | ----------------------------------------------------------- |
| `--strategy.upload_every_n_episodes` | Push to Hub every N episodes (default: 5) |
| `--strategy.target_video_file_size_mb` | Target video file size for episode rotation (default: auto) |
| `--dataset.repo_id` | **Required.** Hub repository for the recorded dataset |
| `--dataset.push_to_hub` | Whether to push to Hub on teardown (default: true) |
### Highlight (`--strategy.type=highlight`)
Autonomous rollout with on-demand recording via a memory-bounded ring buffer. The robot runs continuously while the buffer captures the last N seconds of telemetry. Press the save key to flush the buffer and start live recording; press it again to save the episode.
```bash
lerobot-rollout \
--strategy.type=highlight \
--strategy.ring_buffer_seconds=30 \
--strategy.save_key=s \
--strategy.push_key=h \
--policy.path=${HF_USER}/my_policy \
--robot.type=koch_follower \
--robot.port=/dev/ttyACM0 \
--dataset.repo_id=${HF_USER}/rollout_highlight_data \
--dataset.single_task="Pick up the red cube"
```
**Keyboard controls:**
| Key | Action |
| ------------------ | -------------------------------------------------------- |
| `s` (configurable) | Start recording (flushes buffer) / stop and save episode |
| `h` (configurable) | Push dataset to Hub |
| `ESC` | Stop the session |
| Flag | Description |
| -------------------------------------- | ---------------------------------------------- |
| `--strategy.ring_buffer_seconds` | Duration of buffered telemetry (default: 30) |
| `--strategy.ring_buffer_max_memory_mb` | Memory cap for the ring buffer (default: 2048) |
| `--strategy.save_key` | Key to toggle recording (default: `s`) |
| `--strategy.push_key` | Key to push to Hub (default: `h`) |
### DAgger (`--strategy.type=dagger`)
Human-in-the-loop data collection. Alternates between autonomous policy execution and human intervention via a teleoperator. Intervention frames are tagged with `intervention=True`. Requires a teleoperator (`--teleop.type`).
See the [Human-In-the-Loop Data Collection](./hil_data_collection) guide for a detailed walkthrough.
**Corrections-only mode** (default): Only human correction windows are recorded. Each correction becomes one episode.
```bash
lerobot-rollout \
--strategy.type=dagger \
--strategy.num_episodes=20 \
--policy.path=outputs/pretrain/checkpoints/last/pretrained_model \
--robot.type=bi_openarm_follower \
--teleop.type=bi_openarm_mini \
--dataset.repo_id=${HF_USER}/rollout_hil_data \
--dataset.single_task="Fold the T-shirt"
```
**Continuous recording mode** (`--strategy.record_autonomous=true`): Both autonomous and correction frames are recorded with time-based episode rotation (same as Sentry).
```bash
lerobot-rollout \
--strategy.type=dagger \
--strategy.record_autonomous=true \
--strategy.num_episodes=50 \
--policy.path=${HF_USER}/my_policy \
--robot.type=so100_follower \
--robot.port=/dev/ttyACM0 \
--teleop.type=so101_leader \
--teleop.port=/dev/ttyACM1 \
--dataset.repo_id=${HF_USER}/rollout_dagger_data \
--dataset.single_task="Grasp the block"
```
**Keyboard controls** (default input device):
| Key | Action |
| ------- | ------------------------------------------- |
| `Space` | Pause / resume policy execution |
| `Tab` | Start / stop human correction |
| `Enter` | Push dataset to Hub (corrections-only mode) |
| `ESC` | Stop the session |
Foot pedal input is also supported via `--strategy.input_device=pedal`. Configure pedal codes with `--strategy.pedal.*` flags.
| Flag | Description |
| ------------------------------------ | ------------------------------------------------------- |
| `--strategy.num_episodes` | Number of correction episodes to record (default: 10) |
| `--strategy.record_autonomous` | Record autonomous frames too (default: false) |
| `--strategy.upload_every_n_episodes` | Push to Hub every N episodes (default: 5) |
| `--strategy.input_device` | Input device: `keyboard` or `pedal` (default: keyboard) |
| `--teleop.type` | **Required.** Teleoperator type |
### Episodic (`--strategy.type=episodic`)
Episode-oriented recording that mirrors the behavior of `lerobot-record`. The policy drives the robot for each episode; an optional teleoperator can drive the robot during the reset phase between episodes.
```bash
lerobot-rollout \
--strategy.type=episodic \
--policy.path=${HF_USER}/my_policy \
--robot.type=so100_follower \
--robot.port=/dev/ttyACM0 \
--teleop.type=so100_leader \
--teleop.port=/dev/ttyACM1 \
--dataset.repo_id=${HF_USER}/my_eval_data \
--dataset.num_episodes=20 \
--dataset.episode_time_s=30 \
--dataset.reset_time_s=10 \
--dataset.single_task="Pick up the red cube"
```
Teleop is optional — if omitted the robot holds its position during the reset phase.
**Keyboard controls:**
| Key | Action |
| ----------- | -------------------------------- |
| `→` (right) | End the current episode early |
| `←` (left) | Discard episode and re-record it |
| `ESC` | Stop the recording session |
| Flag | Description |
| ----------------------------------------------- | -------------------------------------------------------------------------- |
| `--dataset.num_episodes` | Number of episodes to record |
| `--dataset.episode_time_s` | Duration of each recording episode in seconds |
| `--dataset.reset_time_s` | Duration of the reset phase between episodes in seconds |
| `--teleop.type` | Optional. Teleoperator to drive the robot during resets |
| `--strategy.reset_to_initial_position` | Whether to reset the robot to its initial position between episodes |
| `--strategy.smooth_leader_to_follower_handover` | Whether to turn on or off the leader -> follower smooth handover behavior. |
---
## Inference Backends
Select a backend with `--inference.type=<name>`. All strategies work with both backends.
### Sync (default)
One policy call per control tick. The main loop blocks until the action is computed.
Works with all policies. No extra flags needed.
### Real-Time Chunking (`--inference.type=rtc`)
A background thread produces action chunks asynchronously. The main control loop polls for the next ready action while the policy computes the next chunk in parallel.
Use RTC with large, slow VLA models (Pi0, Pi0.5, SmolVLA) for smooth, continuous motion despite high inference latency.
```bash
lerobot-rollout \
--strategy.type=base \
--inference.type=rtc \
--inference.rtc.execution_horizon=10 \
--inference.rtc.max_guidance_weight=10.0 \
--policy.path=${HF_USER}/pi0_policy \
--robot.type=so100_follower \
--robot.port=/dev/ttyACM0 \
--robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
--task="Pick up the cube" \
--duration=60 \
--device=cuda
```
| Flag | Description |
| ------------------------------------------- | -------------------------------------------------------------- |
| `--inference.rtc.execution_horizon` | Steps to blend with previous chunk (default: varies by policy) |
| `--inference.rtc.max_guidance_weight` | Consistency enforcement strength (default: varies by policy) |
| `--inference.rtc.prefix_attention_schedule` | Blend schedule: `LINEAR`, `EXP`, `ONES`, `ZEROS` |
| `--inference.queue_threshold` | Max queue size before backpressure (default: 30) |
See the [Real-Time Chunking](./rtc) guide for details on tuning RTC parameters.
---
## Common Flags
| Flag | Description | Default |
| --------------------------------- | ----------------------------------------------------------------- | ------- |
| `--policy.path` | **Required.** HF Hub model ID or local checkpoint path | -- |
| `--robot.type` | **Required.** Robot type (e.g. `so100_follower`, `koch_follower`) | -- |
| `--robot.port` | Serial port for the robot | -- |
| `--robot.cameras` | Camera configuration (JSON dict) | -- |
| `--fps` | Control loop frequency | 30 |
| `--duration` | Run time in seconds (0 = infinite) | 0 |
| `--device` | Torch device (`cpu`, `cuda`, `mps`) | auto |
| `--task` | Task description (used when no dataset is provided) | -- |
| `--display_data` | Stream telemetry to Rerun visualization | false |
| `--display_ip` / `--display_port` | Remote Rerun server address | -- |
| `--interpolation_multiplier` | Action interpolation factor | 1 |
| `--use_torch_compile` | Enable `torch.compile` for inference | false |
| `--resume` | Resume a previous recording session | false |
| `--play_sounds` | Vocal synthesis for events | true |
---
## Programmatic Usage
For custom deployments (e.g. with kinematics processors), use the rollout module API directly:
```python
from lerobot.rollout import BaseStrategyConfig, RolloutConfig, build_rollout_context
from lerobot.rollout.inference import SyncInferenceConfig
from lerobot.rollout.strategies import BaseStrategy
from lerobot.utils.process import ProcessSignalHandler
cfg = RolloutConfig(
robot=my_robot_config,
policy=my_policy_config,
strategy=BaseStrategyConfig(),
inference=SyncInferenceConfig(),
fps=30,
duration=60,
task="my task",
)
signal_handler = ProcessSignalHandler(use_threads=True)
ctx = build_rollout_context(
cfg,
signal_handler.shutdown_event,
robot_action_processor=my_custom_action_processor, # optional
robot_observation_processor=my_custom_obs_processor, # optional
)
strategy = BaseStrategy(cfg.strategy)
try:
strategy.setup(ctx)
strategy.run(ctx)
finally:
strategy.teardown(ctx)
```
See `examples/so100_to_so100_EE/rollout.py` and `examples/phone_to_so100/rollout.py` for full examples with kinematics processors.
docs/source/installation.mdx
+300
View File
@@ -0,0 +1,300 @@
# Installation
This guide uses `conda` (via miniforge) to manage environments (recommended). If you prefer another environment manager (e.g. `uv`, `venv`), ensure you have Python >=3.12 and support PyTorch >= 2.10, then skip ahead to [Environment Setup](#step-2-environment-setup).
## Step 1 (`conda` only): Install [`miniforge`](https://conda-forge.org/download/)
```bash
wget "https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-$(uname)-$(uname -m).sh"
bash Miniforge3-$(uname)-$(uname -m).sh
```
## Step 2: Environment Setup
Create a virtual environment with Python 3.12:
<!-- prettier-ignore-start -->
<hfoptions id="create_venv">
<hfoption id="conda">
```bash
conda create -y -n lerobot python=3.12
```
</hfoption>
<hfoption id="uv (PyTorch >= 2.10 only)">
```bash
uv python install 3.12
uv venv --python 3.12
```
</hfoption>
</hfoptions>
<!-- prettier-ignore-end -->
Then activate your virtual environment, you have to do this each time you open a shell to use lerobot:
<!-- prettier-ignore-start -->
<hfoptions id="activate_venv">
<hfoption id="conda">
```bash
conda activate lerobot
```
> [!NOTE]
> When installing LeRobot inside WSL (Windows Subsystem for Linux), make sure to also install `evdev`:
>
> ```bash
> conda install evdev -c conda-forge
> ```
</hfoption>
<hfoption id="uv (PyTorch >= 2.10 only)">
```bash
# Linux/macOS
source .venv/bin/activate
# Windows PowerShell
.venv\Scripts\activate
```
> [!NOTE]
> When installing LeRobot inside WSL (Windows Subsystem for Linux), make sure to also install `evdev`:
>
> ```bash
> sudo apt install libevdev-dev
> uv pip install evdev
> ```
</hfoption>
</hfoptions>
<!-- prettier-ignore-end -->
### Install `ffmpeg` (for video decoding)
LeRobot uses [TorchCodec](https://github.com/meta-pytorch/torchcodec) for video decoding by default, which requires `ffmpeg`.
> [!NOTE]
> **Platform support:** TorchCodec is **not available** on macOS Intel (x86_64), Linux ARM (aarch64, arm64, armv7l), or Windows with PyTorch < 2.8. On these platforms, LeRobot automatically falls back to `pyav` — so you do not need to install `ffmpeg` and can skip to Step 3.
If your platform supports TorchCodec, install `ffmpeg` using one of the methods below:
<!-- prettier-ignore-start -->
<hfoptions id="install_ffmpeg">
<hfoption id="conda (any PyTorch version)">
Install `ffmpeg` in your conda environment. This works with **all PyTorch versions** and is **required for PyTorch < 2.10**:
```bash
conda install ffmpeg -c conda-forge
```
> [!TIP]
> This usually installs `ffmpeg 8.X` with the `libsvtav1` encoder. If you run into issues (e.g. `libsvtav1` missing — check with `ffmpeg -encoders` — or a version mismatch with `torchcodec`), you can explicitly install `ffmpeg 7.1.1` using:
>
> ```bash
> conda install ffmpeg=7.1.1 -c conda-forge
> ```
</hfoption>
<hfoption id="uv (PyTorch >= 2.10 only)">
Starting with **PyTorch >= 2.10** (TorchCodec ≥ 0.10), TorchCodec can dynamically link to a system-wide `ffmpeg` installation. This is useful when using `uv` or other non-`conda` environment managers:
```bash
# Ubuntu/Debian
sudo apt install ffmpeg
# macOS (Apple Silicon)
brew install ffmpeg
```
> [!IMPORTANT]
> System-wide `ffmpeg` is **only supported with PyTorch >= 2.10** (TorchCodec ≥ 0.10). For older PyTorch versions, you **must** use `conda install ffmpeg -c conda-forge` instead.
</hfoption>
</hfoptions>
<!-- prettier-ignore-end -->
## Step 3: Install LeRobot 🤗
The base `lerobot` install is intentionally **lightweight** — it includes only core ML dependencies (PyTorch, torchvision, numpy, opencv, einops, draccus, huggingface-hub, gymnasium, safetensors). Heavier dependencies are gated behind optional extras so you only install what you need.
### From Source
First, clone the repository and navigate into the directory:
```bash
git clone https://github.com/huggingface/lerobot.git
cd lerobot
```
Then, install the library in editable mode. This is useful if you plan to contribute to the code.
<!-- prettier-ignore-start -->
<hfoptions id="install_lerobot_src">
<hfoption id="conda">
```bash
pip install -e ".[core_scripts]" # For robot workflows (recording, replaying, calibrate)
pip install -e ".[training]" # For training policies
pip install -e ".[all]" # Everything (all policies, envs, hardware, dev tools)
```
</hfoption>
<hfoption id="uv">
```bash
uv pip install -e ".[core_scripts]" # For robot workflows (recording, replaying, calibrate)
uv pip install -e ".[training]" # For training policies
uv pip install -e ".[all]" # Everything (all policies, envs, hardware, dev tools)
```
</hfoption>
</hfoptions>
<!-- prettier-ignore-end -->
### Installation from PyPI
**Core Library:**
Install the base package with:
<!-- prettier-ignore-start -->
<hfoptions id="install_lerobot_pypi">
<hfoption id="conda">
```bash
pip install lerobot
```
</hfoption>
<hfoption id="uv">
```bash
uv pip install lerobot
```
</hfoption>
</hfoptions>
<!-- prettier-ignore-end -->
_This installs only the core ML dependencies. You will need to add extras for most workflows._
**Feature Extras:**
LeRobot provides **feature-scoped extras** that map to common workflows. If you are using `uv`, replace `pip install` with `uv pip install` in the commands below.
| Extra | What it adds | Typical use case |
| ---------- | ------------------------------------------- | ----------------------------------- |
| `dataset` | `datasets`, `av`, `torchcodec`, `jsonlines` | Loading & creating datasets |
| `training` | `dataset` + `accelerate`, `wandb` | Training policies |
| `hardware` | `pynput`, `pyserial`, `deepdiff` | Connecting to real robots |
| `viz` | `rerun-sdk` | Visualization during recording/eval |
**Composite Extras** combine feature extras for common CLI scripts:
| Extra | Includes | Typical use case |
| -------------- | ------------------------------ | ------------------------------------------------------- |
| `core_scripts` | `dataset` + `hardware` + `viz` | `lerobot-record`, `lerobot-replay`, `lerobot-calibrate` |
| `evaluation` | `av` | `lerobot-eval` (add policy + env extras as needed) |
| `dataset_viz` | `dataset` + `viz` | `lerobot-dataset-viz`, `lerobot-imgtransform-viz` |
```bash
pip install 'lerobot[core_scripts]' # Record, replay, calibrate
pip install 'lerobot[training]' # Train policies
pip install 'lerobot[core_scripts,training]' # Record + train
pip install 'lerobot[all]' # Everything
```
**Policy, environment, and hardware extras** are still available for specific dependencies:
```bash
pip install 'lerobot[pi]' # Pi0/Pi0.5/Pi0-FAST policy deps
pip install 'lerobot[smolvla]' # SmolVLA policy deps
pip install 'lerobot[diffusion]' # Diffusion policy deps (diffusers)
pip install 'lerobot[aloha,pusht]' # Simulation environments
pip install 'lerobot[feetech]' # Feetech motor support
```
_Multiple extras can be combined (e.g., `.[core_scripts,pi,pusht]`). For a full list of available extras, refer to `pyproject.toml`._
### PyTorch CUDA variant (Linux only)
On Linux, the install path determines which CUDA wheel you get. macOS and Windows installs use the PyPI default (MPS / CPU / CUDA-Windows wheel respectively) and can skip this section.
<!-- prettier-ignore-start -->
<hfoptions id="cuda_variant">
<hfoption id="uv-source">
**Source install via `uv` (`uv sync` or `uv pip install -e .`)**
`torch` and `torchvision` are pinned by the project to the **CUDA 12.8** PyTorch index (`https://download.pytorch.org/whl/cu128`, driver floor **570.86**) — covers Ampere/Ada/Hopper/Blackwell GPUs. No action needed for typical NVIDIA setups.
To override for a different CUDA variant:
```bash
uv pip install --force-reinstall torch torchvision \
--index-url https://download.pytorch.org/whl/cu126 # older drivers; or cu130 for Blackwell on driver ≥ 580
```
</hfoption>
<hfoption id="pip-conda">
**Source install via `pip`/`conda`, or `pip install lerobot` from PyPI**
PyPI default torch wheel is currently a cu130-bundled Linux wheel, driver floor **580.65**.
To pick a specific CUDA variant:
**Using `pip` or `conda`** — install torch first with an explicit index, then lerobot:
```bash
pip install --index-url https://download.pytorch.org/whl/cu128 torch torchvision
pip install -e ".[all]" # source
# — or —
pip install lerobot # from PyPI
```
**Using `uv` to install from PyPI** — one-liner via `--torch-backend` (uv ≥ 0.6):
```bash
uv pip install --torch-backend cu128 lerobot
```
Supported values include `auto`, `cpu`, `cu126`, `cu128`, `cu129`, `cu130`, plus various `rocm*` and `xpu`. Swap as needed for your driver.
</hfoption>
</hfoptions>
<!-- prettier-ignore-end -->
### Troubleshooting
If you encounter build errors, you may need to install additional system dependencies: `cmake`, `build-essential`, and `ffmpeg libs`.
To install these for Linux run:
```bash
sudo apt-get install cmake build-essential python3-dev pkg-config libavformat-dev libavcodec-dev libavdevice-dev libavutil-dev libswscale-dev libswresample-dev libavfilter-dev
```
For other systems, see: [Compiling PyAV](https://pyav.org/docs/develop/overview/installation.html#bring-your-own-ffmpeg)
## Optional dependencies
LeRobot provides optional extras for specific functionalities. Multiple extras can be combined (e.g., `.[aloha,feetech]`). For all available extras, refer to `pyproject.toml`. If you are using `uv`, replace `pip install` with `uv pip install` in the commands below.
### Simulations
Install environment packages: `aloha` ([gym-aloha](https://github.com/huggingface/gym-aloha)), or `pusht` ([gym-pusht](https://github.com/huggingface/gym-pusht)).
These automatically include the `dataset` extra.
```bash
pip install -e ".[aloha]" # or "[pusht]" for example
```
### Motor Control
For Koch v1.1 install the Dynamixel SDK, for SO100/SO101/Moss install the Feetech SDK.
```bash
pip install -e ".[feetech]" # or "[dynamixel]" for example
```
### Experiment Tracking
Weights and Biases is included in the `training` extra. To use [Weights and Biases](https://docs.wandb.ai/quickstart) for experiment tracking, log in with:
```bash
wandb login
```
You can now assemble your robot if it's not ready yet, look for your robot type on the left. Then follow the link below to use Lerobot with your robot.
docs/source/integrate_hardware.mdx
+476
View File
@@ -0,0 +1,476 @@
# Bring Your Own Hardware
This tutorial will explain how to integrate your own robot design into the LeRobot ecosystem and have it access all of our tools (data collection, control pipelines, policy training and inference).
To that end, we provide the [`Robot`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/robots/robot.py) base class in the LeRobot which specifies a standard interface for physical robot integration. Let's see how to implement it.
## Prerequisites
- Your own robot which exposes a communication interface (e.g. serial, CAN, TCP)
- A way to read sensor data and send motor commands programmatically, e.g. manufacturer's SDK or API, or your own protocol implementation.
- LeRobot installed in your environment. Follow our [Installation Guide](./installation).
## Choose your motors
If you're using Feetech or Dynamixel motors, LeRobot provides built-in bus interfaces:
- [`FeetechMotorsBus`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/motors/feetech/feetech.py) – for controlling Feetech servos
- [`DynamixelMotorsBus`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/motors/dynamixel/dynamixel.py) – for controlling Dynamixel servos
Please refer to the [`MotorsBus`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/motors/motors_bus.py) abstract class to learn about its API.
For a good example of how it can be used, you can have a look at our own [SO101 follower implementation](https://github.com/huggingface/lerobot/blob/main/src/lerobot/robots/so_follower/so101_follower/so101_follower.py)
Use these if compatible. Otherwise, you'll need to find or write a Python interface (not covered in this tutorial):
- Find an existing SDK in Python (or use bindings to C/C++)
- Or implement a basic communication wrapper (e.g., via pyserial, socket, or CANopen)
You're not alone—many community contributions use custom boards or firmware!
For Feetech and Dynamixel, we currently support these servos: - Feetech: - STS & SMS series (protocol 0): `sts3215`, `sts3250`, `sm8512bl` - SCS series (protocol 1): `scs0009` - Dynamixel (protocol 2.0 only): `xl330-m077`, `xl330-m288`, `xl430-w250`, `xm430-w350`, `xm540-w270`, `xc430-w150`
If you are using Feetech or Dynamixel servos that are not in this list, you can add those in the [Feetech table](https://github.com/huggingface/lerobot/blob/main/src/lerobot/motors/feetech/tables.py) or [Dynamixel table](https://github.com/huggingface/lerobot/blob/main/src/lerobot/motors/dynamixel/tables.py). Depending on the model, this will require you to add model-specific information. In most cases though, there shouldn't be a lot of additions to do.
In the next sections, we'll use a `FeetechMotorsBus` as the motors interface for the examples. Replace it and adapt to your motors if necessary.
## Step 1: Subclass the `Robot` Interface
You’ll first need to specify the config class and a string identifier (`name`) for your robot. If your robot has special needs that you'd like to be able to change easily, it should go here (e.g. port/address, baudrate).
Here, we'll add the port name and one camera by default for our robot:
<!-- prettier-ignore-start -->
```python
from dataclasses import dataclass, field
from lerobot.cameras import CameraConfig
from lerobot.cameras.opencv import OpenCVCameraConfig
from lerobot.robots import RobotConfig
@RobotConfig.register_subclass("my_cool_robot")
@dataclass
class MyCoolRobotConfig(RobotConfig):
port: str
cameras: dict[str, CameraConfig] = field(
default_factory={
"cam_1": OpenCVCameraConfig(
index_or_path=2,
fps=30,
width=480,
height=640,
),
}
)
```
<!-- prettier-ignore-end -->
[Cameras tutorial](./cameras) to understand how to detect and add your camera.
Next, we'll create our actual robot class which inherits from `Robot`. This abstract class defines a contract you must follow for your robot to be usable with the rest of the LeRobot tools.
Here we'll create a simple 5-DoF robot with one camera. It could be a simple arm but notice that the `Robot` abstract class does not assume anything on your robot's form factor. You can let you imagination run wild when designing new robots!
<!-- prettier-ignore-start -->
```python
from lerobot.cameras import make_cameras_from_configs
from lerobot.motors import Motor, MotorNormMode
from lerobot.motors.feetech import FeetechMotorsBus
from lerobot.robots import Robot
class MyCoolRobot(Robot):
config_class = MyCoolRobotConfig
name = "my_cool_robot"
def __init__(self, config: MyCoolRobotConfig):
super().__init__(config)
self.bus = FeetechMotorsBus(
port=self.config.port,
motors={
"joint_1": Motor(1, "sts3250", MotorNormMode.RANGE_M100_100),
"joint_2": Motor(2, "sts3215", MotorNormMode.RANGE_M100_100),
"joint_3": Motor(3, "sts3215", MotorNormMode.RANGE_M100_100),
"joint_4": Motor(4, "sts3215", MotorNormMode.RANGE_M100_100),
"joint_5": Motor(5, "sts3215", MotorNormMode.RANGE_M100_100),
},
calibration=self.calibration,
)
self.cameras = make_cameras_from_configs(config.cameras)
```
<!-- prettier-ignore-end -->
## Step 2: Define Observation and Action Features
These two properties define the _interface contract_ between your robot and tools that consume it (such as data collection or learning pipelines).
> [!WARNING]
> Note that these properties must be callable even if the robot is not yet connected, so avoid relying on runtime hardware state to define them.
### `observation_features`
This property should return a dictionary describing the structure of sensor outputs from your robot. The keys match what `get_observation()` returns, and the values describe either the shape (for arrays/images) or the type (for simple values).
Example for our 5-DoF arm with one camera:
<!-- prettier-ignore-start -->
```python
@property
def _motors_ft(self) -> dict[str, type]:
return {
"joint_1.pos": float,
"joint_2.pos": float,
"joint_3.pos": float,
"joint_4.pos": float,
"joint_5.pos": float,
}
@property
def _cameras_ft(self) -> dict[str, tuple]:
return {
cam: (self.cameras[cam].height, self.cameras[cam].width, 3) for cam in self.cameras
}
@property
def observation_features(self) -> dict:
return {**self._motors_ft, **self._cameras_ft}
```
<!-- prettier-ignore-end -->
In this case, observations consist of a simple dict storing each motor's position and a camera image.
### `action_features`
This property describes the commands your robot expects via `send_action()`. Again, keys must match the expected input format, and values define the shape/type of each command.
Here, we simply use the same joints proprioceptive features (`self._motors_ft`) as with `observation_features`: the action sent will simply the goal position for each motor.
<!-- prettier-ignore-start -->
```python
def action_features(self) -> dict:
return self._motors_ft
```
<!-- prettier-ignore-end -->
## Step 3: Handle Connection and Disconnection
These methods should handle opening and closing communication with your hardware (e.g. serial ports, CAN interfaces, USB devices, cameras).
### `is_connected`
This property should simply reflect that communication with the robot's hardware is established. When this property is `True`, it should be possible to read and write to the hardware using `get_observation()` and `send_action()`.
<!-- prettier-ignore-start -->
```python
@property
def is_connected(self) -> bool:
return self.bus.is_connected and all(cam.is_connected for cam in self.cameras.values())
```
<!-- prettier-ignore-end -->
### `connect()`
This method should establish communication with the hardware. Moreover, if your robot needs calibration and is not calibrated, it should start a calibration procedure by default. If your robot needs some specific configuration, this should also be called here.
<!-- prettier-ignore-start -->
```python
def connect(self, calibrate: bool = True) -> None:
self.bus.connect()
if not self.is_calibrated and calibrate:
self.calibrate()
for cam in self.cameras.values():
cam.connect()
self.configure()
```
<!-- prettier-ignore-end -->
### `disconnect()`
This method should gracefully terminate communication with the hardware: free any related resources (threads or processes), close ports, etc.
Here, we already handle this in our `MotorsBus` and `Camera` classes so we just need to call their own `disconnect()` methods:
<!-- prettier-ignore-start -->
```python
def disconnect(self) -> None:
self.bus.disconnect()
for cam in self.cameras.values():
cam.disconnect()
```
<!-- prettier-ignore-end -->
## Step 4: Support Calibration and Configuration
LeRobot supports saving and loading calibration data automatically. This is useful for joint offsets, zero positions, or sensor alignment.
> Note that depending on your hardware, this may not apply. If that's the case, you can simply leave these methods as no-ops:
<!-- prettier-ignore-start -->
```python
@property
def is_calibrated(self) -> bool:
return True
def calibrate(self) -> None:
pass
```
<!-- prettier-ignore-end -->
### `is_calibrated`
This should reflect whether your robot has the required calibration loaded.
<!-- prettier-ignore-start -->
```python
@property
def is_calibrated(self) -> bool:
return self.bus.is_calibrated
```
<!-- prettier-ignore-end -->
### `calibrate()`
The goal of the calibration is twofold:
- Know the physical range of motion of each motors in order to only send commands within this range.
- Normalize raw motors positions to sensible continuous values (e.g. percentages, degrees) instead of arbitrary discrete value dependant on the specific motor used that will not replicate elsewhere.
It should implement the logic for calibration (if relevant) and update the `self.calibration` dictionary. If you are using Feetech or Dynamixel motors, our bus interfaces already include methods to help with this.
<!-- prettier-ignore-start -->
```python
def calibrate(self) -> None:
self.bus.disable_torque()
for motor in self.bus.motors:
self.bus.write("Operating_Mode", motor, OperatingMode.POSITION.value)
input(f"Move {self} to the middle of its range of motion and press ENTER....")
homing_offsets = self.bus.set_half_turn_homings()
print(
"Move all joints sequentially through their entire ranges "
"of motion.\nRecording positions. Press ENTER to stop..."
)
range_mins, range_maxes = self.bus.record_ranges_of_motion()
self.calibration = {}
for motor, m in self.bus.motors.items():
self.calibration[motor] = MotorCalibration(
id=m.id,
drive_mode=0,
homing_offset=homing_offsets[motor],
range_min=range_mins[motor],
range_max=range_maxes[motor],
)
self.bus.write_calibration(self.calibration)
self._save_calibration()
print("Calibration saved to", self.calibration_fpath)
```
<!-- prettier-ignore-end -->
### `configure()`
Use this to set up any configuration for your hardware (servos control modes, controller gains, etc.). This should usually be run at connection time and be idempotent.
<!-- prettier-ignore-start -->
```python
def configure(self) -> None:
with self.bus.torque_disabled():
self.bus.configure_motors()
for motor in self.bus.motors:
self.bus.write("Operating_Mode", motor, OperatingMode.POSITION.value)
self.bus.write("P_Coefficient", motor, 16)
self.bus.write("I_Coefficient", motor, 0)
self.bus.write("D_Coefficient", motor, 32)
```
<!-- prettier-ignore-end -->
## Step 5: Implement Sensors Reading and Action Sending
These are the most important runtime functions: the core I/O loop.
### `get_observation()`
Returns a dictionary of sensor values from the robot. These typically include motor states, camera frames, various sensors, etc. In the LeRobot framework, these observations are what will be fed to a policy in order to predict the actions to take. The dictionary keys and structure must match `observation_features`.
<!-- prettier-ignore-start -->
```python
def get_observation(self) -> dict[str, Any]:
if not self.is_connected:
raise ConnectionError(f"{self} is not connected.")
# Read arm position
obs_dict = self.bus.sync_read("Present_Position")
obs_dict = {f"{motor}.pos": val for motor, val in obs_dict.items()}
# Capture images from cameras
for cam_key, cam in self.cameras.items():
obs_dict[cam_key] = cam.async_read()
return obs_dict
```
<!-- prettier-ignore-end -->
### `send_action()`
Takes a dictionary that matches `action_features`, and sends it to your hardware. You can add safety limits (clipping, smoothing) and return what was actually sent.
For simplicity, we won't be adding any modification of the actions in our example here.
<!-- prettier-ignore-start -->
```python
def send_action(self, action: dict[str, Any]) -> dict[str, Any]:
goal_pos = {key.removesuffix(".pos"): val for key, val in action.items()}
# Send goal position to the arm
self.bus.sync_write("Goal_Position", goal_pos)
return action
```
<!-- prettier-ignore-end -->
## Adding a Teleoperator
For implementing teleoperation devices, we also provide a [`Teleoperator`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/teleoperators/teleoperator.py) base class. This class is very similar to the `Robot` base class and also doesn't assume anything on form factor.
The main differences are in the I/O functions: a teleoperator allows you to produce action via `get_action` and can receive feedback actions via `send_feedback`. Feedback could be anything controllable on the teleoperation device that could help the person controlling it understand the consequences of the actions sent. Think motion/force feedback on a leader arm, vibrations on a gamepad controller for example. To implement a teleoperator, you can follow this same tutorial and adapt it for these two methods.
## Using Your Own `LeRobot` Devices 🔌
You can easily extend `lerobot` with your own custom hardware—be it a camera, robot, or teleoperation device—by creating a separate, installable Python package. If you follow a few simple conventions, the `lerobot` command-line tools (like `lerobot-teleop` and `lerobot-record`) will **automatically discover and integrate your creations** without requiring any changes to the `lerobot` source code.
This guide outlines the conventions your plugin must follow.
### The 4 Core Conventions
To ensure your custom device is discoverable, you must adhere to the following four rules.
#### 1\. Create an Installable Package with a Specific Prefix
Your project must be a standard, installable Python package. Crucially, the name of your package (as defined in `pyproject.toml` or `setup.py`) must begin with one of these prefixes:
- `lerobot_robot_` for a robot.
- `lerobot_camera_` for a camera.
- `lerobot_teleoperator_` for a teleoperation device.
This prefix system is how `lerobot` automatically finds your plugin in the Python environment.
#### 2\. Follow the `SomethingConfig`/`Something` Naming Pattern
Your device's implementation class must be named after its configuration class, simply by removing the `Config` suffix.
- **Config Class:** `MyAwesomeTeleopConfig`
- **Device Class:** `MyAwesomeTeleop`
#### 3\. Place Your Files in a Predictable Structure
The device class (`MyAwesomeTeleop`) must be located in a predictable module relative to its configuration class (`MyAwesomeTeleopConfig`). `lerobot` will automatically search in these locations:
- In the **same module** as the config class.
- In a **submodule named after the device** (e.g., `my_awesome_teleop.py`).
The recommended and simplest structure is to place them in separate, clearly named files within the same directory.
#### 4\. Expose Classes in `__init__.py`
Your package's `__init__.py` file should import and expose both the configuration and the device classes, making them easily accessible.
### Putting It All Together: A Complete Example
Let's create a new teleoperator called `my_awesome_teleop`.
#### Directory Structure
Here is what the project folder should look like. The package name, `lerobot_teleoperator_my_awesome_teleop`, follows **Convention \#1**.
```
lerobot_teleoperator_my_awesome_teleop/
├── pyproject.toml # (or setup.py) lists lerobot as a dependency
└── lerobot_teleoperator_my_awesome_teleop/
├── __init__.py
├── config_my_awesome_teleop.py
└── my_awesome_teleop.py
```
#### File Contents
- **`config_my_awesome_teleop.py`**: Defines the configuration class. Note the `Config` suffix (**Convention \#2**).
```python
from dataclasses import dataclass
from lerobot.teleoperators.config import TeleoperatorConfig
@TeleoperatorConfig.register_subclass("my_awesome_teleop")
@dataclass
class MyAwesomeTeleopConfig(TeleoperatorConfig):
# Your configuration fields go here
port: str = "192.168.1.1"
```
- **`my_awesome_teleop.py`**: Implements the device. The class name `MyAwesomeTeleop` matches its config class name (**Convention \#2**). This file structure adheres to **Convention \#3**.
```python
from lerobot.teleoperators.teleoperator import Teleoperator
from .config_my_awesome_teleop import MyAwesomeTeleopConfig
class MyAwesomeTeleop(Teleoperator):
config_class = MyAwesomeTeleopConfig
name = "my_awesome_teleop"
def __init__(self, config: MyAwesomeTeleopConfig):
super().__init__(config)
self.config = config
# Your device logic (e.g., connect) goes here
```
- **`__init__.py`**: Exposes the key classes (**Convention \#4**).
```python
from .config_my_awesome_teleop import MyAwesomeTeleopConfig
from .my_awesome_teleop import MyAwesomeTeleop
```
### Installation and Usage
1. **Install your new plugin in your Python environment.** You can install your local plugin package using `pip`'s editable mode or from PyPi.
```bash
# Locally
# Navigate to your plugin's root directory and install it
cd lerobot_teleoperator_my_awesome_teleop
pip install -e .
# From PyPi
pip install lerobot_teleoperator_my_awesome_teleop
```
2. **Use it directly from the command line.** Now, you can use your custom device by referencing its type.
```bash
lerobot-teleoperate --teleop.type=my_awesome_teleop \
# other arguments
```
And that's it\! Your custom device is now fully integrated.
### Looking for an example ?
Check out these two packages from the community:
- https://github.com/SpesRobotics/lerobot-robot-xarm
- https://github.com/SpesRobotics/lerobot-teleoperator-teleop
## Wrapping Up
Once your robot class is complete, you can leverage the LeRobot ecosystem:
- Control your robot with available teleoperators or integrate directly your teleoperating device
- Record training data and visualize it
- Integrate it into RL or imitation learning pipelines
Don't hesitate to reach out to the community for help on our [Discord](https://discord.gg/s3KuuzsPFb) 🤗
docs/source/introduction_processors.mdx
+314
View File
@@ -0,0 +1,314 @@
# Introduction to Processors
In robotics, there's a fundamental mismatch between the data that robots and humans produce and what machine learning models expect.
Robots output raw sensor data like camera images and joint positions that need normalization, batching, and device placement before models can process them.
Language instructions from humans must be tokenized into numerical representations, and different robots use different coordinate systems that need standardization.
The challenge extends to model outputs as well.
Models might output end-effector positions while robots need joint-space commands, or teleoperators produce relative movements while robots expect absolute commands.
Model predictions are often normalized and need conversion back to real-world scales.
Cross-domain translation adds another layer of complexity.
Training data from one robot setup needs adaptation for deployment on different hardware, models trained with specific camera configurations must work with new arrangements, and datasets with different naming conventions need harmonization.
**That's where processors come in.** They serve as universal translators that bridge these gaps, ensuring seamless data flow from sensors to models to actuators.
Processors handle all the preprocessing and postprocessing steps needed to convert raw environment data into model-ready inputs and vice versa.
This means that your favorite policy can be used like this:
```python
import torch
from lerobot.datasets import LeRobotDataset
from lerobot.policies import make_pre_post_processors
from lerobot.policies.your_policy import YourPolicy
from lerobot.processor import RobotProcessorPipeline, PolicyProcessorPipeline
dataset = LeRobotDataset("hf_user/dataset", episodes=[0])
sample = dataset[10]
model = YourPolicy.from_pretrained(
"hf_user/model",
)
model.eval()
model.to("cuda")
preprocessor, postprocessor = make_pre_post_processors(model.config, pretrained_path="hf_user/model", dataset_stats=dataset.meta.stats)
preprocessed_sample = preprocessor(sample)
action = model.select_action(preprocessed_sample)
postprocessed_action = postprocessor(action)
```
## What are Processors?
In robotics, data comes in many forms: images from cameras, joint positions from sensors, text instructions from users, and more. Each type of data requires specific transformations before a model can use it effectively. Models need this data to be:
- **Normalized**: Scaled to appropriate ranges for neural network processing
- **Batched**: Organized with proper dimensions for batch processing
- **Tokenized**: Text converted to numerical representations
- **Device-placed**: Moved to the right hardware (CPU/GPU)
- **Type-converted**: Cast to appropriate data types
Processors handle these transformations through composable, reusable steps that can be chained together into pipelines. Think of them as a modular assembly line where each station performs a specific transformation on your data.
## Core Concepts
### EnvTransition: The Universal Data Container
The `EnvTransition` is the fundamental data structure that flows through all processors.
It's a typed dictionary that represents a complete robot-environment interaction:
- **OBSERVATION**: All sensor data (images, states, proprioception)
- **ACTION**: The action to execute or that was executed
- **REWARD**: Reinforcement learning signal
- **DONE/TRUNCATED**: Episode boundary indicators
- **INFO**: Arbitrary metadata
- **COMPLEMENTARY_DATA**: Task descriptions, indices, padding flags, inter-step data
### ProcessorStep: The Building Block
A `ProcessorStep` is a single transformation unit that processes transitions. It's an abstract base class with two required methods:
```python
from lerobot.processor import ProcessorStep, EnvTransition
class MyProcessorStep(ProcessorStep):
"""Example processor step - inherit and implement abstract methods."""
def __call__(self, transition: EnvTransition) -> EnvTransition:
"""Transform the transition - REQUIRED abstract method."""
# Your processing logic here
return transition
def transform_features(self, features):
"""Declare how this step transforms feature shapes/types - REQUIRED abstract method."""
return features # Most processors return features unchanged
```
`__call__` is the core of your processor step. It takes an `EnvTransition` and returns a modified `EnvTransition`.
`transform_features` is used to declare how this step transforms feature shapes/types.
### DataProcessorPipeline: The Generic Orchestrator
The `DataProcessorPipeline[TInput, TOutput]` chains multiple `ProcessorStep` instances together:
```python
from lerobot.processor import RobotProcessorPipeline, PolicyProcessorPipeline
# For robot hardware (unbatched data)
robot_processor = RobotProcessorPipeline[RobotAction, RobotAction](
steps=[step1, step2, step3],
name="robot_pipeline"
)
# For model training/inference (batched data)
policy_processor = PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
steps=[step1, step2, step3],
name="policy_pipeline"
)
```
## RobotProcessorPipeline vs PolicyProcessorPipeline
The key distinction is in the data structures they handle:
| Aspect | RobotProcessorPipeline | PolicyProcessorPipeline |
| --------------- | -------------------------------------------- | ---------------------------------------- |
| **Input** | `dict[str, Any]` - Individual robot values | `dict[str, Any]` - Batched tensors |
| **Output** | `dict[str, Any]` - Individual robot commands | `torch.Tensor` - Policy predictions |
| **Use Case** | Real-time robot control | Model training/inference |
| **Data Format** | Unbatched, heterogeneous | Batched, homogeneous |
| **Examples** | `{"joint_1": 0.5}` | `{"observation.state": tensor([[0.5]])}` |
**Use `RobotProcessorPipeline`** for robot hardware interfaces:
```python
# Robot data structures: dict[str, Any] for observations and actions
robot_obs: dict[str, Any] = {
"joint_1": 0.5, # Individual joint values
"joint_2": -0.3,
"camera_0": image_array # Raw camera data
}
robot_action: dict[str, Any] = {
"joint_1": 0.2, # Target joint positions
"joint_2": 0.1,
"gripper": 0.8
}
```
**Use `PolicyProcessorPipeline`** for model training and batch processing:
```python
# Policy data structures: batch dicts and tensors
policy_batch: dict[str, Any] = {
"observation.state": torch.tensor([[0.5, -0.3]]), # Batched states
"observation.images.camera0": torch.tensor(...), # Batched images
"action": torch.tensor([[0.2, 0.1, 0.8]]) # Batched actions
}
policy_action: torch.Tensor = torch.tensor([[0.2, 0.1, 0.8]]) # Model output tensor
```
## Converter Functions
LeRobot provides converter functions to bridge different data formats in `lerobot.processor.converters`. These functions handle the crucial translations between robot hardware data structures, policy model formats, and the internal `EnvTransition` representation that flows through processor pipelines.
| Category | Function | Description |
| ------------------------------ | ----------------------------- | ------------------------------- |
| **Robot Hardware Converters** | `robot_action_to_transition` | Robot dict → EnvTransition |
| | `observation_to_transition` | Robot obs → EnvTransition |
| | `transition_to_robot_action` | EnvTransition → Robot dict |
| **Policy/Training Converters** | `batch_to_transition` | Batch dict → EnvTransition |
| | `transition_to_batch` | EnvTransition → Batch dict |
| | `policy_action_to_transition` | Policy tensor → EnvTransition |
| | `transition_to_policy_action` | EnvTransition → Policy tensor |
| **Utilities** | `create_transition` | Build transitions with defaults |
| | `identity_transition` | Pass-through converter |
The key insight is that **robot hardware converters** work with individual values and dictionaries, while **policy/training converters** work with batched tensors and model outputs. The converter functions automatically handle the structural differences, so your processor steps can focus on the core transformations without worrying about data format compatibility.
## Processor Examples
The following examples demonstrate real-world processor configurations for policy training and inference.
Here is an example processor for policy training and inference:
```python
# Training data preprocessing (optimized order for GPU performance)
training_preprocessor = PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
steps=[
RenameObservationsProcessorStep(rename_map={}), # Standardize keys
AddBatchDimensionProcessorStep(), # Add batch dims
TokenizerProcessorStep(tokenizer_name="...", ...), # Tokenize language
DeviceProcessorStep(device="cuda"), # Move to GPU first
NormalizerProcessorStep(features=..., stats=...), # Normalize on GPU
]
)
# Model output postprocessing
training_postprocessor = PolicyProcessorPipeline[torch.Tensor, torch.Tensor](
steps=[
DeviceProcessorStep(device="cpu"), # Move to CPU
UnnormalizerProcessorStep(features=..., stats=...), # Denormalize
]
to_transition=policy_action_to_transition,
to_output=transition_to_policy_action,
)
```
### An interaction between a robot and a policy with processors
The most common real-world scenario combines both pipeline types robot hardware generates observations that need policy processing, and policy outputs need robot-compatible postprocessing:
```python
# Real deployment: Robot sensors → Model → Robot commands
with torch.no_grad():
while not done:
raw_obs = robot.get_observation() # dict[str, Any]
# Add your robot observation to policy observation processor
policy_input = policy_preprocessor(raw_obs) # Batched dict
policy_output = policy.select_action(policy_input) # Policy tensor
policy_action = policy_postprocessor(policy_output)
# Add your robot action to policy action processor
robot.send_action(policy_action)
```
## Feature Contracts: Shape and Type Transformation
Processors don't just transform data - they can also **change the data structure itself**. The `transform_features()` method declares these changes, which is crucial for dataset recording and policy creation.
### Why Feature Contracts Matter
When building datasets or policies, LeRobot needs to know:
- **What data fields will exist** after processing
- **What shapes and types** each field will have
- **How to configure models** for the expected data structure
```python
# Example: A processor that adds velocity to observations
class VelocityProcessor(ObservationProcessorStep):
def observation(self, obs):
new_obs = obs.copy()
if "observation.state" in obs:
# concatenate computed velocity field to the state
new_obs["observation.state"] = self._compute_velocity(obs["observation.state"])
return new_obs
def transform_features(self, features):
"""Declare the new velocity field we're adding."""
state_feature = features[PipelineFeatureType.OBSERVATION].get("observation.state")
if state_feature:
double_shape = (state_feature.shape[0] * 2,) if state_feature.shape else (2,)
features[PipelineFeatureType.OBSERVATION]["observation.state"] = PolicyFeature(
type=FeatureType.STATE, shape=double_shape
)
return features
```
### Feature Specification Functions
`create_initial_features()` and `aggregate_pipeline_dataset_features()` solve a critical dataset creation problem: determining the exact final data structure before any data is processed.
Since processor pipelines can add new features (like velocity fields), change tensor shapes (like cropping images), or rename keys, datasets need to know the complete output specification upfront to allocate proper storage and define schemas.
These functions work together by starting with robot hardware specifications (`create_initial_features()`) then simulating the entire pipeline transformation (`aggregate_pipeline_dataset_features()`) to compute the final feature dictionary that gets passed to `LeRobotDataset.create()`, ensuring perfect alignment between what processors output and what datasets expect to store.
```python
from lerobot.datasets import aggregate_pipeline_dataset_features
# Start with robot's raw features
initial_features = create_initial_features(
observation=robot.observation_features, # {"joint_1.pos": float, "camera_0": (480,640,3)}
action=robot.action_features # {"joint_1.pos": float, "gripper.pos": float}
)
# Apply processor pipeline to compute final features
final_features = aggregate_pipeline_dataset_features(
pipeline=my_processor_pipeline,
initial_features=initial_features,
use_videos=True
)
# Use for dataset creation
dataset = LeRobotDataset.create(
repo_id="my_dataset",
features=final_features, # Knows exactly what data to expect
...
)
```
## Common Processor Steps
LeRobot provides many registered processor steps. Here are the most commonly used core processors:
### Essential Processors
- **`normalizer_processor`**: Normalize observations/actions using dataset statistics (mean/std or min/max)
- **`device_processor`**: Move tensors to CPU/GPU with optional dtype conversion
- **`to_batch_processor`**: Add batch dimensions to transitions for model compatibility
- **`rename_observations_processor`**: Rename observation keys using mapping dictionaries
- **`tokenizer_processor`**: Tokenize natural language task descriptions into tokens and attention masks
### Next Steps
- **[Implement Your Own Processor](./implement_your_own_processor)** - Create custom processor steps
- **[Debug Your Pipeline](./debug_processor_pipeline)** - Troubleshoot and optimize pipelines
- **[Processors for Robots and Teleoperators](./processors_robots_teleop)** - Real-world integration patterns
## Summary
Processors solve the data translation problem in robotics by providing:
- **Modular transformations**: Composable, reusable processing steps
- **Type safety**: Generic pipelines with compile-time checking
- **Performance optimization**: GPU-accelerated operations
- **Robot/Policy distinction**: Separate pipelines for different data structures
- **Comprehensive ecosystem**: 30+ registered processors for common tasks
The key insight: `RobotProcessorPipeline` handles unbatched robot hardware data, while `PolicyProcessorPipeline` handles batched model data. Choose the right tool for your data structure!
docs/source/isaac_teleop.mdx
+397
View File
@@ -0,0 +1,397 @@
# Isaac Teleop
Control your robot with NVIDIA [Isaac Teleop](https://github.com/NVIDIA/IsaacTeleop), a
multi-modal teleoperation framework. Isaac Teleop drives a single `TeleopSession` from a range
of input devices — XR (VR) controllers, hand tracking, full-body tracking, Manus gloves, foot
pedals, and more.
In LeRobot, Isaac Teleop ships as a self-contained example under
[`examples/isaac_teleop_to_so101/`](https://github.com/huggingface/lerobot/tree/main/examples/isaac_teleop_to_so101).
Each Isaac Teleop input device is its own `Teleoperator` subclass in the example's
`isaac_teleop` package, sharing one session lifecycle (see `IsaacTeleopTeleoperator`). The
devices available today are the **XR controller** (`XRController`) and a back-drivable
**SO-101 leader arm** (`SO101LeaderArm`); Manus gloves and hand/full-body tracking are the
natural next devices. This guide focuses on the XR controller; the SO-101 leader is summarized
under [Run the example](#step-3-run-the-example).
**In this guide you'll learn:**
- How an Isaac Teleop device drives a robot end‑effector (EE) target
- How the _clutch_ (squeeze/grip on the XR controller) engages teleoperation without jerking the arm
- How to run the SO‑101 teleoperation example and tune motion / gripper / IK
## Installation
The example lives in the LeRobot repository (it is not part of the `lerobot` pip package), so
clone the repo and install from source. The canonical, always-up-to-date install and usage
reference is the example's
[`README.md`](https://github.com/huggingface/lerobot/tree/main/examples/isaac_teleop_to_so101/README.md);
in short:
```bash
git clone https://github.com/huggingface/lerobot.git
cd lerobot
uv pip install -e ".[feetech,kinematics,dataset]" "huggingface_hub>=1.5"
uv pip install "isaacteleop[cloudxr,retargeters-lite]~=1.3.131" "scipy>=1.14"
```
`isaacteleop` is published on public PyPI (Linux only). The `cloudxr` extra brings the CloudXR
runtime bindings; `retargeters-lite` is the scipy-based retargeter path that resolves on both
x86_64 and ARM (on aarch64 — e.g. a DGX Spark — the full `retargeters` extra does not resolve
because of its `dex-retargeting`/`nlopt` pins, which is why it is not the default here). On
x86_64 you can additionally install the full retargeter stack:
```bash
uv pip install "isaacteleop[retargeters]~=1.3.131"
```
### Set up CloudXR and connect a headset
Isaac Teleop streams the headset to your machine over **NVIDIA CloudXR**, which provides the
OpenXR runtime the session connects to. By default LeTeleop **auto-launches the CloudXR runtime
for you** when you call `teleop_device.connect()` — you no longer have to run `python -m
isaacteleop.cloudxr` and `source cloudxr.env` in a separate shell. All you need is a supported
headset connected and the CloudXR firewall ports open. Follow the Isaac Teleop
[Quick Start](https://nvidia.github.io/IsaacTeleop/main/getting_started/quick_start.html) for the
headset-pairing and firewall details.
**First run (EULA).** The very first launch must accept the NVIDIA CloudXR EULA. The auto-launch
prompts for it **on stdin**, so on a headless machine it will hang waiting for input. Bootstrap
the EULA once, interactively, with:
```bash
python -m isaacteleop.cloudxr --accept-eula # one-time: accept the CloudXR EULA
```
After that, `connect()` launches the runtime non-interactively. The launch **blocks for ~30s**
while the runtime comes up.
**Configuration.** Two fields on `IsaacTeleopConfig` (shared by every device) control this:
- `auto_launch_cloudxr` (default `True`) — whether `connect()` starts the runtime. Set `False`
when CloudXR is already running externally.
- `cloudxr_env_file` (default `None`) — an optional CloudXR device-profile `.env` selecting the
headset transport (e.g. an Apple Vision Pro profile). This is launcher **input**; it is not the
`~/.cloudxr/run/cloudxr.env` **output** file the old manual flow told you to `source`. `None`
keeps the default auto-WebRTC profile — though the SO-101 example overrides it to the
`default.env` shipped next to `teleoperate.py` unless you pass `--teleop.cloudxr_env_file`.
**Opting out.** To skip the auto-launch (CloudXR already running), either set
`auto_launch_cloudxr=False` or export:
```bash
export LEROBOT_CLOUDXR_SKIP_AUTOLAUNCH=1
```
The **env var takes precedence over the config field**: if `LEROBOT_CLOUDXR_SKIP_AUTOLAUNCH=1` is
set, the auto-launch is skipped even when `auto_launch_cloudxr=True`. This variable is
**independent** of Isaac Lab's `ISAACLAB_CXR_SKIP_AUTOLAUNCH` — setting one does not affect the
other.
**One teleoperator per process.** The CloudXR runtime configures the environment process-wide (a
singleton), so run a single Isaac Teleop teleoperator per process.
**Shutting down.** Always call `teleop_device.disconnect()` on exit — including on Ctrl-C. Wrap
your teleoperation loop in `try/finally` and call `disconnect()` in the `finally`. This tears down
the OpenXR session **before** the CloudXR runtime, which is the required order; the launcher's
`atexit` hook only reaps the runtime and does not run the session's `__exit__`, so without an
explicit `disconnect()` an interrupted run shuts down in the wrong order.
```python
teleop_device.connect()
try:
while True:
action = teleop_device.get_action()
# ... drive the robot ...
finally:
teleop_device.disconnect()
```
See [System Requirements](https://nvidia.github.io/IsaacTeleop/main/references/requirements.html)
for supported OS / GPU / CloudXR versions and headsets.
## How it works
The XR controller is one Isaac Teleop **input** device. `XRController` is a deliberately thin
reader: it exposes the **raw** controller grip pose — already statically rebased into the robot
base frame — plus the squeeze and trigger analog values. It has **no** retargeters and **no**
clutch logic of its own. The clutch (engage latch + delta rebasing onto the EE) and the gripper
mapping live downstream in the example loop, which then feeds LeRobot's existing closed‑loop
Cartesian IK pipeline — the same one the phone teleoperator uses. The device‑specific pieces are
`XRController`, the loop's `Clutch`, and `MapXRControllerActionToRobotAction`; everything downstream
(`EEBoundsAndSafety`, `InverseKinematicsEEToJoints`) is shared, and a future device (e.g. Manus
gloves) would swap in its own `teleop_<device>.py` + processor while reusing the rest.
`XRController._build_pipeline` wires Isaac Teleop's `ControllersSource` — statically rebased into
the robot base frame by the native `ControllerTransform` (`base_T_anchor`) — and exposes the
transformed controller stream verbatim. `get_action()` reads the grip pose, squeeze, and trigger
straight off it; the session is always stepped `RUNNING` (there is no clutch retargeter to gate).
The `Clutch` class (in `examples/isaac_teleop_to_so101/isaac_teleop/clutch.py`, driven by the
loop in `common.py`) mirrors Isaac Teleop's `SO101ClutchRetargeter`, but lives in-loop so the
device can stay a thin reader:
- It latches its engage origin on the squeeze **engage edge** (the frame the squeeze first crosses
`clutch_threshold`) and rebases both position and orientation around it, so engaging does not
teleport the arm. `Clutch.rebase` returns the absolute base-frame target as a `(pos, quat)`
pair, which the loop concatenates into the 7D `ee_pose` fed to the processor.
- The analog trigger becomes a gripper `closedness` in `[0, 1]` (0 = open, 1 = closed),
proportional to the trigger pull, which `MapXRControllerActionToRobotAction` maps to a jaw target.
See the Isaac Teleop
[Retargeting interface](https://nvidia.github.io/IsaacTeleop/main/references/retargeting/index.html)
and [architecture overview](https://nvidia.github.io/IsaacTeleop/main/overview/architecture.html)
for how source nodes and retargeters compose.
```text
VR controller (OpenXR)
│
▼
XRController.get_action() ── raw base-frame grip_pos / grip_quat + squeeze + trigger
│ (TeleopSession always stepped RUNNING; clutch lives downstream)
▼
Clutch.rebase(grip_pos, grip_quat) ── engage-relative delta applied to the EE home (pos + orient)
│ ee_pose (7) / closedness → absolute ee_pose; closedness = trigger
▼
MapXRControllerActionToRobotAction ── absolute ee.x/y/z; ee.w* = orientation rotvec target;
│ ee.x/y/z / ee.w* / ee.gripper_pos ee.gripper_pos = (1 - closedness) * 100
▼
EEBoundsAndSafety ── workspace clip + per-frame step clamp (clamp+warn)
│
▼
InverseKinematicsEEToJoints ── closed-loop Placo IK; position + soft-orientation
│ (orientation_weight=0.01) (passes ee.gripper_pos → gripper.pos)
▼
SO-101 follower joint targets
```
### The clutch: owned by the example loop
Unlike the phone pipeline (which splits the clutch across `MapPhoneActionToRobotAction` and
`EEReferenceAndDelta`), the XR clutch lives entirely in the example loop's `Clutch` class. It emits
an **absolute** EE pose, so there is no `EEReferenceAndDelta` stage and no delta accumulation in the
processor — `MapXRControllerActionToRobotAction` is a pure, stateless per‑frame mapping.
The clutch latches its engage origin on the squeeze **engage edge** (the moment the squeeze crosses
`clutch_threshold`) and drives the EE from the motion _relative_ to that origin, so the arm does not
teleport on engage. On **every** engage — startup and mid‑task re‑clutch alike — the home
_position_ is latched from forward kinematics on the arm's **measured joints**, so the home equals
where the arm physically is even if it moved while disengaged, and the engage is jump‑free. The
home _orientation_ keeps the last commanded rotation: the 5‑DOF arm tracks orientation only
softly, so latching the measured wrist orientation would inject its tracking offset into the
command on every re‑clutch.
## Controls
- **Squeeze / grip** — the **clutch** (deadman). Hold it past `clutch_threshold` to engage
teleoperation; release to pause. Each engage re‑captures the origin, so you can reposition
your hand while paused and re‑engage without the arm jumping (index/clutch style).
- **Trigger** — the **gripper**, controlled **analog**. The jaw tracks the trigger
proportionally — a half‑pressed trigger leaves the jaw half‑closed — via a closedness in
`[0, 1]` (0 = open, 1 = closed) that maps to an absolute gripper joint target.
- **Controller orientation** — the **wrist**. The clutch rebases the controller orientation
(engage‑relative, base‑frame) into a soft IK orientation target the wrist tracks alongside
position. On the 5‑DOF SO‑101 the wrist follows the hand only partially by design — see
`orientation_weight` below.
## Get started
### Step 1: Create the teleoperator
```python
# Run from the repo root so the `examples` package is importable.
from examples.isaac_teleop_to_so101.isaac_teleop import XRController, XRControllerConfig
teleop_config = XRControllerConfig(
hand_side="right", # "left" or "right" controller
clutch_threshold=0.5, # squeeze value above which the clutch engages
)
teleop_device = XRController(teleop_config)
```
`XRController.get_action()` returns the **raw** base‑frame controller pose, not a clutch‑rebased
target: `grip_pos` (3,) `[x, y, z]` [m] and `grip_quat` (4,) `[qx, qy, qz, qw]` in the robot base
frame, plus scalar `squeeze` and `trigger` analog values in `[0, 1]`. The example loop's `Clutch`
turns these into the absolute `ee_pose`, and the squeeze is thresholded by the loop against
`clutch_threshold` to engage.
### Step 2: Connect
Calling `teleop_device.connect()` first auto-launches the CloudXR runtime (unless you opted out —
see [Set up CloudXR and connect a headset](#set-up-cloudxr-and-connect-a-headset); this blocks for
~30s and on the first run prompts for the EULA on stdin), then starts the Isaac Teleop
[`TeleopSession`](https://nvidia.github.io/IsaacTeleop/main/getting_started/teleop_session.html)
(opens the OpenXR session and discovers the controllers). XR controllers are self‑calibrating, so
there is no manual calibration step — the clutch handles re‑centering each time you engage. Pair
`connect()` with a `try/finally` that calls `disconnect()` so the session tears down before the
runtime on exit/Ctrl-C.
### Step 3: Run the example
The example assumes you configured your robot (SO‑101 follower) and set the correct serial port.
The **robot URDF and its meshes are fetched automatically** on first run: the XR device downloads
the SO-101 URDF from the
[`lerobot/robot-urdfs` Hugging Face bucket](https://huggingface.co/buckets/lerobot/robot-urdfs/tree/so101)
into the LeRobot cache (`HF_LEROBOT_HOME/robot-urdfs/so101/`) and reuses it after, so there is no
separate download step :
```bash
python -m examples.isaac_teleop_to_so101.teleoperate --robot.type=so101_follower --robot.port=/dev/ttyACM0 \
--robot.id=so101_follower_arm --teleop.type=xr_controller
```
The CLI is `lerobot-teleoperate`-style (draccus): `--robot.*` configures the SO-101 follower and
`--teleop.type` selects the Isaac input device (`xr_controller` | `so101_leader`), with
`--teleop.*` its device knobs. `--teleop.type=xr_controller` runs the XR-controller path described
above. The startup safety contract: by default it slews all joints to a default reset pose over
`--reset_duration` seconds (`--reset_to_origin=false` keeps the arm where it is), then seeds the
clutch home from the arm's measured pose so the first engage is jump-free; the follower is
commanded only while the clutch is engaged.
**Customizing the reset pose.** The reset pose ships as a built-in default (a comfortable mid-range
pose) and works out of the box — you do **not** need to record anything. To tailor it to your setup,
back-drive the arm to the pose you want and run
`python -m examples.isaac_teleop_to_so101.override_reset_pose --id <robot.id>`; it writes the
current joints to a per-arm file in the LeRobot cache
(`HF_LEROBOT_HOME/reset_poses/<robot.name>/<robot.id>.json`, keyed like calibration), which then takes
priority over the built-in default on the next run. Because it lives in the user-local cache (not
the repo), your override stays on your machine, and both `teleoperate` and `record` honor it
when launched with the same `--robot.id`.
The other device, `--teleop.type=so101_leader`, mirrors the follower 1:1 from a back-drivable
SO-101 _leader arm_ whose joints are streamed by Isaac Teleop's native `so101_leader` plugin (no
clutch, no IK — the leader and follower share the SO-101 kinematics).
The `so101_leader_plugin` binary is a C++ plugin that is **not** part of the `isaacteleop` pip
package — you build it from the Isaac Teleop source tree. Follow
[Build Isaac Teleop from source](https://nvidia.github.io/IsaacTeleop/main/getting_started/build_from_source/index.html)
(in short, from your Isaac Teleop checkout: `cmake -B build && cmake --build build --parallel &&
cmake --install build`); the build installs the plugins under `<IsaacTeleop>/install/plugins/`, so
the binary lands at `install/plugins/so101_leader/so101_leader_plugin` — the `--launch_plugin` path
below. See the plugin's own `README.md` (next to the binary) for its serial/calibration details.
Point `--teleop.port` at the physical leader's serial port and `--launch_plugin` at that plugin
binary to have the script spawn it after CloudXR is up:
```bash
python -m examples.isaac_teleop_to_so101.teleoperate --robot.type=so101_follower --robot.port=/dev/ttyACM0 \
--robot.id=so101_follower_arm --teleop.type=so101_leader \
--teleop.port=/dev/ttyACM1 --teleop.id=so101_leader_arm \
--launch_plugin=/code/Teleop/install/plugins/so101_leader/so101_leader_plugin
```
(Note `so101_leader` here is the _Isaac_ leader, resolved against the Isaac Teleop device
registry, distinct from `lerobot-teleoperate`'s serial `so101_leader`.) When a `--teleop.port` is
set, the plugin's tick→radian calibration is inferred from `--teleop.id` and passed to the plugin
as its third positional arg — the LeRobot-format JSON at
`HF_LEROBOT_CALIBRATION/teleoperators/so_leader/<id>.json`, the same file the serial SO-101 leader
uses (`lerobot-calibrate --teleop.type=so101_leader --teleop.id=<id>`). If it is missing the script
warns and the plugin uses built-in defaults. Run `python -m examples.isaac_teleop_to_so101.teleoperate --help` for all flags. Its
startup safety contract: by default the follower is
slewed to the leader's first reading over `--align_duration` seconds (`--align=false` to skip) so
the arm does not snap when the mirror begins, and while the leader stream is stale the follower is
held at its measured pose.
The URDF fetch uses `huggingface_hub` (already a LeRobot dependency) against the public
`lerobot/robot-urdfs` bucket, so it needs no login. It is cached under
`HF_LEROBOT_HOME/robot-urdfs/so101/`; delete that folder to force a re‑download.
Then, in your headset: squeeze and hold the grip to engage, move the controller to drive the
arm, twist/tilt it to orient the wrist, and press the trigger to close the gripper
(proportionally — release to open).
To record a dataset (not just teleoperate), use `record.py` in the same folder. It dispatches on
`--teleop.type` (`xr_controller` | `so101_leader`) exactly like `teleoperate.py`, so either device
can drive the follower, and it saves the commanded joints to a LeRobot dataset (`lerobot-record`-style
`--dataset.*` flags). See its module docstring for the full CLI and the keyboard recording shortcuts.
## Important pipeline steps and options
The clutch already produces an absolute base‑frame pose, so the processor side is a thin
**absolute‑pose** path — there is no frame remap, no delta accumulation, and no
`EEReferenceAndDelta` stage.
- `MapXRControllerActionToRobotAction` is a stateless per‑frame mapping from the device output to
the IK input contract. It writes the absolute base‑frame position, encodes the absolute
orientation as a rotvec target, and inverts the closedness into a motor gripper target:
```python
action["ee.x"], action["ee.y"], action["ee.z"] = ee_pose[:3] # absolute, base frame [m]
action["ee.wx"], action["ee.wy"], action["ee.wz"] = orient_rotvec # orientation target (rotvec)
action["ee.gripper_pos"] = (1 - closedness) * 100 # motor units; SO-101 calibrates 100 = open
```
The gripper polarity (`100 = open, 0 = closed`) is a hardware‑calibration convention in the source — flip it there if the jaw opens when it should close.
- `EEBoundsAndSafety` clamps the EE to a workspace and rate‑limits per‑frame jumps. The clutch's
no‑teleport keeps frames small, so `max_ee_step_m` mostly catches transient controller tracking
glitches. The z floor is `0.0` (the table plane) so a stray target cannot drive the EE below the
table; x/y stay at the loose `[-1, 1]` m box. Set `raise_on_jump=False` so an over‑limit frame is
**clamped and warned** instead of raising — a crash mid‑loop would leave the arm uncontrolled:
```python
EEBoundsAndSafety(
end_effector_bounds={"min": [-1.0, -1.0, 0.0], "max": [1.0, 1.0, 1.0]},
max_ee_step_m=0.10,
raise_on_jump=False,
)
```
- `InverseKinematicsEEToJoints(initial_guess_current_joints=False, orientation_weight=0.01)` solves
closed‑loop Placo IK. SO‑101 is a 5‑DOF arm, so the IK is position‑dominant; the small
`orientation_weight` lets it softly track the orientation target carried in `ee.w*` so the wrist
follows the hand, while the under‑determined roll stays partial by design. There is **no**
`GripperVelocityToJoint`: the absolute `ee.gripper_pos` is passed straight to `gripper.pos`.
`initial_guess_current_joints=False` warm‑starts each solve from the **previous IK solution**
rather than re‑seeding from the measured joints, so the joint trajectory stays continuous
frame‑to‑frame. Tune `orientation_weight` on hardware — too high fights position tracking, too
low ignores the orientation command.
The example also gates safety at the loop level: after the startup reset slew (on by default —
pass `--reset_to_origin=false` to keep the arm where it is), it commands the robot **only while
the clutch is engaged**, and re‑sends the measured joints while disengaged, so releasing the
clutch freezes the arm in place.
See the [Processors for Robots and Teleoperators](./processors_robots_teleop) guide for more on
adapting the pipeline to other robots.
## Troubleshooting
- **`ModuleNotFoundError: isaacteleop`** — the `isaacteleop` package is not installed in the
active environment. Re-run the install command at the top of this guide:
`uv pip install "isaacteleop[cloudxr,retargeters-lite]~=1.3.131"`.
- **No controllers found** — make sure the CloudXR runtime is running, the firewall ports are
whitelisted, and the headset is connected (see
[Set up CloudXR and connect a headset](#set-up-cloudxr-and-connect-a-headset) and the Isaac
Teleop [Quick Start](https://nvidia.github.io/IsaacTeleop/main/getting_started/quick_start.html)).
- **CloudXR auto-launch failed** — `connect()` raises a `RuntimeError` if the runtime does not
come up within its startup timeout. Check the launcher logs under `~/.cloudxr/logs`. Common
causes: the EULA was never accepted (run `python -m isaacteleop.cloudxr --accept-eula` once,
interactively — the auto-launch prompts on stdin and hangs headless), or the runtime is already
running externally (set `LEROBOT_CLOUDXR_SKIP_AUTOLAUNCH=1` or `auto_launch_cloudxr=False` to
skip the auto-launch).
- **Arm does not move** — the clutch is a deadman: you must hold the squeeze/grip past
`clutch_threshold`. Lower the threshold if your controller's squeeze is reported softly.
- **Motion feels misaligned** — confirm the headset/play space orientation. The controller stream
is rebased into the robot base frame by the `base_T_anchor` transform on `XRControllerConfig`
(default: standard OpenXR → robot axis convention); adjust it if your anchor frame differs.
## Learn more
NVIDIA Isaac Teleop documentation ([docs home](https://nvidia.github.io/IsaacTeleop/),
[GitHub](https://github.com/NVIDIA/IsaacTeleop)):
- [Quick Start](https://nvidia.github.io/IsaacTeleop/main/getting_started/quick_start.html) —
install, run the CloudXR server, connect a headset, run a teleop example.
- [TeleopSession](https://nvidia.github.io/IsaacTeleop/main/getting_started/teleop_session.html) —
the session API `XRController` wraps.
- [Retargeting interface](https://nvidia.github.io/IsaacTeleop/main/references/retargeting/index.html)
and [architecture overview](https://nvidia.github.io/IsaacTeleop/main/overview/architecture.html) —
how source nodes and retargeters compose into a pipeline.
- [Build from source](https://nvidia.github.io/IsaacTeleop/main/getting_started/build_from_source/index.html) —
build `isaacteleop` (and its C++ plugins, including the `so101_leader` plugin used above) from a
local checkout.
- [System Requirements](https://nvidia.github.io/IsaacTeleop/main/references/requirements.html) and
the [CloudXR SDK docs](https://docs.nvidia.com/cloudxr-sdk) — supported platforms, GPUs,
CloudXR/OpenXR runtime versions, and headsets.
docs/source/koch.mdx
+283
View File
@@ -0,0 +1,283 @@
# Koch v1.1
In the steps below, we explain how to assemble the Koch v1.1 robot.
## Order and assemble the parts
Follow the sourcing and assembling instructions provided in this [README](https://github.com/jess-moss/koch-v1-1). This will guide you through setting up both the follower and leader arms, as shown in the image below.
For a visual walkthrough of the assembly process, you can refer to [this video tutorial](https://youtu.be/8nQIg9BwwTk).
> [!WARNING]
> Since the production of this video, we simplified the configuration phase. Because of this, two things differ from the instructions in that video:
>
> - Don't plug in all the motor cables right away and wait to be instructed to do so in [Configure the motors](#configure-the-motors).
> - Don't screw in the controller board (PCB) to the base right away and wait for being instructed to do so in [Configure the motors](#configure-the-motors).
## Install LeRobot 🤗
To install LeRobot follow, our [Installation Guide](./installation)
In addition to these instructions, you need to install the Dynamixel SDK:
```bash
pip install -e ".[dynamixel]"
```
## Configure the motors
### 1. Find the USB ports associated with each arm
To find the port for each bus servo adapter, run this script:
```bash
lerobot-find-port
```
<hfoptions id="example">
<hfoption id="Mac">
Example output:
```
Finding all available ports for the MotorBus.
['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
Remove the USB cable from your MotorsBus and press Enter when done.
[...Disconnect corresponding leader or follower arm and press Enter...]
The port of this MotorsBus is /dev/tty.usbmodem575E0032081
Reconnect the USB cable.
```
Where the found port is: `/dev/tty.usbmodem575E0032081` corresponding to your leader or follower arm.
</hfoption>
<hfoption id="Linux">
On Linux, you might need to give access to the USB ports by running:
```bash
sudo chmod 666 /dev/ttyACM0
sudo chmod 666 /dev/ttyACM1
```
Example output:
```
Finding all available ports for the MotorBus.
['/dev/ttyACM0', '/dev/ttyACM1']
Remove the usb cable from your MotorsBus and press Enter when done.
[...Disconnect corresponding leader or follower arm and press Enter...]
The port of this MotorsBus is /dev/ttyACM1
Reconnect the USB cable.
```
Where the found port is: `/dev/ttyACM1` corresponding to your leader or follower arm.
</hfoption>
</hfoptions>
### 2. Set the motors ids and baudrates
Each motor is identified by a unique id on the bus. When brand new, motors usually come with a default id of `1`. For the communication to work properly between the motors and the controller, we first need to set a unique, different id to each motor. Additionally, the speed at which data is transmitted on the bus is determined by the baudrate. In order to talk to each other, the controller and all the motors need to be configured with the same baudrate.
To that end, we first need to connect to each motor individually with the controller in order to set these. Since we will write these parameters in the non-volatile section of the motors' internal memory (EEPROM), we'll only need to do this once.
If you are repurposing motors from another robot, you will probably also need to perform this step, as the ids and baudrate likely won't match.
#### Follower
Connect the usb cable from your computer and the 5V power supply to the follower arm's controller board. Then, run the following command or run the API example with the port you got from the previous step. You'll also need to give your leader arm a name with the `id` parameter.
For a visual reference on how to set the motor ids please refer to [this video](https://huggingface.co/docs/lerobot/en/so101#setup-motors-video) where we follow the process for the SO101 arm.
<hfoptions id="setup_motors">
<hfoption id="Command">
```bash
lerobot-setup-motors \
--robot.type=koch_follower \
--robot.port=/dev/tty.usbmodem575E0031751 # <- paste here the port found at previous step
```
</hfoption>
<hfoption id="API example">
<!-- prettier-ignore-start -->
```python
from lerobot.robots.koch_follower import KochFollower, KochFollowerConfig
config = KochFollowerConfig(
port="/dev/tty.usbmodem575E0031751",
id="my_awesome_follower_arm",
)
follower = KochFollower(config)
follower.setup_motors()
```
<!-- prettier-ignore-end -->
</hfoption>
</hfoptions>
You should see the following instruction.
```
Connect the controller board to the 'gripper' motor only and press enter.
```
As instructed, plug the gripper's motor. Make sure it's the only motor connected to the board, and that the motor itself is not yet daisy-chained to any other motor. As you press `[Enter]`, the script will automatically set the id and baudrate for that motor.
<details>
<summary>Troubleshooting</summary>
If you get an error at that point, check your cables and make sure they are plugged in properly:
<ul>
<li>Power supply</li>
<li>USB cable between your computer and the controller board</li>
<li>The 3-pin cable from the controller board to the motor</li>
</ul>
If you are using a Waveshare controller board, make sure that the two jumpers are set on the `B` channel (USB).
</details>
You should then see the following message:
```
'gripper' motor id set to 6
```
Followed by the next instruction:
```
Connect the controller board to the 'wrist_roll' motor only and press enter.
```
You can disconnect the 3-pin cable from the controller board but you can leave it connected to the gripper motor on the other end as it will already be in the right place. Now, plug in another 3-pin cable to the wrist roll motor and connect it to the controller board. As with the previous motor, make sure it is the only motor connected to the board and that the motor itself isn't connected to any other one.
Repeat the operation for each motor as instructed.
> [!TIP]
> Check your cabling at each step before pressing Enter. For instance, the power supply cable might disconnect as you manipulate the board.
When you are done, the script will simply finish, at which point the motors are ready to be used. You can now plug the 3-pin cable from each motor to the next one, and the cable from the first motor (the 'shoulder pan' with id=1) to the controller board, which can now be attached to the base of the arm.
#### Leader
Do the same steps for the leader arm but modify the command or script accordingly.
<hfoptions id="setup_motors">
<hfoption id="Command">
```bash
lerobot-setup-motors \
--teleop.type=koch_leader \
--teleop.port=/dev/tty.usbmodem575E0031751 \ # <- paste here the port found at previous step
```
</hfoption>
<hfoption id="API example">
<!-- prettier-ignore-start -->
```python
from lerobot.teleoperators.koch_leader import KochLeader, KochLeaderConfig
config = KochLeaderConfig(
port="/dev/tty.usbmodem575E0031751",
id="my_awesome_leader_arm",
)
leader = KochLeader(config)
leader.setup_motors()
```
<!-- prettier-ignore-end -->
</hfoption>
</hfoptions>
## Calibrate
Next, you'll need to calibrate your robot to ensure that the leader and follower arms have the same position values when they are in the same physical position.
The calibration process is very important because it allows a neural network trained on one robot to work on another.
#### Follower
Run the following command or API example to calibrate the follower arm:
<hfoptions id="calibrate_follower">
<hfoption id="Command">
```bash
lerobot-calibrate \
--robot.type=koch_follower \
--robot.port=/dev/tty.usbmodem58760431551 \ # <- The port of your robot
--robot.id=my_awesome_follower_arm # <- Give the robot a unique name
```
</hfoption>
<hfoption id="API example">
<!-- prettier-ignore-start -->
```python
from lerobot.robots.koch_follower import KochFollowerConfig, KochFollower
config = KochFollowerConfig(
port="/dev/tty.usbmodem585A0076891",
id="my_awesome_follower_arm",
)
follower = KochFollower(config)
follower.connect(calibrate=False)
follower.calibrate()
follower.disconnect()
```
<!-- prettier-ignore-end -->
</hfoption>
</hfoptions>
We unified the calibration method for most robots. Thus, the calibration steps for this Koch arm are the same as the steps for the SO100 and SO101. First, we have to move the robot to the position where each joint is in the middle of its range, then we press `Enter`. Secondly, we move all joints through their full range of motion. A video of this same process for the SO101 as reference can be found [here](https://huggingface.co/docs/lerobot/en/so101#calibration-video).
#### Leader
Do the same steps to calibrate the leader arm, run the following command or API example:
<hfoptions id="calibrate_leader">
<hfoption id="Command">
```bash
lerobot-calibrate \
--teleop.type=koch_leader \
--teleop.port=/dev/tty.usbmodem58760431551 \ # <- The port of your robot
--teleop.id=my_awesome_leader_arm # <- Give the robot a unique name
```
</hfoption>
<hfoption id="API example">
<!-- prettier-ignore-start -->
```python
from lerobot.teleoperators.koch_leader import KochLeaderConfig, KochLeader
config = KochLeaderConfig(
port="/dev/tty.usbmodem575E0031751",
id="my_awesome_leader_arm",
)
leader = KochLeader(config)
leader.connect(calibrate=False)
leader.calibrate()
leader.disconnect()
```
<!-- prettier-ignore-end -->
</hfoption>
</hfoptions>
Congrats 🎉, your robot is all set to learn a task on its own. Start training it by following this tutorial: [Getting started with real-world robots](./il_robots)
> [!TIP]
> If you have any questions or need help, please reach out on [Discord](https://discord.com/invite/s3KuuzsPFb).
docs/source/language_and_recipes.mdx
+147
View File
@@ -0,0 +1,147 @@
# Language columns and recipes
Most LeRobot datasets ship with a single `task` string per episode — fine for
short, single-instruction skills, but not enough for the longer-horizon,
multi-modal robot policies the field is moving toward (high-level planning,
memory, interjections, VQA, tool use). To support those policies without
forking the dataset format, LeRobot extends `LeRobotDataset` with two optional
language columns and a small recipe layer that turns those rows into
chat-style training samples on the fly.
The design splits cleanly into three layers:
1. **Data in the dataset** — language annotations stored next to frames in
`data/chunk-*/file-*.parquet` as two optional columns (`language_persistent`
and `language_events`). Datasets without these columns keep their existing
behavior.
2. **Recipe** — a YAML file that declares which annotation rows to bind and
how to lay them out as chat turns (`role`, `content`, optional images,
optional tool calls). Recipes are pure config; no Python required to add a
new one.
3. **Training format** — at sample time, `RenderMessagesStep` resolves the
recipe against the per-frame annotations and emits HF-style `messages` plus
LeRobot-specific sidecars (`message_streams`, `target_message_indices`)
that policy processors consume.
This page describes each layer in turn.
## Layer 1 — language columns in the dataset
The two optional columns live next to frame data in
`data/chunk-*/file-*.parquet`:
- `language_persistent`: a list of rows broadcast across every frame in an episode for state that remains active, such as `subtask`, `plan`, and `memory`.
- `language_events`: a list of rows only on the exact frame where an event was emitted, such as `interjection`, `vqa`, and speech tool calls.
Both columns share the same row shape (event rows omit `timestamp` because the
frame the row sits on already provides it):
```text
role: string
content: string | null
style: string | null
timestamp: float32 # persistent rows only
camera: string | null # observation.images.* feature key, view-dependent rows only
tool_calls: list[Json] | null
```
The `camera` field tags rows whose `content` is grounded in a specific camera
view. Rows of view-dependent styles (`vqa` and `trace`) MUST set `camera` to
the matching `observation.images.*` feature key. Rows of every other style —
including `motion`, which describes robot-frame primitives in joint / Cartesian
terms — MUST leave `camera` as `null`. Pipeline writers and the validator
enforce this via `validate_camera_field(style, camera)`.
`meta/tasks.parquet` remains the canonical source for the task. The special `${task}` recipe binding always reads that task string and does not depend on language annotations.
### Architecture
The language stack itself has three internal modules backing layer 1:
1. `lerobot.datasets.language` defines the schema, style registry, and `column_for_style`.
2. `lerobot.datasets.language_render` resolves rows and renders messages.
3. `RenderMessagesStep` turns dataset samples into `messages`, `message_streams`, and `target_message_indices`.
`LeRobotDataset` stays recipe-agnostic. It passes `language_persistent` and `language_events` through when present, and unannotated datasets keep their existing behavior.
## Layer 2 — recipe anatomy
Recipes are YAML files backed by `TrainingRecipe` and `MessageTurn`. They
declare which annotation rows to pull (via `bindings`) and how to compose them
into chat turns (`messages`).
```yaml
messages:
- { role: user, content: "${task}", stream: high_level }
- { role: assistant, content: "${subtask}", stream: low_level, target: true }
```
A recipe can also branch into a weighted **blend** of sub-recipes. At sample
time, exactly one branch is selected deterministically from the sample index,
so different frames train different objectives (e.g. memory updates vs.
low-level execution vs. VQA) without any Python wiring.
### Temporal semantics
Persistent styles are active after emission until replaced:
- `active_at(t, style=subtask)`
- `nth_prev(style=memory, offset=1)`
- `nth_next(style=subtask, offset=1)`
Event styles only exist on their exact timestamp:
- `emitted_at(t, style=interjection)`
- `emitted_at(t, style=vqa, role=user, camera=observation.images.top)`
- `emitted_at(t, role=assistant, tool_name=say)`
Exact event matching has no tolerance window, so writers must stamp event rows with frame timestamps from the parquet data.
### View-dependent resolution
For view-dependent styles (`vqa` and `trace`), the resolver gains a
`camera=` filter parallel to `role=` and `tool_name=`. Datasets with multiple
cameras typically emit one (`vqa`, `user`) + (`vqa`, `assistant`) pair per
camera at the same timestamp; without `camera=`, those resolvers see two
matches and raise an ambiguity error. Recipes consume each camera through its
own binding plus a matching image block, e.g.
```yaml
ask_vqa_top:
bindings:
vqa_query: "emitted_at(t, style=vqa, role=user, camera=observation.images.top)"
vqa: "emitted_at(t, style=vqa, role=assistant, camera=observation.images.top)"
messages:
- role: user
stream: high_level
if_present: vqa_query
content:
- { type: image, feature: observation.images.top }
- { type: text, text: "${vqa_query}" }
- {
role: assistant,
content: "${vqa}",
stream: high_level,
target: true,
if_present: vqa,
}
```
Add one such sub-recipe per camera the dataset records.
## Layer 3 — training format
Rendered samples use HF-style chat messages plus LeRobot sidecars:
```python
sample["messages"]
sample["message_streams"]
sample["target_message_indices"]
```
The renderer does not apply a tokenizer chat template. Policy processors decide how to serialize the messages for their backbone, which keeps the same dataset usable across SmolVLA, Pi0.5, and any future VLM that expects OpenAI-style chat messages.
## Graceful absence
If both language columns are missing, `None`, or empty, `RenderMessagesStep` is a no-op.
If an event-scoped branch is selected on a frame without the required event row, rendering returns `None`, allowing a loader to retry another sample.
docs/source/lekiwi.mdx
+343
View File
@@ -0,0 +1,343 @@
# LeKiwi
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/1740517739083.jpeg"
alt="LeKiwi"
width="70%"
/>
In the steps below, we explain how to assemble the LeKiwi mobile robot.
## Source the parts
Follow this [README](https://github.com/SIGRobotics-UIUC/LeKiwi). It contains the bill of materials, with a link to source the parts, as well as the instructions to 3D print the parts.
And advise if it's your first time printing or if you don't own a 3D printer.
### Wired version
If you have the **wired** LeKiwi version, you can skip the installation of the Raspberry Pi and setting up SSH. You can also run all commands directly on your PC for both the LeKiwi scripts and the leader arm scripts for teleoperating.
## Install software on Pi
Now we have to set up the remote PC that will run on the LeKiwi Robot. This is normally a Raspberry Pi, but can be any PC that can run on 5V and has enough usb ports (2 or more) for the cameras and motor control board.
### Install OS
For setting up the Raspberry Pi and its SD-card see: [Setup PI](https://www.raspberrypi.com/documentation/computers/getting-started.html). Here is explained how to download the [Imager](https://www.raspberrypi.com/software/) to install Raspberry Pi OS or Ubuntu.
### Setup SSH
After setting up your Pi, you should enable and set up [SSH](https://www.raspberrypi.com/news/coding-on-raspberry-pi-remotely-with-visual-studio-code/) (Secure Shell Protocol) so you can log in to the Pi from your laptop without requiring a screen, keyboard, and mouse on the Pi. A great tutorial on how to do this can be found [here](https://www.raspberrypi.com/documentation/computers/remote-access.html#ssh). Logging into your Pi can be done in your Command Prompt (cmd) or, if you use VSCode you can use [this](https://marketplace.visualstudio.com/items?itemName=ms-vscode-remote.remote-ssh) extension.
### Install LeRobot on Pi 🤗
On your Raspberry Pi install LeRobot using our [Installation Guide](./installation)
In addition to these instructions, you need to install the Feetech SDK & ZeroMQ on your Pi:
```bash
pip install -e ".[lekiwi]"
```
## Install LeRobot locally
If you already have installed LeRobot on your laptop/pc you can skip this step; otherwise, please follow along as we do the same steps we did on the Pi.
Follow our [Installation Guide](./installation)
In addition to these instructions, you need to install the Feetech SDK & ZeroMQ on your laptop/pc:
```bash
pip install -e ".[lekiwi]"
```
Great :hugs:! You are now done installing LeRobot, and we can begin assembling the SO100/SO101 arms and the mobile base :robot:.
Every time you now want to use LeRobot, you can go to the `~/lerobot` folder where we installed LeRobot and run one of the commands.
# Step-by-Step Assembly Instructions
First, we will assemble the two SO100/SO101 arms. One to attach to the mobile base and one for teleoperation. Then we will assemble the mobile base. The instructions for assembling can be found on these two pages:
- [Assemble SO101](./so101#step-by-step-assembly-instructions)
- [Assemble LeKiwi](https://github.com/SIGRobotics-UIUC/LeKiwi/blob/main/Assembly.md)
### Find the USB ports associated with motor board
To find the port for each bus servo adapter, run this script:
```bash
lerobot-find-port
```
<hfoptions id="example">
<hfoption id="Mac">
Example output:
```
Finding all available ports for the MotorBus.
['/dev/tty.usbmodem575E0032081']
Remove the USB cable from your MotorsBus and press Enter when done.
[...Disconnect corresponding leader or follower arm and press Enter...]
The port of this MotorsBus is /dev/tty.usbmodem575E0032081
Reconnect the USB cable.
```
Where the found port is: `/dev/tty.usbmodem575E0032081` corresponding to your board.
</hfoption>
<hfoption id="Linux">
On Linux, you might need to give access to the USB ports by running:
```bash
sudo chmod 666 /dev/ttyACM0
sudo chmod 666 /dev/ttyACM1
```
Example output:
```
Finding all available ports for the MotorBus.
['/dev/ttyACM0']
Remove the usb cable from your MotorsBus and press Enter when done.
[...Disconnect corresponding leader or follower arm and press Enter...]
The port of this MotorsBus is /dev/ttyACM0
Reconnect the USB cable.
```
Where the found port is: `/dev/ttyACM0` corresponding to your board.
</hfoption>
</hfoptions>
### Configure motors
The instructions for configuring the motors can be found in the SO101 [docs](./so101#configure-the-motors). Besides the ids for the arm motors, we also need to set the motor ids for the mobile base. These need to be in a specific order to work. Below an image of the motor ids and motor mounting positions for the mobile base. Note that we only use one Motor Control board on LeKiwi. This means the motor ids for the wheels are 7, 8 and 9.
You can run this command to setup motors for LeKiwi. It will first setup the motors for arm (id 6..1) and then setup motors for wheels (9,8,7)
```bash
lerobot-setup-motors \
--robot.type=lekiwi \
--robot.port=/dev/tty.usbmodem58760431551 # <- paste here the port found at previous step
```
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/motor_ids.webp" alt="Motor ID's for mobile robot" title="Motor ID's for mobile robot" width="60%">
### Troubleshoot communication
If you are having trouble connecting to the Mobile SO100, follow these steps to diagnose and resolve the issue.
#### 1. Verify IP Address Configuration
Make sure that the correct IP for the Pi is used in the commands or in your code. To check the Raspberry Pi's IP address, run (on the Pi command line):
```bash
hostname -I
```
#### 2. Check if Pi is reachable from laptop/pc
Try pinging the Raspberry Pi from your laptop:
```bach
ping <your_pi_ip_address>
```
If the ping fails:
- Ensure the Pi is powered on and connected to the same network.
- Check if SSH is enabled on the Pi.
#### 3. Try SSH connection
If you can't SSH into the Pi, it might not be properly connected. Use:
```bash
ssh <your_pi_user_name>@<your_pi_ip_address>
```
If you get a connection error:
- Ensure SSH is enabled on the Pi by running:
```bash
sudo raspi-config
```
Then navigate to: **Interfacing Options -> SSH** and enable it.
### Calibration
Now we have to calibrate the leader arm and the follower arm. The wheel motors don't have to be calibrated.
The calibration process is very important because it allows a neural network trained on one robot to work on another.
### Calibrate follower arm (on mobile base)
Make sure the arm is connected to the Raspberry Pi and run this script or API example (on the Raspberry Pi via SSH) to launch calibration of the follower arm:
```bash
lerobot-calibrate \
--robot.type=lekiwi \
--robot.id=my_awesome_kiwi # <- Give the robot a unique name
```
We unified the calibration method for most robots, thus, the calibration steps for this SO100 arm are the same as the steps for the Koch and SO101. First, we have to move the robot to the position where each joint is in the middle of its range, then we press `Enter`. Secondly, we move all joints through their full range of motion. A video of this same process for the SO101 as reference can be found [here](https://huggingface.co/docs/lerobot/en/so101#calibration-video).
### Wired version
If you have the **wired** LeKiwi version, please run all commands on your laptop.
### Calibrate leader arm
Then, to calibrate the leader arm (which is attached to the laptop/pc). Run the following command of API example on your laptop:
<hfoptions id="calibrate_leader">
<hfoption id="Command">
```bash
lerobot-calibrate \
--teleop.type=so100_leader \
--teleop.port=/dev/tty.usbmodem58760431551 \ # <- The port of your robot
--teleop.id=my_awesome_leader_arm # <- Give the robot a unique name
```
</hfoption>
<hfoption id="API example">
<!-- prettier-ignore-start -->
```python
from lerobot.teleoperators.so_leader import SO100LeaderConfig, SO100Leader
config = SO100LeaderConfig(
port="/dev/tty.usbmodem58760431551",
id="my_awesome_leader_arm",
)
leader = SO100Leader(config)
leader.connect(calibrate=False)
leader.calibrate()
leader.disconnect()
```
<!-- prettier-ignore-end -->
</hfoption>
</hfoptions>
## Teleoperate LeKiwi
> [!TIP]
> If you're using a Mac, you might need to give Terminal permission to access your keyboard for teleoperation. Go to System Preferences > Security & Privacy > Input Monitoring and check the box for Terminal.
To teleoperate, SSH into your Raspberry Pi, and run `conda activate lerobot` and this command:
```bash
python -m lerobot.robots.lekiwi.lekiwi_host --robot.id=my_awesome_kiwi
```
Then on your laptop, also run `conda activate lerobot` and run the API example, make sure you set the correct `remote_ip` and `port` in `examples/lekiwi/teleoperate.py`.
```bash
python examples/lekiwi/teleoperate.py
```
You should see on your laptop something like this: `[INFO] Connected to remote robot at tcp://172.17.133.91:5555 and video stream at tcp://172.17.133.91:5556.` Now you can move the leader arm and use the keyboard (w,a,s,d) to drive forward, left, backwards, right. And use (z,x) to turn left or turn right. You can use (r,f) to increase and decrease the speed of the mobile robot. There are three speed modes, see the table below:
| Speed Mode | Linear Speed (m/s) | Rotation Speed (deg/s) |
| ---------- | ------------------ | ---------------------- |
| Fast | 0.4 | 90 |
| Medium | 0.25 | 60 |
| Slow | 0.1 | 30 |
| Key | Action |
| --- | -------------- |
| W | Move forward |
| A | Move left |
| S | Move backward |
| D | Move right |
| Z | Turn left |
| X | Turn right |
| R | Increase speed |
| F | Decrease speed |
> [!TIP]
> If you use a different keyboard, you can change the keys for each command in the [`LeKiwiClientConfig`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/robots/lekiwi/config_lekiwi.py).
### Wired version
If you have the **wired** LeKiwi version, please run all commands on your laptop.
## Record a dataset
Once you're familiar with teleoperation, you can record your first dataset.
We use the Hugging Face hub features for uploading your dataset. If you haven't previously used the Hub, make sure you can login via the cli using a write-access token, this token can be generated from the [Hugging Face settings](https://huggingface.co/settings/tokens).
Add your token to the CLI by running this command:
```bash
hf auth login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
```
Then store your Hugging Face repository name in a variable:
```bash
HF_USER=$(hf auth whoami | awk -F': *' 'NR==1 {print $2}')
echo $HF_USER
```
Now you can record a dataset. To record episodes and upload your dataset to the hub, execute this API example tailored for LeKiwi. Make sure to first adapt the `remote_ip`, `repo_id`, `port` and `task` in the script. If you would like to run the script for longer you can increase `NB_CYCLES_CLIENT_CONNECTION`.
```bash
python examples/lekiwi/record.py
```
#### Dataset upload
Locally, your dataset is stored in this folder: `~/.cache/huggingface/lerobot/{repo-id}`. At the end of data recording, your dataset will be uploaded on your Hugging Face page (e.g. https://huggingface.co/datasets/cadene/so101_test) that you can obtain by running:
```bash
echo https://huggingface.co/datasets/${HF_USER}/so101_test
```
Your dataset will be automatically tagged with `LeRobot` for the community to find it easily, and you can also add custom tags (in this case `tutorial` for example).
You can look for other LeRobot datasets on the hub by searching for `LeRobot` [tags](https://huggingface.co/datasets?other=LeRobot).
#### Tips for gathering data
Once you're comfortable with data recording, you can create a larger dataset for training. A good starting task is grasping an object at different locations and placing it in a bin. We suggest recording at least 50 episodes, with 10 episodes per location. Keep the cameras fixed and maintain consistent grasping behavior throughout the recordings. Also make sure the object you are manipulating is visible on the camera's. A good rule of thumb is you should be able to do the task yourself by only looking at the camera images.
In the following sections, you’ll train your neural network. After achieving reliable grasping performance, you can start introducing more variations during data collection, such as additional grasp locations, different grasping techniques, and altering camera positions.
Avoid adding too much variation too quickly, as it may hinder your results.
If you want to dive deeper into this important topic, you can check out the [blog post](https://huggingface.co/blog/lerobot-datasets#what-makes-a-good-dataset) we wrote on what makes a good dataset.
#### Troubleshooting:
- On Linux, the recording control-flow keys (arrow keys, Escape) work on X11, Wayland, and headless/SSH sessions as long as you run the recording from an interactive terminal (keep it focused) — no `$DISPLAY` setup is needed; the letter equivalents `n` / `r` / `q` also work. Note that **keyboard teleoperation of the LeKiwi base** is different: it relies on a global key backend and therefore works only on an X11 session, a Windows desktop, or macOS with Accessibility/Input Monitoring granted — not on Wayland or headless machines. See [pynput limitations](https://pynput.readthedocs.io/en/latest/limitations.html#linux).
## Replay an episode
To replay an episode run the API example below, make sure to change `remote_ip`, `port`, LeRobotDatasetId and episode index.
```bash
python examples/lekiwi/replay.py
```
Congrats 🎉, your robot is all set to learn a task on its own. Start training it by the training part of this tutorial: [Getting started with real-world robots](./il_robots)
## Evaluate your policy
To evaluate your policy run the `evaluate.py` API example, make sure to change `remote_ip`, `port`, model..
```bash
python examples/lekiwi/evaluate.py
```
> [!TIP]
> If you have any questions or need help, please reach out on [Discord](https://discord.com/invite/s3KuuzsPFb).
docs/source/lelab.mdx
+29
View File
@@ -0,0 +1,29 @@
# LeLab - LeRobot Guide
LeLab is a graphical user interface built on top of the LeRobot library, designed to make robotics accessible without needing to memorize CLI commands. From a single app you can configure your robot, teleoperate it, collect datasets, train policies locally or on cloud GPUs via HF Jobs, and deploy trained models back onto your robot. It's the easiest way to go from an unboxed SO-101 to a working policy, and a great companion for anyone learning the LeRobot workflow. Source code and issues live on GitHub: [huggingface/leLab](https://github.com/huggingface/leLab).
> [!TIP]
> For now LeLab is compatible only with SO-ARM101
<Youtube id="VqyKUuW9V1g" />
### Installation
Requires [`uv`](https://docs.astral.sh/uv/getting-started/installation/). Install and launch in one command:
```
uv tool install git+https://github.com/huggingface/leLab.git && lelab
```
After install, run `lelab` from your terminal anytime to start the app.
### Features
- **Add robots** — Select arm type (leader/follower), calibrate each joint from the middle position, and attach cameras.
- **Teleoperation** — Control the follower arm with the leader and see a live 3D visualization of the arms.
- **Dataset recording** — Define a task description, number of episodes, and episode/reset durations. Press spacebar to advance between episodes. 30+ episodes recommended.
- **Local training** — Train a policy directly on your own machine with a selected dataset, policy type, batch size, and step count.
- **Cloud training with HF Jobs** — Train on powerful GPUs via [HF Jobs](https://huggingface.co/docs/huggingface_hub/en/guides/jobs) with transparent pricing. Run `hf auth login` first. See the [Compute HW Guide](hardware_guide) for hardware/batch size tips.
- **Training visualization** — Watch progress live in the app, with checkpoints saved automatically.
- **Run trained policies** — Pick any model from your jobs list and run inference on your robot with one click.
- **Use community datasets** — Provide any Hugging Face dataset ID to train on datasets you didn't record yourself.
docs/source/lerobot-dataset-v3.mdx
+354
View File
@@ -0,0 +1,354 @@
# LeRobotDataset v3.0
`LeRobotDataset v3.0` is a standardized format for robot learning data. It provides unified access to multi-modal time-series data, sensorimotor signals and multi‑camera video, as well as rich metadata for indexing, search, and visualization on the Hugging Face Hub.
This docs will guide you to:
- Understand the v3.0 design and directory layout
- Record a dataset and push it to the Hub
- Load datasets for training with `LeRobotDataset`
- Stream datasets without downloading using `StreamingLeRobotDataset`
- Apply image transforms for data augmentation during training
- Migrate existing `v2.1` datasets to `v3.0`
- Experiment with other `LeRobotDataset` formats and implementations like Lance
## What’s new in `v3`
- **File-based storage**: Many episodes per Parquet/MP4 file (v2 used one file per episode).
- **Relational metadata**: Episode boundaries and lookups are resolved through metadata, not filenames.
- **Hub-native streaming**: Consume datasets directly from the Hub with `StreamingLeRobotDataset`.
- **Lower file-system pressure**: Fewer, larger files ⇒ faster initialization and fewer issues at scale.
- **Unified organization**: Clean directory layout with consistent path templates across data and videos.
## Installation
`LeRobotDataset v3.0` will be included in `lerobot >= 0.4.0`.
Until that stable release, you can use the main branch by following the [build from source instructions](./installation#from-source).
## Record a dataset
Run the command below to record a dataset with the SO-101 and push to the Hub:
```bash
lerobot-record \
--robot.type=so101_follower \
--robot.port=/dev/tty.usbmodem585A0076841 \
--robot.id=my_awesome_follower_arm \
--robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 1920, height: 1080, fps: 30}}" \
--teleop.type=so101_leader \
--teleop.port=/dev/tty.usbmodem58760431551 \
--teleop.id=my_awesome_leader_arm \
--display_data=true \
--dataset.repo_id=${HF_USER}/record-test \
--dataset.num_episodes=5 \
--dataset.single_task="Grab the black cube" \
--dataset.streaming_encoding=true \
# --dataset.rgb_encoder.vcodec=auto \
--dataset.encoder_threads=2
```
See the [recording guide](./il_robots#record-a-dataset) for more details.
## Format design
A core v3 principle is **decoupling storage from the user API**: data is stored efficiently (few large files), while the public API exposes intuitive episode-level access.
`v3` has three pillars:
1. **Tabular data**: Low‑dimensional, high‑frequency signals (states, actions, timestamps) stored in **Apache Parquet**. Access is memory‑mapped or streamed via the `datasets` stack.
2. **Visual data**: Camera frames concatenated and encoded into **MP4**. Frames from the same episode are grouped; videos are sharded per camera for practical sizes.
3. **Metadata**: JSON/Parquet records describing schema (feature names, dtypes, shapes), frame rates, normalization stats, and **episode segmentation** (start/end offsets into shared Parquet/MP4 files).
> To scale to millions of episodes, tabular rows and video frames from multiple episodes are **concatenated** into larger files. Episode‑specific views are reconstructed **via metadata**, not file boundaries.
<div style="display:flex; justify-content:center; gap:12px; flex-wrap:wrap;">
<figure style="margin:0; text-align:center;">
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobotdataset-v3/asset1datasetv3.png"
alt="LeRobotDataset v3 diagram"
width="220"
/>
<figcaption style="font-size:0.9em; color:#666;">
From episode‑based to file‑based datasets
</figcaption>
</figure>
</div>
### Directory layout (simplified)
- **`meta/info.json`**: canonical schema (features, shapes/dtypes), FPS, codebase version, and **path templates** to locate data/video shards.
- **`meta/stats.json`**: global feature statistics (mean/std/min/max) used for normalization; exposed as `dataset.meta.stats`.
- **`meta/tasks.jsonl`**: natural‑language task descriptions mapped to integer IDs for task‑conditioned policies.
- **`meta/episodes/`**: per‑episode records (lengths, tasks, offsets) stored as **chunked Parquet** for scalability.
- **`data/`**: frame‑by‑frame **Parquet** shards; each file typically contains **many episodes**.
- **`videos/`**: **MP4** shards per camera; each file typically contains **many episodes**.
## Load a dataset for training
`LeRobotDataset` returns Python dictionaries of PyTorch tensors and integrates with `torch.utils.data.DataLoader`. Here is a code example showing its use:
```python
import torch
from lerobot.datasets import LeRobotDataset
repo_id = "yaak-ai/L2D-v3"
# 1) Load from the Hub (cached locally)
dataset = LeRobotDataset(repo_id)
# 2) Random access by index
sample = dataset[100]
print(sample)
# {
# 'observation.state': tensor([...]),
# 'action': tensor([...]),
# 'observation.images.front_left': tensor([C, H, W]),
# 'timestamp': tensor(1.234),
# ...
# }
# 3) Temporal windows via delta_timestamps (seconds relative to t)
delta_timestamps = {
"observation.images.front_left": [-0.2, -0.1, 0.0] # 0.2s and 0.1s before current frame
}
dataset = LeRobotDataset(repo_id, delta_timestamps=delta_timestamps)
# Accessing an index now returns a stack for the specified key(s)
sample = dataset[100]
print(sample["observation.images.front_left"].shape) # [T, C, H, W], where T=3
# 4) Wrap with a DataLoader for training
batch_size = 16
data_loader = torch.utils.data.DataLoader(dataset, batch_size=batch_size)
device = "cuda" if torch.cuda.is_available() else "cpu"
for batch in data_loader:
observations = batch["observation.state"].to(device)
actions = batch["action"].to(device)
images = batch["observation.images.front_left"].to(device)
# model.forward(batch)
```
## Stream a dataset (no downloads)
Use `StreamingLeRobotDataset` to iterate directly from the Hub without local copies. This allows to stream large datasets without the need to downloading them onto disk or loading them onto memory, and is a key feature of the new dataset format.
```python
from lerobot.datasets import StreamingLeRobotDataset
repo_id = "yaak-ai/L2D-v3"
dataset = StreamingLeRobotDataset(repo_id) # streams directly from the Hub
```
<div style="display:flex; justify-content:center; gap:12px; flex-wrap:wrap;">
<figure style="margin:0; text-align:center;">
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobotdataset-v3/streaming-lerobot.png"
alt="StreamingLeRobotDataset"
width="520"
/>
<figcaption style="font-size:0.9em; color:#666;">
Stream directly from the Hub for on‑the‑fly training.
</figcaption>
</figure>
</div>
## Image transforms
Image transforms are data augmentations applied to camera frames during training to improve model robustness and generalization. LeRobot supports various transforms including brightness, contrast, saturation, hue, and sharpness adjustments.
### Using transforms during dataset creation/recording
Currently, transforms are applied during **training time only**, not during recording. When you create or record a dataset, the raw images are stored without transforms. This allows you to experiment with different augmentations later without re-recording data.
### Adding transforms to existing datasets (API)
Use the `image_transforms` parameter when loading a dataset for training:
```python
from lerobot.datasets import LeRobotDataset
from lerobot.transforms import ImageTransforms, ImageTransformsConfig, ImageTransformConfig
# Option 1: Use default transform configuration (disabled by default)
transforms_config = ImageTransformsConfig(
enable=True, # Enable transforms
max_num_transforms=3, # Apply up to 3 transforms per frame
random_order=False, # Apply in standard order
)
transforms = ImageTransforms(transforms_config)
dataset = LeRobotDataset(
repo_id="your-username/your-dataset",
image_transforms=transforms
)
# Option 2: Create custom transform configuration
custom_transforms_config = ImageTransformsConfig(
enable=True,
max_num_transforms=2,
random_order=True,
tfs={
"brightness": ImageTransformConfig(
weight=1.0,
type="ColorJitter",
kwargs={"brightness": (0.7, 1.3)} # Adjust brightness range
),
"contrast": ImageTransformConfig(
weight=2.0, # Higher weight = more likely to be selected
type="ColorJitter",
kwargs={"contrast": (0.8, 1.2)}
),
"sharpness": ImageTransformConfig(
weight=0.5, # Lower weight = less likely to be selected
type="SharpnessJitter",
kwargs={"sharpness": (0.3, 2.0)}
),
}
)
dataset = LeRobotDataset(
repo_id="your-username/your-dataset",
image_transforms=ImageTransforms(custom_transforms_config)
)
# Option 3: Use pure torchvision transforms
from torchvision.transforms import v2
torchvision_transforms = v2.Compose([
v2.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2, hue=0.1),
v2.GaussianBlur(kernel_size=3, sigma=(0.1, 2.0)),
])
dataset = LeRobotDataset(
repo_id="your-username/your-dataset",
image_transforms=torchvision_transforms
)
```
### Available transform types
LeRobot provides several transform types:
- **`ColorJitter`**: Adjusts brightness, contrast, saturation, and hue
- **`SharpnessJitter`**: Randomly adjusts image sharpness
- **`Identity`**: No transformation (useful for testing)
You can also use any `torchvision.transforms.v2` transform by passing it directly to the `image_transforms` parameter.
### Configuration options
- **`enable`**: Enable/disable transforms (default: `False`)
- **`max_num_transforms`**: Maximum number of transforms applied per frame (default: `3`)
- **`random_order`**: Apply transforms in random order vs. standard order (default: `False`)
- **`weight`**: Sampling probability for each transform (higher = more likely, if sum of weights is not 1, they will be normalized)
- **`kwargs`**: Transform-specific parameters (e.g., brightness range)
### Visualizing transforms
Use the visualization script to preview how transforms affect your data:
```bash
lerobot-imgtransform-viz \
--repo-id=your-username/your-dataset \
--output-dir=./transform_examples \
--n-examples=5
```
This saves example images showing the effect of each transform, helping you tune parameters.
### Best practices
- **Start conservative**: Begin with small ranges (e.g., brightness 0.9-1.1) and increase gradually
- **Test first**: Use the visualization script to ensure transforms look reasonable
- **Monitor training**: Strong augmentations can hurt performance if too aggressive
- **Match your domain**: If your robot operates in varying lighting, use brightness/contrast transforms
- **Combine wisely**: Using too many transforms simultaneously can make training unstable
## Migrate `v2.1` → `v3.0`
A converter aggregates per‑episode files into larger shards and writes episode offsets/metadata. Convert your dataset using the instructions below.
```bash
# Pre-release build with v3 support:
pip install "https://github.com/huggingface/lerobot/archive/33cad37054c2b594ceba57463e8f11ee374fa93c.zip"
# Convert an existing v2.1 dataset hosted on the Hub:
python -m lerobot.scripts.convert_dataset_v21_to_v30 --repo-id=<HF_USER/DATASET_ID>
```
**What it does**
- Aggregates parquet files: `episode-0000.parquet`, `episode-0001.parquet`, … → **`file-0000.parquet`**, …
- Aggregates mp4 files: `episode-0000.mp4`, `episode-0001.mp4`, … → **`file-0000.mp4`**, …
- Updates `meta/episodes/*` (chunked Parquet) with per‑episode lengths, tasks, and byte/frame offsets.
## Common Issues
### Always call `finalize()` before pushing
When creating or recording datasets, you **must** call `dataset.finalize()` to properly close parquet writers. See the [PR #1903](https://github.com/huggingface/lerobot/pull/1903) for more details.
```python
from lerobot.datasets import LeRobotDataset
# Create dataset and record episodes
dataset = LeRobotDataset.create(...)
for episode in range(num_episodes):
# Record frames
for frame in episode_data:
dataset.add_frame(frame)
dataset.save_episode()
# Call finalize() when done recording and before push_to_hub()
dataset.finalize() # Closes parquet writers, writes metadata footers
dataset.push_to_hub()
```
**Why is this necessary?**
Dataset v3.0 uses incremental parquet writing with buffered metadata for efficiency. The `finalize()` method:
- Flushes any buffered episode metadata to disk
- Closes parquet writers to write footer metadata, otherwise the parquet files will be corrupt
- Ensures the dataset is valid for loading
Without calling `finalize()`, your parquet files will be incomplete and the dataset won't load properly.
## Other formats and implementations
### Lance
Lance is a useful format for multimodal AI datasets, especially for large-scale training requiring high performance IO and random access.
The `lerobot-lancedb` package implements `LeRobotLanceDataset` (for JPEG images) and `LeRobotLanceVideoDataset` (for mp4 videos).
Those two storage layouts both subclass LeRobotDataset and can provide data loading speed ups.
`LeRobotLanceDataset` is a drop-in replacement for `LeRobotDataset`:
```python
from lerobot.datasets import LeRobotDatasetMetadata
from lerobot.policies.diffusion.configuration_diffusion import DiffusionConfig
from lerobot_lancedb import LeRobotLanceDataset, LeRobotLanceVideoDataset
cfg = DiffusionConfig(...)
meta = LeRobotDatasetMetadata(root=local_dataset_path) # or use repo_id=... to load metadata from the Hub
delta_timestamps = {...}
# Use LeRobotLanceDataset for image datasets
dataset = LeRobotLanceDataset(
root=local_dataset_path, # or use repo_id=... to stream from the Hub
delta_timestamps=delta_timestamps,
return_uint8=True,
)
# Or use LeRobotLanceVideoDataset for video datasets:
dataset = LeRobotLanceVideoDataset(
root=local_dataset_path, # or use repo_id=... to stream from the Hub
delta_timestamps=delta_timestamps,
return_uint8=True,
)
```
Join the discussion on [Github](https://github.com/huggingface/lerobot/issues/3608) and explore the `lerobot-lancedb` documentation [here](https://lancedb.github.io/lerobot-lancedb/).
docs/source/libero.mdx
+181
View File
@@ -0,0 +1,181 @@
# LIBERO
LIBERO is a benchmark designed to study **lifelong robot learning** — the idea that robots need to keep learning and adapting with their users over time, not just be pretrained once. It provides a set of standardized manipulation tasks that focus on **knowledge transfer**: how well a robot can apply what it has already learned to new situations. By evaluating on LIBERO, different algorithms can be compared fairly and researchers can build on each other's work.
- Paper: [Benchmarking Knowledge Transfer for Lifelong Robot Learning](https://arxiv.org/abs/2306.03310)
- GitHub: [Lifelong-Robot-Learning/LIBERO](https://github.com/Lifelong-Robot-Learning/LIBERO)
- Project website: [libero-project.github.io](https://libero-project.github.io)
![An overview of the LIBERO benchmark](https://libero-project.github.io/assets/img/libero/fig1.png)
## Available tasks
LIBERO includes **five task suites** covering **130 tasks**, ranging from simple object manipulations to complex multi-step scenarios:
| Suite | CLI name | Tasks | Description |
| -------------- | ---------------- | ----- | -------------------------------------------------- |
| LIBERO-Spatial | `libero_spatial` | 10 | Tasks requiring reasoning about spatial relations |
| LIBERO-Object | `libero_object` | 10 | Tasks centered on manipulating different objects |
| LIBERO-Goal | `libero_goal` | 10 | Goal-conditioned tasks with changing targets |
| LIBERO-90 | `libero_90` | 90 | Short-horizon tasks from the LIBERO-100 collection |
| LIBERO-Long | `libero_10` | 10 | Long-horizon tasks from the LIBERO-100 collection |
## Installation
After following the LeRobot installation instructions:
```bash
pip install -e ".[libero]"
```
<Tip>
LIBERO requires Linux (`sys_platform == 'linux'`). LeRobot uses MuJoCo for simulation — set the rendering backend before training or evaluation:
```bash
export MUJOCO_GL=egl # for headless servers (HPC, cloud)
```
</Tip>
## Evaluation
### Default evaluation (recommended)
Evaluate across the four standard suites (10 episodes per task):
```bash
lerobot-eval \
--policy.path="your-policy-id" \
--env.type=libero \
--env.task=libero_spatial,libero_object,libero_goal,libero_10 \
--eval.batch_size=1 \
--eval.n_episodes=10 \
--env.max_parallel_tasks=1
```
### Single-suite evaluation
Evaluate on one LIBERO suite:
```bash
lerobot-eval \
--policy.path="your-policy-id" \
--env.type=libero \
--env.task=libero_object \
--eval.batch_size=2 \
--eval.n_episodes=3
```
- `--env.task` picks the suite (`libero_object`, `libero_spatial`, etc.).
- `--env.task_ids` restricts to specific task indices (`[0]`, `[1,2,3]`, etc.). Omit to run all tasks in the suite.
- `--eval.batch_size` controls how many environments run in parallel.
- `--eval.n_episodes` sets how many episodes to run per task.
### Multi-suite evaluation
Benchmark a policy across multiple suites at once by passing a comma-separated list:
```bash
lerobot-eval \
--policy.path="your-policy-id" \
--env.type=libero \
--env.task=libero_object,libero_spatial \
--eval.batch_size=1 \
--eval.n_episodes=2
```
### Control mode
LIBERO supports two control modes — `relative` (default) and `absolute`. Different VLA checkpoints are trained with different action parameterizations, so make sure the mode matches your policy:
```bash
--env.control_mode=relative # or "absolute"
```
### Policy inputs and outputs
**Observations:**
- `observation.state` — 8-dim proprioceptive features (eef position, axis-angle orientation, gripper qpos)
- `observation.images.image` — main camera view (`agentview_image`), HWC uint8
- `observation.images.image2` — wrist camera view (`robot0_eye_in_hand_image`), HWC uint8
<Tip warning={true}>
LeRobot enforces the `.images.*` prefix for visual features. Ensure your
policy config `input_features` use the same naming keys, and that your dataset
metadata keys follow this convention. If your data contains different keys,
you must rename the observations to match what the policy expects, since
naming keys are encoded inside the normalization statistics layer.
</Tip>
**Actions:**
- Continuous control in `Box(-1, 1, shape=(7,))` — 6D end-effector delta + 1D gripper
### Recommended evaluation episodes
For reproducible benchmarking, use **10 episodes per task** across all four standard suites (Spatial, Object, Goal, Long). This gives 400 total episodes and matches the protocol used for published results.
## Training
### Dataset
We provide a preprocessed LIBERO dataset fully compatible with LeRobot:
- [HuggingFaceVLA/libero](https://huggingface.co/datasets/HuggingFaceVLA/libero)
For reference, the original dataset published by Physical Intelligence:
- [physical-intelligence/libero](https://huggingface.co/datasets/physical-intelligence/libero)
### Example training command
```bash
lerobot-train \
--policy.type=smolvla \
--policy.repo_id=${HF_USER}/libero-test \
--policy.load_vlm_weights=true \
--dataset.repo_id=HuggingFaceVLA/libero \
--env.type=libero \
--env.task=libero_10 \
--output_dir=./outputs/ \
--steps=100000 \
--batch_size=4 \
--eval.batch_size=1 \
--eval.n_episodes=1 \
--env_eval_freq=1000
```
## Reproducing published results
We reproduce the results of Pi0.5 on the LIBERO benchmark. We take the Physical Intelligence LIBERO base model (`pi05_libero`) and finetune for an additional 6k steps in bfloat16, with batch size of 256 on 8 H100 GPUs using the [HuggingFace LIBERO dataset](https://huggingface.co/datasets/HuggingFaceVLA/libero).
The finetuned model: [lerobot/pi05_libero_finetuned](https://huggingface.co/lerobot/pi05_libero_finetuned)
### Evaluation command
```bash
lerobot-eval \
--output_dir=./eval_logs/ \
--env.type=libero \
--env.task=libero_spatial,libero_object,libero_goal,libero_10 \
--eval.batch_size=1 \
--eval.n_episodes=10 \
--policy.path=pi05_libero_finetuned \
--policy.n_action_steps=10 \
--env.max_parallel_tasks=1
```
We set `n_action_steps=10`, matching the original OpenPI implementation.
### Results
| Model | LIBERO Spatial | LIBERO Object | LIBERO Goal | LIBERO 10 | Average |
| ------------------- | -------------- | ------------- | ----------- | --------- | -------- |
| **Pi0.5 (LeRobot)** | 97.0 | 99.0 | 98.0 | 96.0 | **97.5** |
These results are consistent with the [original results](https://github.com/Physical-Intelligence/openpi/tree/main/examples/libero#results) reported by Physical Intelligence:
| Model | LIBERO Spatial | LIBERO Object | LIBERO Goal | LIBERO 10 | Average |
| ------------------ | -------------- | ------------- | ----------- | --------- | --------- |
| **Pi0.5 (OpenPI)** | 98.8 | 98.2 | 98.0 | 92.4 | **96.85** |
docs/source/libero_plus.mdx
+188
View File
@@ -0,0 +1,188 @@
# LIBERO-plus
LIBERO-plus is a **robustness benchmark** for Vision-Language-Action (VLA) models built on top of [LIBERO](./libero). It systematically stress-tests policies by applying **seven independent perturbation dimensions** to the original LIBERO task set, exposing failure modes that standard benchmarks miss.
- Paper: [In-depth Robustness Analysis of Vision-Language-Action Models](https://arxiv.org/abs/2510.13626)
- GitHub: [sylvestf/LIBERO-plus](https://github.com/sylvestf/LIBERO-plus)
- Dataset: [lerobot/libero_plus](https://huggingface.co/datasets/lerobot/libero_plus)
![An overview of the LIBERO-plus benchmark perturbation dimensions](https://github.com/sylvestf/LIBERO-plus/raw/main/static/images/libero-plus.jpg)
## Perturbation dimensions
LIBERO-plus creates ~10 000 task variants by perturbing each original LIBERO task along these axes:
| Dimension | What changes |
| --------------------- | ----------------------------------------------------- |
| Objects layout | Target position, presence of confounding objects |
| Camera viewpoints | Camera position, orientation, field-of-view |
| Robot initial states | Manipulator start pose |
| Language instructions | LLM-rewritten task description (paraphrase / synonym) |
| Light conditions | Intensity, direction, color, shadow |
| Background textures | Scene surface and object appearance |
| Sensor noise | Photometric distortions and image degradation |
## Available task suites
LIBERO-plus covers the same five suites as LIBERO:
| Suite | CLI name | Tasks | Max steps | Description |
| -------------- | ---------------- | ----- | --------- | -------------------------------------------------- |
| LIBERO-Spatial | `libero_spatial` | 10 | 280 | Tasks requiring reasoning about spatial relations |
| LIBERO-Object | `libero_object` | 10 | 280 | Tasks centered on manipulating different objects |
| LIBERO-Goal | `libero_goal` | 10 | 300 | Goal-conditioned tasks with changing targets |
| LIBERO-90 | `libero_90` | 90 | 400 | Short-horizon tasks from the LIBERO-100 collection |
| LIBERO-Long | `libero_10` | 10 | 520 | Long-horizon tasks from the LIBERO-100 collection |
<Tip warning={true}>
Installing LIBERO-plus **replaces** vanilla LIBERO — it uninstalls `hf-libero`
so that `import libero` resolves to the LIBERO-plus fork. You cannot have both
installed at the same time. To switch back to vanilla LIBERO, uninstall the
fork and reinstall with `pip install -e ".[libero]"`.
</Tip>
## Installation
### System dependencies (Linux only)
```bash
sudo apt install libexpat1 libfontconfig1-dev libmagickwand-dev
```
### Python package
```bash
pip install -e ".[libero]" "robosuite==1.4.1" bddl easydict mujoco wand scikit-image gym
git clone https://github.com/sylvestf/LIBERO-plus.git
cd LIBERO-plus && pip install --no-deps -e .
pip uninstall -y hf-libero # so `import libero` resolves to the fork
```
LIBERO-plus is installed from its GitHub fork rather than a pyproject extra — the fork ships as a namespace package that pip can't handle, so it must be cloned and added to `PYTHONPATH`. See `docker/Dockerfile.benchmark.libero_plus` for the canonical install. MuJoCo is required, so only Linux is supported.
<Tip>
Set the MuJoCo rendering backend before running evaluation:
```bash
export MUJOCO_GL=egl # headless / HPC / cloud
```
</Tip>
### Download LIBERO-plus assets
LIBERO-plus ships its extended asset pack separately. Download `assets.zip` from the [Hugging Face dataset](https://huggingface.co/datasets/Sylvest/LIBERO-plus/tree/main) and extract it into the LIBERO-plus package directory:
```bash
# After installing the package, find where it was installed:
python -c "import libero; print(libero.__file__)"
# Then extract assets.zip into <package_root>/libero/assets/
```
## Evaluation
### Default evaluation (recommended)
Evaluate across the four standard suites (10 episodes per task):
```bash
lerobot-eval \
--policy.path="your-policy-id" \
--env.type=libero_plus \
--env.task=libero_spatial,libero_object,libero_goal,libero_10 \
--eval.batch_size=1 \
--eval.n_episodes=10 \
--env.max_parallel_tasks=1
```
### Single-suite evaluation
Evaluate on one LIBERO-plus suite:
```bash
lerobot-eval \
--policy.path="your-policy-id" \
--env.type=libero_plus \
--env.task=libero_spatial \
--eval.batch_size=1 \
--eval.n_episodes=10
```
- `--env.task` picks the suite (`libero_spatial`, `libero_object`, etc.).
- `--env.task_ids` restricts to specific task indices (`[0]`, `[1,2,3]`, etc.). Omit to run all tasks in the suite.
- `--eval.batch_size` controls how many environments run in parallel.
- `--eval.n_episodes` sets how many episodes to run per task.
### Multi-suite evaluation
Benchmark a policy across multiple suites at once by passing a comma-separated list:
```bash
lerobot-eval \
--policy.path="your-policy-id" \
--env.type=libero_plus \
--env.task=libero_spatial,libero_object \
--eval.batch_size=1 \
--eval.n_episodes=10
```
### Control mode
LIBERO-plus supports two control modes — `relative` (default) and `absolute`. Different VLA checkpoints are trained with different action parameterizations, so make sure the mode matches your policy:
```bash
--env.control_mode=relative # or "absolute"
```
### Policy inputs and outputs
**Observations:**
- `observation.state` — 8-dim proprioceptive features (eef position, axis-angle orientation, gripper qpos)
- `observation.images.image` — main camera view (`agentview_image`), HWC uint8
- `observation.images.image2` — wrist camera view (`robot0_eye_in_hand_image`), HWC uint8
**Actions:**
- Continuous control in `Box(-1, 1, shape=(7,))` — 6D end-effector delta + 1D gripper
### Recommended evaluation episodes
For reproducible benchmarking, use **10 episodes per task** across all four standard suites (Spatial, Object, Goal, Long). This gives 400 total episodes and matches the protocol used for published results.
## Training
### Dataset
A LeRobot-format training dataset for LIBERO-plus is available at:
- [lerobot/libero_plus](https://huggingface.co/datasets/lerobot/libero_plus)
### Example training command
```bash
lerobot-train \
--policy.type=smolvla \
--policy.repo_id=${HF_USER}/smolvla_libero_plus \
--policy.load_vlm_weights=true \
--dataset.repo_id=lerobot/libero_plus \
--env.type=libero_plus \
--env.task=libero_spatial \
--output_dir=./outputs/ \
--steps=100000 \
--batch_size=4 \
--eval.batch_size=1 \
--eval.n_episodes=1 \
--env_eval_freq=1000
```
## Relationship to LIBERO
LIBERO-plus is a drop-in extension of LIBERO:
- Same Python gym interface (`LiberoEnv`, `LiberoProcessorStep`)
- Same camera names and observation/action format
- Same task suite names
- Installs under the same `libero` Python package name (different GitHub repo)
To use the original LIBERO benchmark, see [LIBERO](./libero) and use `--env.type=libero`.
docs/source/lingbot_va.mdx
+187
View File
@@ -0,0 +1,187 @@
# LingBot-VA
LingBot-VA is an **autoregressive video-action world-model policy** built on the **Wan2.2**
video-diffusion stack. It interleaves, in one autoregressive sequence, the prediction of
future **video latents** and **robot actions** ("VA" = Video-Action). The LeRobot
integration wires LingBot-VA into the standard training, evaluation and processor
interfaces.
## Model Overview
LingBot-VA is a **dual-stream "mixture-of-transformers"**: a video/latent stream
(`patch_embedding_mlp → blocks → proj_out`) and an action stream
(`action_embedder → blocks → action_proj_out`) share the same 30 transformer blocks and
text conditioning.
| Component | Class | Role |
| ------------------------ | ----------------------- | ----------------------------------------------------------- |
| DiT backbone (trainable) | `WanTransformer3DModel` | ~5B-param dual-stream transformer. |
| VAE (frozen) | `AutoencoderKLWan` | Wan2.2 VAE, `z_dim=48`. Lazy-pulled from the source repo. |
| Text encoder (frozen) | `UMT5EncoderModel` | UMT5-XXL, `d_model=4096`. Lazy-pulled from the source repo. |
At inference the policy runs an autoregressive loop per chunk: it denoises the video-latent
stream (CFG, ~20 steps) and the action stream (~50 steps) with two independent
flow-matching schedulers, maintaining a KV cache across chunks. Real observed keyframes are
fed back into the KV cache as the chunk is executed (closed-loop world modeling).
### What the LeRobot Integration Covers
- Standard `policy.type=lingbot_va` configuration through LeRobot.
- Ready-to-use LeRobot-format checkpoints on the Hub (converted from the released upstream ones).
- Autoregressive dual-stream inference behind the standard `select_action` interface
(single-environment eval, `--eval.batch_size=1`).
- Opt-in saving of the policy's **predicted (imagined) videos** during eval / training.
- Evaluation with `lerobot-eval` on LIBERO and RoboTwin.
- Training / fine-tuning via the dual-stream flow-matching loss (`policy.forward`), see below.
## Installation
1. Install LeRobot by following the [Installation Guide](./installation).
2. Install the LingBot-VA extra:
```bash
pip install -e ".[lingbot_va]"
```
## Checkpoints
The released upstream checkpoints have been converted to LeRobot format and pushed to the Hub:
| Variant | LeRobot checkpoint |
| ---------------------- | -------------------------------- |
| LIBERO-Long post-train | `lerobot/lingbot_va_libero_long` |
| RoboTwin post-train | `lerobot/lingbot_va_robotwin` |
| Pretrained base | `lerobot/lingbot_va_base` |
Only the trainable ~5B transformer is stored in the LeRobot
`model.safetensors`. The frozen VAE + UMT5 + tokenizer (~20 GB) are pulled from
`config.wan_pretrained_path` at load time (defaults to the source `robbyant/*` repo). The
UMT5-XXL text encoder runs on CPU by default (`config.text_encoder_device`) so the 5B
transformer + VAE fit on a single 24–32 GB GPU.
## Evaluation (LIBERO)
```bash
lerobot-eval \
--policy.path=lerobot/lingbot_va_libero_long \
--policy.device=cuda \
--env.type=libero --env.task=libero_10 \
--env.observation_height=128 --env.observation_width=128 \
--eval.n_episodes=50 --eval.batch_size=1 \
--output_dir=outputs/eval/lingbot_va_libero
```
LingBot-VA's streaming inference (KV cache + observed-keyframe feedback) is implemented for
single-environment eval; use `--eval.batch_size=1`.
## Evaluation (RoboTwin)
RoboTwin 2.0 needs the SAPIEN + CuRobo simulator stack. You can use the benchmark Docker image
(`docker/Dockerfile.benchmark.robotwin`, which also needs `warp-lang==1.3.1` and CuRobo built
with the GPU's compute capability in `TORCH_CUDA_ARCH_LIST`). RoboTwin uses **end-effector-pose
control**, so run with `--env.action_mode=ee`: the policy predicts per-arm `xyz+quaternion+gripper`
deltas (`robotwin_tshape` latent layout) that are composed onto the episode's initial eef pose and
executed via CuRobo IK.
```bash
lerobot-eval \
--policy.path=lerobot/lingbot_va_robotwin \
--policy.device=cuda \
--env.type=robotwin --env.task=beat_block_hammer --env.action_mode=ee \
--eval.n_episodes=10 --eval.batch_size=1 \
--output_dir=outputs/eval/lingbot_va_robotwin
```
### Saving predicted (imagined) videos
Set `--policy.save_predicted_video=true` to additionally VAE-decode the predicted video
latents and write `pred_episode_*.mp4` next to the env-rendered `eval_episode_*.mp4` videos.
The same flag works for the periodic eval during `lerobot-train`.
## Training / fine-tuning
`LingBotVAPolicy.forward(batch)` implements the dual-stream **flow-matching** loss
(`latent_loss + action_loss`, timestep-weighted, action-masked) from the paper: it VAE-encodes
the camera clips into video latents, UMT5-encodes the task, noises both streams, runs the
transformer's block-causal training pass and returns `(loss, metrics)`. Optimizer preset is AdamW
with a linear-warmup-then-constant schedule (matching upstream).
Requirements:
- The block-causal masks use PyTorch **flex-attention**, so build the policy with
`--policy.attn_mode=flex` for training (the default `torch` SDPA is inference-only).
- The full 5B DiT does not fit a single 24–32 GB GPU under AdamW; fine-tune with **LoRA**
(`--policy.use_peft=true`) and/or optimizer offload. `get_optim_params` returns only the
trainable (e.g. adapter) parameters; the VAE + UMT5 text encoder stay frozen.
```bash
lerobot-train \
--policy.path=lerobot/lingbot_va_libero_long --policy.attn_mode=flex \
--policy.use_peft=true \
--dataset.repo_id=<your LeRobot-format dataset> \
--batch_size=1 --steps=... --output_dir=outputs/train/lingbot_va
```
The dataset must provide camera clips (a temporal window per camera, VAE-encoded to
`frame_chunk_size` latent frames) and `frame_chunk_size * action_per_frame` action steps per item.
## Data format (action channels & camera order)
LingBot-VA is an **end-effector (Cartesian) pose** policy, it predicts EEF poses + gripper, not
joint positions. Actions live in a fixed multi-embodiment **30-dim** layout; map your robot's
action dimensions into these channels and pad the rest with `0` (`used_action_channel_ids` selects
the channels a given checkpoint actually uses):
| channels | meaning |
| -------- | ----------------------------------------------------- |
| 0–6 | Left-arm end-effector pose |
| 7–13 | Right-arm end-effector pose |
| 14–20 | Left-arm joints (unused by the released checkpoints) |
| 21–27 | Right-arm joints (unused by the released checkpoints) |
| 28 | Left gripper |
| 29 | Right gripper |
- **LIBERO** uses channels `0–6`: a 6-DoF EEF delta (xyz + rotation) + gripper (single arm).
- **RoboTwin** uses channels `[0–6, 28, 7–13, 29]`: left EEF (xyz + quaternion) + left gripper +
right EEF + right gripper (16 dims). The env converts these poses to joint trajectories via
CuRobo IK — joints are never predicted.
Joint-space datasets (or a different EEF convention) must be remapped into this schema before
fine-tuning these checkpoints.
**Camera order is fixed and order-sensitive**, per-camera latents are concatenated spatially in
`obs_cam_keys` order, so the physical camera→slot mapping must match training:
| benchmark | `obs_cam_keys` (in order) | `camera_layout` |
| --------- | ----------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------- |
| LIBERO | `observation.images.image` (agentview / 3rd-person), `observation.images.image2` (eye-in-hand wrist) | `width_concat` (latents concatenated on width) |
| RoboTwin | `observation.images.head_camera`, `observation.images.left_camera`, `observation.images.right_camera` | `robotwin_tshape` (full-res head below, two half-res wrists on top) |
The first camera is the exterior/head view and the rest are wrist views.
## Inference Hyperparameters (LIBERO)
| Key | Value |
| -------------------------------------- | --------------------------------------------------------------------------------- |
| height × width | 128 × 128 |
| cameras | `observation.images.image` (agentview), `observation.images.image2` (eye-in-hand) |
| action channels used | 0–6 (7-DoF arm + gripper) |
| action_per_frame / frame_chunk_size | 4 / 4 |
| attn_window | 30 |
| video / action denoising steps | 20 / 50 |
| guidance_scale / action_guidance_scale | 5 / 1 |
| snr_shift / action_snr_shift | 5.0 / 0.05 |
These are the defaults of `LingBotVAConfig`; override any of them via `--policy.<name>=...`.
## Notes
- **Attention backend:** inference uses the `torch` SDPA backend (always available). The
`flashattn` and `flex` backends are optional; `flex` is only needed for training.
- **Model size:** the DiT is ~5B params and the frozen VAE+UMT5 add ~20 GB; inference needs
roughly 18–24 GB of VRAM.
## License
LingBot-VA is released under Apache-2.0. See the
[upstream repository](https://github.com/Robbyant/lingbot-va).
docs/source/metaworld.mdx
+130
View File
@@ -0,0 +1,130 @@
# Meta-World
Meta-World is an open-source simulation benchmark for **multi-task and meta reinforcement learning** in continuous-control robotic manipulation. It bundles 50 diverse manipulation tasks using everyday objects and a common tabletop Sawyer arm, providing a standardized playground to test whether algorithms can learn many different tasks and generalize quickly to new ones.
- Paper: [Meta-World: A Benchmark and Evaluation for Multi-Task and Meta Reinforcement Learning paper](https://arxiv.org/abs/1910.10897)
- GitHub: [Farama-Foundation/Metaworld](https://github.com/Farama-Foundation/Metaworld)
- Project website: [metaworld.farama.org](https://metaworld.farama.org)
![MetaWorld MT10 demo](https://meta-world.github.io/figures/ml45.gif)
## Available tasks
Meta-World provides 50 tasks organized into difficulty groups. In LeRobot, you can evaluate on individual tasks, difficulty groups, or the full MT50 suite:
| Group | CLI name | Tasks | Description |
| ---------- | -------------------- | ----- | ------------------------------------------------------ |
| Easy | `easy` | 28 | Tasks with simple dynamics and single-step goals |
| Medium | `medium` | 11 | Tasks requiring multi-step reasoning |
| Hard | `hard` | 6 | Tasks with complex contacts and precise manipulation |
| Very Hard | `very_hard` | 5 | The most challenging tasks in the suite |
| MT50 (all) | Comma-separated list | 50 | All 50 tasks — the most challenging multi-task setting |
You can also pass individual task names directly (e.g., `assembly-v3`, `dial-turn-v3`).
We provide a LeRobot-ready dataset for Meta-World MT50 on the HF Hub: [lerobot/metaworld_mt50](https://huggingface.co/datasets/lerobot/metaworld_mt50). This dataset is formatted for the MT50 evaluation that uses all 50 tasks with fixed object/goal positions and one-hot task vectors for consistency.
## Installation
After following the LeRobot installation instructions:
```bash
pip install -e ".[metaworld]"
```
<Tip warning={true}>
If you encounter an `AssertionError: ['human', 'rgb_array', 'depth_array']` when running Meta-World environments, this is a mismatch between Meta-World and your Gymnasium version. Fix it with:
```bash
pip install "gymnasium==1.1.0"
```
</Tip>
## Evaluation
### Default evaluation (recommended)
Evaluate on the medium difficulty split (a good balance of coverage and compute):
```bash
lerobot-eval \
--policy.path="your-policy-id" \
--env.type=metaworld \
--env.task=medium \
--eval.batch_size=1 \
--eval.n_episodes=10
```
### Single-task evaluation
Evaluate on a specific task:
```bash
lerobot-eval \
--policy.path="your-policy-id" \
--env.type=metaworld \
--env.task=assembly-v3 \
--eval.batch_size=1 \
--eval.n_episodes=10
```
### Multi-task evaluation
Evaluate across multiple tasks or difficulty groups:
```bash
lerobot-eval \
--policy.path="your-policy-id" \
--env.type=metaworld \
--env.task=assembly-v3,dial-turn-v3,handle-press-side-v3 \
--eval.batch_size=1 \
--eval.n_episodes=10
```
- `--env.task` accepts explicit task lists (comma-separated) or difficulty groups (e.g., `easy`, `medium`, `hard`, `very_hard`).
- `--eval.batch_size` controls how many environments run in parallel.
- `--eval.n_episodes` sets how many episodes to run per task.
### Policy inputs and outputs
**Observations:**
- `observation.image` — single camera view (`corner2`), 480x480 HWC uint8
- `observation.state` — 4-dim proprioceptive state (end-effector position + gripper)
**Actions:**
- Continuous control in `Box(-1, 1, shape=(4,))` — 3D end-effector delta + 1D gripper
### Recommended evaluation episodes
For reproducible benchmarking, use **10 episodes per task**. For the full MT50 suite this gives 500 total episodes. If you care about generalization, run on the full MT50 — it is intentionally challenging and reveals strengths/weaknesses better than a few narrow tasks.
## Training
### Example training command
Train a SmolVLA policy on a subset of Meta-World tasks:
```bash
lerobot-train \
--policy.type=smolvla \
--policy.repo_id=${HF_USER}/metaworld-test \
--policy.load_vlm_weights=true \
--dataset.repo_id=lerobot/metaworld_mt50 \
--env.type=metaworld \
--env.task=assembly-v3,dial-turn-v3,handle-press-side-v3 \
--output_dir=./outputs/ \
--steps=100000 \
--batch_size=4 \
--eval.batch_size=1 \
--eval.n_episodes=1 \
--env_eval_freq=1000
```
## Practical tips
- Use the one-hot task conditioning for multi-task training (MT10/MT50 conventions) so policies have explicit task context.
- Inspect the dataset task descriptions and the `info["is_success"]` keys when writing post-processing or logging so your success metrics line up with the benchmark.
- Adjust `batch_size`, `steps`, and `env_eval_freq` to match your compute budget.
docs/source/molmoact2.mdx
+495
View File
@@ -0,0 +1,495 @@
# MolmoAct2 Policy
MolmoAct2 is the LeRobot policy implementation of
[MolmoAct2](https://allenai.org/blog/molmoact2), ported into the LeRobot
training, evaluation, checkpointing, and dataset interfaces for easier use with
LeRobot datasets.
This implementation currently supports training and evaluation for the regular
MolmoAct2 model. MolmoAct2-Think, which supports adaptive depth reasoning, is
not included in this LeRobot policy yet and is coming soon.
For the original MolmoAct2 training code used for the experiments reported in
the paper, see [allenai/molmoact2](https://github.com/allenai/molmoact2).
## Installation Requirements
Install LeRobot with the MolmoAct2 optional dependencies:
```bash
uv sync --locked --extra molmoact2
```
To run the models in this repository, you need an NVIDIA GPU. The measurements
below were taken on a single NVIDIA H100 80GB with bf16 model loading, LIBERO with two RGB cameras. MolmoAct2 rows use `chunk_size=10`, action dim 7
padded to `expected_max_action_dim=32`, and `num_flow_timesteps=8`. Training measurements use
`gradient_checkpointing=true` and include the forward pass, backward pass,
gradient clipping, optimizer step, and optimizer state allocation. Values are
peak GPU memory sampled with `nvidia-smi`. Leave a few GiB of headroom for
dataloader workers, CUDA context, and fragmentation.
Multi-GPU training through `accelerate` increases throughput and global batch
size, but this LeRobot port does not currently expose the original MolmoAct2
`fsdp_devices` model-parallel training path. The current training script has
not been tested for multi-node training.
| Mode | Peak Memory, bs=8 | Peak Memory, bs=16 | Peak Memory, bs=32 |
| ------------------------------------------------ | ----------------: | -----------------: | -----------------: |
| Inference, continuous, CUDA graph enabled (bs=1) | 12.1 GiB | - | - |
| Fine-tuning, action expert only, continuous | 16.5 GiB | 18.3 GiB | 21.4 GiB |
| Fine-tuning, LoRA VLM, both action modes | 20.2 GiB | 26.8 GiB | 41.3 GiB |
| Fine-tuning, full model, both action modes | 48.3 GiB | 49.8 GiB | 60.1 GiB |
The repo has been tested with Ubuntu 22.04.
## Usage
To use MolmoAct2 in a LeRobot training config, set:
```bash
--policy.type=molmoact2
```
## Training
MolmoAct2 can be fine-tuned from either the released MolmoAct2 Hugging Face
checkpoint format or from a checkpoint already saved by LeRobot. Both routes use
the same LeRobot training loop, dataset transforms, checkpoint saving, and
logging. The difference is only how the initial policy weights and processor
state are loaded.
### Training With Original MolmoAct2 Weight
Use `policy.checkpoint_path` when starting from a released MolmoAct2 checkpoint,
for example `allenai/MolmoAct2` or `allenai/MolmoAct2-LIBERO`. LeRobot will load
the original HF model files, then build its own policy processor from the
dataset metadata and the policy options below.
The command below shows full fine-tuning on the merged LIBERO dataset. It uses
bf16 model loading, 8 flow timesteps, LeRobot dataset statistics, image
augmentation, and LeRobot's checkpointing/logging path.
```bash
accelerate launch \
--num_processes=8 \
--mixed_precision=bf16 \
-m lerobot.scripts.lerobot_train \
--dataset.repo_id=allenai/MolmoAct2-LIBERO-Dataset \
--dataset.root=/path/to/lerobot/data/allenai/MolmoAct2-LIBERO-Dataset \
--dataset.video_backend=pyav \
--dataset.image_transforms.enable=true \
--policy.type=molmoact2 \
--policy.checkpoint_path=allenai/MolmoAct2-LIBERO \
--policy.device=cuda \
--policy.action_mode=both \
--policy.chunk_size=10 \
--policy.n_action_steps=10 \
--policy.setup_type="single franka robotic arm in libero" \
--policy.control_mode="delta end-effector pose" \
--policy.image_keys='["observation.images.image","observation.images.wrist_image"]' \
--policy.model_dtype=bfloat16 \
--policy.num_flow_timesteps=8 \
--policy.gradient_checkpointing=true \
--policy.freeze_embedding=true \
--policy.normalize_gripper=false \
--policy.enable_knowledge_insulation=false \
--policy.push_to_hub=false \
--wandb.enable=true \
--wandb.entity=<wandb_entity> \
--wandb.project=<wandb_project> \
--job_name=<job_name> \
--output_dir=outputs/<job_name> \
--steps=10000 \
--batch_size=32 \
--num_workers=4 \
--log_freq=20 \
--env_eval_freq=-1 \
--save_checkpoint=true \
--save_freq=2000
```
### Training With LeRobot MolmoAct2 Weight
Use `policy.path` when starting from a MolmoAct2 checkpoint that was saved by
LeRobot, either from a local `pretrained_model` directory or from the Hub. This
restores the saved LeRobot policy config, model weights, processor, and
normalization statistics. You can still override training-time options such as
`batch_size`, `steps`, LoRA flags, or `policy.action_mode`.
```bash
accelerate launch \
--num_processes=8 \
--mixed_precision=bf16 \
-m lerobot.scripts.lerobot_train \
--dataset.repo_id=allenai/MolmoAct2-LIBERO-Dataset \
--dataset.root=/path/to/lerobot/data/allenai/MolmoAct2-LIBERO-Dataset \
--dataset.video_backend=pyav \
--dataset.image_transforms.enable=true \
--policy.path=/path/to/pretrained_model \
--policy.device=cuda \
--policy.action_mode=both \
--policy.chunk_size=10 \
--policy.n_action_steps=10 \
--policy.model_dtype=bfloat16 \
--policy.num_flow_timesteps=8 \
--policy.gradient_checkpointing=true \
--wandb.enable=true \
--wandb.entity=<wandb_entity> \
--wandb.project=<wandb_project> \
--job_name=<job_name> \
--output_dir=outputs/<job_name> \
--steps=10000 \
--batch_size=32 \
--num_workers=4 \
--log_freq=20 \
--env_eval_freq=-1 \
--save_checkpoint=true \
--save_freq=2000
```
### Common Practices
For fine-tuning on a comparatively small dataset, such as a single LIBERO suite
or a real-world dataset with less than 200 demonstrations, a global batch size of
16 to 32 is a good starting point. In these settings, `policy.enable_lora_vlm=true` or `policy.train_action_expert_only=true` is also a practical choice. In both
cases, we intentionally keep the action expert fully trainable, which we found
to be crucial for model performance. For larger fine-tuning datasets, larger
global batch sizes and full fine-tuning are usually preferred.
### Common Policy Options
- `policy.checkpoint_path`: original MolmoAct2 HF checkpoint to initialize from.
Use this for released MolmoAct2 weights.
- `policy.path`: LeRobot checkpoint to initialize from. Use this for checkpoints
created by LeRobot training.
- `policy.action_mode`: training target, one of `continuous`, `discrete`, or
`both`. `both` trains the flow-matching action expert and the discrete
action-token loss.
- `policy.train_action_expert_only`: trains only parameters whose names contain
`action_expert`. It requires `policy.action_mode=continuous`.
- `policy.enable_lora_vlm`: enables LoRA on VLM linear layers. Use
`policy.enable_lora_action_expert=true` only if LoRA should also cover action
expert linear layers. When `policy.enable_lora_action_expert=false`, the
action expert base weights remain fully trainable while the VLM is trained
through LoRA adapters. When `policy.enable_lora_action_expert=true`, the
action expert is also adapter-tuned instead of fully fine-tuned.
- `policy.enable_knowledge_insulation`: when `true`, detaches action-expert
context K/V states before the action loss. The default is `false`.
- `policy.chunk_size`: action horizon used by the policy. For LIBERO we use
`10`. This LeRobot port overrides the loaded checkpoint's
`max_action_horizon` with this value.
- `policy.n_action_steps`: number of actions consumed from each predicted
chunk before querying the policy again. For LIBERO, set it to `chunk_size`.
- `policy.setup_type`: text inserted into the prompt to describe the robot and
scene, e.g. `single franka robotic arm in libero`. More examples are listed
in the `metadata_by_tag` entries of
[`norm_stats.json`](https://huggingface.co/allenai/MolmoAct2/blob/main/norm_stats.json).
- `policy.control_mode`: text inserted into the prompt to describe the action
space, e.g. `delta end-effector pose` or `absolute joint pose`.
- `policy.image_keys`: ordered LeRobot image observation keys passed to the
processor.
- `policy.model_dtype`: checkpoint/forward dtype, one of `float32`,
`bfloat16`, or `float16`. Use `bfloat16` for normal training.
- `policy.num_flow_timesteps`: number of flow-matching timesteps sampled per
example during training. We use `8` for fine-tuning.
- `policy.num_inference_steps`: optional override for continuous action
generation steps at inference time.
- `policy.gradient_checkpointing`: enables checkpointing in the VLM/action path
to reduce activation memory.
- `policy.freeze_embedding`: freezes input embeddings. The default is `true`.
- `policy.normalize_gripper`: controls whether gripper dimensions are included
in state/action quantile normalization. The default is `false`.
- `policy.normalize_language`: normalizes task strings before prompt
construction. The default is `true`.
- `policy.mask_action_dim_padding`: masks padded dimensions in the flow loss.
Released checkpoints use `policy.expected_max_action_dim=32`.
- `policy.max_sequence_length`: optional manual sequence cap. Leave unset to
infer it from images, state dimension, action dimension, action horizon, and
discrete-action mode.
### Learning Rates
MolmoAct2 uses parameter-group learning rates to match the original MolmoAct2
fine-tuning experiments.
- Full fine-tuning uses `policy.optimizer_lr=1e-5` for the VLM,
`policy.optimizer_vit_lr=5e-6` for the vision tower,
`policy.optimizer_connector_lr=5e-6` for image connector layers, and
`policy.optimizer_action_expert_lr=5e-5` for the action expert.
- LoRA VLM fine-tuning sets the VLM, vision, and connector LoRA parameter
groups to `5e-5` when `policy.enable_lora_vlm=true`. By default,
`policy.enable_lora_action_expert=false`, so the action expert is still fully
fine-tuned with `policy.optimizer_action_expert_lr`. If
`policy.enable_lora_action_expert=true`, the action expert is trained through
LoRA adapters instead.
- Action-expert-only fine-tuning trains only the action expert and uses
`policy.optimizer_action_expert_lr=5e-5`.
You can override the full fine-tuning and action-expert learning rates with
`policy.optimizer_lr`, `policy.optimizer_vit_lr`,
`policy.optimizer_connector_lr`, and `policy.optimizer_action_expert_lr`.
Scheduler settings can be changed with `policy.scheduler_warmup_steps`,
`policy.scheduler_decay_steps`, and `policy.scheduler_decay_lr`.
### Dataset Quantile Statistics
MolmoAct2 defaults to quantile normalization for state and action features. If
your dataset has not been converted with quantile statistics, you can add them
with:
```bash
python src/lerobot/scripts/augment_dataset_quantile_stats.py \
--repo-id=your_dataset
```
Alternatively, train MolmoAct2 with mean/std normalization:
```bash
--policy.normalization_mapping='{"ACTION": "MEAN_STD", "STATE": "MEAN_STD", "VISUAL": "IDENTITY"}'
```
## Evaluation
Evaluation also supports both LeRobot-saved checkpoints and original MolmoAct2
HF checkpoints. For LIBERO replication, keep the EGL rendering environment
fixed and use `policy.per_episode_seed=true`.
**Important:** We found that `num_steps_wait=10` does not reliably let the
LIBERO scene stabilize and can degrade measured success. All LIBERO evaluation
results reported here use `num_steps_wait=50`.
### Evaluation With LeRobot MolmoAct2 Weight
Use `policy.path` for a checkpoint saved by LeRobot. The saved processor and
normalization statistics are restored together with the model.
```bash
export MUJOCO_GL=egl
export PYOPENGL_PLATFORM=egl
export OMP_NUM_THREADS=1
export MKL_NUM_THREADS=1
lerobot-eval \
--policy.path=allenai/MolmoAct2-LIBERO-LeRobot \
--policy.inference_action_mode=continuous \
--policy.model_dtype=bfloat16 \
--policy.use_amp=true \
--policy.enable_inference_cuda_graph=true \
--policy.device=cuda \
--policy.per_episode_seed=true \
--policy.eval_seed=1000 \
--env.type=libero \
--env.task=libero_10,libero_goal,libero_object,libero_spatial \
--env.camera_name_mapping='{"agentview_image":"image","robot0_eye_in_hand_image":"wrist_image"}' \
--eval.batch_size=1 \
--eval.n_episodes=50 \
--seed=1000
```
### Evaluation With Original MolmoAct2 Weight
You can evaluate a released Hugging Face checkpoint directly without first
converting it to a LeRobot checkpoint. In this case, set
`policy.checkpoint_path` to the HF model repo and provide `policy.norm_tag`.
For LIBERO, `policy.norm_tag=libero` loads the LIBERO action/state
normalization statistics, action horizon, prompt metadata, and image-key order
from the checkpoint's `norm_stats.json`.
To fully replicate the MolmoAct2 paper results with released Hugging Face
checkpoints, we recommend using the v0.5.1-pinned
[`allenai/lerobot` `molmoact2-hf-inference`](https://github.com/allenai/lerobot/tree/molmoact2-hf-inference)
branch. That branch matches the original evaluation settings used for the
reported numbers.
```bash
export MUJOCO_GL=egl
export PYOPENGL_PLATFORM=egl
export OMP_NUM_THREADS=1
export MKL_NUM_THREADS=1
lerobot-eval \
--policy.type=molmoact2 \
--policy.checkpoint_path=allenai/MolmoAct2-LIBERO \
--policy.norm_tag=libero \
--policy.inference_action_mode=continuous \
--policy.model_dtype=float32 \
--policy.use_amp=false \
--policy.enable_inference_cuda_graph=true \
--policy.device=cuda \
--policy.per_episode_seed=true \
--policy.eval_seed=1000 \
--env.type=libero \
--env.task=libero_goal \
--env.camera_name_mapping='{"agentview_image":"image","robot0_eye_in_hand_image":"wrist_image"}' \
--eval.batch_size=1 \
--eval.n_episodes=50 \
--seed=1000
```
Use `--env.task=libero_10,libero_goal,libero_object,libero_spatial` to run the
full LIBERO suite. The same command works for other released MolmoAct2
checkpoints as long as the requested `policy.norm_tag` exists in that
checkpoint's `norm_stats.json`.
### Common Evaluation Options
- `policy.inference_action_mode`: required for rollout. Use `continuous` for
flow-matching inference or `discrete` for action-token inference. It must be
compatible with the training-time `policy.action_mode` saved in the
checkpoint.
- `policy.path`: LeRobot checkpoint path or Hub repo. Use this for checkpoints
saved by LeRobot.
- `policy.checkpoint_path`: original MolmoAct2 HF checkpoint path or Hub repo.
Use this with `policy.type=molmoact2` and `policy.norm_tag`.
- `policy.norm_tag`: selects normalization statistics, prompt metadata,
image-key order, and action horizon from the original checkpoint's
`norm_stats.json`. It is required for direct original-HF checkpoint
evaluation.
- `policy.model_dtype`: model load/forward dtype. Use `bfloat16` for normal
GPU evaluation. Use `float32` only when you explicitly want fp32 inference.
- `policy.use_amp`: runs the policy forward under autocast during eval. For
`model_dtype=bfloat16`, keep this enabled.
- `policy.enable_inference_cuda_graph`: enables the MolmoAct2 inference CUDA
graph path for faster repeated continuous-action rollout.
- `policy.per_episode_seed` and `policy.eval_seed`: make stochastic continuous
action generation deterministic per episode for replication.
- `env.task`: comma-separated LIBERO suites or a single suite. Use
`libero_10,libero_goal,libero_object,libero_spatial` for the full benchmark.
- `env.camera_name_mapping`: maps LIBERO camera names to the image keys expected
by the policy processor.
## Performance Results
### LIBERO Benchmark Results
MolmoAct2 has demonstrated strong performance on the LIBERO benchmark suite. To
compare and test its LeRobot implementation, we fine-tuned
[`allenai/MolmoAct2-LIBERO`](https://huggingface.co/allenai/MolmoAct2-LIBERO)
for an additional 10k steps on the LIBERO dataset with per-GPU batch size 32 on
8 H100 GPUs, then compared the results to the original MolmoAct2 reference
results.
The LeRobot fine-tuned checkpoint reported here is available at
[`allenai/MolmoAct2-LIBERO-LeRobot`](https://huggingface.co/allenai/MolmoAct2-LIBERO-LeRobot)
and was trained on
[`allenai/MolmoAct2-LIBERO-Dataset`](https://huggingface.co/datasets/allenai/MolmoAct2-LIBERO-Dataset).
| Benchmark | LeRobot Implementation | MolmoAct2 Original |
| -------------- | ---------------------: | -----------------: |
| LIBERO Spatial | 98.4% | 97.8% |
| LIBERO Object | 100.0% | 100.0% |
| LIBERO Goal | 98.0% | 97.8% |
| LIBERO 10 | 96.6% | 93.2% |
| Average | 98.25% | 97.20% |
These results demonstrate MolmoAct2's strong performance across diverse robotic
manipulation tasks. To reproduce them, follow the instructions in the LIBERO
evaluation section.
## Hardware Deployment (lerobot-rollout)
LeRobot-format checkpoints are available on the Hub for direct use with
`lerobot-rollout`. Each checkpoint uses specific camera names that must
match your robot's camera configuration.
### Camera naming convention
Each checkpoint expects specific `observation.images.*` keys.
If your robot cameras have different names, use `--rename_map` to map them:
| Checkpoint | Camera keys | Description |
| ----------------------------- | ---------------------- | ------------------------ |
| MolmoAct2-LIBERO-LeRobot | `image`, `wrist_image` | LIBERO sim cameras |
| MolmoAct2-BimanualYAM-LeRobot | `top`, `left`, `right` | YAM 3-camera setup |
| MolmoAct2-DROID-LeRobot | `cam0`, `cam1` | External + wrist |
| MolmoAct2-SO100_101-LeRobot | `cam0`, `cam1` | Primary + secondary view |
Example with an SO-100 robot using top and side cameras:
```bash
lerobot-rollout \
--policy.path=lerobot/MolmoAct2-SO100_101-LeRobot \
--rename_map='{"observation.images.top": "observation.images.cam0", "observation.images.side": "observation.images.cam1"}' \
--robot.type=so100_follower \
--robot.port=/dev/ttyACM0 \
--robot.cameras='{
top: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30},
side: {type: opencv, index_or_path: 2, width: 640, height: 480, fps: 30}
}' \
--task="pick up the red cube" --duration=30
```
To use a wrist camera instead, just change the rename mapping:
```bash
--rename_map='{"observation.images.top": "observation.images.cam0", "observation.images.wrist": "observation.images.cam1"}'
```
### Joint frame transform (SO-100/101 zero-shot)
<Tip warning={true}>
The MolmoAct2-SO100_101 checkpoint was trained on data that uses a different
joint calibration convention than LeRobot >= 0.5.0. Without a frame
correction, the arm may move in the wrong direction.
This affects both **zero-shot deployment** and **fine-tuning** from the
original checkpoint. The pretrained weights expect the old convention, so
all joint data (observations and actions) must be transformed to match.
The converted LeRobot checkpoint (`lerobot/MolmoAct2-SO100_101-LeRobot`)
already includes this correction in its processor pipeline. If you convert
or fine-tune the checkpoint yourself, set the following in the policy config (`configuration_molmoact2.py`):
- `joint_signs`: `[1, -1, 1, 1, 1, 1]` (flips shoulder_lift direction)
- `joint_offsets`: `[0, 90, 90, 0, 0, 0]` (shifts shoulder_lift and elbow_flex by 90°)
See the [backward compatibility guide](./backwardcomp) for details on the
calibration change.
</Tip>
## Differences From the Original Implementation
This LeRobot port is intended to match MolmoAct2 behavior while using LeRobot's
dataset, training, evaluation, checkpoint, and logging infrastructure. The main
differences from the original training repository are:
- The original paper training stack loads the model in fp32 and trains under
mixed precision. This LeRobot port usually loads the checkpoint directly in
`policy.model_dtype=bfloat16` for lower memory use.
- The original repository uses its own FSDP/model-parallel training path. The
LeRobot port uses the standard LeRobot/Accelerate training path and has not
been tested for multi-node training.
- The original repository supports sequence packing. The LeRobot port trains on
one LeRobot sample per item and pads to an inferred fixed sequence budget.
- The LeRobot port follows LeRobot's optimizer, scheduler, checkpoint saving,
dataset transforms, image augmentation, and Weights & Biases logging
conventions.
- The original training path supports mixed action horizons by padding to
`max_action_horizon` and masking padded horizon slots in the action expert
self-attention. This is useful when training across datasets with different
control frequencies. The LeRobot port currently targets single-dataset
fine-tuning, so `policy.chunk_size` overrides the checkpoint
`max_action_horizon` and horizon masking is not implemented yet. Support for
this mixed-horizon path is planned.
## Citation
```bibtex
@misc{fang2026molmoact2actionreasoningmodels,
title={MolmoAct2: Action Reasoning Models for Real-world Deployment},
author={Haoquan Fang and Jiafei Duan and Donovan Clay and Sam Wang and Shuo Liu and Weikai Huang and Xiang Fan and Wei-Chuan Tsai and Shirui Chen and Yi Ru Wang and Shanli Xing and Jaemin Cho and Jae Sung Park and Ainaz Eftekhar and Peter Sushko and Karen Farley and Angad Wadhwa and Cole Harrison and Winson Han and Ying-Chun Lee and Eli VanderBilt and Rose Hendrix and Suveen Ellawela and Lucas Ngoo and Joyce Chai and Zhongzheng Ren and Ali Farhadi and Dieter Fox and Ranjay Krishna},
year={2026},
eprint={2605.02881},
archivePrefix={arXiv},
primaryClass={cs.RO},
url={https://arxiv.org/abs/2605.02881},
}
```
## License
This model is licensed under Apache 2.0. It is intended for research and
educational use in accordance with
[Ai2's Responsible Use Guidelines](https://allenai.org/responsible-use),
consistent with [allenai/molmoact2](https://github.com/allenai/molmoact2).
docs/source/multi_gpu_training.mdx
+180
View File
@@ -0,0 +1,180 @@
# Multi-GPU Training
This guide shows you how to train policies on multiple GPUs using [Hugging Face Accelerate](https://huggingface.co/docs/accelerate).
## Installation
`accelerate` is included in the `training` extra. Install it with:
```bash
pip install 'lerobot[training]'
```
## Training with Multiple GPUs
You can launch training in two ways:
### Option 1: Without config (specify parameters directly)
You can specify all parameters directly in the command without running `accelerate config`:
```bash
accelerate launch \
--multi_gpu \
--num_processes=2 \
$(which lerobot-train) \
--dataset.repo_id=${HF_USER}/my_dataset \
--policy.type=act \
--policy.repo_id=${HF_USER}/my_trained_policy \
--output_dir=outputs/train/act_multi_gpu \
--job_name=act_multi_gpu \
--wandb.enable=true
```
**Key accelerate parameters:**
- `--multi_gpu`: Enable multi-GPU training
- `--num_processes=2`: Number of GPUs to use
- `--mixed_precision=fp16`: Use fp16 mixed precision (or `bf16` if supported)
### Option 2: Using accelerate config
If you prefer to save your configuration, you can optionally configure accelerate for your hardware setup by running:
```bash
accelerate config
```
This interactive setup will ask you questions about your training environment (number of GPUs, mixed precision settings, etc.) and saves the configuration for future use. For a simple multi-GPU setup on a single machine, you can use these recommended settings:
- Compute environment: This machine
- Number of machines: 1
- Number of processes: (number of GPUs you want to use)
- GPU ids to use: (leave empty to use all)
- Mixed precision: fp16 or bf16 (recommended for faster training)
Then launch training with:
```bash
accelerate launch $(which lerobot-train) \
--dataset.repo_id=${HF_USER}/my_dataset \
--policy.type=act \
--policy.repo_id=${HF_USER}/my_trained_policy \
--output_dir=outputs/train/act_multi_gpu \
--job_name=act_multi_gpu \
--wandb.enable=true
```
## How It Works
When you launch training with accelerate:
1. **Automatic detection**: LeRobot automatically detects if it's running under accelerate
2. **Data distribution**: Your batch is automatically split across GPUs
3. **Gradient synchronization**: Gradients are synchronized across GPUs during backpropagation
4. **Single process logging**: Only the main process logs to wandb and saves checkpoints
## Learning Rate and Training Steps Scaling
**Important:** LeRobot does **NOT** automatically scale learning rates or training steps based on the number of GPUs. This gives you full control over your training hyperparameters.
### Why No Automatic Scaling?
Many distributed training frameworks automatically scale the learning rate by the number of GPUs (e.g., `lr = base_lr × num_gpus`).
However, LeRobot keeps the learning rate exactly as you specify it.
### When and How to Scale
If you want to scale your hyperparameters when using multiple GPUs, you should do it manually:
**Learning Rate Scaling:**
```bash
# Example: 2 GPUs with linear LR scaling
# Base LR: 1e-4, with 2 GPUs -> 2e-4
accelerate launch --num_processes=2 $(which lerobot-train) \
--optimizer.lr=2e-4 \
--dataset.repo_id=lerobot/pusht \
--policy.type=act
```
**Training Steps Scaling:**
Since the effective batch size `bs` increases with multiple GPUs (batch_size × num_gpus), you may want to reduce the number of training steps proportionally:
```bash
# Example: 2 GPUs with effective batch size 2x larger
# Original: batch_size=8, steps=100000
# With 2 GPUs: batch_size=8 (16 in total), steps=50000
accelerate launch --num_processes=2 $(which lerobot-train) \
--batch_size=8 \
--steps=50000 \
--dataset.repo_id=lerobot/pusht \
--policy.type=act
```
## Training Large Models with FSDP
DDP replicates the full model on every GPU, so a model that doesn't fit on one GPU won't fit under
DDP either. For large models, use **FSDP** (Fully Sharded Data Parallel), which shards parameters,
gradients, and optimizer state across GPUs. See the [accelerate FSDP guide](https://huggingface.co/docs/accelerate/usage_guides/fsdp) for background.
An example on how to launch LeRobot training with FSDP across 4 GPUs (1 machine):
```bash
accelerate launch --config_file fsdp.yaml --num_processes=4 $(which lerobot-train) \
--dataset.repo_id=${HF_USER}/my_dataset \
--policy.type=<your_policy> \
--output_dir=outputs/train/my_policy_fsdp
```
A minimal `fsdp.yaml` (FSDP1; shards params/grads/optimizer — ZeRO-3-equivalent):
```yaml
compute_environment: LOCAL_MACHINE
distributed_type: FSDP
mixed_precision: bf16
num_machines: 1
num_processes: 4
fsdp_config:
fsdp_version: 1
fsdp_sharding_strategy: FULL_SHARD # params + grads + optimizer (ZeRO-3)
fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
fsdp_transformer_layer_cls_to_wrap: <YourTransformerBlock> # repeated block class to shard
fsdp_use_orig_params: true # required: optimizer is built pre-prepare
fsdp_state_dict_type: FULL_STATE_DICT
```
Set `fsdp_transformer_layer_cls_to_wrap` to your model's repeated transformer-block class so each
block is sharded as its own unit. `fsdp_use_orig_params: true` is required because LeRobot builds the
optimizer before `accelerator.prepare()`.
### FSDP checkpoints
LeRobot gathers the full state dict across all ranks and the main process writes it as a single
`model.safetensors`, loadable as usual with `Policy.from_pretrained(...)`. Two things to look out for:
- **Checkpoints store fp32 weights.** Under mixed precision (`bf16`/`fp16`) FSDP keeps an fp32 master
copy, and the checkpoint saves it (~2× the bf16 size on disk) so training can resume consistently
with the fp32 optimizer state; `from_pretrained` casts back to the policy dtype on load. FSDP-specific
caveat: an fp32 checkpoint is materialized in full precision on the target device _before_ casting,
so loading it for inference on a tight GPU can OOM even when the bf16 model would fit — load on CPU
first, or cast `model.safetensors` to the deployment dtype offline.
- The sharded optimizer state is gathered into a full (world-size-independent) state dict and saved
alongside the model in the same `optimizer_state.safetensors` / `optimizer_param_groups.json`
format as single-GPU training, so **resume-from-checkpoint is supported** with `--resume=true`.
Resume reshards both the model and the optimizer state to the _current_ FSDP topology, so you can
resume an FSDP checkpoint on a different number of GPUs. Note that the data sampler is only
sample-exact when the world size and batch size match the original run (a warning is logged
otherwise); the optimizer/model state itself is unaffected.
## Notes
- The `--policy.use_amp` flag in `lerobot-train` is only used when **not** running with accelerate. When using accelerate, mixed precision is controlled by accelerate's configuration.
- Training logs, checkpoints, and hub uploads are only done by the main process to avoid conflicts. Non-main processes have console logging disabled to prevent duplicate output.
- The effective batch size is `batch_size × num_gpus`. If you use 4 GPUs with `--batch_size=8`, your effective batch size is 32.
- Learning rate scheduling is handled correctly across multiple processes—LeRobot sets `step_scheduler_with_optimizer=False` to prevent accelerate from adjusting scheduler steps based on the number of processes.
- When saving or pushing models, LeRobot automatically unwraps the model from accelerate's distributed wrapper to ensure compatibility.
- WandB integration automatically initializes only on the main process, preventing multiple runs from being created.
For more advanced configurations and troubleshooting, see the [Accelerate documentation](https://huggingface.co/docs/accelerate). If you want to learn more about how to train on a large number of GPUs, checkout this awesome guide: [Ultrascale Playbook](https://huggingface.co/spaces/nanotron/ultrascale-playbook).
docs/source/multi_task_dit.mdx
+388
View File
@@ -0,0 +1,388 @@
# Multitask DiT Policy
Multitask Diffusion Transformer (DiT) Policy is an evolution of the original Diffusion Policy architecture, which leverages a large DiT with text and vision conditioning for multitask robot learning. This implementation supports both diffusion and flow matching objectives for action generation, enabling robots to perform diverse manipulation tasks conditioned on language instructions.
## Model Overview
The model uses:
- **CLIP Vision Encoder**: Processes RGB images from multiple camera views
- **CLIP Text Encoder**: Encodes language task instructions (frozen weights with learnable projection)
- **Diffusion Transformer**: Predicts action sequences conditioned on observations and language
- **Two Objectives**: Supports both diffusion (DDPM/DDIM) and flow matching for action generation
This model is exciting because you can achieve extremely high dexterity, competitive with multi-billion parameter
VLAs, with only ~450M parameters and significantly less training.
## Installation Requirements
Multitask DiT Policy has additional dependencies. Install it with:
```bash
pip install lerobot[multi_task_dit]
```
This will install all necessary dependencies including the HuggingFace Transformers library for CLIP models.
## Usage
To use Multitask DiT in your LeRobot configuration, specify the policy type as:
```python
policy.type=multi_task_dit
```
## Training
### Basic Training Command
Here's a complete training command for training Multitask DiT on your dataset:
```bash
lerobot-train \
--dataset.repo_id=YOUR_DATASET \
--output_dir=./outputs/multitask_dit_training \
--batch_size=32 \
--steps=5000 \
--save_freq=500 \
--log_freq=100 \
--policy.type=multi_task_dit \
--policy.device=cuda \
--policy.repo_id="HF_USER/multitask-dit-your-robot" \
--wandb.enable=true
```
### Recommended Hyperparameters and Dataset Details (30Hz Control Frequency)
For reliable performance, start with these suggested default hyperparameters:
```bash
lerobot-train \
--dataset.repo_id=YOUR_DATASET \
--output_dir=./outputs/mutitask_dit_training \
--batch_size=320 \
--steps=30000 \
--policy.type=multi_task_dit \
--policy.device=cuda \
--policy.horizon=32 \
--policy.n_action_steps=24 \
--policy.objective=diffusion \
--policy.noise_scheduler_type=DDPM \
--policy.num_train_timesteps=100 \
--policy.repo_id="HF_USER/multitask-dit-your-robot" \
--wandb.enable=true
```
**Key Parameters:**
- **Batch Size**: 192-320 - If you have access to a GPU that can support this, you will get the best training dynamics
- **Horizon**: 32 - number of action steps to predict, ~1.0 sec at 30Hz
- **n_action_steps**: 24 - ~0.8 seconds at 30Hz
- **Objective**: `diffusion` - start with diffusion and experiment with flow matching if generation quality is poor
- **Training Steps**: >30k steps recommended for a single task
### Training Configuration Parameters
#### Objective Selection
Choose between diffusion and flow matching:
```bash
# Diffusion objective (default)
--policy.objective=diffusion \
--policy.noise_scheduler_type=DDPM \ # or "DDIM"
--policy.num_train_timesteps=100 \
--policy.num_inference_steps=10 \ # For faster inference
--policy.beta_schedule=squaredcos_cap_v2 \ # Noise schedule type
--policy.prediction_type=epsilon \ # "epsilon" (predict noise) or "sample" (predict clean)
--policy.clip_sample=true \ # Clip samples during denoising
--policy.clip_sample_range=1.0 # Clipping range [-x, x]
# Flow matching objective
--policy.objective=flow_matching \
--policy.timestep_sampling_strategy=beta \ # or "uniform" | the beta sampling strategy performance appears much better in practice
--policy.num_integration_steps=100 \
--policy.integration_method=euler \ # or "rk4"
--policy.sigma_min=0.0 # Minimum noise in flow interpolation path
```
#### Transformer Architecture
Adjust model capacity based on dataset size:
```bash
# Small datasets (< 100 examples)
--policy.num_layers=4 \
--policy.hidden_dim=512 \
--policy.num_heads=8 # should ideally be hidden_dim // 64
# Medium datasets (100-5k examples) - default
--policy.num_layers=6 \
--policy.hidden_dim=512 \
--policy.num_heads=8 # should ideally be hidden_dim // 64
# Large datasets (> 5k examples)
--policy.num_layers=8 \
--policy.hidden_dim=512 \
--policy.num_heads=8 # should ideally be hidden_dim // 64
```
**Positional Encoding Options:**
The model supports two positional encoding methods for action sequences:
```bash
# Rotary Position Embedding (RoPE) - default, recommended
--policy.use_rope=true \
--policy.rope_base=10000.0 # Base frequency for RoPE
# Absolute positional encoding
--policy.use_positional_encoding=true # Disables RoPE when true
```
**Other Transformer Parameters:**
```bash
--policy.dropout=0.1 # Dropout rate for DiT blocks (0.0-1.0)
--policy.timestep_embed_dim=256 # Timestep embedding dimension
```
#### Vision Encoder Configuration
```bash
# Use different CLIP model for more expressivity at the cost of inference time
# experiment with larger or smaller models depending on the complexity of your tasks and size of dataset
--policy.vision_encoder_name=openai/clip-vit-large-patch14
# Use separate vision encoder per camera
# This may be useful when cameras have significantly different characteristics, but
# be wary of increased VRAM footprint.
--policy.use_separate_rgb_encoder_per_camera=true
# Image preprocessing
--policy.image_resize_shape=[XXX,YYY] \ # you may need to resize your images for inference speed ups
--policy.image_crop_shape=[224,224] \
--policy.image_crop_is_random=true # Random during training, center at inference
```
#### Text Encoder Configuration
```bash
# Use different CLIP text encoder model
# same as vision: experiment with larger or smaller models depending on the
# complexity of your tasks and size of dataset
--policy.text_encoder_name=openai/clip-vit-large-patch14
```
#### Learning Rate Configuration
The vision encoder uses a separate learning rate multiplier, where 1/10th is suggested to be the ideal staritng point:
```bash
--policy.optimizer_lr=2e-5 \
--policy.vision_encoder_lr_multiplier=0.1 # Vision encoder LR = 0.1 * optimizer_lr
```
### Training Tuning Guidelines
#### 1. Flow Matching with Beta Sampling
The original diffusion implementation here is based on the work described in [TRI's LBM paper](https://arxiv.org/abs/2507.05331)
Additionally, we have implemented a flow-matching objective, which is described at a high-level in [Boston Dynamics blog post](https://bostondynamics.com/blog/large-behavior-models-atlas-find-new-footing/).
Consider testing the flow-matching objective and evaluating performance differences for your task:
```bash
--policy.objective=flow_matching \
--policy.timestep_sampling_strategy=beta \
--policy.timestep_sampling_alpha=1.5 \
--policy.timestep_sampling_beta=1.0 \
--policy.timestep_sampling_s=0.999
```
This hasn't been shown to be a silver bullet across every user case, but it occasionally results in smoother and more consistent actions.
#### 2. Number of Transformer Layers
Match model capacity to your dataset size:
- **Small datasets** (< 100 examples): Reduce to 4 layers
- **Large datasets** (> 5k examples): Increase to 8 layers
#### 3. `horizon` Tuning
The model can be sensitive to the horizon you choose. Start with around a 1 second horizon based on your control frequency:
- **30 Hz frequency**: `horizon=30`
- **10 Hz frequency**: `horizon=10`
Then experiment with increasing from there. The horizon determines how far into the future the model predicts actions.
#### 4. `n_action_steps` Sensitivity
The model can also be very sensitive to `n_action_steps`. Start with it being around 0.8 seconds based on your control frequency and tune from there:
- **Lower values**: More reactive but potentially less stable for long-horizon tasks
- **Higher values**: Better for long-horizon execution but open-loop failures are limited in their recovery
### Inference Tuning
For faster inference, use DDIM with fewer sampling steps:
```bash
--policy.noise_scheduler_type=DDIM \
--policy.num_inference_steps=10
```
### Resuming Training
To resume training from a checkpoint:
```bash
lerobot-train \
--config_path=./outputs/mutitask_dit_training/checkpoints/last/pretrained_model/train_config.json \
--resume=true
```
The checkpoint directory should contain `model.safetensors` and `config.json` files (saved automatically during training). When resuming, the configuration is loaded from the checkpoint, so you don't need to specify other parameters.
## Common Failure Modes and Debugging
Training these models can be finicky. Here are common failure modes and debugging approaches:
### Idling / No Motion
The model may "collapse" during inference, resulting in static or no motion. This can occur when:
1. **Insufficient training data**: If you only have 20-50 examples, try to roughly double your dataset size. Once you have above 300 examples, if you're still seeing this, the task may be too complex.
2. **Multiple similar tasks**: When your dataset contains multiple similar tasks (e.g., picking up 2 different objects), the model may rely too heavily on language conditioning which might not be rich enough.
**Debugging tips:**
- Increase dataset size (double until you get to over 300 examples)
- Train for longer, up to 100k steps, even when the loss flatlines
- Check if the model is receiving proper language instructions or increase diversity of instruction
### Executing the Wrong Task
Sometimes the robot will completely ignore your instruction and perform some other task. This generally only happens if you have trained on multiple tasks.
**Potential causes:**
- Language instruction ambiguity
- Insufficient task-specific training data
- Model confusion between similar tasks in the multitask dataset
**Debugging tips:**
- Verify language instruction specificity, especially if descriptions are similar between multiple tasks
- Check task distribution in your training dataset and add weighting to the failing/ignored task
- Consider task-specific fine-tuning
### Training Instability
If training loss is unstable or diverging:
- Try adjusting learning rate between `1e-5` and `3e-4`
- Increase batch size if possible
- Check that your dataset normalization is correct
- Verify image preprocessing is working correctly
## Performance Considerations
### GPU Requirements
- **Inference**: At least an RTX 5070 Ti (or equivalent GPU) is recommended for reasonable speed performance
- **Training**: A GPU with enough VRAM to load batch sizes of >64 is ideal, which will vary depending on the number of image observations, etc
### Batch Size Recommendations
- **Minimum**: 64 (less than this may result in unstable training)
- **Recommended**: 256-320 (best performance, requires larger GPU)
## Example: Training on Custom Dataset
Here's a complete example training on a custom dataset:
```bash
lerobot-train \
--dataset.repo_id=YOUR_DATASET \
--output_dir=./outputs/mutitask_dit_training \
--batch_size=320 \
--steps=30000 \
--save_freq=1000 \
--log_freq=100 \
--env_eval_freq=1000 \
--policy.type=multi_task_dit \
--policy.device=cuda \
--policy.horizon=32 \
--policy.n_action_steps=24 \
--policy.objective=diffusion \
--policy.noise_scheduler_type=DDPM \
--policy.num_layers=6 \
--policy.hidden_dim=512 \
--policy.vision_encoder_name=openai/clip-vit-base-patch16 \
--policy.image_resize_shape=[320,240] \
--policy.image_crop_shape=[224,224] \
--policy.repo_id="HF_USER/multitask-dit-your-robot" \
--wandb.enable=true \
--wandb.project=multitask_dit
```
## Libero Results
```
python -m lerobot.scripts.lerobot_train \
--dataset.repo_id=HuggingFaceVLA/libero \
--policy.type=multi_task_dit \
--policy.push_to_hub=false \
--output_dir="./outputs/multitask_dit_libero" \
--job_name="multitask-dit-libero" \
--wandb.enable=true \
--wandb.project=multitask_dit_libero \
--dataset.image_transforms.enable=true \
--dataset.image_transforms.max_num_transforms=4 \
--dataset.image_transforms.tfs='{"brightness":{"type":"ColorJitter","kwargs":{"brightness":[0.75,1.25]}},"contrast":{"type":"ColorJitter","kwargs":{"contrast":[0.6,1.4]}},"saturation":{"type":"ColorJitter","kwargs":{"saturation":[0.8,1.2]}},"hue":{"type":"ColorJitter","kwargs":{"hue":[-0.05,0.05]}},"sharpness":{"type":"SharpnessJitter","kwargs":{"sharpness":[0.6,1.4]}},"rotation":{"type":"RandomRotation","kwargs":{"degrees":[-5,5]}},"translation":{"type":"RandomAffine","kwargs":{"degrees":0,"translate":[0.1,0.1]}}}' \
--dataset.video_backend=torchcodec \
--policy.use_amp=true \
--policy.horizon=48 \
--policy.n_obs_steps=2 \
--policy.use_rope=true \
--policy.use_positional_encoding=false \
--policy.hidden_dim=768 \
--policy.num_layers=8 \
--policy.num_heads=12 \
--policy.dropout=0.1 \
--policy.timestep_embed_dim=256 \
--policy.objective=diffusion \
--policy.optimizer_lr=3e-4 \
--policy.optimizer_weight_decay=0 \
--policy.scheduler_warmup_steps=0 \
--policy.vision_encoder_name=openai/clip-vit-base-patch16 \
--policy.image_resize_shape=[256,256] \
--policy.image_crop_is_random=true \
--policy.text_encoder_name=openai/clip-vit-base-patch16 \
--policy.vision_encoder_lr_multiplier=0.1 \
--policy.device=cuda \
--num_workers=8 \
--save_freq=4000 \
--log_freq=100 \
--steps=100000 \
--batch_size=320
```
Results:
| LIBERO Spatial | LIBERO Object | LIBERO Goal | LIBERO 10 | Average |
| -------------- | ------------- | ----------- | --------- | ------- |
| 87.0 | 98.2 | 93.8 | 83.2 | 90.6 |
## References
For more details on the technical implementation and architecture, see:
- [A Careful Examination of Large Behavior Models for Multitask Dexterous Manipulation](https://arxiv.org/abs/2507.05331)
- [Large Behavior Models and Atlas Find New Footing](https://bostondynamics.com/blog/large-behavior-models-atlas-find-new-footing/)
- [Dissecting and Open-Sourcing Multitask Diffusion Transformer Policy](https://brysonkjones.substack.com/p/dissecting-and-open-sourcing-multitask-diffusion-transformer-policy)
docs/source/notebooks.mdx
+29
View File
@@ -0,0 +1,29 @@
# 🤗 LeRobot Notebooks
This repository contains example notebooks for using LeRobot. These notebooks demonstrate how to train policies on real or simulation datasets using standardized policies.
---
### Training ACT
[ACT](https://huggingface.co/papers/2304.13705) (Action Chunking Transformer) is a transformer-based policy architecture for imitation learning that processes robot states and camera inputs to generate smooth, chunked action sequences.
We provide a ready-to-run Google Colab notebook to help you train ACT policies using datasets from the Hugging Face Hub, with optional logging to Weights & Biases.
| Notebook | Colab |
| :------------------------------------------------------------------------------------------------------ | :-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| [Train ACT with LeRobot](https://github.com/huggingface/notebooks/blob/main/lerobot/training-act.ipynb) | [![Open in Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/lerobot/training-act.ipynb) |
Expected training time for 100k steps: ~1.5 hours on an NVIDIA A100 GPU with batch size of `64`.
### Training SmolVLA
[SmolVLA](https://huggingface.co/papers/2506.01844) is a small but efficient Vision-Language-Action model. It is compact in size with 450 M-parameter and is developed by Hugging Face.
We provide a ready-to-run Google Colab notebook to help you train SmolVLA policies using datasets from the Hugging Face Hub, with optional logging to Weights & Biases.
| Notebook | Colab |
| :-------------------------------------------------------------------------------------------------------------- | :------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
| [Train SmolVLA with LeRobot](https://github.com/huggingface/notebooks/blob/main/lerobot/training-smolvla.ipynb) | [![Open in Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/lerobot/training-smolvla.ipynb) |
Expected training time for 20k steps: ~5 hours on an NVIDIA A100 GPU with batch size of `64`.
docs/source/omx.mdx
+197
View File
@@ -0,0 +1,197 @@
## Order and Assemble the parts
First, assemble the OMX hardware following the official assembly guide.
OMX Assembly Guide: https://ai.robotis.com/omx/assembly_guide_omx.html
OMX robots are shipped preconfigured from the factory. Motor IDs, communication parameters, and joint offsets are already set, so no additional motor setup or calibration is required before using LeRobot.
## Install LeRobot 🤗
To install LeRobot, follow our [Installation Guide](./installation)
In addition to these instructions, you need to install the Dynamixel SDK:
```bash
pip install -e ".[dynamixel]"
```
## Connect the robot
To find the port for each bus servo adapter, run this script:
```bash
lerobot-find-port
```
This command runs and when prompted, disconnect the USB cable from either the leader or follower arm and press Enter. The output will show 'The port of this MotorsBus is [port]'. This identifies the port for the disconnected arm. Repeat for the other arm to identify both ports.
<hfoptions id="find_port">
<hfoption id="Mac">
Example output on macOS:
```
Finding all available ports for the MotorBus.
['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
Remove the USB cable from your MotorsBus and press Enter when done.
[...Disconnect corresponding leader or follower arm and press Enter...]
The port of this MotorsBus is /dev/tty.usbmodem575E0032081
Reconnect the USB cable.
```
Where the found port is: `/dev/tty.usbmodem575E0032081` corresponding to your leader or follower arm.
</hfoption>
<hfoption id="Linux">
On Linux, we strongly recommend using udev rules to assign persistent and human-readable device names to the OMX leader and follower arms. This avoids issues where device names such as ttyACM0 and ttyACM1 change when the robot is unplugged, replugged, or when the system is rebooted.
#### 1. Find your device serial numbers
You should have obtained the port numbers like ../../ttyACM? for the leader and follower using `lerobot-find-port`. You can match those results with the serial numbers using the `ls -l /dev/serial/by-id/` command.
To create udev rules, you need the unique serial number for each OMX device. The easiest way is to list devices under:
```bash
ls -l /dev/serial/by-id/
```
You will see output similar to:
```bash
usb-ROBOTIS_OpenRB-150_228BDD7B503059384C2E3120FF0A2B19-if00 -> ../../ttyACM0
usb-ROBOTIS_OpenRB-150_67E1ED68503059384C2E3120FF092234-if00 -> ../../ttyACM1
```
In each line, the serial number is the long string after `usb-ROBOTIS_OpenRB-150_` and before `-if00`.
Follower serial: `228BDD7B503059384C2E3120FF0A2B19`
Leader serial: `67E1ED68503059384C2E3120FF092234`
#### 2. Create the udev rule
Create a new udev rule file:
```bash
sudo nano /etc/udev/rules.d/99-omx.rules
```
Paste the following lines, replacing the serial numbers with the values you found above:
```bash
SUBSYSTEM=="tty", ATTRS{idVendor}=="0403", ATTRS{serial}=="228BDD7B503059384C2E3120FF0A2B19", SYMLINK+="omx_follower"
SUBSYSTEM=="tty", ATTRS{idVendor}=="0403", ATTRS{serial}=="67E1ED68503059384C2E3120FF092234", SYMLINK+="omx_leader"
```
Save the file and reload udev rules:
```bash
sudo udevadm control --reload-rules
sudo udevadm trigger
```
Now unplug and replug both devices once.
#### 3. Verify the symlinks
Check that the persistent device names exist:
```bash
ls -l /dev/omx_follower /dev/omx_leader
```
You should see them pointing to ttyACM\* devices:
```bash
/dev/omx_follower -> ttyACM*
/dev/omx_leader -> ttyACM*
```
These names remain stable across reboots and reconnections.
</hfoption>
</hfoptions>
## Teleoperate
After identifying the correct ports, you can directly teleoperate the follower arm using the leader arm.
<hfoptions id="teleoperate">
<hfoption id="Mac">
### Teleoperate without camera
```bash
lerobot-teleoperate \
--robot.type=omx_follower \
--robot.port=<your_follower_port> \
--robot.id=omx_follower_arm \
--teleop.type=omx_leader \
--teleop.port=<your_leader_port> \
--teleop.id=omx_leader_arm
```
During teleoperation, motions of the leader arm are mirrored in real time by the follower arm. OMX is already preconfigured, teleoperation can begin immediately without any calibration steps.
### Teleoperate with camera
You can also enable camera input during teleoperation by providing a camera configuration for the follower arm.
```bash
lerobot-teleoperate \
--robot.type=omx_follower \
--robot.port=<your_follower_port> \
--robot.id=omx_follower_arm \
--robot.cameras="{front: {type: opencv, index_or_path: '/dev/video0', width: 640, height: 480, fps: 30}}" \
--teleop.type=omx_leader \
--teleop.port=<your_leader_port> \
--teleop.id=omx_leader_arm \
--display_data=true
```
When the camera is enabled, the camera stream is displayed in real time and synchronized with the robot state. This setup is useful for visual monitoring and can be reused later for demonstration recording and imitation learning.
</hfoption>
<hfoption id="Linux">
### Teleoperate without camera
```bash
lerobot-teleoperate \
--robot.type=omx_follower \
--robot.port=/dev/omx_follower \
--robot.id=omx_follower_arm \
--teleop.type=omx_leader \
--teleop.port=/dev/omx_leader \
--teleop.id=omx_leader_arm
```
During teleoperation, motions of the leader arm are mirrored in real time by the follower arm. OMX is already preconfigured, teleoperation can begin immediately without any calibration steps.
### Teleoperate with camera
You can also enable camera input during teleoperation by providing a camera configuration for the follower arm.
```bash
lerobot-teleoperate \
--robot.type=omx_follower \
--robot.port=/dev/omx_follower \
--robot.id=omx_follower_arm \
--robot.cameras="{front: {type: opencv, index_or_path: '/dev/video0', width: 640, height: 480, fps: 30}}" \
--teleop.type=omx_leader \
--teleop.port=/dev/omx_leader \
--teleop.id=omx_leader_arm \
--display_data=true
```
When the camera is enabled, the camera stream is displayed in real time and synchronized with the robot state. This setup is useful for visual monitoring and can be reused later for demonstration recording and imitation learning.
</hfoption>
</hfoptions>
Congrats 🎉, your robot is all set to learn a task on its own.
> If you have any questions or need help, please reach out on [Discord](https://discord.com/invite/robotis).
docs/source/openarm.mdx
+276
View File
@@ -0,0 +1,276 @@
# OpenArm
[OpenArm](https://openarm.dev) is an open-source 7DOF humanoid arm designed for physical AI research and deployment.
To get your OpenArm, assembled or DIY, and join the global community, browse verified and certified manufacturers worldwide at [openarm.dev](https://openarm.dev).
## What's Unique?
- **Human-Scale Design**: OpenArm is designed with human-like proportions, scaled for a person around 160-165cm tall. This provides an optimal balance between practical reach and manageable inertia for safe, responsive operation.
- **Safety-First Architecture**: Built with QDD backdrivable motors and high compliance, OpenArm prioritizes safe human-robot interaction while maintaining practical payload capabilities (6.0kg peak / 4.1kg nominal) for real-world tasks.
- **Built for Durability**: Critical structural components use aluminum and stainless steel construction, ensuring robust performance for repetitive data collection and continuous research use.
- **Fully Accessible & Buildable**: Every component, from CNC parts and 3D-printed casings to electrical wiring is designed to be purchasable and buildable by individual researchers and labs, with complete fabrication data provided.
- **Practical & Affordable**: At $6,500 USD for a complete bimanual system, OpenArm delivers research-grade capabilities at a fraction of traditional humanoid robot costs.
## Platform Requirements
<Tip warning={true}>
**Linux Only**: OpenArm currently only works on Linux. The CAN bus USB adapter
does not have macOS drivers and has not been tested on Windows.
</Tip>
## Safety Guide
Before operating OpenArm, please read the [official safety guide](https://docs.openarm.dev/getting-started/safety-guide). Key points:
- **Secure installation**: Fasten the arm to a flat, stable surface with screws or clamps
- **Safe distance**: Keep body parts and objects outside the range of motion during operation
- **Protective equipment**: Always wear safety goggles; use additional PPE as needed
- **Payload limits**: Do not exceed specified payload limits (6.0kg peak / 4.1kg nominal per arm)
- **Emergency stop**: Know the location and operation of the emergency stop device
- **Regular inspection**: Check for loose screws, damaged mechanical limits, unusual noises, and wiring damage
## Hardware Setup
Follow the official [OpenArm hardware documentation](https://docs.openarm.dev) for:
- Bill of materials and sourcing
- 3D printing instructions
- Mechanical assembly
- Electrical wiring
The hardware repositories are available at [github.com/enactic/openarm](https://github.com/enactic/openarm).
## CAN Bus Setup
OpenArm uses CAN bus communication with Damiao motors. Once you have the CAN bus USB adapter plugged into your Linux PC, follow the [Damiao Motors and CAN Bus guide](./damiao) to configure the interface.
Quick setup:
```bash
# Setup CAN interfaces
lerobot-setup-can --mode=setup --interfaces=can0,can1
# Test motor communication
lerobot-setup-can --mode=test --interfaces=can0,can1
```
## Install LeRobot 🤗
Follow our [Installation Guide](./installation), then install the Damiao motor support:
```bash
pip install -e ".[damiao]"
```
## Usage
### Follower Arm (Robot)
<hfoptions id="follower">
<hfoption id="Command">
```bash
lerobot-calibrate \
--robot.type=openarm_follower \
--robot.port=can0 \
--robot.side=right \
--robot.id=my_openarm_follower
```
</hfoption>
<hfoption id="API example">
```python
from lerobot.robots.openarm_follower import OpenArmFollower, OpenArmFollowerConfig
config = OpenArmFollowerConfig(
port="can0",
side="right", # or "left" for left arm
id="my_openarm_follower",
)
follower = OpenArmFollower(config)
follower.connect()
# Read current state
obs = follower.get_observation()
print(obs)
# Send action (position in degrees)
action = {
"joint_1.pos": 0.0,
"joint_2.pos": 0.0,
"joint_3.pos": 0.0,
"joint_4.pos": 45.0,
"joint_5.pos": 0.0,
"joint_6.pos": 0.0,
"joint_7.pos": 0.0,
"gripper.pos": 0.0,
}
follower.send_action(action)
follower.disconnect()
```
</hfoption>
</hfoptions>
### Leader Arm (Teleoperator)
The leader arm is used for teleoperation - manually moving it to control the follower arm.
<hfoptions id="leader">
<hfoption id="Command">
```bash
lerobot-calibrate \
--teleop.type=openarm_leader \
--teleop.port=can1 \
--teleop.id=my_openarm_leader
```
</hfoption>
<hfoption id="API example">
```python
from lerobot.teleoperators.openarm_leader import OpenArmLeader, OpenArmLeaderConfig
config = OpenArmLeaderConfig(
port="can1",
id="my_openarm_leader",
manual_control=True, # Disable torque for manual movement
)
leader = OpenArmLeader(config)
leader.connect()
# Read current position (as action to send to follower)
action = leader.get_action()
print(action)
leader.disconnect()
```
</hfoption>
</hfoptions>
### Teleoperation
To teleoperate OpenArm with leader-follower control:
```bash
lerobot-teleoperate \
--robot.type=openarm_follower \
--robot.port=can0 \
--robot.side=right \
--robot.id=my_follower \
--teleop.type=openarm_leader \
--teleop.port=can1 \
--teleop.id=my_leader
```
### Bimanual Teleoperation
To teleoperate a bimanual OpenArm setup with two leader and two follower arms:
```bash
lerobot-teleoperate \
--robot.type=bi_openarm_follower \
--robot.left_arm_config.port=can0 \
--robot.left_arm_config.side=left \
--robot.right_arm_config.port=can1 \
--robot.right_arm_config.side=right \
--robot.id=my_bimanual_follower \
--teleop.type=bi_openarm_leader \
--teleop.left_arm_config.port=can2 \
--teleop.right_arm_config.port=can3 \
--teleop.id=my_bimanual_leader
```
### Recording Data
To record a dataset during teleoperation:
```bash
lerobot-record \
--robot.type=openarm_follower \
--robot.port=can0 \
--robot.side=right \
--robot.id=my_follower \
--teleop.type=openarm_leader \
--teleop.port=can1 \
--teleop.id=my_leader \
--repo-id=my_hf_username/my_openarm_dataset \
--fps=30 \
--num-episodes=10
```
## Configuration Options
### Follower Configuration
| Parameter | Default | Description |
| --------------------- | --------- | ---------------------------------------------------------- |
| `port` | - | CAN interface (e.g., `can0`) |
| `side` | `None` | Arm side: `"left"`, `"right"`, or `None` for custom limits |
| `use_can_fd` | `True` | Enable CAN FD for higher data rates |
| `can_bitrate` | `1000000` | Nominal bitrate (1 Mbps) |
| `can_data_bitrate` | `5000000` | CAN FD data bitrate (5 Mbps) |
| `max_relative_target` | `None` | Safety limit for relative target positions |
| `position_kp` | Per-joint | Position control proportional gains |
| `position_kd` | Per-joint | Position control derivative gains |
### Leader Configuration
| Parameter | Default | Description |
| ------------------ | --------- | ----------------------------------- |
| `port` | - | CAN interface (e.g., `can1`) |
| `manual_control` | `True` | Disable torque for manual movement |
| `use_can_fd` | `True` | Enable CAN FD for higher data rates |
| `can_bitrate` | `1000000` | Nominal bitrate (1 Mbps) |
| `can_data_bitrate` | `5000000` | CAN FD data bitrate (5 Mbps) |
## Motor Configuration
OpenArm uses Damiao motors with the following default configuration:
| Joint | Motor Type | Send ID | Recv ID |
| --------------------------- | ---------- | ------- | ------- |
| joint_1 (Shoulder pan) | DM8009 | 0x01 | 0x11 |
| joint_2 (Shoulder lift) | DM8009 | 0x02 | 0x12 |
| joint_3 (Shoulder rotation) | DM4340 | 0x03 | 0x13 |
| joint_4 (Elbow flex) | DM4340 | 0x04 | 0x14 |
| joint_5 (Wrist roll) | DM4310 | 0x05 | 0x15 |
| joint_6 (Wrist pitch) | DM4310 | 0x06 | 0x16 |
| joint_7 (Wrist rotation) | DM4310 | 0x07 | 0x17 |
| gripper | DM4310 | 0x08 | 0x18 |
## Troubleshooting
### No Response from Motors
1. Check power supply connections
2. Verify CAN wiring (CAN-H, CAN-L, GND)
3. Run diagnostics: `lerobot-setup-can --mode=test --interfaces=can0`
4. See the [Damiao troubleshooting guide](./damiao#troubleshooting) for more details
### CAN Interface Not Found
Ensure the CAN interface is configured:
```bash
ip link show can0
```
## Resources
- [OpenArm Website](https://openarm.dev)
- [OpenArm Documentation](https://docs.openarm.dev)
- [OpenArm GitHub](https://github.com/enactic/openarm)
- [Safety Guide](https://docs.openarm.dev/getting-started/safety-guide)
- [Damiao Motors and CAN Bus](./damiao)
docs/source/peft_training.mdx
+64
View File
@@ -0,0 +1,64 @@
# Parameter efficient fine-tuning with 🤗 PEFT
[🤗 PEFT](https://github.com/huggingface/peft) (Parameter-Efficient Fine-Tuning) is a library for efficiently adapting
large pretrained models such as pre-trained policies (e.g., SmolVLA, π₀, ...) to new tasks without training all
of the model's parameters while yielding comparable performance.
Install the `lerobot[peft]` optional package to enable PEFT support.
To read about all the possible methods of adaption, please refer to the [🤗 PEFT docs](https://huggingface.co/docs/peft/index).
## Training SmolVLA
In this section we'll show you how to train a pre-trained SmolVLA policy with PEFT on the libero dataset.
For brevity we're only training on the `libero_spatial` subset. We will use `lerobot/smolvla_base` as the model
to parameter efficiently fine-tune:
```
lerobot-train \
--policy.path=lerobot/smolvla_base \
--policy.repo_id=your_hub_name/my_libero_smolvla \
--dataset.repo_id=HuggingFaceVLA/libero \
--policy.output_features=null \
--policy.input_features=null \
--policy.optimizer_lr=1e-3 \
--policy.scheduler_decay_lr=1e-4 \
--env.type=libero \
--env.task=libero_spatial \
--steps=100000 \
--batch_size=32 \
--peft.method_type=LORA \
--peft.r=64 \
--peft.lora_alpha=64
```
Note the `--peft.method_type` parameter that let's you select which PEFT method to use. Here we use
[LoRA](https://huggingface.co/docs/peft/main/en/package_reference/lora) (Low-Rank Adapter) which is probably the most
popular fine-tuning method to date. Low-rank adaption means that we only fine-tune a matrix with comparably low rank
instead of the full weight matrix. This rank can be specified using the `--peft.r` parameter, and the LoRA scaling factor with
`--peft.lora_alpha` (where `scaling = lora_alpha / r`). The higher the rank
the closer you get to full fine-tuning
There are more complex methods that have more parameters. These are not yet supported, feel free to raise an issue
if you want to see a specific PEFT method supported.
By default, PEFT will target the `q_proj` and `v_proj` layers of the LM expert in SmolVLA. It will also target the
state and action projection matrices as they are most likely task-dependent. If you need to target different layers
you can use `--peft.target_modules` to specify which layers to target. You can refer to the respective PEFT method's
documentation to see what inputs are supported, (e.g., [LoRA's target_modules documentation](https://huggingface.co/docs/peft/main/en/package_reference/lora#peft.LoraConfig.target_modules)).
Usually a list of suffixes or a regex are supported. For example, to target the MLPs of the `lm_expert` instead of
the `q` and `v` projections, use:
```
--peft.target_modules='(model\.vlm_with_expert\.lm_expert\..*\.(down|gate|up)_proj|.*\.(state_proj|action_in_proj|action_out_proj|action_time_mlp_in|action_time_mlp_out))'
```
In case you need to fully fine-tune a layer instead of just adapting it, you can supply a list of layer suffixes
to the `--peft.full_training_modules` parameter:
```
--peft.full_training_modules=["state_proj"]
```
The learning rate and the scheduled target learning rate can usually be scaled by a factor of 10 compared to the
learning rate used for full fine-tuning (e.g., 1e-4 normal, so 1e-3 using LoRA).
docs/source/phone_teleop.mdx
+196
View File
@@ -0,0 +1,196 @@
# Phone
Use your phone (iOS or Android) to control your robot.
**In this guide you'll learn:**
- How to connect an iOS/Android phone
- How phone pose is mapped to robot end‑effector (EE) targets
- How to tweak safety limits, gripper control, and IK settings
To use phone to control your robot, install the relevant dependencies with:
```bash
pip install lerobot[phone]
```
## Get started
### Supported platforms
- iOS: Uses the HEBI Mobile I/O app (ARKit pose + buttons). Download the app first, open it and the examples will discover it on your network and stream the phone pose and inputs.
- Android: Uses the `teleop` package (WebXR). When you start the Python process, it prints a local URL. Open the link on your phone, tap Start, then use Move to stream pose.
Links:
- Android WebXR library: [`teleop` on PyPI](https://pypi.org/project/teleop/)
- iOS app: [HEBI Mobile I/O](https://docs.hebi.us/tools.html#mobile-io)
### Phone orientation and controls
- Orientation: hold the phone with the screen facing up and the top edge pointing in the same direction as the robot gripper. This ensures calibration aligns the phone’s frame with the robot frame so motion feels natural, see the image below for reference.
- Enable/disable:
- iOS: Hold `B1` to enable teleoperation, release to stop. The first press captures a reference pose.
- Android: Press and hold the `Move` button, release to stop. The first press captures a reference pose.
- Gripper control:
- iOS: Analog input `A3` controls the gripper as velocity input.
- Android: Buttons `A` and `B` act like increment/decrement (A opens, B closes). You can tune velocity in the `GripperVelocityToJoint` step.
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/phone_teleop.webp" alt="Phone teleop orientation" title="Phone teleop orientation" width="40%">
### Step 1: Choose the platform
Modify the examples to use `PhoneOS.IOS` or `PhoneOS.ANDROID` in `PhoneConfig`. The API is identical across platforms, only the input source differs. All examples are under `examples/` and have `phone_so100_*.py` variants.
Teleoperation example:
```python
from lerobot.teleoperators.phone import Phone, PhoneConfig
from lerobot.teleoperators.phone.config_phone import PhoneOS
teleop_config = PhoneConfig(phone_os=PhoneOS.IOS) # or PhoneOS.ANDROID
teleop_device = Phone(teleop_config)
```
### Step 2: Connect and calibrate
When `Phone(teleop_config)` is created and `connect()` is called, calibration is prompted automatically. Hold the phone in the orientation described above, then:
- iOS: press and hold `B1` to capture the reference pose.
- Android: press `Move` button on the WebXR page to capture the reference pose.
Why calibrate? We capture the current pose so subsequent poses are expressed in a robot aligned frame. When you again press the button to enable control, the position is recaptured to avoid drift when your phone is repositioned while it was disabled.
### Step 3: Run an example
Run on of the examples scripts to teleoperate, record a dataset, replay a dataset or evaluate a policy.
All scripts assume you configured your robot (e.g., SO-100 follower) and set the correct serial port.
Additionally you need to **copy the URDF of the robot into the examples folder**. For the examples in this tutorial (using SO100/SO101), copy the `SO101` folder from the [SO-ARM100 repo](https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101) into the `examples/phone_to_so100/` directory, so that the URDF file path becomes `examples/phone_to_so100/SO101/so101_new_calib.urdf`.
- Run this example to teleoperate:
```bash
cd examples/phone_to_so100
python teleoperate.py
```
After running the example:
- Android: after starting the script, open the printed local URL on your phone, tap Start, then press and hold Move.
- iOS: open HEBI Mobile I/O first; B1 enables motion. A3 controls the gripper.
Additionally you can customize mapping or safety limits by editing the processor steps shown in the examples. You can also remap inputs (e.g., use a different analog input) or adapt the pipeline to other robots (e.g., LeKiwi) by modifying the input and kinematics steps. More about this in the [Processors for Robots and Teleoperators](./processors_robots_teleop) guide.
- Run this example to record a dataset, which saves absolute end effector observations and actions:
```bash
cd examples/phone_to_so100
python record.py
```
- Run this example to replay recorded episodes:
```bash
cd examples/phone_to_so100
python replay.py
```
- Run this example to evaluate a pretrained policy:
```bash
cd examples/phone_to_so100
python evaluate.py
```
### Important pipeline steps and options
- Kinematics are used in multiple steps. We use [Placo](https://github.com/Rhoban/placo) which is a wrapper around Pinocchio for handling our kinematics. We construct the kinematics object by passing the robot's URDF and target frame. We set `target_frame_name` to the gripper frame.
```python
kinematics_solver = RobotKinematics(
urdf_path="./SO101/so101_new_calib.urdf",
target_frame_name="gripper_frame_link",
joint_names=list(robot.bus.motors.keys()),
)
```
- The `MapPhoneActionToRobotAction` step converts the calibrated phone pose and inputs into target deltas and gripper commands, below is shown what the step outputs.
```python
action["enabled"] = enabled
action["target_x"] = -pos[1] if enabled else 0.0
action["target_y"] = pos[0] if enabled else 0.0
action["target_z"] = pos[2] if enabled else 0.0
action["target_wx"] = rotvec[1] if enabled else 0.0
action["target_wy"] = rotvec[0] if enabled else 0.0
action["target_wz"] = -rotvec[2] if enabled else 0.0
action["gripper_vel"] = gripper_vel # Still send gripper action when disabled
```
- The `EEReferenceAndDelta` step converts target deltas to an absolute desired EE pose, storing a reference on enable, the `end_effector_step_sizes` are the step sizes for the EE pose and can be modified to change the motion speed.
```python
EEReferenceAndDelta(
kinematics=kinematics_solver,
end_effector_step_sizes={"x": 0.5, "y": 0.5, "z": 0.5},
motor_names=list(robot.bus.motors.keys()),
use_latched_reference=True,
),
```
- The `EEBoundsAndSafety` step clamps EE motion to a workspace and checks for large ee step jumps to ensure safety. The `end_effector_bounds` are the bounds for the EE pose and can be modified to change the workspace. The `max_ee_step_m` are the step limits for the EE pose and can be modified to change the safety limits.
```python
EEBoundsAndSafety(
end_effector_bounds={"min": [-1.0, -1.0, -1.0], "max": [1.0, 1.0, 1.0]},
max_ee_step_m=0.10,
)
```
- The `GripperVelocityToJoint` step turns a velocity‑like gripper input into absolute gripper position using the current measured state. The `speed_factor` is the factor by which the velocity is multiplied.
```python
GripperVelocityToJoint(speed_factor=20.0)
```
#### Different IK initial guesses
We use different IK initial guesses in the kinematic steps. As initial guess either the current measured joints or the previous IK solution is used.
- Closed loop (used in record/eval): sets `initial_guess_current_joints=True` so IK starts from the measured joints each frame.
```python
InverseKinematicsEEToJoints(
kinematics=kinematics_solver,
motor_names=list(robot.bus.motors.keys()),
initial_guess_current_joints=True, # closed loop
)
```
- Open loop (used in replay): sets `initial_guess_current_joints=False` so IK continues from the previous IK solution rather than the measured state. This preserves action stability when we replay without feedback.
```python
InverseKinematicsEEToJoints(
kinematics=kinematics_solver,
motor_names=list(robot.bus.motors.keys()),
initial_guess_current_joints=False, # open loop
)
```
### Pipeline steps explained
- MapPhoneActionToRobotAction: converts calibrated phone pose and inputs into target deltas and a gripper command. Motion is gated by an enable signal (B1 on iOS, Move on Android).
- EEReferenceAndDelta: latches a reference EE pose on enable and combines it with target deltas to produce an absolute desired EE pose each frame. When disabled, it keeps sending the last commanded pose.
- EEBoundsAndSafety: clamps the EE pose to a workspace and rate‑limits jumps for safety. Also declares `action.ee.*` features.
- InverseKinematicsEEToJoints: turns an EE pose into joint positions with IK. `initial_guess_current_joints=True` is recommended for closed‑loop control; set `False` for open‑loop replay for stability.
- GripperVelocityToJoint: integrates a velocity‑like gripper input into an absolute gripper position using the current measured state.
- ForwardKinematicsJointsToEE: computes `observation.state.ee.*` from observed joints for logging and training on EE state.
### Troubleshooting
- iOS not discovered: ensure HEBI Mobile I/O is open and your laptop/phone are on the same network.
- Android URL not reachable: check local you used `https` instead of `http`, use the exact IP printed by the script and allow your browser to enter and ignore the certificate issue.
- Motion feels inverted: adjust the sign flips in `MapPhoneActionToRobotAction` or swap axes to match your setup.
docs/source/pi0.mdx
+135
View File
@@ -0,0 +1,135 @@
# π₀ (Pi0)
π₀ is a **Vision-Language-Action model for general robot control**, from Physical Intelligence. The LeRobot implementation is adapted from their open source [OpenPI](https://github.com/Physical-Intelligence/openpi) repository.
## Model Overview
π₀ represents a breakthrough in robotics as the first general-purpose robot foundation model developed by [Physical Intelligence](https://www.physicalintelligence.company/blog/pi0). Unlike traditional robot programs that are narrow specialists programmed for repetitive motions, π₀ is designed to be a generalist policy that can understand visual inputs, interpret natural language instructions, and control a variety of different robots across diverse tasks.
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/lerobot-pi0%20(1).png"
alt="An overview of Pi0"
width="85%"
/>
### The Vision for Physical Intelligence
As described by Physical Intelligence, while AI has achieved remarkable success in digital domains, from chess-playing to drug discovery, human intelligence still dramatically outpaces AI in the physical world. To paraphrase Moravec's paradox, winning a game of chess represents an "easy" problem for AI, but folding a shirt or cleaning up a table requires solving some of the most difficult engineering problems ever conceived. π₀ represents a first step toward developing artificial physical intelligence that enables users to simply ask robots to perform any task they want, just like they can with large language models.
### Architecture and Approach
π₀ combines several key innovations:
- **Flow Matching**: Uses a novel method to augment pre-trained VLMs with continuous action outputs via flow matching (a variant of diffusion models)
- **Cross-Embodiment Training**: Trained on data from 8 distinct robot platforms including UR5e, Bimanual UR5e, Franka, Bimanual Trossen, Bimanual ARX, Mobile Trossen, and Mobile Fibocom
- **Internet-Scale Pre-training**: Inherits semantic knowledge from a pre-trained 3B parameter Vision-Language Model
- **High-Frequency Control**: Outputs motor commands at up to 50 Hz for real-time dexterous manipulation
## Installation Requirements
1. Install LeRobot by following our [Installation Guide](./installation).
2. Install Pi0 dependencies by running:
```bash
pip install -e ".[pi]"
```
## Training Data and Capabilities
π₀ is trained on the largest robot interaction dataset to date, combining three key data sources:
1. **Internet-Scale Pre-training**: Vision-language data from the web for semantic understanding
2. **Open X-Embodiment Dataset**: Open-source robot manipulation datasets
3. **Physical Intelligence Dataset**: Large and diverse dataset of dexterous tasks across 8 distinct robots
## Usage
To use π₀ in LeRobot, specify the policy type as:
```python
policy.type=pi0
```
## Training
For training π₀, you can use the standard LeRobot training script with the appropriate configuration:
```bash
lerobot-train \
--dataset.repo_id=your_dataset \
--policy.type=pi0 \
--output_dir=./outputs/pi0_training \
--job_name=pi0_training \
--policy.pretrained_path=lerobot/pi0_base \
--policy.repo_id=your_repo_id \
--policy.compile_model=true \
--policy.gradient_checkpointing=true \
--policy.dtype=bfloat16 \
--policy.freeze_vision_encoder=false \
--policy.train_expert_only=false \
--steps=3000 \
--policy.device=cuda \
--batch_size=32
```
### Key Training Parameters
- **`--policy.compile_model=true`**: Enables model compilation for faster training
- **`--policy.gradient_checkpointing=true`**: Reduces memory usage significantly during training
- **`--policy.dtype=bfloat16`**: Use mixed precision training for efficiency
- **`--batch_size=32`**: Batch size for training, adapt this based on your GPU memory
- **`--policy.pretrained_path=lerobot/pi0_base`**: The base π₀ model you want to finetune, options are:
- [lerobot/pi0_base](https://huggingface.co/lerobot/pi0_base)
- [lerobot/pi0_libero](https://huggingface.co/lerobot/pi0_libero) (specifically trained on the Libero dataset)
### Training Parameters Explained
| Parameter | Default | Description |
| ----------------------- | ------- | ------------------------------------------- |
| `freeze_vision_encoder` | `false` | Do not freeze the vision encoder |
| `train_expert_only` | `false` | Do not freeze the VLM, train all parameters |
**💡 Tip**: Setting `train_expert_only=true` freezes the VLM and trains only the action expert and projections, allowing finetuning with reduced memory usage.
## Relative Actions
By default, π₀ predicts absolute actions. You can enable **relative actions** so the model predicts offsets relative to the current robot state. This can improve training stability for certain setups.
To use relative actions, first recompute your dataset stats in relative space via the CLI:
```bash
lerobot-edit-dataset \
--repo_id your_dataset \
--operation.type recompute_stats \
--operation.relative_action true \
--operation.chunk_size 50 \
--operation.relative_exclude_joints "['gripper']" \
--push_to_hub true
```
Or equivalently in Python:
```python
from lerobot.datasets import LeRobotDataset, recompute_stats
dataset = LeRobotDataset("your_dataset")
recompute_stats(dataset, relative_action=True, chunk_size=50, relative_exclude_joints=["gripper"])
dataset.push_to_hub()
```
The `chunk_size` should match your policy's `chunk_size` (default 50 for π₀). `relative_exclude_joints` lists joint names that should remain in absolute space (e.g. gripper commands). Use `--push_to_hub true` to upload the updated stats to the Hub.
Then train with relative actions enabled:
```bash
lerobot-train \
--dataset.repo_id=your_dataset \
--policy.type=pi0 \
--policy.use_relative_actions=true \
--policy.relative_exclude_joints='["gripper"]' \
...
```
## License
This model follows the **Apache 2.0 License**, consistent with the original [OpenPI repository](https://github.com/Physical-Intelligence/openpi).
docs/source/pi05.mdx
+157
View File
@@ -0,0 +1,157 @@
# π₀.₅ (Pi05) Policy
π₀.₅ is a **Vision-Language-Action model with open-world generalization**, from Physical Intelligence. The LeRobot implementation is adapted from their open source [OpenPI](https://github.com/Physical-Intelligence/openpi) repository.
## Model Overview
π₀.₅ represents a significant evolution from π₀, developed by [Physical Intelligence](https://www.physicalintelligence.company/blog/pi05) to address a big challenge in robotics: **open-world generalization**. While robots can perform impressive tasks in controlled environments, π₀.₅ is designed to generalize to entirely new environments and situations that were never seen during training.
### The Generalization Challenge
As Physical Intelligence explains, the fundamental challenge isn't performing tasks of agility or dexterity, but generalization, the ability to correctly perform tasks in new settings with new objects. Consider a robot cleaning different homes: each home has different objects in different places. Generalization must occur at multiple levels:
- **Physical Level**: Understanding how to pick up a spoon (by the handle) or plate (by the edge), even with unseen objects in cluttered environments
- **Semantic Level**: Understanding task semantics, where to put clothes and shoes (laundry hamper, not on the bed), and what tools are appropriate for cleaning spills
- **Environmental Level**: Adapting to "messy" real-world environments like homes, grocery stores, offices, and hospitals
### Co-Training on Heterogeneous Data
The breakthrough innovation in π₀.₅ is **co-training on heterogeneous data sources**. The model learns from:
1. **Multimodal Web Data**: Image captioning, visual question answering, object detection
2. **Verbal Instructions**: Humans coaching robots through complex tasks step-by-step
3. **Subtask Commands**: High-level semantic behavior labels (e.g., "pick up the pillow" for an unmade bed)
4. **Cross-Embodiment Robot Data**: Data from various robot platforms with different capabilities
5. **Multi-Environment Data**: Static robots deployed across many different homes
6. **Mobile Manipulation Data**: ~400 hours of mobile robot demonstrations
This diverse training mixture creates a "curriculum" that enables generalization across physical, visual, and semantic levels simultaneously.
## Installation Requirements
1. Install LeRobot by following our [Installation Guide](./installation).
2. Install Pi0.5 dependencies by running:
```bash
pip install -e ".[pi]"
```
## Usage
To use π₀.₅ in your LeRobot configuration, specify the policy type as:
```python
policy.type=pi05
```
## Training
### Training Command Example
Here's a complete training command for finetuning the base π₀.₅ model on your own dataset:
```bash
lerobot-train \
--dataset.repo_id=your_dataset \
--policy.type=pi05 \
--output_dir=./outputs/pi05_training \
--job_name=pi05_training \
--policy.repo_id=your_repo_id \
--policy.pretrained_path=lerobot/pi05_base \
--policy.compile_model=true \
--policy.gradient_checkpointing=true \
--wandb.enable=true \
--policy.dtype=bfloat16 \
--policy.freeze_vision_encoder=false \
--policy.train_expert_only=false \
--steps=3000 \
--policy.device=cuda \
--batch_size=32
```
### Key Training Parameters
- **`--policy.compile_model=true`**: Enables model compilation for faster training
- **`--policy.gradient_checkpointing=true`**: Reduces memory usage significantly during training
- **`--policy.dtype=bfloat16`**: Use mixed precision training for efficiency
- **`--batch_size=32`**: Batch size for training, adapt this based on your GPU memory
- **`--policy.pretrained_path=lerobot/pi05_base`**: The base π₀.₅ model you want to finetune, options are:
- [lerobot/pi05_base](https://huggingface.co/lerobot/pi05_base)
- [lerobot/pi05_libero](https://huggingface.co/lerobot/pi05_libero) (specifically trained on the Libero dataset)
### Training Parameters Explained
| Parameter | Default | Description |
| ----------------------- | ------- | ------------------------------------------- |
| `freeze_vision_encoder` | `false` | Do not freeze the vision encoder |
| `train_expert_only` | `false` | Do not freeze the VLM, train all parameters |
**💡 Tip**: Setting `train_expert_only=true` freezes the VLM and trains only the action expert and projections, allowing finetuning with reduced memory usage.
If your dataset is not converted with `quantiles`, you can convert it with the following command:
```bash
python src/lerobot/scripts/augment_dataset_quantile_stats.py \
--repo-id=your_dataset \
```
Or train pi05 with this normalization mapping: `--policy.normalization_mapping='{"ACTION": "MEAN_STD", "STATE": "MEAN_STD", "VISUAL": "IDENTITY"}'`
## Relative Actions
By default, π₀.₅ predicts absolute actions. You can enable **relative actions** so the model predicts offsets relative to the current robot state. This can improve training stability for certain setups.
To use relative actions, first recompute your dataset stats in relative space via the CLI:
```bash
lerobot-edit-dataset \
--repo_id your_dataset \
--operation.type recompute_stats \
--operation.relative_action true \
--operation.chunk_size 50 \
--operation.relative_exclude_joints "['gripper']" \
--push_to_hub true
```
Or equivalently in Python:
```python
from lerobot.datasets import LeRobotDataset, recompute_stats
dataset = LeRobotDataset("your_dataset")
recompute_stats(dataset, relative_action=True, chunk_size=50, relative_exclude_joints=["gripper"])
dataset.push_to_hub()
```
The `chunk_size` should match your policy's `chunk_size` (default 50 for π₀.₅). `relative_exclude_joints` lists joint names that should remain in absolute space (e.g. gripper commands). Use `--push_to_hub true` to upload the updated stats to the Hub.
Then train with relative actions enabled:
```bash
lerobot-train \
--dataset.repo_id=your_dataset \
--policy.type=pi05 \
--policy.use_relative_actions=true \
--policy.relative_exclude_joints='["gripper"]' \
...
```
## Performance Results
### Libero Benchmark Results
π₀.₅ has demonstrated strong performance on the Libero benchmark suite. To compare and test its LeRobot implementation, we finetuned the libero base model for an additional 6k steps on the Libero dataset and compared the results to the OpenPI reference results.
| Benchmark | LeRobot Implementation | OpenPI Reference |
| ------------------ | ---------------------- | ---------------- |
| **Libero Spatial** | 97.0% | 98.8% |
| **Libero Object** | 99.0% | 98.2% |
| **Libero Goal** | 98.0% | 98.0% |
| **Libero 10** | 96.0% | 92.4% |
| **Average** | 97.5% | 96.85% |
These results demonstrate π₀.₅'s strong generalization capabilities across diverse robotic manipulation tasks. To reproduce these results, you can follow the instructions in the [Libero](https://huggingface.co/docs/lerobot/libero) section.
## License
This model follows the **Apache 2.0 License**, consistent with the original [OpenPI repository](https://github.com/Physical-Intelligence/openpi).
docs/source/pi0fast.mdx
+241
View File
@@ -0,0 +1,241 @@
# π₀-FAST (Pi0-FAST)
π₀-FAST is a **Vision-Language-Action model for general robot control** that uses autoregressive next-token prediction to model continuous robot actions.
## Model Overview
π₀-FAST combines the power of Vision-Language Models with a novel action tokenization approach called **FAST (Frequency-space Action Sequence Tokenization)**. This enables training autoregressive VLAs on highly dexterous tasks that are impossible with standard binning-based discretization, while training **up to 5x faster** than diffusion-based approaches like π₀.
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/lerobot-pifast.png"
alt="An overview of Pi0-FAST"
width="85%"
/>
### Why FAST?
Standard approaches for robot action tokenization use simple per-dimension, per-timestep binning schemes. While passable for simple behaviors, this rapidly breaks down for complex and dexterous skills that require precision and high-frequency control.
FAST solves this by compressing action sequences using signal processing techniques, resulting in a dense sequence of action tokens that can be predicted autoregressively—just like language tokens.
### How FAST Tokenization Works
The FAST tokenizer compresses action sequences through the following steps:
1. **Normalize**: Take a continuous action chunk of shape `(H, D)` where `H` is the horizon and `D` is the action dimension. Normalize using one of the supported normalization methods (Quantiles recommended to handle outliers).
2. **Discrete Cosine Transform (DCT)**: Apply DCT (via scipy) to each action dimension separately. DCT is a compression algorithm commonly used in image and audio codecs (JPEG, MP3).
3. **Quantization**: Round and remove insignificant coefficients for each action dimension, producing a sparse frequency matrix.
4. **Flatten**: Flatten the matrix into a 1D vector, with low-frequency components first.
5. **Byte Pair Encoding (BPE)**: Train a BPE tokenizer to compress the DCT coefficients into dense action tokens, typically achieving **10x compression** over prior tokenization approaches.
This approach can transform **any existing VLM** into a VLA by training it to predict these FAST tokens.
## Installation Requirements
1. Install LeRobot by following our [Installation Guide](./installation).
2. Install π₀-FAST dependencies by running:
```bash
pip install -e ".[pi]"
```
## Training a Custom FAST Tokenizer
You have two options for the FAST tokenizer:
1. **Use the pre-trained tokenizer**: The `lerobot/fast-action-tokenizer` tokenizer was trained on 1M+ real robot action sequences and works as a general-purpose tokenizer.
2. **Train your own tokenizer**: For maximum performance on your specific dataset, you can finetune the tokenizer on your own data.
### Training Your Own Tokenizer
```bash
lerobot-train-tokenizer \
--repo_id "user/my-lerobot-dataset" \
--action_horizon 10 \
--encoded_dims "0:6" \
--vocab_size 1024 \
--scale 10.0 \
--normalization_mode QUANTILES \
--output_dir "./my_fast_tokenizer" \
--push_to_hub \
--hub_repo_id "username/my-action-tokenizer"
```
### Key Tokenizer Parameters
| Parameter | Description | Default |
| ---------------------- | --------------------------------------------------------------------------------- | ------------ |
| `--repo_id` | LeRobot dataset repository ID | Required |
| `--action_horizon` | Number of future actions in each chunk | `10` |
| `--encoded_dims` | Comma-separated dimension ranges to encode (e.g., `"0:6,7:23"`) | `"0:6,7:23"` |
| `--vocab_size` | BPE vocabulary size | `1024` |
| `--scale` | DCT scaling factor for quantization | `10.0` |
| `--normalization_mode` | Normalization mode (`MEAN_STD`, `MIN_MAX`, `QUANTILES`, `QUANTILE10`, `IDENTITY`) | `QUANTILES` |
| `--sample_fraction` | Fraction of chunks to sample per episode | `0.1` |
## Usage
To use π₀-FAST in LeRobot, specify the policy type as:
```python
policy.type=pi0_fast
```
## Training
For training π₀-FAST, you can use the LeRobot training script:
```bash
lerobot-train \
--dataset.repo_id=your_dataset \
--policy.type=pi0_fast \
--output_dir=./outputs/pi0fast_training \
--job_name=pi0fast_training \
--policy.pretrained_path=lerobot/pi0fast-base \
--policy.dtype=bfloat16 \
--policy.gradient_checkpointing=true \
--policy.chunk_size=10 \
--policy.n_action_steps=10 \
--policy.max_action_tokens=256 \
--steps=100000 \
--batch_size=4 \
--policy.device=cuda
```
### Key Training Parameters
| Parameter | Description | Default |
| -------------------------------------- | -------------------------------------------------- | ------------------------------- |
| `--policy.gradient_checkpointing=true` | Reduces memory usage significantly during training | `false` |
| `--policy.dtype=bfloat16` | Use mixed precision training for efficiency | `float32` |
| `--policy.chunk_size` | Number of action steps to predict (action horizon) | `50` |
| `--policy.n_action_steps` | Number of action steps to execute | `50` |
| `--policy.max_action_tokens` | Maximum number of FAST tokens per action chunk | `256` |
| `--policy.action_tokenizer_name` | FAST tokenizer to use | `lerobot/fast-action-tokenizer` |
| `--policy.compile_model=true` | Enable torch.compile for faster training | `false` |
## Inference
### KV-Caching for Fast Inference
π₀-FAST supports **KV-caching**, a widely used optimization in LLM inference. This caches the key-value pairs from the attention mechanism, avoiding redundant computation during autoregressive decoding.
```python
# KV-caching is enabled by default
policy.use_kv_cache=true
```
### Inference Example
```python
from lerobot.policies.pi0_fast import PI0FastPolicy, PI0FastConfig
# Load the policy
policy = PI0FastPolicy.from_pretrained("your-model-path")
# During inference
actions = policy.predict_action_chunk(batch)
```
## Model Architecture
π₀-FAST uses a PaliGemma-based architecture:
- **Vision Encoder**: SigLIP vision tower for image understanding
- **Language Model**: Gemma 2B for processing language instructions and predicting action tokens
The model takes images, text instructions, and robot state as input, and outputs discrete FAST tokens that are decoded back to continuous actions.
## Configuration Options
| Parameter | Description | Default |
| -------------------- | ----------------------------------------------- | ---------- |
| `paligemma_variant` | VLM backbone variant (`gemma_300m`, `gemma_2b`) | `gemma_2b` |
| `max_state_dim` | Maximum state vector dimension (padded) | `32` |
| `max_action_dim` | Maximum action vector dimension (padded) | `32` |
| `temperature` | Sampling temperature (0.0 for greedy) | `0.0` |
| `max_decoding_steps` | Maximum decoding steps | `256` |
| `use_kv_cache` | Enable KV caching for faster inference | `true` |
## Comparison with π₀
| Feature | π₀ | π₀-FAST |
| --------------------- | ------------------------- | ---------------------------- |
| Action Representation | Flow Matching (Diffusion) | Autoregressive Tokens (FAST) |
| Training Speed | 1x | **5x faster** |
| Dexterity | High | High |
| Inference Method | Iterative Denoising | Autoregressive Decoding |
| KV-Caching | N/A | Supported |
## Reproducing π₀Fast results
We reproduce the results of π₀Fast on the LIBERO benchmark using the LeRobot implementation. We take the LeRobot PiFast base model [lerobot/pi0fast-base](https://huggingface.co/lerobot/pi0fast-base) and finetune for an additional 40kk steps in bfloat16, with batch size of 256 on 8 H100 GPUs using the [HuggingFace LIBERO dataset](https://huggingface.co/datasets/HuggingFaceVLA/libero).
The finetuned model can be found here:
- **π₀Fast LIBERO**: [lerobot/pi0fast-libero](https://huggingface.co/lerobot/pi0fast-libero)
With the following training command:
```bash
lerobot-train \
--dataset.repo_id=lerobot/libero \
--output_dir=outputs/libero_pi0fast \
--job_name=libero_pi0fast \
--policy.path=lerobot/pi0fast-base \
--policy.dtype=bfloat16 \
--steps=100000 \
--save_freq=20000 \
--batch_size=4 \
--policy.device=cuda \
--policy.scheduler_warmup_steps=4000 \
--policy.scheduler_decay_steps=100000 \
--policy.scheduler_decay_lr=1e-5 \
--policy.gradient_checkpointing=true \
--policy.chunk_size=10 \
--policy.n_action_steps=10 \
--policy.max_action_tokens=256 \
--policy.empty_cameras=1 \
```
We then evaluate the finetuned model using the LeRobot LIBERO implementation, by running the following command:
```bash
tasks="libero_object,libero_spatial,libero_goal,libero_10"
lerobot-eval \
--policy.path=lerobot/pi0fast-libero \
--policy.max_action_tokens=256 \
--env.type=libero \
--policy.gradient_checkpointing=false \
--env.task=${tasks} \
--eval.batch_size=1 \
--eval.n_episodes=1 \
--rename_map='{"observation.images.image":"observation.images.base_0_rgb","observation.images.image2":"observation.images.left_wrist_0_rgb"}'
```
**Note:** We set `n_action_steps=10`, similar to the original OpenPI implementation.
### Results
We obtain the following results on the LIBERO benchmark:
| Model | LIBERO Spatial | LIBERO Object | LIBERO Goal | LIBERO 10 | Average |
| ----------- | -------------- | ------------- | ----------- | --------- | -------- |
| **π₀-fast** | 70.0 | 100.0 | 100.0 | 60.0 | **82.5** |
The full evaluation output folder, including videos, is available [here](https://drive.google.com/drive/folders/1HXpwPTRm4hx6g1sF2P7OOqGG0TwPU7LQ?usp=sharing)
## License
This model follows the **Apache 2.0 License**, consistent with the original [OpenPI repository](https://github.com/Physical-Intelligence/openpi).
## References
- [FAST: Efficient Robot Action Tokenization](https://www.physicalintelligence.company/research/fast) - Physical Intelligence Blog
- [OpenPI Repository](https://github.com/Physical-Intelligence/openpi) - Original implementation
- [FAST Tokenizer on Hugging Face](https://huggingface.co/physical-intelligence/fast) - Pre-trained tokenizer
docs/source/policy_act_README.md
+14
View File
@@ -0,0 +1,14 @@
## Paper
https://tonyzhaozh.github.io/aloha
## Citation
```bibtex
@article{zhao2023learning,
title={Learning fine-grained bimanual manipulation with low-cost hardware},
author={Zhao, Tony Z and Kumar, Vikash and Levine, Sergey and Finn, Chelsea},
journal={arXiv preprint arXiv:2304.13705},
year={2023}
}
```
docs/source/policy_diffusion_README.md
+14
View File
@@ -0,0 +1,14 @@
## Paper
https://diffusion-policy.cs.columbia.edu
## Citation
```bibtex
@article{chi2024diffusionpolicy,
author = {Cheng Chi and Zhenjia Xu and Siyuan Feng and Eric Cousineau and Yilun Du and Benjamin Burchfiel and Russ Tedrake and Shuran Song},
title ={Diffusion Policy: Visuomotor Policy Learning via Action Diffusion},
journal = {The International Journal of Robotics Research},
year = {2024},
}
```
docs/source/policy_evo1_README.md
+18
View File
@@ -0,0 +1,18 @@
# EVO1
EVO1 is a Vision-Language-Action policy for robot control. The LeRobot
integration uses an InternVL3 vision-language backbone with a flow-matching
action head, and supports staged training through the standard LeRobot policy
APIs.
The upstream EVO1 project is available at
[MINT-SJTU/Evo-1](https://github.com/MINT-SJTU/Evo-1).
```bibtex
@misc{evo1,
title = {EVO1},
author = {{MINT-SJTU}},
year = {2025},
howpublished = {\url{https://github.com/MINT-SJTU/Evo-1}},
}
```
docs/source/policy_fastwam_README.md
+56
View File
@@ -0,0 +1,56 @@
## Research Paper
Paper: https://arxiv.org/abs/2603.16666
## Repository
Code: https://github.com/yuantianyuan01/FastWAM
Project page: https://yuantianyuan01.github.io/FastWAM/
## Citation
```bibtex
@article{yuan2026fastwam,
title = {Fast-WAM: Do World Action Models Need Test-time Future Imagination?},
author = {Tianyuan Yuan and Zibin Dong and Yicheng Liu and Hang Zhao},
journal = {arXiv preprint arXiv:2603.16666},
year = {2026},
url = {https://arxiv.org/abs/2603.16666}
}
```
## Additional Resources
Base video model: https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B
Released upstream checkpoints: https://huggingface.co/yuanty/fastwam
## Results
Evaluated on LIBERO with [`ZibinDong/fastwam_libero_uncond_2cam224`](https://huggingface.co/ZibinDong/fastwam_libero_uncond_2cam224):
| Suite | Success rate | n_episodes |
| -------------- | -----------: | ---------: |
| libero_spatial | 97.6% | 500 |
| libero_object | 99.0% | 500 |
| libero_goal | 95.0% | 500 |
| libero_10 | 94.0% | 500 |
| **average** | **96.4%** | 2000 |
Reproduce: `lerobot-eval --policy.path=ZibinDong/fastwam_libero_uncond_2cam224 --policy.device=cuda --policy.torch_dtype=float32 --policy.n_action_steps=10 --env.type=libero --env.task=libero_spatial --env.observation_height=256 --env.observation_width=256 --eval.batch_size=1 --eval.n_episodes=50 --seed=0 --env.episode_length=300`.
For LIBERO-10, use `--env.task=libero_10 --env.episode_length=600`:
```bash
lerobot-eval \
--policy.path=ZibinDong/fastwam_libero_uncond_2cam224 \
--policy.device=cuda \
--policy.torch_dtype=float32 \
--policy.n_action_steps=10 \
--env.type=libero \
--env.task=libero_10 --env.observation_height=256 --env.observation_width=256 \
--eval.batch_size=1 \
--eval.n_episodes=50 \
--seed=0 --env.episode_length=600
```
docs/source/policy_groot_README.md
+138
View File
@@ -0,0 +1,138 @@
## Research Paper
GR00T N1 technical report (covers the GR00T N1.x family, including N1.7): https://arxiv.org/abs/2503.14734
GR00T N1.7 model card: https://huggingface.co/nvidia/GR00T-N1.7-3B
GR00T N1.5 research page (earlier version): https://research.nvidia.com/labs/gear/gr00t-n1_5/
> GR00T N1.5 support was removed from LeRobot; the last release supporting it is `lerobot==0.5.1`.
> Current releases support GR00T N1.7 only.
## Repository
Code: https://github.com/NVIDIA/Isaac-GR00T
## Citation
```bibtex
@inproceedings{gr00tn1_2025,
archivePrefix = {arxiv},
eprint = {2503.14734},
title = {{GR00T} {N1}: An Open Foundation Model for Generalist Humanoid Robots},
author = {NVIDIA and Johan Bjorck andFernando Castañeda, Nikita Cherniadev and Xingye Da and Runyu Ding and Linxi "Jim" Fan and Yu Fang and Dieter Fox and Fengyuan Hu and Spencer Huang and Joel Jang and Zhenyu Jiang and Jan Kautz and Kaushil Kundalia and Lawrence Lao and Zhiqi Li and Zongyu Lin and Kevin Lin and Guilin Liu and Edith Llontop and Loic Magne and Ajay Mandlekar and Avnish Narayan and Soroush Nasiriany and Scott Reed and You Liang Tan and Guanzhi Wang and Zu Wang and Jing Wang and Qi Wang and Jiannan Xiang and Yuqi Xie and Yinzhen Xu and Zhenjia Xu and Seonghyeon Ye and Zhiding Yu and Ao Zhang and Hao Zhang and Yizhou Zhao and Ruijie Zheng and Yuke Zhu},
month = {March},
year = {2025},
booktitle = {ArXiv Preprint},
}
```
## Additional Resources
Blog: https://developer.nvidia.com/isaac/gr00t
Hugging Face Models:
- GR00T N1.7: https://huggingface.co/nvidia/GR00T-N1.7-3B
- GR00T N1.7 LIBERO checkpoints: https://huggingface.co/nvidia/GR00T-N1.7-LIBERO
<details>
<summary><b>Original-vs-LeRobot parity test</b></summary>
## Original-vs-LeRobot parity test
`tests/policies/groot/test_groot_vs_original.py` verifies this LeRobot
reimplementation of GR00T N1.7 (Qwen3-VL backbone + flow-matching action head)
against NVIDIA's original `gr00t` package with two comparisons, each parametrized
over every embodiment tag present in the checkpoint:
1. **Model parity** — given byte-identical pre-processed inputs and the same
flow-matching seed (recorded in each artifact), both implementations must produce
the **same raw model output** (`get_action(...)["action_pred"]`, the normalized
flow-matching prediction). Output shapes must match exactly; any action-horizon
or action-dim mismatch fails the test.
2. **Preprocessor parity** — given the identical raw observations (per-camera
frames, state vectors, language instruction), LeRobot's own preprocessor pipeline
(real Qwen3-VL chat template / tokenizer / image packing + checkpoint-driven
state normalization, no mocks) must produce the **same collated model inputs**
(`input_ids`, `attention_mask`, `pixel_values`, `image_grid_thw`, `state`,
`embodiment_id`) as the original package's processor.
### Why two environments
The original `gr00t` package pins `transformers==4.57.3` (Python 3.10); this
integration requires `transformers>=5.x` (Qwen3-VL). Under 5.x, `PretrainedConfig`
is itself a defaulted dataclass, so the original config dataclasses fail to import
(`non-default argument follows default argument`). The two implementations therefore
**cannot be imported in the same Python process**.
So the test uses a **producer / consumer** split across two venvs:
1. **Producer** — `tests/policies/groot/utils/dump_original_n1_7.py`, run in the _original_
gr00t venv. For each embodiment it builds dummy inputs generically from the
checkpoint metadata (state dims from `statistics.json`; camera/language keys from
the processor modality configs), runs the original model, and saves to one `.npz`
per tag: the raw observations (`raw::` keys), the exact collated inputs
(`in::` keys), the seed, and the raw `action_pred`.
2. **Consumer** — the pytest above, run in the _LeRobot_ venv. It discovers every
`.npz`; the model-parity case replays the byte-identical collated inputs through
the LeRobot model with the recorded seed and asserts the outputs match, and the
preprocessor-parity case replays the raw observations through LeRobot's full
preprocessor pipeline and asserts the collated tensors match.
> Artifacts generated by older versions of the dump script contain no `raw::`
> fields; the preprocessor-parity case then **skips** with a regeneration hint.
> Re-run the producer to refresh them.
### Fairness controls
- **Same pre-processed inputs (model parity)** — the original processor's `input_ids`,
`pixel_values`, `image_grid_thw`, `attention_mask`, `state`, `embodiment_id` are
fed verbatim to the LeRobot model (no re-tokenization / re-normalization), so the
model comparison isolates the model. LeRobot's own tokenization / image packing is
covered separately by the preprocessor-parity case, which compares its output
against those same collated tensors from identical raw observations.
- **Same precision + attention kernel** — both sides run **fp32 + SDPA**. The
original defaults to `use_flash_attention=True` (flash_attention_2 + bf16); the
producer forces SDPA + fp32. (With the defaults the gap is ~3e-2 — pure
kernel/rounding noise, not an implementation difference.)
- **Same flow-matching seed** — fixed right before sampling on both sides; the
producer records it in each artifact (`--seed`, default 42) and the consumer
replays the recorded value.
### How to run
```bash
# Resolve a local checkpoint (GR00T-N1.7-LIBERO / libero_10)
CKPT=$(python - <<'PY'
import os
from huggingface_hub import snapshot_download
print(os.path.join(snapshot_download("nvidia/GR00T-N1.7-LIBERO",
allow_patterns=["libero_10/*"]), "libero_10"))
PY
)
# 1) Produce the original-side artifacts for all embodiments (original gr00t venv, CUDA)
CUDA_VISIBLE_DEVICES=0 /path/to/Isaac-GR00T/.venv-original/bin/python \
tests/policies/groot/utils/dump_original_n1_7.py \
--ckpt "$CKPT" --out-dir tests/policies/groot/artifacts --device cuda --seed 42
# 2) Run the parity test (LeRobot venv) — one parametrized case per embodiment
CUDA_VISIBLE_DEVICES=0 GROOT_PARITY_DEVICE=cuda \
uv run pytest tests/policies/groot/test_groot_vs_original.py -v -s
```
The `.npz` artifacts are local-only (gitignored, ~6–10 MB each) and are regenerated by
the producer; they are never committed. The tests **skip** (do not fail) on CI or
when the checkpoint / artifacts are absent.
#### Env knobs (all optional)
| Var | Default | Purpose |
| ----------------------------------------- | -------------------------------- | ------------------------------------- |
| `GROOT_N1_7_PARITY_DIR` | `tests/policies/groot/artifacts` | directory of per-tag `.npz` artifacts |
| `GROOT_N1_7_LIBERO_CKPT` | auto (HF cache) | override checkpoint dir |
| `GROOT_PARITY_DEVICE` | `cuda` if available | `cpu` or `cuda` |
| `GROOT_PARITY_ATOL` / `GROOT_PARITY_RTOL` | `1e-3` | comparison tolerance |
</details>
docs/source/policy_molmoact2_README.md
+39
View File
@@ -0,0 +1,39 @@
# MolmoAct2
This repository contains the LeRobot policy implementation of
[MolmoAct2](https://allenai.org/blog/molmoact2), ported into LeRobot for
training, evaluation, checkpointing, and dataset compatibility.
This implementation currently supports training and evaluation for the regular
MolmoAct2 model. MolmoAct2-Think, which supports adaptive depth reasoning, is
not included in this LeRobot policy yet and is coming soon.
For the original MolmoAct2 training code used for the experiments reported in
the paper, see [allenai/molmoact2](https://github.com/allenai/molmoact2).
## LIBERO Evaluation
Important: we found that `num_steps_wait=10` does not reliably let the LIBERO
scene stabilize and can degrade measured success. All LIBERO evaluation results
reported for this LeRobot implementation use `num_steps_wait=50`.
## Citation
```bibtex
@misc{fang2026molmoact2actionreasoningmodels,
title={MolmoAct2: Action Reasoning Models for Real-world Deployment},
author={Haoquan Fang and Jiafei Duan and Donovan Clay and Sam Wang and Shuo Liu and Weikai Huang and Xiang Fan and Wei-Chuan Tsai and Shirui Chen and Yi Ru Wang and Shanli Xing and Jaemin Cho and Jae Sung Park and Ainaz Eftekhar and Peter Sushko and Karen Farley and Angad Wadhwa and Cole Harrison and Winson Han and Ying-Chun Lee and Eli VanderBilt and Rose Hendrix and Suveen Ellawela and Lucas Ngoo and Joyce Chai and Zhongzheng Ren and Ali Farhadi and Dieter Fox and Ranjay Krishna},
year={2026},
eprint={2605.02881},
archivePrefix={arXiv},
primaryClass={cs.RO},
url={https://arxiv.org/abs/2605.02881},
}
```
## License
This model is licensed under Apache 2.0. It is intended for research and
educational use in accordance with
[Ai2's Responsible Use Guidelines](https://allenai.org/responsible-use),
consistent with [allenai/molmoact2](https://github.com/allenai/molmoact2).
docs/source/policy_multi_task_dit_README.md
+37
View File
@@ -0,0 +1,37 @@
# Multitask DiT Policy
## Citation
If you use this work, please cite the following works:
```bibtex
@misc{jones2025multitaskditpolicy,
author = {Bryson Jones},
title = {Dissecting and Open-Sourcing Multitask Diffusion Transformer Policy},
year = {2025},
url = {https://brysonkjones.substack.com/p/dissecting-and-open-sourcing-multitask-diffusion-transformer-policy},
note = {Blog post}
}
```
```bibtex
@misc{trilbmteam2025carefulexaminationlargebehaviormodels,
author = {TRI LBM Team},
title = {A Careful Examination of Large Behavior Models for Multitask Dexterous Manipulation},
year = {2025},
eprint = {arXiv:2507.05331},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2507.05331}
}
```
```bibtex
@misc{bostondynamics2025largebehaviormodelsatlas,
author = {Boston Dynamics and TRI Research Team},
title = {Large Behavior Models and Atlas Find New Footing},
year = {2025},
url = {https://bostondynamics.com/blog/large-behavior-models-atlas-find-new-footing/},
note = {Blog post}
}
```
docs/source/policy_pi05_README.md
+91
View File
@@ -0,0 +1,91 @@
# π₀.₅ (pi05)
This repository contains the Hugging Face port of **π₀.₅**, adapted from [OpenPI](https://github.com/Physical-Intelligence/openpi) by the Physical Intelligence.
It is designed as a **Vision-Language-Action model with open-world generalization**.
---
## Model Overview
| Feature | π₀ | π₀.₅ |
| -------------------- | ------------------------------------------------------ | ----------------------------------------- |
| Time Conditioning | Concatenates time with actions via `action_time_mlp_*` | Uses `time_mlp_*` for AdaRMS conditioning |
| AdaRMS | Not used | Used in action expert |
| Tokenizer Length | 48 tokens | 200 tokens |
| Discrete State Input | False (Uses `state_proj` layer) | True |
| Parameter Count | Higher (includes state embedding) | Lower (no state embedding) |
---
## Relative Actions
π₀.₅ supports training with **relative actions**, where the model learns relative offsets
from the current robot state instead of absolute joint positions. This mirrors the
relative-action transform in OpenPI (`DeltaActions`) and can improve performance.
### How it works
1. **During preprocessing**, absolute actions are converted to relative offsets:
`relative = action - state` (for selected joints).
2. The relative actions are normalized using statistics computed from the relative distribution.
3. **During postprocessing**, predicted relative actions are converted back to absolute:
`absolute = relative + state`.
Joints listed in `relative_exclude_joints` (e.g., gripper) are kept absolute.
### Configuration
| Parameter | Type | Default | Description |
| ------------------------- | ----------- | ------------- | ---------------------------------------------------------------- |
| `use_relative_actions` | `bool` | `False` | Enable relative-action training |
| `relative_exclude_joints` | `list[str]` | `["gripper"]` | Joint names to keep absolute (matched by substring) |
| `action_feature_names` | `list[str]` | `None` | Auto-populated from dataset metadata at runtime by `make_policy` |
### Training example
```bash
python -m lerobot.scripts.lerobot_train \
--policy.type=pi05 \
--dataset.repo_id=your_org/your_dataset \
--policy.use_relative_actions=true \
--policy.relative_exclude_joints='["gripper"]'
```
When `use_relative_actions=true`, the training script automatically:
- Computes relative action statistics from the dataset (sampled chunk-level relative actions)
- Replaces the standard action stats with relative stats for normalization
- Broadcasts these stats across all ranks in distributed training
---
## Citation
If you use this work, please cite both **OpenPI** and the π₀.₅ paper:
```bibtex
@misc{openpi2024,
author = {Physical Intelligence Lab},
title = {OpenPI: PyTorch Implementation of π0 and π0.5 Policies},
year = {2024},
publisher = {GitHub},
howpublished = {\url{https://github.com/Physical-Intelligence/openpi}},
license = {Apache-2.0}
}
@misc{intelligence2025pi05visionlanguageactionmodelopenworld,
title = {π₀.₅: a Vision-Language-Action Model with Open-World Generalization},
author = {Physical Intelligence and Kevin Black and Noah Brown and James Darpinian and Karan Dhabalia and Danny Driess and Adnan Esmail and Michael Equi and Chelsea Finn and Niccolo Fusai and Manuel Y. Galliker and Dibya Ghosh and Lachy Groom and Karol Hausman and Brian Ichter and Szymon Jakubczak and Tim Jones and Liyiming Ke and Devin LeBlanc and Sergey Levine and Adrian Li-Bell and Mohith Mothukuri and Suraj Nair and Karl Pertsch and Allen Z. Ren and Lucy Xiaoyang Shi and Laura Smith and Jost Tobias Springenberg and Kyle Stachowicz and James Tanner and Quan Vuong and Homer Walke and Anna Walling and Haohuan Wang and Lili Yu and Ury Zhilinsky},
year = {2025},
eprint = {2504.16054},
archivePrefix= {arXiv},
primaryClass = {cs.LG},
url = {https://arxiv.org/abs/2504.16054},
}
```
---
## License
This port follows the **Apache 2.0 License**, consistent with the original [OpenPI repository](https://github.com/Physical-Intelligence/openpi).
docs/source/policy_pi0_README.md
+107
View File
@@ -0,0 +1,107 @@
# π₀ (pi0)
This repository contains the Hugging Face port of **π₀**, adapted from [OpenPI](https://github.com/Physical-Intelligence/openpi) by the Physical Intelligence.
It is designed as a **Vision-Language-Action model for general robot control**.
---
## Model Overview
| Feature | π₀ | π₀.₅ |
| -------------------- | ------------------------------------------------------ | ----------------------------------------- |
| Time Conditioning | Concatenates time with actions via `action_time_mlp_*` | Uses `time_mlp_*` for AdaRMS conditioning |
| AdaRMS | Not used | Used in action expert |
| Tokenizer Length | 48 tokens | 200 tokens |
| Discrete State Input | False (Uses `state_proj` layer) | True |
| Parameter Count | Higher (includes state embedding) | Lower (no state embedding) |
---
## Relative Actions
π₀ supports training with **relative actions**, where the model learns relative offsets
from the current robot state instead of absolute joint positions. This mirrors the
relative-action transform in OpenPI (`DeltaActions`) and can improve performance.
### How it works
1. **During preprocessing**, absolute actions are converted to relative offsets:
`relative = action - state` (for selected joints).
2. The relative actions are normalized using statistics computed from the relative distribution.
3. **During postprocessing**, predicted relative actions are converted back to absolute:
`absolute = relative + state`.
Joints listed in `relative_exclude_joints` (e.g., gripper) are kept absolute.
### Configuration
| Parameter | Type | Default | Description |
| ------------------------- | ----------- | ------------- | ---------------------------------------------------------------- |
| `use_relative_actions` | `bool` | `False` | Enable relative-action training |
| `relative_exclude_joints` | `list[str]` | `["gripper"]` | Joint names to keep absolute (matched by substring) |
| `action_feature_names` | `list[str]` | `None` | Auto-populated from dataset metadata at runtime by `make_policy` |
### Training example
```bash
python -m lerobot.scripts.lerobot_train \
--policy.type=pi0 \
--dataset.repo_id=your_org/your_dataset \
--policy.use_relative_actions=true \
--policy.relative_exclude_joints='["gripper"]'
```
When `use_relative_actions=true`, the training script automatically:
- Computes relative action statistics from the dataset (sampled chunk-level relative actions)
- Replaces the standard action stats with relative stats for normalization
- Broadcasts these stats across all ranks in distributed training
### Recomputing stats for an existing dataset
If you want to precompute relative action stats offline, use `recompute_stats` from
`lerobot.datasets`:
```python
from lerobot.datasets import LeRobotDataset, recompute_stats
dataset = LeRobotDataset("your_org/your_dataset")
dataset = recompute_stats(
dataset,
relative_action=True,
relative_exclude_joints=["gripper"],
)
```
---
## Citation
If you use this work, please cite both **OpenPI** and the π₀ paper:
```bibtex
@misc{openpi2024,
author = {Physical Intelligence Lab},
title = {OpenPI: PyTorch Implementation of π0 and π0.5 Policies},
year = {2024},
publisher = {GitHub},
howpublished = {\url{https://github.com/Physical-Intelligence/openpi}},
license = {Apache-2.0}
}
@misc{black2024pi0visionlanguageactionflowmodel,
title = {π₀: A Vision-Language-Action Flow Model for General Robot Control},
author = {Kevin Black and Noah Brown and Danny Driess and Adnan Esmail and Michael Equi and Chelsea Finn and Niccolo Fusai and Lachy Groom and Karol Hausman and Brian Ichter and Szymon Jakubczak and Tim Jones and Liyiming Ke and Sergey Levine and Adrian Li-Bell and Mohith Mothukuri and Suraj Nair and Karl Pertsch and Lucy Xiaoyang Shi and James Tanner and Quan Vuong and Anna Walling and Haohuan Wang and Ury Zhilinsky},
year = {2024},
eprint = {2410.24164},
archivePrefix= {arXiv},
primaryClass = {cs.LG},
url = {https://arxiv.org/abs/2410.24164},
}
```
---
## License
This port follows the **Apache 2.0 License**, consistent with the original [OpenPI repository](https://github.com/Physical-Intelligence/openpi).
docs/source/policy_rtc_README.md
+38
View File
@@ -0,0 +1,38 @@
# Real-Time Chunking (RTC)
This module contains the LeRobot implementation of **Real-Time Chunking (RTC)**, an inference-time technique for flow-matching based policies.
**Note**: RTC is not a policy itself, but rather an inference enhancement that works with flow-matching based policies including [π₀](../pi0/), [π₀.₅](../pi05/), and [SmolVLA](../smolvla/).
---
## Citation
If you use Real-Time Chunking in your work, please cite:
```bibtex
@misc{openpi2024,
author = {Physical Intelligence Lab},
title = {OpenPI: PyTorch Implementation of π0 and π0.5 Policies},
year = {2024},
publisher = {GitHub},
howpublished = {\url{https://github.com/Physical-Intelligence/openpi}},
license = {Apache-2.0}
}
@misc{black2025realtimeexecutionactionchunking,
title={Real-Time Execution of Action Chunking Flow Policies},
author={Kevin Black and Manuel Y. Galliker and Sergey Levine},
year={2025},
eprint={2506.07339},
archivePrefix={arXiv},
primaryClass={cs.RO},
url={https://arxiv.org/abs/2506.07339},
}
```
---
## License
This implementation follows the **Apache 2.0 License**, consistent with the LeRobot project.
docs/source/policy_sarm_README.md
+14
View File
@@ -0,0 +1,14 @@
## Paper
https://arxiv.org/abs/2509.25358
## Citation
```bibtex
@article{chen2025sarm,
title={SARM: Stage-Aware Reward Modeling for Long Horizon Robot Manipulation},
author={Chen, Qianzhong and Yu, Justin and Schwager, Mac and Abbeel, Pieter and Shentu, Yide and Wu, Philipp},
journal={arXiv preprint arXiv:2509.25358},
year={2025}
}
```
docs/source/policy_smolvla_README.md
+14
View File
@@ -0,0 +1,14 @@
## Paper
https://arxiv.org/abs/2506.01844
## Citation
```bibtex
@article{shukor2025smolvla,
title={SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics},
author={Shukor, Mustafa and Aubakirova, Dana and Capuano, Francesco and Kooijmans, Pepijn and Palma, Steven and Zouitine, Adil and Aractingi, Michel and Pascal, Caroline and Russi, Martino and Marafioti, Andres and Alibert, Simon and Cord, Matthieu and Wolf, Thomas and Cadene, Remi},
journal={arXiv preprint arXiv:2506.01844},
year={2025}
}
```
docs/source/policy_tdmpc_README.md
+14
View File
@@ -0,0 +1,14 @@
## Paper
https://www.nicklashansen.com/td-mpc/
## Citation
```bibtex
@inproceedings{Hansen2022tdmpc,
title={Temporal Difference Learning for Model Predictive Control},
author={Nicklas Hansen and Xiaolong Wang and Hao Su},
booktitle={ICML},
year={2022}
}
```
docs/source/policy_vla_jepa_README.md
+39
View File
@@ -0,0 +1,39 @@
# VLA-JEPA
This repository contains the LeRobot port of **VLA-JEPA**, a Vision-Language-Action model that combines a Qwen3-VL language backbone with a self-supervised video world model (V-JEPA2) and a flow-matching DiT action head.
Converted from [ginwind/VLA-JEPA](https://huggingface.co/ginwind/VLA-JEPA).
---
## Architecture Overview
| Component | Module | Role |
| ----------------------- | --------------------------------- | ------------------------------------------------------- |
| **Qwen3-VL backbone** | `Qwen3VLInterface` | Fuses images + language instruction into context tokens |
| **DiT-B action head** | `VLAJEPAActionHead` | Flow-matching diffusion over the action chunk |
| **V-JEPA2 world model** | `ActionConditionedVideoPredictor` | Self-supervised video prediction loss (training only) |
At inference time only the Qwen backbone and action head are used; the world model is not needed.
---
## Citation
```bibtex
@misc{sun2026vlajepaenhancingvisionlanguageactionmodel,
title = {VLA-JEPA: Enhancing Vision-Language-Action Model with Latent World Model},
author = {Jingwen Sun and Wenyao Zhang and Zekun Qi and Shaojie Ren and Zezhi Liu and Hanxin Zhu and Guangzhong Sun and Xin Jin and Zhibo Chen},
year = {2026},
eprint = {2602.10098},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2602.10098},
}
```
---
## License
Weights are distributed under the license terms of the original [ginwind/VLA-JEPA](https://huggingface.co/ginwind/VLA-JEPA) repository (**Apache 2.0 License**). The LeRobot integration code follows the **Apache 2.0 License**.
docs/source/policy_vqbet_README.md
+14
View File
@@ -0,0 +1,14 @@
## Paper
https://sjlee.cc/vq-bet/
## Citation
```bibtex
@article{lee2024behavior,
title={Behavior generation with latent actions},
author={Lee, Seungjae and Wang, Yibin and Etukuru, Haritheja and Kim, H Jin and Shafiullah, Nur Muhammad Mahi and Pinto, Lerrel},
journal={arXiv preprint arXiv:2403.03181},
year={2024}
}
```
docs/source/policy_walloss_README.md
+45
View File
@@ -0,0 +1,45 @@
# WALL-OSS
This repository contains the Hugging Face port of [**WALL-OSS**](https://x2robot.com/en/research/68bc2cde8497d7f238dde690), a Vision-Language-Action model for cross-embodiment robotic control based on Qwen2.5-VL with flow matching/FAST action prediction.
---
## Model Overview
| Feature | Description |
| ------------------ | ----------------------------------------------------- |
| Base Model | Qwen2.5-VL (Vision-Language Model) |
| Action Prediction | Flow Matching (diffusion) or FAST (discrete tokens) |
| Architecture | Mixture of Experts (MoE) with action-specific routing |
| Multi-Modal Inputs | Vision (images/videos), Language, Proprioception |
---
## Additional Resources
Paper: https://arxiv.org/pdf/2509.11766
Official Repository: https://github.com/X-Square-Robot/wall-x
Hugging Face: https://huggingface.co/x-square-robot
---
## Citation
If you use this work, please cite:
```bibtex
@article{zhai2025igniting,
title = {Igniting VLMs Toward the Embodied Space},
author = {Zhai, Andy and Liu, Brae and Fang, Bruno and Cai, Chalse and Ma, Ellie and Yin, Ethan and Wang, Hao and Zhou, Hugo and Wang, James and Shi, Lights and Liang, Lucy and Wang, Make and Wang, Qian and Gan, Roy and Yu, Ryan and Li, Shalfun and Liu, Starrick and Chen, Sylas and Chen, Vincent and Xu, Zach},
journal = {arXiv preprint arXiv:2509.11766},
year = {2025}
}
```
---
## License
This model follows the **Apache 2.0 License**, consistent with the original [WallX repository](https://github.com/X-Square-Robot/wall-x).
docs/source/porting_datasets_v3.mdx
+321
View File
@@ -0,0 +1,321 @@
# Porting Large Datasets to LeRobot Dataset v3.0
This tutorial explains how to port large-scale robotic datasets to the LeRobot Dataset v3.0 format. We'll use the **DROID 1.0.1** dataset as our primary example, which demonstrates handling multi-terabyte datasets with thousands of shards across SLURM clusters.
## File Organization: v2.1 vs v3.0
Dataset v3.0 fundamentally changes how data is organized and stored:
**v2.1 Structure (Episode-based)**:
```
dataset/
├── data/chunk-000/episode_000000.parquet
├── data/chunk-000/episode_000001.parquet
├── videos/chunk-000/camera/episode_000000.mp4
└── meta/episodes.jsonl
```
**v3.0 Structure (File-based)**:
```
dataset/
├── data/chunk-000/file-000.parquet # Multiple episodes per file
├── videos/camera/chunk-000/file-000.mp4 # Consolidated video chunks
└── meta/episodes/chunk-000/file-000.parquet # Structured metadata
```
This transition from individual episode files to file-based chunks dramatically improves performance and reduces storage overhead.
## What's New in Dataset v3.0
Dataset v3.0 introduces significant improvements for handling large datasets:
### 🏗️ **Enhanced File Organization**
- **File-based structure**: Episodes are now grouped into chunked files rather than individual episode files
- **Configurable file sizes**: for data and video files
- **Improved storage efficiency**: Better compression and reduced overhead
### 📊 **Modern Metadata Management**
- **Parquet-based metadata**: Replaced JSON Lines with efficient parquet format
- **Structured episode access**: Direct pandas DataFrame access via `dataset.meta.episodes`
- **Per-episode statistics**: Enhanced statistics tracking at episode level
### 🚀 **Performance Enhancements**
- **Memory-mapped access**: Improved RAM usage through PyArrow memory mapping
- **Faster loading**: Significantly reduced dataset initialization time
- **Better scalability**: Designed for datasets with millions of episodes
## Prerequisites
Before porting large datasets, ensure you have:
- **LeRobot installed** with v3.0 support. Follow our [Installation Guide](./installation).
- **Sufficient storage**: Raw datasets can be very large (e.g., DROID requires 2TB)
- **Cluster access** (recommended for large datasets): SLURM or similar job scheduler
- **Dataset-specific dependencies**: For DROID, you'll need TensorFlow Dataset utilities
## Understanding the DROID Dataset
[DROID 1.0.1](https://droid-dataset.github.io/droid/the-droid-dataset) is an excellent example of a large-scale robotic dataset:
- **Size**: 1.7TB (RLDS format), 8.7TB (raw data)
- **Structure**: 2048 pre-defined TensorFlow dataset shards
- **Content**: 76,000+ robot manipulation trajectories from Franka Emika Panda robots
- **Scope**: Real-world manipulation tasks across multiple environments and objects
- **Format**: Originally in TensorFlow Records/RLDS format, requiring conversion to LeRobot format
- **Hosting**: Google Cloud Storage with public access via `gsutil`
The dataset contains diverse manipulation demonstrations with:
- Multiple camera views (wrist camera, exterior cameras)
- Natural language task descriptions
- Robot proprioceptive state and actions
- Success/failure annotations
### DROID Features Schema
```python
DROID_FEATURES = {
# Episode markers
"is_first": {"dtype": "bool", "shape": (1,)},
"is_last": {"dtype": "bool", "shape": (1,)},
"is_terminal": {"dtype": "bool", "shape": (1,)},
# Language instructions
"language_instruction": {"dtype": "string", "shape": (1,)},
"language_instruction_2": {"dtype": "string", "shape": (1,)},
"language_instruction_3": {"dtype": "string", "shape": (1,)},
# Robot state
"observation.state.gripper_position": {"dtype": "float32", "shape": (1,)},
"observation.state.cartesian_position": {"dtype": "float32", "shape": (6,)},
"observation.state.joint_position": {"dtype": "float32", "shape": (7,)},
# Camera observations
"observation.images.wrist_left": {"dtype": "image"},
"observation.images.exterior_1_left": {"dtype": "image"},
"observation.images.exterior_2_left": {"dtype": "image"},
# Actions
"action.gripper_position": {"dtype": "float32", "shape": (1,)},
"action.cartesian_position": {"dtype": "float32", "shape": (6,)},
"action.joint_position": {"dtype": "float32", "shape": (7,)},
# Standard LeRobot format
"observation.state": {"dtype": "float32", "shape": (8,)}, # joints + gripper
"action": {"dtype": "float32", "shape": (8,)}, # joints + gripper
}
```
## Approach 1: Single Computer Porting
### Step 1: Install Dependencies
For DROID specifically:
```bash
pip install tensorflow
pip install tensorflow_datasets
```
For other datasets, install the appropriate readers for your source format.
### Step 2: Download Raw Data
Download DROID from Google Cloud Storage using `gsutil`:
```bash
# Install Google Cloud SDK if not already installed
# https://cloud.google.com/sdk/docs/install
# Download the full RLDS dataset (1.7TB)
gsutil -m cp -r gs://gresearch/robotics/droid/1.0.1 /your/data/
# Or download just the 100-episode sample (2GB) for testing
gsutil -m cp -r gs://gresearch/robotics/droid_100 /your/data/
```
> [!WARNING]
> Large datasets require substantial time and storage:
>
> - **Full DROID (1.7TB)**: Several days to download depending on bandwidth
> - **Processing time**: 7+ days for local porting of full dataset
> - **Upload time**: 3+ days to push to Hugging Face Hub
> - **Local storage**: ~400GB for processed LeRobot format
### Step 3: Port the Dataset
```bash
python examples/port_datasets/port_droid.py \
--raw-dir /your/data/droid/1.0.1 \
--repo-id your_id/droid_1.0.1 \
--push-to-hub
```
### Development and Testing
For development, you can port a single shard:
```bash
python examples/port_datasets/port_droid.py \
--raw-dir /your/data/droid/1.0.1 \
--repo-id your_id/droid_1.0.1_test \
--num-shards 2048 \
--shard-index 0
```
This approach works for smaller datasets or testing, but large datasets require cluster computing.
## Approach 2: SLURM Cluster Porting (Recommended)
For large datasets like DROID, parallel processing across multiple nodes dramatically reduces processing time.
### Step 1: Install Cluster Dependencies
```bash
pip install datatrove # Hugging Face's distributed processing library
```
### Step 2: Configure Your SLURM Environment
Find your partition information:
```bash
sinfo --format="%R" # List available partitions
sinfo -N -p your_partition -h -o "%N cpus=%c mem=%m" # Check resources
```
Choose a **CPU partition** - no GPU needed for dataset porting.
### Step 3: Launch Parallel Porting Jobs
```bash
python examples/port_datasets/slurm_port_shards.py \
--raw-dir /your/data/droid/1.0.1 \
--repo-id your_id/droid_1.0.1 \
--logs-dir /your/logs \
--job-name port_droid \
--partition your_partition \
--workers 2048 \
--cpus-per-task 8 \
--mem-per-cpu 1950M
```
#### Parameter Guidelines
- **`--workers`**: Number of parallel jobs (max 2048 for DROID's shard count)
- **`--cpus-per-task`**: 8 CPUs recommended for frame encoding parallelization
- **`--mem-per-cpu`**: ~16GB total RAM (8×1950M) for loading raw frames
> [!TIP]
> Start with fewer workers (e.g., 100) to test your cluster configuration before launching thousands of jobs.
### Step 4: Monitor Progress
Check running jobs:
```bash
squeue -u $USER
```
Monitor overall progress:
```bash
jobs_status /your/logs
```
Inspect individual job logs:
```bash
less /your/logs/port_droid/slurm_jobs/JOB_ID_WORKER_ID.out
```
Debug failed jobs:
```bash
failed_logs /your/logs/port_droid
```
### Step 5: Aggregate Shards
Once all porting jobs complete:
```bash
python examples/port_datasets/slurm_aggregate_shards.py \
--repo-id your_id/droid_1.0.1 \
--logs-dir /your/logs \
--job-name aggr_droid \
--partition your_partition \
--workers 2048 \
--cpus-per-task 8 \
--mem-per-cpu 1950M
```
### Step 6: Upload to Hub
```bash
python examples/port_datasets/slurm_upload.py \
--repo-id your_id/droid_1.0.1 \
--logs-dir /your/logs \
--job-name upload_droid \
--partition your_partition \
--workers 50 \
--cpus-per-task 4 \
--mem-per-cpu 1950M
```
> [!NOTE]
> Upload uses fewer workers (50) since it's network-bound rather than compute-bound.
## Dataset v3.0 File Structure
Your completed dataset will have this modern structure:
```
dataset/
├── meta/
│ ├── episodes/
│ │ └── chunk-000/
│ │ └── file-000.parquet # Episode metadata
│ ├── tasks.parquet # Task definitions
│ ├── stats.json # Aggregated statistics
│ └── info.json # Dataset information
├── data/
│ └── chunk-000/
│ └── file-000.parquet # Consolidated episode data
└── videos/
└── camera_key/
└── chunk-000/
└── file-000.mp4 # Consolidated video files
```
This replaces the old episode-per-file structure with efficient, optimally-sized chunks.
## Migrating from Dataset v2.1
If you have existing datasets in v2.1 format, use the migration tool:
```bash
python src/lerobot/scripts/convert_dataset_v21_to_v30.py \
--repo-id your_id/existing_dataset
```
This automatically:
- Converts file structure to v3.0 format
- Migrates metadata from JSON Lines to parquet
- Aggregates statistics and creates per-episode stats
- Updates version information
## Performance Benefits
Dataset v3.0 provides significant improvements for large datasets:
- **Faster loading**: 3-5x reduction in initialization time
- **Memory efficiency**: Better RAM usage through memory mapping
- **Scalable processing**: Handles millions of episodes efficiently
- **Storage optimization**: Reduced file count and improved compression
docs/source/processors_robots_teleop.mdx
+151
View File
@@ -0,0 +1,151 @@
# Processors for Robots and Teleoperators
This guide shows how to build and modify processing pipelines that connect teleoperators (e.g., phone) to robots and datasets. Pipelines standardize conversions between different action/observation spaces so you can swap teleops and robots without rewriting glue code.
We use the Phone to SO‑100 follower examples for concreteness, but the same patterns apply to other robots.
**What you'll learn**
- Absolute vs. relative EE control: What each means, trade‑offs, and how to choose for your task.
- Three-pipeline pattern: How to map teleop actions → dataset actions → robot commands, and robot observations → dataset observations.
- Adapters (`to_transition` / `to_output`): How these convert raw dicts to `EnvTransition` and back to reduce boilerplate.
- Dataset feature contracts: How steps declare features via `transform_features(...)`, and how to aggregate/merge them for recording.
- Choosing a representation: When to store joints, absolute EE poses, or relative EE deltas—and how that affects training.
- Pipeline customization guidance: How to swap robots/URDFs safely and tune bounds, step sizes, and options like IK initialization.
### Absolute vs relative EE control
The examples in this guide use absolute end effector (EE) poses because they are easy to reason about. In practice, relative EE deltas or joint position are often preferred as learning features.
With processors, you choose the learning features you want to use for your policy. This could be joints positions/velocities, absolute EE, or relative EE positions. You can also choose to store other features, such as joint torques, motor currents, etc.
## Three pipelines
We often compose three pipelines. Depending on your setup, some can be empty if action and observation spaces already match.
Each of these pipelines handle different conversions between different action and observation spaces. Below is a quick explanation of each pipeline.
1. Pipeline 1: Teleop action space → dataset action space (phone pose → EE targets)
2. Pipeline 2: Dataset action space → robot command space (EE targets → joints)
3. Pipeline 3: Robot observation space → dataset observation space (joints → EE pose)
Below is an example of the three pipelines that we use in the phone to SO-100 follower examples:
```python
phone_to_robot_ee_pose_processor = RobotProcessorPipeline[RobotAction, RobotAction]( # teleop -> dataset action
steps=[
MapPhoneActionToRobotAction(platform=teleop_config.phone_os),
EEReferenceAndDelta(
kinematics=kinematics_solver, end_effector_step_sizes={"x": 0.5, "y": 0.5, "z": 0.5}, motor_names=list(robot.bus.motors.keys()),
),
EEBoundsAndSafety(
end_effector_bounds={"min": [-1.0, -1.0, -1.0], "max": [1.0, 1.0, 1.0]}, max_ee_step_m=0.20,
),
GripperVelocityToJoint(),
],
to_transition=robot_action_to_transition,
to_output=transition_to_robot_action,
)
robot_ee_to_joints_processor = RobotProcessorPipeline[RobotAction, RobotAction]( # dataset action -> robot
steps=[
InverseKinematicsEEToJoints(
kinematics=kinematics_solver, motor_names=list(robot.bus.motors.keys()), initial_guess_current_joints=True,
),
],
to_transition=robot_action_to_transition,
to_output=transition_to_robot_action,
)
robot_joints_to_ee_pose = RobotProcessorPipeline[RobotObservation, RobotObservation]( # robot obs -> dataset obs
steps=[
ForwardKinematicsJointsToEE(kinematics=kinematics_solver, motor_names=list(robot.bus.motors.keys()))
],
to_transition=observation_to_transition,
to_output=transition_to_observation,
)
```
## Why to_transition / to_output
To convert from robot/teleoperator to pipeline and back, we use the `to_transition` and `to_output` pipeline adapters.
They standardize conversions to reduce boilerplate code, and form the bridge between the robot and teleoperators raw dictionaries and the pipeline’s `EnvTransition` format.
In the phone to SO-100 follower examples we use the following adapters:
- `robot_action_to_transition`: transforms the teleop action dict to a pipeline transition.
- `transition_to_robot_action`: transforms the pipeline transition to a robot action dict.
- `observation_to_transition`: transforms the robot observation dict to a pipeline transition.
- `transition_to_observation`: transforms the pipeline transition to a observation dict.
Checkout [src/lerobot/processor/converters.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/processor/converters.py) for more details.
## Dataset feature contracts
Dataset features are determined by the keys saved in the dataset. Each step can declare what features it modifies in a contract called `transform_features(...)`. Once you build a processor, the processor can then aggregate all of these features with `aggregate_pipeline_dataset_features()` and merge multiple feature dicts with `combine_feature_dicts(...)`.
Below is and example of how we declare features with the `transform_features` method in the phone to SO-100 follower examples:
```python
def transform_features(
self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
# We only use the ee pose in the dataset, so we don't need the joint positions
for n in self.motor_names:
features[PipelineFeatureType.ACTION].pop(f"{n}.pos", None)
# We specify the dataset features of this step that we want to be stored in the dataset
for k in ["x", "y", "z", "wx", "wy", "wz", "gripper_pos"]:
features[PipelineFeatureType.ACTION][f"ee.{k}"] = PolicyFeature(
type=FeatureType.STATE, shape=(1,)
)
return features
```
Here we declare what PolicyFeatures we modify in this step, so we know what features we can expect when we run the processor. These features can then be aggregated and used to create the dataset features.
Below is an example of how we aggregate and merge features in the phone to SO-100 record example:
```python
features=combine_feature_dicts(
# Run the feature contract of the pipelines
# This tells you how the features would look like after the pipeline steps
aggregate_pipeline_dataset_features(
pipeline=phone_to_robot_ee_pose_processor,
initial_features=create_initial_features(action=phone.action_features), # <- Action features we can expect, these come from our teleop device (phone) and action processor
use_videos=True,
),
aggregate_pipeline_dataset_features(
pipeline=robot_joints_to_ee_pose,
initial_features=create_initial_features(observation=robot.observation_features), # <- Observation features we can expect, these come from our robot and observation processor
use_videos=True,
patterns=["observation.state.ee"], # <- Here you could optionally filter the features we want to store in the dataset, with a specific pattern
),
),
```
How it works:
- `aggregate_pipeline_dataset_features(...)`: applies `transform_features` across the pipeline and filters by patterns (images included when `use_videos=True`, and state features included when `patterns` is specified).
- `combine_feature_dicts(...)`: combine multiple feature dicts.
- Recording with `record_loop(...)` uses `build_dataset_frame(...)` to build frames consistent with `dataset.features` before we call `add_frame(...)` to add the frame to the dataset.
## Guidance when customizing robot pipelines
You can store any of the following features as your action/observation space:
- Joint positions
- Absolute EE poses
- Relative EE deltas
- Other features: joint velocity, torques, etc.
Pick what you want to use for your policy action and observation space and configure/modify the pipelines and steps accordingly.
### Different robots
- You can easily reuse pipelines, for example to use another robot with phone teleop, modify the examples and swap the robot `RobotKinematics` (URDF) and `motor_names` to use your own robot with Phone teleop. Additionally you should ensure `target_frame_name` points to your gripper/wrist.
### Safety first
- When changing pipelines, start with tight bounds, implement safety steps when working with real robots.
- Its advised to start with simulation first and then move to real robots.
Thats it! We hope this guide helps you get started with customizing your robot pipelines, If you run into any issues at any point, jump into our [Discord community](https://discord.com/invite/s3KuuzsPFb) for support.
docs/source/reachy2.mdx
+309
View File
@@ -0,0 +1,309 @@
# Reachy 2
Reachy 2 is an open-source humanoid robot made by Pollen Robotics, specifically designed for the development of embodied AI and real-world applications.
Check out [Pollen Robotics website](https://www.pollen-robotics.com/reachy/), or access [Reachy 2 documentation](https://docs.pollen-robotics.com/) for more information on the platform!
## Teleoperate Reachy 2
Currently, there are two ways to teleoperate Reachy 2:
- Pollen Robotics’ VR teleoperation (not included in LeRobot).
- Robot-to-robot teleoperation (use one Reachy 2 to control another).
## Reachy 2 Simulation
**(Linux only)** You can run Reachy 2 in simulation (Gazebo or MuJoCo) using the provided [Docker image](https://hub.docker.com/r/pollenrobotics/reachy2_core).
1. Install [Docker Engine](https://docs.docker.com/engine/).
2. Run (for MuJoCo):
```
docker run --rm -it \
--name reachy \
--privileged \
--network host \
--ipc host \
--device-cgroup-rule='c 189:* rwm' \
--group-add audio \
-e ROS_DOMAIN_ID="$ROS_DOMAIN_ID" \
-e DISPLAY="$DISPLAY" \
-e RCUTILS_CONSOLE_OUTPUT_FORMAT="[{severity}]: {message}" \
-e REACHY2_CORE_SERVICE_FAKE="${REACHY2_CORE_SERVICE_FAKE:-true}" \
-v /dev:/dev \
-v "$HOME/.reachy_config":/home/reachy/.reachy_config_override \
-v "$HOME/.reachy.log":/home/reachy/.ros/log \
-v /usr/lib/x86_64-linux-gnu:/opt/host-libs \
--entrypoint /package/launch.sh \
pollenrobotics/reachy2_core:1.7.5.9_deploy \
start_rviz:=true start_sdk_server:=true mujoco:=true
```
> [!NOTE]
> If MuJoCo runs slowly (low simulation frequency), append `-e LD_LIBRARY_PATH="/opt/host-libs:$LD_LIBRARY_PATH" \` to the previous command to improve performance:
>
> ```
> docker run --rm -it \
> --name reachy \
> --privileged \
> --network host \
> --ipc host \
> --device-cgroup-rule='c 189:* rwm' \
> --group-add audio \
> -e ROS_DOMAIN_ID="$ROS_DOMAIN_ID" \
> -e DISPLAY="$DISPLAY" \
> -e RCUTILS_CONSOLE_OUTPUT_FORMAT="[{severity}]: {message}" \
> -e REACHY2_CORE_SERVICE_FAKE="${REACHY2_CORE_SERVICE_FAKE:-true}" \
> -e LD_LIBRARY_PATH="/opt/host-libs:$LD_LIBRARY_PATH" \
> -v /dev:/dev \
> -v "$HOME/.reachy_config":/home/reachy/.reachy_config_override \
> -v "$HOME/.reachy.log":/home/reachy/.ros/log \
> -v /usr/lib/x86_64-linux-gnu:/opt/host-libs \
> --entrypoint /package/launch.sh \
> pollenrobotics/reachy2_core:1.7.5.9_deploy \
> start_rviz:=true start_sdk_server:=true mujoco:=true
> ```
## Setup
### Prerequisites
- On your robot, check the **service images** meet the minimum versions:
- **reachy2-core >= 1.7.5.2**
- **webrtc >= 2.0.1.1**
Then, if you want to use VR teleoperation:
- Install the [Reachy 2 teleoperation application](https://docs.pollen-robotics.com/teleoperation/teleoperation-introduction/discover-teleoperation/).
Use version **>=v1.2.0**
We recommend using two computers: one for teleoperation (Windows required) and another for recording with LeRobot.
### Install LeRobot
Follow the [installation instructions](https://github.com/huggingface/lerobot#installation) to install LeRobot.
Install LeRobot with Reachy 2 dependencies:
```bash
pip install -e ".[reachy2]"
```
### (Optional but recommended) Install pollen_data_acquisition_server
How you manage Reachy 2 recording sessions is up to you, but the **easiest** way is to use this server so you can control sessions directly from the VR teleoperation app.
> **Note:** Currently, only the VR teleoperation application works as a client for this server, so this step primarily targets teleoperation. You’re free to develop custom clients to manage sessions to your needs.
In your LeRobot environment, install the server from source:
```bash
git clone https://github.com/pollen-robotics/pollen_data_acquisition_server.git
cd pollen_data_acquisition_server
pip install -e .
```
Find the [pollen_data_acquisition_server documentation here](https://github.com/pollen-robotics/pollen_data_acquisition_server).
## Step 1: Recording
### Get Reachy 2 IP address
Before starting teleoperation and data recording, find the [robot's IP address](https://docs.pollen-robotics.com/getting-started/setup-reachy2/connect-reachy2/).
We strongly recommend connecting all devices (PC and robot) via **Ethernet**.
### Launch recording
There are two ways to manage recording sessions when using the Reachy 2 VR teleoperation application:
- **Using the data acquisition server (recommended for VR teleop)**: The VR app orchestrates sessions (via the server it tells LeRobot when to create datasets, start/stop episodes) while also controlling the robot’s motions.
- **Using LeRobot’s record script**: LeRobot owns session control and decides when to start/stop episodes. If you also use the VR teleop app, it’s only for motion control.
### Option 1: Using Pollen data acquisition server (recommended for VR teleop)
Make sure you have installed pollen_data_acquisition_server, as explained in the Setup section.
Launch the data acquisition server to be able to manage your session directly from the teleoperation application:
```bash
python -m pollen_data_acquisition_server.server
```
Then get into the teleoperation application and choose "Data acquisition session".
You can finally setup your session by following the screens displayed.
> Even without the VR app, you can use the `pollen_data_acquisition_server` with your own client implementation.
### Option 2: Using lerobot.record
Reachy 2 is fully supported by LeRobot’s recording features.
If you choose this option but still want to use the VR teleoperation application, select "Standard session" in the app.
**Example: start a recording without the mobile base:**
First add reachy2 and reachy2_teleoperator to the imports of the record script. Then you can use the following command:
```bash
lerobot-record \
--robot.type=reachy2 \
--robot.ip_address=192.168.0.200 \
--robot.id=r2-0000 \
--robot.use_external_commands=true \
--robot.with_mobile_base=false \
--teleop.type=reachy2_teleoperator \
--teleop.ip_address=192.168.0.200 \
--teleop.with_mobile_base=false \
--robot.with_torso_camera=true \
--dataset.repo_id=pollen_robotics/record_test \
--dataset.single_task="Reachy 2 recording test" \
--dataset.num_episodes=1 \
--dataset.episode_time_s=5 \
--dataset.fps=15 \
--dataset.push_to_hub=true \
--dataset.private=true \
--dataset.streaming_encoding=true \
--dataset.encoder_threads=2 \
# --dataset.rgb_encoder.vcodec=auto \
--display_data=true
```
#### Specific Options
**Extended setup overview (all options included):**
```bash
lerobot-record \
--robot.type=reachy2 \
--robot.ip_address=192.168.0.200 \
--robot.use_external_commands=true \
--robot.with_mobile_base=true \
--robot.with_l_arm=true \
--robot.with_r_arm=true \
--robot.with_neck=true \
--robot.with_antennas=true \
--robot.with_left_teleop_camera=true \
--robot.with_right_teleop_camera=true \
--robot.with_torso_camera=false \
--robot.camera_width=640 \
--robot.camera_height=480 \
--robot.disable_torque_on_disconnect=false \
--robot.max_relative_target=5.0 \
--teleop.type=reachy2_teleoperator \
--teleop.ip_address=192.168.0.200 \
--teleop.use_present_position=false \
--teleop.with_mobile_base=false \
--teleop.with_l_arm=true \
--teleop.with_r_arm=true \
--teleop.with_neck=true \
--teleop.with_antennas=true \
--dataset.repo_id=pollen_robotics/record_test \
--dataset.single_task="Reachy 2 recording test" \
--dataset.num_episodes=1 \
--dataset.episode_time_s=5 \
--dataset.fps=15 \
--dataset.push_to_hub=true \
--dataset.private=true \
--dataset.streaming_encoding=true \
--dataset.encoder_threads=2 \
# --dataset.rgb_encoder.vcodec=auto \
--display_data=true
```
##### `--robot.use_external_commands`
Determine whether LeRobot robot.send_action() sends commands to the robot.
**Must** be set to false while using the VR teleoperation application, as the app already sends commands.
##### `--teleop.use_present_position`
Determine whether the teleoperator reads the goal or present position of the robot.
Must be set to true if a compliant Reachy 2 is used to control another one.
##### Use the relevant parts
From our initial tests, recording **all** joints when only some are moving can reduce model quality with certain policies.
To avoid this, you can exclude specific parts from recording and replay using:
```bash
--robot.with_<part>=false
```
with `<part>` being one of : `mobile_base`, `l_arm`, `r_arm", `neck`, `antennas`.
It determine whether the corresponding part is recorded in the observations. True if not set.
By default, **all parts are recorded**.
The same per-part mechanism is available in `reachy2_teleoperator` as well.
```bash
--teleop.with\_<part>
```
with `<part>` being one of : `mobile_base`, `l_arm`, `r_arm", `neck`, `antennas`.
Determine whether the corresponding part is recorded in the actions. True if not set.
> **Important:** In a given session, the **enabled parts must match** on both the robot and the teleoperator.
> For example, if the robot runs with `--robot.with_mobile_base=false`, the teleoperator must disable the same part `--teleoperator.with_mobile_base=false`.
##### Use the relevant cameras
You can do the same for **cameras**. Enable or disable each camera with default parameters using:
```bash
--robot.with_left_teleop_camera=<true|false> \
--robot.with_right_teleop_camera=<true|false> \
--robot.with_torso_camera=<true|false>
```
By default, no camera is recorded, all camera arguments are set to `false`.
If you want to, you can use custom `width` and `height` parameters for Reachy 2's cameras using the `--robot.camera_width` & `--robot.camera_height` argument:
```bash
--robot.camera_width=1920 \
--robot.camera_height=1080
```
This will change the resolution of all 3 default robot cameras (enabled by the above bool arguments).
If you want, you can add additional cameras other than the ones in the robot as usual with:
```bash
--robot.cameras="{ extra: {type: opencv, index_or_path: 42, width: 640, height: 480, fps: 30}}" \
```
## Step 2: Replay
Make sure the robot is configured with the same parts as the dataset:
```bash
lerobot-replay \
--robot.type=reachy2 \
--robot.ip_address=192.168.0.200 \
--robot.use_external_commands=false \
--robot.with_mobile_base=false \
--dataset.repo_id=pollen_robotics/record_test \
--dataset.episode=0
```
## Step 3: Train
```bash
lerobot-train \
--dataset.repo_id=pollen_robotics/record_test \
--policy.type=act \
--output_dir=outputs/train/reachy2_test \
--job_name=reachy2 \
--policy.device=mps \
--wandb.enable=true \
--policy.repo_id=pollen_robotics/record_test_policy
```
## Step 4: Evaluate
```bash
lerobot-eval \
--robot.type=reachy2 \
--robot.ip_address=192.168.0.200 \
--dataset.repo_id=pollen_robotics/eval_record_test \
--dataset.single_task="Evaluate reachy2 policy" \
--dataset.num_episodes=10 \
--policy.path=outputs/train/reachy2_test/checkpoints/last/pretrained_model
```
docs/source/rebot_b601.mdx
+186
View File
@@ -0,0 +1,186 @@
# reBot B601-DM
[reBot B601-DM](https://wiki.seeedstudio.com/rebot_arm_b601_dm_lerobot/) is an open-source, low-cost robot arm from Seeed Studio for embodied-AI and imitation learning. It comes as a **follower** arm (the `B601-DM`, a 6-DOF arm plus gripper driven by Damiao CAN motors) and a **leader** arm (the `StarArm102` / `reBot Arm 102`, driven by FashionStar UART smart servos) used to teleoperate it.
This page covers **calibration** and **teleoperation** for both single-arm and bimanual (dual-arm) setups.
<div style="display: flex; align-items: center; gap: 10px;">
<img
src="https://files.seeedstudio.com/wiki/robotics/projects/lerobot/b601dm_zeroposition.jpg"
alt="reBot B601-DM follower arm at its zero position"
width="48%"
/>
<img
src="https://files.seeedstudio.com/wiki/robotics/projects/lerobot/102_zeroposition.jpg"
alt="reBot Arm 102 leader arm at its zero position"
width="48%"
/>
</div>
_Left: the B601-DM follower at its zero position. Right: the reBot Arm 102 leader at its zero position. Images courtesy of [Seeed Studio](https://wiki.seeedstudio.com/rebot_arm_b601_dm_lerobot/)._
## Install LeRobot 🤗
Follow our [Installation Guide](./installation), then install the reBot support:
```bash
pip install -e ".[rebot]"
```
This pulls in `motorbridge` (CAN motor control for the B601-DM follower) and `motorbridge-smart-servo` (FashionStar UART servos for the reBot Arm 102 leader).
## Registered device types
| Type | Kind |
| ------------------------ | -------------------------------------------- |
| `rebot_b601_follower` | single-arm B601-DM follower robot |
| `bi_rebot_b601_follower` | bimanual (dual-arm) follower robot |
| `rebot_102_leader` | single-arm reBot Arm 102 leader teleoperator |
| `bi_rebot_102_leader` | bimanual (dual-arm) leader teleoperator |
The bimanual types compose two single-arm instances and namespace each arm's
observation/action keys with a `left_` / `right_` prefix. Per-arm settings are
passed through nested `left_arm_config.*` / `right_arm_config.*` arguments.
## Find the USB ports
For each device, find the USB port associated with its motor bus using:
```bash
lerobot-find-port
```
<Tip warning={true}>
On Linux, remove `brltty` (`sudo apt remove brltty`) so it does not hold the
leader's USB serial port. You may also need to grant access to the serial
devices: `sudo chmod 666 /dev/ttyACM* /dev/ttyUSB*`.
</Tip>
## Calibration
Neither arm stores a persistent hardware calibration: every time it connects, the motors are re-zeroed against the pose the arm is physically holding. Calibration simply records that zero pose. When prompted, **manually move the arm to its zero position** (the default sit-down pose shown above, gripper fully closed) and press <kbd>ENTER</kbd>.
### Follower (B601-DM)
<hfoptions id="calibrate-follower">
<hfoption id="Single arm">
```bash
lerobot-calibrate \
--robot.type=rebot_b601_follower \
--robot.port=/dev/ttyACM0 \
--robot.id=follower \
--robot.can_adapter=damiao
```
</hfoption>
<hfoption id="Dual arm">
Connect the bimanual follower; calibration runs for the left arm, then the right arm.
```bash
lerobot-calibrate \
--robot.type=bi_rebot_b601_follower \
--robot.id=bi_follower \
--robot.left_arm_config.port=/dev/ttyACM0 \
--robot.left_arm_config.can_adapter=damiao \
--robot.right_arm_config.port=/dev/ttyACM1 \
--robot.right_arm_config.can_adapter=damiao
```
Per-arm calibration files are saved with `_left` / `_right` suffixes on the id.
</hfoption>
</hfoptions>
### Leader (reBot Arm 102)
<hfoptions id="calibrate-leader">
<hfoption id="Single arm">
```bash
lerobot-calibrate \
--teleop.type=rebot_102_leader \
--teleop.port=/dev/ttyUSB0 \
--teleop.id=leader
```
</hfoption>
<hfoption id="Dual arm">
```bash
lerobot-calibrate \
--teleop.type=bi_rebot_102_leader \
--teleop.id=bi_leader \
--teleop.left_arm_config.port=/dev/ttyUSB0 \
--teleop.right_arm_config.port=/dev/ttyUSB1
```
</hfoption>
</hfoptions>
## Teleoperation
Once both arms are calibrated, drive the follower with the leader. The follower talks to its CAN bus through a Damiao serial bridge (`can_adapter=damiao`, the default) or a SocketCAN adapter (`can_adapter=socketcan`). See the [OpenArm page](./openarm) for more details on the SocketCAN adapter configuration.
<hfoptions id="teleoperate">
<hfoption id="Single arm">
```bash
lerobot-teleoperate \
--robot.type=rebot_b601_follower \
--robot.port=/dev/ttyACM0 \
--robot.id=follower \
--robot.can_adapter=damiao \
--teleop.type=rebot_102_leader \
--teleop.port=/dev/ttyUSB0 \
--teleop.id=leader
```
</hfoption>
<hfoption id="Dual arm">
The bimanual leader and follower reuse the single-arm classes; each arm is
configured through nested `left_arm_config.*` / `right_arm_config.*` arguments,
so a bimanual reBot Arm 102 leader drives a bimanual B601-DM follower.
```bash
lerobot-teleoperate \
--robot.type=bi_rebot_b601_follower \
--robot.id=bi_follower \
--robot.left_arm_config.port=/dev/ttyACM0 \
--robot.left_arm_config.can_adapter=damiao \
--robot.right_arm_config.port=/dev/ttyACM1 \
--robot.right_arm_config.can_adapter=damiao \
--teleop.type=bi_rebot_102_leader \
--teleop.id=bi_leader \
--teleop.left_arm_config.port=/dev/ttyUSB0 \
--teleop.right_arm_config.port=/dev/ttyUSB1
```
</hfoption>
</hfoptions>
<Tip>
The leader and follower share the same joint names (`shoulder_pan,
shoulder_lift, elbow_flex, wrist_flex, wrist_yaw, wrist_roll, gripper`), so
leader actions map directly onto the follower.
</Tip>
If the motion of a joint is reversed, flip its sign in the leader's `joint_directions` (the gripper also carries a scale to widen its range to the follower):
```bash
lerobot-teleoperate \
--robot.type=rebot_b601_follower \
--robot.port=/dev/ttyACM0 \
--robot.can_adapter=damiao \
--teleop.type=rebot_102_leader \
--teleop.port=/dev/ttyUSB0 \
--teleop.joint_directions='{"shoulder_pan":-1,"shoulder_lift":-1,"elbow_flex":1,"wrist_flex":1,"wrist_yaw":1,"wrist_roll":-1,"gripper":-6}'
```
## Recording datasets
Swap `lerobot-teleoperate` for `lerobot-record` (with the same `--robot.*` / `--teleop.*` arguments, plus `--dataset.*`) to record demonstrations for training. See [Imitation Learning for Robots](./il_robots) for the full workflow.
For hardware assembly and wiring, see the [Seeed Studio reBot wiki](https://wiki.seeedstudio.com/rebot_arm_b601_dm_lerobot/).
docs/source/rename_map.mdx
+103
View File
@@ -0,0 +1,103 @@
# Rename Map and Empty Cameras
When you train, evaluate, or record with a robot policy, your **dataset** or **environment** provides observations under one set of keys (e.g. `observation.images.front`, `observation.images.eagle`), while your **policy** expects another (e.g. `observation.images.image`, `observation.images.image2`). The **rename map** bridges that gap without changing the policy or data source.
> **Scope:** The rename map only renames **observation** keys (images and state). Action keys are not affected.
## Why observation keys don't always match
Policies have a fixed set of **input feature names** baked into their pretrained config. For example:
- [pi0fast-libero](https://huggingface.co/lerobot/pi0fast-libero) expects `observation.images.base_0_rgb` and `observation.images.left_wrist_0_rgb`.
- [xvla-base](https://huggingface.co/lerobot/xvla-base) expects `observation.images.image`, `observation.images.image2`, and `observation.images.image3`.
Your dataset might use different names entirely (e.g. `observation.images.front`, `observation.images.eagle`, `observation.images.glove`), and your eval environment might use yet another set. Rather than editing the policy config or renaming columns in the dataset, you pass a **rename map**: a JSON dictionary that maps source keys to the keys the policy expects. Renaming happens inside the preprocessor pipeline, so the policy always sees its expected keys.
## Using the rename map
Pass the mapping as a JSON string on the command line. The convention is always:
```
--rename_map='{"source_key": "policy_key", ...}'
```
where **source_key** is what the dataset or environment provides, and **policy_key** is what the policy expects.
Only listed keys are renamed; everything else passes through unchanged. Order of entries doesn't matter.
Supported policies: **PI0**, **PI05**, **PI0Fast**, **SmolVLA**, and **XVLA**.
### Training
Suppose you fine-tune [lerobot/xvla-base](https://huggingface.co/lerobot/xvla-base) on a dataset with images under `observation.images.front`, `observation.images.eagle`, and `observation.images.glove`. XVLA expects `observation.images.image`, `observation.images.image2`, and `observation.images.image3`:
```bash
lerobot-train \
--dataset.repo_id=YOUR_DATASET \
--output_dir=./outputs/xvla_training \
--job_name=xvla_training \
--policy.path="lerobot/xvla-base" \
--policy.repo_id="HF_USER/xvla-your-robot" \
--policy.dtype=bfloat16 \
--policy.action_mode=auto \
--steps=20000 \
--policy.device=cuda \
--policy.freeze_vision_encoder=false \
--policy.freeze_language_encoder=false \
--policy.train_policy_transformer=true \
--policy.train_soft_prompts=true \
--rename_map='{"observation.images.front": "observation.images.image", "observation.images.eagle": "observation.images.image2", "observation.images.glove": "observation.images.image3"}'
```
### Evaluation
A policy that expects `observation.images.base_0_rgb` and `observation.images.left_wrist_0_rgb` (e.g. [pi0fast-libero](https://huggingface.co/lerobot/pi0fast-libero)), but the LIBERO environment returns `observation.images.image` and `observation.images.image2`:
```bash
lerobot-eval \
--policy.path=lerobot/pi0fast-libero \
--env.type=libero \
... \
--rename_map='{"observation.images.image": "observation.images.base_0_rgb", "observation.images.image2": "observation.images.left_wrist_0_rgb"}'
```
## Alternative: edit the policy config directly
If you always use the same dataset or environment, you can **edit the policy's `config.json`** so its observation keys match your data source. Then no rename map is needed.
The tradeoff: modifying the policy config ties it to one data source. A rename map keeps one policy usable across many datasets and environments.
## Empty cameras: fewer views than the policy expects
Some policies are built for a fixed number of image inputs. If your dataset has fewer cameras, you can set **`empty_cameras`** in the policy config instead of modifying the model architecture.
### How it works
Setting `empty_cameras=N` adds N placeholder image features to the policy config, named:
```
observation.images.empty_camera_0
observation.images.empty_camera_1
...
```
At runtime, these keys have no corresponding data in the batch. The policy fills them with masked dummy tensors (padded with `-1` for SigLIP-based vision encoders, with a zero attention mask), so the extra image slots are effectively ignored during training and inference.
### Example
XVLA-base has three visual inputs and `empty_cameras=0` by default. Your dataset only has two cameras:
1. Set `--policy.empty_cameras=1`.
2. The config adds a third key: `observation.images.empty_camera_0`.
3. Use the rename map for your two real cameras as usual.
4. The third slot is masked out — no fake images needed in your dataset.
## Quick reference
| Goal | What to do |
| --------------------------------------- | --------------------------------------------------------------------------- |
| Dataset keys ≠ policy keys | `--rename_map='{"dataset_key": "policy_key", ...}'` |
| Env keys ≠ policy keys (eval) | `--rename_map='{"env_key": "policy_key", ...}'` |
| Rollout with different keys (inference) | `--rename_map='{"source_key": "policy_key", ...}'`. |
| Fewer cameras than policy expects | `--policy.empty_cameras=N` (supported by PI0, PI05, PI0Fast, SmolVLA, XVLA) |
| Avoid passing a rename map | Edit the policy's `config.json` so its keys match your data source |
docs/source/robocasa.mdx
+188
View File
@@ -0,0 +1,188 @@
# RoboCasa365
[RoboCasa365](https://robocasa.ai) is a large-scale simulation framework for training and benchmarking **generalist robots** in everyday kitchen tasks. It ships 365 diverse manipulation tasks across 2,500 kitchen environments, 3,200+ object assets and 600+ hours of human demonstration data, on a PandaOmron 12-DOF mobile manipulator (Franka arm on a holonomic base).
- Paper: [RoboCasa: Large-Scale Simulation of Everyday Tasks for Generalist Robots](https://arxiv.org/abs/2406.02523)
- GitHub: [robocasa/robocasa](https://github.com/robocasa/robocasa)
- Project website: [robocasa.ai](https://robocasa.ai)
- Pretrained policy: [`lerobot/smolvla_robocasa`](https://huggingface.co/lerobot/smolvla_robocasa)
- Single-task dataset (CloseFridge): [`pepijn223/robocasa_CloseFridge`](https://huggingface.co/datasets/pepijn223/robocasa_CloseFridge)
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/robocasa-banner.webp"
alt="RoboCasa365 benchmark overview"
width="85%"
/>
## Available tasks
RoboCasa365 organizes its 365 tasks into two families and three upstream benchmark groups that LeRobot exposes as first-class `--env.task` shortcuts:
| Family | Tasks | Description |
| --------- | ----- | ------------------------------------------------------------------------------- |
| Atomic | ~65 | Single-skill tasks: pick-and-place, door/drawer manipulation, appliance control |
| Composite | ~300 | Multi-step tasks across 60+ categories: cooking, cleaning, organizing, etc. |
**Atomic task examples:** `CloseFridge`, `OpenDrawer`, `OpenCabinet`, `TurnOnMicrowave`, `TurnOffStove`, `NavigateKitchen`, `PickPlaceCounterToStove`.
**Composite task categories:** baking, boiling, brewing, chopping, clearing table, defrosting food, loading dishwasher, making tea, microwaving food, washing dishes, and more.
`--env.task` accepts three forms:
- a single task name (`CloseFridge`)
- a comma-separated list (`CloseFridge,OpenBlenderLid,PickPlaceCoffee`)
- a benchmark-group shortcut — `atomic_seen`, `composite_seen`, `composite_unseen`, `pretrain50`, `pretrain100`, `pretrain200`, `pretrain300` — which auto-expands to the upstream task list and auto-sets the dataset `split` (`target` or `pretrain`).
## Installation
RoboCasa and its dependency `robosuite` are not published on PyPI, and RoboCasa's own `setup.py` hardcodes `lerobot==0.3.3`, which conflicts with this repo's `lerobot`. LeRobot therefore does **not** expose a `robocasa` extra — install the two packages manually as editable clones (using `--no-deps` on `robocasa` to skip its shadowed `lerobot` pin):
```bash
# After following the standard LeRobot installation instructions.
git clone https://github.com/robocasa/robocasa.git ~/robocasa
git clone https://github.com/ARISE-Initiative/robosuite.git ~/robosuite
pip install -e ~/robocasa --no-deps
pip install -e ~/robosuite
# Robocasa's runtime deps (the ones its setup.py would have pulled, minus
# the bad lerobot pin).
pip install numpy numba scipy mujoco pygame Pillow opencv-python \
pyyaml pynput tqdm termcolor imageio h5py lxml hidapi \
tianshou gymnasium
python -m robocasa.scripts.setup_macros
# Lightweight assets (lightwheel object meshes + textures). Enough for
# the default env out of the box.
python -m robocasa.scripts.download_kitchen_assets \
--type tex tex_generative fixtures_lw objs_lw
# Optional: full objaverse/aigen registries (~30GB) for richer object
# variety. Enable at eval time via --env.obj_registries (see below).
# python -m robocasa.scripts.download_kitchen_assets --type objs_objaverse
```
<Tip>
RoboCasa requires MuJoCo. Set the rendering backend before training or evaluation:
```bash
export MUJOCO_GL=egl # for headless servers (HPC, cloud)
```
</Tip>
### Object registries
By default the env samples objects only from the `lightwheel` registry (what `--type objs_lw` ships), which avoids a `Probabilities contain NaN` crash when the objaverse / aigen packs aren't on disk. If you've downloaded the full asset set, enable the full registry at runtime:
```bash
--env.obj_registries='[objaverse,lightwheel]'
```
## Evaluation
All eval snippets below mirror the CI command (see `.github/workflows/benchmark_tests.yml`). The `--rename_map` argument maps RoboCasa's native camera keys (`robot0_agentview_left` / `robot0_eye_in_hand` / `robot0_agentview_right`) onto the three-camera (`camera1` / `camera2` / `camera3`) input layout the released `smolvla_robocasa` policy was trained on.
### Single-task evaluation (recommended for quick iteration)
```bash
lerobot-eval \
--policy.path=lerobot/smolvla_robocasa \
--env.type=robocasa \
--env.task=CloseFridge \
--eval.batch_size=1 \
--eval.n_episodes=20 \
--eval.use_async_envs=false \
--policy.device=cuda \
'--rename_map={"observation.images.robot0_agentview_left": "observation.images.camera1", "observation.images.robot0_eye_in_hand": "observation.images.camera2", "observation.images.robot0_agentview_right": "observation.images.camera3"}'
```
### Multi-task evaluation
Pass a comma-separated list of tasks:
```bash
lerobot-eval \
--policy.path=lerobot/smolvla_robocasa \
--env.type=robocasa \
--env.task=CloseFridge,OpenCabinet,OpenDrawer,TurnOnMicrowave,TurnOffStove \
--eval.batch_size=1 \
--eval.n_episodes=20 \
--eval.use_async_envs=false \
--policy.device=cuda \
'--rename_map={"observation.images.robot0_agentview_left": "observation.images.camera1", "observation.images.robot0_eye_in_hand": "observation.images.camera2", "observation.images.robot0_agentview_right": "observation.images.camera3"}'
```
### Benchmark-group evaluation
Run an entire upstream group (e.g. all 18 `atomic_seen` tasks with `split=target`):
```bash
lerobot-eval \
--policy.path=lerobot/smolvla_robocasa \
--env.type=robocasa \
--env.task=atomic_seen \
--eval.batch_size=1 \
--eval.n_episodes=20 \
--eval.use_async_envs=false \
--policy.device=cuda \
'--rename_map={"observation.images.robot0_agentview_left": "observation.images.camera1", "observation.images.robot0_eye_in_hand": "observation.images.camera2", "observation.images.robot0_agentview_right": "observation.images.camera3"}'
```
### Recommended evaluation episodes
**20 episodes per task** for reproducible benchmarking. Matches the protocol used in published results.
## Policy inputs and outputs
**Observations** (raw RoboCasa camera names are preserved verbatim):
- `observation.state` — 16-dim proprioceptive state (base position, base quaternion, relative end-effector position, relative end-effector quaternion, gripper qpos)
- `observation.images.robot0_agentview_left` — left agent view, 256×256 HWC uint8
- `observation.images.robot0_eye_in_hand` — wrist camera view, 256×256 HWC uint8
- `observation.images.robot0_agentview_right` — right agent view, 256×256 HWC uint8
**Actions:**
- Continuous control in `Box(-1, 1, shape=(12,))` — base motion (4D) + control mode (1D) + end-effector position (3D) + end-effector rotation (3D) + gripper (1D).
## Training
### Single-task example
A ready-to-use single-task dataset is on the Hub:
[`pepijn223/robocasa_CloseFridge`](https://huggingface.co/datasets/pepijn223/robocasa_CloseFridge).
Fine-tune a SmolVLA base on `CloseFridge`:
```bash
lerobot-train \
--policy.type=smolvla \
--policy.repo_id=${HF_USER}/smolvla_robocasa_CloseFridge \
--policy.load_vlm_weights=true \
--policy.push_to_hub=true \
--dataset.repo_id=pepijn223/robocasa_CloseFridge \
--env.type=robocasa \
--env.task=CloseFridge \
--output_dir=./outputs/smolvla_robocasa_CloseFridge \
--steps=100000 \
--batch_size=4 \
--env_eval_freq=5000 \
--eval.batch_size=1 \
--eval.n_episodes=5 \
--save_freq=10000
```
Evaluate the resulting checkpoint:
```bash
lerobot-eval \
--policy.path=${HF_USER}/smolvla_robocasa_CloseFridge \
--env.type=robocasa \
--env.task=CloseFridge \
--eval.batch_size=1 \
--eval.n_episodes=20
```
## Reproducing published results
The released checkpoint [`lerobot/smolvla_robocasa`](https://huggingface.co/lerobot/smolvla_robocasa) is evaluated with the commands in the [Evaluation](#evaluation) section. CI runs a 10-atomic-task smoke eval (one episode each) on every PR touching the benchmark, picking fixture-centric tasks that don't require the objaverse asset pack.
docs/source/robocerebra.mdx
+99
View File
@@ -0,0 +1,99 @@
# RoboCerebra
[RoboCerebra](https://robocerebra-project.github.io/) is a long-horizon manipulation benchmark that evaluates **high-level reasoning, planning, and memory** in VLAs. Episodes chain multiple sub-goals with language-grounded intermediate instructions, built on top of LIBERO's simulator stack (MuJoCo + robosuite, Franka Panda 7-DOF).
- Paper: [RoboCerebra: A Large-scale Benchmark for Long-horizon Robotic Manipulation Evaluation](https://arxiv.org/abs/2506.06677)
- Project website: [robocerebra-project.github.io](https://robocerebra-project.github.io/)
- Dataset: [`lerobot/robocerebra_unified`](https://huggingface.co/datasets/lerobot/robocerebra_unified) — LeRobot v3.0, 6,660 episodes / 571,116 frames at 20 fps, 1,728 language-grounded sub-tasks.
- Pretrained policy: [`lerobot/smolvla_robocerebra`](https://huggingface.co/lerobot/smolvla_robocerebra)
## Available tasks
RoboCerebra reuses LIBERO's simulator, so evaluation runs against the LIBERO `libero_10` long-horizon suite:
| Suite | CLI name | Tasks | Description |
| --------- | ----------- | ----- | ------------------------------------------------------------- |
| LIBERO-10 | `libero_10` | 10 | Long-horizon kitchen/living room tasks chaining 3–6 sub-goals |
Each RoboCerebra episode in the dataset is segmented into multiple sub-tasks with natural-language instructions, which the unified dataset exposes as independent supervision signals.
## Installation
RoboCerebra piggybacks on LIBERO, so the `libero` extra is all you need:
```bash
pip install -e ".[libero]"
```
<Tip>
RoboCerebra requires Linux (MuJoCo / robosuite). Set the rendering backend before training or evaluation:
```bash
export MUJOCO_GL=egl # for headless servers (HPC, cloud)
```
</Tip>
## Evaluation
RoboCerebra eval runs against LIBERO's `libero_10` suite with RoboCerebra's camera naming (`image` + `wrist_image`) and an extra empty-camera slot so a three-view-trained policy receives the expected input layout:
```bash
lerobot-eval \
--policy.path=lerobot/smolvla_robocerebra \
--env.type=libero \
--env.task=libero_10 \
--env.fps=20 \
--env.obs_type=pixels_agent_pos \
--env.observation_height=256 \
--env.observation_width=256 \
'--env.camera_name_mapping={"agentview_image": "image", "robot0_eye_in_hand_image": "wrist_image"}' \
--eval.batch_size=1 \
--eval.n_episodes=10 \
--eval.use_async_envs=false \
--policy.device=cuda \
'--rename_map={"observation.images.image": "observation.images.camera1", "observation.images.wrist_image": "observation.images.camera2"}' \
--policy.empty_cameras=1
```
### Recommended evaluation episodes
**10 episodes per task** across the `libero_10` suite (100 total) for reproducible benchmarking. Matches the protocol used in the RoboCerebra paper.
## Policy inputs and outputs
**Observations:**
- `observation.state` — 8-dim proprioceptive state (7 joint positions + gripper)
- `observation.images.image` — third-person view, 256×256 HWC uint8
- `observation.images.wrist_image` — wrist-mounted camera view, 256×256 HWC uint8
**Actions:**
- Continuous control in `Box(-1, 1, shape=(7,))` — end-effector delta (6D) + gripper (1D)
## Training
The unified dataset at [`lerobot/robocerebra_unified`](https://huggingface.co/datasets/lerobot/robocerebra_unified) exposes two RGB streams and language-grounded sub-task annotations:
| Feature | Shape | Description |
| -------------------------------- | ------------- | -------------------- |
| `observation.images.image` | (256, 256, 3) | Third-person view |
| `observation.images.wrist_image` | (256, 256, 3) | Wrist-mounted camera |
| `observation.state` | (8,) | Joint pos + gripper |
| `action` | (7,) | EEF delta + gripper |
Fine-tune a SmolVLA base on it:
```bash
lerobot-train \
--policy.path=lerobot/smolvla_base \
--dataset.repo_id=lerobot/robocerebra_unified \
--env.type=libero \
--env.task=libero_10 \
--output_dir=outputs/smolvla_robocerebra
```
## Reproducing published results
The released checkpoint [`lerobot/smolvla_robocerebra`](https://huggingface.co/lerobot/smolvla_robocerebra) was trained on `lerobot/robocerebra_unified` and evaluated with the command in the [Evaluation](#evaluation) section. CI runs the same command with `--eval.n_episodes=1` as a smoke test on every PR touching the benchmark.
docs/source/robometer.mdx
+185
View File
@@ -0,0 +1,185 @@
# ROBOMETER
ROBOMETER is a **general-purpose video-language robotic reward model**. It predicts dense, frame-level task progress and frame-level success from a trajectory video and a task description.
**Paper**: [ROBOMETER: Scaling General-Purpose Robotic Reward Models via Trajectory Comparisons](https://arxiv.org/abs/2603.02115)
**Project**: [robometer.github.io](https://robometer.github.io/)
**Original code**: [github.com/robometer/robometer](https://github.com/robometer/robometer)
**Checkpoint**: [lerobot/Robometer-4B](https://huggingface.co/lerobot/Robometer-4B)
## Overview
ROBOMETER builds on `Qwen/Qwen3-VL-4B-Instruct` and adds three lightweight prediction heads:
- **Progress head**: predicts per-frame task progress in `[0, 1]`.
- **Success head**: predicts per-frame task success probability.
- **Preference head**: predicts which of two trajectories better completes the task during training.
The paper trains ROBOMETER with a composite objective:
```text
L = L_pref + L_prog + L_succ
```
The LeRobot integration is currently **inference-only**. It preserves the preference head so that the published `Robometer-4B` checkpoint loads without remapping, but `compute_reward()` queries the progress or success head only.
## What the LeRobot Integration Covers
- Standard `reward_model.type=robometer` configuration through LeRobot.
- Qwen3-VL image and text preprocessing through `RobometerEncoderProcessorStep`.
- LeRobot reward-model save/load APIs through `PreTrainedRewardModel`.
- Dense, frame-level progress and success predictions internally.
- A scalar reward through `compute_reward()` for downstream LeRobot reward-model usage.
This page focuses on using the published ROBOMETER checkpoint as a zero-shot reward model. Training ROBOMETER from scratch is outside the current LeRobot integration.
## Installation Requirements
1. Install LeRobot by following the [Installation Guide](./installation).
2. Install the ROBOMETER dependencies:
```bash
pip install -e ".[robometer]"
```
If you use `uv` directly from a source checkout:
```bash
uv sync --extra robometer
```
ROBOMETER uses a Qwen3-VL-4B backbone, so GPU inference is strongly recommended.
## Model Inputs and Outputs
ROBOMETER expects:
- A trajectory video or sequence of frames.
- A natural-language task description.
In LeRobot datasets, the preprocessor reads:
| Config field | Default | Meaning |
| ------------------------- | ------------------------ | ----------------------------------------------------- |
| `reward_model.image_key` | `observation.images.top` | Camera/video observation used by ROBOMETER |
| `reward_model.task_key` | `task` | Key in complementary data that stores the task string |
| `reward_model.max_frames` | `8` | Maximum number of frames passed to ROBOMETER |
The model predicts per-frame progress and success internally. The LeRobot reward API returns a scalar per sample:
- `reward_output="progress"` (default): return the last-frame progress, clamped to `[0, 1]`.
- `reward_output="success"`: return `1.0` if the last-frame success probability is above `success_threshold`, otherwise `0.0`.
## Usage
### Load the Reward Model Directly
```python
from lerobot.rewards.robometer import RobometerConfig, RobometerRewardModel
cfg = RobometerConfig(
pretrained_path="lerobot/Robometer-4B",
device="cuda",
reward_output="progress",
)
reward_model = RobometerRewardModel.from_pretrained(cfg.pretrained_path, config=cfg)
```
### Encode Frames and Compute a Reward
For a direct Python call, provide frames as `uint8` arrays with shape `(T, H, W, C)` and a task string:
```python
from lerobot.rewards.robometer.modeling_robometer import ROBOMETER_FEATURE_PREFIX
from lerobot.rewards.robometer.processor_robometer import RobometerEncoderProcessorStep
# frames: np.ndarray, shape (T, H, W, C), dtype uint8
# task: str
encoder = RobometerEncoderProcessorStep(
base_model_id=cfg.base_model_id,
use_multi_image=cfg.use_multi_image,
use_per_frame_progress_token=cfg.use_per_frame_progress_token,
max_frames=cfg.max_frames,
)
encoded = encoder.encode_samples([(frames, task)])
batch = {f"{ROBOMETER_FEATURE_PREFIX}{key}": value for key, value in encoded.items()}
reward = reward_model.compute_reward(batch)
```
`reward` is a tensor of shape `(batch_size,)`.
### Use the Reward Factory
You can also instantiate ROBOMETER through the reward factory:
```python
from lerobot.rewards import make_reward_model, make_reward_model_config, make_reward_pre_post_processors
cfg = make_reward_model_config(
"robometer",
pretrained_path="lerobot/Robometer-4B",
device="cuda",
image_key="observation.images.top",
)
reward_model = make_reward_model(cfg)
preprocessor, postprocessor = make_reward_pre_post_processors(cfg)
```
The preprocessor writes Qwen-VL tensors under the `observation.robometer.*` namespace, and `compute_reward()` reads those encoded tensors.
## Configuration Notes
### Backbone and Vocabulary
The published checkpoint uses a Qwen3-VL-4B backbone. ROBOMETER adds five special tokens to the tokenizer in a fixed order:
```text
<|split_token|>
<|reward_token|>
<|pref_token|>
<|sim_token|>
<|prog_token|>
```
`<|prog_token|>` is inserted after each frame and is the hidden-state position used for per-frame progress and success prediction. `<|split_token|>` and `<|pref_token|>` are used by the paper's pairwise trajectory preference objective. `<|reward_token|>` and `<|sim_token|>` are preserved for checkpoint compatibility.
The LeRobot config stores a serialized `vlm_config` with the post-resize vocabulary so the model can reload from `config.json` without downloading the base Qwen weights first. For `Qwen/Qwen3-VL-4B-Instruct`, the tokenizer length is `151669`, and the five ROBOMETER tokens produce the checkpoint vocabulary size `151674`.
### Progress Prediction
In the published checkpoint, progress is discrete. The progress head outputs logits over `progress_discrete_bins=10` uniformly spaced bin centers in `[0, 1]`. LeRobot converts these logits into a continuous value by applying a softmax and taking the expectation over bin centers, matching the upstream ROBOMETER implementation.
### Success Prediction
The success head outputs raw logits per frame. LeRobot converts them to probabilities with `sigmoid`. When `reward_output="success"`, `compute_reward()` thresholds the last-frame success probability using `success_threshold`.
## Limitations
- The current LeRobot integration is inference-only; it does not implement ROBOMETER training or preference-pair training.
- `compute_reward()` returns a scalar per sample for the LeRobot reward-model API, even though ROBOMETER predicts per-frame progress and success internally.
- ROBOMETER is video-language based; it does not use privileged robot state such as contact forces or object poses.
## References
- [ROBOMETER project](https://robometer.github.io/)
- [ROBOMETER paper](https://arxiv.org/abs/2603.02115)
- [Original ROBOMETER code](https://github.com/robometer/robometer)
- [Published ROBOMETER-4B checkpoint](https://huggingface.co/lerobot/Robometer-4B)
- [Qwen3-VL-4B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct)
## Citation
```bibtex
@inproceedings{liang2026robometer,
title = {Robometer: Scaling General-Purpose Robotic Reward Models via Trajectory Comparisons},
author={Anthony Liang and Yigit Korkmaz and Jiahui Zhang and Minyoung Hwang and Abrar Anwar and Sidhant Kaushik and Aditya Shah and Alex S. Huang and Luke Zettlemoyer and Dieter Fox and Yu Xiang and Anqi Li and Andreea Bobu and Abhishek Gupta and Stephen Tu and Erdem Biyik and Jesse Zhang},
year={2026},
booktitle={Robotics: Science and Systems 2026},
}
```
## License
This LeRobot integration follows the **Apache 2.0 License** used by LeRobot. Check the upstream ROBOMETER code and model pages for the licenses of the original implementation and released checkpoints.
docs/source/robomme.mdx
+130
View File
@@ -0,0 +1,130 @@
# RoboMME
[RoboMME](https://robomme.github.io) is a memory-augmented manipulation benchmark built on ManiSkill (SAPIEN). It evaluates a robot's ability to retain and use information across an episode — counting, object permanence, reference, and imitation.
- **16 tasks** across 4 memory-skill suites
- **1,600 training demos** (100 per task, 50 val, 50 test)
- **Dataset**: [`lerobot/robomme`](https://huggingface.co/datasets/lerobot/robomme) — LeRobot v3.0, 768K frames at 10 fps
- **Simulator**: ManiSkill / SAPIEN, Panda arm, Linux only
![RoboMME benchmark tasks overview](https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.04639/gradient.png)
## Tasks
| Suite | Tasks |
| --------------------------------- | ------------------------------------------------------------- |
| **Counting** (temporal memory) | BinFill, PickXtimes, SwingXtimes, StopCube |
| **Permanence** (spatial memory) | VideoUnmask, VideoUnmaskSwap, ButtonUnmask, ButtonUnmaskSwap |
| **Reference** (object memory) | PickHighlight, VideoRepick, VideoPlaceButton, VideoPlaceOrder |
| **Imitation** (procedural memory) | MoveCube, InsertPeg, PatternLock, RouteStick |
## Installation
> RoboMME requires **Linux** (ManiSkill/SAPIEN uses Vulkan rendering). Docker is recommended to isolate dependency conflicts.
### Native (Linux)
```bash
pip install --override <(printf 'gymnasium==0.29.1\nnumpy==1.26.4\n') \
-e '.[smolvla,av-dep]' \
'robomme @ git+https://github.com/RoboMME/robomme_benchmark.git@main'
```
> **Dependency note**: `mani-skill` (pulled by `robomme`) pins `gymnasium==0.29.1` and `numpy<2.0.0`, which conflict with lerobot's base `numpy>=2.0.0`. That's why `robomme` is not a pyproject extra — use the override install above, or the Docker approach below to avoid conflicts entirely.
### Docker (recommended)
```bash
# Build base image first (from repo root)
docker build -f docker/Dockerfile.eval-base -t lerobot-eval-base .
# Build RoboMME eval image (applies gymnasium + numpy pin overrides)
docker build -f docker/Dockerfile.benchmark.robomme -t lerobot-robomme .
```
The `docker/Dockerfile.benchmark.robomme` image overrides `gymnasium==0.29.1` and `numpy==1.26.4` after lerobot's install. Both versions are runtime-safe for lerobot's actual API usage.
## Running Evaluation
### Default (single task, single episode)
```bash
lerobot-eval \
--policy.path=<your_policy_repo> \
--env.type=robomme \
--env.task=PickXtimes \
--env.dataset_split=test \
--env.task_ids=[0] \
--eval.batch_size=1 \
--eval.n_episodes=1
```
### Multi-task evaluation
Evaluate multiple tasks in one run by comma-separating task names. Use `task_ids` to control which episodes are evaluated per task. Recommended: 50 episodes per task for the test split.
```bash
lerobot-eval \
--policy.path=<your_policy_repo> \
--env.type=robomme \
--env.task=PickXtimes,BinFill,StopCube,MoveCube,InsertPeg \
--env.dataset_split=test \
--env.task_ids=[0,1,2,3,4,5,6,7,8,9] \
--eval.batch_size=1 \
--eval.n_episodes=50
```
### Key CLI options for `env.type=robomme`
| Option | Default | Description |
| -------------------- | ------------- | -------------------------------------------------- |
| `env.task` | `PickXtimes` | Any of the 16 task names above (comma-separated) |
| `env.dataset_split` | `test` | `train`, `val`, or `test` |
| `env.action_space` | `joint_angle` | `joint_angle` (8-D) or `ee_pose` (7-D) |
| `env.episode_length` | `300` | Max steps per episode |
| `env.task_ids` | `null` | List of episode indices to evaluate (null = `[0]`) |
## Dataset
The dataset [`lerobot/robomme`](https://huggingface.co/datasets/lerobot/robomme) is in **LeRobot v3.0 format** and can be loaded directly:
```python
from lerobot.datasets.lerobot_dataset import LeRobotDataset
dataset = LeRobotDataset("lerobot/robomme")
```
### Dataset features
| Feature | Shape | Description |
| ------------------ | ------------- | ------------------------------- |
| `image` | (256, 256, 3) | Front camera RGB |
| `wrist_image` | (256, 256, 3) | Wrist camera RGB |
| `actions` | (8,) | Joint angles + gripper |
| `state` | (8,) | Joint positions + gripper state |
| `simple_subgoal` | str | High-level language annotation |
| `grounded_subgoal` | str | Grounded language annotation |
| `episode_index` | int | Episode ID |
| `frame_index` | int | Frame within episode |
### Feature key alignment (training)
The env wrapper exposes `pixels/image` and `pixels/wrist_image` as observation keys. The `features_map` in `RoboMMEEnv` maps these to `observation.images.image` and `observation.images.wrist_image` for the policy. State is exposed as `agent_pos` and maps to `observation.state`.
The dataset's `image` and `wrist_image` columns already align with the policy input keys, so no renaming is needed when fine-tuning.
## Action Spaces
| Type | Dim | Description |
| ------------- | --- | --------------------------------------------------------- |
| `joint_angle` | 8 | 7 joint angles + 1 gripper (−1 closed, +1 open, absolute) |
| `ee_pose` | 7 | xyz + roll/pitch/yaw + gripper |
Set via `--env.action_space=joint_angle` (default) or `--env.action_space=ee_pose`.
## Platform Notes
- **Linux only**: ManiSkill requires SAPIEN/Vulkan. macOS and Windows are not supported.
- **GPU recommended**: Rendering is CPU-capable but slow; CUDA + Vulkan gives full speed.
- **gymnasium / numpy conflict**: See installation note above. Docker image handles this automatically.
- **ManiSkill fork**: `robomme` depends on a specific ManiSkill fork (`YinpeiDai/ManiSkill`), pulled in automatically via the `robomme` package.
docs/source/robotwin.mdx
+223
View File
@@ -0,0 +1,223 @@
# RoboTwin 2.0
RoboTwin 2.0 is a **large-scale dual-arm manipulation benchmark** built on the SAPIEN physics engine. It provides a standardized evaluation protocol for bimanual robotic policies across 50 tasks (as of upstream `main`) with strong domain randomization (clutter, lighting, background, tabletop height, and language instructions).
- Paper: [RoboTwin 2.0: A Scalable Data Generator and Benchmark with Strong Domain Randomization for Robust Bimanual Robotic Manipulation](https://arxiv.org/abs/2506.18088)
- GitHub: [RoboTwin-Platform/RoboTwin](https://github.com/RoboTwin-Platform/RoboTwin)
- Leaderboard: [robotwin-platform.github.io/leaderboard](https://robotwin-platform.github.io/leaderboard)
- Dataset: [lerobot/robotwin_unified](https://huggingface.co/datasets/lerobot/robotwin_unified)
![RoboTwin 2.0 benchmark overview](https://www.aitntnews.com/pictures/2025/7/8/9a7f79cb-5ba9-11f0-8581-fa163e47d677.png)
## Overview
| Property | Value |
| ------------- | -------------------------------------------------------- |
| Tasks | 50 dual-arm manipulation tasks |
| Robot | Aloha-AgileX bimanual (14 DOF, 7 per arm) |
| Action space | 14-dim joint-space, continuous in `[-1, 1]` |
| Cameras | `head_camera`, `left_camera`, `right_camera` |
| Simulator | SAPIEN (not MuJoCo) |
| Eval protocol | 100 episodes/task, 50 demo_clean demonstrations |
| Eval settings | **Easy** (`demo_clean`) and **Hard** (`demo_randomized`) |
## Available tasks
RoboTwin 2.0 ships 50 dual-arm manipulation tasks in its upstream `envs/` directory. The canonical list is the `ROBOTWIN_TASKS` tuple in `src/lerobot/envs/robotwin.py`, mirrored verbatim from the upstream repo. Example tasks:
| Task | CLI name | Category |
| ------------------------ | ------------------------ | ----------------- |
| Beat block with hammer | `beat_block_hammer` | Tool use |
| Click bell / alarm clock | `click_bell` | Precision press |
| Stack blocks (2 / 3) | `stack_blocks_two/three` | Stacking |
| Stack bowls (2 / 3) | `stack_bowls_two/three` | Stacking |
| Handover block / mic | `handover_block` | Bimanual coord. |
| Lift pot | `lift_pot` | Bimanual lift |
| Shake bottle | `shake_bottle` | Continuous motion |
| Turn switch | `turn_switch` | Articulated obj |
| Stamp seal | `stamp_seal` | Precision place |
| Scan object | `scan_object` | Mobile manip. |
Pass a comma-separated list to `--env.task` to run multiple tasks in a single eval sweep.
<Tip warning={true}>
`open_laptop` is currently broken upstream (its `check_success()` uses
`self.arm_tag`, which is only set inside the scripted-expert `play_once()`
path and therefore unavailable during normal policy eval). Avoid it until the
upstream bug is fixed, or patch the task to default `self.arm_tag = "left"` in
`load_actors()`.
</Tip>
## Dataset
The RoboTwin 2.0 dataset is available in **LeRobot v3.0 format** on the Hugging Face Hub:
```
lerobot/robotwin_unified
```
It contains over 100,000 pre-collected trajectories across all 50 tasks (79.6 GB, Apache 2.0 license). No format conversion is needed — it is already in the correct LeRobot v3.0 schema with video observations and action labels.
You can load it directly with the HF Datasets library:
```python
from datasets import load_dataset
ds = load_dataset("lerobot/robotwin_unified", split="train")
```
## Installation
RoboTwin 2.0 requires **Linux** with an NVIDIA GPU (CUDA 12.1 recommended). Installation takes approximately 20 minutes.
### 1. Create a conda environment
```bash
conda create -n robotwin python=3.10 -y
conda activate robotwin
```
### 2. Install LeRobot
```bash
git clone https://github.com/huggingface/lerobot.git
cd lerobot
pip install -e "."
```
### 3. Install RoboTwin 2.0
```bash
git clone https://github.com/RoboTwin-Platform/RoboTwin.git
cd RoboTwin
bash script/_install.sh
bash script/_download_assets.sh
```
The install script handles all Python dependencies including SAPIEN, CuRobo, mplib, and pytorch3d.
<Tip warning={true}>
If the automated install fails, install manually:
```bash
pip install -r requirements.txt
pip install "git+https://github.com/facebookresearch/pytorch3d.git@stable"
cd envs && git clone https://github.com/NVlabs/curobo.git && cd curobo
pip install -e . --no-build-isolation
```
Then apply the required mplib fix: in `mplib/planner.py` line 807, remove `or collide` from the conditional.
</Tip>
### 4. Add RoboTwin to PYTHONPATH
The RoboTwin task modules must be importable by LeRobot. From within the `RoboTwin/` directory:
```bash
export PYTHONPATH="${PYTHONPATH}:$(pwd)"
```
Add this to your shell profile to make it permanent.
## Evaluation
### Standard evaluation (recommended)
Evaluate a policy on a single task with the official protocol (100 episodes):
```bash
lerobot-eval \
--policy.path="your-hf-policy-id" \
--env.type=robotwin \
--env.task=beat_block_hammer \
--eval.batch_size=1 \
--eval.n_episodes=100
```
### Single-task quick check
```bash
lerobot-eval \
--policy.path="your-hf-policy-id" \
--env.type=robotwin \
--env.task=beat_block_hammer \
--eval.batch_size=1 \
--eval.n_episodes=5
```
### Multi-task sweep
Evaluate on several tasks in one run:
```bash
lerobot-eval \
--policy.path="your-hf-policy-id" \
--env.type=robotwin \
--env.task=beat_block_hammer,click_bell,handover_block,stack_blocks_two \
--eval.batch_size=1 \
--eval.n_episodes=100
```
### Full benchmark (all 50 tasks)
```bash
lerobot-eval \
--policy.path="your-hf-policy-id" \
--env.type=robotwin \
--env.task=adjust_bottle,beat_block_hammer,blocks_ranking_rgb,blocks_ranking_size,click_alarmclock,click_bell,dump_bin_bigbin,grab_roller,handover_block,handover_mic,hanging_mug,lift_pot,move_can_pot,move_pillbottle_pad,move_playingcard_away,move_stapler_pad,open_microwave,pick_diverse_bottles,pick_dual_bottles,place_a2b_left,place_a2b_right,place_bread_basket,place_bread_skillet,place_burger_fries,place_can_basket,place_cans_plasticbox,place_container_plate,place_dual_shoes,place_empty_cup,place_fan,place_mouse_pad,place_object_basket,place_object_scale,place_object_stand,place_phone_stand,place_shoe,press_stapler,put_bottles_dustbin,put_object_cabinet,rotate_qrcode,scan_object,shake_bottle,shake_bottle_horizontally,stack_blocks_three,stack_blocks_two,stack_bowls_three,stack_bowls_two,stamp_seal,turn_switch \
--eval.batch_size=1 \
--eval.n_episodes=100
```
<Tip>
`open_laptop` is intentionally omitted above because of the upstream
`self.arm_tag` bug (see the **Available tasks** section). Re-add it once the
upstream fix lands.
</Tip>
## Camera configuration
By default, all three cameras are included:
| Camera key | Description |
| -------------- | ------------------------------ |
| `head_camera` | Torso-mounted overhead view |
| `left_camera` | Left arm wrist-mounted camera |
| `right_camera` | Right arm wrist-mounted camera |
To use a subset of cameras, override `--env.camera_names`:
```bash
lerobot-eval \
--policy.path="your-hf-policy-id" \
--env.type=robotwin \
--env.task=beat_block_hammer \
--env.camera_names="head_camera,left_camera" \
--eval.batch_size=1 \
--eval.n_episodes=10
```
## Environment config reference
Key parameters for `RoboTwinEnvConfig`:
| Parameter | Default | Description |
| -------------------- | ---------------------------------------- | ---------------------------------- |
| `task` | `"beat_block_hammer"` | Comma-separated task name(s) |
| `fps` | `25` | Simulation FPS |
| `episode_length` | `300` | Max steps per episode |
| `obs_type` | `"pixels_agent_pos"` | `"pixels"` or `"pixels_agent_pos"` |
| `camera_names` | `"head_camera,left_camera,right_camera"` | Comma-separated active cameras |
| `observation_height` | `240` | Camera pixel height |
| `observation_width` | `320` | Camera pixel width |
## Leaderboard submission
Results can be submitted to the [RoboTwin 2.0 leaderboard](https://robotwin-platform.github.io/leaderboard). The official protocol requires:
- Training on 50 `demo_clean` demonstrations per task
- Evaluating 100 episodes per task
- Reporting success rate separately for **Easy** (`demo_clean`) and **Hard** (`demo_randomized`) settings
For submission instructions, refer to the [RoboTwin 2.0 documentation](https://robotwin-platform.github.io/doc/).
docs/source/rtc.mdx
+191
View File
@@ -0,0 +1,191 @@
# Real-Time Chunking (RTC)
Real-Time Chunking (RTC) is an inference-time method that allows large, flow-matching based robotic policies, such as [Pi0](./pi0), [Pi0.5](./pi05), and [SmolVLA](./smolvla), to produce smooth, continuous, and reactive motion despite having high inference latency.
These policies generate chunks of future actions (e.g., 50 steps at a time) instead of single actions.
Because the models are large, producing each chunk takes longer than the time it takes the robot to execute it.
Naively executing chunks leads to problems such as pauses, jerky transitions, or sudden changes in strategy whenever the next chunk arrives late or disagrees with the previously executed actions.
RTC solves this by asynchronously generating the next chunk while the robot continues executing the current one, and by guiding the new chunk so it aligns smoothly with the portion of the previous chunk that has already been executed.
## How RTC Works (simplified)
RTC lets the robot think ahead while it’s still moving. When the robot is carrying out one chunk of actions, RTC starts creating the next chunk early.
But since the robot has already moved a bit by the time the new chunk is ready, RTC has to make sure the new chunk still lines up smoothly with what the robot is currently doing.
To do this, RTC treats the beginning of the new chunk like an inpainting or “fill-in-the-gaps” problem:
it gently adjusts the first part of the new chunk so it blends naturally with the robot’s ongoing motion. The result is no pauses, no sudden jumps.
In technical terms, RTC adds a guidance term to the flow-matching denoising process that forces the overlapping timesteps of the new chunk to stay close to the executed portion of the previous chunk, typically using a soft transition mask.
## Quick Start
### Installation
RTC is built into LeRobot. Just install the policy dependencies you need:
```bash
# For Pi0 or Pi0.5
pip install -e ".[pi]"
# For SmolVLA
pip install -e ".[smolvla]"
```
### Using RTC with Pi0
You can use `lerobot-rollout --strategy.type=base --inference.type=rtc` for RTC deployment on real robots.
The snippet below provides a simplified pseudo-example of how RTC operates with Pi0 in your pipeline:
```python
from lerobot.policies.pi0 import PI0Policy, PI0Config
from lerobot.configs import RTCAttentionSchedule
from lerobot.policies.rtc import RTCConfig, ActionQueue
# Load Pi0 with RTC enabled
policy_cfg = PI0Config()
# Enable RTC
policy_cfg.rtc_config = RTCConfig(
enabled=True,
execution_horizon=10, # How many steps to blend with previous chunk
max_guidance_weight=10.0, # How strongly to enforce consistency
prefix_attention_schedule=RTCAttentionSchedule.EXP, # Exponential blend
)
# Load the policy
policy = PI0Policy.from_pretrained("lerobot/pi0_base", policy_cfg=policy_cfg, device="cuda")
# Now use predict_action_chunk with RTC parameters
inference_delay = 4 # How many steps of inference latency, this values should be calculated based on the inference latency of the policy
# Initialize the action queue
action_queue = ActionQueue(policy_cfg.rtc_config)
# Start in a separate thread with the following function
def get_actions():
while True:
if should_get_actions:
prev_actions = action_queue.get_left_over()
obs = get_robot_observations(robot)
# Generate actions WITH RTC
actions = policy.predict_action_chunk(
obs,
inference_delay=inference_delay,
prev_chunk_left_over=prev_actions,
)
action_queue.merge(
actions, actions, inference_delay
)
for step in range(num_steps):
action = action_queue.get()
# Execute the first N actions
execute_actions(action)
```
## Key Parameters
`RTCConfig` has the following parameters to tune:
**`execution_horizon`**: How many timesteps from the previous chunk to maintain consistency with. Higher values mean smoother transitions but potentially less reactivity.
Typical values: 8-12 steps
```python
RTCConfig(execution_horizon=10)
```
**`max_guidance_weight`**: How strongly to enforce consistency with the previous chunk. This is a hyperparameter that can be tuned to balance the smoothness of the transitions and the reactivity of the policy. For 10 steps flow matching (SmolVLA, Pi0, Pi0.5), a value of 10.0 is a optimal value.
**`prefix_attention_schedule`**: How to weight consistency across the overlap region.
- `LINEAR`: Linear decay from inference_delay to execution_horizon
- `EXP`: Exponential decay (recommended for getting started)
- `ONES`: Full weight across entire execution_horizon
- `ZEROS`: Binary (full weight up to inference_delay, then zero)
**`inference_delay`**: How many timesteps of inference latency your system has. This is passed to `predict_action_chunk()` rather than the config, since it may vary at runtime.
## Testing RTC Offline
Before running on a real robot, test RTC with dataset samples to visualize how it works:
```bash
python examples/rtc/eval_dataset.py \
--policy.path=lerobot/pi0_libero_finetuned \
--dataset.repo_id=HuggingFaceVLA/libero \
--rtc.execution_horizon=10 \
--rtc.max_guidance_weight=10.0 \
--device=cuda
```
The script generates a visualization of the denoising process, comparing standard generation (left) with RTC (right). In the RTC plots, you can see how the first few steps (blue/purple lines) are guided to match the red ground truth trajectory (previous chunk's tail), ensuring a smooth transition between chunks.
<p align="center">
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/flow_matching.png"
alt="Denoising steps with and without RTC"
width="100%"
/>
</p>
## Testing RTC with a Real Robot
```bash
lerobot-rollout \
--strategy.type=base \
--policy.path=${HF_USERNAME}/policy_repo_id \
--inference.type=rtc \
--inference.rtc.execution_horizon=10 \
--inference.rtc.max_guidance_weight=10.0 \
--robot.type=so100_follower \
--robot.port=/dev/tty.usbmodem58FA0834591 \
--robot.cameras="{ gripper: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30}, front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
--task="Move green small object into the purple platform" \
--duration=120 \
--device=cuda
```
## How It Differs from the Async Inference in LeRobot
Both RTC and [async inference](./async) improve real-time robot control, but they solve different problems.
| Aspect | Async Inference | RTC |
| ------------- | -------------------------------------------------------------------------- | --------------------------------------------------- |
| **Problem** | Idle frames while waiting for inference | Discontinuities between action chunks |
| **Solution** | Decouple prediction from execution | Guide new chunks to continue smoothly from previous |
| **Benefit** | No waiting, continuous action | Smooth transitions, natural motion |
| **Best Used** | Async inference is best used with large models with high inference latency | Flow-matching based policies |
**Use both together** for maximum smoothness and reactivity!
## Advanced: Debug Tracking
RTC includes built-in debug tracking to help you understand what's happening during inference:
```python
# Enable debug tracking
policy_cfg.rtc_config.debug = True
policy_cfg.rtc_config.debug_maxlen = 100
# After inference, access debug data
debug_data = policy.rtc_processor.get_debug_data()
# Visualize denoising steps, corrections, etc.
from lerobot.policies.rtc.debug_visualizer import RTCDebugVisualizer
visualizer = RTCDebugVisualizer()
# ... create plots
```
See `examples/rtc/eval_dataset.py` for a complete example of offline RTC visualization.
## References
- [Smooth-As-Butter Robot Policies](https://alexander-soare.github.io/robotics/2025/08/05/smooth-as-butter-robot-policies.html) - Excellent technical explanation with real robot results
- [Physical Intelligence - Real-Time Chunking](https://www.physicalintelligence.company/research/real_time_chunking) - Original paper and research
- [Kinetix RTC Implementation](https://github.com/Physical-Intelligence/real-time-chunking-kinetix) - Reference implementation from Physical Intelligence
docs/source/sarm.mdx
+593
View File
@@ -0,0 +1,593 @@
# SARM: Stage-Aware Reward Modeling
SARM (Stage-Aware Reward Modeling) is a video-based reward modeling framework for long-horizon robot manipulation tasks. This guide covers how to train SARM reward models and optionally use them with Reward-Aligned Behavior Cloning (RA-BC).
**Paper**: [SARM: Stage-Aware Reward Modeling for Long Horizon Robot Manipulation](https://arxiv.org/abs/2509.25358)
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/lerobot-sarm.png"
alt="An overview of SARM"
width="80%"
/>
## Why Reward Models?
Standard behavior cloning treats all demonstration frames equally, but real-world robot datasets are messy. They contain hesitations, corrections, and variable-quality trajectories. Reward models solve this by learning a generalizable notion of **task progress** from demonstrations: given video frames and a task description, they predict how close the robot is to completing the task (0→1). This learned "progress signal" can be used in multiple ways, two promising applications are: (1) **weighted imitation learning** (RA-BC), where high-progress frames receive more weight during policy training, and (2) **reinforcement learning**, where the reward model provides dense rewards for online or offline policy improvement.
## Overview
SARM has following features:
1. **Stage-aware architecture**: Jointly predicts the high-level task stage and fine-grained progress within each stage
2. **Subtask annotations**: Uses natural language subtask annotations to derive consistent progress labels
3. **Temporal proportions**: Computes dataset-level priors (α̅\_k) for each subtask to normalize progress across variable-length demonstrations
SARM trains on a compact **stage+tau** target for each frame:
- **stage**: integer stage index `k ∈ {0, ..., K-1}`
- **τ (tau)**: within-stage progress `τ ∈ [0, 1]`
- **target encoding**: `y = k + τ` (this is what the dataset processor produces)
At inference time (and in downstream RA-BC), SARM converts the raw `k + τ` value into a **normalized progress** in `[0, 1]` using dataset-level **temporal proportions** `α̅_k` (stored in `meta/temporal_proportions_*.json`).
This matches **Formula (2)** from the paper:
```
progress_t = P_{k-1} + α̅_k × τ_t
```
Where:
- `τ_t = (t - s_k) / (e_k - s_k)` is within-subtask normalized time
- `P_{k-1}` is cumulative prior (sum of previous subtask proportions)
- `α̅_k` is the temporal proportion for subtask k
This ensures identical task states map to consistent progress values, even across demonstrations of different lengths.
## Inputs and Targets (What the new code expects)
SARM is trained through its processor (`src/lerobot/rewards/sarm/processor_sarm.py`), which:
- **Encodes** images and task text with CLIP (ViT-B/32) into `video_features` and `text_features`
- **Pads/truncates** robot state into `state_features` (up to `max_state_dim`)
- **Builds targets** as `sparse_targets` (and `dense_targets` in `dense_only`/`dual`) using the stage+tau encoding `y = k + τ`
- **Masks rewind frames** using a per-sample `lengths` tensor (rewind is a training-time augmentation)
At minimum, each training sample needs:
- `task` (string): task description
- `policy.image_key` images and `policy.state_key` states from the dataset
---
## Annotation Modes
You can choose from **3 annotation modes** that determine how progress labels are computed:
| Mode | Annotations Required | Heads | Use Case |
| -------------- | -------------------- | ---------------------------- | ------------------------------------------------------------ |
| `single_stage` | None | Sparse only | Simple tasks, quick experiments, no VLM needed |
| `dense_only` | Dense (VLM) | Dual (sparse auto-generated) | Detailed subtask tracking without defining high-level stages |
| `dual` | Sparse + Dense (VLM) | Dual | Full SARM paper setup with both granularities |
### Mode Details
<hfoptions id="mode_explanation">
<hfoption id="single_stage">
**No annotations required.** The entire episode is treated as a single stage called `"task"`, and progress is linear from 0 to 1 over the episode duration.
- **Sparse head**: 1 stage ("task"), linear progress
- **Dense head**: Not used
- **Best for**: Simple tasks, quick experiments, or when VLM annotation is not available
## Set Up Your Environment
1. Install LeRobot by following our [Installation Guide](./installation).
2. Install SARM dependencies by running:
```bash
pip install -e ".[sarm]"
```
Workflow:
```
1. Train SARM → 2. Visualize predictions → 3. (Optional) Train policy with RA-BC
```
</hfoption>
<hfoption id="dense_only">
**Only dense (fine-grained) annotations from a VLM.** The sparse head automatically uses a single `"task"` stage covering the full episode, while the dense head learns detailed subtask progression.
- **Sparse head**: 1 stage ("task"), linear progress (auto-generated)
- **Dense head**: Multiple fine-grained stages from VLM annotations
- **Best for**: When you want detailed subtask tracking but don't need to define high-level stages
Workflow:
```
1. Annotate (dense) → 2. Verify → 3. Train SARM → 4. Visualize → 5. (Optional) Train policy with RA-BC
```
</hfoption>
<hfoption id="dual">
**Both sparse and dense annotations from VLM.** Full dual-head mode as described in the SARM paper, with both high-level (sparse) and fine-grained (dense) stage predictions.
- **Sparse head**: High-level stages from VLM annotations
- **Dense head**: Fine-grained stages from VLM annotations
- **Best for**: Complex multi-stage tasks where both granularities are useful
Workflow:
```
1. Annotate (sparse+dense) → 2. Verify → 3. Train SARM → 4. Visualize → 5. (Optional) Train policy with RA-BC
```
</hfoption>
</hfoptions>
---
## Step 1: Subtask Annotation
<hfoptions id="annotation_mode">
<hfoption id="single_stage">
**No annotation required!** Skip this step entirely. The model will use the episode's task description and compute linear progress automatically.
</hfoption>
<hfoption id="dense_only">
Generate **dense (fine-grained) annotations only** using a VLM. The sparse stage will be auto-generated.
```bash
python src/lerobot/data_processing/sarm_annotations/subtask_annotation.py \
--repo-id your-username/your-dataset \
--dense-only \
--dense-subtasks "Bring robot arms up from starting position,Grab near side and do 1st fold,Grab side and do 2nd fold,Grab side and do 3rd fold to finish folding" \
--video-key observation.images.base \
--num-workers 4 \
--push-to-hub
```
**What gets saved:**
- `meta/temporal_proportions_sparse.json` - Auto-generated sparse proportions (`{"task": 1.0}`)
- `meta/temporal_proportions_dense.json` - Dense temporal proportions
- Per-episode columns in `episodes/*.parquet`:
- `dense_subtask_names`, `dense_subtask_start_frames`, `dense_subtask_end_frames`
- (also time-based columns: `dense_subtask_start_times`, `dense_subtask_end_times`)
</hfoption>
<hfoption id="dual">
Generate **both sparse (high-level) and dense (fine-grained) annotations** using a VLM.
```bash
python src/lerobot/data_processing/sarm_annotations/subtask_annotation.py \
--repo-id your-username/your-dataset \
--sparse-subtasks "Bring arms up from starting position,Fold the towel (3 folds in total)" \
--dense-subtasks "Bring robot arms up from starting position,Grab near side and do 1st fold,Grab side and do 2nd fold,Grab side and do 3rd fold to finish folding" \
--video-key observation.images.base \
--num-workers 4 \
--push-to-hub
```
**What gets saved:**
- `meta/temporal_proportions_sparse.json` - Sparse temporal proportions
- `meta/temporal_proportions_dense.json` - Dense temporal proportions
- Per-episode columns in `episodes/*.parquet`:
- `sparse_subtask_names`, `sparse_subtask_start_frames`, `sparse_subtask_end_frames`
- `dense_subtask_names`, `dense_subtask_start_frames`, `dense_subtask_end_frames`
- (also time-based columns: `*_subtask_start_times`, `*_subtask_end_times`)
</hfoption>
</hfoptions>
### Annotation Arguments
| Argument | Description |
| ---------------------- | ------------------------------------------------------------------------------- |
| `--repo-id` | HuggingFace dataset repository ID |
| `--sparse-subtasks` | Comma-separated list of high-level subtask names |
| `--dense-subtasks` | Comma-separated list of fine-grained subtask names |
| `--dense-only` | Generate only dense annotations (auto-creates sparse "task" stage) |
| `--video-key` | Camera/video key to use (e.g., `observation.images.top`) |
| `--num-workers` | Number of parallel GPU workers (default: 1) |
| `--episodes` | Specific episode indices to annotate (default: all) |
| `--skip-existing` | Skip episodes that already have annotations |
| `--model` | VLM model (default: `Qwen/Qwen3-VL-30B-A3B-Instruct`) |
| `--num-visualizations` | Number of episodes to visualize after annotation (default: 5, set to 0 to skip) |
> **Note**: After annotation completes, 5 episodes are automatically visualized by default. Use `--num-visualizations 0` to skip this step.
---
## Step 2: Verify Annotations
<hfoptions id="verify_mode">
<hfoption id="single_stage">
**No verification needed!** Skip this step.
</hfoption>
<hfoption id="dense_only">
Visualize annotations using the `--visualize-only` flag:
```bash
python src/lerobot/data_processing/sarm_annotations/subtask_annotation.py \
--repo-id your-username/your-dataset \
--visualize-only \
--visualize-type dense \
--num-visualizations 5 \
--video-key observation.images.base \
--output-dir ./subtask_viz
```
</hfoption>
<hfoption id="dual">
Visualize annotations using the `--visualize-only` flag:
```bash
python src/lerobot/data_processing/sarm_annotations/subtask_annotation.py \
--repo-id your-username/your-dataset \
--visualize-only \
--visualize-type both \
--num-visualizations 5 \
--video-key observation.images.base \
--output-dir ./subtask_viz
```
</hfoption>
</hfoptions>
This generates visualizations showing video frames with subtask boundaries overlaid and timeline of subtasks.
### Visualization Arguments
| Argument | Description |
| ---------------------- | -------------------------------------------------------------- |
| `--visualize-only` | Only visualize existing annotations (no generation) |
| `--num-visualizations` | Number of episodes to visualize (default: 5) |
| `--visualize-type` | Type of annotations to visualize: `sparse`, `dense`, or `both` |
**Tip**: If annotations are inaccurate, adjust your subtask descriptions to be more specific and re-run.
---
## Step 3: Train SARM
<hfoptions id="train_mode">
<hfoption id="single_stage">
Train with **no annotations** - uses linear progress from 0 to 1:
```bash
lerobot-train \
--dataset.repo_id=your-username/your-dataset \
--policy.type=sarm \
--policy.annotation_mode=single_stage \
--policy.image_key=observation.images.base \
--output_dir=outputs/train/sarm_single \
--batch_size=32 \
--steps=5000 \
--wandb.enable=true \
--wandb.project=sarm \
--policy.repo_id=your-username/your-model-name
```
</hfoption>
<hfoption id="dense_only">
Train with **dense annotations only** (sparse auto-generated):
```bash
lerobot-train \
--dataset.repo_id=your-username/your-dataset \
--policy.type=sarm \
--policy.annotation_mode=dense_only \
--policy.image_key=observation.images.base \
--output_dir=outputs/train/sarm_dense \
--batch_size=32 \
--steps=5000 \
--wandb.enable=true \
--wandb.project=sarm \
--policy.repo_id=your-username/your-model-name
```
</hfoption>
<hfoption id="dual">
Train with **both sparse and dense annotations**:
```bash
lerobot-train \
--dataset.repo_id=your-username/your-dataset \
--policy.type=sarm \
--policy.annotation_mode=dual \
--policy.image_key=observation.images.base \
--output_dir=outputs/train/sarm_dual \
--batch_size=32 \
--steps=5000 \
--wandb.enable=true \
--wandb.project=sarm \
--policy.repo_id=your-username/your-model-name
```
</hfoption>
</hfoptions>
### Multi-GPU Training
Add `accelerate launch --multi_gpu --num_processes=4` to use multiple GPUs for training.
### Training Arguments
| Argument | Description | Default |
| -------------------------- | ----------------------------------------------------------------- | ------------------------ |
| `--policy.annotation_mode` | `single_stage`, `dense_only`, or `dual` | `single_stage` |
| `--policy.image_key` | Camera key for images | `observation.images.top` |
| `--policy.state_key` | Key for joint states | `observation.state` |
| `--policy.n_obs_steps` | Observation history steps (total obs frames = `n_obs_steps + 1`) | `8` |
| `--policy.frame_gap` | Gap (in frames) between sampled observations (at 30 fps: 30 ≈ 1s) | `30` |
---
## Step 4: Visualize Predictions
Use `compute_rabc_weights.py` with `--visualize-only` to visualize model predictions (and, if available, annotation-derived targets) without writing a parquet file.
<hfoptions id="viz_mode">
<hfoption id="single_stage">
```bash
python -m lerobot.rewards.sarm.compute_rabc_weights \
--dataset-repo-id your-username/your-dataset \
--reward-model-path your-username/sarm-model \
--visualize-only \
--num-visualizations 5 \
--head-mode sparse \
--output-dir ./sarm_viz
```
</hfoption>
<hfoption id="dense_only">
```bash
python -m lerobot.rewards.sarm.compute_rabc_weights \
--dataset-repo-id your-username/your-dataset \
--reward-model-path your-username/sarm-model \
--visualize-only \
--num-visualizations 5 \
--head-mode dense \
--output-dir ./sarm_viz
```
</hfoption>
<hfoption id="dual">
```bash
python -m lerobot.rewards.sarm.compute_rabc_weights \
--dataset-repo-id your-username/your-dataset \
--reward-model-path your-username/sarm-model \
--visualize-only \
--num-visualizations 5 \
--head-mode both \
--output-dir ./sarm_viz
```
</hfoption>
</hfoptions>
The visualization shows:
- **Progress plot**: Predicted progress (and optional annotation-derived “GT” when available and `--stride 1`)
- **Stage probabilities**: Stacked area plot of predicted stage probabilities
- **Sample frames**: Key frames from the episode with progress/stage labels
### Visualization Arguments
| Argument | Description |
| ---------------------- | --------------------------------------------------------- |
| `--visualize-only` | Only visualize predictions (no RABC computation) |
| `--num-visualizations` | Number of episodes to visualize (default: 5) |
| `--head-mode` | SARM head to use: `sparse`, `dense`, or `both` |
| `--stride` | Compute every N frames, interpolate the rest (default: 1) |
---
## Step 5 (Optional): Train Policy with RA-BC
Reward-Aligned Behavior Cloning (RA-BC) uses the trained SARM model to weight training samples based on predicted progress improvement. This requires two steps:
1. **Precompute progress values** for all frames using the trained SARM model
2. **Train policy** with RA-BC weighting using the precomputed values
### How RA-BC Works
For each training sample, RA-BC computes the progress delta:
```
r_i = φ(o_{t+Δ}) - φ(o_t)
```
Where `φ` is the SARM progress prediction and `Δ` is the policy's `chunk_size`. Samples with positive progress (good demonstrations) get higher weights, while samples with negative or zero progress get down-weighted.
The weighting follows **Equations 8-9** from the paper:
- **Soft weight**: `w̃_i = clip((r_i − (μ − 2σ)) / (4σ + ε), 0, 1)`
- **Final weight**: `w_i = 𝟙{r_i > κ} + 𝟙{0 ≤ r_i ≤ κ} × w̃_i`
### Step 5a: Compute SARM Progress Values
First, run the SARM model on all frames in your dataset to compute progress values:
```bash
python -m lerobot.rewards.sarm.compute_rabc_weights \
--dataset-repo-id your-username/your-dataset \
--reward-model-path your-username/sarm-model \
--head-mode sparse \
--num-visualizations 5 \
--push-to-hub
```
This script:
- Processes all frames and computes progress values
- Saves progress values to a parquet file next to the dataset on disk (defaults to `<dataset_root>/sarm_progress.parquet`)
- Generates visualizations of the first N episodes (default: 5)
**Arguments:**
| Argument | Description | Default |
| ---------------------- | -------------------------------------------------------------- | ---------- |
| `--reward-model-path` | Path to trained SARM model | (required) |
| `--head-mode` | SARM head to use: `sparse`, `dense`, or `both` | `sparse` |
| `--device` | Device for inference | `cuda` |
| `--visualize-only` | Only visualize predictions (no RA-BC computation) | `false` |
| `--num-visualizations` | Number of episodes to visualize (default: 5, set to 0 to skip) | `5` |
**Output format** (`sarm_progress.parquet`):
| Column | Description |
| ----------------- | ---------------------------------------------- |
| `index` | Global frame index in dataset |
| `episode_index` | Episode number |
| `frame_index` | Local frame index within episode |
| `progress_sparse` | Sparse head progress value [0, 1] |
| `progress_dense` | Dense head progress value [0, 1] (if computed) |
### Step 5b: Train Policy with RA-BC
Once you have the progress file, train your policy with RA-BC weighting. The progress file is auto-detected from the dataset path (`sarm_progress.parquet`) if not explicitly provided. Currently PI0, PI0.5 and SmolVLA are supported with RA-BC:
```bash
lerobot-train \
--dataset.repo_id=your-username/your-dataset \
--policy.type=pi0 \
--sample_weighting.type=rabc \
--sample_weighting.head_mode=sparse \
--sample_weighting.kappa=0.01 \
--output_dir=outputs/train/policy_rabc \
--batch_size=32 \
--steps=40000
```
The training script automatically:
- Loads the precomputed progress values from the parquet file
- Uses the policy's `chunk_size` to compute progress deltas (Δ)
- Computes sample weights based on progress improvement
- Applies weighted loss during training
**RA-BC Arguments:**
| Argument | Description | Default |
| ---------------------------------- | ------------------------------------------------------ | ----------------------- |
| `--sample_weighting.type` | Weighting strategy type (`rabc` or `uniform`) | `rabc` |
| `--sample_weighting.progress_path` | Path to progress parquet file | `sarm_progress.parquet` |
| `--sample_weighting.head_mode` | Which SARM head's progress to use: `sparse` or `dense` | `sparse` |
| `--sample_weighting.kappa` | Threshold κ for high-quality samples | `0.01` |
| `--sample_weighting.epsilon` | Small constant for numerical stability | `1e-6` |
### Tuning RA-BC Kappa
The `kappa` parameter is the threshold that determines which samples get full weight (w=1). Understanding how to tune it is critical for RA-BC to work effectively.
**How the weighting works:**
| Condition | Weight |
| ------------------- | ----------------------- |
| `delta > kappa` | 1.0 (hard threshold) |
| `0 ≤ delta ≤ kappa` | Soft weight from Eq. 8 |
| `delta < 0` | 0.0 (negative progress) |
**Diagnosing kappa issues:**
Monitor these WandB metrics during training:
| Metric | Healthy Range | Problem Indicator |
| ----------------------------- | ------------- | ------------------------- |
| `sample_weight_mean_weight` | 0.3 - 0.8 | ≈ 1.0 means kappa too low |
| `sample_weighting/delta_mean` | > 0 | Should be positive |
| `sample_weighting/delta_std` | > 0 | Variance in data quality |
**If `sample_weight_mean_weight ≈ 1.0`:** Your kappa is too low. Most samples have `delta > kappa` and bypass the soft-weighting entirely. RA-BC becomes equivalent to vanilla BC.
**Setting kappa based on your data:**
The default `kappa=0.01` was tuned for the paper's T-shirt folding task (~90s episodes at 30fps). For your dataset, check the logged `sample_weighting/delta_mean` and `sample_weighting/delta_std`:
```
# If delta_mean ≈ 0.03 and delta_std ≈ 0.02:
# Most deltas fall in range [0.01, 0.05]
# Option 1: Set kappa = delta_mean (medium selectivity)
--sample_weighting.kappa=0.03
# Option 2: Set kappa = delta_mean + delta_std (high selectivity)
--sample_weighting.kappa=0.05
# Option 3: Set kappa = delta_mean + 2*delta_std (very selective)
--sample_weighting.kappa=0.07
```
**When RA-BC may not help:**
If your dataset is already high quality (consistent progress across all demonstrations), RA-BC won't provide much benefit since there's nothing to filter.
### Multi-GPU Training with RA-BC
```bash
accelerate launch \
--multi_gpu \
--num_processes=4 \
src/lerobot/scripts/lerobot_train.py \
--dataset.repo_id=your-username/your-dataset \
--policy.type=pi0 \
--sample_weighting.type=rabc \
--sample_weighting.kappa=0.01 \
--output_dir=outputs/train/policy_rabc \
--batch_size=32 \
--steps=40000
```
---
## Tips & Best Practices
### Choosing a Mode
- **Start with `single_stage`** for quick experiments - no annotation overhead
- Use **`dense_only`** when you want detailed progress tracking but tasks don't have clear high-level stages
- Use **`dual`** for complex tasks where both coarse and fine-grained progress is meaningful
### Annotation Quality
1. **Be specific with subtask names**: Instead of "fold", use "grab near side and fold toward center"
2. **Verify with visualization**: Always check a few episodes before training
3. **Consistent naming**: Use the same subtask names across all episodes
### RA-BC
1. **Train SARM first**: RA-BC quality depends entirely on SARM quality
2. **Monitor `sample_weight_mean_weight`**: If it's ≈ 1.0, increase kappa (see [Tuning RA-BC Kappa](#tuning-ra-bc-kappa))
---
## Citation
```bibtex
@article{chen2025sarm,
title={SARM: Stage-Aware Reward Modeling for Long Horizon Robot Manipulation},
author={Chen, Qianzhong and Yu, Justin and Schwager, Mac and Abbeel, Pieter and Shentu, Yide and Wu, Philipp},
journal={arXiv preprint arXiv:2509.25358},
year={2025}
}
```
docs/source/smolvla.mdx
+119
View File
@@ -0,0 +1,119 @@
# SmolVLA
SmolVLA is Hugging Face’s lightweight foundation model for robotics. Designed for easy fine-tuning on LeRobot datasets, it helps accelerate your development!
<p align="center">
<img
src="https://cdn-uploads.huggingface.co/production/uploads/640e21ef3c82bd463ee5a76d/aooU0a3DMtYmy_1IWMaIM.png"
alt="SmolVLA architecture."
width="500"
/>
<br />
<em>
Figure 1. SmolVLA takes as input (i) multiple cameras views, (ii) the
robot’s current sensorimotor state, and (iii) a natural language
instruction, encoded into contextual features used to condition the action
expert when generating an action chunk.
</em>
</p>
## Set Up Your Environment
1. Install LeRobot by following our [Installation Guide](./installation).
2. Install SmolVLA dependencies by running:
```bash
pip install -e ".[smolvla]"
```
## Collect a dataset
SmolVLA is a base model, so fine-tuning on your own data is required for optimal performance in your setup.
We recommend recording ~50 episodes of your task as a starting point. Follow our guide to get started: [Recording a Dataset](./il_robots)
<Tip>
In your dataset, make sure to have enough demonstrations per each variation (e.g. the cube position on the table if it is cube pick-place task) you are introducing.
We recommend checking out the dataset linked below for reference that was used in the [SmolVLA paper](https://huggingface.co/papers/2506.01844):
🔗 [SVLA SO100 PickPlace](https://huggingface.co/spaces/lerobot/visualize_dataset?path=%2Flerobot%2Fsvla_so100_pickplace%2Fepisode_0)
In this dataset, we recorded 50 episodes across 5 distinct cube positions. For each position, we collected 10 episodes of pick-and-place interactions. This structure, repeating each variation several times, helped the model generalize better. We tried similar dataset with 25 episodes, and it was not enough leading to a bad performance. So, the data quality and quantity is definitely a key.
After you have your dataset available on the Hub, you are good to go to use our finetuning script to adapt SmolVLA to your application.
</Tip>
## Finetune SmolVLA on your data
Use [`smolvla_base`](https://hf.co/lerobot/smolvla_base), our pretrained 450M model, and fine-tune it on your data.
Training the model for 20k steps will roughly take ~4 hrs on a single A100 GPU. You should tune the number of steps based on performance and your use-case.
If you don't have a gpu device, you can train using our notebook on [![Google Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/lerobot/training-smolvla.ipynb)
Pass your dataset to the training script using `--dataset.repo_id`. If you want to test your installation, run the following command where we use one of the datasets we collected for the [SmolVLA Paper](https://huggingface.co/papers/2506.01844).
```bash
cd lerobot && lerobot-train \
--policy.path=lerobot/smolvla_base \
--dataset.repo_id=${HF_USER}/mydataset \
--batch_size=64 \
--steps=20000 \
--output_dir=outputs/train/my_smolvla \
--job_name=my_smolvla_training \
--policy.device=cuda \
--wandb.enable=true
```
<Tip>
You can start with a small batch size and increase it incrementally, if the
GPU allows it, as long as loading times remain short.
</Tip>
Fine-tuning is an art. For a complete overview of the options for finetuning, run
```bash
lerobot-train --help
```
<p align="center">
<img
src="https://cdn-uploads.huggingface.co/production/uploads/640e21ef3c82bd463ee5a76d/S-3vvVCulChREwHDkquoc.gif"
alt="Comparison of SmolVLA across task variations."
width="500"
/>
<br />
<em>
Figure 2: Comparison of SmolVLA across task variations. From left to right:
(1) pick-place cube counting, (2) pick-place cube counting, (3) pick-place
cube counting under perturbations, and (4) generalization on pick-and-place
of the lego block with real-world SO101.
</em>
</p>
## Evaluate the finetuned model and run it in real-time
Similarly for when recording an episode, it is recommended that you are logged in to the HuggingFace Hub. You can follow the corresponding steps: [Record a dataset](./il_robots).
Once you are logged in, you can run inference in your setup by doing:
```bash
lerobot-rollout \
--strategy.type=base \
--robot.type=so101_follower \
--robot.port=/dev/ttyACM0 \ # <- Use your port
--robot.id=my_blue_follower_arm \ # <- Use your robot id
--robot.cameras="{ front: {type: opencv, index_or_path: 8, width: 640, height: 480, fps: 30}}" \ # <- Use your cameras
--task="Grasp a lego block and put it in the bin." \ # <- Use the same task description you used in your dataset recording
# <- RTC optional, use when running on low power hardware \
# --inference.type=rtc \
# --inference.rtc.execution_horizon=10 \
# --inference.rtc.max_guidance_weight=10.0 \
# <- Teleop optional if you want to teleoperate in between episodes \
# --teleop.type=so100_leader \
# --teleop.port=/dev/ttyACM0 \
# --teleop.id=my_red_leader_arm \
# --display_data=true #optional use if you want to see the camera stream \
--policy.path=HF_USER/FINETUNE_MODEL_NAME # <- Use your fine-tuned model
```
Depending on your evaluation setup, you can configure the duration and the number of episodes to record for your evaluation suite.
docs/source/so100.mdx
+640
View File
@@ -0,0 +1,640 @@
# SO-100
In the steps below, we explain how to assemble the SO-100 robot.
## Source the parts
Follow this [README](https://github.com/TheRobotStudio/SO-ARM100/blob/main/SO100.md). It contains the bill of materials, with a link to source the parts, as well as the instructions to 3D print the parts. And advise if it's your first time printing or if you don't own a 3D printer.
## Install LeRobot 🤗
To install LeRobot, follow our [Installation Guide](./installation)
In addition to these instructions, you need to install the Feetech SDK:
```bash
pip install -e ".[feetech]"
```
## Configure the motors
**Note:**
Unlike the SO-101, the motor connectors are not easily accessible once the arm is assembled, so the configuration step must be done beforehand.
### 1. Find the USB ports associated with each arm
To find the port for each bus servo adapter, run this script:
```bash
lerobot-find-port
```
<hfoptions id="example">
<hfoption id="Mac">
Example output:
```
Finding all available ports for the MotorBus.
['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
Remove the USB cable from your MotorsBus and press Enter when done.
[...Disconnect corresponding leader or follower arm and press Enter...]
The port of this MotorsBus is /dev/tty.usbmodem575E0032081
Reconnect the USB cable.
```
Where the found port is: `/dev/tty.usbmodem575E0032081` corresponding to your leader or follower arm.
</hfoption>
<hfoption id="Linux">
On Linux, you might need to give access to the USB ports by running:
```bash
sudo chmod 666 /dev/ttyACM0
sudo chmod 666 /dev/ttyACM1
```
Example output:
```
Finding all available ports for the MotorBus.
['/dev/ttyACM0', '/dev/ttyACM1']
Remove the usb cable from your MotorsBus and press Enter when done.
[...Disconnect corresponding leader or follower arm and press Enter...]
The port of this MotorsBus is /dev/ttyACM1
Reconnect the USB cable.
```
Where the found port is: `/dev/ttyACM1` corresponding to your leader or follower arm.
</hfoption>
</hfoptions>
### 2. Set the motors ids and baudrates
Each motor is identified by a unique id on the bus. When brand new, motors usually come with a default id of `1`. For the communication to work properly between the motors and the controller, we first need to set a unique, different id to each motor. Additionally, the speed at which data is transmitted on the bus is determined by the baudrate. In order to talk to each other, the controller and all the motors need to be configured with the same baudrate.
To that end, we first need to connect to each motor individually with the controller in order to set these. Since we will write these parameters in the non-volatile section of the motors' internal memory (EEPROM), we'll only need to do this once.
If you are repurposing motors from another robot, you will probably also need to perform this step as the ids and baudrate likely won't match.
#### Follower
Connect the usb cable from your computer and the power supply to the follower arm's controller board. Then, run the following command or run the API example with the port you got from the previous step. You'll also need to give your leader arm a name with the `id` parameter.
For a visual reference on how to set the motor ids please refer to [this video](https://huggingface.co/docs/lerobot/en/so101#setup-motors-video) where we follow the process for the SO101 arm.
<hfoptions id="setup_motors">
<hfoption id="Command">
```bash
lerobot-setup-motors \
--robot.type=so100_follower \
--robot.port=/dev/tty.usbmodem585A0076841 # <- paste here the port found at previous step
```
</hfoption>
<hfoption id="API example">
<!-- prettier-ignore-start -->
```python
from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
config = SO100FollowerConfig(
port="/dev/tty.usbmodem585A0076841",
id="my_awesome_follower_arm",
)
follower = SO100Follower(config)
follower.setup_motors()
```
<!-- prettier-ignore-end -->
</hfoption>
</hfoptions>
You should see the following instruction
```
Connect the controller board to the 'gripper' motor only and press enter.
```
As instructed, plug the gripper's motor. Make sure it's the only motor connected to the board, and that the motor itself is not yet daisy-chained to any other motor. As you press `[Enter]`, the script will automatically set the id and baudrate for that motor.
<details>
<summary>Troubleshooting</summary>
If you get an error at that point, check your cables and make sure they are plugged in properly:
<ul>
<li>Power supply</li>
<li>USB cable between your computer and the controller board</li>
<li>The 3-pin cable from the controller board to the motor</li>
</ul>
If you are using a Waveshare controller board, make sure that the two jumpers are set on the `B` channel (USB).
</details>
You should then see the following message:
```
'gripper' motor id set to 6
```
Followed by the next instruction:
```
Connect the controller board to the 'wrist_roll' motor only and press enter.
```
You can disconnect the 3-pin cable from the controller board, but you can leave it connected to the gripper motor on the other end, as it will already be in the right place. Now, plug in another 3-pin cable to the wrist roll motor and connect it to the controller board. As with the previous motor, make sure it is the only motor connected to the board and that the motor itself isn't connected to any other one.
Repeat the operation for each motor as instructed.
> [!TIP]
> Check your cabling at each step before pressing Enter. For instance, the power supply cable might disconnect as you manipulate the board.
When you are done, the script will simply finish, at which point the motors are ready to be used. You can now plug the 3-pin cable from each motor to the next one, and the cable from the first motor (the 'shoulder pan' with id=1) to the controller board, which can now be attached to the base of the arm.
#### Leader
Do the same steps for the leader arm.
<hfoptions id="setup_motors">
<hfoption id="Command">
```bash
lerobot-setup-motors \
--teleop.type=so100_leader \
--teleop.port=/dev/tty.usbmodem575E0031751 # <- paste here the port found at previous step
```
</hfoption>
<hfoption id="API example">
<!-- prettier-ignore-start -->
```python
from lerobot.teleoperators.so_leader import SO100Leader, SO100LeaderConfig
config = SO100LeaderConfig(
port="/dev/tty.usbmodem585A0076841",
id="my_awesome_leader_arm",
)
leader = SO100Leader(config)
leader.setup_motors()
```
<!-- prettier-ignore-end -->
</hfoption>
</hfoptions>
## Step-by-Step Assembly Instructions
## Remove the gears of the 6 leader motors
<details>
<summary><strong>Video removing gears</strong></summary>
<div class="video-container">
<video controls width="600">
<source
src="https://github.com/user-attachments/assets/0c95b88c-5b85-413d-ba19-aee2f864f2a7"
type="video/mp4"
/>
</video>
</div>
</details>
Follow the video for removing gears. You need to remove the gear for the motors of the leader arm. As a result, you will only use the position encoding of the motor and reduce friction to more easily operate the leader arm.
### Clean Parts
Remove all support material from the 3D-printed parts. The easiest way to do this is using a small screwdriver to get underneath the support material.
### Additional Guidance
<details>
<summary><strong>Video assembling arms</strong></summary>
<div class="video-container">
<video controls width="600">
<source
src="https://github.com/user-attachments/assets/488a39de-0189-4461-9de3-05b015f90cca"
type="video/mp4"
/>
</video>
</div>
</details>
**Note:**
This video provides visual guidance for assembling the arms, but it doesn't specify when or how to do the wiring. Inserting the cables beforehand is much easier than doing it afterward. The first arm may take a bit more than 1 hour to assemble, but once you get used to it, you can assemble the second arm in under 1 hour.
---
### First Motor
**Step 2: Insert Wires**
- Insert two wires into the first motor.
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/so100_assembly_1.webp"
style="height:300px;"
/>
**Step 3: Install in Base**
- Place the first motor into the base.
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/so100_assembly_2.webp"
style="height:300px;"
/>
**Step 4: Secure Motor**
- Fasten the motor with 4 screws. Two from the bottom and two from top.
**Step 5: Attach Motor Holder**
- Slide over the first motor holder and fasten it using two screws (one on each side).
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/so100_assembly_4.webp"
style="height:300px;"
/>
**Step 6: Attach Motor Horns**
- Install both motor horns, securing the top horn with a screw. Try not to move the motor position when attaching the motor horn, especially for the leader arms, where we removed the gears.
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/so100_assembly_5.webp"
style="height:300px;"
/>
<details>
<summary>
<strong>Video adding motor horn</strong>
</summary>
<video src="https://github.com/user-attachments/assets/ef3391a4-ad05-4100-b2bd-1699bf86c969"></video>
</details>
**Step 7: Attach Shoulder Part**
- Route one wire to the back of the robot and the other to the left or towards you (see photo).
- Attach the shoulder part.
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/so100_assembly_6.webp"
style="height:300px;"
/>
**Step 8: Secure Shoulder**
- Tighten the shoulder part with 4 screws on top and 4 on the bottom
_(access bottom holes by turning the shoulder)._
---
### Second Motor Assembly
**Step 9: Install Motor 2**
- Slide the second motor in from the top and link the wire from motor 1 to motor 2.
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/so100_assembly_8.webp"
style="height:300px;"
/>
**Step 10: Attach Shoulder Holder**
- Add the shoulder motor holder.
- Ensure the wire from motor 1 to motor 2 goes behind the holder while the other wire is routed upward (see photo).
- This part can be tight to assemble, you can use a workbench like the image or a similar setup to push the part around the motor.
<div style="display: flex;">
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/so100_assembly_9.webp"
style="height:250px;"
/>
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/so100_assembly_10.webp"
style="height:250px;"
/>
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/so100_assembly_12.webp"
style="height:250px;"
/>
</div>
**Step 11: Secure Motor 2**
- Fasten the second motor with 4 screws.
**Step 12: Attach Motor Horn**
- Attach both motor horns to motor 2, again use the horn screw.
**Step 13: Attach Base**
- Install the base attachment using 2 screws.
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/so100_assembly_11.webp" style="height:300px;">
**Step 14: Attach Upper Arm**
- Attach the upper arm with 4 screws on each side.
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/so100_assembly_13.webp" style="height:300px;">
---
### Third Motor Assembly
**Step 15: Install Motor 3**
- Route the motor cable from motor 2 through the cable holder to motor 3, then secure motor 3 with 4 screws.
**Step 16: Attach Motor Horn**
- Attach both motor horns to motor 3 and secure one again with a horn screw.
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/so100_assembly_14.webp"
style="height:300px;"
/>
**Step 17: Attach Forearm**
- Connect the forearm to motor 3 using 4 screws on each side.
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/so100_assembly_15.webp"
style="height:300px;"
/>
---
### Fourth Motor Assembly
**Step 18: Install Motor 4**
- Slide in motor 4, attach the cable from motor 3, and secure the cable in its holder with a screw.
<div style="display: flex;">
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/so100_assembly_16.webp"
style="height:300px;"
/>
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/so100_assembly_19.webp"
style="height:300px;"
/>
</div>
**Step 19: Attach Motor Holder 4**
- Install the fourth motor holder (a tight fit). Ensure one wire is routed upward and the wire from motor 3 is routed downward (see photo).
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/so100_assembly_17.webp"
style="height:300px;"
/>
**Step 20: Secure Motor 4 & Attach Horn**
- Fasten motor 4 with 4 screws and attach its motor horns, use for one a horn screw.
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/so100_assembly_18.webp"
style="height:300px;"
/>
---
### Wrist Assembly
**Step 21: Install Motor 5**
- Insert motor 5 into the wrist holder and secure it with 2 front screws.
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/so100_assembly_20.webp"
style="height:300px;"
/>
**Step 22: Attach Wrist**
- Connect the wire from motor 4 to motor 5. And already insert the other wire for the gripper.
- Secure the wrist to motor 4 using 4 screws on both sides.
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/so100_assembly_22.webp"
style="height:300px;"
/>
**Step 23: Attach Wrist Horn**
- Install only one motor horn on the wrist motor and secure it with a horn screw.
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/so100_assembly_23.webp"
style="height:300px;"
/>
---
### Follower Configuration
**Step 24: Attach Gripper**
- Attach the gripper to motor 5.
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/so100_assembly_24.webp"
style="height:300px;"
/>
**Step 25: Install Gripper Motor**
- Insert the gripper motor, connect the motor wire from motor 5 to motor 6, and secure it with 3 screws on each side.
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/so100_assembly_25.webp"
style="height:300px;"
/>
**Step 26: Attach Gripper Horn & Claw**
- Attach the motor horns and again use a horn screw.
- Install the gripper claw and secure it with 4 screws on both sides.
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/so100_assembly_26.webp"
style="height:300px;"
/>
**Step 27: Mount Controller**
- Attach the motor controller to the back of the robot.
<div style="display: flex;">
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/so100_assembly_27.webp"
style="height:300px;"
/>
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/so100_assembly_28.webp"
style="height:300px;"
/>
</div>
_Assembly complete – proceed to Leader arm assembly._
---
### Leader Configuration
For the leader configuration, perform **Steps 1–23**. Make sure that you removed the motor gears from the motors.
**Step 24: Attach Leader Holder**
- Mount the leader holder onto the wrist and secure it with a screw.
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/so100_assembly_29.webp"
style="height:300px;"
/>
**Step 25: Attach Handle**
- Attach the handle to motor 5 using 4 screws.
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/so100_assembly_30.webp"
style="height:300px;"
/>
**Step 26: Install Gripper Motor**
- Insert the gripper motor, secure it with 3 screws on each side, attach a motor horn using a horn screw, and connect the motor wire.
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/so100_assembly_31.webp"
style="height:300px;"
/>
**Step 27: Attach Trigger**
- Attach the follower trigger with 4 screws.
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/so100_assembly_32.webp"
style="height:300px;"
/>
**Step 28: Mount Controller**
- Attach the motor controller to the back of the robot.
<div style="display: flex;">
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/so100_assembly_27.webp"
style="height:300px;"
/>
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/so100_assembly_28.webp"
style="height:300px;"
/>
</div>
## Calibrate
Next, you'll need to calibrate your robot to ensure that the leader and follower arms have the same position values when they are in the same physical position.
The calibration process is very important because it allows a neural network trained on one robot to work on another.
#### Follower
Run the following command or API example to calibrate the follower arm:
<hfoptions id="calibrate_follower">
<hfoption id="Command">
```bash
lerobot-calibrate \
--robot.type=so100_follower \
--robot.port=/dev/tty.usbmodem58760431551 \ # <- The port of your robot
--robot.id=my_awesome_follower_arm # <- Give the robot a unique name
```
</hfoption>
<hfoption id="API example">
<!-- prettier-ignore-start -->
```python
from lerobot.robots.so_follower import SO100FollowerConfig, SO100Follower
config = SO100FollowerConfig(
port="/dev/tty.usbmodem585A0076891",
id="my_awesome_follower_arm",
)
follower = SO100Follower(config)
follower.connect(calibrate=False)
follower.calibrate()
follower.disconnect()
```
<!-- prettier-ignore-end -->
</hfoption>
</hfoptions>
We unified the calibration method for most robots. Thus, the calibration steps for this SO100 arm are the same as the steps for the Koch and SO101. First, we have to move the robot to the position where each joint is in the middle of its range, then we press `Enter`. Secondly, we move all joints through their full range of motion. A video of this same process for the SO101 as reference can be found [here](https://huggingface.co/docs/lerobot/en/so101#calibration-video)
#### Leader
Do the same steps to calibrate the leader arm, run the following command or API example:
<hfoptions id="calibrate_leader">
<hfoption id="Command">
```bash
lerobot-calibrate \
--teleop.type=so100_leader \
--teleop.port=/dev/tty.usbmodem58760431551 \ # <- The port of your robot
--teleop.id=my_awesome_leader_arm # <- Give the robot a unique name
```
</hfoption>
<hfoption id="API example">
<!-- prettier-ignore-start -->
```python
from lerobot.teleoperators.so_leader import SO100LeaderConfig, SO100Leader
config = SO100LeaderConfig(
port="/dev/tty.usbmodem58760431551",
id="my_awesome_leader_arm",
)
leader = SO100Leader(config)
leader.connect(calibrate=False)
leader.calibrate()
leader.disconnect()
```
<!-- prettier-ignore-end -->
</hfoption>
</hfoptions>
Congrats 🎉, your robot is all set to learn a task on its own. Start training it by following this tutorial: [Getting started with real-world robots](./il_robots)
> [!TIP]
> If you have any questions or need help, please reach out on [Discord](https://discord.com/invite/s3KuuzsPFb).
docs/source/so101.mdx
+450
View File
@@ -0,0 +1,450 @@
# SO-101
<div style="display: flex; align-items: center; gap: 10px;">
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/SO101_Follower.webp"
alt="SO-101"
width="60%"
/>
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/SO101_Leader.webp"
alt="SO-101"
width="60%"
/>
</div>
In the steps below, we explain how to assemble our flagship robot, the SO-101.
## Source the parts
Follow this [README](https://github.com/TheRobotStudio/SO-ARM100). It contains the bill of materials, with a link to source the parts, as well as the instructions to 3D print the parts.
And advise if it's your first time printing or if you don't own a 3D printer.
## Install LeRobot 🤗
To install LeRobot, follow our [Installation Guide](./installation)
In addition to these instructions, you need to install the Feetech SDK:
```bash
pip install -e ".[feetech]"
```
## Step-by-Step Assembly Instructions
The follower arm uses 6x STS3215 motors with 1/345 gearing. The leader, however, uses three differently geared motors to make sure it can both sustain its own weight and it can be moved without requiring much force. Which motor is needed for which joint is shown in the table below.
| Leader-Arm Axis | Motor | Gear Ratio |
| ------------------- | :---: | :--------: |
| Base / Shoulder Pan | 1 | 1 / 191 |
| Shoulder Lift | 2 | 1 / 345 |
| Elbow Flex | 3 | 1 / 191 |
| Wrist Flex | 4 | 1 / 147 |
| Wrist Roll | 5 | 1 / 147 |
| Gripper | 6 | 1 / 147 |
## Configure the motors
### 1. Find the USB ports associated with each arm
To find the port for each bus servo adapter, connect MotorBus to your computer via USB and power. Run the following script and disconnect the MotorBus when prompted:
```bash
lerobot-find-port
```
<hfoptions id="example">
<hfoption id="Mac">
Example output:
```
Finding all available ports for the MotorBus.
['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
Remove the USB cable from your MotorsBus and press Enter when done.
[...Disconnect corresponding leader or follower arm and press Enter...]
The port of this MotorsBus is /dev/tty.usbmodem575E0032081
Reconnect the USB cable.
```
Where the found port is: `/dev/tty.usbmodem575E0032081` corresponding to your leader or follower arm.
</hfoption>
<hfoption id="Linux">
On Linux, you might need to give access to the USB ports by running:
```bash
sudo chmod 666 /dev/ttyACM0
sudo chmod 666 /dev/ttyACM1
```
Example output:
```
Finding all available ports for the MotorBus.
['/dev/ttyACM0', '/dev/ttyACM1']
Remove the usb cable from your MotorsBus and press Enter when done.
[...Disconnect corresponding leader or follower arm and press Enter...]
The port of this MotorsBus is /dev/ttyACM1
Reconnect the USB cable.
```
Where the found port is: `/dev/ttyACM1` corresponding to your leader or follower arm.
</hfoption>
</hfoptions>
### 2. Set the motors ids and baudrates
Each motor is identified by a unique id on the bus. When brand new, motors usually come with a default id of `1`. For the communication to work properly between the motors and the controller, we first need to set a unique, different id to each motor. Additionally, the speed at which data is transmitted on the bus is determined by the baudrate. In order to talk to each other, the controller and all the motors need to be configured with the same baudrate.
To that end, we first need to connect to each motor individually with the controller in order to set these. Since we will write these parameters in the non-volatile section of the motors' internal memory (EEPROM), we'll only need to do this once.
If you are repurposing motors from another robot, you will probably also need to perform this step as the ids and baudrate likely won't match.
The video below shows the sequence of steps for setting the motor ids.
##### Setup motors video
<div class="video-container">
<video controls width="600">
<source
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/setup_motors_so101_2.mp4"
type="video/mp4"
/>
</video>
</div>
#### Follower
Connect the usb cable from your computer and the power supply to the follower arm's controller board. Then, run the following command or run the API example with the port you got from the previous step. You'll also need to give your follower arm a name with the `id` parameter.
<hfoptions id="setup_motors">
<hfoption id="Command">
```bash
lerobot-setup-motors \
--robot.type=so101_follower \
--robot.port=/dev/tty.usbmodem585A0076841 # <- paste here the port found at previous step
```
</hfoption>
<hfoption id="API example">
<!-- prettier-ignore-start -->
```python
from lerobot.robots.so_follower import SO101Follower, SO101FollowerConfig
config = SO101FollowerConfig(
port="/dev/tty.usbmodem585A0076841",
id="my_awesome_follower_arm",
)
follower = SO101Follower(config)
follower.setup_motors()
```
<!-- prettier-ignore-end -->
</hfoption>
</hfoptions>
You should see the following instruction
```bash
Connect the controller board to the 'gripper' motor only and press enter.
```
As instructed, plug the gripper's motor. Make sure it's the only motor connected to the board, and that the motor itself is not yet daisy-chained to any other motor. As you press `[Enter]`, the script will automatically set the id and baudrate for that motor.
<details>
<summary>Troubleshooting</summary>
If you get an error at that point, check your cables and make sure they are plugged in properly:
<ul>
<li>Power supply</li>
<li>USB cable between your computer and the controller board</li>
<li>The 3-pin cable from the controller board to the motor</li>
</ul>
If you are using a Waveshare controller board, make sure that the two jumpers are set on the `B` channel (USB).
</details>
You should then see the following message:
```bash
'gripper' motor id set to 6
```
Followed by the next instruction:
```bash
Connect the controller board to the 'wrist_roll' motor only and press enter.
```
You can disconnect the 3-pin cable from the controller board, but you can leave it connected to the gripper motor on the other end, as it will already be in the right place. Now, plug in another 3-pin cable to the wrist roll motor and connect it to the controller board. As with the previous motor, make sure it is the only motor connected to the board and that the motor itself isn't connected to any other one.
Repeat the operation for each motor as instructed.
> [!TIP]
> Check your cabling at each step before pressing Enter. For instance, the power supply cable might disconnect as you manipulate the board.
When you are done, the script will simply finish, at which point the motors are ready to be used. You can now plug the 3-pin cable from each motor to the next one, and the cable from the first motor (the 'shoulder pan' with id=1) to the controller board, which can now be attached to the base of the arm.
#### Leader
Do the same steps for the leader arm.
<hfoptions id="setup_motors">
<hfoption id="Command">
```bash
lerobot-setup-motors \
--teleop.type=so101_leader \
--teleop.port=/dev/tty.usbmodem575E0031751 # <- paste here the port found at previous step
```
</hfoption>
<hfoption id="API example">
<!-- prettier-ignore-start -->
```python
from lerobot.teleoperators.so_leader import SO101Leader, SO101LeaderConfig
config = SO101LeaderConfig(
port="/dev/tty.usbmodem585A0076841",
id="my_awesome_leader_arm",
)
leader = SO101Leader(config)
leader.setup_motors()
```
<!-- prettier-ignore-end -->
</hfoption>
</hfoptions>
### Clean Parts
Remove all support material from the 3D-printed parts. The easiest way to do this is using a small screwdriver to get underneath the support material.
It is advisable to install one 3-pin cable in the motor after placing them before continuing assembly.
### Joint 1
- Install both motor horns. Secure the top horn with a M3x6mm screw. No screws are required for the bottom horn.
- Place the first motor into the base.
- Fasten the motor with 4 M2x6mm screws (smallest screws). Two from the top and two from the bottom.
- Slide over the first motor holder and fasten it using two M2x6mm screws (one on each side).
- Attach the shoulder part.
- Tighten the shoulder part with 4 M3x6mm screws on top and 4 M3x6mm screws on the bottom
- Add the shoulder motor holder.
<div class="video-container">
<video controls width="600">
<source
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Joint1_v2.mp4"
type="video/mp4"
/>
</video>
</div>
### Joint 2
- Install both motor horns. Secure the top horn with a M3x6mm screw. No screws are required for the bottom horn.
- Slide the second motor in from the top.
- Fasten the second motor with 4 M2x6mm screws.
- Attach the upper arm with 4 M3x6mm screws on each side.
<div class="video-container">
<video controls width="600">
<source
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Joint2_v2.mp4"
type="video/mp4"
/>
</video>
</div>
### Joint 3
- Install both motor horns. Secure the top horn with a M3x6mm screw. No screws are required for the bottom horn.
- Insert motor 3 and fasten using 4 M2x6mm screws.
- Connect the forearm to motor 3 using 4 M3x6mm screws on each side.
<div class="video-container">
<video controls width="600">
<source
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Joint3_v2.mp4"
type="video/mp4"
/>
</video>
</div>
### Joint 4
- Install both motor horns. Secure the top horn with a M3x6mm screw. No screws are required for the bottom horn.
- Slide over motor holder 4.
- Slide in motor 4.
- Fasten motor 4 with 4 M2x6mm screws.
<div class="video-container">
<video controls width="600">
<source
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Joint4_v2.mp4"
type="video/mp4"
/>
</video>
</div>
### Joint 5
- Insert motor 5 into the wrist holder and secure it with 2 M2x6mm front screws.
- Install only one motor horn on the wrist motor and secure it with a M3x6mm horn screw.
- Secure the wrist to motor 4 using 4 M3x6mm screws on both sides.
<div class="video-container">
<video controls width="600">
<source
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Joint5_v2.mp4"
type="video/mp4"
/>
</video>
</div>
### Gripper / Handle
<hfoptions id="assembly">
<hfoption id="Follower">
- Attach the gripper to motor 5, attach it to the motor horn on the wrist using 4 M3x6mm screws.
- Insert the gripper motor and secure it with 2 M2x6mm screws on each side.
- Install both motor horns on the gripper motor. Secure the top horn with a M3x6mm screw; no screws are required for the bottom horn.
- Install the gripper claw and secure it with 4 M3x6mm screws on both sides.
<div class="video-container">
<video controls width="600">
<source
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Gripper_v2.mp4"
type="video/mp4"
/>
</video>
</div>
</hfoption>
<hfoption id="Leader">
- Mount the leader holder onto the wrist and secure it with 4 M3x6mm screws.
- Attach the handle to motor 5 using 1 M2x6mm screw.
- Insert the gripper motor, secure it with 2 M2x6mm screws on each side, attach a motor horn using a M3x6mm horn screw.
- Attach the follower trigger with 4 M3x6mm screws.
<div class="video-container">
<video controls width="600">
<source
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Leader_v2.mp4"
type="video/mp4"
/>
</video>
</div>
</hfoption>
</hfoptions>
## Calibrate
Next, you'll need to calibrate your robot to ensure that the leader and follower arms have the same position values when they are in the same physical position.
The calibration process is very important because it allows a neural network trained on one robot to work on another.
#### Follower
Run the following command or API example to calibrate the follower arm:
<hfoptions id="calibrate_follower">
<hfoption id="Command">
```bash
lerobot-calibrate \
--robot.type=so101_follower \
--robot.port=/dev/tty.usbmodem58760431551 \ # <- The port of your robot
--robot.id=my_awesome_follower_arm # <- Give the robot a unique name
```
</hfoption>
<hfoption id="API example">
<!-- prettier-ignore-start -->
```python
from lerobot.robots.so_follower import SO101FollowerConfig, SO101Follower
config = SO101FollowerConfig(
port="/dev/tty.usbmodem585A0076891",
id="my_awesome_follower_arm",
)
follower = SO101Follower(config)
follower.connect(calibrate=False)
follower.calibrate()
follower.disconnect()
```
<!-- prettier-ignore-end -->
</hfoption>
</hfoptions>
The video below shows how to perform the calibration. First you need to move the robot to the position where all joints are in the middle of their ranges. Then after pressing enter you have to move each joint through its full range of motion.
##### Calibration video
<div class="video-container">
<video controls width="600">
<source
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/calibrate_so101_2.mp4"
type="video/mp4"
/>
</video>
</div>
#### Leader
Do the same steps to calibrate the leader arm, run the following command or API example:
<hfoptions id="calibrate_leader">
<hfoption id="Command">
```bash
lerobot-calibrate \
--teleop.type=so101_leader \
--teleop.port=/dev/tty.usbmodem58760431551 \ # <- The port of your robot
--teleop.id=my_awesome_leader_arm # <- Give the robot a unique name
```
</hfoption>
<hfoption id="API example">
<!-- prettier-ignore-start -->
```python
from lerobot.teleoperators.so_leader import SO101LeaderConfig, SO101Leader
config = SO101LeaderConfig(
port="/dev/tty.usbmodem58760431551",
id="my_awesome_leader_arm",
)
leader = SO101Leader(config)
leader.connect(calibrate=False)
leader.calibrate()
leader.disconnect()
```
<!-- prettier-ignore-end -->
</hfoption>
</hfoptions>
Congrats 🎉, your robot is all set to learn a task on its own. Start training it by following this tutorial: [Getting started with real-world robots](./il_robots)
> [!TIP]
> If you have any questions or need help, please reach out on [Discord](https://discord.com/invite/s3KuuzsPFb).
docs/source/streaming_video_encoding.mdx
+155
View File
@@ -0,0 +1,155 @@
# Streaming Video Encoding Guide
## 1. Overview
Streaming video encoding eliminates the traditional PNG round-trip during video dataset recording. Instead of:
1. Capture frame -> write PNG to disk -> (at episode end) read PNG's -> encode to MP4 -> delete PNG's
Frames can be encoded in real-time during capture:
1. Capture frame -> queue to encoder thread -> encode to MP4 directly
This makes `save_episode()` near-instant (the video is already encoded by the time the episode ends) and removes the blocking wait that previously occurred between episodes, especially with multiple cameras in long episodes.
## 2. Tuning Parameters
| Parameter | CLI Flag | Type | Default | Description |
| ----------------------- | --------------------------------- | ------------- | ------------- | ----------------------------------------------------------------- |
| `streaming_encoding` | `--dataset.streaming_encoding` | `bool` | `True` | Enable real-time encoding during capture |
| `vcodec` | `--dataset.rgb_encoder.vcodec` | `str` | `"libsvtav1"` | Video codec. `"auto"` detects best HW encoder |
| `encoder_threads` | `--dataset.encoder_threads` | `int \| None` | `None` (auto) | Threads per encoder instance. `None` will leave the vcoded decide |
| `encoder_queue_maxsize` | `--dataset.encoder_queue_maxsize` | `int` | `30` | Max buffered frames per camera (~1s at 30fps). Consumes RAM |
## 3. Performance Considerations
Streaming encoding means the CPU is encoding video **during** the capture loop, not after. This creates a CPU budget that must be shared between:
- **Control loop** (reading cameras, control the robot, writing non-video data)
- **Encoder threads** (one pool per camera)
- **Rerun visualization** (if enabled)
- **OS and other processes**
### Resolution & Number of Cameras Impact
| Setup | Throughput (px/sec) | CPU Encoding Load | Notes |
| ------------------------- | ------------------- | ----------------- | ------------------------------ |
| 2camsx 640x480x3 @30fps | 55M | Low | Works on most systems |
| 2camsx 1280x720x3 @30fps | 165M | Moderate | Comfortable on modern systems |
| 2camsx 1920x1080x3 @30fps | 373M | High | Requires powerful high-end CPU |
### `encoder_threads` Tuning
This parameter controls how many threads each encoder instance uses internally:
- **Higher values** (e.g., 4-5): Faster encoding, but uses more CPU cores per camera. Good for high-end systems with many cores.
- **Lower values** (e.g., 1-2): Less CPU per camera, freeing cores for capture and visualization. Good for low-res images and capable CPUs.
- **`None` (default)**: Lets the codec decide. Information available in the codec logs.
### Backpressure and Frame Dropping
Each camera has a bounded queue (`encoder_queue_maxsize`, default 30 frames). When the encoder can't keep up:
1. The queue fills up (consuming RAM)
2. New frames are **dropped** (not blocked) — the capture loop continues uninterrupted
3. A warning is logged: `"Encoder queue full for {camera}, dropped N frame(s)"`
4. At episode end, total dropped frames per camera are reported
### Symptoms of Encoder Falling Behind
- **System feels laggy and freezes**: all CPUs are at 100%
- **Dropped frame warnings** in the log or lower frames/FPS than expected in the recorded dataset
- **Choppy robot movement**: If CPU is severely overloaded, even the capture loop may be affected
- **Accumulated rerun lag**: Visualization falls behind real-time
## 4. Hardware-Accelerated Encoding
### When to Use
Use HW encoding when:
- CPU is the bottleneck (dropped frames, choppy robot, rerun lag)
- You have compatible hardware (GPU or dedicated encoder)
- You're recording at high throughput (high resolution or with many cameras)
### Choosing a Codec
| Codec | CPU Usage | File Size | Quality | Notes |
| --------------------- | --------- | -------------- | ------- | ---------------------------------------------------------------- |
| `libsvtav1` (default) | High | Smallest | Best | Default. Best compression but most CPU-intensive |
| `h264` | Medium | ~30-50% larger | Good | Software H.264. Lower CPU |
| HW encoders | Very Low | Largest | Good | Offloads to dedicated hardware. Best for CPU-constrained systems |
### Available HW Encoders
| Encoder | Platform | Hardware | CLI Value |
| ------------------- | ------------- | ------------------------------------------------------------------------------------------------ | ------------------------------------------------ |
| `h264_videotoolbox` | macOS | Apple Silicon / Intel | `--dataset.rgb_encoder.vcodec=h264_videotoolbox` |
| `hevc_videotoolbox` | macOS | Apple Silicon / Intel | `--dataset.rgb_encoder.vcodec=hevc_videotoolbox` |
| `h264_nvenc` | Linux/Windows | NVIDIA GPU | `--dataset.rgb_encoder.vcodec=h264_nvenc` |
| `hevc_nvenc` | Linux/Windows | NVIDIA GPU | `--dataset.rgb_encoder.vcodec=hevc_nvenc` |
| `h264_vaapi` | Linux | Intel/AMD GPU | `--dataset.rgb_encoder.vcodec=h264_vaapi` |
| `h264_qsv` | Linux/Windows | Intel Quick Sync | `--dataset.rgb_encoder.vcodec=h264_qsv` |
| `auto` | Any | Probes the system for available HW encoders. Falls back to `libsvtav1` if no HW encoder is found | `--dataset.rgb_encoder.vcodec=auto` |
> [!NOTE]
> In order to use the HW accelerated encoders you might need to upgrade your GPU drivers.
> [!NOTE]
> `libsvtav1` is the default because it provides the best training performance; other vcodecs can reduce CPU usage and be faster, but they typically produce larger files and may affect training time.
## 5. Troubleshooting
| Symptom | Likely Cause | Fix |
| ------------------------------------------------------------------ | -------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
| System freezes or choppy robot movement or Rerun visualization lag | CPU starved (100% load usage) | Close other apps, reduce encoding throughput, lower `encoder_threads`, use `h264`, use `display_data=False`. If the CPU continues to be at 100% then it might be insufficient for your setup, consider `--dataset.streaming_encoding=false` or HW encoding (`--dataset.rgb_encoder.vcodec=auto`) |
| "Encoder queue full" warnings or dropped frames in dataset | Encoder can't keep up (Queue overflow) | If CPU is not at 100%: Increase `encoder_threads`, increase `encoder_queue_maxsize` or use HW encoding (`--dataset.rgb_encoder.vcodec=auto`). |
| High RAM usage | Queue filling faster than encoding | `encoder_threads` too low or CPU insufficient. Reduce `encoder_queue_maxsize` or use HW encoding |
| Large video files | Using HW encoder or H.264 | Expected trade-off. Switch to `libsvtav1` if CPU allows |
| `save_episode()` still slow | `streaming_encoding` is `False` | Set `--dataset.streaming_encoding=true` |
| Encoder thread crash | Codec not available or invalid settings | Check `vcodec` is installed, try `--dataset.rgb_encoder.vcodec=auto` |
| Recorded dataset is missing frames | CPU/GPU starvation or occasional load spikes | If ~5% of frames are missing, your system is likely overloaded — follow the recommendations above. If fewer frames are missing (~2%), they are probably due to occasional transient load spikes (often at startup) and can be considered expected. |
## 6. Recommended Configurations
These estimates are conservative; we recommend testing them on your setup—start with a low load and increase it gradually.
### High-End Systems: modern 12+ cores (24+ threads)
A throughput between ~250-500M px/sec should be comfortable in CPU. For even better results try HW encoding if available.
```bash
# 3camsx 1280x720x3 @30fps: Defaults work well. Optionally increase encoder parallelism.
# 2camsx 1920x1080x3 @30fps: Defaults work well. Optionally increase encoder parallelism.
lerobot-record --dataset.encoder_threads=5 ...
# 3camsx 1920x1080x3 @30fps: Might require some tuning.
```
### Mid-Range Systems: modern 8+ cores (16+ threads) or Apple Silicon
A throughput between ~80-300M px/sec should be possible in CPU.
```bash
# 3camsx 640x480x3 @30fps: Defaults work well. Optionally decrease encoder parallelism.
# 2camsx 1280x720x3 @30fps: Defaults work well. Optionally decrease encoder parallelism.
lerobot-record --dataset.encoder_threads=2 ...
# 2camsx 1920x1080x3 @30fps: Might require some tuning.
```
### Low-Resource Systems: modern 4+ cores (8+ threads) or Raspberry Pi 5
On very constrained systems, streaming encoding may compete too heavily with the capture loop. Disabling it falls back to the PNG-based approach where encoding happens between episodes (blocking, but doesn't interfere with capture). Alternatively, record at a lower throughput to reduce both capture and encoding load. Consider also changing codec to `h264` and using batch encoding.
```bash
# 2camsx 640x480x3 @30fps: Requires some tuning.
# Use H.264, disable streaming, consider batching encoding
lerobot-record --dataset.rgb_encoder.vcodec=h264 --dataset.streaming_encoding=false ...
```
## 7. Closing note
Performance ultimately depends on your exact setup — frames-per-second, resolution, CPU cores and load, available memory, episode length, and the encoder you choose. Always test with your target workload, be mindful about your CPU & system capabilities and tune `encoder_threads`, `encoder_queue_maxsize`, and
`vcodec` reasonably. That said, a common practical configuration (for many applications) is three cameras at 640×480x3 @30fps; this usually runs fine with the default streaming video encoding settings in modern systems. Always verify your recorded dataset is healthy by comparing the video duration to the CLI episode duration and confirming the row count equals FPS × CLI duration.
docs/source/tools.mdx
+210
View File
@@ -0,0 +1,210 @@
# Tools
LeRobot v3.1 supports **tool calls** in policies — assistant messages can
emit structured invocations like `say(text="OK, starting now")` that the
runtime dispatches to a real implementation (TTS, controller, logger, …).
This page covers:
1. Where the tool catalog lives.
2. How the annotation pipeline produces tool-call atoms.
3. How to add your own tool.
## Where tools are declared
Two layers.
**The catalog** — a list of OpenAI-style function schemas — lives at
`meta/info.json["tools"]` on each dataset. Example:
```json
{
"features": { "...": "..." },
"tools": [
{
"type": "function",
"function": {
"name": "say",
"description": "Speak a short utterance to the user via the TTS executor.",
"parameters": {
"type": "object",
"properties": {
"text": {
"type": "string",
"description": "The verbatim text to speak."
}
},
"required": ["text"]
}
}
}
]
}
```
Read it via the dataset metadata accessor:
```python
from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
meta = LeRobotDatasetMetadata(repo_id="pepijn/super_poulain_final_annotations")
tools = meta.tools # list[dict] — OpenAI tool schemas
```
If the dataset's `info.json` doesn't declare any tools, `meta.tools`
returns `DEFAULT_TOOLS` from `lerobot.datasets.language` — currently a
single-entry list with the canonical `say` schema. So unannotated
datasets and chat-template consumers keep working without any
configuration:
```python
prompt_str = tokenizer.apply_chat_template(
sample["messages"],
tools=meta.tools, # works either way
add_generation_prompt=False,
tokenize=False,
)
```
**The implementations** — runnable Python — will live under
`src/lerobot/tools/`, one file per tool. The runtime dispatcher and
the canonical `say` implementation (wrapping Kyutai's pocket-tts) are
not part of the catalog layer described here; today this layer ships
only the schema storage and the `DEFAULT_TOOLS` fallback constant.
## Per-row tool _invocations_
The catalog above describes _what can be called_. The actual _call_ — the
function name plus the argument values — is stored per-row, on the
assistant atoms in `language_events`:
```python
{
"role": "assistant",
"content": null,
"style": null,
"timestamp": 12.4,
"camera": null,
"tool_calls": [
{ "type": "function",
"function": { "name": "say", "arguments": { "text": "On it." } } }
]
}
```
Recipes splice these into rendered messages via `tool_calls_from`:
```yaml
user_interjection_response:
bindings:
speech: "emitted_at(t, role=assistant, tool_name=say)"
messages:
- { role: user, content: "${task}", stream: high_level }
- {
role: assistant,
content: "${current_plan}",
stream: high_level,
target: true,
tool_calls_from: speech,
}
```
The model's training target is one assistant turn that carries both the
plan text _and_ the `say` tool call. At inference, the runtime parses
the generated text back into structured `tool_calls` and dispatches to
the matching implementation.
## How to add your own tool
> **Note:** Steps 2 and 3 below describe the runtime layer
> (`src/lerobot/tools/`, the `Tool` protocol, `TOOL_REGISTRY`,
> `get_tools(meta)`) which is not part of the catalog layer shipped
> today — those modules don't yet exist in the tree. Step 1 alone is
> enough to make the tool visible to the chat template via
> `meta.tools` so the model can learn to _generate_ the call;
> executing the call at inference requires the runtime layer.
Three steps. Concrete example: a `record_observation` tool the policy
can call to capture an extra observation outside the regular control
loop.
### Step 1 — declare the schema
Add an entry under `meta/info.json["tools"]`. Either edit the file
directly on disk _before_ running the annotation pipeline (it'll be
preserved) or hand it to `lerobot-annotate` via a config flag.
```json
{
"tools": [
{ "type": "function", "function": { "name": "say", "...": "..." } },
{
"type": "function",
"function": {
"name": "record_observation",
"description": "Capture a high-resolution still image for the user.",
"parameters": {
"type": "object",
"properties": {
"label": {
"type": "string",
"description": "Short label for the saved image."
}
},
"required": ["label"]
}
}
}
]
}
```
The schema follows OpenAI's function-calling convention exactly, so the
chat template can render it natively.
### Step 2 — implement the call
Create `src/lerobot/tools/record_observation.py`:
```python
from .base import Tool
from typing import Any
RECORD_OBSERVATION_SCHEMA: dict[str, Any] = { "...": "..." } # mirrors the JSON above
class RecordObservationTool:
name = "record_observation"
schema = RECORD_OBSERVATION_SCHEMA
def __init__(self, schema: dict | None = None, output_dir: str = "."):
self.output_dir = output_dir
def call(self, arguments: dict) -> str:
label = arguments["label"]
# ... save the latest camera frame to <output_dir>/<label>.png ...
return f"saved {label}.png"
```
One file per tool keeps dependencies isolated — `record_observation`
might pull `pillow`, while `say` pulls `pocket-tts`. Users installing
only the tools they need avoid heavy transitive deps.
### Step 3 — register it
Add to `src/lerobot/tools/registry.py`:
```python
from .record_observation import RecordObservationTool
TOOL_REGISTRY["record_observation"] = RecordObservationTool
```
That's it. At runtime `get_tools(meta)` looks up each schema in
`meta.tools`, instantiates the matching registered class, and returns
a name → instance dict the dispatcher can route into.
If you want to use a tool _without_ writing an implementation (e.g. for
training-time chat-template formatting only), step 1 alone is enough —
the model still learns to _generate_ the call. Steps 2 and 3 are only
needed to actually _execute_ it at inference.
docs/source/topreward.mdx
+177
View File
@@ -0,0 +1,177 @@
# TOPReward
TOPReward is a **zero-shot reward model** that extracts token log-probabilities from an off-the-shelf vision-language model (VLM) as a robotic reward signal. Given a video trajectory and a task instruction, it returns the VLM's log-likelihood that the instruction is true — no fine-tuning required.
**Paper**: [TOPReward: Token Probabilities as Hidden Zero-Shot Rewards for Robotics](https://arxiv.org/abs/2602.19313)
**Project**: [topreward.github.io](https://topreward.github.io/webpage/)
**Original code**: [github.com/TOPReward/TOPReward](https://github.com/TOPReward/TOPReward)
**Default backbone**: [Qwen/Qwen3-VL-8B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct)
## Overview
TOPReward asks a generic VLM how likely a task instruction is, **conditioned on the video** of a robot trying to complete that task. Concretely, given:
- A trajectory video (a sequence of frames).
- A task instruction (e.g. _"open the drawer"_).
it builds a chat prompt of the form
```text
<video>
"The above video shows a robot manipulation trajectory that completes the
following task: <instruction> Decide whether the above statement is True
or not. The answer is: True"
```
forwards it through the VLM, label-masks everything except the very last token, and reads back the log-probability of that token — by default the literal `"True"` that closes the suffix template. The resulting `log P("True" | video + prompt + instruction)` is the reward.
Because the method only depends on a frozen VLM, TOPReward is **zero-shot**: there are no fine-tuned weights to host. The "model" in LeRobot is a small wrapper around `transformers`' `Qwen3VLForConditionalGeneration` plus the label-masking logic. The processor owns the tokeniser and builds the full chat prompt (EO-1/Robometer pattern).
## What the LeRobot integration covers
- Standard `reward_model.type=topreward` configuration through LeRobot.
- VLM loading via the `transformers` `Qwen3VLForConditionalGeneration` API.
- Prompt assembly + tokenisation in the processor (matching upstream `QwenClient.compute_instruction_reward`).
- `compute_reward()` returns one scalar log-prob per sample.
- LeRobot reward-model save/load — `save_pretrained` writes only `config.json` (the VLM is identified by `vlm_name`).
- An offline labeling script that writes a `topreward_progress.parquet` (SARM-compatible schema) for RA-BC and overlay.
The current LeRobot port supports the **Qwen3-VL client only**. Other upstream clients (Gemini, OpenAI, Gemma, Molmo) can be added as follow-up extras.
## Installation Requirements
1. Install LeRobot following the [Installation Guide](./installation).
2. Install the TOPReward optional extra:
```bash
pip install -e ".[topreward]"
```
or, with `uv` from a source checkout:
```bash
uv sync --extra topreward
```
This pulls in `transformers`. The first time you run TOPReward, Hugging Face will also download the VLM weights from the Hub (~16 GB for Qwen3-VL-8B-Instruct). A GPU is strongly recommended.
## Model Inputs and Outputs
TOPReward expects:
- A trajectory video or sequence of frames.
- A natural-language task description.
In LeRobot datasets the preprocessor reads:
| Config field | Default | Meaning |
| ------------------------- | --------------------------- | --------------------------------------------- |
| `reward_model.image_key` | `observation.images.top` | Camera observation used by TOPReward |
| `reward_model.task_key` | `task` | Key in complementary data for the task string |
| `reward_model.max_frames` | `16` | Cap on frames per sample |
| `reward_model.fps` | `2.0` | Metadata passed to the Qwen video processor |
| `reward_model.vlm_name` | `Qwen/Qwen3-VL-8B-Instruct` | Hugging Face Hub id of the underlying VLM |
The model returns:
- `compute_reward(batch)`: one log-probability per sample. Higher = better task-video alignment. When `success_threshold` is finite, returns the binary thresholded value instead.
## Usage
### Load the reward model directly
```python
from lerobot.rewards.topreward import TOPRewardConfig, TOPRewardModel
cfg = TOPRewardConfig(
vlm_name="Qwen/Qwen3-VL-8B-Instruct",
device="cuda",
)
reward_model = TOPRewardModel(cfg)
```
### Use the reward factory
```python
from lerobot.rewards import make_reward_model, make_reward_model_config, make_reward_pre_post_processors
cfg = make_reward_model_config(
"topreward",
vlm_name="Qwen/Qwen3-VL-8B-Instruct",
device="cuda",
image_key="observation.images.top",
)
reward_model = make_reward_model(cfg)
preprocessor, postprocessor = make_reward_pre_post_processors(cfg)
```
The preprocessor tokenises the full prompt (video + prefix + instruction suffix), writes Qwen-VL tensors + `prompt_length` under `observation.topreward.*`. The model reads those tensors, label-masks based on `prompt_length`, and extracts the log-prob reward.
### Offline dataset labeling
Write a `topreward_progress.parquet` for RA-BC training and overlay videos:
```bash
# Sparse-dense (15 anchors per episode, matches upstream)
uv run python -m lerobot.rewards.topreward.compute_rabc_weights \
--dataset-repo-id lerobot/libero_10_image \
--num-samples 15 \
--device cuda
```
Then render the progress overlay for any episode:
```bash
uv run examples/dataset/create_progress_videos.py \
--repo-id lerobot/libero_10_image \
--episode 0 \
--progress-file topreward_progress.parquet \
--gif
```
## Configuration Notes
### Prompt knobs
The default prompt mirrors the upstream paper:
```text
prompt_prefix = "The above video shows a robot manipulation trajectory that completes the following task: "
prompt_suffix_template = "{instruction} Decide whether the above statement is True or not. The answer is: True"
```
Both are exposed on `TOPRewardConfig` for ablation. The suffix template **must** contain `{instruction}`.
### Chat template
`add_chat_template=True` wraps the full prompt (including instruction) with the tokenizer's chat template before tokenisation. Default is `False`, matching the upstream paper's main experiments.
## Limitations
- The current LeRobot port is **inference-only and zero-shot**; `forward()` is not overridden and `is_trainable` returns `False`.
- Only the **Qwen3-VL family** is supported; other upstream clients are out of scope.
- TOPReward inherits the underlying VLM's biases.
## References
- [TOPReward project page](https://topreward.github.io/webpage/)
- [TOPReward paper](https://arxiv.org/abs/2602.19313)
- [Original TOPReward code](https://github.com/TOPReward/TOPReward)
- [Qwen3-VL-8B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct)
## Citation
```bibtex
@article{chen2026topreward,
title={TOPReward: Token Probabilities as Hidden Zero-Shot Rewards for Robotics},
author={Chen, Shirui and Harrison, Cole and Lee, Ying-Chun and Yang, Angela Jin and
Ren, Zhongzheng and Ratliff, Lillian J and Duan, Jiafei and Fox, Dieter and
Krishna, Ranjay},
journal={arXiv preprint arXiv:2602.19313},
year={2026}
}
```
## License
The original TOPReward codebase is MIT-licensed. The LeRobot port follows the LeRobot Apache 2.0 license; the wrapped Qwen3-VL weights are subject to the original Qwen license.
docs/source/torch_accelerators.mdx
+42
View File
@@ -0,0 +1,42 @@
# PyTorch accelerators
LeRobot supports multiple hardware acceleration options for both training and inference.
These options include:
- **CPU**: CPU executes all computations, no dedicated accelerator is used
- **CUDA**: acceleration with NVIDIA & AMD GPUs
- **MPS**: acceleration with Apple Silicon GPUs
- **XPU**: acceleration with Intel integrated and discrete GPUs
## Getting Started
To use particular accelerator, a suitable version of PyTorch should be installed.
For CPU, CUDA, and MPS backends follow instructions provided on [PyTorch installation page](https://pytorch.org/get-started/locally).
For XPU backend, follow instructions from [PyTorch documentation](https://docs.pytorch.org/docs/stable/notes/get_start_xpu.html).
### Verifying the installation
After installation, accelerator availability can be verified by running
```python
import torch
print(torch.<backend_name>.is_available()) # <backend_name> is cuda, mps, or xpu
```
## How to run training or evaluation
To select the desired accelerator, use the `--policy.device` flag when running `lerobot-train` or `lerobot-eval`. For example, to use MPS on Apple Silicon, run:
```bash
lerobot-train
--policy.device=mps ...
```
```bash
lerobot-eval \
--policy.device=mps ...
```
However, in most cases, presence of an accelerator is detected automatically and `policy.device` parameter can be omitted from CLI commands.
docs/source/unitree_g1.mdx
+303
View File
@@ -0,0 +1,303 @@
# Unitree G1
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/unitree_thumbnail.jpg"
alt="Unitree G1 locomanipulation demo"
style={{ width: "100%" }}
/>
The Unitree G1 humanoid is now supported in LeRobot! You can teleoperate, train locomanipulation policies, test in sim, and more. Both 29 and 23 DoF variants are supported.
---
## Part 1: Getting Started
### Install the Unitree SDK
Follow the [unitree_sdk2_python installation guide](https://github.com/unitreerobotics/unitree_sdk2_python#installation). Tested with `unitree_sdk2py==1.0.1` and `cyclonedds==0.10.2`:
```bash
conda create -y -n lerobot python=3.12
conda activate lerobot
git clone https://github.com/unitreerobotics/unitree_sdk2_python.git
cd unitree_sdk2_python
pip install -e .
cd ..
```
### Install LeRobot
```bash
conda install ffmpeg -c conda-forge
conda install -c conda-forge "pinocchio>=3.0.0,<4.0.0"
git clone https://github.com/huggingface/lerobot.git
cd lerobot
pip install -e '.[unitree_g1]'
```
<Tip>
For now, pinocchio must be installed from conda-forge (not pip) to include the
CasADi bindings needed for arm IK.
</Tip>
### Test the Installation (Simulation)
The simulation environment has its own dependencies. Check the Simulation environment dependencies: [Unitree G1 Mujoco EnvHub](https://huggingface.co/lerobot/unitree-g1-mujoco/tree/main).
```bash
pip install mujoco loguru msgpack msgpack-numpy
```
```bash
lerobot-teleoperate \
--robot.type=unitree_g1 \
--robot.is_simulation=true \
--teleop.type=unitree_g1 \
--teleop.id=wbc_unitree \
--robot.cameras='{"global_view": {"type": "zmq", "server_address": "localhost", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30, "warmup_s": 5}}' \
--display_data=true \
--robot.controller=GrootLocomotionController
```
This will launch a [MuJoCo sim instance](https://huggingface.co/lerobot/unitree-g1-mujoco/tree/main) for the G1. You can connect a gamepad to your machine before launching in order to control the robot's locomotion in sim. We support both [HolosomaLocomotionController](https://github.com/amazon-far/holosoma) and [GrootLocomotionController](https://github.com/NVlabs/GR00T-WholeBodyControl) via `--robot.controller`.
- Press `9` to release the robot
- Press `7` / `8` to increase / decrease waist height
### Connect to the Physical Robot
The G1's Ethernet IP is fixed at `192.168.123.164`. Your machine must have a static IP on the same subnet: `192.168.123.x` where `x ≠ 164`.
```bash
# Replace 'enp131s0' with your ethernet interface name (check with `ip a`)
sudo ip addr flush dev enp131s0
sudo ip addr add 192.168.123.200/24 dev enp131s0
sudo ip link set enp131s0 up
```
### SSH into the Robot
```bash
ssh unitree@192.168.123.164
# Password: 123
```
### Share Internet via Ethernet
The G1 needs internet access to clone repos and install packages. Share your laptop's connection over Ethernet:
**On your laptop:**
```bash
sudo sysctl -w net.ipv4.ip_forward=1
# Replace wlp132s0f0 with your WiFi interface name
sudo iptables -t nat -A POSTROUTING -o wlp132s0f0 -s 192.168.123.0/24 -j MASQUERADE
sudo iptables -A FORWARD -i wlp132s0f0 -o enp131s0 -m state --state RELATED,ESTABLISHED -j ACCEPT
sudo iptables -A FORWARD -i enp131s0 -o wlp132s0f0 -j ACCEPT
```
**On the G1:**
```bash
sudo ip route del default 2>/dev/null || true
sudo ip route add default via 192.168.123.200 dev eth0
echo "nameserver 8.8.8.8" | sudo tee /etc/resolv.conf
# Verify
ping -c 3 8.8.8.8
```
### Install the Unitree SDK on the G1
Follow the [unitree_sdk2_python installation guide](https://github.com/unitreerobotics/unitree_sdk2_python#installation):
```bash
conda create -y -n lerobot python=3.12
conda activate lerobot
git clone https://github.com/unitreerobotics/unitree_sdk2_python.git
cd unitree_sdk2_python
python -m pip install -e .
cd ..
```
### Install LeRobot on the G1
```bash
git clone https://github.com/huggingface/lerobot.git
cd lerobot
conda install -c conda-forge "pinocchio>=3.0.0,<4.0.0"
python -m pip install -e '.[unitree_g1]'
```
<Tip>
For now, pinocchio must be installed from conda-forge (not pip) to include the
CasADi bindings needed for arm IK.
</Tip>
### (Optional) Enable WiFi on the Robot
For wireless SSH access, you can enable WiFi on the G1 (it's blocked by default):
```bash
sudo rfkill unblock all
sudo ip link set wlan0 up
sudo nmcli radio wifi on
sudo nmcli device set wlan0 managed yes
sudo systemctl restart NetworkManager
```
**Connect to a WiFi network:**
```bash
nmcli device wifi list
sudo nmcli connection add type wifi ifname wlan0 con-name "YourNetwork" ssid "YourNetwork"
sudo nmcli connection modify "YourNetwork" wifi-sec.key-mgmt wpa-psk
sudo nmcli connection modify "YourNetwork" wifi-sec.psk "YourPassword"
sudo nmcli connection modify "YourNetwork" connection.autoconnect yes
sudo nmcli connection up "YourNetwork"
ip a show wlan0
```
You can then SSH over WiFi instead of Ethernet:
```bash
ssh unitree@<ROBOT_WIFI_IP>
# Password: 123
```
---
## Part 2: Teleoperation & Locomotion
### Run the Robot Server
On the robot (from `~/lerobot`):
```bash
cd ~/lerobot
python src/lerobot/robots/unitree_g1/run_g1_server.py --camera
```
### Run the Locomotion Policy
You can run the teleoperation client from your laptop over Ethernet, over WiFi (experimental), or directly on the robot itself. Mind potential latency introduced by your network.
**From your laptop:**
```bash
lerobot-teleoperate \
--robot.type=unitree_g1 \
--robot.is_simulation=false \
--robot.robot_ip=<ROBOT_IP> \
--teleop.type=unitree_g1 \
--teleop.id=wbc_unitree \
--robot.cameras='{"global_view": {"type": "zmq", "server_address": "<ROBOT_IP>", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
--display_data=true \
--robot.controller=HolosomaLocomotionController
```
We support both [GrootLocomotionController](https://github.com/NVlabs/GR00T-WholeBodyControl) and [HolosomaLocomotionController](https://github.com/amazon-far/holosoma) via `--robot.controller`.
---
## Part 3: Loco-Manipulation with the Homunculus Exoskeleton
We provide a loco-manipulation solution via the Homunculus Exoskeleton — an open-source 7 DoF exoskeleton for whole-body control. Check it out [here](https://github.com/nepyope/hmc_exo).
### Calibrate
```bash
lerobot-calibrate \
--teleop.type=unitree_g1 \
--teleop.left_arm_config.port=/dev/ttyACM1 \
--teleop.right_arm_config.port=/dev/ttyACM0 \
--teleop.id=exo
```
During calibration move each joint through its entire range. After fitting, move the joint in a neutral position and press `n` to advance.
### Record a Dataset
```bash
lerobot-record \
--robot.type=unitree_g1 \
--robot.is_simulation=true \
--robot.cameras='{"global_view": {"type": "zmq", "server_address": "localhost", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
--teleop.type=unitree_g1 \
--teleop.left_arm_config.port=/dev/ttyACM1 \
--teleop.right_arm_config.port=/dev/ttyACM0 \
--teleop.id=exo \
--dataset.repo_id=your-username/dataset-name \
--dataset.single_task="Test" \
--dataset.num_episodes=2 \
--dataset.episode_time_s=5 \
--dataset.reset_time_s=5 \
--dataset.push_to_hub=true \
--dataset.streaming_encoding=true \
--dataset.encoder_threads=2
```
> **Note:** Omit `--teleop.left_arm_config.port` and `--teleop.right_arm_config.port` if you're only using the joystick.
Example dataset: [nepyope/unitree_box_move_blue_full](https://huggingface.co/datasets/nepyope/unitree_box_move_blue_full)
---
## Part 4: Training & Inference
### Train
```bash
python src/lerobot/scripts/lerobot_train.py \
--dataset.repo_id=your-username/dataset-name \
--policy.type=pi05 \
--output_dir=./outputs/pi05_training \
--job_name=pi05_training \
--policy.repo_id=your-username/your-repo-id \
--policy.pretrained_path=lerobot/pi05_base \
--policy.compile_model=true \
--policy.gradient_checkpointing=true \
--wandb.enable=true \
--policy.dtype=bfloat16 \
--policy.freeze_vision_encoder=false \
--policy.train_expert_only=false \
--steps=3000 \
--policy.device=cuda \
--batch_size=32
```
### Inference with RTC
Once trained, we recommend deploying policies using inference-time RTC:
```bash
lerobot-rollout \
--strategy.type=base \
--policy.path=your-username/your-repo-id \
--policy.device=cuda \
--robot.type=unitree_g1 \
--robot.is_simulation=false \
--robot.controller=HolosomaLocomotionController \
--robot.cameras='{"global_view": {"type": "zmq", "server_address": "<ROBOT_IP>", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
--task="task_description" \
--duration=1000 \
--fps=30 \
--inference.type=rtc
```
---
## Additional Resources
- [Unitree SDK Documentation](https://github.com/unitreerobotics/unitree_sdk2_python)
- [GR00T-WholeBodyControl](https://github.com/NVlabs/GR00T-WholeBodyControl)
- [Holosoma](https://github.com/amazon-far/holosoma)
- [LeRobot Documentation](https://github.com/huggingface/lerobot)
- [Unitree IL LeRobot](https://github.com/unitreerobotics/unitree_IL_lerobot)
---
_Last updated: March 2026_
docs/source/using_dataset_tools.mdx
+274
View File
@@ -0,0 +1,274 @@
# Using Dataset Tools
This guide covers the dataset tools utilities available in LeRobot for modifying and editing existing datasets.
## Overview
LeRobot provides several utilities for manipulating datasets:
1. **Delete Episodes** - Remove specific episodes from a dataset
2. **Split Dataset** - Divide a dataset into multiple smaller datasets
3. **Merge Datasets** - Combine multiple datasets into one. The datasets must have identical features, and episodes are concatenated in the order specified in `repo_ids`
4. **Add Features** - Add new features to a dataset
5. **Remove Features** - Remove features from a dataset
6. **Convert to Video** - Convert image-based datasets to video format for efficient storage (RGB and depth cameras are encoded with separate encoders)
7. **Re-encode Videos** - Re-encode an existing video dataset's RGB and/or depth streams with new encoder settings
8. **Show the Info of Datasets** - Show the summary of datasets information such as number of episode etc.
The core implementation is in `lerobot.datasets.dataset_tools`.
An example script detailing how to use the tools API is available in `examples/dataset/use_dataset_tools.py`.
## Command-Line Tool: lerobot-edit-dataset
`lerobot-edit-dataset` is a command-line script for editing datasets. It can be used to delete episodes, split datasets, merge datasets, add features, remove features, and convert image datasets to video format.
Run `lerobot-edit-dataset --help` for more information on the configuration of each operation.
### Usage Examples
#### Delete Episodes
Remove specific episodes from a dataset. This is useful for filtering out undesired data.
```bash
# Delete episodes 0, 2, and 5 (modifies original dataset)
lerobot-edit-dataset \
--repo_id lerobot/pusht \
--operation.type delete_episodes \
--operation.episode_indices "[0, 2, 5]"
# Delete episodes and save to a new dataset (preserves original dataset)
lerobot-edit-dataset \
--repo_id lerobot/pusht \
--new_repo_id lerobot/pusht_after_deletion \
--operation.type delete_episodes \
--operation.episode_indices "[0, 2, 5]"
```
#### Split Dataset
Divide a dataset into multiple subsets.
```bash
# Split by fractions (e.g. 80% train, 20% test, 20% val)
lerobot-edit-dataset \
--repo_id lerobot/pusht \
--operation.type split \
--operation.splits '{"train": 0.8, "test": 0.2, "val": 0.2}'
# Split by specific episode indices
lerobot-edit-dataset \
--repo_id lerobot/pusht \
--operation.type split \
--operation.splits '{"task1": [0, 1, 2, 3], "task2": [4, 5]}'
```
There are no constraints on the split names, they can be determined by the user. Resulting datasets are saved under the repo id with the split name appended, e.g. `lerobot/pusht_train`, `lerobot/pusht_task1`, `lerobot/pusht_task2`.
#### Merge Datasets
Combine multiple datasets into a single dataset.
```bash
# Merge train and validation splits back into one dataset
lerobot-edit-dataset \
--repo_id lerobot/pusht_merged \
--operation.type merge \
--operation.repo_ids "['lerobot/pusht_train', 'lerobot/pusht_val']"
```
#### Remove Features
Remove features from a dataset.
```bash
# Remove a camera feature
lerobot-edit-dataset \
--repo_id lerobot/pusht \
--operation.type remove_feature \
--operation.feature_names "['observation.images.top']"
```
#### Convert to Video
Convert an image-based dataset to video format, creating a new LeRobotDataset where images are stored as videos. This is useful for reducing storage requirements and improving data loading performance. The new dataset will have the exact same structure as the original, but with images encoded as MP4 videos in the proper LeRobot format.
```bash
# Local-only: Save to a custom output directory (no hub push)
lerobot-edit-dataset \
--repo_id lerobot/pusht_image \
--operation.type convert_image_to_video \
--operation.output_dir /path/to/output/pusht_video
# Save with new repo_id (local storage)
lerobot-edit-dataset \
--repo_id lerobot/pusht_image \
--new_repo_id lerobot/pusht_video \
--operation.type convert_image_to_video
# Convert and push to Hugging Face Hub
lerobot-edit-dataset \
--repo_id lerobot/pusht_image \
--new_repo_id lerobot/pusht_video \
--operation.type convert_image_to_video \
--push_to_hub true
# Convert with custom video codec and quality settings
lerobot-edit-dataset \
--repo_id lerobot/pusht_image \
--operation.type convert_image_to_video \
--operation.output_dir outputs/pusht_video \
--operation.rgb_encoder.vcodec libsvtav1 \
--operation.rgb_encoder.pix_fmt yuv420p \
--operation.rgb_encoder.g 2 \
--operation.rgb_encoder.crf 30
# Convert a dataset that includes depth maps, customizing the depth encoder
lerobot-edit-dataset \
--repo_id lerobot/pusht_image \
--operation.type convert_image_to_video \
--operation.output_dir outputs/pusht_video \
--operation.depth_encoder.depth_min 0.01 \
--operation.depth_encoder.depth_max 10.0 \
--operation.depth_encoder.use_log true
# Convert only specific episodes
lerobot-edit-dataset \
--repo_id lerobot/pusht_image \
--operation.type convert_image_to_video \
--operation.output_dir outputs/pusht_video \
--operation.episode_indices "[0, 1, 2, 5, 10]"
# Convert with multiple workers for parallel processing
lerobot-edit-dataset \
--repo_id lerobot/pusht_image \
--operation.type convert_image_to_video \
--operation.output_dir outputs/pusht_video \
--operation.num_workers 8
# For memory-constrained systems, users can now specify limits:
lerobot-edit-dataset \
--repo_id lerobot/pusht_image \
--operation.type convert_to_video \
--operation.max_episodes_per_batch 50 \
--operation.max_frames_per_batch 10000
```
**Parameters:**
- `output_dir`: Custom output directory (optional - by default uses `new_repo_id` or `{repo_id}_video`)
- `rgb_encoder`: Video encoder settings applied to RGB cameras — all sub-fields accessible via `--operation.rgb_encoder.<field>`. See [Video Encoding Parameters](./video_encoding_parameters) for more details.
- `depth_encoder`: Video encoder settings applied to depth-map cameras (e.g. from an Intel RealSense). In addition to the standard encoder fields it exposes the depth quantization knobs (`depth_min`, `depth_max`, `shift`, `use_log`), accessible via `--operation.depth_encoder.<field>`. These quantization settings are persisted to the dataset metadata so depth can be dequantized back to physical units on load. See the [Depth streams](./video_encoding_parameters#depth-streams) section for details.
- `episode_indices`: List of specific episodes to convert (default: all episodes)
- `num_workers`: Number of parallel workers for processing (default: 4)
**Note:** The resulting dataset will be a proper LeRobotDataset with all cameras encoded as videos in the `videos/` directory, with parquet files containing only metadata (no raw image data). Depth-map cameras are detected automatically and routed to the `depth_encoder`, while RGB cameras use the `rgb_encoder`. All episodes, stats, and tasks are preserved.
#### Re-encode Videos
Re-encode the videos of an existing video dataset with different encoder settings, without going back to raw frames. RGB videos use the `rgb_encoder` and depth videos use the `depth_encoder`. Provide only the encoder(s) you want to re-encode; the other stream type is left untouched.
```bash
# Re-encode all RGB videos with new settings (saves to lerobot/pusht_reencoded by default)
lerobot-edit-dataset \
--repo_id lerobot/pusht \
--operation.type reencode_videos \
--operation.rgb_encoder.vcodec h264 \
--operation.rgb_encoder.pix_fmt yuv420p \
--operation.rgb_encoder.crf 23
# Re-encode both RGB and depth videos in a dataset with depth maps
lerobot-edit-dataset \
--repo_id lerobot/pusht_depth \
--operation.type reencode_videos \
--operation.rgb_encoder.vcodec h264 \
--operation.depth_encoder.crf 50
```
**Parameters:**
- `rgb_encoder`: Encoder settings applied to every RGB video. Omit to skip re-encoding RGB videos.
- `depth_encoder`: Encoder settings applied to every depth video. Omit to skip re-encoding depth videos.
- `num_workers`: Number of parallel workers for processing.
> [!NOTE]
> When re-encoding depth videos, the existing depth quantization parameters (`depth_min`, `depth_max`, `shift`, `use_log`) and the `is_depth_map` flag are **preserved** — re-encoding only changes the codec/quality of the stored stream, not how depth is dequantized on load.
### Show the information of datasets
Show the information of datasets such as number of episode, number of frame, File size and so on.
No change will be made to the dataset
```bash
# Show dataset information without feature details
lerobot-edit-dataset \
--repo_id lerobot/pusht_image \
--operation.type info \
# Show dataset information with feature details
lerobot-edit-dataset \
--repo_id lerobot/pusht_image \
--operation.type info \
--operation.show_features true
```
**Parameters:**
- `parameters`: The flag to control show or no show dataset information with feature details.(default=false)
### Push to Hub
Add the `--push_to_hub true` flag to any command to automatically upload the resulting dataset to the Hugging Face Hub:
```bash
lerobot-edit-dataset \
--repo_id lerobot/pusht \
--new_repo_id lerobot/pusht_after_deletion \
--operation.type delete_episodes \
--operation.episode_indices "[0, 2, 5]" \
--push_to_hub true
```
There is also a tool for adding features to a dataset that is not yet covered in `lerobot-edit-dataset`.
# Dataset Visualization
## Online Visualization
When you record a dataset using `lerobot`, it automatically uploads to the Hugging Face Hub unless you specify otherwise. To view the dataset online, use our **LeRobot Dataset Visualizer**, available at:
https://huggingface.co/spaces/lerobot/visualize_dataset
## Local Visualization
You can also visualize episodes from a dataset locally using our command-line tool.
**From the Hugging Face Hub:**
```bash
lerobot-dataset-viz \
--repo-id lerobot/pusht \
--episode-index 0
```
**From a local folder:**
Add the `--root` option and set `--mode local`. For example, to search in `./my_local_data_dir/lerobot/pusht`:
```bash
lerobot-dataset-viz \
--repo-id lerobot/pusht \
--root ./my_local_data_dir \
--mode local \
--episode-index 0
```
Once executed, the tool opens `rerun.io` and displays the camera streams, robot states, and actions for the selected episode.
To use [Foxglove](https://foxglove.dev) instead of Rerun, install the extra add `--display-mode foxglove`. This starts a WebSocket server (connect the Foxglove app to `ws://127.0.0.1:8765`) that serves the episode as a seekable timeline you can play/pause and scrub.
For advanced usage—including visualizing datasets stored on a remote server—run:
```bash
lerobot-dataset-viz --help
```
docs/source/video_encoding_parameters.mdx
+255
View File
@@ -0,0 +1,255 @@
# Video encoding parameters
When video storage is enabled, LeRobot stores each camera stream as an **MP4** file instead of saving one image file per timestep. Video encoding compresses across time, which usually cuts dataset size and I/O compared to a pile of PNG, while keeping MP4 — a format every player and loader understands.
Encoding frames into an MP4 is a full FFmpeg pipeline: choice of encoder, pixel format, GOP/keyframes, quality vs. speed, and optional extra encoder flags. Most of these knobs are user-tunable through `rgb_encoder`, a nested `RGBEncoderConfig` (`lerobot.configs.video.RGBEncoderConfig`) passed through PyAV.
You can set these parameters from the CLI with `--dataset.rgb_encoder.<field>` (e.g. with `lerobot-record` or `lerobot-rollout`). The same block applies to every camera video stream in that run.
> [!TIP]
> Video storage must be on for `rgb_encoder` to have any effect —
> `use_videos=True` in Python APIs, or `--dataset.video=true` on the CLI (the
> recording default). With video off, inputs stay as images and `rgb_encoder` is
> ignored.
For details on **when** frames are written vs. encoded (streaming vs. post-episode), queues, and other top-level `--dataset.*` switches, see [Streaming Video Encoding](./streaming_video_encoding). For an encoding-parameter comparison and experiments, see the [video-benchmark Space](https://huggingface.co/spaces/lerobot/video-benchmark).
---
## Example
```bash
lerobot-record \
--robot.type=so100_follower \
--robot.port=/dev/tty.usbmodem58760431541 \
--robot.cameras="{laptop: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
--robot.id=black \
--teleop.type=so100_leader \
--teleop.port=/dev/tty.usbmodem58760431551 \
--teleop.id=blue \
--dataset.repo_id=<my_username>/<my_dataset_name> \
--dataset.num_episodes=2 \
--dataset.single_task="Grab the cube" \
--dataset.streaming_encoding=true \
--dataset.encoder_threads=2 \
--dataset.rgb_encoder.vcodec=h264 \
--dataset.rgb_encoder.preset=fast \
--dataset.rgb_encoder.extra_options={"tune": "film", "profile:v": "high", "bf": 2} \
--display_data=true
```
---
## Tuning parameters
> [!WARNING]
> The defaults are tuned to balance **compression ratio**, **visual quality**, and **decoding/seek speed** for typical robotics datasets. Changing them can affect both recording (CPU load, frame drops) and training (decoding throughput, image quality).
>
> Only override these parameters if you have a specific reason to, and measure the impact on your pipeline before relying on the new settings.
All flags below are prefixed with `--dataset.rgb_encoder.` on the CLI.
| Parameter | Type | Default | Description |
| --------------- | ---------------- | ------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `vcodec` | `str` | `"libsvtav1"` | Video codec name. `"auto"` picks the first available hardware encoder from a fixed preference list, falling back to `libsvtav1`. |
| `pix_fmt` | `str` | `"yuv420p"` | Output pixel format. Must be supported by the chosen codec in your FFmpeg build. |
| `g` | `int` | `2` | GOP size — a keyframe every `g` frames. Emitted as FFmpeg option `g`. |
| `crf` | `int` or `float` | `30` | Abstract quality value, mapped per codec (see the [mapping](#mapping-videoencoderconfig--ffmpeg-options) below). Lower → higher quality / larger output where the mapping is monotone. |
| `preset` | `int` or `str` | `12` \* | Encoder speed preset; meaning depends on the codec. <br/>\* When unset and `vcodec=libsvtav1`, LeRobot defaults to `12`. |
| `fast_decode` | `int` | `0` | `libsvtav1`: `0–2`, passed via `svtav1-params`. <br/>`h264` / `hevc` (software): if `>0`, sets `tune=fastdecode`. <br/>Other codecs: usually unused. |
| `video_backend` | `str` | `"pyav"` | Only `"pyav"` is currently implemented for video encoding. |
| `extra_options` | `dict` | `{}` | Extra FFmpeg or codec specific options merged after the structured fields above. Cannot override keys already set by those fields. |
---
## Depth streams
Depth maps (Intel RealSense, Reachy 2) are stored as their **own video streams** alongside the RGB streams. Raw depth (`uint16` millimetres or `float32` metres) can't survive an 8-bit codec, so LeRobot **quantizes** each map to a 12-bit code (`[0, 4095]`) — logarithmically by default, to match the `1/depth` error profile of depth sensors — then packs it into a high-bit-depth pixel format (`gray12le`) and encodes it with a 12-bit codec.
<div style="margin:28px 0;padding:14px 0;">
<div style="margin:0 auto;display:flex;flex-wrap:wrap;justify-content:center;align-items:stretch;gap:6px;font-family:'Source Sans 3',ui-sans-serif,system-ui,sans-serif;font-size:14px;font-weight:600;color:#1B1B1D;">
<span style="display:flex;flex-direction:column;justify-content:center;align-items:center;text-align:center;gap:2px;background:#DBEAFE;color:#1D4ED8;border-radius:9px;padding:8px 12px;">
<span>Raw depth</span>
<span style="font-size:11px;font-weight:400;color:#3B6FD4;white-space:nowrap;">
uint16 mm
<br />
float32 m
</span>
</span>
<span style="display:flex;align-items:center;font-size:16px;color:#C3CBD9;">
→
</span>
<div style="border:2px dashed #C4B5FD;border-radius:13px;padding:18px 12px 12px;position:relative;display:flex;align-items:stretch;gap:6px;">
<span style="position:absolute;top:-10px;left:12px;background:#fff;padding:0 6px;font-size:11px;font-weight:700;color:#7E22CE;text-transform:uppercase;letter-spacing:0.5px;white-space:nowrap;">
Record time
</span>
<span style="display:flex;flex-direction:column;justify-content:center;align-items:center;text-align:center;gap:2px;background:#F3E8FF;color:#7E22CE;border-radius:9px;padding:8px 12px;">
<span>Clip</span>
<span style="font-size:11px;font-weight:400;color:#9061C2;white-space:nowrap;">
to [depth_min,
<br />
depth_max]
</span>
</span>
<span style="display:flex;align-items:center;font-size:16px;color:#C3CBD9;">
→
</span>
<span style="display:flex;flex-direction:column;justify-content:center;align-items:center;text-align:center;gap:2px;background:#F3E8FF;color:#7E22CE;border-radius:9px;padding:8px 12px;">
<span>Quantize</span>
<span style="font-size:11px;font-weight:400;color:#9061C2;white-space:nowrap;">
12-bit codes 0–4095
<br />
log (default) or linear
</span>
</span>
<span style="display:flex;align-items:center;font-size:16px;color:#C3CBD9;">
→
</span>
<span style="display:flex;flex-direction:column;justify-content:center;align-items:center;text-align:center;gap:2px;background:#F3E8FF;color:#7E22CE;border-radius:9px;padding:8px 12px;">
<span>Pack</span>
<span style="font-size:11px;font-weight:400;color:#9061C2;white-space:nowrap;">
into gray12le
<br />
plane
</span>
</span>
<span style="display:flex;align-items:center;font-size:16px;color:#C3CBD9;">
→
</span>
<span style="display:flex;flex-direction:column;justify-content:center;align-items:center;text-align:center;gap:2px;background:#F3E8FF;color:#7E22CE;border-radius:9px;padding:8px 12px;">
<span>Encode</span>
<span style="font-size:11px;font-weight:400;color:#9061C2;white-space:nowrap;">
HEVC
<br />
Main 12
</span>
</span>
</div>
<span style="display:flex;align-items:center;font-size:16px;color:#C3CBD9;">
→
</span>
<span style="display:flex;flex-direction:column;justify-content:center;align-items:center;text-align:center;gap:2px;background:#FEF3C7;color:#B45309;border-radius:9px;padding:8px 12px;">
<span>MP4</span>
<span style="font-size:11px;font-weight:400;color:#C77D18;white-space:nowrap;">
stored
<br />
stream
</span>
</span>
<span style="display:flex;align-items:center;font-size:16px;color:#34A06B;">
→
</span>
<div style="border:2px dashed #6EE7B7;border-radius:13px;padding:18px 12px 12px;position:relative;display:flex;align-items:center;gap:6px;">
<span style="position:absolute;top:-10px;left:12px;background:#fff;padding:0 6px;font-size:11px;font-weight:700;color:#047857;text-transform:uppercase;letter-spacing:0.5px;white-space:nowrap;">
Load time
</span>
<span style="display:flex;flex-direction:column;justify-content:center;align-items:center;text-align:center;gap:2px;background:#D1FAE5;color:#047857;border-radius:9px;padding:8px 12px;">
<span>Dequantize</span>
<span style="font-size:11px;font-weight:400;color:#059669;white-space:nowrap;">
to mm / m
</span>
</span>
</div>
</div>
</div>
Configure the depth pipeline through a parallel **`depth_encoder`** block (`DepthEncoderConfig`). It shares every `RGBEncoderConfig` field (`vcodec`, `pix_fmt`, `crf`, …) and adds four quantizer knobs, set via `--dataset.depth_encoder.<field>`:
```bash
lerobot-record \
... \
--dataset.depth_encoder.vcodec=hevc \
--dataset.depth_encoder.depth_min=0.05 \
--dataset.depth_encoder.depth_max=5.0 \
--dataset.depth_encoder.use_log=true
```
| Parameter | Type | Default | Description |
| --------------- | ------- | ------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------- |
| `vcodec` | `str` | `"hevc"` | HEVC Main 12 (a 12-bit-capable codec, MP4-compatible). |
| `extra_options` | `dict` | `{"x265-params": "lossless=1"}` | **Depth defaults to lossless** (exact round-trip); `crf` is ignored. Pass `extra_options={}` and set `crf` for a smaller lossy stream. |
| `pix_fmt` | `str` | `"gray12le"` | Single-channel 12-bit pixel format used to carry the quantized codes. |
| `depth_min` | `float` | `0.01` | Depth in metres mapped to quantum `0`. Values below are clipped on decode. |
| `depth_max` | `float` | `10.0` | Depth in metres mapped to quantum `4095`. Values above are clipped on decode. |
| `shift` | `float` | `3.5` | Pre-log offset (metres) used in logarithmic quantization for numerical stability near zero. Must satisfy `depth_min + shift > 0`. |
| `use_log` | `bool` | `True` | If `true`, quantize in log-space (recommended for typical depth sensors). Set to `false` for uniform/linear quantization. |
> [!TIP]
> `depth_min`, `depth_max`, and `shift` are always interpreted in **metres**, regardless of the input depth's unit. Inputs are auto-detected: integer arrays (e.g. `uint16` millimetres straight from a RealSense) are treated as millimetres, floating arrays as metres.
> Pick `depth_min` / `depth_max` to bracket the actual working range of your sensor — quanta outside that range saturate, which can crush detail at the boundaries.
Depth features are flagged with `"is_depth_map": true` in `meta/info.json`, and their quantizer settings (`video.depth_min`, `video.depth_max`, `video.shift`, `video.use_log`) are persisted — which is what lets depth be **dequantized back to physical units** on load.
### Output unit at load time
`depth_encoder` is a **record-time** concern. The unit that depth maps are dequantized to on _load_ (e.g. during training) is set separately by the read-time flag `--dataset.depth_output_unit`:
```bash
lerobot-train \
--dataset.repo_id=<my_username>/<my_dataset_name> \
--dataset.depth_output_unit=m \
--policy.type=act
```
| Parameter | Type | Default | Description |
| ------------------- | ----- | ------- | -------------------------------------------------------------------------------------------- |
| `depth_output_unit` | `str` | `"mm"` | Physical unit depth maps are dequantized to on load: `"mm"` (millimetres) or `"m"` (metres). |
> [!TIP]
> This is purely a decode-time presentation choice — it does **not** alter the stored video or its metadata, so the same dataset can be read as `mm` or `m` without re-encoding. It has no effect on datasets without depth cameras.
---
## Persistence in dataset metadata
After the first episode of a video stream is encoded, the encoder configuration is **persisted into the dataset metadata** (`meta/info.json`) under each video feature, alongside the values probed from the file itself. For a video feature `observation.images.<camera>`, the layout in `info.json` is:
```json
{
"features": {
"observation.images.laptop": {
"dtype": "video",
"shape": [480, 640, 3],
"info": {
"video.height": 480,
"video.width": 640,
"video.codec": "h264",
"video.pix_fmt": "yuv420p",
"video.fps": 30,
"video.channels": 3,
"is_depth_map": false,
"video.g": 2,
"video.crf": 30,
"video.preset": "fast",
"video.fast_decode": 0,
"video.video_backend": "pyav",
"video.extra_options": { "tune": "film", "profile:v": "high", "bf": 2 }
}
}
}
}
```
Two sources contribute to the `info` block:
| Source | Where it comes from | Fields |
| ------------------- | ----------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------- |
| **Stream-derived** | Read back from the encoded MP4 with PyAV. | `video.height`, `video.width`, `video.codec`, `video.pix_fmt`, `video.fps`, `video.channels`, `is_depth_map`, `audio.*` |
| **Encoder-derived** | Taken from `RGBEncoderConfig` / `DepthEncoderConfig`. | `video.g`, `video.crf`, `video.preset`, `video.fast_decode`, `video.video_backend`, `video.extra_options` |
> [!IMPORTANT]
> This block is populated **once**, from the **first** episode. It assumes every
> episode in the dataset was encoded with the same `rgb_encoder`. Changing
> encoder settings partway through a recording is not supported — the
> `info.json` will only reflect the parameters used for the first episode.
---
## Merging datasets
When aggregating datasets with `merge_datasets`, video files are concatenated as-is (no re-encoding), and encoder fields in `info.json` are merged per-key:
| Merge rule | Fields | Behaviour |
| ------------------ | ---------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Must match** | `video.codec`, `video.pix_fmt`, `video.height`, `video.width`, `video.fps` | Stream-derived fields must match across sources, otherwise FFmpeg's concat demuxer fails. |
| **Merged loosely** | `video.g`, `video.crf`, `video.preset`, `video.fast_decode`, `video.extra_options` | Encoder-tuning fields. If every source agrees, the value is kept; if not, it's set to `null` (or `{}` for `video.extra_options`) and a warning is logged. |
docs/source/vla_jepa.mdx
+235
View File
@@ -0,0 +1,235 @@
# VLA-JEPA
This is the LeRobot port of **VLA-JEPA**, a Vision-Language-Action model that combines a Qwen3-VL language backbone with a self-supervised video world model (V-JEPA2) and a flow-matching DiT action head.
---
## Architecture Overview
VLA-JEPA has three main components:
| Component | Module | Role |
| ----------------------- | --------------------------------- | ------------------------------------------------------- |
| **Qwen3-VL backbone** | `Qwen3VLInterface` | Fuses images + language instruction into context tokens |
| **DiT-B action head** | `VLAJEPAActionHead` | Flow-matching diffusion over the action chunk |
| **V-JEPA2 world model** | `ActionConditionedVideoPredictor` | Self-supervised video prediction loss (training only) |
### Data flow
**Training:**
1. A video clip of `num_video_frames` frames is encoded by V-JEPA2 into per-frame patch tokens.
2. The Qwen3-VL backbone processes multi-view images + the task instruction and produces a sequence of context tokens that includes special action tokens (for world model conditioning) and embodied tokens.
3. The action head receives those context tokens as cross-attention keys/values and predicts a denoised action chunk via flow matching.
4. The world model predictor uses the action tokens extracted from Qwen to predict future V-JEPA2 frame embeddings; a regression loss on those predictions is added to the action loss.
**Inference:**
Only Qwen + the action head are used. The world model is not needed at inference time.
### Action head details
Available presets via `action_model_type`:
| Preset | Hidden dim | Heads | Head dim |
| ------- | ---------- | ----- | -------- |
| `DiT-B` | 768 | 12 | 64 |
| `DiT-L` | 1536 | 32 | 48 |
### World model details
The video predictor is a ViT-style transformer (`ActionConditionedVideoPredictor`) that takes:
- **Frame tokens**: V-JEPA2 patch embeddings projected to `predictor_embed_dim`
- **Action tokens**: Qwen action token embeddings projected to `predictor_embed_dim`
It uses block-causal attention so each temporal step can attend to all previous steps. The predictor's input `embed_dim` equals `num_views × video_encoder_hidden_size` (e.g. 2 views × 1024 = 2048 for the pretrained checkpoints).
---
## Pretrained Checkpoints
Three checkpoints are available directly inside the LeRobot org here: [`lerobot/VLA-JEPA`](https://huggingface.co/collections/lerobot/vla-jepa), converted from [ginwind/VLA-JEPA](https://huggingface.co/ginwind/VLA-JEPA):
| Checkpoint | Dataset | Cameras | World model | Action dim |
| ----------------------------- | ----------------- | ----------------------- | ----------- | ---------- |
| `lerobot/VLA-JEPA-LIBERO` | LIBERO-10 | 2 (agentview + wrist) | Enabled | 7 |
| `lerobot/VLA-JEPA-Pretrain` | DROID 1.0.1 | 2 (exterior left views) | Enabled | 7 |
| `lerobot/VLA-JEPA-SimplerEnv` | OXE Bridge / RT-1 | 1 (view duplicated ×2) | Enabled | 7 |
All checkpoints use `Qwen/Qwen3-VL-2B-Instruct` as the language backbone.
---
## Configuration
Key parameters in `VLAJEPAConfig`:
| Parameter | Default | Description |
| ------------------------- | ------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `chunk_size` | 7 | Number of actions predicted per inference call |
| `n_action_steps` | 7 | Steps executed from the predicted chunk before re-planning |
| `num_video_frames` | 8 | Video clip length fed to the world model |
| `enable_world_model` | `True` | Whether to load and train the V-JEPA2 predictor |
| `world_model_loss_weight` | 0.1 | Weight of the JEPA prediction loss relative to the action loss |
| `num_inference_timesteps` | 4 | Euler integration steps for action denoising |
| `freeze_qwen` | `False` | Freeze the Qwen3-VL backbone and only train the action head |
| `reinit_modules` | `None` | Key prefixes allowed to be randomly re-initialised on load (for cross-embodiment transfer, see [Fine-tuning on a different embodiment](#fine-tuning-on-a-different-embodiment)) |
| `gripper_dim` | 6 | Index of the gripper dimension in the action vector (e.g. 6 for a 7-DoF arm with gripper as the last joint) |
| `gripper_threshold` | 0.5 | Threshold used by `pre_snap_gripper_action` and `binarize_gripper_action` to binarize the gripper dimension |
| `pre_snap_gripper_action` | `True` | Snap the gripper dim to {0, 1} before unnormalization. Set to `False` for robots without a binary gripper |
| `binarize_gripper_action` | `True` | Binarize the gripper dim to {-1, 1} after unnormalization. Set to `False` for robots without a binary gripper |
---
## Training
Number of training steps may vary based on dataset size and compute budget. The original paper pretrained for 50k on ssv2 + droid jointly, then additional 30k steps for LIBERO, but fewer steps may still yield good performance when fine-tuning from the provided pretrained checkpoints.
### Full training from scratch
```bash
lerobot-train \
policy.type=vla_jepa \
policy.repo_id=your_org/your_repo \
dataset.repo_id=your_org/your_dataset
```
### Fine-tuning from a pretrained checkpoint
```bash
lerobot-train \
--policy.path=lerobot/VLA-JEPA-Pretrain \
--policy.repo_id=your_org/your_repo \
--dataset.repo_id=your_org/your_dataset
```
If you want to freeze the Qwen backbone and only train the action head, set `policy.freeze_qwen=True`:
```bash
lerobot-train \
--policy.path=lerobot/VLA-JEPA-Pretrain \
--policy.repo_id=your_org/your_repo \
--policy.freeze_qwen=true \
--dataset.repo_id=your_org/your_dataset
```
### Fine-tuning on a different embodiment
When the target robot has a different action or state dimensionality than the pretrained checkpoint, the input/output projection layers of the action head will have mismatched shapes and cannot be loaded directly. `reinit_modules` lets you list the key prefixes that are allowed to mismatch — those layers are randomly re-initialised while every other weight is reused from the checkpoint. Any shape mismatch outside the listed prefixes raises an error.
The layers that depend on `action_dim` and `state_dim` are:
| Layer | Key prefix |
| ----------------------------------------- | ----------------------------------- |
| Action encoder (action_dim → inner_dim) | `model.action_model.action_encoder` |
| Action decoder (hidden_size → action_dim) | `model.action_model.action_decoder` |
| State encoder (state_dim → inner_dim) | `model.action_model.state_encoder` |
```bash
lerobot-train \
--policy.path=lerobot/VLA-JEPA-Pretrain \
--policy.repo_id=your_org/your_repo \
--policy.freeze_qwen=true \
--policy.reinit_modules='["model.action_model.action_encoder", "model.action_model.action_decoder", "model.action_model.state_encoder"]' \
--dataset.repo_id=your_org/your_dataset
```
If your robot has no proprioceptive state, omit `model.action_model.state_encoder` from the list.
### Reproducing the LIBERO results
**Training on LIBERO:**
starts the training from the Pretrain checkpoint, trains for 30k steps on the LIBERO dataset.
Original paper mentions training across 8 GPUs with a batch size of 32, meaning global batch size of 256.
```bash
lerobot-train \
--policy.path=lerobot/VLA-JEPA-Pretrain \
--policy.repo_id=your_org/your_repo \
--dataset.repo_id=HuggingFaceVLA/libero \
--steps=30000
```
**Evaluating the pretrained LIBERO-10 checkpoint:**
```bash
lerobot-eval \
--policy.path=lerobot/VLA-JEPA-LIBERO \
--env.type=libero \
--env.task=libero_spatial,libero_object,libero_goal,libero_10 \
--eval.n_episodes=10 \
--eval.batch_size=5
```
To evaluate a subset of tasks only:
```bash
lerobot-eval \
--policy.path=lerobot/VLA-JEPA-LIBERO \
--env.type=libero \
--env.task=libero_10 \
--env.task_ids='[0,1,2]' \
--eval.n_episodes=10 \
--eval.batch_size=5
```
**Expected results:**
| Suite | Episodes | Successes | Success Rate |
| -------------- | -------- | --------- | ------------ |
| libero_spatial | 100 | 93 | **95.0%** |
| libero_object | 100 | 100 | **100.0%** |
| libero_goal | 100 | 98 | **98.0%** |
| libero_10 | 100 | 96 | **93.0%** |
| **Overall** | **400** | **387** | **96.5%** |
---
## Fine-tuning on datasets with a different number of cameras
The pretrained world model predictor was trained with `embed_dim = jepa_tubelet_size × 1024` (default `jepa_tubelet_size=2`).
**Default behaviour — view padding / trimming (no action required)**
When fine-tuning from `VLA-JEPA-Pretrain` the model automatically adjusts the number of views fed to the world model to match `jepa_tubelet_size`:
- **Single-view datasets (e.g. BridgeV2):** the single-view latent is duplicated to produce a two-view world-model input, preserving the JEPA self-supervised signal without any weight mismatch.
- **>2-view datasets (e.g. DROID with 3 views):** all views are passed to the Qwen backbone (for richer context), but only the first `jepa_tubelet_size` views (one wrist + one third-person, following the configured view order) are used for the world model.
**Option 1 — Disable the world model**
Set `enable_world_model=False` to skip the JEPA loss entirely. Only the Qwen backbone and action head are loaded and trained. This is sufficient for good action performance.
```bash
lerobot-train \
--policy.path=lerobot/VLA-JEPA-Pretrain \
--policy.enable_world_model=false \
--policy.repo_id=your_org/your_repo \
--dataset.repo_id=your_org/single_camera_dataset
```
**Option 2 — Reinitialize the predictor input projection**
If you want to change `jepa_tubelet_size` to a value other than 2, load the checkpoint with `strict=False` and reinitialize `model.video_predictor.predictor_embed` for the new `embed_dim`. All other predictor block weights (attention, MLP, norm, output projection) are camera-count-agnostic and can be reused from the pretrained checkpoint.
---
## Citation
```bibtex
@misc{sun2026vlajepaenhancingvisionlanguageactionmodel,
title = {VLA-JEPA: Enhancing Vision-Language-Action Model with Latent World Model},
author = {Jingwen Sun and Wenyao Zhang and Zekun Qi and Shaojie Ren and Zezhi Liu and Hanxin Zhu and Guangzhong Sun and Xin Jin and Zhibo Chen},
year = {2026},
eprint = {2602.10098},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2602.10098},
}
```
---
## License
Weights are distributed under the license terms of the original [ginwind/VLA-JEPA](https://huggingface.co/ginwind/VLA-JEPA) repository (**Apache 2.0 License**). The LeRobot integration code follows the **Apache 2.0 License**.
docs/source/vlabench.mdx
+176
View File
@@ -0,0 +1,176 @@
# VLABench
[VLABench](https://github.com/OpenMOSS/VLABench) is a large-scale benchmark for **language-conditioned robotic manipulation with long-horizon reasoning**. The upstream suite covers 100 task categories across 2,000+ objects and evaluates six dimensions of robot intelligence: mesh & texture understanding, spatial reasoning, world-knowledge transfer, semantic instruction comprehension, physical-law understanding, and long-horizon planning. Built on MuJoCo / dm_control with a Franka Panda 7-DOF arm. LeRobot exposes **43 of these tasks** through `--env.task` (21 primitives + 22 composites, see [Available tasks](#available-tasks) below).
- Paper: [VLABench: A Large-Scale Benchmark for Language-Conditioned Robotics Manipulation with Long-Horizon Reasoning](https://arxiv.org/abs/2412.18194)
- GitHub: [OpenMOSS/VLABench](https://github.com/OpenMOSS/VLABench)
- Project website: [vlabench.github.io](https://vlabench.github.io)
- Pretrained policy: [`lerobot/smolvla_vlabench`](https://huggingface.co/lerobot/smolvla_vlabench)
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/vlabench.png"
alt="VLABench benchmark overview"
width="85%"
/>
## Available tasks
VLABench ships two task suites covering **43 task categories** in LeRobot's `--env.task` surface:
| Suite | CLI name | Tasks | Description |
| --------- | ----------- | ----- | ---------------------------------------------------------------- |
| Primitive | `primitive` | 21 | Single / few-skill combinations (select, insert, physics QA) |
| Composite | `composite` | 22 | Multi-step reasoning and long-horizon planning (cook, rearrange) |
**Primitive tasks:** `select_fruit`, `select_toy`, `select_chemistry_tube`, `add_condiment`, `select_book`, `select_painting`, `select_drink`, `insert_flower`, `select_billiards`, `select_ingredient`, `select_mahjong`, `select_poker`, and physical-reasoning tasks (`density_qa`, `friction_qa`, `magnetism_qa`, `reflection_qa`, `simple_cuestick_usage`, `simple_seesaw_usage`, `sound_speed_qa`, `thermal_expansion_qa`, `weight_qa`).
**Composite tasks:** `cluster_billiards`, `cluster_book`, `cluster_drink`, `cluster_toy`, `cook_dishes`, `cool_drink`, `find_unseen_object`, `get_coffee`, `hammer_nail`, `heat_food`, `make_juice`, `play_mahjong`, `play_math_game`, `play_poker`, `play_snooker`, `rearrange_book`, `rearrange_chemistry_tube`, `set_dining_table`, `set_study_table`, `store_food`, `take_chemistry_experiment`, `use_seesaw_complex`.
`--env.task` accepts three forms:
- a single task name (`select_fruit`)
- a comma-separated list (`select_fruit,heat_food`)
- a suite shortcut (`primitive`, `composite`, or `primitive,composite`)
## Installation
VLABench is **not on PyPI** — its only distribution is the [OpenMOSS/VLABench](https://github.com/OpenMOSS/VLABench) GitHub repo — so LeRobot does not expose a `vlabench` extra. Install it manually as an editable clone, alongside the MuJoCo / dm_control pins VLABench needs, then fetch the mesh assets:
```bash
# After following the standard LeRobot installation instructions.
git clone https://github.com/OpenMOSS/VLABench.git ~/VLABench
git clone https://github.com/motion-planning/rrt-algorithms.git ~/rrt-algorithms
pip install -e ~/VLABench -e ~/rrt-algorithms
pip install "mujoco==3.2.2" "dm-control==1.0.22" \
open3d colorlog scikit-learn openai gdown
python ~/VLABench/scripts/download_assets.py
```
<Tip>
VLABench requires Linux (`sys_platform == 'linux'`) and Python 3.10+. Set the MuJoCo rendering backend before running:
```bash
export MUJOCO_GL=egl # for headless servers (HPC, cloud)
```
</Tip>
## Evaluation
All eval snippets below mirror the command CI runs (see `.github/workflows/benchmark_tests.yml`). The `--rename_map` argument maps VLABench's `image` / `second_image` / `wrist_image` camera keys onto the three-camera (`camera1` / `camera2` / `camera3`) input layout the released `smolvla_vlabench` policy was trained on.
### Single-task evaluation (recommended for quick iteration)
```bash
lerobot-eval \
--policy.path=lerobot/smolvla_vlabench \
--env.type=vlabench \
--env.task=select_fruit \
--eval.batch_size=1 \
--eval.n_episodes=10 \
--eval.use_async_envs=false \
--policy.device=cuda \
'--rename_map={"observation.images.image": "observation.images.camera1", "observation.images.second_image": "observation.images.camera2", "observation.images.wrist_image": "observation.images.camera3"}'
```
### Multi-task evaluation
Pass a comma-separated list of tasks:
```bash
lerobot-eval \
--policy.path=lerobot/smolvla_vlabench \
--env.type=vlabench \
--env.task=select_fruit,select_toy,add_condiment,heat_food \
--eval.batch_size=1 \
--eval.n_episodes=10 \
--eval.use_async_envs=false \
--policy.device=cuda \
'--rename_map={"observation.images.image": "observation.images.camera1", "observation.images.second_image": "observation.images.camera2", "observation.images.wrist_image": "observation.images.camera3"}'
```
### Suite-wide evaluation
Run an entire suite (all 21 primitives or all 22 composites):
```bash
lerobot-eval \
--policy.path=lerobot/smolvla_vlabench \
--env.type=vlabench \
--env.task=primitive \
--eval.batch_size=1 \
--eval.n_episodes=10 \
--eval.use_async_envs=false \
--policy.device=cuda \
--env.max_parallel_tasks=1 \
'--rename_map={"observation.images.image": "observation.images.camera1", "observation.images.second_image": "observation.images.camera2", "observation.images.wrist_image": "observation.images.camera3"}'
```
Or both suites:
```bash
lerobot-eval \
--policy.path=lerobot/smolvla_vlabench \
--env.type=vlabench \
--env.task=primitive,composite \
--eval.batch_size=1 \
--eval.n_episodes=10 \
--eval.use_async_envs=false \
--policy.device=cuda \
--env.max_parallel_tasks=1 \
'--rename_map={"observation.images.image": "observation.images.camera1", "observation.images.second_image": "observation.images.camera2", "observation.images.wrist_image": "observation.images.camera3"}'
```
### Recommended evaluation episodes
**10 episodes per task** for reproducible benchmarking (210 total for the full primitive suite, 220 for composite). Matches the protocol in the VLABench paper.
## Policy inputs and outputs
**Observations:**
- `observation.state` — 7-dim end-effector state (position xyz + Euler xyz + gripper)
- `observation.images.image` — front camera, 480×480 HWC uint8
- `observation.images.second_image` — second camera, 480×480 HWC uint8
- `observation.images.wrist_image` — wrist camera, 480×480 HWC uint8
**Actions:**
- Continuous control in `Box(-1, 1, shape=(7,))` — 3D position + 3D Euler orientation + 1D gripper.
## Training
### Datasets
Pre-collected VLABench datasets in LeRobot format on the Hub:
- [`VLABench/vlabench_primitive_ft_lerobot_video`](https://huggingface.co/datasets/VLABench/vlabench_primitive_ft_lerobot_video) — 5,000 episodes, 128 tasks, 480×480 images.
- [`VLABench/vlabench_composite_ft_lerobot_video`](https://huggingface.co/datasets/VLABench/vlabench_composite_ft_lerobot_video) — 5,977 episodes, 167 tasks, 224×224 images.
### Example training command
Fine-tune a SmolVLA base on the primitive suite:
```bash
lerobot-train \
--policy.type=smolvla \
--policy.repo_id=${HF_USER}/smolvla_vlabench_primitive \
--policy.load_vlm_weights=true \
--policy.push_to_hub=true \
--dataset.repo_id=VLABench/vlabench_primitive_ft_lerobot_video \
--env.type=vlabench \
--env.task=select_fruit \
--output_dir=./outputs/smolvla_vlabench_primitive \
--steps=100000 \
--batch_size=4 \
--env_eval_freq=5000 \
--eval.batch_size=1 \
--eval.n_episodes=1 \
--save_freq=10000
```
## Reproducing published results
The released checkpoint [`lerobot/smolvla_vlabench`](https://huggingface.co/lerobot/smolvla_vlabench) was trained on the primitive-suite dataset above and is evaluated with the [Single-task](#single-task-evaluation-recommended-for-quick-iteration) / [Suite-wide](#suite-wide-evaluation) commands. CI runs a 10-primitive-task smoke eval (one episode each) on every PR touching the benchmark.
docs/source/walloss.mdx
+80
View File
@@ -0,0 +1,80 @@
# WALL-OSS
WALL-OSS is an open-source foundation model for embodied intelligence, proposed by the [XSquare Robot](https://x2robot.com/en/research/68bc2cde8497d7f238dde690) team in 2025. The LeRobot implementation is adapted from their open-source [WallX](https://github.com/X-Square-Robot/wall-x) repository.
X Square Robot’s WALL-OSS is now integrated into Hugging Face’s LeRobot ecosystem. This is an exciting collaborative project between the LeRobot and X Square Robot teams. You can now post-train, evaluate, and deploy WALL-OSS directly through LeRobot. With this, we’re aiming to make it easier for the open-source robotics community to customize and deploy WALL-OSS foundation models. Read and explore WALL-OSS [paper](https://arxiv.org/pdf/2509.11766) and [code](https://github.com/X-Square-Robot/wall-x).
## Model Overview
The WALL-OSS team is building the embodied foundation model to capture and compress the world's most valuable data: the continuous, high-fidelity stream of physical interaction. By creating a direct feedback loop between the model's decisions and the body's lived experience, the emergence of a truly generalizable intelligence is enabled—one that understands not just how the world works, but how to act effectively within it.
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/walloss-lerobot-paper.png"
alt="An overview of WALL-OSS"
width="85%"
/>
Technically, WALL-OSS introduces a tightly coupled multimodal architecture (tightly-coupled MoE structure) that integrates both discrete and continuous action modeling strategies. Through a two-stage training pipeline (Inspiration → Integration), the model gradually unifies semantic reasoning and high-frequency action generation. Its core innovations include:
- **Embodied perception–enhanced multimodal pretraining**: Large-scale training on unified vision–language–action data to strengthen spatial, causal, and manipulation understanding.
- **Unified Cross-Level Chain-of-Thought (Uni-CoT)**: A single differentiable framework that unifies high-level instruction reasoning, sub-task decomposition, and fine-grained action synthesis, forming a continuous chain from “understanding” to “execution.”
- **Mixture-of-Experts (MoE) action heads**: Dynamically activating experts depending on the task phase and modeling actions in discrete or continuous space to maintain stable VLM priors.
- **Two-stage training paradigm**:
- **Inspiration stage**: Injecting discrete action priors to strengthen spatial understanding and semantic-action alignment.
- **Integration stage**: Using flow matching to achieve high-frequency continuous control.
## Installation Requirements
1. Install LeRobot by following our [Installation Guide](./installation).
2. Install WallX dependencies by running:
```bash
pip install -e ".[wallx]"
```
## Usage
To use WallX in LeRobot, specify the policy type as:
```python
policy.type=wall_x
```
## Training
For training WallX, you can use the standard LeRobot training script with the appropriate configuration:
```bash
lerobot-train \
--dataset.repo_id=your_dataset \
--policy.type=wall_x \
--output_dir=./outputs/wallx_training \
--job_name=wallx_training \
--policy.repo_id=your_repo_id \
--policy.pretrained_name_or_path=x-square-robot/wall-oss-flow \
--policy.prediction_mode=diffusion \
--policy.attn_implementation=eager \
--steps=3000 \
--policy.device=cuda \
--batch_size=32
```
### Training Arguments
| Argument | Description |
| ------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `--dataset.repo_id` | The Hugging Face Hub repository ID for your training dataset (e.g., `lerobot/aloha_sim_insertion_human`) |
| `--policy.type` | Specifies using the WallX policy architecture |
| `--output_dir` | Local directory where training checkpoints and logs will be saved |
| `--job_name` | A name identifier for this training run (used in logging/tracking) |
| `--policy.repo_id` | Your Hugging Face Hub repo ID where the trained model will be pushed |
| `--policy.pretrained_path` | Path to pretrained WallX weights to initialize from (the official WALL-OSS checkpoint) |
| `--policy.prediction_mode` | The action prediction strategy: `diffusion` or `fast` - `diffusion` uses iterative denoising for action generation, `fast` uses next token prediction instead |
| `--policy.attn_implementation` | Attention implementation backend - `eager` uses standard PyTorch attention (alternatives include `flash_attention_2` or `sdpa`) |
| `--steps` | Total number of training steps to run |
| `--policy.device` | Device to train on (`cuda` for GPU, `cpu` for CPU) |
| `--batch_size` | Number of samples per training batch |
## License
This model follows the **Apache 2.0 License**, consistent with the original [WallX repository](https://github.com/X-Square-Robot/wall-x).
docs/source/xvla.mdx
+528
View File
@@ -0,0 +1,528 @@
# X-VLA: The First Soft-Prompted Robot Foundation Model for Any Robot, Any Task
## Overview
For years, robotics has aspired to build agents that can follow natural human instructions and operate dexterously across many environments and robot bodies. Recent breakthroughs in LLMs and VLMs suggest a path forward: extend these foundation-model architectures to embodied control by grounding them in actions. This has led to the rise of Vision-Language-Action (VLA) models, with the hope that a single generalist model could combine broad semantic understanding with robust manipulation skills.
But training such models is difficult. Robot data is fragmented across platforms, sensors, embodiments, and collection protocols. Heterogeneity appears everywhere: different arm configurations, different action spaces, different camera setups, different visual domains, and different task distributions. These inconsistencies create major distribution shifts that make pretraining unstable and adaptation unreliable.
Inspired by meta-learning and prompt learning, we ask: **"What if a VLA model could learn the structure of each robot and dataset the same way LLMs learn tasks, through prompts?"**
**X-VLA** is a soft-prompted, flow-matching VLA framework that treats each hardware setup as a "task" and encodes it using a small set of learnable embeddings. These **Soft Prompts** capture embodiment and domain-specific variations, guiding the Transformer from the earliest stages of multimodal fusion. With this mechanism, X-VLA can reconcile diverse robot morphologies, data types, and sensor setups within a single unified architecture.
<p align="center">
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/xvla-architecture.png"
alt="XVLA Architecture"
style="max-width: 100%; height: auto; width: 800px;"
/>
</p>
Built from pure Transformer encoders, X-VLA scales naturally with model size and dataset diversity. Across 6 simulation benchmarks and 3 real robots, Soft Prompts consistently outperform existing methods in handling hardware and domain differences. X-VLA-0.9B, trained on 290K episodes spanning seven robotic platforms, learns an embodiment-agnostic generalist policy in Phase I, and adapts efficiently to new robots in Phase II simply by learning a new set of prompts, while keeping the backbone frozen.
<p align="center">
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/xvla-architecture2.png"
alt="XVLA Architecture 2"
style="width: 60%; height: auto;"
/>
</p>
With only 1% of parameters tuned (9M), X-VLA-0.9B achieves near-π₀ performance on LIBERO and Simpler-WidowX, despite using **300× fewer trainable parameters**. It also demonstrates strong real-world dexterity with minimal demonstrations, including folding cloths in under two minutes.
<p align="center">
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/xvla-fold.png"
alt="XVLA fold visualization"
style="width: 95%; max-width: 1100px; height: auto;"
/>
</p>
X-VLA shows that generalist robot intelligence does not require increasingly complex architectures, only the right way to absorb heterogeneity. Soft Prompts offer a simple, scalable mechanism for unifying diverse robotic data, paving the way toward adaptable, cross-embodiment robot foundation models.
## Installation
After installing LeRobot, install the X-VLA dependencies:
```bash
pip install -e .[xvla]
```
After the new release, you'll be able to do:
```bash
pip install lerobot[xvla]
```
## Quick Start
### Basic Usage
To use X-VLA in your LeRobot configuration, specify the policy type as:
```bash
policy.type=xvla
```
### Evaluating Pre-trained Checkpoints
Example evaluation with LIBERO:
```bash
lerobot-eval \
--policy.path="lerobot/xvla-libero" \
--env.type=libero \
--env.task=libero_spatial,libero_goal,libero_10 \
--env.control_mode=absolute \
--eval.batch_size=1 \
--eval.n_episodes=1 \
--env.episode_length=800 \
--seed=142
```
## Available Checkpoints
### 🎯 Base Model
**[lerobot/xvla-base](https://huggingface.co/lerobot/xvla-base)**
A 0.9B parameter instantiation of X-VLA, trained with a carefully designed data processing and learning recipe. The training pipeline consists of two phases:
- **Phase I: Pretraining** - Pretrained on 290K episodes from Droid, Robomind, and Agibot, spanning seven platforms across five types of robotic arms (single-arm to bi-manual setups). By leveraging soft prompts to absorb embodiment-specific variations, the model learns an embodiment-agnostic generalist policy.
- **Phase II: Domain Adaptation** - Adapted to deployable policies for target domains. A new set of soft prompts is introduced and optimized to encode the hardware configuration of the novel domain, while the pretrained backbone remains frozen.
### Simulation Checkpoints
**[lerobot/xvla-libero](https://huggingface.co/lerobot/xvla-libero)**
Achieves 93% success rate on LIBERO benchmarks. Fine-tuned from the base model for simulation tasks.
**[lerobot/xvla-widowx](https://huggingface.co/lerobot/xvla-widowx)**
Fine-tuned on BridgeData for pick-and-place experiments on compact WidowX platforms. Demonstrates robust manipulation capabilities.
### 🤖 Real-World Checkpoints
**[lerobot/xvla-folding](https://huggingface.co/lerobot/xvla-folding)**
A fine-tuned dexterous manipulation model trained on the high-quality Soft-FOLD cloth folding dataset. Achieves 100% success rate over 2 hours of continuous cloth folding.
**[lerobot/xvla-agibot-world](https://huggingface.co/lerobot/xvla-agibot-world)**
Optimized for AgileX robot dexterous manipulation tasks.
**[lerobot/xvla-google-robot](https://huggingface.co/lerobot/xvla-google-robot)**
Adapted for Google Robot platforms.
## Training X-VLA
### Recommended Training Configuration
When fine-tuning X-VLA for a new embodiment or task, we recommend not freezing the VLM, and also setting the `policy.dtype=bfloat16` to not hit OOM errors.
```bash
lerobot-train \
--dataset.repo_id=YOUR_DATASET \
--output_dir=./outputs/xvla_training \
--job_name=xvla_training \
--policy.path="lerobot/xvla-base" \
--policy.repo_id="HF_USER/xvla-your-robot" \
--policy.dtype=bfloat16 \
--policy.action_mode=auto \
--steps=20000 \
--policy.device=cuda \
--policy.freeze_vision_encoder=false \
--policy.freeze_language_encoder=false \
--policy.train_policy_transformer=true \
--policy.train_soft_prompts=true \
```
### Training Parameters Explained
| Parameter | Default | Description |
| -------------------------- | ------- | ---------------------------------------------- |
| `freeze_vision_encoder` | `false` | Do not freeze the VLM vision encoder weights |
| `freeze_language_encoder` | `false` | Do not freeze the VLM language encoder weights |
| `train_policy_transformer` | `true` | Allow policy transformer layers to train |
| `train_soft_prompts` | `true` | Allow soft prompts to train |
**💡 Best Practice**: For Phase II adaptation to new embodiments, do not freeze the VLM encoders and also train the policy transformer and soft prompts.
### Example: Training on Bimanual Robot
```bash
lerobot-train \
--dataset.repo_id=<USER>/bimanual-so100-handover-cube \
--output_dir=./outputs/xvla_bimanual \
--job_name=xvla_so101_training \
--policy.path="lerobot/xvla-base" \
--policy.dtype=bfloat16 \
--policy.repo_id="YOUR_USERNAME/xvla-biso101" \
--steps=3000 \
--policy.device=cuda \
--policy.action_mode=so101_bimanual \
--policy.freeze_vision_encoder=false \
--policy.freeze_language_encoder=false \
--policy.train_policy_transformer=true \
--policy.train_soft_prompts=true
```
💡 **Best Performance:** If you have sufficient computational resources and want to achieve best X-VLA finetuning performance, you should follow the official finetuning strategy:
**🔥 Full-finetune all components with a custom learning-rate scheme**
To ensure stable optimization, the Vision-Language Model (VLM) must be trained with only 1/10 of the base learning rate, while all other components use the full LR.
This LR ratio is crucial for achieving strong and stable finetuning performance. This is already done for you by default.
❕Note
Completely matching the official reported performance may require an additional warm-up LR schedule for soft-prompts, which can bring minor improvements.
We encourage implementing this in your customized training pipeline for optimal results.
## Core Concepts
### 1. Action Modes
X-VLA uses an **Action Registry** system to handle different action spaces and embodiments. The `action_mode` parameter defines how actions are processed, what loss functions are used, and how predictions are post-processed.
#### Available Action Modes
| Action Mode | Action Dim | Description | Use Case |
| ---------------- | ----------------------- | ------------------------------------------- | ------------------------------------ |
| `ee6d` | 20 | End-effector with xyz, 6D rotation, gripper | Dual-arm setups with spatial control |
| `joint` | 14 | Joint-space with gripper | Direct joint control robots |
| `agibot_ee6d` | 20 | AGI-bot variant with MSE loss | AGI-bot platforms |
| `so101_bimanual` | 20 (model), 12 (real) | SO101 bimanual robot | Bimanual manipulation tasks |
| `auto` | 20 (model), auto (real) | Auto-detects action dim from dataset | **Recommended** for new robots |
#### Why Action Modes Matter
When you have a pretrained checkpoint like `lerobot/xvla-base` trained with `action_dim=20`, and you want to train on a dataset with a different action dimension (e.g., 14 for bimanual arms), you can't simply trim the action dimension. The action mode orchestrates:
1. **Loss Computation**: Different loss functions for different action components (MSE for joints, BCE for grippers, etc.)
2. **Preprocessing**: Zeroing out gripper channels, padding dimensions
3. **Postprocessing**: Applying sigmoid to gripper logits, trimming padding
#### Example: BimanualSO101 Action Space
The `so101_bimanual` action mode handles the mismatch between model output (20D) and real robot control (12D):
```python
# Model outputs 20 dimensions for compatibility
dim_action = 20
# Real robot only needs 12 dimensions
# [left_arm (6), right_arm (6)] = [joints (5) + gripper (1)] × 2
REAL_DIM = 12
# Preprocessing: Pad 12D actions to 20D for training
# Postprocessing: Trim 20D predictions to 12D for deployment
```
See the [action_hub.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/xvla/action_hub.py) implementation for details.
#### Auto Action Mode (Recommended)
The `auto` action mode is the easiest way to use X-VLA with any robot. It automatically detects your dataset's action dimension and handles padding/trimming:
```bash
lerobot-train \
--policy.path="lerobot/xvla-base" \
--policy.action_mode=auto \
--policy.max_action_dim=20 \
...
```
**How it works:**
- Reads `action_feature.shape[-1]` from your dataset (e.g., 7 for Franka)
- Model outputs `max_action_dim` (default 20) for pretrained compatibility
- Loss is computed **only on the real dimensions**: `MSE(pred[:,:,:real_dim], target[:,:,:real_dim])`
- Postprocess trims output back to `real_dim` for robot control
This eliminates the need to create custom action modes for most robots.
### 2. Domain IDs
Domain IDs are learnable identifiers for different robot configurations and camera setups. They allow X-VLA to distinguish between:
- Different robots (Robot 1 vs Robot 2)
- Different camera configurations (cam1 vs cam2)
- Different combinations (Robot1-cam1-cam2 vs Robot1-cam1 vs Robot2-cam1)
#### Setting Domain IDs
**During Training**: By default, domain_id is set to 0 for general training.
**During Evaluation**: Specify the domain_id that matches your checkpoint's training configuration.
```python
# Example: LIBERO checkpoint uses domain_id=3
domain_id = 3
```
The domain_id is automatically added to observations by the `XVLAAddDomainIdProcessorStep` in the preprocessing pipeline.
The `lerobot/xvla-base` model has been trained on the following domain IDs. It is recommended to choose one that most resembles your robot/configuration:
#### Fine-tuning Datasets
| Dataset Name | Domain ID |
| ---------------- | --------- |
| Bridge | 0 |
| RT1 | 1 |
| Calvin | 2 |
| libero | 3 |
| widowx-air | 4 |
| AIR-AGILEX-HQ | 5 |
| robotwin2_abs_ee | 6 |
| robotwin2_clean | 6 |
| robocasa-human | 7 |
| VLABench | 8 |
| AGIBOT-challenge | 9 |
| AIR-AGILEX | 10 |
| AIRBOT | 18 |
### 3. Processor Steps
X-VLA requires specific preprocessing and postprocessing steps for proper operation.
#### Required Preprocessing Steps
1. **XVLAImageToFloatProcessorStep**: Converts images from [0, 255] to [0, 1] range
2. **XVLAImageNetNormalizeProcessorStep**: Applies ImageNet normalization (required for VLM backbone)
3. **XVLAAddDomainIdProcessorStep**: Adds domain_id to observations
#### Example Custom Processor
For LIBERO environments, a custom processor handles the specific observation format:
```python
from lerobot.policies.xvla.processor_xvla import LiberoProcessorStep
processor = LiberoProcessorStep()
# Handles robot_state dictionary, converts rotation matrices to 6D representation
# Applies 180° image rotation for camera convention
```
### 4. Configuration Parameters
Key configuration parameters for X-VLA:
```python
# Observation and action
n_obs_steps: int = 1 # Number of observation timesteps
chunk_size: int = 32 # Action sequence length
n_action_steps: int = 32 # Number of action steps to execute
# Model architecture
hidden_size: int = 1024 # Transformer hidden dimension
depth: int = 24 # Number of transformer layers
num_heads: int = 16 # Number of attention heads
num_domains: int = 30 # Maximum number of domain IDs
len_soft_prompts: int = 32 # Length of soft prompt embeddings
# Action space
action_mode: str = "ee6d" # Action space type (use "auto" for auto-detection)
use_proprio: bool = True # Use proprioceptive state
max_state_dim: int = 32 # Maximum state dimension
max_action_dim: int = 20 # Max action dim for padding (used by "auto" mode)
# Vision
num_image_views: int | None # Number of camera views
resize_imgs_with_padding: tuple[int, int] | None # Target image size with padding
# Training
num_denoising_steps: int = 10 # Flow matching denoising steps
```
## Creating Custom Action Modes
If your robot has a unique action space, you can create a custom action mode:
### Step 1: Define Your Action Space
```python
from lerobot.policies.xvla.action_hub import BaseActionSpace, register_action
import torch.nn as nn
@register_action("my_custom_robot")
class MyCustomActionSpace(BaseActionSpace):
"""Custom action space for my robot."""
dim_action = 15 # Your robot's action dimension
gripper_idx = (7, 14) # Gripper channel indices
def __init__(self):
super().__init__()
self.mse = nn.MSELoss()
self.bce = nn.BCEWithLogitsLoss()
def compute_loss(self, pred, target):
"""Define your loss computation."""
# Example: MSE for joints, BCE for grippers
joints_loss = self.mse(pred[:, :, :7], target[:, :, :7])
gripper_loss = self.bce(pred[:, :, self.gripper_idx],
target[:, :, self.gripper_idx])
return {
"joints_loss": joints_loss,
"gripper_loss": gripper_loss,
}
def preprocess(self, proprio, action, mode="train"):
"""Preprocess actions before training."""
# Example: Zero out grippers in proprioception
proprio_m = proprio.clone()
action_m = action.clone() if action is not None else None
proprio_m[..., self.gripper_idx] = 0.0
if action_m is not None:
action_m[..., self.gripper_idx] = 0.0
return proprio_m, action_m
def postprocess(self, action):
"""Post-process predictions for deployment."""
# Example: Apply sigmoid to gripper logits
action[..., self.gripper_idx] = torch.sigmoid(action[..., self.gripper_idx])
return action
```
### Step 2: Use Your Custom Action Mode
```bash
lerobot-train \
--policy.action_mode=my_custom_robot \
--dataset.repo_id=YOUR_DATASET \
--policy.path="lerobot/xvla-base" \
...
```
## Advanced Topics
### Multi-Camera Support
X-VLA supports multiple camera views through the `num_image_views` parameter:
```python
# Configure for 3 camera views
policy.num_image_views=3
# Add empty cameras if you have fewer physical cameras
policy.empty_cameras=1 # Adds 1 zero-padded camera view
```
### Custom Preprocessing Pipeline
Create a custom preprocessing pipeline for your environment:
```python
from lerobot.processor import PolicyProcessorPipeline
from lerobot.policies.xvla import (
XVLAImageToFloatProcessorStep,
XVLAImageNetNormalizeProcessorStep,
XVLAAddDomainIdProcessorStep,
)
# Build custom pipeline
preprocessor = PolicyProcessorPipeline(
steps=[
YourCustomProcessorStep(), # Your custom processing
XVLAImageToFloatProcessorStep(), # Required: convert to float
XVLAImageNetNormalizeProcessorStep(), # Required: ImageNet norm
XVLAAddDomainIdProcessorStep(domain_id=5), # Your domain ID
]
)
```
### Handling Different Action Dimensions
When your dataset has fewer action dimensions than the pretrained model:
**Option 1 (Recommended)**: Use `auto` action mode
```bash
# Automatically detects your dataset's action dimension
# Works with any robot without custom code
policy.action_mode=auto
policy.max_action_dim=20 # Match pretrained model
```
**Option 2**: Use a predefined action mode with built-in padding
```python
# Model expects 20D, dataset has 12D
# Action mode handles padding internally
action_mode = "so101_bimanual" # Pads 12 → 20
```
**Option 2**: Create a custom action mode that maps dimensions explicitly
```python
@register_action("my_mapped_action")
class MappedActionSpace(BaseActionSpace):
dim_action = 20
REAL_DIM = 12
def _pad_to_model_dim(self, x):
# Custom padding logic
...
```
## Troubleshooting
### Common Issues
**Issue**: "Action dimension mismatch"
- **Solution**: Check that your `action_mode` matches your robot's action space. Create a custom action mode if needed.
**Issue**: "Image values outside [0, 1] range"
- **Solution**: Ensure images are preprocessed with `XVLAImageToFloatProcessorStep` before normalization.
**Issue**: "Domain ID not found"
- **Solution**: Make sure `XVLAAddDomainIdProcessorStep` is in your preprocessing pipeline with the correct domain_id.
**Issue**: "Low success rate on new embodiment"
- **Solution**:
1. Verify your action_mode is correct
2. Check that soft prompts are being trained (`train_soft_prompts=True`)
3. Ensure proper preprocessing (ImageNet normalization, domain_id)
4. Consider increasing training steps
**Issue**: "Out of memory during training"
- **Solution**:
1. Reduce `chunk_size` (e.g., from 32 to 16)
2. Enable gradient checkpointing
3. Reduce batch size
4. Freeze more components
## Citation
If you use X-VLA in your research, please cite:
```bibtex
@article{zheng2025x,
title = {X-VLA: Soft-Prompted Transformer as Scalable Cross-Embodiment Vision-Language-Action Model},
author = {Zheng, Jinliang and Li, Jianxiong and Wang, Zhihao and Liu, Dongxiu and Kang, Xirui
and Feng, Yuchun and Zheng, Yinan and Zou, Jiayin and Chen, Yilun and Zeng, Jia and others},
journal = {arXiv preprint arXiv:2510.10274},
year = {2025}
}
```
## Additional Resources
- [X-VLA Paper](https://arxiv.org/pdf/2510.10274)
- [LeRobot Documentation](https://github.com/huggingface/lerobot)
- [Action Registry Implementation](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/xvla/action_hub.py)
- [Processor Implementation](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/xvla/processor_xvla.py)
- [Model Configuration](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/xvla/configuration_xvla.py)
## Contributing
We welcome contributions! If you've implemented a new action mode or processor for your robot, please consider submitting a PR to help the community.