add check for cfg.policy in force_cpu line

* Fix SDK link and enhance setup instructions

Updated the Frodobots SDK link and added credential setup instructions.

Signed-off-by: ./c² <cagataycali@icloud.com>

* Update docs/source/earthrover_mini_plus.mdx

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
Signed-off-by: ./c² <cagataycali@icloud.com>

* Update docs/source/earthrover_mini_plus.mdx

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
Signed-off-by: Steven Palma <imstevenpmwork@ieee.org>

---------

Signed-off-by: ./c² <cagataycali@icloud.com>
Signed-off-by: Steven Palma <imstevenpmwork@ieee.org>
Co-authored-by: Steven Palma <imstevenpmwork@ieee.org>

.github/PULL_REQUEST_TEMPLATE.md

@@ -22,20 +22,21 @@ Short, imperative summary (e.g., "fix(robots): handle None in sensor parser"). S
 - Short, concrete bullets of the modifications (files/behaviour).
 - Short note if this introduces breaking changes and migration steps.
 
-## How was this tested
+## How was this tested (or how to run locally)
 
 - Tests added: list new tests or test files.
 - Manual checks / dataset runs performed.
+- Instructions for the reviewer
 
-## How to run locally (reviewer)
+Example:
 
-- Run the relevant tests:
+- Ran the relevant tests:
 
   ```bash
   pytest -q tests/ -k <keyword>
   ```
 
-- Run a quick example or CLI (if applicable):
+- Reproduce with a quick example or CLI (if applicable):
 
   ```bash
   lerobot-train --some.option=true

.github/workflows/full_tests.yml

@@ -186,15 +186,18 @@ jobs:
     steps:
       - name: Get Docker Hub Token and Delete Image
         # zizmor: ignore[template-injection]
+        env:
+          DOCKERHUB_LEROBOT_USERNAME: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
+          DOCKERHUB_LEROBOT_PASSWORD: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
+          IMAGE_FULL: ${{ needs.build-and-push-docker.outputs.image_tag }}
         run: |
-          IMAGE_NAME=$(echo "${{ needs.build-and-push-docker.outputs.image_tag }}" | cut -d':' -f1)
-          IMAGE_TAG=$(echo "${{ needs.build-and-push-docker.outputs.image_tag }}" | cut -d':' -f2)
-
+          IMAGE_NAME=$(echo "$IMAGE_FULL" | cut -d':' -f1)
+          IMAGE_TAG=$(echo "$IMAGE_FULL" | cut -d':' -f2-)
           echo "Attempting to delete image: $IMAGE_NAME:$IMAGE_TAG"
 
           TOKEN=$(curl -s -H "Content-Type: application/json" \
                        -X POST \
-                       -d '{"username": "${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}", "password": "${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}"}' \
+                       -d "{\"username\": \"$DOCKERHUB_LEROBOT_USERNAME\", \"password\": \"$DOCKERHUB_LEROBOT_PASSWORD\"}" \
                        https://hub.docker.com/v2/users/login/ | jq -r .token)
 
           if [ "$TOKEN" == "null" ] || [ -z "$TOKEN" ]; then
@@ -205,7 +208,7 @@ jobs:
           HTTP_RESPONSE=$(curl -s -o /dev/null -w "%{http_code}" \
                                -H "Authorization: JWT ${TOKEN}" \
                                -X DELETE \
-                               https://hub.docker.com/v2/repositories/${IMAGE_NAME}/tags/${IMAGE_TAG}/)
+                               https://hub.docker.com/v2/repositories/${IMAGE_NAME}/tags/$IMAGE_TAG)
 
           if [ "$HTTP_RESPONSE" -eq 204 ]; then
             echo "Successfully deleted Docker image tag: $IMAGE_NAME:$IMAGE_TAG"

.github/workflows/unbound_deps_tests.yml

@@ -162,15 +162,19 @@ jobs:
     steps:
       - name: Get Docker Hub Token and Delete Image
         # zizmor: ignore[template-injection]
+        env:
+          DOCKERHUB_LEROBOT_USERNAME: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
+          DOCKERHUB_LEROBOT_PASSWORD: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
+          IMAGE_FULL: ${{ needs.build-and-push-docker.outputs.image_tag }}
         run: |
-          IMAGE_NAME=$(echo "${{ needs.build-and-push-docker.outputs.image_tag }}" | cut -d':' -f1)
-          IMAGE_TAG=$(echo "${{ needs.build-and-push-docker.outputs.image_tag }}" | cut -d':' -f2)
+          IMAGE_NAME=$(echo "$IMAGE_FULL" | cut -d':' -f1)
+          IMAGE_TAG=$(echo "$IMAGE_FULL" | cut -d':' -f2)
 
           echo "Attempting to delete image: $IMAGE_NAME:$IMAGE_TAG"
 
           TOKEN=$(curl -s -H "Content-Type: application/json" \
                        -X POST \
-                       -d '{"username": "${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}", "password": "${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}"}' \
+                       -d "{\"username\": \"$DOCKERHUB_LEROBOT_USERNAME\", \"password\": \"$DOCKERHUB_LEROBOT_PASSWORD\"}" \
                        https://hub.docker.com/v2/users/login/ | jq -r .token)
 
           if [ "$TOKEN" == "null" ] || [ -z "$TOKEN" ]; then
@@ -181,7 +185,7 @@ jobs:
           HTTP_RESPONSE=$(curl -s -o /dev/null -w "%{http_code}" \
                                -H "Authorization: JWT ${TOKEN}" \
                                -X DELETE \
-                               https://hub.docker.com/v2/repositories/${IMAGE_NAME}/tags/${IMAGE_TAG}/)
+                               https://hub.docker.com/v2/repositories/${IMAGE_NAME}/tags/$IMAGE_TAG)
 
           if [ "$HTTP_RESPONSE" -eq 204 ]; then
             echo "Successfully deleted Docker image tag: $IMAGE_NAME:$IMAGE_TAG"

README.md

@@ -10,6 +10,7 @@
 [![Status](https://img.shields.io/pypi/status/lerobot)](https://pypi.org/project/lerobot/)
 [![Version](https://img.shields.io/pypi/v/lerobot)](https://pypi.org/project/lerobot/)
 [![Contributor Covenant](https://img.shields.io/badge/Contributor%20Covenant-v2.1-ff69b4.svg)](https://github.com/huggingface/lerobot/blob/main/CODE_OF_CONDUCT.md)
+[![Discord](https://img.shields.io/badge/Discord-Join_Us-5865F2?style=flat&logo=discord&logoColor=white)](https://discord.gg/q8Dzzpym3f)
 
 </div>
 
@@ -99,11 +100,11 @@ lerobot-train \
   --dataset.repo_id=lerobot/aloha_mobile_cabinet
 ```
 
-| Category                   | Models                                                                                                                                                                 |
-| -------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| **Imitation Learning**     | [ACT](./docs/source/policy_act_README.md), [Diffusion](./docs/source/policy_diffusion_README.md), [VQ-BeT](./docs/source/policy_vqbet_README.md)                       |
-| **Reinforcement Learning** | [HIL-SERL](./docs/source/hilserl.mdx), [TDMPC](./docs/source/policy_tdmpc_README.md) & QC-FQL (coming soon)                                                            |
-| **VLAs Models**            | [Pi0.5](./docs/source/pi05.mdx), [GR00T N1.5](./docs/source/policy_groot_README.md), [SmolVLA](./docs/source/policy_smolvla_README.md), [XVLA](./docs/source/xvla.mdx) |
+| Category                   | Models                                                                                                                                                                                                       |
+| -------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
+| **Imitation Learning**     | [ACT](./docs/source/policy_act_README.md), [Diffusion](./docs/source/policy_diffusion_README.md), [VQ-BeT](./docs/source/policy_vqbet_README.md)                                                             |
+| **Reinforcement Learning** | [HIL-SERL](./docs/source/hilserl.mdx), [TDMPC](./docs/source/policy_tdmpc_README.md) & QC-FQL (coming soon)                                                                                                  |
+| **VLAs Models**            | [Pi0Fast](./docs/source/pi0fast.mdx), [Pi0.5](./docs/source/pi05.mdx), [GR00T N1.5](./docs/source/policy_groot_README.md), [SmolVLA](./docs/source/policy_smolvla_README.md), [XVLA](./docs/source/xvla.mdx) |
 
 Similarly to the hardware, you can easily implement your own policy & leverage LeRobot's data collection, training, and visualization tools, and share your model to the HF Hub
 
@@ -127,7 +128,7 @@ Learn how to implement your own simulation environment or benchmark and distribu
 ## Resources
 
 - **[Documentation](https://huggingface.co/docs/lerobot/index):** The complete guide to tutorials & API.
-- **[Discord](https://discord.gg/3gxM6Avj):** Join the `LeRobot` server to discuss with the community.
+- **[Discord](https://discord.gg/q8Dzzpym3f):** Join the `LeRobot` server to discuss with the community.
 - **[X](https://x.com/LeRobotHF):** Follow us on X to stay up-to-date with the latest developments.
 - **[Robot Learning Tutorial](https://huggingface.co/spaces/lerobot/robot-learning-tutorial):** A free, hands-on course to learn robot learning using LeRobot.
 

docker/Dockerfile.internal

@@ -73,7 +73,7 @@ ENV HOME=/home/user_lerobot \
 RUN uv venv --python python${PYTHON_VERSION}
 
 # Install Python dependencies for caching
-COPY --chown=user_lerobot:user_lerobot pyproject.toml README.md MANIFEST.in ./
+COPY --chown=user_lerobot:user_lerobot setup.py pyproject.toml README.md MANIFEST.in ./
 COPY --chown=user_lerobot:user_lerobot src/ src/
 
 ARG UNBOUND_DEPS=false

docker/Dockerfile.user

@@ -59,7 +59,7 @@ ENV HOME=/home/user_lerobot \
 RUN uv venv
 
 # Install Python dependencies for caching
-COPY --chown=user_lerobot:user_lerobot pyproject.toml README.md MANIFEST.in ./
+COPY --chown=user_lerobot:user_lerobot setup.py pyproject.toml README.md MANIFEST.in ./
 COPY --chown=user_lerobot:user_lerobot src/ src/
 
 ARG UNBOUND_DEPS=false

docs/source/_toctree.yml

@@ -19,6 +19,8 @@
     title: Train RL in Simulation
   - local: multi_gpu_training
     title: Multi GPU training
+  - local: peft_training
+    title: Training with PEFT (e.g., LoRA)
   title: "Tutorials"
 - sections:
   - local: lerobot-dataset-v3
@@ -35,6 +37,8 @@
     title: SmolVLA
   - local: pi0
     title: π₀ (Pi0)
+  - local: pi0fast
+    title: π₀-FAST (Pi0Fast)
   - local: pi05
     title: π₀.₅ (Pi05)
   - local: groot
@@ -59,6 +63,8 @@
     title: Environments from the Hub
   - local: envhub_leisaac
     title: Control & Train Robots in Sim (LeIsaac)
+  - local: envhub_isaaclab_arena
+    title: NVIDIA IsaacLab Arena Environments
   - local: libero
     title: Using Libero
   - local: metaworld

docs/source/async.mdx

@@ -169,7 +169,7 @@ python -m lerobot.async_inference.robot_client \
 <!-- prettier-ignore-start -->
 ```python
 import threading
-from lerobot.robots.so100_follower import SO100FollowerConfig
+from lerobot.robots.so_follower import SO100FollowerConfig
 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
 from lerobot.async_inference.configs import RobotClientConfig
 from lerobot.async_inference.robot_client import RobotClient

docs/source/earthrover_mini_plus.mdx

@@ -12,23 +12,42 @@ The EarthRover Mini Plus is a fully open source mobile robot that connects throu
 
 ### Setting Up the Frodobots SDK
 
-The robot needs the [Frodobots SDK](https://github.com/Frodobots/earth-rovers-sdk) running on your computer. Here's how:
+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/earth-rovers-sdk.git
+git clone https://github.com/frodobots-org/earth-rovers-sdk.git
 cd earth-rovers-sdk
 pip install -r requirements.txt
 ```
 
-2. Start the SDK:
+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
 ```
 
-3. Open your web browser and go to `http://localhost:8000`, then click "Join"
+4. Open your web browser and go to `http://localhost:8000`, then click "Join"
 
 The SDK gives you:
 

docs/source/envhub.mdx

@@ -2,14 +2,32 @@
 
 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.
 
-## Overview
+## What is EnvHub?
 
-With EnvHub, you can:
+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.
 
-- Load environments from the Hub instantly
-- Share your custom simulation tasks with the community
-- Version control your environments using Git
-- Distribute complex physics simulations without packaging hassles
+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
 
@@ -29,17 +47,6 @@ env = make_env("lerobot/cartpole-env", trust_remote_code=True)
   hash for reproducibility and security.
 </Tip>
 
-## What is EnvHub?
-
-EnvHub is a framework that allows researchers and developers to:
-
-1. **Publish environments** to the Hugging Face Hub as Git repositories
-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, without worrying about dependency conflicts or complex installation procedures.
-
 ## Repository Structure
 
 To make your environment loadable from the Hub, your repository must contain at minimum:

docs/source/envhub_isaaclab_arena.mdx

@@ -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
+git clone https://github.com/huggingface/lerobot.git
+cd lerobot
+pip install -e .
+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 N1.5](./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.factory 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

@@ -137,7 +137,8 @@ from lerobot.teleoperators import (  # noqa: F401
     Teleoperator,
     TeleoperatorConfig,
     make_teleoperator_from_config,
-    so101_leader,
+    so_leader,
+    bi_so_leader,
 )
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import init_logging
@@ -196,7 +197,7 @@ def teleop_loop(teleop: Teleoperator, env: gym.Env, fps: int):
             obs, info = env.reset()
 
         dt_s = time.perf_counter() - loop_start
-        precise_sleep(1 / fps - dt_s)
+        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)")
 
@@ -222,7 +223,7 @@ def teleoperate(cfg: TeleoperateConfig):
 
 def main():
     teleoperate(TeleoperateConfig(
-        teleop=so101_leader.SO101LeaderConfig(
+        teleop=so_leader.SO101LeaderConfig(
             port="/dev/ttyACM0",
             id='leader',
             use_degrees=False,

docs/source/groot.mdx

@@ -12,6 +12,12 @@ Developers and researchers can post-train GR00T N1.5 with their own real or synt
 
 GR00T N1.5 (specifically the GR00T-N1.5-3B model) is built using pre-trained vision and language encoders. It utilizes 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.
@@ -103,7 +109,7 @@ Once you have trained your model using your parameters you can run inference in
 
 ```bash
 lerobot-record \
-  --robot.type=bi_so100_follower \
+  --robot.type=bi_so_follower \
   --robot.left_arm_port=/dev/ttyACM1 \
   --robot.right_arm_port=/dev/ttyACM0 \
   --robot.id=bimanual_follower \

docs/source/il_robots.mdx

@@ -58,8 +58,8 @@ lerobot-teleoperate \
 
 <!-- prettier-ignore-start -->
 ```python
-from lerobot.teleoperators.so101_leader import SO101LeaderConfig, SO101Leader
-from lerobot.robots.so101_follower import SO101FollowerConfig, SO101Follower
+from lerobot.teleoperators.so_leader import SO101LeaderConfig, SO101Leader
+from lerobot.robots.so_follower import SO101FollowerConfig, SO101Follower
 
 robot_config = SO101FollowerConfig(
     port="/dev/tty.usbmodem58760431541",
@@ -195,9 +195,9 @@ lerobot-record \
 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.datasets.utils import hw_to_dataset_features
-from lerobot.robots.so100_follower import SO100Follower, SO100FollowerConfig
-from lerobot.teleoperators.so100_leader.config_so100_leader import SO100LeaderConfig
-from lerobot.teleoperators.so100_leader.so100_leader import SO100Leader
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
+from lerobot.teleoperators.so_leader.config_so100_leader import SO100LeaderConfig
+from lerobot.teleoperators.so_leader.so100_leader import SO100Leader
 from lerobot.utils.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun
@@ -408,8 +408,8 @@ lerobot-replay \
 import time
 
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
-from lerobot.robots.so100_follower.so100_follower import SO100Follower
+from lerobot.robots.so_follower.config_so100_follower import SO100FollowerConfig
+from lerobot.robots.so_follower.so100_follower import SO100Follower
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import log_say
 
@@ -432,7 +432,7 @@ for idx in range(dataset.num_frames):
     }
     robot.send_action(action)
 
-    precise_sleep(1.0 / dataset.fps - (time.perf_counter() - t0))
+    precise_sleep(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
 
 robot.disconnect()
 ```
@@ -531,8 +531,8 @@ from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.datasets.utils import hw_to_dataset_features
 from lerobot.policies.act.modeling_act import ACTPolicy
 from lerobot.policies.factory import make_pre_post_processors
-from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
-from lerobot.robots.so100_follower.so100_follower import SO100Follower
+from lerobot.robots.so_follower.config_so100_follower import SO100FollowerConfig
+from lerobot.robots.so_follower.so100_follower import SO100Follower
 from lerobot.scripts.lerobot_record import record_loop
 from lerobot.utils.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say

docs/source/integrate_hardware.mdx

@@ -18,7 +18,7 @@ If you're using Feetech or Dynamixel motors, LeRobot provides built-in bus inter
 - [`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/so101_follower/so101_follower.py)
+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):
 

docs/source/lekiwi.mdx

@@ -204,7 +204,7 @@ lerobot-calibrate \
 
 <!-- prettier-ignore-start -->
 ```python
-from lerobot.teleoperators.so100_leader import SO100LeaderConfig, SO100Leader
+from lerobot.teleoperators.so_leader import SO100LeaderConfig, SO100Leader
 
 config = SO100LeaderConfig(
     port="/dev/tty.usbmodem58760431551",

docs/source/peft_training.mdx

@@ -0,0 +1,62 @@
+# 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
+```
+
+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. 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

@@ -44,7 +44,7 @@ Modify the examples to use `PhoneOS.IOS` or `PhoneOS.ANDROID` in `PhoneConfig`.
 
 Teleoperation example:
 
-```36:43:examples/phone_so100_teleop.py
+```python
 from lerobot.teleoperators.phone.config_phone import PhoneConfig, PhoneOS
 
 teleop_config = PhoneConfig(phone_os=PhoneOS.IOS)  # or PhoneOS.ANDROID
@@ -103,7 +103,7 @@ Additionally you can customize mapping or safety limits by editing the processor
 
 - 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.
 
-  ```examples/phone_to_so100/teleoperate.py
+  ```python
   kinematics_solver = RobotKinematics(
     urdf_path="./SO101/so101_new_calib.urdf",
     target_frame_name="gripper_frame_link",
@@ -114,7 +114,7 @@ Additionally you can customize mapping or safety limits by editing the processor
 
 - The `MapPhoneActionToRobotAction` step converts the calibrated phone pose and inputs into target deltas and gripper commands, below is shown what the step outputs.
 
-  ```src/lerobot/teleoperators/phone/phone_processor.py
+  ```python
   action["enabled"] = enabled
         action["target_x"] = -pos[1] if enabled else 0.0
         action["target_y"] = pos[0] if enabled else 0.0
@@ -127,7 +127,7 @@ Additionally you can customize mapping or safety limits by editing the processor
 
 - 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.
 
-  ```examples/phone_to_so100/teleoperate.py
+  ```python
   EEReferenceAndDelta(
       kinematics=kinematics_solver,
       end_effector_step_sizes={"x": 0.5, "y": 0.5, "z": 0.5},
@@ -138,7 +138,7 @@ Additionally you can customize mapping or safety limits by editing the processor
 
 - 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.
 
-  ```examples/phone_to_so100/teleoperate.py
+  ```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,
@@ -147,7 +147,7 @@ Additionally you can customize mapping or safety limits by editing the processor
 
 - 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.
 
-  ```examples/phone_to_so100/teleoperate.py
+  ```python
   GripperVelocityToJoint(speed_factor=20.0)
   ```
 
@@ -157,7 +157,7 @@ We use different IK initial guesses in the kinematic steps. As initial guess eit
 
 - Closed loop (used in record/eval): sets `initial_guess_current_joints=True` so IK starts from the measured joints each frame.
 
-  ```examples/phone_to_so100/record.py
+  ```python
   InverseKinematicsEEToJoints(
       kinematics=kinematics_solver,
       motor_names=list(robot.bus.motors.keys()),
@@ -167,7 +167,7 @@ We use different IK initial guesses in the kinematic steps. As initial guess eit
 
 - 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.
 
-  ```examples/phone_to_so100/replay.py
+  ```python
   InverseKinematicsEEToJoints(
       kinematics=kinematics_solver,
       motor_names=list(robot.bus.motors.keys()),

docs/source/pi0.mdx

@@ -6,6 +6,12 @@
 
 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.
@@ -64,6 +70,8 @@ python src/lerobot/scripts/lerobot_train.py \
     --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
@@ -79,6 +87,15 @@ python src/lerobot/scripts/lerobot_train.py \
   - [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.
+
 ## 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

@@ -67,6 +67,8 @@ python src/lerobot/scripts/lerobot_train.py\
     --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
@@ -82,6 +84,15 @@ python src/lerobot/scripts/lerobot_train.py\
   - [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

docs/source/pi0fast.mdx

@@ -0,0 +1,246 @@
+# 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]"
+   ```
+
+   > [!NOTE]
+   > For lerobot 0.4.0, if you want to install the pi tag, you will have to do: `pip install "lerobot[pi]@git+https://github.com/huggingface/lerobot.git"`.
+   >
+   > This will be solved in the next patch release
+
+## Training a Custom FAST Tokenizer
+
+You have two options for the FAST tokenizer:
+
+1. **Use the pre-trained tokenizer**: The `physical-intelligence/fast` 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/pi0_fast_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                              | `physical-intelligence/fast` |
+| `--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_walloss_README.md

@@ -1,20 +1,30 @@
 # WALL-OSS
 
-This repository contains the Hugging Face port of **WALL-OSS**, a Vision-Language-Action model for cross-embodiment robotic control based on Qwen2.5-VL with flow matching/FAST action prediction.
+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 |     |
+| 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:
@@ -32,4 +42,4 @@ If you use this work, please cite:
 
 ## License
 
-This port follows the **Apache 2.0 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/processors_robots_teleop.mdx

@@ -30,7 +30,7 @@ Each of these pipelines handle different conversions between different action an
 
 Below is an example of the three pipelines that we use in the phone to SO-100 follower examples:
 
-```69:90:examples/phone_so100_record.py
+```python
 phone_to_robot_ee_pose_processor = RobotProcessorPipeline[RobotAction, RobotAction]( # teleop -> dataset action
     steps=[
         MapPhoneActionToRobotAction(platform=teleop_config.phone_os),
@@ -84,7 +84,7 @@ Dataset features are determined by the keys saved in the dataset. Each step can
 
 Below is and example of how we declare features with the `transform_features` method in the phone to SO-100 follower examples:
 
-```src/lerobot/robots/so100_follower/robot_kinematic_processor.py
+```python
     def transform_features(
         self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
     ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
@@ -103,7 +103,7 @@ Here we declare what PolicyFeatures we modify in this step, so we know what feat
 
 Below is an example of how we aggregate and merge features in the phone to SO-100 record example:
 
-```121:145:examples/phone_so100_record.py
+```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

docs/source/reachy2.mdx

@@ -38,6 +38,7 @@ docker run --rm -it \
   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:
 >
 > ```
@@ -141,7 +142,7 @@ If you choose this option but still want to use the VR teleoperation application
 First add reachy2 and reachy2_teleoperator to the imports of the record script. Then you can use the following command:
 
 ```bash
-python -m lerobot.record \
+lerobot-record \
     --robot.type=reachy2 \
     --robot.ip_address=192.168.0.200 \
     --robot.id=r2-0000 \
@@ -150,6 +151,7 @@ python -m lerobot.record \
     --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 \
@@ -165,7 +167,7 @@ python -m lerobot.record \
 **Extended setup overview (all options included):**
 
 ```bash
-python -m lerobot.record \
+lerobot-record \
     --robot.type=reachy2 \
     --robot.ip_address=192.168.0.200 \
     --robot.use_external_commands=true \
@@ -177,6 +179,8 @@ python -m lerobot.record \
     --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 \
@@ -212,9 +216,10 @@ Must be set to true if a compliant Reachy 2 is used to control another one.
 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.
 
@@ -222,49 +227,60 @@ 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`.
+> 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**. By default, only the **teleoperation cameras** are recorded (both `left_teleop_camera` and `right_teleop_camera`). Enable or disable each camera with:
+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>
 ```
 
---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
-python -m lerobot.replay \
+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
-    --display_data=true
-````
+```
 
 ## Step 3: Train
 
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
   --dataset.repo_id=pollen_robotics/record_test \
   --policy.type=act \
   --output_dir=outputs/train/reachy2_test \
@@ -277,10 +293,9 @@ python -m lerobot.scripts.train \
 ## Step 4: Evaluate
 
 ```bash
-python -m lerobot.record \
+lerobot-eval \
   --robot.type=reachy2 \
   --robot.ip_address=192.168.0.200 \
-  --display_data=false \
   --dataset.repo_id=pollen_robotics/eval_record_test \
   --dataset.single_task="Evaluate reachy2 policy" \
   --dataset.num_episodes=10 \

docs/source/sarm.mdx

@@ -4,6 +4,12 @@ SARM (Stage-Aware Reward Modeling) is a video-based reward modeling framework fo
 
 **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.

docs/source/so100.mdx

@@ -103,7 +103,7 @@ lerobot-setup-motors \
 
 <!-- prettier-ignore-start -->
 ```python
-from lerobot.robots.so100_follower import SO100Follower, SO100FollowerConfig
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
 
 config = SO100FollowerConfig(
     port="/dev/tty.usbmodem585A0076841",
@@ -177,7 +177,7 @@ lerobot-setup-motors \
 
 <!-- prettier-ignore-start -->
 ```python
-from lerobot.teleoperators.so100_leader import SO100Leader, SO100LeaderConfig
+from lerobot.teleoperators.so_leader import SO100Leader, SO100LeaderConfig
 
 config = SO100LeaderConfig(
     port="/dev/tty.usbmodem585A0076841",
@@ -579,7 +579,7 @@ lerobot-calibrate \
 
 <!-- prettier-ignore-start -->
 ```python
-from lerobot.robots.so100_follower import SO100FollowerConfig, SO100Follower
+from lerobot.robots.so_follower import SO100FollowerConfig, SO100Follower
 
 config = SO100FollowerConfig(
     port="/dev/tty.usbmodem585A0076891",
@@ -617,7 +617,7 @@ lerobot-calibrate \
 
 <!-- prettier-ignore-start -->
 ```python
-from lerobot.teleoperators.so100_leader import SO100LeaderConfig, SO100Leader
+from lerobot.teleoperators.so_leader import SO100LeaderConfig, SO100Leader
 
 config = SO100LeaderConfig(
     port="/dev/tty.usbmodem58760431551",

docs/source/so101.mdx

@@ -125,7 +125,7 @@ lerobot-setup-motors \
 
 <!-- prettier-ignore-start -->
 ```python
-from lerobot.robots.so101_follower import SO101Follower, SO101FollowerConfig
+from lerobot.robots.so_follower import SO101Follower, SO101FollowerConfig
 
 config = SO101FollowerConfig(
     port="/dev/tty.usbmodem585A0076841",
@@ -201,7 +201,7 @@ lerobot-setup-motors \
 
 <!-- prettier-ignore-start -->
 ```python
-from lerobot.teleoperators.so101_leader import SO101Leader, SO101LeaderConfig
+from lerobot.teleoperators.so_leader import SO101Leader, SO101LeaderConfig
 
 config = SO101LeaderConfig(
     port="/dev/tty.usbmodem585A0076841",
@@ -364,7 +364,7 @@ lerobot-calibrate \
 
 <!-- prettier-ignore-start -->
 ```python
-from lerobot.robots.so101_follower import SO101FollowerConfig, SO101Follower
+from lerobot.robots.so_follower import SO101FollowerConfig, SO101Follower
 
 config = SO101FollowerConfig(
     port="/dev/tty.usbmodem585A0076891",
@@ -413,7 +413,7 @@ lerobot-calibrate \
 
 <!-- prettier-ignore-start -->
 ```python
-from lerobot.teleoperators.so101_leader import SO101LeaderConfig, SO101Leader
+from lerobot.teleoperators.so_leader import SO101LeaderConfig, SO101Leader
 
 config = SO101LeaderConfig(
     port="/dev/tty.usbmodem58760431551",

docs/source/unitree_g1.mdx

@@ -1,21 +1,21 @@
-# Unitree G1 Robot Setup and Control
+# Unitree G1
 
 This guide covers the complete setup process for the Unitree G1 humanoid, from initial connection to running gr00t_wbc locomotion.
 
-## About the Unitree G1
+## About
 
-We offer support for both 29 and 23 DOF G1. We introduce:
+We support both 29 and 23 DOF G1 EDU version. We introduce:
 
-- **`unitree g1` robot class, handling low level communication with the humanoid**
-- **ZMQ socket bridge** for remote communication over WiFi, allowing one to deploy policies remotely instead of over ethernet or directly on the Orin
-- **GR00T locomotion policy** for bipedal walking and balance
-- **MuJoCo simulation mode** for testing policies without the physical robot
+- **`unitree g1` robot class, handling low level read/write from/to the humanoid**
+- **ZMQ socket bridge** for remote communication and camera streaming, allowing for remote policy deployment over wlan, eth or directly on the robot
+- **Locomotion policies** from NVIDIA gr00t and Amazon FAR Holosoma
+- **Simulation mode** for testing policies without the physical robot in mujoco
 
 ---
 
-## Part 1: Connect to Robot over Ethernet
+## Connection guide
 
-### Step 1: Configure Your Computer's Ethernet Interface
+### Step 1: Configure Ethernet Interface
 
 Set a static IP on the same subnet as the robot:
 
@@ -26,7 +26,7 @@ sudo ip addr add 192.168.123.200/24 dev enp131s0
 sudo ip link set enp131s0 up
 ```
 
-**Note**: The robot's Ethernet IP is fixed at `192.168.123.164`. Your computer must use `192.168.123.x` where x ≠ 164.
+**Note**: The G1's Ethernet IP is fixed at `192.168.123.164`. Your computer must use `192.168.123.x` with x ≠ 164.
 
 ### Step 2: SSH into the Robot
 
@@ -35,25 +35,24 @@ ssh unitree@192.168.123.164
 # Password: 123
 ```
 
-You should now be connected to the robot's onboard computer.
+You should now be connected to the G1's Orin.
 
 ---
 
 ## Part 2: Enable WiFi on the Robot
 
-Once connected via Ethernet, follow these steps to enable WiFi:
+Wlan0 is disabled by default on the G1. To enable it:
 
 ### Step 1: Enable WiFi Hardware
 
 ```bash
-# Unblock WiFi radio
 sudo rfkill unblock wifi
 sudo rfkill unblock all
 
-# Bring up WiFi interface
+# Bring up wlan0
 sudo ip link set wlan0 up
 
-# Enable NetworkManager control
+# Enable NetworkManager control of wlan0
 sudo nmcli radio wifi on
 sudo nmcli device set wlan0 managed yes
 sudo systemctl restart NetworkManager
@@ -73,7 +72,7 @@ sudo iptables -A FORWARD -i wlp132s0f0 -o enp131s0 -m state --state RELATED,ESTA
 sudo iptables -A FORWARD -i enp131s0 -o wlp132s0f0 -j ACCEPT
 ```
 
-**On the robot:**
+**On the G1:**
 
 ```bash
 # Add laptop as default gateway
@@ -111,7 +110,7 @@ ssh unitree@<YOUR_ROBOT_IP>
 # Password: 123
 ```
 
-Replace `<YOUR_ROBOT_IP>` with your robot's actual WiFi IP address (e.g., `172.18.129.215`).
+Replace `<YOUR_ROBOT_IP>` with your robot's actual WiFi IP address.
 
 ---
 
@@ -147,9 +146,9 @@ python src/lerobot/robots/unitree_g1/run_g1_server.py
 
 ---
 
-## Part 4: Running GR00T Locomotion
+## Part 4: Controlling the robot
 
-With the robot server running, you can now control the robot from your laptop.
+With the robot server running, you can now control the robot remotely. Let's launch a locomotion policy
 
 ### Step 1: Install LeRobot on your machine
 
@@ -172,34 +171,30 @@ Edit the config file to match your robot's WiFi IP:
 robot_ip: str = "<YOUR_ROBOT_IP>"  # Replace with your robot's WiFi IP.
 ```
 
-**Note**: When running directly on the G1 (not remotely), set `robot_ip: str = "127.0.0.1"` instead.
-
 ### Step 3: Run the Locomotion Policy
 
 ```bash
 # Run GR00T locomotion controller
 python examples/unitree_g1/gr00t_locomotion.py --repo-id "nepyope/GR00T-WholeBodyControl_g1"
+
+# Run Holosoma locomotion controller
+python examples/unitree_g1/holosoma_locomotion.py
+
 ```
 
-### Step 4: Control with Remote
-
-- **Left stick**: Forward/backward and left/right movement
-- **Right stick**: Rotation
-- **R1 button**: Raise waist height
-- **R2 button**: Lower waist height
-
 Press `Ctrl+C` to stop the policy.
 
 ---
 
-## Extra: Running in Simulation Mode (MuJoCo)
+## Running in Simulation Mode (MuJoCo)
 
-You can now test and develop policies without a physical robot using MuJoCo. to do so set `is_simulation=True` in config.
+You can now test policies before unleashing them on the physical robot using MuJoCo. To do so simply set `is_simulation=True` in config.
 
 ## Additional Resources
 
 - [Unitree SDK Documentation](https://github.com/unitreerobotics/unitree_sdk2_python)
-- [GR00T Policy Repository](https://huggingface.co/nepyope/GR00T-WholeBodyControl_g1)
+- [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)
 

docs/source/walloss.mdx

@@ -8,6 +8,12 @@ X Square Robot’s WALL-OSS is now integrated into Hugging Face’s LeRobot ecos
 
 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.

examples/backward_compatibility/replay.py

@@ -41,8 +41,7 @@ from lerobot.robots import (  # noqa: F401
     RobotConfig,
     koch_follower,
     make_robot_from_config,
-    so100_follower,
-    so101_follower,
+    so_follower,
 )
 from lerobot.utils.constants import ACTION
 from lerobot.utils.robot_utils import precise_sleep
@@ -97,7 +96,7 @@ def replay(cfg: ReplayConfig):
         robot.send_action(action)
 
         dt_s = time.perf_counter() - start_episode_t
-        precise_sleep(1 / dataset.fps - dt_s)
+        precise_sleep(max(1 / dataset.fps - dt_s, 0.0))
 
     robot.disconnect()
 

examples/lekiwi/record.py

@@ -21,7 +21,7 @@ from lerobot.robots.lekiwi.config_lekiwi import LeKiwiClientConfig
 from lerobot.robots.lekiwi.lekiwi_client import LeKiwiClient
 from lerobot.scripts.lerobot_record import record_loop
 from lerobot.teleoperators.keyboard import KeyboardTeleop, KeyboardTeleopConfig
-from lerobot.teleoperators.so100_leader import SO100Leader, SO100LeaderConfig
+from lerobot.teleoperators.so_leader import SO100Leader, SO100LeaderConfig
 from lerobot.utils.constants import ACTION, OBS_STR
 from lerobot.utils.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say

examples/lekiwi/teleoperate.py

@@ -18,7 +18,7 @@ import time
 
 from lerobot.robots.lekiwi import LeKiwiClient, LeKiwiClientConfig
 from lerobot.teleoperators.keyboard.teleop_keyboard import KeyboardTeleop, KeyboardTeleopConfig
-from lerobot.teleoperators.so100_leader import SO100Leader, SO100LeaderConfig
+from lerobot.teleoperators.so_leader import SO100Leader, SO100LeaderConfig
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
 

examples/phone_to_so100/evaluate.py

@@ -34,12 +34,11 @@ from lerobot.processor.converters import (
     transition_to_observation,
     transition_to_robot_action,
 )
-from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
-from lerobot.robots.so100_follower.robot_kinematic_processor import (
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
+from lerobot.robots.so_follower.robot_kinematic_processor import (
     ForwardKinematicsJointsToEE,
     InverseKinematicsEEToJoints,
 )
-from lerobot.robots.so100_follower.so100_follower import SO100Follower
 from lerobot.scripts.lerobot_record import record_loop
 from lerobot.utils.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say

examples/phone_to_so100/record.py

@@ -26,15 +26,14 @@ from lerobot.processor.converters import (
     transition_to_observation,
     transition_to_robot_action,
 )
-from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
-from lerobot.robots.so100_follower.robot_kinematic_processor import (
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
+from lerobot.robots.so_follower.robot_kinematic_processor import (
     EEBoundsAndSafety,
     EEReferenceAndDelta,
     ForwardKinematicsJointsToEE,
     GripperVelocityToJoint,
     InverseKinematicsEEToJoints,
 )
-from lerobot.robots.so100_follower.so100_follower import SO100Follower
 from lerobot.scripts.lerobot_record import record_loop
 from lerobot.teleoperators.phone.config_phone import PhoneConfig, PhoneOS
 from lerobot.teleoperators.phone.phone_processor import MapPhoneActionToRobotAction

examples/phone_to_so100/replay.py

@@ -23,11 +23,10 @@ from lerobot.processor.converters import (
     robot_action_observation_to_transition,
     transition_to_robot_action,
 )
-from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
-from lerobot.robots.so100_follower.robot_kinematic_processor import (
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
+from lerobot.robots.so_follower.robot_kinematic_processor import (
     InverseKinematicsEEToJoints,
 )
-from lerobot.robots.so100_follower.so100_follower import SO100Follower
 from lerobot.utils.constants import ACTION
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import log_say
@@ -96,7 +95,7 @@ def main():
         # Send action to robot
         _ = robot.send_action(joint_action)
 
-        precise_sleep(1.0 / dataset.fps - (time.perf_counter() - t0))
+        precise_sleep(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
 
     # Clean up
     robot.disconnect()

examples/phone_to_so100/teleoperate.py

@@ -21,14 +21,13 @@ from lerobot.processor.converters import (
     robot_action_observation_to_transition,
     transition_to_robot_action,
 )
-from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
-from lerobot.robots.so100_follower.robot_kinematic_processor import (
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
+from lerobot.robots.so_follower.robot_kinematic_processor import (
     EEBoundsAndSafety,
     EEReferenceAndDelta,
     GripperVelocityToJoint,
     InverseKinematicsEEToJoints,
 )
-from lerobot.robots.so100_follower.so100_follower import SO100Follower
 from lerobot.teleoperators.phone.config_phone import PhoneConfig, PhoneOS
 from lerobot.teleoperators.phone.phone_processor import MapPhoneActionToRobotAction
 from lerobot.teleoperators.phone.teleop_phone import Phone

examples/rtc/eval_with_real_robot.py

@@ -94,9 +94,9 @@ from lerobot.rl.process import ProcessSignalHandler
 from lerobot.robots import (  # noqa: F401
     Robot,
     RobotConfig,
+    bi_so_follower,
     koch_follower,
-    so100_follower,
-    so101_follower,
+    so_follower,
 )
 from lerobot.robots.utils import make_robot_from_config
 from lerobot.utils.constants import OBS_IMAGES
@@ -455,7 +455,18 @@ def demo_cli(cfg: RTCDemoConfig):
     if cfg.policy.type == "pi05" or cfg.policy.type == "pi0":
         config.compile_model = cfg.use_torch_compile
 
-    policy = policy_class.from_pretrained(cfg.policy.pretrained_path, config=config)
+    if config.use_peft:
+        from peft import PeftConfig, PeftModel
+
+        peft_pretrained_path = cfg.policy.pretrained_path
+        peft_config = PeftConfig.from_pretrained(peft_pretrained_path)
+
+        policy = policy_class.from_pretrained(
+            pretrained_name_or_path=peft_config.base_model_name_or_path, config=config
+        )
+        policy = PeftModel.from_pretrained(policy, peft_pretrained_path, config=peft_config)
+    else:
+        policy = policy_class.from_pretrained(cfg.policy.pretrained_path, config=config)
 
     # Turn on RTC
     policy.config.rtc_config = cfg.rtc

examples/so100_to_so100_EE/evaluate.py

@@ -34,12 +34,11 @@ from lerobot.processor.converters import (
     transition_to_observation,
     transition_to_robot_action,
 )
-from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
-from lerobot.robots.so100_follower.robot_kinematic_processor import (
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
+from lerobot.robots.so_follower.robot_kinematic_processor import (
     ForwardKinematicsJointsToEE,
     InverseKinematicsEEToJoints,
 )
-from lerobot.robots.so100_follower.so100_follower import SO100Follower
 from lerobot.scripts.lerobot_record import record_loop
 from lerobot.utils.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say

examples/so100_to_so100_EE/record.py

@@ -27,16 +27,14 @@ from lerobot.processor.converters import (
     transition_to_observation,
     transition_to_robot_action,
 )
-from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
-from lerobot.robots.so100_follower.robot_kinematic_processor import (
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
+from lerobot.robots.so_follower.robot_kinematic_processor import (
     EEBoundsAndSafety,
     ForwardKinematicsJointsToEE,
     InverseKinematicsEEToJoints,
 )
-from lerobot.robots.so100_follower.so100_follower import SO100Follower
 from lerobot.scripts.lerobot_record import record_loop
-from lerobot.teleoperators.so100_leader.config_so100_leader import SO100LeaderConfig
-from lerobot.teleoperators.so100_leader.so100_leader import SO100Leader
+from lerobot.teleoperators.so_leader import SO100Leader, SO100LeaderConfig
 from lerobot.utils.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun

examples/so100_to_so100_EE/replay.py

@@ -24,11 +24,10 @@ from lerobot.processor.converters import (
     robot_action_observation_to_transition,
     transition_to_robot_action,
 )
-from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
-from lerobot.robots.so100_follower.robot_kinematic_processor import (
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
+from lerobot.robots.so_follower.robot_kinematic_processor import (
     InverseKinematicsEEToJoints,
 )
-from lerobot.robots.so100_follower.so100_follower import SO100Follower
 from lerobot.utils.constants import ACTION
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import log_say
@@ -97,7 +96,7 @@ def main():
         # Send action to robot
         _ = robot.send_action(joint_action)
 
-        precise_sleep(1.0 / dataset.fps - (time.perf_counter() - t0))
+        precise_sleep(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
 
     # Clean up
     robot.disconnect()

examples/so100_to_so100_EE/teleoperate.py

@@ -23,15 +23,13 @@ from lerobot.processor.converters import (
     robot_action_to_transition,
     transition_to_robot_action,
 )
-from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
-from lerobot.robots.so100_follower.robot_kinematic_processor import (
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
+from lerobot.robots.so_follower.robot_kinematic_processor import (
     EEBoundsAndSafety,
     ForwardKinematicsJointsToEE,
     InverseKinematicsEEToJoints,
 )
-from lerobot.robots.so100_follower.so100_follower import SO100Follower
-from lerobot.teleoperators.so100_leader.config_so100_leader import SO100LeaderConfig
-from lerobot.teleoperators.so100_leader.so100_leader import SO100Leader
+from lerobot.teleoperators.so_leader import SO100Leader, SO100LeaderConfig
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
 

examples/tutorial/act/act_using_example.py

@@ -5,8 +5,7 @@ from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
 from lerobot.policies.act.modeling_act import ACTPolicy
 from lerobot.policies.factory import make_pre_post_processors
 from lerobot.policies.utils import build_inference_frame, make_robot_action
-from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
-from lerobot.robots.so100_follower.so100_follower import SO100Follower
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
 
 MAX_EPISODES = 5
 MAX_STEPS_PER_EPISODE = 20

examples/tutorial/async-inf/robot_client.py

@@ -4,7 +4,7 @@ from lerobot.async_inference.configs import RobotClientConfig
 from lerobot.async_inference.helpers import visualize_action_queue_size
 from lerobot.async_inference.robot_client import RobotClient
 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.robots.so100_follower import SO100FollowerConfig
+from lerobot.robots.so_follower import SO100FollowerConfig
 
 
 def main():

examples/tutorial/diffusion/diffusion_using_example.py

@@ -5,8 +5,7 @@ from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
 from lerobot.policies.diffusion.modeling_diffusion import DiffusionPolicy
 from lerobot.policies.factory import make_pre_post_processors
 from lerobot.policies.utils import build_inference_frame, make_robot_action
-from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
-from lerobot.robots.so100_follower.so100_follower import SO100Follower
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
 
 MAX_EPISODES = 5
 MAX_STEPS_PER_EPISODE = 20

examples/tutorial/pi0/using_pi0_example.py

@@ -5,8 +5,7 @@ from lerobot.datasets.utils import hw_to_dataset_features
 from lerobot.policies.factory import make_pre_post_processors
 from lerobot.policies.pi0.modeling_pi0 import PI0Policy
 from lerobot.policies.utils import build_inference_frame, make_robot_action
-from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
-from lerobot.robots.so100_follower.so100_follower import SO100Follower
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
 
 MAX_EPISODES = 5
 MAX_STEPS_PER_EPISODE = 20

examples/tutorial/rl/hilserl_example.py

@@ -14,8 +14,8 @@ from lerobot.policies.sac.modeling_sac import SACPolicy
 from lerobot.policies.sac.reward_model.modeling_classifier import Classifier
 from lerobot.rl.buffer import ReplayBuffer
 from lerobot.rl.gym_manipulator import make_robot_env
-from lerobot.robots.so100_follower import SO100FollowerConfig
-from lerobot.teleoperators.so100_leader import SO100LeaderConfig
+from lerobot.robots.so_follower import SO100FollowerConfig
+from lerobot.teleoperators.so_leader import SO100LeaderConfig
 from lerobot.teleoperators.utils import TeleopEvents
 
 LOG_EVERY = 10

examples/tutorial/smolvla/using_smolvla_example.py

@@ -5,8 +5,7 @@ from lerobot.datasets.utils import hw_to_dataset_features
 from lerobot.policies.factory import make_pre_post_processors
 from lerobot.policies.smolvla.modeling_smolvla import SmolVLAPolicy
 from lerobot.policies.utils import build_inference_frame, make_robot_action
-from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
-from lerobot.robots.so100_follower.so100_follower import SO100Follower
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
 
 MAX_EPISODES = 5
 MAX_STEPS_PER_EPISODE = 20

examples/unitree_g1/gr00t_locomotion.py

@@ -13,16 +13,9 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-"""
-Example: GR00T Locomotion with Pre-loaded Policies
-
-This example demonstrates the NEW pattern for loading GR00T policies externally
-and passing them to the robot class.
-"""
 
 import argparse
 import logging
-import threading
 import time
 from collections import deque
 
@@ -31,24 +24,26 @@ import onnxruntime as ort
 from huggingface_hub import hf_hub_download
 
 from lerobot.robots.unitree_g1.config_unitree_g1 import UnitreeG1Config
+from lerobot.robots.unitree_g1.g1_utils import G1_29_JointIndex
 from lerobot.robots.unitree_g1.unitree_g1 import UnitreeG1
 
+logging.basicConfig(level=logging.INFO)
 logger = logging.getLogger(__name__)
 
+
 GROOT_DEFAULT_ANGLES = np.zeros(29, dtype=np.float32)
-GROOT_DEFAULT_ANGLES[[0, 6]] = -0.1  # hip pitch
-GROOT_DEFAULT_ANGLES[[3, 9]] = 0.3  # knee
-GROOT_DEFAULT_ANGLES[[4, 10]] = -0.2  # ankle pitch
+GROOT_DEFAULT_ANGLES[[0, 6]] = -0.1  # Hip pitch
+GROOT_DEFAULT_ANGLES[[3, 9]] = 0.3  # Knee
+GROOT_DEFAULT_ANGLES[[4, 10]] = -0.2  # Ankle pitch
 
 MISSING_JOINTS = []
-G1_MODEL = "g1_23"  # or "g1_29"
+G1_MODEL = "g1_23"  # Or "g1_29"
 if G1_MODEL == "g1_23":
-    MISSING_JOINTS = [12, 14, 20, 21, 27, 28]  # waist yaw/pitch, wrist pitch/yaw
-
-LOCOMOTION_ACTION_SCALE = 0.25
-
-LOCOMOTION_CONTROL_DT = 0.02
+    MISSING_JOINTS = [12, 14, 20, 21, 27, 28]  # Waist yaw/pitch, wrist pitch/yaw
 
+# Control parameters
+ACTION_SCALE = 0.25
+CONTROL_DT = 0.02  # 50Hz
 ANG_VEL_SCALE: float = 0.25
 DOF_POS_SCALE: float = 1.0
 DOF_VEL_SCALE: float = 0.05
@@ -61,12 +56,12 @@ DEFAULT_GROOT_REPO_ID = "nepyope/GR00T-WholeBodyControl_g1"
 def load_groot_policies(
     repo_id: str = DEFAULT_GROOT_REPO_ID,
 ) -> tuple[ort.InferenceSession, ort.InferenceSession]:
-    """Load GR00T dual-policy system (Balance + Walk) from Hugging Face Hub.
+    """Load GR00T dual-policy system (Balance + Walk) from the hub.
 
     Args:
         repo_id: Hugging Face Hub repository ID containing the ONNX policies.
     """
-    logger.info(f"Loading GR00T dual-policy system from Hugging Face Hub ({repo_id})...")
+    logger.info(f"Loading GR00T dual-policy system from the hub ({repo_id})...")
 
     # Download ONNX policies from Hugging Face Hub
     balance_path = hf_hub_download(
@@ -88,15 +83,7 @@ def load_groot_policies(
 
 
 class GrootLocomotionController:
-    """
-    Handles GR00T-style locomotion control for the Unitree G1 robot.
-
-    This controller manages:
-    - Dual-policy system (Balance + Walk)
-    - 29-joint observation processing
-    - 15D action output (legs + waist)
-    - Policy inference and motor command generation
-    """
+    """GR00T lower-body locomotion controller for the Unitree G1."""
 
     def __init__(self, policy_balance, policy_walk, robot, config):
         self.policy_balance = policy_balance
@@ -104,9 +91,9 @@ class GrootLocomotionController:
         self.robot = robot
         self.config = config
 
-        self.locomotion_cmd = np.array([0.0, 0.0, 0.0], dtype=np.float32)  # vx, vy, theta_dot
+        self.cmd = np.array([0.0, 0.0, 0.0], dtype=np.float32)  # vx, vy, theta_dot
 
-        # GR00T-specific state
+        # Robot state
         self.groot_qj_all = np.zeros(29, dtype=np.float32)
         self.groot_dqj_all = np.zeros(29, dtype=np.float32)
         self.groot_action = np.zeros(15, dtype=np.float32)
@@ -116,47 +103,39 @@ class GrootLocomotionController:
         self.groot_height_cmd = 0.74  # Default base height
         self.groot_orientation_cmd = np.array([0.0, 0.0, 0.0], dtype=np.float32)
 
-        # input to gr00t is 6 frames (6*86D=516)
+        # Input to GR00T is 6 frames (6*86D=516)
         for _ in range(6):
             self.groot_obs_history.append(np.zeros(86, dtype=np.float32))
 
-        # Thread management
-        self.locomotion_running = False
-        self.locomotion_thread = None
-
         logger.info("GrootLocomotionController initialized")
 
-    def groot_locomotion_run(self):
-        # get current observation
-        robot_state = self.robot.get_observation()
+    def run_step(self):
+        # Get current observation
+        obs = self.robot.get_observation()
 
-        if robot_state is None:
+        if not obs:
             return
 
-        # get command from remote controller
-        if robot_state.wireless_remote is not None:
-            self.robot.remote_controller.set(robot_state.wireless_remote)
-            if self.robot.remote_controller.button[0]:  # R1 - raise waist
-                self.groot_height_cmd += 0.001
-                self.groot_height_cmd = np.clip(self.groot_height_cmd, 0.50, 1.00)
-            if self.robot.remote_controller.button[4]:  # R2 - lower waist
-                self.groot_height_cmd -= 0.001
-                self.groot_height_cmd = np.clip(self.groot_height_cmd, 0.50, 1.00)
-        else:
-            self.robot.remote_controller.lx = 0.0
-            self.robot.remote_controller.ly = 0.0
-            self.robot.remote_controller.rx = 0.0
-            self.robot.remote_controller.ry = 0.0
+        # Get command from remote controller
+        if obs["remote.buttons"][0]:  # R1 - raise waist
+            self.groot_height_cmd += 0.001
+            self.groot_height_cmd = np.clip(self.groot_height_cmd, 0.50, 1.00)
+        if obs["remote.buttons"][4]:  # R2 - lower waist
+            self.groot_height_cmd -= 0.001
+            self.groot_height_cmd = np.clip(self.groot_height_cmd, 0.50, 1.00)
 
-        self.locomotion_cmd[0] = self.robot.remote_controller.ly  # forward/backward
-        self.locomotion_cmd[1] = self.robot.remote_controller.lx * -1  # left/right
-        self.locomotion_cmd[2] = self.robot.remote_controller.rx * -1  # rotation rate
+        self.cmd[0] = obs["remote.ly"]  # Forward/backward
+        self.cmd[1] = obs["remote.lx"] * -1  # Left/right
+        self.cmd[2] = obs["remote.rx"] * -1  # Rotation rate
 
-        for i in range(29):
-            self.groot_qj_all[i] = robot_state.motor_state[i].q
-            self.groot_dqj_all[i] = robot_state.motor_state[i].dq
+        # Get joint positions and velocities from flat dict
+        for motor in G1_29_JointIndex:
+            name = motor.name
+            idx = motor.value
+            self.groot_qj_all[idx] = obs[f"{name}.q"]
+            self.groot_dqj_all[idx] = obs[f"{name}.dq"]
 
-        # adapt observation for g1_23dof
+        # Adapt observation for g1_23dof
         for idx in MISSING_JOINTS:
             self.groot_qj_all[idx] = 0.0
             self.groot_dqj_all[idx] = 0.0
@@ -165,18 +144,18 @@ class GrootLocomotionController:
         qj_obs = self.groot_qj_all.copy()
         dqj_obs = self.groot_dqj_all.copy()
 
-        # express imu data in gravity frame of reference
-        quat = robot_state.imu_state.quaternion
-        ang_vel = np.array(robot_state.imu_state.gyroscope, dtype=np.float32)
+        # Express IMU data in gravity frame of reference
+        quat = [obs["imu.quat.w"], obs["imu.quat.x"], obs["imu.quat.y"], obs["imu.quat.z"]]
+        ang_vel = np.array([obs["imu.gyro.x"], obs["imu.gyro.y"], obs["imu.gyro.z"]], dtype=np.float32)
         gravity_orientation = self.robot.get_gravity_orientation(quat)
 
-        # scale joint positions and velocities before policy inference
+        # Scale joint positions and velocities before policy inference
         qj_obs = (qj_obs - GROOT_DEFAULT_ANGLES) * DOF_POS_SCALE
         dqj_obs = dqj_obs * DOF_VEL_SCALE
         ang_vel_scaled = ang_vel * ANG_VEL_SCALE
 
-        # build single frame observation
-        self.groot_obs_single[:3] = self.locomotion_cmd * np.array(CMD_SCALE)
+        # Build single frame observation
+        self.groot_obs_single[:3] = self.cmd * np.array(CMD_SCALE)
         self.groot_obs_single[3] = self.groot_height_cmd
         self.groot_obs_single[4:7] = self.groot_orientation_cmd
         self.groot_obs_single[7:10] = ang_vel_scaled
@@ -194,113 +173,76 @@ class GrootLocomotionController:
             end_idx = start_idx + 86
             self.groot_obs_stacked[start_idx:end_idx] = obs_frame
 
-        # Run policy inference (ONNX) with 516D stacked observation
-
-        cmd_magnitude = np.linalg.norm(self.locomotion_cmd)
-
+        cmd_magnitude = np.linalg.norm(self.cmd)
         selected_policy = (
             self.policy_balance if cmd_magnitude < 0.05 else self.policy_walk
-        )  # balance/standing policy for small commands, walking policy for movement commands
+        )  # Balance/standing policy for small commands, walking policy for movement commands
 
-        # run policy inference
+        # Run policy inference
         ort_inputs = {selected_policy.get_inputs()[0].name: np.expand_dims(self.groot_obs_stacked, axis=0)}
         ort_outs = selected_policy.run(None, ort_inputs)
         self.groot_action = ort_outs[0].squeeze()
 
-        # transform action back to target joint positions
-        target_dof_pos_15 = GROOT_DEFAULT_ANGLES[:15] + self.groot_action * LOCOMOTION_ACTION_SCALE
+        # Transform action back to target joint positions
+        target_dof_pos_15 = GROOT_DEFAULT_ANGLES[:15] + self.groot_action * ACTION_SCALE
 
-        # command motors
+        # Build action dict (only first 15 joints for GR00T)
+        action_dict = {}
         for i in range(15):
-            motor_idx = i
-            self.robot.msg.motor_cmd[motor_idx].q = target_dof_pos_15[i]
-            self.robot.msg.motor_cmd[motor_idx].qd = 0
-            self.robot.msg.motor_cmd[motor_idx].kp = self.robot.kp[motor_idx]
-            self.robot.msg.motor_cmd[motor_idx].kd = self.robot.kd[motor_idx]
-            self.robot.msg.motor_cmd[motor_idx].tau = 0
+            motor_name = G1_29_JointIndex(i).name
+            action_dict[f"{motor_name}.q"] = float(target_dof_pos_15[i])
 
-        # adapt action for g1_23dof
+        # Zero out missing joints for g1_23dof
         for joint_idx in MISSING_JOINTS:
-            self.robot.msg.motor_cmd[joint_idx].q = 0.0
-            self.robot.msg.motor_cmd[joint_idx].qd = 0
-            self.robot.msg.motor_cmd[joint_idx].kp = self.robot.kp[joint_idx]
-            self.robot.msg.motor_cmd[joint_idx].kd = self.robot.kd[joint_idx]
-            self.robot.msg.motor_cmd[joint_idx].tau = 0
+            motor_name = G1_29_JointIndex(joint_idx).name
+            action_dict[f"{motor_name}.q"] = 0.0
 
-        # send action to robot
-        self.robot.send_action(self.robot.msg)
+        # Send action to robot
+        self.robot.send_action(action_dict)
 
-    def _locomotion_thread_loop(self):
-        """Background thread that runs the locomotion policy at specified rate."""
-        logger.info("Locomotion thread started")
-        while self.locomotion_running:
+
+def run(repo_id: str = DEFAULT_GROOT_REPO_ID) -> None:
+    """Main function to run the GR00T locomotion controller.
+
+    Args:
+        repo_id: Hugging Face Hub repository ID for GR00T policies.
+    """
+    # Load policies
+    policy_balance, policy_walk = load_groot_policies(repo_id=repo_id)
+
+    # Initialize robot
+    config = UnitreeG1Config()
+    robot = UnitreeG1(config)
+
+    robot.connect()
+
+    # Initialize gr00T locomotion controller
+    groot_controller = GrootLocomotionController(
+        policy_balance=policy_balance,
+        policy_walk=policy_walk,
+        robot=robot,
+        config=config,
+    )
+
+    try:
+        robot.reset(CONTROL_DT, GROOT_DEFAULT_ANGLES)
+
+        logger.info("Use joystick: LY=fwd/back, LX=left/right, RX=rotate, R1=raise waist, R2=lower waist")
+        logger.info("Press Ctrl+C to stop")
+
+        # Run step
+        while not robot._shutdown_event.is_set():
             start_time = time.time()
-            try:
-                self.groot_locomotion_run()
-            except Exception as e:
-                logger.error(f"Error in locomotion loop: {e}")
-
-            # Sleep to maintain control rate
+            groot_controller.run_step()
             elapsed = time.time() - start_time
-            sleep_time = max(0, LOCOMOTION_CONTROL_DT - elapsed)
+            sleep_time = max(0, CONTROL_DT - elapsed)
             time.sleep(sleep_time)
-        logger.info("Locomotion thread stopped")
-
-    def start_locomotion_thread(self):
-        if self.locomotion_running:
-            logger.warning("Locomotion thread already running")
-            return
-
-        logger.info("Starting locomotion control thread...")
-        self.locomotion_running = True
-        self.locomotion_thread = threading.Thread(target=self._locomotion_thread_loop, daemon=True)
-        self.locomotion_thread.start()
-
-        logger.info("Locomotion control thread started!")
-
-    def stop_locomotion_thread(self):
-        if not self.locomotion_running:
-            return
-
-        logger.info("Stopping locomotion control thread...")
-        self.locomotion_running = False
-        if self.locomotion_thread:
-            self.locomotion_thread.join(timeout=2.0)
-        logger.info("Locomotion control thread stopped")
-
-    def reset_robot(self):
-        """Move robot legs to default standing position over 2 seconds (arms are not moved)."""
-        total_time = 3.0
-        num_step = int(total_time / self.robot.control_dt)
-
-        # Only control legs, not arms (first 12 joints)
-        default_pos = GROOT_DEFAULT_ANGLES  # First 12 values are leg angles
-        dof_size = len(default_pos)
-
-        # Get current lowstate
-        robot_state = self.robot.get_observation()
-
-        # Record the current leg positions
-        init_dof_pos = np.zeros(dof_size, dtype=np.float32)
-        for i in range(dof_size):
-            init_dof_pos[i] = robot_state.motor_state[i].q
-
-        # Move legs to default pos
-        for i in range(num_step):
-            alpha = i / num_step
-            for motor_idx in range(dof_size):
-                target_pos = default_pos[motor_idx]
-                self.robot.msg.motor_cmd[motor_idx].q = (
-                    init_dof_pos[motor_idx] * (1 - alpha) + target_pos * alpha
-                )
-                self.robot.msg.motor_cmd[motor_idx].qd = 0
-                self.robot.msg.motor_cmd[motor_idx].kp = self.robot.kp[motor_idx]
-                self.robot.msg.motor_cmd[motor_idx].kd = self.robot.kd[motor_idx]
-                self.robot.msg.motor_cmd[motor_idx].tau = 0
-            self.robot.msg.crc = self.robot.crc.Crc(self.robot.msg)
-            self.robot.lowcmd_publisher.Write(self.robot.msg)
-            time.sleep(self.robot.control_dt)
-        logger.info("Reached default position (legs only)")
+    except KeyboardInterrupt:
+        logger.info("Stopping locomotion...")
+    finally:
+        if robot.is_connected:
+            robot.disconnect()
+        logger.info("Done!")
 
 
 if __name__ == "__main__":
@@ -313,35 +255,4 @@ if __name__ == "__main__":
     )
     args = parser.parse_args()
 
-    # load policies
-    policy_balance, policy_walk = load_groot_policies(repo_id=args.repo_id)
-
-    # initialize robot
-    config = UnitreeG1Config()
-    robot = UnitreeG1(config)
-
-    # initialize gr00t locomotion controller
-    groot_controller = GrootLocomotionController(
-        policy_balance=policy_balance,
-        policy_walk=policy_walk,
-        robot=robot,
-        config=config,
-    )
-
-    # reset legs and start locomotion thread
-    try:
-        groot_controller.reset_robot()
-        groot_controller.start_locomotion_thread()
-
-        # log status
-        logger.info("Robot initialized with GR00T locomotion policies")
-        logger.info("Locomotion controller running in background thread")
-        logger.info("Press Ctrl+C to stop")
-
-        # keep robot alive
-        while True:
-            time.sleep(1.0)
-    except KeyboardInterrupt:
-        print("\nStopping locomotion...")
-        groot_controller.stop_locomotion_thread()
-        print("Done!")
+    run(repo_id=args.repo_id)

examples/unitree_g1/holosoma_locomotion.py

@@ -0,0 +1,264 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import json
+import logging
+import time
+
+import numpy as np
+import onnx
+import onnxruntime as ort
+from huggingface_hub import hf_hub_download
+
+from lerobot.robots.unitree_g1.config_unitree_g1 import UnitreeG1Config
+from lerobot.robots.unitree_g1.g1_utils import G1_29_JointIndex
+from lerobot.robots.unitree_g1.unitree_g1 import UnitreeG1
+
+logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger(__name__)
+
+DEFAULT_ANGLES = np.zeros(29, dtype=np.float32)
+DEFAULT_ANGLES[[0, 6]] = -0.312  # Hip pitch
+DEFAULT_ANGLES[[3, 9]] = 0.669  # Knee
+DEFAULT_ANGLES[[4, 10]] = -0.363  # Ankle pitch
+DEFAULT_ANGLES[[15, 22]] = 0.2  # Shoulder pitch
+DEFAULT_ANGLES[16] = 0.2  # Left shoulder roll
+DEFAULT_ANGLES[23] = -0.2  # Right shoulder roll
+DEFAULT_ANGLES[[18, 25]] = 0.6  # Elbow
+
+MISSING_JOINTS = []
+G1_MODEL = "g1_23"  # Or "g1_29"
+if G1_MODEL == "g1_23":
+    MISSING_JOINTS = [12, 14, 20, 21, 27, 28]  # Waist yaw/pitch, wrist pitch/yaw
+
+# Control parameters
+ACTION_SCALE = 0.25
+CONTROL_DT = 0.02  # 50Hz
+ANG_VEL_SCALE = 0.25
+DOF_POS_SCALE = 1.0
+DOF_VEL_SCALE = 0.05
+GAIT_PERIOD = 1.0
+
+
+DEFAULT_HOLOSOMA_REPO_ID = "nepyope/holosoma_locomotion"
+
+# Policy filename mapping
+POLICY_FILES = {
+    "fastsac": "fastsac_g1_29dof.onnx",
+    "ppo": "ppo_g1_29dof.onnx",
+}
+
+
+def load_policy(
+    repo_id: str = DEFAULT_HOLOSOMA_REPO_ID,
+    policy_type: str = "fastsac",
+) -> tuple[ort.InferenceSession, np.ndarray, np.ndarray]:
+    """Load Holosoma locomotion policy and extract KP/KD from metadata.
+
+    Args:
+        repo_id: Hugging Face Hub repo ID
+        policy_type: Either "fastsac" (default) or "ppo"
+
+    Returns:
+        (policy, kp, kd) tuple
+    """
+    if policy_type not in POLICY_FILES:
+        raise ValueError(f"Unknown policy type: {policy_type}. Choose from: {list(POLICY_FILES.keys())}")
+
+    filename = POLICY_FILES[policy_type]
+    logger.info(f"Loading {policy_type.upper()} policy from: {repo_id}/{filename}")
+    policy_path = hf_hub_download(repo_id=repo_id, filename=filename)
+
+    policy = ort.InferenceSession(policy_path)
+    logger.info(f"Policy loaded: {policy.get_inputs()[0].shape} → {policy.get_outputs()[0].shape}")
+
+    # Extract KP/KD from ONNX metadata
+    model = onnx.load(policy_path)
+    metadata = {prop.key: prop.value for prop in model.metadata_props}
+
+    if "kp" not in metadata or "kd" not in metadata:
+        raise ValueError("ONNX model must contain 'kp' and 'kd' in metadata")
+
+    kp = np.array(json.loads(metadata["kp"]), dtype=np.float32)
+    kd = np.array(json.loads(metadata["kd"]), dtype=np.float32)
+    logger.info(f"Loaded KP/KD from ONNX ({len(kp)} joints)")
+
+    return policy, kp, kd
+
+
+class HolosomaLocomotionController:
+    """Holosoma whole-body locomotion controller for Unitree G1."""
+
+    def __init__(self, policy, robot, kp: np.ndarray, kd: np.ndarray):
+        self.policy = policy
+        self.robot = robot
+
+        # Override robot's PD gains with policy gains
+        self.robot.kp = kp
+        self.robot.kd = kd
+
+        self.cmd = np.zeros(3, dtype=np.float32)
+
+        # Robot state
+        self.qj = np.zeros(29, dtype=np.float32)
+        self.dqj = np.zeros(29, dtype=np.float32)
+        self.obs = np.zeros(100, dtype=np.float32)
+        self.last_action = np.zeros(29, dtype=np.float32)
+
+        # Gait phase
+        self.phase = np.array([[0.0, np.pi]], dtype=np.float32)
+        self.phase_dt = 2 * np.pi / ((1.0 / CONTROL_DT) * GAIT_PERIOD)
+        self.is_standing = True
+
+    def run_step(self):
+        # Get current observation
+        obs = self.robot.get_observation()
+
+        if not obs:
+            return
+
+        # Get command from remote controller
+        ly = obs["remote.ly"] if abs(obs["remote.ly"]) > 0.1 else 0.0
+        lx = obs["remote.lx"] if abs(obs["remote.lx"]) > 0.1 else 0.0
+        rx = obs["remote.rx"] if abs(obs["remote.rx"]) > 0.1 else 0.0
+        self.cmd[:] = [ly, -lx, -rx]
+
+        # Get joint positions and velocities
+        for motor in G1_29_JointIndex:
+            name = motor.name
+            idx = motor.value
+            self.qj[idx] = obs[f"{name}.q"]
+            self.dqj[idx] = obs[f"{name}.dq"]
+
+        # Adapt observation for g1_23dof
+        for idx in MISSING_JOINTS:
+            self.qj[idx] = 0.0
+            self.dqj[idx] = 0.0
+
+        # Express IMU data in gravity frame of reference
+        quat = [obs["imu.quat.w"], obs["imu.quat.x"], obs["imu.quat.y"], obs["imu.quat.z"]]
+        ang_vel = np.array([obs["imu.gyro.x"], obs["imu.gyro.y"], obs["imu.gyro.z"]], dtype=np.float32)
+        gravity = self.robot.get_gravity_orientation(quat)
+
+        # Scale joint positions and velocities before policy inference
+        qj_obs = (self.qj - DEFAULT_ANGLES) * DOF_POS_SCALE
+        dqj_obs = self.dqj * DOF_VEL_SCALE
+        ang_vel_s = ang_vel * ANG_VEL_SCALE
+
+        # Update gait phase
+        if np.linalg.norm(self.cmd[:2]) < 0.01 and abs(self.cmd[2]) < 0.01:
+            self.phase[0, :] = np.pi
+            self.is_standing = True
+        elif self.is_standing:
+            self.phase = np.array([[0.0, np.pi]], dtype=np.float32)
+            self.is_standing = False
+        else:
+            self.phase = np.fmod(self.phase + self.phase_dt + np.pi, 2 * np.pi) - np.pi
+
+        sin_ph = np.sin(self.phase[0])
+        cos_ph = np.cos(self.phase[0])
+
+        # Build observations
+        self.obs[0:29] = self.last_action
+        self.obs[29:32] = ang_vel_s
+        self.obs[32] = self.cmd[2]
+        self.obs[33:35] = self.cmd[:2]
+        self.obs[35:37] = cos_ph
+        self.obs[37:66] = qj_obs
+        self.obs[66:95] = dqj_obs
+        self.obs[95:98] = gravity
+        self.obs[98:100] = sin_ph
+
+        # Run policy inference
+        ort_in = {self.policy.get_inputs()[0].name: self.obs.reshape(1, -1).astype(np.float32)}
+        raw_action = self.policy.run(None, ort_in)[0].squeeze()
+        action = np.clip(raw_action, -100.0, 100.0)
+        self.last_action = action.copy()
+
+        # Transform action back to target joint positions
+        target = DEFAULT_ANGLES + action * ACTION_SCALE
+
+        # Build action dict
+        action_dict = {}
+        for motor in G1_29_JointIndex:
+            action_dict[f"{motor.name}.q"] = float(target[motor.value])
+
+        # Zero out missing joints for g1_23dof
+        for joint_idx in MISSING_JOINTS:
+            motor_name = G1_29_JointIndex(joint_idx).name
+            action_dict[f"{motor_name}.q"] = 0.0
+
+        # Send action to robot
+        self.robot.send_action(action_dict)
+
+
+def run(repo_id: str = DEFAULT_HOLOSOMA_REPO_ID, policy_type: str = "fastsac") -> None:
+    """Main function to run the Holosoma locomotion controller.
+
+    Args:
+        repo_id: Hugging Face Hub repository ID for Holosoma policies.
+        policy_type: Policy type to use ('fastsac' or 'ppo').
+    """
+    # Load policy and gains
+    policy, kp, kd = load_policy(repo_id=repo_id, policy_type=policy_type)
+
+    # Initialize robot
+    config = UnitreeG1Config()
+    robot = UnitreeG1(config)
+    robot.connect()
+
+    holosoma_controller = HolosomaLocomotionController(policy, robot, kp, kd)
+
+    try:
+        robot.reset(CONTROL_DT, DEFAULT_ANGLES)
+
+        logger.info("Use joystick: LY=fwd/back, LX=left/right, RX=rotate")
+        logger.info("Press Ctrl+C to stop")
+
+        # Run step
+        while not robot._shutdown_event.is_set():
+            start_time = time.time()
+            holosoma_controller.run_step()
+            elapsed = time.time() - start_time
+            sleep_time = max(0, CONTROL_DT - elapsed)
+            time.sleep(sleep_time)
+    except KeyboardInterrupt:
+        logger.info("Stopping locomotion...")
+    finally:
+        if robot.is_connected:
+            robot.disconnect()
+        logger.info("Done!")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Holosoma Locomotion Controller for Unitree G1")
+    parser.add_argument(
+        "--repo-id",
+        type=str,
+        default=DEFAULT_HOLOSOMA_REPO_ID,
+        help=f"Hugging Face Hub repo ID for Holosoma policies (default: {DEFAULT_HOLOSOMA_REPO_ID})",
+    )
+    parser.add_argument(
+        "--policy",
+        type=str,
+        choices=["fastsac", "ppo"],
+        default="fastsac",
+        help="Policy type to use: 'fastsac' (default) or 'ppo'",
+    )
+    args = parser.parse_args()
+
+    run(repo_id=args.repo_id, policy_type=args.policy)

pyproject.toml

@@ -27,7 +27,7 @@ discord = "https://discord.gg/s3KuuzsPFb"
 name = "lerobot"
 version = "0.4.3"
 description = "🤗 LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch"
-readme = "README.md"
+dynamic = ["readme"]
 license = { text = "Apache-2.0" }
 requires-python = ">=3.10"
 authors = [
@@ -74,7 +74,7 @@ dependencies = [
     "packaging>=24.2,<26.0",
     "pynput>=1.7.7,<1.9.0",
     "pyserial>=3.5,<4.0",
-    "wandb>=0.20.0,<0.22.0", # TODO: Bumb dependency (compatible with protobuf)
+    "wandb>=0.24.0,<0.25.0",
 
     "torch>=2.2.1,<2.8.0", # TODO: Bumb dependency
     "torchcodec>=0.2.1,<0.6.0; sys_platform != 'win32' and (sys_platform != 'linux' or (platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')) and (sys_platform != 'darwin' or platform_machine != 'x86_64')", # TODO: Bumb dependency
@@ -97,7 +97,7 @@ dependencies = [
 pygame-dep = ["pygame>=2.5.1,<2.7.0"]
 placo-dep = ["placo>=0.9.6,<0.10.0"]
 transformers-dep = ["transformers>=4.57.1,<5.0.0"]
-grpcio-dep = ["grpcio==1.73.1", "protobuf==6.31.0"] # TODO: Bumb dependency (compatible with wandb)
+grpcio-dep = ["grpcio==1.73.1", "protobuf>=6.31.1,<6.32.0"]
 
 # Motors
 feetech = ["feetech-servo-sdk>=1.0.0,<2.0.0"]
@@ -109,9 +109,9 @@ hopejr = ["lerobot[feetech]", "lerobot[pygame-dep]"]
 lekiwi = ["lerobot[feetech]", "pyzmq>=26.2.1,<28.0.0"]
 unitree_g1 = [
     "pyzmq>=26.2.1,<28.0.0",
-    "onnxruntime>=1.16.0"
+    "onnxruntime>=1.16.0,<2.0.0"
 ]
-reachy2 = ["reachy2_sdk>=1.0.14,<1.1.0"]
+reachy2 = ["reachy2_sdk>=1.0.15,<1.1.0"]
 kinematics = ["lerobot[placo-dep]"]
 intelrealsense = [
     "pyrealsense2>=2.55.1.6486,<2.57.0 ; sys_platform != 'darwin'",
@@ -127,7 +127,7 @@ wallx = [
     "torchdiffeq==0.2.5",
     "qwen_vl_utils==0.0.11"
 ]
-pi = ["transformers @ git+https://github.com/huggingface/transformers.git@fix/lerobot_openpi"]
+pi = ["transformers @ git+https://github.com/huggingface/transformers.git@fix/lerobot_openpi", "scipy>=1.10.1,<1.15"]
 smolvla = ["lerobot[transformers-dep]", "num2words>=0.5.14,<0.6.0", "accelerate>=1.7.0,<2.0.0", "safetensors>=0.4.3,<1.0.0"]
 groot = [
     "lerobot[transformers-dep]",
@@ -140,12 +140,13 @@ groot = [
     "ninja>=1.11.1,<2.0.0",
     "flash-attn>=2.5.9,<3.0.0 ; sys_platform != 'darwin'"
 ]
-sarm = ["lerobot[transformers-dep]", "faker>=33.0.0,<35.0.0", "matplotlib>=3.10.3,<4.0.0", "qwen-vl-utils>=0.0.14"]
+sarm = ["lerobot[transformers-dep]", "faker>=33.0.0,<35.0.0", "matplotlib>=3.10.3,<4.0.0", "qwen-vl-utils>=0.0.14,<0.1.0"]
 xvla = ["lerobot[transformers-dep]"]
 hilserl = ["lerobot[transformers-dep]", "gym-hil>=0.1.13,<0.2.0", "lerobot[grpcio-dep]", "lerobot[placo-dep]"]
 
 # Features
 async = ["lerobot[grpcio-dep]", "matplotlib>=3.10.3,<4.0.0"]
+peft = ["lerobot[transformers-dep]", "peft>=0.18.0,<1.0.0"]
 
 # Development
 dev = ["pre-commit>=3.7.0,<5.0.0", "debugpy>=1.8.1,<1.9.0", "lerobot[grpcio-dep]", "grpcio-tools==1.73.1", "mypy>=1.19.1"]
@@ -182,7 +183,8 @@ all = [
     "lerobot[phone]",
     "lerobot[libero]",
     "lerobot[metaworld]",
-    "lerobot[sarm]"
+    "lerobot[sarm]",
+    "lerobot[peft]",
 ]
 
 [project.scripts]
@@ -195,6 +197,7 @@ lerobot-setup-motors="lerobot.scripts.lerobot_setup_motors:main"
 lerobot-teleoperate="lerobot.scripts.lerobot_teleoperate:main"
 lerobot-eval="lerobot.scripts.lerobot_eval:main"
 lerobot-train="lerobot.scripts.lerobot_train:main"
+lerobot-train-tokenizer="lerobot.scripts.lerobot_train_tokenizer:main"
 lerobot-dataset-viz="lerobot.scripts.lerobot_dataset_viz:main"
 lerobot-info="lerobot.scripts.lerobot_info:main"
 lerobot-find-joint-limits="lerobot.scripts.lerobot_find_joint_limits:main"
@@ -417,6 +420,10 @@ conflicts = [
         { extra = "wallx" },
         { extra = "libero" },
     ],
+    [
+        { extra = "wallx" },
+        { extra = "peft" },
+    ],
     [
         { extra = "wallx" },
         { extra = "all" },
@@ -450,6 +457,10 @@ conflicts = [
         { extra = "pi" },
         { extra = "libero" },
     ],
+    [
+        { extra = "pi" },
+        { extra = "peft" },
+    ],
     [
         { extra = "pi" },
         { extra = "all" },

setup.py

@@ -0,0 +1,72 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from setuptools import setup
+
+
+def get_version_from_toml() -> str:
+    """Return the project's version string parsed from `pyproject.toml`.
+
+    The function scans `pyproject.toml` line-by-line looking for a line
+    that starts with ``version`` (for example: ``version = "1.2.3"``)
+    and returns the value without surrounding quotes. If no such line is
+    found a :class:`ValueError` is raised.
+
+    Returns:
+        The version string from `pyproject.toml` (e.g. ``"1.2.3"`` ->
+        ``1.2.3``).
+    """
+
+    version = None
+    with open("pyproject.toml", encoding="utf-8") as f:
+        for line in f:
+            if line.strip().startswith("version"):
+                version = line.split("=")[1].strip().strip('"')
+                break
+    if version is None:
+        raise ValueError("Version not found in pyproject.toml")
+    return version
+
+
+def read_long_description() -> str:
+    """Read and return the project's long description for setup.
+
+    This function reads `README.md` and replaces image links that point
+    to the local `./media/` directory with absolute raw GitHub URLs that
+    reference the release tag corresponding to the version parsed from
+    `pyproject.toml` (for example, ``v1.2.3``). The modified README
+    content is returned as a string suitable for passing to
+    ``setuptools.setup(long_description=...)``.
+
+    Returns:
+        The README content with rewritten media links.
+    """
+
+    with open("README.md", encoding="utf-8") as f:
+        content = f.read()
+
+    version = get_version_from_toml()
+    git_tag = f"v{version}"
+
+    base_raw_url = f"https://raw.githubusercontent.com/huggingface/lerobot/{git_tag}/"
+    content = content.replace('src="./media/', f'src="{base_raw_url}media/')
+
+    return content
+
+
+setup(
+    long_description=read_long_description(),
+    long_description_content_type="text/markdown",
+)

src/lerobot/async_inference/constants.py

@@ -26,4 +26,4 @@ DEFAULT_OBS_QUEUE_TIMEOUT = 2
 SUPPORTED_POLICIES = ["act", "smolvla", "diffusion", "tdmpc", "vqbet", "pi0", "pi05"]
 
 # TODO: Add all other robots
-SUPPORTED_ROBOTS = ["so100_follower", "so101_follower", "bi_so100_follower", "omx_follower"]
+SUPPORTED_ROBOTS = ["so100_follower", "so101_follower", "bi_so_follower", "omx_follower"]

src/lerobot/async_inference/robot_client.py

@@ -40,7 +40,6 @@ from collections.abc import Callable
 from dataclasses import asdict
 from pprint import pformat
 from queue import Queue
-from typing import Any
 
 import draccus
 import grpc
@@ -48,15 +47,15 @@ import torch
 
 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig  # noqa: F401
 from lerobot.cameras.realsense.configuration_realsense import RealSenseCameraConfig  # noqa: F401
+from lerobot.processor import RobotAction
 from lerobot.robots import (  # noqa: F401
     Robot,
     RobotConfig,
-    bi_so100_follower,
+    bi_so_follower,
     koch_follower,
     make_robot_from_config,
     omx_follower,
-    so100_follower,
-    so101_follower,
+    so_follower,
 )
 from lerobot.transport import (
     services_pb2,  # type: ignore
@@ -352,7 +351,7 @@ class RobotClient:
         action = {key: action_tensor[i].item() for i, key in enumerate(self.robot.action_features)}
         return action
 
-    def control_loop_action(self, verbose: bool = False) -> dict[str, Any]:
+    def control_loop_action(self, verbose: bool = False) -> RobotAction:
         """Reading and performing actions in local queue"""
 
         # Lock only for queue operations

src/lerobot/cameras/reachy2_camera/configuration_reachy2_camera.py

@@ -35,18 +35,19 @@ class Reachy2CameraConfig(CameraConfig):
         name="teleop",
         image_type="left",
         ip_address="192.168.0.200",  # IP address of the robot
-        fps=15,
+        port=50065,  # Port of the camera server
         width=640,
         height=480,
+        fps=30,  # Not configurable for Reachy 2 cameras
         color_mode=ColorMode.RGB,
-    )  # Left teleop camera, 640x480 @ 15FPS
+    )  # Left teleop camera, 640x480 @ 30FPS
     ```
 
     Attributes:
         name: Name of the camera device. Can be "teleop" or "depth".
         image_type: Type of image stream. For "teleop" camera, can be "left" or "right".
                     For "depth" camera, can be "rgb" or "depth". (depth is not supported yet)
-        fps: Requested frames per second for the color stream.
+        fps: Requested frames per second for the color stream. Not configurable for Reachy 2 cameras.
         width: Requested frame width in pixels for the color stream.
         height: Requested frame height in pixels for the color stream.
         color_mode: Color mode for image output (RGB or BGR). Defaults to RGB.
@@ -62,7 +63,6 @@ class Reachy2CameraConfig(CameraConfig):
     color_mode: ColorMode = ColorMode.RGB
     ip_address: str | None = "localhost"
     port: int = 50065
-    # use_depth: bool = False
 
     def __post_init__(self) -> None:
         if self.name not in ["teleop", "depth"]:

src/lerobot/cameras/reachy2_camera/reachy2_camera.py

@@ -16,12 +16,13 @@
 Provides the Reachy2Camera class for capturing frames from Reachy 2 cameras using Reachy 2's CameraManager.
 """
 
+from __future__ import annotations
+
 import logging
 import os
 import platform
 import time
-from threading import Event, Lock, Thread
-from typing import Any
+from typing import TYPE_CHECKING, Any
 
 from numpy.typing import NDArray  # type: ignore  # TODO: add type stubs for numpy.typing
 
@@ -30,10 +31,19 @@ if platform.system() == "Windows" and "OPENCV_VIDEOIO_MSMF_ENABLE_HW_TRANSFORMS"
     os.environ["OPENCV_VIDEOIO_MSMF_ENABLE_HW_TRANSFORMS"] = "0"
 import cv2  # type: ignore  # TODO: add type stubs for OpenCV
 import numpy as np  # type: ignore  # TODO: add type stubs for numpy
-from reachy2_sdk.media.camera import CameraView  # type: ignore  # TODO: add type stubs for reachy2_sdk
-from reachy2_sdk.media.camera_manager import (  # type: ignore  # TODO: add type stubs for reachy2_sdk
-    CameraManager,
-)
+
+from lerobot.utils.import_utils import _reachy2_sdk_available
+
+if TYPE_CHECKING or _reachy2_sdk_available:
+    from reachy2_sdk.media.camera import CameraView
+    from reachy2_sdk.media.camera_manager import CameraManager
+else:
+    CameraManager = None
+
+    class CameraView:
+        LEFT = 0
+        RIGHT = 1
+
 
 from lerobot.utils.errors import DeviceNotConnectedError
 
@@ -69,17 +79,10 @@ class Reachy2Camera(Camera):
 
         self.config = config
 
-        self.fps = config.fps
         self.color_mode = config.color_mode
 
         self.cam_manager: CameraManager | None = None
 
-        self.thread: Thread | None = None
-        self.stop_event: Event | None = None
-        self.frame_lock: Lock = Lock()
-        self.latest_frame: NDArray[Any] | None = None
-        self.new_frame_event: Event = Event()
-
     def __str__(self) -> str:
         return f"{self.__class__.__name__}({self.config.name}, {self.config.image_type})"
 
@@ -100,44 +103,23 @@ class Reachy2Camera(Camera):
     def connect(self, warmup: bool = True) -> None:
         """
         Connects to the Reachy2 CameraManager as specified in the configuration.
+
+        Raises:
+            DeviceNotConnectedError: If the camera is not connected.
         """
         self.cam_manager = CameraManager(host=self.config.ip_address, port=self.config.port)
+        if self.cam_manager is None:
+            raise DeviceNotConnectedError(f"Could not connect to {self}.")
         self.cam_manager.initialize_cameras()
 
         logger.info(f"{self} connected.")
 
     @staticmethod
-    def find_cameras(ip_address: str = "localhost", port: int = 50065) -> list[dict[str, Any]]:
+    def find_cameras() -> list[dict[str, Any]]:
         """
-        Detects available Reachy 2 cameras.
-
-        Returns:
-            List[Dict[str, Any]]: A list of dictionaries,
-            where each dictionary contains 'name', 'stereo',
-            and the default profile properties (width, height, fps).
+        Detection not implemented for Reachy2 cameras.
         """
-        initialized_cameras = []
-        camera_manager = CameraManager(host=ip_address, port=port)
-
-        for camera in [camera_manager.teleop, camera_manager.depth]:
-            if camera is None:
-                continue
-
-            height, width, _, _, _, _, _ = camera.get_parameters()
-
-            camera_info = {
-                "name": camera._cam_info.name,
-                "stereo": camera._cam_info.stereo,
-                "default_profile": {
-                    "width": width,
-                    "height": height,
-                    "fps": 30,
-                },
-            }
-            initialized_cameras.append(camera_info)
-
-        camera_manager.disconnect()
-        return initialized_cameras
+        raise NotImplementedError("Camera detection is not implemented for Reachy2 cameras.")
 
     def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
         """
@@ -155,95 +137,49 @@ class Reachy2Camera(Camera):
                        (height, width, channels), using the specified or default
                        color mode and applying any configured rotation.
         """
+        start_time = time.perf_counter()
+
         if not self.is_connected:
             raise DeviceNotConnectedError(f"{self} is not connected.")
 
-        start_time = time.perf_counter()
+        if self.cam_manager is None:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
 
         frame: NDArray[Any] = np.empty((0, 0, 3), dtype=np.uint8)
 
-        if self.cam_manager is None:
-            raise DeviceNotConnectedError(f"{self} is not connected.")
+        if self.config.name == "teleop" and hasattr(self.cam_manager, "teleop"):
+            if self.config.image_type == "left":
+                frame = self.cam_manager.teleop.get_frame(
+                    CameraView.LEFT, size=(self.config.width, self.config.height)
+                )[0]
+            elif self.config.image_type == "right":
+                frame = self.cam_manager.teleop.get_frame(
+                    CameraView.RIGHT, size=(self.config.width, self.config.height)
+                )[0]
+        elif self.config.name == "depth" and hasattr(self.cam_manager, "depth"):
+            if self.config.image_type == "depth":
+                frame = self.cam_manager.depth.get_depth_frame()[0]
+            elif self.config.image_type == "rgb":
+                frame = self.cam_manager.depth.get_frame(size=(self.config.width, self.config.height))[0]
         else:
-            if self.config.name == "teleop" and hasattr(self.cam_manager, "teleop"):
-                if self.config.image_type == "left":
-                    frame = self.cam_manager.teleop.get_frame(CameraView.LEFT, size=(640, 480))[0]
-                elif self.config.image_type == "right":
-                    frame = self.cam_manager.teleop.get_frame(CameraView.RIGHT, size=(640, 480))[0]
-            elif self.config.name == "depth" and hasattr(self.cam_manager, "depth"):
-                if self.config.image_type == "depth":
-                    frame = self.cam_manager.depth.get_depth_frame()[0]
-                elif self.config.image_type == "rgb":
-                    frame = self.cam_manager.depth.get_frame(size=(640, 480))[0]
+            raise ValueError(f"Invalid camera name '{self.config.name}'. Expected 'teleop' or 'depth'.")
 
-            if frame is None:
-                return np.empty((0, 0, 3), dtype=np.uint8)
+        if frame is None:
+            return np.empty((0, 0, 3), dtype=np.uint8)
 
-            if self.config.color_mode == "rgb":
-                frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+        if self.config.color_mode == "rgb":
+            frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
 
         read_duration_ms = (time.perf_counter() - start_time) * 1e3
         logger.debug(f"{self} read took: {read_duration_ms:.1f}ms")
 
         return frame
 
-    def _read_loop(self) -> None:
-        """
-        Internal loop run by the background thread for asynchronous reading.
-
-        On each iteration:
-        1. Reads a color frame
-        2. Stores result in latest_frame (thread-safe)
-        3. Sets new_frame_event to notify listeners
-
-        Stops on DeviceNotConnectedError, logs other errors and continues.
-        """
-        if self.stop_event is None:
-            raise RuntimeError(f"{self}: stop_event is not initialized before starting read loop.")
-
-        while not self.stop_event.is_set():
-            try:
-                color_image = self.read()
-
-                with self.frame_lock:
-                    self.latest_frame = color_image
-                self.new_frame_event.set()
-
-            except DeviceNotConnectedError:
-                break
-            except Exception as e:
-                logger.warning(f"Error reading frame in background thread for {self}: {e}")
-
-    def _start_read_thread(self) -> None:
-        """Starts or restarts the background read thread if it's not running."""
-        if self.thread is not None and self.thread.is_alive():
-            self.thread.join(timeout=0.1)
-        if self.stop_event is not None:
-            self.stop_event.set()
-
-        self.stop_event = Event()
-        self.thread = Thread(target=self._read_loop, args=(), name=f"{self}_read_loop")
-        self.thread.daemon = True
-        self.thread.start()
-
-    def _stop_read_thread(self) -> None:
-        """Signals the background read thread to stop and waits for it to join."""
-        if self.stop_event is not None:
-            self.stop_event.set()
-
-        if self.thread is not None and self.thread.is_alive():
-            self.thread.join(timeout=2.0)
-
-        self.thread = None
-        self.stop_event = None
-
     def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
         """
-        Reads the latest available frame asynchronously.
+        Reads the latest available frame.
 
-        This method retrieves the most recent frame captured by the background
-        read thread. It does not block waiting for the camera hardware directly,
-        but may wait up to timeout_ms for the background thread to provide a frame.
+        This method retrieves the most recent frame available in Reachy 2's low-level software.
 
         Args:
             timeout_ms (float): Maximum time in milliseconds to wait for a frame
@@ -261,22 +197,10 @@ class Reachy2Camera(Camera):
         if not self.is_connected:
             raise DeviceNotConnectedError(f"{self} is not connected.")
 
-        if self.thread is None or not self.thread.is_alive():
-            self._start_read_thread()
-
-        if not self.new_frame_event.wait(timeout=timeout_ms / 1000.0):
-            thread_alive = self.thread is not None and self.thread.is_alive()
-            raise TimeoutError(
-                f"Timed out waiting for frame from camera {self} after {timeout_ms} ms. "
-                f"Read thread alive: {thread_alive}."
-            )
-
-        with self.frame_lock:
-            frame = self.latest_frame
-            self.new_frame_event.clear()
+        frame = self.read()
 
         if frame is None:
-            raise RuntimeError(f"Internal error: Event set but no frame available for {self}.")
+            raise RuntimeError(f"Internal error: No frame available for {self}.")
 
         return frame
 
@@ -287,12 +211,9 @@ class Reachy2Camera(Camera):
         Raises:
             DeviceNotConnectedError: If the camera is already disconnected.
         """
-        if not self.is_connected and self.thread is None:
+        if not self.is_connected:
             raise DeviceNotConnectedError(f"{self} not connected.")
 
-        if self.thread is not None:
-            self._stop_read_thread()
-
         if self.cam_manager is not None:
             self.cam_manager.disconnect()
 

src/lerobot/cameras/utils.py

@@ -43,6 +43,11 @@ def make_cameras_from_configs(camera_configs: dict[str, CameraConfig]) -> dict[s
 
             cameras[key] = Reachy2Camera(cfg)
 
+        elif cfg.type == "zmq":
+            from .zmq.camera_zmq import ZMQCamera
+
+            cameras[key] = ZMQCamera(cfg)
+
         else:
             try:
                 cameras[key] = cast(Camera, make_device_from_device_class(cfg))

src/lerobot/cameras/zmq/__init__.py

@@ -1,6 +1,6 @@
 #!/usr/bin/env python
 
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -14,5 +14,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from .bi_so100_follower import BiSO100Follower
-from .config_bi_so100_follower import BiSO100FollowerConfig
+from .camera_zmq import ZMQCamera
+from .configuration_zmq import ZMQCameraConfig
+
+__all__ = ["ZMQCamera", "ZMQCameraConfig"]

src/lerobot/cameras/zmq/camera_zmq.py

@@ -0,0 +1,235 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+ZMQCamera - Captures frames from remote cameras via ZeroMQ using JSON protocol in the
+following format:
+    {
+        "timestamps": {"camera_name": float},
+        "images": {"camera_name": "<base64-jpeg>"}
+    }
+"""
+
+import base64
+import json
+import logging
+import time
+from threading import Event, Lock, Thread
+from typing import Any
+
+import cv2
+import numpy as np
+from numpy.typing import NDArray
+
+from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+
+from ..camera import Camera
+from ..configs import ColorMode
+from .configuration_zmq import ZMQCameraConfig
+
+logger = logging.getLogger(__name__)
+
+
+class ZMQCamera(Camera):
+    """
+    Example usage:
+        ```python
+        from lerobot.cameras.zmq import ZMQCamera, ZMQCameraConfig
+
+        config = ZMQCameraConfig(server_address="192.168.123.164", port=5555, camera_name="head_camera")
+        camera = ZMQCamera(config)
+        camera.connect()
+        frame = camera.read()
+        camera.disconnect()
+        ```
+    """
+
+    def __init__(self, config: ZMQCameraConfig):
+        super().__init__(config)
+        import zmq
+
+        self.config = config
+        self.server_address = config.server_address
+        self.port = config.port
+        self.camera_name = config.camera_name
+        self.color_mode = config.color_mode
+        self.timeout_ms = config.timeout_ms
+
+        self.context: zmq.Context | None = None
+        self.socket: zmq.Socket | None = None
+        self._connected = False
+
+        self.thread: Thread | None = None
+        self.stop_event: Event | None = None
+        self.frame_lock: Lock = Lock()
+        self.latest_frame: NDArray[Any] | None = None
+        self.new_frame_event: Event = Event()
+
+    def __str__(self) -> str:
+        return f"ZMQCamera({self.camera_name}@{self.server_address}:{self.port})"
+
+    @property
+    def is_connected(self) -> bool:
+        return self._connected and self.context is not None and self.socket is not None
+
+    def connect(self, warmup: bool = True) -> None:
+        """Connect to ZMQ camera server."""
+        if self.is_connected:
+            raise DeviceAlreadyConnectedError(f"{self} is already connected.")
+
+        logger.info(f"Connecting to {self}...")
+
+        try:
+            import zmq
+
+            self.context = zmq.Context()
+            self.socket = self.context.socket(zmq.SUB)
+            self.socket.setsockopt_string(zmq.SUBSCRIBE, "")
+            self.socket.setsockopt(zmq.RCVTIMEO, self.timeout_ms)
+            self.socket.setsockopt(zmq.CONFLATE, True)
+            self.socket.connect(f"tcp://{self.server_address}:{self.port}")
+            self._connected = True
+
+            # Auto-detect resolution
+            if self.width is None or self.height is None:
+                h, w = self.read().shape[:2]
+                self.height = h
+                self.width = w
+                logger.info(f"{self} resolution: {w}x{h}")
+
+            logger.info(f"{self} connected.")
+
+            if warmup:
+                time.sleep(0.1)
+
+        except Exception as e:
+            self._cleanup()
+            raise RuntimeError(f"Failed to connect to {self}: {e}") from e
+
+    def _cleanup(self):
+        """Clean up ZMQ resources."""
+        self._connected = False
+        if self.socket:
+            self.socket.close()
+            self.socket = None
+        if self.context:
+            self.context.term()
+            self.context = None
+
+    @staticmethod
+    def find_cameras() -> list[dict[str, Any]]:
+        """ZMQ cameras require manual configuration (server address/port)."""
+        return []
+
+    def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
+        """
+        Read a single frame from the ZMQ camera.
+
+        Returns:
+            np.ndarray: Decoded frame (height, width, 3)
+        """
+        if not self.is_connected or self.socket is None:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+
+        try:
+            message = self.socket.recv_string()
+        except Exception as e:
+            if type(e).__name__ == "Again":
+                raise TimeoutError(f"{self} timeout after {self.timeout_ms}ms") from e
+            raise
+
+        # Decode JSON message
+        data = json.loads(message)
+
+        if "images" not in data:
+            raise RuntimeError(f"{self} invalid message: missing 'images' key")
+
+        images = data["images"]
+
+        # Get image by camera name or first available
+        if self.camera_name in images:
+            img_b64 = images[self.camera_name]
+        elif images:
+            img_b64 = next(iter(images.values()))
+        else:
+            raise RuntimeError(f"{self} no images in message")
+
+        # Decode base64 JPEG
+        img_bytes = base64.b64decode(img_b64)
+        frame = cv2.imdecode(np.frombuffer(img_bytes, np.uint8), cv2.IMREAD_COLOR)
+
+        if frame is None:
+            raise RuntimeError(f"{self} failed to decode image")
+
+        return frame
+
+    def _read_loop(self) -> None:
+        while self.stop_event and not self.stop_event.is_set():
+            try:
+                frame = self.read()
+                with self.frame_lock:
+                    self.latest_frame = frame
+                self.new_frame_event.set()
+            except DeviceNotConnectedError:
+                break
+            except TimeoutError:
+                pass
+            except Exception as e:
+                logger.warning(f"Read error: {e}")
+
+    def _start_read_thread(self) -> None:
+        if self.thread and self.thread.is_alive():
+            return
+        self.stop_event = Event()
+        self.thread = Thread(target=self._read_loop, daemon=True)
+        self.thread.start()
+
+    def _stop_read_thread(self) -> None:
+        if self.stop_event:
+            self.stop_event.set()
+        if self.thread and self.thread.is_alive():
+            self.thread.join(timeout=2.0)
+        self.thread = None
+        self.stop_event = None
+
+    def async_read(self, timeout_ms: float = 10000) -> NDArray[Any]:
+        """Read latest frame asynchronously (non-blocking)."""
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+
+        if not self.thread or not self.thread.is_alive():
+            self._start_read_thread()
+
+        if not self.new_frame_event.wait(timeout=timeout_ms / 1000.0):
+            raise TimeoutError(f"{self} async_read timeout after {timeout_ms}ms")
+
+        with self.frame_lock:
+            frame = self.latest_frame
+            self.new_frame_event.clear()
+
+        if frame is None:
+            raise RuntimeError(f"{self} no frame available")
+
+        return frame
+
+    def disconnect(self) -> None:
+        """Disconnect from ZMQ camera."""
+        if not self.is_connected and not self.thread:
+            raise DeviceNotConnectedError(f"{self} not connected.")
+
+        self._stop_read_thread()
+        self._cleanup()
+        logger.info(f"{self} disconnected.")

src/lerobot/cameras/zmq/configuration_zmq.py

@@ -0,0 +1,46 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+
+from ..configs import CameraConfig, ColorMode
+
+__all__ = ["ZMQCameraConfig", "ColorMode"]
+
+
+@CameraConfig.register_subclass("zmq")
+@dataclass
+class ZMQCameraConfig(CameraConfig):
+    server_address: str
+    port: int = 5555
+    camera_name: str = "zmq_camera"
+    color_mode: ColorMode = ColorMode.RGB
+    timeout_ms: int = 5000
+
+    def __post_init__(self) -> None:
+        if self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
+            raise ValueError(
+                f"`color_mode` is expected to be {ColorMode.RGB.value} or {ColorMode.BGR.value}, but {self.color_mode} is provided."
+            )
+
+        if self.timeout_ms <= 0:
+            raise ValueError(f"`timeout_ms` must be positive, but {self.timeout_ms} is provided.")
+
+        if not self.server_address:
+            raise ValueError("`server_address` cannot be empty.")
+
+        if self.port <= 0 or self.port > 65535:
+            raise ValueError(f"`port` must be between 1 and 65535, but {self.port} is provided.")

src/lerobot/cameras/zmq/image_server.py

@@ -0,0 +1,114 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Streams camera images over ZMQ.
+Uses lerobot's OpenCVCamera for capture, encodes images to base64 and sends them over ZMQ.
+"""
+
+import base64
+import contextlib
+import json
+import logging
+import time
+from collections import deque
+
+import cv2
+import numpy as np
+import zmq
+
+from lerobot.cameras.configs import ColorMode
+from lerobot.cameras.opencv import OpenCVCamera, OpenCVCameraConfig
+
+logger = logging.getLogger(__name__)
+
+
+def encode_image(image: np.ndarray, quality: int = 80) -> str:
+    """Encode RGB image to base64 JPEG string."""
+    _, buffer = cv2.imencode(".jpg", image, [int(cv2.IMWRITE_JPEG_QUALITY), quality])
+    return base64.b64encode(buffer).decode("utf-8")
+
+
+class ImageServer:
+    def __init__(self, config: dict, port: int = 5555):
+        self.fps = config.get("fps", 30)
+        self.cameras: dict[str, OpenCVCamera] = {}
+
+        for name, cfg in config.get("cameras", {}).items():
+            shape = cfg.get("shape", [480, 640])
+            cam_config = OpenCVCameraConfig(
+                index_or_path=cfg.get("device_id", 0),
+                fps=self.fps,
+                width=shape[1],
+                height=shape[0],
+                color_mode=ColorMode.RGB,
+            )
+            camera = OpenCVCamera(cam_config)
+            camera.connect()
+            self.cameras[name] = camera
+            logger.info(f"Camera {name}: {shape[1]}x{shape[0]}")
+
+        # ZMQ PUB socket
+        self.context = zmq.Context()
+        self.socket = self.context.socket(zmq.PUB)
+        self.socket.setsockopt(zmq.SNDHWM, 20)
+        self.socket.setsockopt(zmq.LINGER, 0)
+        self.socket.bind(f"tcp://*:{port}")
+
+        logger.info(f"ImageServer running on port {port}")
+
+    def run(self):
+        frame_count = 0
+        frame_times = deque(maxlen=60)
+
+        try:
+            while True:
+                t0 = time.time()
+
+                # Build message
+                message = {"timestamps": {}, "images": {}}
+                for name, cam in self.cameras.items():
+                    frame = cam.read()  # Returns RGB
+                    message["timestamps"][name] = time.time()
+                    message["images"][name] = encode_image(frame)
+
+                # Send as JSON string (suppress if buffer full)
+                with contextlib.suppress(zmq.Again):
+                    self.socket.send_string(json.dumps(message), zmq.NOBLOCK)
+
+                frame_count += 1
+                frame_times.append(time.time() - t0)
+
+                if frame_count % 60 == 0:
+                    logger.debug(f"FPS: {len(frame_times) / sum(frame_times):.1f}")
+
+                sleep = (1.0 / self.fps) - (time.time() - t0)
+                if sleep > 0:
+                    time.sleep(sleep)
+
+        except KeyboardInterrupt:
+            pass
+        finally:
+            for cam in self.cameras.values():
+                cam.disconnect()
+            self.socket.close()
+            self.context.term()
+
+
+if __name__ == "__main__":
+    logging.basicConfig(level=logging.INFO)
+    config = {"fps": 30, "cameras": {"head_camera": {"device_id": 4, "shape": [480, 640]}}}
+    ImageServer(config, port=5555).run()

src/lerobot/configs/default.py

@@ -67,3 +67,31 @@ class EvalConfig:
                 f"to increase the number of episodes to match the batch size (e.g. `eval.n_episodes={self.batch_size}`), "
                 f"or lower the batch size (e.g. `eval.batch_size={self.n_episodes}`)."
             )
+
+
+@dataclass
+class PeftConfig:
+    # PEFT offers many fine-tuning methods, layer adapters being the most common and currently also the most
+    # effective methods so we'll focus on those in this high-level config interface.
+
+    # Either a string (module name suffix or 'all-linear'), a list of module name suffixes or a regular expression
+    # describing module names to target with the configured PEFT method. Some policies have a default value for this
+    # so that you don't *have* to choose which layers to adapt but it might still be worthwhile depending on your case.
+    target_modules: list[str] | str | None = None
+
+    # Names/suffixes of modules to fully fine-tune and store alongside adapter weights. Useful for layers that are
+    # not part of a pre-trained model (e.g., action state projections). Depending on the policy this defaults to layers
+    # that are newly created in pre-trained policies. If you're fine-tuning an already trained policy you might want
+    # to set this to `[]`. Corresponds to PEFT's `modules_to_save`.
+    full_training_modules: list[str] | None = None
+
+    # The PEFT (adapter) method to apply to the policy. Needs to be a valid PEFT type.
+    method_type: str = "LORA"
+
+    # Adapter initialization method. Look at the specific PEFT adapter documentation for defaults.
+    init_type: str | None = None
+
+    # We expect that all PEFT adapters are in some way doing rank-decomposition therefore this parameter specifies
+    # the rank used for the adapter. In general a higher rank means more trainable parameters and closer to full
+    # fine-tuning.
+    r: int = 16

src/lerobot/configs/eval.py

@@ -38,6 +38,8 @@ class EvalPipelineConfig:
     seed: int | None = 1000
     # Rename map for the observation to override the image and state keys
     rename_map: dict[str, str] = field(default_factory=dict)
+    # Explicit consent to execute remote code from the Hub (required for hub environments).
+    trust_remote_code: bool = False
 
     def __post_init__(self) -> None:
         # HACK: We parse again the cli args here to get the pretrained path if there was one.

src/lerobot/configs/policies.py

@@ -55,14 +55,18 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):  # type: igno
 
     n_obs_steps: int = 1
 
-    input_features: dict[str, PolicyFeature] = field(default_factory=dict)
-    output_features: dict[str, PolicyFeature] = field(default_factory=dict)
+    # `input_features` can be set to None/null in order to infer those values from the dataset.
+    input_features: dict[str, PolicyFeature] | None = field(default_factory=dict)
+    output_features: dict[str, PolicyFeature] | None = field(default_factory=dict)
 
     device: str | None = None  # e.g. "cuda", "cuda:0", "cpu", or "mps"
     # `use_amp` determines whether to use Automatic Mixed Precision (AMP) for training and evaluation. With AMP,
     # automatic gradient scaling is used.
     use_amp: bool = False
 
+    # Whether the policy employed PEFT for training.
+    use_peft: bool = False
+
     push_to_hub: bool = True  # type: ignore[assignment] # TODO: use a different name to avoid override
     repo_id: str | None = None
 
@@ -125,6 +129,8 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):  # type: igno
 
     @property
     def robot_state_feature(self) -> PolicyFeature | None:
+        if not self.input_features:
+            return None
         for ft_name, ft in self.input_features.items():
             if ft.type is FeatureType.STATE and ft_name == OBS_STATE:
                 return ft
@@ -132,6 +138,8 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):  # type: igno
 
     @property
     def env_state_feature(self) -> PolicyFeature | None:
+        if not self.input_features:
+            return None
         for _, ft in self.input_features.items():
             if ft.type is FeatureType.ENV:
                 return ft
@@ -139,10 +147,14 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):  # type: igno
 
     @property
     def image_features(self) -> dict[str, PolicyFeature]:
+        if not self.input_features:
+            return {}
         return {key: ft for key, ft in self.input_features.items() if ft.type is FeatureType.VISUAL}
 
     @property
     def action_feature(self) -> PolicyFeature | None:
+        if not self.output_features:
+            return None
         for ft_name, ft in self.output_features.items():
             if ft.type is FeatureType.ACTION and ft_name == ACTION:
                 return ft

src/lerobot/configs/train.py

@@ -24,7 +24,7 @@ from huggingface_hub.errors import HfHubHTTPError
 
 from lerobot import envs
 from lerobot.configs import parser
-from lerobot.configs.default import DatasetConfig, EvalConfig, WandBConfig
+from lerobot.configs.default import DatasetConfig, EvalConfig, PeftConfig, WandBConfig
 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.optim import OptimizerConfig
 from lerobot.optim.schedulers import LRSchedulerConfig
@@ -65,6 +65,7 @@ class TrainPipelineConfig(HubMixin):
     scheduler: LRSchedulerConfig | None = None
     eval: EvalConfig = field(default_factory=EvalConfig)
     wandb: WandBConfig = field(default_factory=WandBConfig)
+    peft: PeftConfig | None = None
 
     # RA-BC (Reward-Aligned Behavior Cloning) parameters
     use_rabc: bool = False  # Enable reward-weighted training

src/lerobot/datasets/lerobot_dataset.py

@@ -78,6 +78,7 @@ from lerobot.datasets.video_utils import (
 from lerobot.utils.constants import HF_LEROBOT_HOME
 
 CODEBASE_VERSION = "v3.0"
+VALID_VIDEO_CODECS = {"h264", "hevc", "libsvtav1"}
 
 
 class LeRobotDatasetMetadata:
@@ -540,11 +541,13 @@ class LeRobotDatasetMetadata:
         return obj
 
 
-def _encode_video_worker(video_key: str, episode_index: int, root: Path, fps: int) -> Path:
+def _encode_video_worker(
+    video_key: str, episode_index: int, root: Path, fps: int, vcodec: str = "libsvtav1"
+) -> Path:
     temp_path = Path(tempfile.mkdtemp(dir=root)) / f"{video_key}_{episode_index:03d}.mp4"
     fpath = DEFAULT_IMAGE_PATH.format(image_key=video_key, episode_index=episode_index, frame_index=0)
     img_dir = (root / fpath).parent
-    encode_video_frames(img_dir, temp_path, fps, overwrite=True)
+    encode_video_frames(img_dir, temp_path, fps, vcodec=vcodec, overwrite=True)
     shutil.rmtree(img_dir)
     return temp_path
 
@@ -563,6 +566,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         download_videos: bool = True,
         video_backend: str | None = None,
         batch_encoding_size: int = 1,
+        vcodec: str = "libsvtav1",
     ):
         """
         2 modes are available for instantiating this class, depending on 2 different use cases:
@@ -675,8 +679,13 @@ class LeRobotDataset(torch.utils.data.Dataset):
                 You can also use the 'pyav' decoder used by Torchvision, which used to be the default option, or 'video_reader' which is another decoder of Torchvision.
             batch_encoding_size (int, optional): Number of episodes to accumulate before batch encoding videos.
                 Set to 1 for immediate encoding (default), or higher for batched encoding. Defaults to 1.
+            vcodec (str, optional): Video codec for encoding videos during recording. Options: 'h264', 'hevc',
+                'libsvtav1'. Defaults to 'libsvtav1'. Use 'h264' for faster encoding on systems where AV1
+                encoding is CPU-heavy.
         """
         super().__init__()
+        if vcodec not in VALID_VIDEO_CODECS:
+            raise ValueError(f"Invalid vcodec '{vcodec}'. Must be one of: {sorted(VALID_VIDEO_CODECS)}")
         self.repo_id = repo_id
         self.root = Path(root) if root else HF_LEROBOT_HOME / repo_id
         self.image_transforms = image_transforms
@@ -688,6 +697,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         self.delta_indices = None
         self.batch_encoding_size = batch_encoding_size
         self.episodes_since_last_encoding = 0
+        self.vcodec = vcodec
 
         # Unused attributes
         self.image_writer = None
@@ -1211,6 +1221,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
                             episode_index,
                             self.root,
                             self.fps,
+                            self.vcodec,
                         ): video_key
                         for video_key in self.meta.video_keys
                     }
@@ -1526,7 +1537,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         Note: `encode_video_frames` is a blocking call. Making it asynchronous shouldn't speedup encoding,
         since video encoding with ffmpeg is already using multithreading.
         """
-        return _encode_video_worker(video_key, episode_index, self.root, self.fps)
+        return _encode_video_worker(video_key, episode_index, self.root, self.fps, self.vcodec)
 
     @classmethod
     def create(
@@ -1542,8 +1553,11 @@ class LeRobotDataset(torch.utils.data.Dataset):
         image_writer_threads: int = 0,
         video_backend: str | None = None,
         batch_encoding_size: int = 1,
+        vcodec: str = "libsvtav1",
     ) -> "LeRobotDataset":
         """Create a LeRobot Dataset from scratch in order to record data."""
+        if vcodec not in VALID_VIDEO_CODECS:
+            raise ValueError(f"Invalid vcodec '{vcodec}'. Must be one of: {sorted(VALID_VIDEO_CODECS)}")
         obj = cls.__new__(cls)
         obj.meta = LeRobotDatasetMetadata.create(
             repo_id=repo_id,
@@ -1560,6 +1574,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj.image_writer = None
         obj.batch_encoding_size = batch_encoding_size
         obj.episodes_since_last_encoding = 0
+        obj.vcodec = vcodec
 
         if image_writer_processes or image_writer_threads:
             obj.start_image_writer(image_writer_processes, image_writer_threads)

src/lerobot/datasets/pipeline_features.py

@@ -18,12 +18,12 @@ from typing import Any
 
 from lerobot.configs.types import PipelineFeatureType
 from lerobot.datasets.utils import hw_to_dataset_features
-from lerobot.processor import DataProcessorPipeline
+from lerobot.processor import DataProcessorPipeline, RobotAction, RobotObservation
 from lerobot.utils.constants import ACTION, OBS_IMAGES, OBS_STATE, OBS_STR
 
 
 def create_initial_features(
-    action: dict[str, Any] | None = None, observation: dict[str, Any] | None = None
+    action: RobotAction | None = None, observation: RobotObservation | None = None
 ) -> dict[PipelineFeatureType, dict[str, Any]]:
     """
     Creates the initial features dict for the dataset from action and observation specs.

src/lerobot/envs/__init__.py

@@ -12,4 +12,4 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from .configs import AlohaEnv, EnvConfig, PushtEnv  # noqa: F401
+from .configs import AlohaEnv, EnvConfig, HubEnvConfig, PushtEnv  # noqa: F401

src/lerobot/envs/configs.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 import abc
-from dataclasses import dataclass, field
+from dataclasses import dataclass, field, fields
 from typing import Any
 
 import draccus
@@ -68,6 +68,22 @@ class EnvConfig(draccus.ChoiceRegistry, abc.ABC):
         raise NotImplementedError()
 
 
+@dataclass
+class HubEnvConfig(EnvConfig):
+    """Base class for environments that delegate creation to a hub-hosted make_env.
+
+    Hub environments download and execute remote code from the HF Hub.
+    The hub_path points to a repository containing an env.py with a make_env function.
+    """
+
+    hub_path: str | None = None  # required: e.g., "username/repo" or "username/repo@branch:file.py"
+
+    @property
+    def gym_kwargs(self) -> dict:
+        # Not used for hub environments - the hub's make_env handles everything
+        return {}
+
+
 @EnvConfig.register_subclass("aloha")
 @dataclass
 class AlohaEnv(EnvConfig):
@@ -368,3 +384,71 @@ class MetaworldEnv(EnvConfig):
             "obs_type": self.obs_type,
             "render_mode": self.render_mode,
         }
+
+
+@EnvConfig.register_subclass("isaaclab_arena")
+@dataclass
+class IsaaclabArenaEnv(HubEnvConfig):
+    hub_path: str = "nvidia/isaaclab-arena-envs"
+    episode_length: int = 300
+    num_envs: int = 1
+    embodiment: str | None = "gr1_pink"
+    object: str | None = "power_drill"
+    mimic: bool = False
+    teleop_device: str | None = None
+    seed: int | None = 42
+    device: str | None = "cuda:0"
+    disable_fabric: bool = False
+    enable_cameras: bool = False
+    headless: bool = False
+    enable_pinocchio: bool = True
+    environment: str | None = "gr1_microwave"
+    task: str | None = "Reach out to the microwave and open it."
+    state_dim: int = 54
+    action_dim: int = 36
+    camera_height: int = 512
+    camera_width: int = 512
+    video: bool = False
+    video_length: int = 100
+    video_interval: int = 200
+    # Comma-separated keys, e.g., "robot_joint_pos,left_eef_pos"
+    state_keys: str = "robot_joint_pos"
+    # Comma-separated keys, e.g., "robot_pov_cam_rgb,front_cam_rgb"
+    # Set to None or "" for environments without cameras
+    camera_keys: str | None = None
+    features: dict[str, PolicyFeature] = field(default_factory=dict)
+    features_map: dict[str, str] = field(default_factory=dict)
+    kwargs: dict | None = None
+
+    def __post_init__(self):
+        if self.kwargs:
+            # dynamically convert kwargs to fields in the dataclass
+            # NOTE! the new fields will not bee seen by the dataclass repr
+            field_names = {f.name for f in fields(self)}
+            for key, value in self.kwargs.items():
+                if key not in field_names and key != "kwargs":
+                    setattr(self, key, value)
+            self.kwargs = None
+
+        # Set action feature
+        self.features[ACTION] = PolicyFeature(type=FeatureType.ACTION, shape=(self.action_dim,))
+        self.features_map[ACTION] = ACTION
+
+        # Set state feature
+        self.features[OBS_STATE] = PolicyFeature(type=FeatureType.STATE, shape=(self.state_dim,))
+        self.features_map[OBS_STATE] = OBS_STATE
+
+        # Add camera features for each camera key
+        if self.enable_cameras and self.camera_keys:
+            for cam_key in self.camera_keys.split(","):
+                cam_key = cam_key.strip()
+                if cam_key:
+                    self.features[cam_key] = PolicyFeature(
+                        type=FeatureType.VISUAL,
+                        shape=(self.camera_height, self.camera_width, 3),
+                    )
+                    self.features_map[cam_key] = f"{OBS_IMAGES}.{cam_key}"
+
+    @property
+    def gym_kwargs(self) -> dict:
+        return {}

src/lerobot/envs/factory.py

@@ -20,11 +20,11 @@ import gymnasium as gym
 from gymnasium.envs.registration import registry as gym_registry
 
 from lerobot.configs.policies import PreTrainedConfig
-from lerobot.envs.configs import AlohaEnv, EnvConfig, LiberoEnv, PushtEnv
+from lerobot.envs.configs import AlohaEnv, EnvConfig, HubEnvConfig, IsaaclabArenaEnv, LiberoEnv, PushtEnv
 from lerobot.envs.utils import _call_make_env, _download_hub_file, _import_hub_module, _normalize_hub_result
 from lerobot.policies.xvla.configuration_xvla import XVLAConfig
 from lerobot.processor import ProcessorStep
-from lerobot.processor.env_processor import LiberoProcessorStep
+from lerobot.processor.env_processor import IsaaclabArenaProcessorStep, LiberoProcessorStep
 from lerobot.processor.pipeline import PolicyProcessorPipeline
 
 
@@ -73,6 +73,26 @@ def make_env_pre_post_processors(
     if isinstance(env_cfg, LiberoEnv) or "libero" in env_cfg.type:
         preprocessor_steps.append(LiberoProcessorStep())
 
+    # For Isaaclab Arena environments, add the IsaaclabArenaProcessorStep
+    if isinstance(env_cfg, IsaaclabArenaEnv) or "isaaclab_arena" in env_cfg.type:
+        # Parse comma-separated keys (handle None for state-based policies)
+        if env_cfg.state_keys:
+            state_keys = tuple(k.strip() for k in env_cfg.state_keys.split(",") if k.strip())
+        else:
+            state_keys = ()
+        if env_cfg.camera_keys:
+            camera_keys = tuple(k.strip() for k in env_cfg.camera_keys.split(",") if k.strip())
+        else:
+            camera_keys = ()
+        if not state_keys and not camera_keys:
+            raise ValueError("At least one of state_keys or camera_keys must be specified.")
+        preprocessor_steps.append(
+            IsaaclabArenaProcessorStep(
+                state_keys=state_keys,
+                camera_keys=camera_keys,
+            )
+        )
+
     preprocessor = PolicyProcessorPipeline(steps=preprocessor_steps)
     postprocessor = PolicyProcessorPipeline(steps=postprocessor_steps)
 
@@ -98,7 +118,6 @@ def make_env(
         hub_cache_dir (str | None): Optional cache path for downloaded hub files.
         trust_remote_code (bool): **Explicit consent** to execute remote code from the Hub.
             Default False — must be set to True to import/exec hub `env.py`.
-
     Raises:
         ValueError: if n_envs < 1
         ModuleNotFoundError: If the requested env package is not installed
@@ -112,19 +131,35 @@ def make_env(
     """
     # if user passed a hub id string (e.g., "username/repo", "username/repo@main:env.py")
     # simplified: only support hub-provided `make_env`
+    # TODO: (jadechoghari): deprecate string API and remove this check
     if isinstance(cfg, str):
+        hub_path: str | None = cfg
+    elif isinstance(cfg, HubEnvConfig):
+        hub_path = cfg.hub_path
+    else:
+        hub_path = None
+
+    # If hub_path is set, download and call hub-provided `make_env`
+    if hub_path:
         # _download_hub_file will raise the same RuntimeError if trust_remote_code is False
-        repo_id, file_path, local_file, revision = _download_hub_file(cfg, trust_remote_code, hub_cache_dir)
+        repo_id, file_path, local_file, revision = _download_hub_file(
+            hub_path, trust_remote_code, hub_cache_dir
+        )
 
         # import and surface clear import errors
         module = _import_hub_module(local_file, repo_id)
 
         # call the hub-provided make_env
-        raw_result = _call_make_env(module, n_envs=n_envs, use_async_envs=use_async_envs)
+        env_cfg = None if isinstance(cfg, str) else cfg
+        raw_result = _call_make_env(module, n_envs=n_envs, use_async_envs=use_async_envs, cfg=env_cfg)
 
         # normalize the return into {suite: {task_id: vec_env}}
         return _normalize_hub_result(raw_result)
 
+    # At this point, cfg must be an EnvConfig (not a string) since hub_path would have been set otherwise
+    if isinstance(cfg, str):
+        raise TypeError("cfg should be an EnvConfig at this point")
+
     if n_envs < 1:
         raise ValueError("`n_envs` must be at least 1")
 

src/lerobot/envs/libero.py

@@ -29,6 +29,8 @@ from gymnasium import spaces
 from libero.libero import benchmark, get_libero_path
 from libero.libero.envs import OffScreenRenderEnv
 
+from lerobot.processor import RobotObservation
+
 
 def _parse_camera_names(camera_name: str | Sequence[str]) -> list[str]:
     """Normalize camera_name into a non-empty list of strings."""
@@ -237,7 +239,7 @@ class LiberoEnv(gym.Env):
         env.reset()
         return env
 
-    def _format_raw_obs(self, raw_obs: dict[str, Any]) -> dict[str, Any]:
+    def _format_raw_obs(self, raw_obs: RobotObservation) -> RobotObservation:
         images = {}
         for camera_name in self.camera_name:
             image = raw_obs[camera_name]
@@ -313,7 +315,7 @@ class LiberoEnv(gym.Env):
         info = {"is_success": False}
         return observation, info
 
-    def step(self, action: np.ndarray) -> tuple[dict[str, Any], float, bool, bool, dict[str, Any]]:
+    def step(self, action: np.ndarray) -> tuple[RobotObservation, float, bool, bool, dict[str, Any]]:
         if action.ndim != 1:
             raise ValueError(
                 f"Expected action to be 1-D (shape (action_dim,)), "

src/lerobot/envs/metaworld.py

@@ -25,6 +25,8 @@ import metaworld.policies as policies
 import numpy as np
 from gymnasium import spaces
 
+from lerobot.processor import RobotObservation
+
 # ---- Load configuration data from the external JSON file ----
 CONFIG_PATH = Path(__file__).parent / "metaworld_config.json"
 try:
@@ -161,7 +163,7 @@ class MetaworldEnv(gym.Env):
         env._freeze_rand_vec = False  # otherwise no randomization
         return env
 
-    def _format_raw_obs(self, raw_obs: np.ndarray) -> dict[str, Any]:
+    def _format_raw_obs(self, raw_obs: np.ndarray) -> RobotObservation:
         image = None
         if self._env is not None:
             image = self._env.render()
@@ -196,7 +198,7 @@ class MetaworldEnv(gym.Env):
         self,
         seed: int | None = None,
         **kwargs,
-    ) -> tuple[dict[str, Any], dict[str, Any]]:
+    ) -> tuple[RobotObservation, dict[str, Any]]:
         """
         Reset the environment to its initial state.
 
@@ -204,7 +206,7 @@ class MetaworldEnv(gym.Env):
             seed (Optional[int]): Random seed for environment initialization.
 
         Returns:
-            observation (Dict[str, Any]): The initial formatted observation.
+            observation (RobotObservation): The initial formatted observation.
             info (Dict[str, Any]): Additional info about the reset state.
         """
         super().reset(seed=seed)
@@ -216,7 +218,7 @@ class MetaworldEnv(gym.Env):
         info = {"is_success": False}
         return observation, info
 
-    def step(self, action: np.ndarray) -> tuple[dict[str, Any], float, bool, bool, dict[str, Any]]:
+    def step(self, action: np.ndarray) -> tuple[RobotObservation, float, bool, bool, dict[str, Any]]:
         """
         Perform one environment step.
 
@@ -224,7 +226,7 @@ class MetaworldEnv(gym.Env):
             action (np.ndarray): The action to execute, must be 1-D with shape (action_dim,).
 
         Returns:
-            observation (Dict[str, Any]): The formatted observation after the step.
+            observation (RobotObservation): The formatted observation after the step.
             reward (float): The scalar reward for this step.
             terminated (bool): Whether the episode terminated successfully.
             truncated (bool): Whether the episode was truncated due to a time limit.

src/lerobot/envs/utils.py

@@ -29,6 +29,7 @@ from torch import Tensor
 
 from lerobot.configs.types import FeatureType, PolicyFeature
 from lerobot.envs.configs import EnvConfig
+from lerobot.processor import RobotObservation
 from lerobot.utils.constants import OBS_ENV_STATE, OBS_IMAGE, OBS_IMAGES, OBS_STATE, OBS_STR
 from lerobot.utils.utils import get_channel_first_image_shape
 
@@ -46,7 +47,7 @@ def _convert_nested_dict(d):
 
 
 def preprocess_observation(observations: dict[str, np.ndarray]) -> dict[str, Tensor]:
-    # TODO(aliberts, rcadene): refactor this to use features from the environment (no hardcoding)
+    # TODO(jadechoghari, imstevenpmwork): refactor this to use features from the environment (no hardcoding)
     """Convert environment observation to LeRobot format observation.
     Args:
         observation: Dictionary of observation batches from a Gym vector environment.
@@ -98,11 +99,19 @@ def preprocess_observation(observations: dict[str, np.ndarray]) -> dict[str, Ten
 
     if "robot_state" in observations:
         return_observations[f"{OBS_STR}.robot_state"] = _convert_nested_dict(observations["robot_state"])
+
+    # Handle IsaacLab Arena format: observations have 'policy' and 'camera_obs' keys
+    if "policy" in observations:
+        return_observations[f"{OBS_STR}.policy"] = observations["policy"]
+
+    if "camera_obs" in observations:
+        return_observations[f"{OBS_STR}.camera_obs"] = observations["camera_obs"]
+
     return return_observations
 
 
 def env_to_policy_features(env_cfg: EnvConfig) -> dict[str, PolicyFeature]:
-    # TODO(aliberts, rcadene): remove this hardcoding of keys and just use the nested keys as is
+    # TODO(jadechoghari, imstevenpmwork): remove this hardcoding of keys and just use the nested keys as is
     # (need to also refactor preprocess_observation and externalize normalization from policies)
     policy_features = {}
     for key, ft in env_cfg.features.items():
@@ -144,7 +153,7 @@ def check_env_attributes_and_types(env: gym.vector.VectorEnv) -> None:
             )
 
 
-def add_envs_task(env: gym.vector.VectorEnv, observation: dict[str, Any]) -> dict[str, Any]:
+def add_envs_task(env: gym.vector.VectorEnv, observation: RobotObservation) -> RobotObservation:
     """Adds task feature to the observation dict with respect to the first environment attribute."""
     if hasattr(env.envs[0], "task_description"):
         task_result = env.call("task_description")
@@ -302,7 +311,7 @@ def _import_hub_module(local_file: str, repo_id: str) -> Any:
     return module
 
 
-def _call_make_env(module: Any, n_envs: int, use_async_envs: bool) -> Any:
+def _call_make_env(module: Any, n_envs: int, use_async_envs: bool, cfg: EnvConfig | None) -> Any:
     """
     Ensure module exposes make_env and call it.
     """
@@ -311,7 +320,11 @@ def _call_make_env(module: Any, n_envs: int, use_async_envs: bool) -> Any:
             f"The hub module {getattr(module, '__name__', 'hub_module')} must expose `make_env(n_envs=int, use_async_envs=bool)`."
         )
     entry_fn = module.make_env
-    return entry_fn(n_envs=n_envs, use_async_envs=use_async_envs)
+    # Only pass cfg if it's not None (i.e., when an EnvConfig was provided, not a string hub ID)
+    if cfg is not None:
+        return entry_fn(n_envs=n_envs, use_async_envs=use_async_envs, cfg=cfg)
+    else:
+        return entry_fn(n_envs=n_envs, use_async_envs=use_async_envs)
 
 
 def _normalize_hub_result(result: Any) -> dict[str, dict[int, gym.vector.VectorEnv]]:

src/lerobot/policies/__init__.py

@@ -16,6 +16,7 @@ from .act.configuration_act import ACTConfig as ACTConfig
 from .diffusion.configuration_diffusion import DiffusionConfig as DiffusionConfig
 from .groot.configuration_groot import GrootConfig as GrootConfig
 from .pi0.configuration_pi0 import PI0Config as PI0Config
+from .pi0_fast.configuration_pi0_fast import PI0FastConfig as PI0FastConfig
 from .pi05.configuration_pi05 import PI05Config as PI05Config
 from .smolvla.configuration_smolvla import SmolVLAConfig as SmolVLAConfig
 from .smolvla.processor_smolvla import SmolVLANewLineProcessor
@@ -29,6 +30,7 @@ __all__ = [
     "DiffusionConfig",
     "PI0Config",
     "PI05Config",
+    "PI0FastConfig",
     "SmolVLAConfig",
     "SARMConfig",
     "TDMPCConfig",

src/lerobot/policies/factory.py

@@ -51,7 +51,11 @@ from lerobot.processor.converters import (
     transition_to_batch,
     transition_to_policy_action,
 )
-from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
+from lerobot.utils.constants import (
+    ACTION,
+    POLICY_POSTPROCESSOR_DEFAULT_NAME,
+    POLICY_PREPROCESSOR_DEFAULT_NAME,
+)
 
 
 def get_policy_class(name: str) -> type[PreTrainedPolicy]:
@@ -91,6 +95,10 @@ def get_policy_class(name: str) -> type[PreTrainedPolicy]:
         from lerobot.policies.pi0.modeling_pi0 import PI0Policy
 
         return PI0Policy
+    elif name == "pi0_fast":
+        from lerobot.policies.pi0_fast.modeling_pi0_fast import PI0FastPolicy
+
+        return PI0FastPolicy
     elif name == "pi05":
         from lerobot.policies.pi05.modeling_pi05 import PI05Policy
 
@@ -246,7 +254,7 @@ def make_pre_post_processors(
             }
 
             # Also ensure postprocessing slices to env action dim and unnormalizes with dataset stats
-            env_action_dim = policy_cfg.output_features["action"].shape[0]
+            env_action_dim = policy_cfg.output_features[ACTION].shape[0]
             postprocessor_overrides["groot_action_unpack_unnormalize_v1"] = {
                 "stats": kwargs.get("dataset_stats"),
                 "normalize_min_max": True,
@@ -471,11 +479,40 @@ def make_policy(
     if ds_meta is not None:
         kwargs["dataset_meta"] = ds_meta
 
-    if cfg.pretrained_path:
+    if not cfg.pretrained_path and cfg.use_peft:
+        raise ValueError(
+            "Instantiating a policy with `use_peft=True` without a checkpoint is not supported since that requires "
+            "the PEFT config parameters to be set. For training with PEFT, see `lerobot_train.py` on how to do that."
+        )
+
+    if cfg.pretrained_path and not cfg.use_peft:
         # Load a pretrained policy and override the config if needed (for example, if there are inference-time
         # hyperparameters that we want to vary).
         kwargs["pretrained_name_or_path"] = cfg.pretrained_path
         policy = policy_cls.from_pretrained(**kwargs)
+    elif cfg.pretrained_path and cfg.use_peft:
+        # Load a pretrained PEFT model on top of the policy. The pretrained path points to the folder/repo
+        # of the adapter and the adapter's config contains the path to the base policy. So we need the
+        # adapter config first, then load the correct policy and then apply PEFT.
+        from peft import PeftConfig, PeftModel
+
+        logging.info("Loading policy's PEFT adapter.")
+
+        peft_pretrained_path = cfg.pretrained_path
+        peft_config = PeftConfig.from_pretrained(peft_pretrained_path)
+
+        kwargs["pretrained_name_or_path"] = peft_config.base_model_name_or_path
+        if not kwargs["pretrained_name_or_path"]:
+            # This means that there's a bug or we trained a policy from scratch using PEFT.
+            # It is more likely that this is a bug so we'll raise an error.
+            raise ValueError(
+                "No pretrained model name found in adapter config. Can't instantiate the pre-trained policy on which "
+                "the adapter was trained."
+            )
+
+        policy = policy_cls.from_pretrained(**kwargs)
+        policy = PeftModel.from_pretrained(policy, peft_pretrained_path, config=peft_config)
+
     else:
         # Make a fresh policy.
         policy = policy_cls(**kwargs)

src/lerobot/policies/groot/configuration_groot.py

@@ -20,6 +20,7 @@ from lerobot.configs.policies import PreTrainedConfig
 from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
 from lerobot.optim.optimizers import AdamWConfig
 from lerobot.optim.schedulers import CosineDecayWithWarmupSchedulerConfig
+from lerobot.utils.constants import ACTION, OBS_STATE
 
 
 @PreTrainedConfig.register_subclass("groot")
@@ -137,14 +138,14 @@ class GrootConfig(PreTrainedConfig):
                 "No features of type FeatureType.VISUAL found in input_features."
             )
 
-        if "observation.state" not in self.input_features:
+        if OBS_STATE not in self.input_features:
             state_feature = PolicyFeature(
                 type=FeatureType.STATE,
                 shape=(self.max_state_dim,),
             )
-            self.input_features["observation.state"] = state_feature
+            self.input_features[OBS_STATE] = state_feature
         else:
-            state_shape = self.input_features["observation.state"].shape
+            state_shape = self.input_features[OBS_STATE].shape
             state_dim = state_shape[0] if state_shape else 0
             if state_dim > self.max_state_dim:
                 raise ValueError(
@@ -152,14 +153,14 @@ class GrootConfig(PreTrainedConfig):
                     f"Either reduce state dimension or increase max_state_dim in config."
                 )
 
-        if "action" not in self.output_features:
+        if ACTION not in self.output_features:
             action_feature = PolicyFeature(
                 type=FeatureType.ACTION,
                 shape=(self.max_action_dim,),
             )
-            self.output_features["action"] = action_feature
+            self.output_features[ACTION] = action_feature
         else:
-            action_shape = self.output_features["action"].shape
+            action_shape = self.output_features[ACTION].shape
             action_dim = action_shape[0] if action_shape else 0
             if action_dim > self.max_action_dim:
                 raise ValueError(

src/lerobot/policies/groot/groot_n1.py

@@ -46,7 +46,7 @@ from lerobot.policies.groot.action_head.flow_matching_action_head import (
     FlowmatchingActionHeadConfig,
 )
 from lerobot.policies.groot.utils import ensure_eagle_cache_ready
-from lerobot.utils.constants import HF_LEROBOT_HOME
+from lerobot.utils.constants import ACTION, HF_LEROBOT_HOME
 
 DEFAULT_VENDOR_EAGLE_PATH = str((Path(__file__).resolve().parent / "eagle2_hg_model").resolve())
 DEFAULT_TOKENIZER_ASSETS_REPO = "lerobot/eagle2hg-processor-groot-n1p5"
@@ -227,8 +227,8 @@ class GR00TN15(PreTrainedModel):
 
         detected_error = False
         error_msg = ERROR_MSG
-        if "action" in inputs:
-            action = inputs["action"]
+        if ACTION in inputs:
+            action = inputs[ACTION]
             # In inference, action may be omitted or None; validate only when it's a tensor.
             if action is None:
                 pass  # allow None during inference

src/lerobot/policies/groot/modeling_groot.py

@@ -41,6 +41,7 @@ from torch import Tensor
 from lerobot.policies.groot.configuration_groot import GrootConfig
 from lerobot.policies.groot.groot_n1 import GR00TN15
 from lerobot.policies.pretrained import PreTrainedPolicy
+from lerobot.utils.constants import ACTION
 
 
 class GrootPolicy(PreTrainedPolicy):
@@ -147,7 +148,7 @@ class GrootPolicy(PreTrainedPolicy):
 
         actions = outputs.get("action_pred")
 
-        original_action_dim = self.config.output_features["action"].shape[0]
+        original_action_dim = self.config.output_features[ACTION].shape[0]
         actions = actions[:, :, :original_action_dim]
 
         return actions

src/lerobot/policies/groot/processor_groot.py

@@ -51,7 +51,11 @@ from lerobot.processor.converters import (
 )
 from lerobot.processor.core import EnvTransition, TransitionKey
 from lerobot.utils.constants import (
+    ACTION,
     HF_LEROBOT_HOME,
+    OBS_IMAGE,
+    OBS_IMAGES,
+    OBS_STATE,
     POLICY_POSTPROCESSOR_DEFAULT_NAME,
     POLICY_PREPROCESSOR_DEFAULT_NAME,
 )
@@ -107,9 +111,9 @@ def make_groot_pre_post_processors(
     # Define feature specs for optional normalization steps
     _features: dict[str, PolicyFeature] = {
         # Observation features (only add those we may normalize)
-        "observation.state": PolicyFeature(type=FeatureType.STATE, shape=(state_horizon, max_state_dim)),
+        OBS_STATE: PolicyFeature(type=FeatureType.STATE, shape=(state_horizon, max_state_dim)),
         # Action feature
-        "action": PolicyFeature(type=FeatureType.ACTION, shape=(action_horizon, max_action_dim)),
+        ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(action_horizon, max_action_dim)),
     }
 
     # Normalize STATE and ACTION with min_max (SO100-like default)
@@ -120,7 +124,7 @@ def make_groot_pre_post_processors(
 
     # Determine env action dimension from config (simple, object-like PolicyFeature)
     try:
-        env_action_dim = int(config.output_features["action"].shape[0])
+        env_action_dim = int(config.output_features[ACTION].shape[0])
     except Exception:
         env_action_dim = 0
 
@@ -268,9 +272,9 @@ class GrootPackInputsStep(ProcessorStep):
             return torch.where(mask, mapped, torch.zeros_like(mapped))
 
         # 1) Video (B, T=1, V, H, W, C) uint8
-        img_keys = sorted([k for k in obs if k.startswith("observation.images.")])
-        if not img_keys and "observation.image" in obs:
-            img_keys = ["observation.image"]
+        img_keys = sorted([k for k in obs if k.startswith(OBS_IMAGES)])
+        if not img_keys and OBS_IMAGE in obs:
+            img_keys = [OBS_IMAGE]
         if img_keys:
             cams = [_to_uint8_np_bhwc(obs[k]) for k in img_keys]
             video = np.stack(cams, axis=1)  # (B, V, H, W, C)
@@ -294,14 +298,14 @@ class GrootPackInputsStep(ProcessorStep):
         comp["language"] = lang
 
         # 3) State/state_mask -> (B, 1, max_state_dim)
-        if "observation.state" in obs:
-            state = obs["observation.state"]  # (B, D)
+        if OBS_STATE in obs:
+            state = obs[OBS_STATE]  # (B, D)
             if state.dim() != 2:
                 raise ValueError(f"state must be (B, D), got {tuple(state.shape)}")
             bsz, d = state.shape
             # Normalize BEFORE padding
             if self.normalize_min_max:
-                state = _min_max_norm(state, "observation.state")
+                state = _min_max_norm(state, OBS_STATE)
             state = state.unsqueeze(1)  # (B, 1, D)
             if d > self.max_state_dim:
                 state = state[:, :, : self.max_state_dim]
@@ -320,11 +324,11 @@ class GrootPackInputsStep(ProcessorStep):
             # Normalize BEFORE temporal expansion/padding
             if self.normalize_min_max:
                 if action.dim() == 2:
-                    action = _min_max_norm(action, "action")
+                    action = _min_max_norm(action, ACTION)
                 elif action.dim() == 3:
                     b, t, d = action.shape
                     flat = action.reshape(b * t, d)
-                    flat = _min_max_norm(flat, "action")
+                    flat = _min_max_norm(flat, ACTION)
                     action = flat.view(b, t, d)
             if action.dim() == 2:
                 action = action.unsqueeze(1).repeat(1, self.action_horizon, 1)
@@ -590,7 +594,7 @@ class GrootActionUnpackUnnormalizeStep(ProcessorStep):
         # forward: y = 2 * (x - min) / denom - 1, with y=0 when denom==0
         # inverse: x = (y+1)/2 * denom + min, and when denom==0 -> x = min
         if self.normalize_min_max and self.stats is not None:
-            stats_k = self.stats.get("action", {})
+            stats_k = self.stats.get(ACTION, {})
             d = action.shape[-1]
             min_v = torch.as_tensor(
                 stats_k.get("min", torch.zeros(d)), dtype=action.dtype, device=action.device

src/lerobot/policies/pi0/configuration_pi0.py

@@ -21,7 +21,7 @@ from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
 from lerobot.optim.optimizers import AdamWConfig
 from lerobot.optim.schedulers import CosineDecayWithWarmupSchedulerConfig
 from lerobot.policies.rtc.configuration_rtc import RTCConfig
-from lerobot.utils.constants import OBS_IMAGES
+from lerobot.utils.constants import ACTION, OBS_IMAGES, OBS_STATE
 
 DEFAULT_IMAGE_SIZE = 224
 
@@ -76,6 +76,10 @@ class PI0Config(PreTrainedConfig):
     compile_mode: str = "max-autotune"  # Torch compile mode
     device: str | None = None  # Device to use for the model (None = auto-detect)
 
+    # Finetuning settings
+    freeze_vision_encoder: bool = False  # Freeze only the vision encoder
+    train_expert_only: bool = False  # Freeze entire VLM, train only action expert and projections
+
     # Optimizer settings: see openpi `AdamW``
     optimizer_lr: float = 2.5e-5  # see openpi `CosineDecaySchedule: peak_lr`
     optimizer_betas: tuple[float, float] = (0.9, 0.95)
@@ -120,19 +124,19 @@ class PI0Config(PreTrainedConfig):
             )
             self.input_features[key] = empty_camera
 
-        if "observation.state" not in self.input_features:
+        if OBS_STATE not in self.input_features:
             state_feature = PolicyFeature(
                 type=FeatureType.STATE,
                 shape=(self.max_state_dim,),  # Padded to max_state_dim
             )
-            self.input_features["observation.state"] = state_feature
+            self.input_features[OBS_STATE] = state_feature
 
-        if "action" not in self.output_features:
+        if ACTION not in self.output_features:
             action_feature = PolicyFeature(
                 type=FeatureType.ACTION,
                 shape=(self.max_action_dim,),  # Padded to max_action_dim
             )
-            self.output_features["action"] = action_feature
+            self.output_features[ACTION] = action_feature
 
     def get_optimizer_preset(self) -> AdamWConfig:
         return AdamWConfig(

src/lerobot/policies/pi0/modeling_pi0.py

@@ -339,10 +339,14 @@ class PaliGemmaWithExpertModel(
         use_adarms=None,
         precision: Literal["bfloat16", "float32"] = "bfloat16",
         image_size: int = DEFAULT_IMAGE_SIZE,
+        freeze_vision_encoder: bool = False,
+        train_expert_only: bool = False,
     ):
         if use_adarms is None:
             use_adarms = [False, False]
         super().__init__()
+        self.freeze_vision_encoder = freeze_vision_encoder
+        self.train_expert_only = train_expert_only
 
         vlm_config_hf = CONFIG_MAPPING["paligemma"]()
         vlm_config_hf._vocab_size = 257152  # noqa: SLF001
@@ -383,6 +387,7 @@ class PaliGemmaWithExpertModel(
         self.gemma_expert.model.embed_tokens = None
 
         self.to_bfloat16_for_selected_params(precision)
+        self._set_requires_grad()
 
     def to_bfloat16_for_selected_params(self, precision: Literal["bfloat16", "float32"] = "bfloat16"):
         if precision == "bfloat16":
@@ -406,6 +411,23 @@ class PaliGemmaWithExpertModel(
             if any(selector in name for selector in params_to_keep_float32):
                 param.data = param.data.to(dtype=torch.float32)
 
+    def _set_requires_grad(self):
+        if self.freeze_vision_encoder:
+            self.paligemma.vision_tower.eval()
+            for param in self.paligemma.vision_tower.parameters():
+                param.requires_grad = False
+        if self.train_expert_only:
+            self.paligemma.eval()
+            for param in self.paligemma.parameters():
+                param.requires_grad = False
+
+    def train(self, mode: bool = True):
+        super().train(mode)
+        if self.freeze_vision_encoder:
+            self.paligemma.vision_tower.eval()
+        if self.train_expert_only:
+            self.paligemma.eval()
+
     def embed_image(self, image: torch.Tensor):
         return self.paligemma.model.get_image_features(image)
 
@@ -533,6 +555,8 @@ class PI0Pytorch(nn.Module):  # see openpi `PI0Pytorch`
             use_adarms=[False, False],
             precision=config.dtype,
             image_size=config.image_resolution[0],
+            freeze_vision_encoder=config.freeze_vision_encoder,
+            train_expert_only=config.train_expert_only,
         )
 
         self.action_in_proj = nn.Linear(config.max_action_dim, action_expert_config.width)

src/lerobot/policies/pi05/configuration_pi05.py

@@ -21,6 +21,7 @@ from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
 from lerobot.optim.optimizers import AdamWConfig
 from lerobot.optim.schedulers import CosineDecayWithWarmupSchedulerConfig
 from lerobot.policies.rtc.configuration_rtc import RTCConfig
+from lerobot.utils.constants import ACTION, OBS_IMAGES, OBS_STATE
 
 DEFAULT_IMAGE_SIZE = 224
 
@@ -76,6 +77,10 @@ class PI05Config(PreTrainedConfig):
     compile_mode: str = "max-autotune"  # Torch compile mode
     device: str | None = None  # Device to use for the model (None = auto-detect)
 
+    # Finetuning settings
+    freeze_vision_encoder: bool = False  # Freeze only the vision encoder
+    train_expert_only: bool = False  # Freeze entire VLM, train only action expert and projections
+
     # Optimizer settings: see openpi `AdamW`
     optimizer_lr: float = 2.5e-5  # see openpi `CosineDecaySchedule: peak_lr`
     optimizer_betas: tuple[float, float] = (0.9, 0.95)
@@ -113,26 +118,26 @@ class PI05Config(PreTrainedConfig):
     def validate_features(self) -> None:
         """Validate and set up input/output features."""
         for i in range(self.empty_cameras):
-            key = f"observation.images.empty_camera_{i}"
+            key = OBS_IMAGES + f".empty_camera_{i}"
             empty_camera = PolicyFeature(
                 type=FeatureType.VISUAL,
                 shape=(3, *self.image_resolution),  # Use configured image resolution
             )
             self.input_features[key] = empty_camera
 
-        if "observation.state" not in self.input_features:
+        if OBS_STATE not in self.input_features:
             state_feature = PolicyFeature(
                 type=FeatureType.STATE,
                 shape=(self.max_state_dim,),  # Padded to max_state_dim
             )
-            self.input_features["observation.state"] = state_feature
+            self.input_features[OBS_STATE] = state_feature
 
-        if "action" not in self.output_features:
+        if ACTION not in self.output_features:
             action_feature = PolicyFeature(
                 type=FeatureType.ACTION,
                 shape=(self.max_action_dim,),  # Padded to max_action_dim
             )
-            self.output_features["action"] = action_feature
+            self.output_features[ACTION] = action_feature
 
     def get_optimizer_preset(self) -> AdamWConfig:
         return AdamWConfig(

src/lerobot/policies/pi05/modeling_pi05.py

@@ -337,10 +337,14 @@ class PaliGemmaWithExpertModel(
         use_adarms=None,
         precision: Literal["bfloat16", "float32"] = "bfloat16",
         image_size: int = DEFAULT_IMAGE_SIZE,
+        freeze_vision_encoder: bool = False,
+        train_expert_only: bool = False,
     ):
         if use_adarms is None:
             use_adarms = [False, False]
         super().__init__()
+        self.freeze_vision_encoder = freeze_vision_encoder
+        self.train_expert_only = train_expert_only
 
         vlm_config_hf = CONFIG_MAPPING["paligemma"]()
         vlm_config_hf._vocab_size = 257152  # noqa: SLF001
@@ -381,6 +385,7 @@ class PaliGemmaWithExpertModel(
         self.gemma_expert.model.embed_tokens = None
 
         self.to_bfloat16_for_selected_params(precision)
+        self._set_requires_grad()
 
     def to_bfloat16_for_selected_params(self, precision: Literal["bfloat16", "float32"] = "bfloat16"):
         if precision == "bfloat16":
@@ -404,6 +409,23 @@ class PaliGemmaWithExpertModel(
             if any(selector in name for selector in params_to_keep_float32):
                 param.data = param.data.to(dtype=torch.float32)
 
+    def _set_requires_grad(self):
+        if self.freeze_vision_encoder:
+            self.paligemma.vision_tower.eval()
+            for param in self.paligemma.vision_tower.parameters():
+                param.requires_grad = False
+        if self.train_expert_only:
+            self.paligemma.eval()
+            for param in self.paligemma.parameters():
+                param.requires_grad = False
+
+    def train(self, mode: bool = True):
+        super().train(mode)
+        if self.freeze_vision_encoder:
+            self.paligemma.vision_tower.eval()
+        if self.train_expert_only:
+            self.paligemma.eval()
+
     def embed_image(self, image: torch.Tensor):
         return self.paligemma.model.get_image_features(image)
 
@@ -531,6 +553,8 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
             use_adarms=[False, True],
             precision=config.dtype,
             image_size=config.image_resolution[0],
+            freeze_vision_encoder=config.freeze_vision_encoder,
+            train_expert_only=config.train_expert_only,
         )
 
         self.action_in_proj = nn.Linear(config.max_action_dim, action_expert_config.width)

src/lerobot/policies/pi0_fast/__init__.py

@@ -1,6 +1,6 @@
 #!/usr/bin/env python
 
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+# Copyright 2025 Physical Intelligence and The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -14,5 +14,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from .config_so100_follower import SO100FollowerConfig
-from .so100_follower import SO100Follower
+from .configuration_pi0_fast import PI0FastConfig
+from .modeling_pi0_fast import PI0FastPolicy
+from .processor_pi0_fast import make_pi0_fast_pre_post_processors
+
+__all__ = ["PI0FastConfig", "PI0FastPolicy", "make_pi0_fast_pre_post_processors"]

src/lerobot/policies/pi0_fast/configuration_pi0_fast.py

@@ -0,0 +1,162 @@
+#!/usr/bin/env python
+
+# Copyright 2025 Physical Intelligence and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass, field
+
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
+from lerobot.optim.optimizers import AdamWConfig
+from lerobot.optim.schedulers import CosineDecayWithWarmupSchedulerConfig
+from lerobot.policies.rtc.configuration_rtc import RTCConfig
+from lerobot.utils.constants import ACTION, OBS_IMAGES, OBS_STATE
+
+DEFAULT_IMAGE_SIZE = 224
+
+
+@PreTrainedConfig.register_subclass("pi0_fast")
+@dataclass
+class PI0FastConfig(PreTrainedConfig):
+    paligemma_variant: str = "gemma_2b"
+    action_expert_variant: str = "gemma_300m"
+    dtype: str = "float32"  # Options: "bfloat16", "float32"
+
+    chunk_size: int = 50  # Number of action steps to predict, in openpi called "action_horizon"
+    n_action_steps: int = 50  # Number of action steps to execute
+
+    # Shorter state and action vectors will be padded to these dimensions
+    max_state_dim: int = 32
+    max_action_dim: int = 32
+    max_action_tokens: int = 256
+
+    # Real-Time Chunking (RTC) configuration
+    rtc_config: RTCConfig | None = None
+
+    image_resolution: tuple[int, int] = (
+        DEFAULT_IMAGE_SIZE,
+        DEFAULT_IMAGE_SIZE,
+    )  # see openpi `preprocessing_pytorch.py`
+
+    # Add empty images. Used to add empty cameras when no image features are present.
+    empty_cameras: int = 0
+
+    tokenizer_max_length: int = 200  # see openpi `__post_init__`
+    text_tokenizer_name: str = "google/paligemma-3b-pt-224"
+    action_tokenizer_name: str = "physical-intelligence/fast"
+    temperature: float = 0.0
+    max_decoding_steps: int = 256
+    fast_skip_tokens: int = 128
+
+    # Whether to validate that decoded action tokens start with "Action: " prefix
+    validate_action_token_prefix: bool = True
+
+    # Whether to use KV cache for faster autoregressive decoding
+    use_kv_cache: bool = True
+
+    normalization_mapping: dict[str, NormalizationMode] = field(
+        default_factory=lambda: {
+            "VISUAL": NormalizationMode.IDENTITY,
+            "STATE": NormalizationMode.MEAN_STD,  # Pi0Fast uses quantiles for state
+            "ACTION": NormalizationMode.MEAN_STD,  # Pi0Fast uses quantiles for action
+        }
+    )
+
+    # Training settings
+    gradient_checkpointing: bool = False  # Enable gradient checkpointing for memory optimization
+    compile_model: bool = False  # Whether to use torch.compile for model optimization
+    compile_mode: str = "max-autotune"  # Torch compile mode
+    device: str | None = None  # Device to use for the model (None = auto-detect)
+
+    # Optimizer settings: see openpi `AdamW`
+    optimizer_lr: float = 2.5e-5  # see openpi `CosineDecaySchedule: peak_lr`
+    optimizer_betas: tuple[float, float] = (0.9, 0.95)
+    optimizer_eps: float = 1e-8
+    optimizer_weight_decay: float = 0.01
+    optimizer_grad_clip_norm: float = 1.0
+
+    # Scheduler settings: see openpi `CosineDecaySchedule`
+    # Note: These will auto-scale if --steps < scheduler_decay_steps
+    # For example, --steps=3000 will scale warmup to 100 and decay to 3000
+    scheduler_warmup_steps: int = 1_000
+    scheduler_decay_steps: int = 30_000
+    scheduler_decay_lr: float = 2.5e-6
+
+    def __post_init__(self):
+        super().__post_init__()
+
+        # Validate configuration
+        if self.n_action_steps > self.chunk_size:
+            raise ValueError(
+                f"n_action_steps ({self.n_action_steps}) cannot be greater than chunk_size ({self.chunk_size})"
+            )
+
+        if self.paligemma_variant not in ["gemma_300m", "gemma_2b"]:
+            raise ValueError(f"Invalid paligemma_variant: {self.paligemma_variant}")
+
+        if self.dtype not in ["bfloat16", "float32"]:
+            raise ValueError(f"Invalid dtype: {self.dtype}")
+
+    def validate_features(self) -> None:
+        """Validate and set up input/output features."""
+        for i in range(self.empty_cameras):
+            key = OBS_IMAGES + f".empty_camera_{i}"
+            empty_camera = PolicyFeature(
+                type=FeatureType.VISUAL,
+                shape=(3, *self.image_resolution),  # Use configured image resolution
+            )
+            self.input_features[key] = empty_camera
+
+        if OBS_STATE not in self.input_features:
+            state_feature = PolicyFeature(
+                type=FeatureType.STATE,
+                shape=(self.max_state_dim,),  # Padded to max_state_dim
+            )
+            self.input_features[OBS_STATE] = state_feature
+
+        if ACTION not in self.output_features:
+            action_feature = PolicyFeature(
+                type=FeatureType.ACTION,
+                shape=(self.max_action_dim,),  # Padded to max_action_dim
+            )
+            self.output_features[ACTION] = action_feature
+
+    def get_optimizer_preset(self) -> AdamWConfig:
+        return AdamWConfig(
+            lr=self.optimizer_lr,
+            betas=self.optimizer_betas,
+            eps=self.optimizer_eps,
+            weight_decay=self.optimizer_weight_decay,
+            grad_clip_norm=self.optimizer_grad_clip_norm,
+        )
+
+    def get_scheduler_preset(self):
+        return CosineDecayWithWarmupSchedulerConfig(
+            peak_lr=self.optimizer_lr,
+            decay_lr=self.scheduler_decay_lr,
+            num_warmup_steps=self.scheduler_warmup_steps,
+            num_decay_steps=self.scheduler_decay_steps,
+        )
+
+    @property
+    def observation_delta_indices(self) -> None:
+        return None
+
+    @property
+    def action_delta_indices(self) -> list:
+        return list(range(self.chunk_size))
+
+    @property
+    def reward_delta_indices(self) -> None:
+        return None

src/lerobot/policies/pi0_fast/processor_pi0_fast.py

@@ -0,0 +1,177 @@
+#!/usr/bin/env python
+
+# Copyright 2025 Physical Intelligence and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from copy import deepcopy
+from dataclasses import dataclass
+from typing import Any
+
+import numpy as np
+import torch
+
+from lerobot.configs.types import PipelineFeatureType, PolicyFeature
+from lerobot.policies.pi0_fast.configuration_pi0_fast import PI0FastConfig
+from lerobot.policies.pi0_fast.modeling_pi0_fast import pad_vector
+from lerobot.processor import (
+    ActionTokenizerProcessorStep,
+    AddBatchDimensionProcessorStep,
+    DeviceProcessorStep,
+    NormalizerProcessorStep,
+    PolicyAction,
+    PolicyProcessorPipeline,
+    ProcessorStep,
+    ProcessorStepRegistry,
+    RenameObservationsProcessorStep,
+    TokenizerProcessorStep,
+    UnnormalizerProcessorStep,
+)
+from lerobot.processor.converters import policy_action_to_transition, transition_to_policy_action
+from lerobot.processor.core import EnvTransition, TransitionKey
+from lerobot.utils.constants import (
+    OBS_STATE,
+    POLICY_POSTPROCESSOR_DEFAULT_NAME,
+    POLICY_PREPROCESSOR_DEFAULT_NAME,
+)
+
+
+@ProcessorStepRegistry.register(name="pi0_fast_prepare_state_tokenizer_processor_step")
+@dataclass
+class Pi0FastPrepareStateAndLanguageTokenizerProcessorStep(ProcessorStep):
+    """
+    Processor step to prepare the state and tokenize the language input.
+    """
+
+    max_state_dim: int = 32
+    task_key: str = "task"
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        transition = transition.copy()
+
+        state = transition.get(TransitionKey.OBSERVATION, {}).get(OBS_STATE)
+        if state is None:
+            raise ValueError("State is required for PI0Fast")
+        tasks = transition.get(TransitionKey.COMPLEMENTARY_DATA, {}).get(self.task_key)
+        if tasks is None:
+            raise ValueError("No task found in complementary data")
+
+        # TODO: check if this necessary
+        state = deepcopy(state)
+
+        # Prepare state (pad to max_state_dim)
+        state = pad_vector(state, self.max_state_dim)
+
+        # State should already be normalized to [-1, 1] by the NormalizerProcessorStep that runs before this step
+        # Discretize into 256 bins (see openpi `PaligemmaTokenizer.tokenize()`)
+        state_np = state.cpu().numpy()
+        discretized_states = np.digitize(state_np, bins=np.linspace(-1, 1, 256 + 1)[:-1]) - 1
+
+        full_prompts = []
+        for i, task in enumerate(tasks):
+            cleaned_text = task.strip().replace("_", " ").replace("\n", " ")
+            state_str = " ".join(map(str, discretized_states[i]))
+            full_prompt = f"Task: {cleaned_text}, State: {state_str};\n"
+            full_prompts.append(full_prompt)
+
+        transition[TransitionKey.COMPLEMENTARY_DATA][self.task_key] = full_prompts
+        # Normalize state to [-1, 1] range if needed (assuming it's already normalized by normalizer processor step!!)
+        # Discretize into 256 bins (see openpi `PaligemmaTokenizer.tokenize()`)
+        return transition
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        """
+        This step does not alter the feature definitions.
+        """
+        return features
+
+
+def make_pi0_fast_pre_post_processors(
+    config: PI0FastConfig,
+    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
+) -> tuple[
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+    PolicyProcessorPipeline[PolicyAction, PolicyAction],
+]:
+    """
+    Constructs pre-processor and post-processor pipelines for the PI0Fast policy.
+
+    The pre-processing pipeline prepares input data for the model by:
+    1. Renaming features to match pretrained configurations.
+    2. Normalizing input and output features based on dataset statistics.
+    3. Adding a batch dimension.
+    4. Appending a newline character to the task description for tokenizer compatibility.
+    5. Tokenizing the text prompt using the PaliGemma tokenizer.
+    6. Moving all data to the specified device.
+
+    The post-processing pipeline handles the model's output by:
+    1. Moving data to the CPU.
+    2. Unnormalizing the output features to their original scale.
+
+    Args:
+        config: The configuration object for the PI0Fast policy.
+        dataset_stats: A dictionary of statistics for normalization.
+        preprocessor_kwargs: Additional arguments for the pre-processor pipeline.
+        postprocessor_kwargs: Additional arguments for the post-processor pipeline.
+
+    Returns:
+        A tuple containing the configured pre-processor and post-processor pipelines.
+    """
+    # Add remaining processors
+    input_steps: list[ProcessorStep] = [
+        RenameObservationsProcessorStep(rename_map={}),  # To mimic the same processor as pretrained one
+        AddBatchDimensionProcessorStep(),
+        # NOTE: NormalizerProcessorStep MUST come before Pi0FastPrepareStateAndLanguageTokenizerProcessorStep
+        # because the tokenizer step expects normalized state in [-1, 1] range for discretization
+        NormalizerProcessorStep(
+            features={**config.input_features, **config.output_features},
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+        ),
+        Pi0FastPrepareStateAndLanguageTokenizerProcessorStep(max_state_dim=config.max_state_dim),
+        TokenizerProcessorStep(
+            tokenizer_name=config.text_tokenizer_name,
+            max_length=config.tokenizer_max_length,
+            padding_side="right",
+            padding="max_length",
+        ),
+        ActionTokenizerProcessorStep(
+            action_tokenizer_name=config.action_tokenizer_name,
+            max_action_tokens=config.max_action_tokens,
+            fast_skip_tokens=config.fast_skip_tokens,
+            paligemma_tokenizer_name=config.text_tokenizer_name,
+        ),
+        DeviceProcessorStep(device=config.device),
+    ]
+
+    output_steps: list[ProcessorStep] = [
+        UnnormalizerProcessorStep(
+            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
+        ),
+        DeviceProcessorStep(device="cpu"),
+    ]
+
+    return (
+        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
+            steps=input_steps,
+            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
+        ),
+        PolicyProcessorPipeline[PolicyAction, PolicyAction](
+            steps=output_steps,
+            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
+            to_transition=policy_action_to_transition,
+            to_output=transition_to_policy_action,
+        ),
+    )

src/lerobot/policies/pretrained.py

@@ -206,6 +206,7 @@ class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
     def push_model_to_hub(
         self,
         cfg: TrainPipelineConfig,
+        peft_model=None,
     ):
         api = HfApi()
         repo_id = api.create_repo(
@@ -216,7 +217,14 @@ class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
         with TemporaryDirectory(ignore_cleanup_errors=True) as tmp:
             saved_path = Path(tmp) / repo_id
 
-            self.save_pretrained(saved_path)  # Calls _save_pretrained and stores model tensors
+            if peft_model is not None:
+                # Since PEFT just forwards calls to `push_model_to_hub`, `self` is not the PeftModel wrapper
+                # but the actual policy which is why we need the PEFT model passed to us to save the adapter.
+                # That also means that we need to store the policy config ourselves since PEFT can't.
+                peft_model.save_pretrained(saved_path)
+                self.config.save_pretrained(saved_path)
+            else:
+                self.save_pretrained(saved_path)  # Calls _save_pretrained and stores model tensors
 
             card = self.generate_model_card(
                 cfg.dataset.repo_id, self.config.type, self.config.license, self.config.tags

src/lerobot/policies/sarm/configuration_sarm.py

@@ -26,6 +26,7 @@ from lerobot.configs.policies import PreTrainedConfig
 from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
 from lerobot.optim.optimizers import AdamWConfig
 from lerobot.optim.schedulers import CosineDecayWithWarmupSchedulerConfig
+from lerobot.utils.constants import OBS_IMAGES, OBS_STATE
 
 
 @PreTrainedConfig.register_subclass("sarm")
@@ -86,8 +87,8 @@ class SARMConfig(PreTrainedConfig):
 
     pretrained_model_path: str | None = None
     device: str | None = None
-    image_key: str = "observation.images.top"  # Key for image used from the dataset
-    state_key: str = "observation.state"
+    image_key: str = OBS_IMAGES + ".top"  # Key for image used from the dataset
+    state_key: str = OBS_STATE
 
     # Populated by the processor (video_features, state_features, text_features)
     input_features: dict = field(default_factory=lambda: {})

src/lerobot/policies/sarm/modeling_sarm.py

@@ -40,6 +40,7 @@ from lerobot.policies.sarm.sarm_utils import (
     normalize_stage_tau,
     pad_state_to_max_dim,
 )
+from lerobot.utils.constants import OBS_STR
 
 
 class StageTransformer(nn.Module):
@@ -721,7 +722,7 @@ class SARMRewardModel(PreTrainedPolicy):
         Returns:
             Tuple of (total_loss, output_dict with loss components)
         """
-        observation = batch.get("observation", batch)
+        observation = batch.get(OBS_STR, batch)
 
         # Extract features
         video_features = observation["video_features"].to(self.device)

src/lerobot/policies/utils.py

@@ -16,7 +16,6 @@
 
 import logging
 from collections import deque
-from typing import Any
 
 import numpy as np
 import torch
@@ -140,7 +139,7 @@ def prepare_observation_for_inference(
 
 
 def build_inference_frame(
-    observation: dict[str, Any],
+    observation: RobotObservation,
     device: torch.device,
     ds_features: dict[str, dict],
     task: str | None = None,

src/lerobot/policies/wall_x/configuration_wall_x.py

@@ -18,6 +18,7 @@ from lerobot.configs.policies import PreTrainedConfig
 from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
 from lerobot.optim.optimizers import AdamWConfig
 from lerobot.optim.schedulers import CosineDecayWithWarmupSchedulerConfig
+from lerobot.utils.constants import ACTION, OBS_STATE
 
 
 @PreTrainedConfig.register_subclass("wall_x")
@@ -105,14 +106,14 @@ class WallXConfig(PreTrainedConfig):
                 "No features of type FeatureType.VISUAL found in input_features."
             )
 
-        if "observation.state" not in self.input_features:
+        if OBS_STATE not in self.input_features:
             state_feature = PolicyFeature(
                 type=FeatureType.STATE,
                 shape=(self.max_state_dim,),  # Padded to max_state_dim
             )
-            self.input_features["observation.state"] = state_feature
+            self.input_features[OBS_STATE] = state_feature
         else:
-            state_shape = self.input_features["observation.state"].shape
+            state_shape = self.input_features[OBS_STATE].shape
             state_dim = state_shape[0] if state_shape else 0
             if state_dim > self.max_state_dim:
                 raise ValueError(
@@ -120,14 +121,14 @@ class WallXConfig(PreTrainedConfig):
                     f"Either reduce state dimension or increase max_state_dim in config."
                 )
 
-        if "action" not in self.output_features:
+        if ACTION not in self.output_features:
             action_feature = PolicyFeature(
                 type=FeatureType.ACTION,
                 shape=(self.max_action_dim,),  # Padded to max_action_dim
             )
-            self.output_features["action"] = action_feature
+            self.output_features[ACTION] = action_feature
         else:
-            action_shape = self.output_features["action"].shape
+            action_shape = self.output_features[ACTION].shape
             action_dim = action_shape[0] if action_shape else 0
             if action_dim > self.max_action_dim:
                 raise ValueError(

src/lerobot/policies/wall_x/modeling_wall_x.py

@@ -1697,7 +1697,7 @@ class WallXPolicy(PreTrainedPolicy):
     config_class = WallXConfig
     name = "wall_x"
 
-    def __init__(self, config: WallXConfig):
+    def __init__(self, config: WallXConfig, **kwargs):
         super().__init__(config)
         config.validate_features()
         self.config = config
@@ -1861,7 +1861,7 @@ class WallXPolicy(PreTrainedPolicy):
                     dim=-1,
                 )
         else:
-            action_dim = self.config.output_features["action"].shape[0]
+            action_dim = self.config.output_features[ACTION].shape[0]
             dof_mask = torch.cat(
                 [
                     torch.ones(
@@ -1977,7 +1977,7 @@ class WallXPolicy(PreTrainedPolicy):
         elif self.config.prediction_mode == "fast":
             output = self.model(
                 **batch,
-                action_dim=self.config.output_features["action"].shape[0],
+                action_dim=self.config.output_features[ACTION].shape[0],
                 pred_horizon=self.config.chunk_size,
                 mode="predict",
                 predict_mode="fast",
@@ -1989,7 +1989,7 @@ class WallXPolicy(PreTrainedPolicy):
         actions = output["predict_action"]
 
         # Unpad actions to actual action dimension
-        action_dim = self.config.output_features["action"].shape[0]
+        action_dim = self.config.output_features[ACTION].shape[0]
         actions = actions[:, :, :action_dim]
 
         return actions

src/lerobot/policies/xvla/processor_xvla.py

@@ -41,6 +41,7 @@ from lerobot.processor.converters import policy_action_to_transition, transition
 from lerobot.processor.core import EnvTransition, TransitionKey
 from lerobot.utils.constants import (
     OBS_IMAGES,
+    OBS_PREFIX,
     OBS_STATE,
     POLICY_POSTPROCESSOR_DEFAULT_NAME,
     POLICY_PREPROCESSOR_DEFAULT_NAME,
@@ -137,8 +138,9 @@ class LiberoProcessorStep(ObservationProcessorStep):
 
                 processed_obs[key] = img
         # Process robot_state into a flat state vector
-        if "observation.robot_state" in processed_obs:
-            robot_state = processed_obs.pop("observation.robot_state")
+        robot_state_str = OBS_PREFIX + "robot_state"
+        if robot_state_str in processed_obs:
+            robot_state = processed_obs.pop(robot_state_str)
 
             # Extract components
             eef_pos = robot_state["eef"]["pos"]  # (B, 3,)
@@ -174,8 +176,8 @@ class LiberoProcessorStep(ObservationProcessorStep):
         state_feats = {}
 
         # add our new flattened state
-        state_feats["observation.state"] = PolicyFeature(
-            key="observation.state",
+        state_feats[OBS_STATE] = PolicyFeature(
+            key=OBS_STATE,
             shape=(20,),
             dtype="float32",
         )
@@ -247,7 +249,7 @@ class XVLAImageScaleProcessorStep(ProcessorStep):
         keys_to_scale = self.image_keys
         if keys_to_scale is None:
             # Auto-detect image keys
-            keys_to_scale = [k for k in obs if k.startswith("observation.images.")]
+            keys_to_scale = [k for k in obs if k.startswith(OBS_IMAGES)]
 
         # Scale each image
         for key in keys_to_scale:
@@ -303,7 +305,7 @@ class XVLAImageToFloatProcessorStep(ProcessorStep):
         keys_to_convert = self.image_keys
         if keys_to_convert is None:
             # Auto-detect image keys
-            keys_to_convert = [k for k in obs if k.startswith("observation.images.")]
+            keys_to_convert = [k for k in obs if k.startswith(OBS_IMAGES)]
 
         # Convert each image
         for key in keys_to_convert:
@@ -376,7 +378,7 @@ class XVLAImageNetNormalizeProcessorStep(ProcessorStep):
         keys_to_normalize = self.image_keys
         if keys_to_normalize is None:
             # Auto-detect image keys
-            keys_to_normalize = [k for k in obs if k.startswith("observation.images.")]
+            keys_to_normalize = [k for k in obs if k.startswith(OBS_IMAGES)]
 
         # Normalize each image
         for key in keys_to_normalize:

src/lerobot/processor/__init__.py

@@ -49,7 +49,6 @@ from .hil_processor import (
     RewardClassifierProcessorStep,
     TimeLimitProcessorStep,
 )
-from .joint_observations_processor import JointVelocityProcessorStep, MotorCurrentProcessorStep
 from .normalize_processor import NormalizerProcessorStep, UnnormalizerProcessorStep, hotswap_stats
 from .observation_processor import VanillaObservationProcessorStep
 from .pipeline import (
@@ -75,7 +74,7 @@ from .policy_robot_bridge import (
     RobotActionToPolicyActionProcessorStep,
 )
 from .rename_processor import RenameObservationsProcessorStep
-from .tokenizer_processor import TokenizerProcessorStep
+from .tokenizer_processor import ActionTokenizerProcessorStep, TokenizerProcessorStep
 
 __all__ = [
     "ActionProcessorStep",
@@ -94,14 +93,12 @@ __all__ = [
     "ImageCropResizeProcessorStep",
     "InfoProcessorStep",
     "InterventionActionProcessorStep",
-    "JointVelocityProcessorStep",
     "make_default_processors",
     "make_default_teleop_action_processor",
     "make_default_robot_action_processor",
     "make_default_robot_observation_processor",
     "MapDeltaActionToRobotActionStep",
     "MapTensorToDeltaActionDictStep",
-    "MotorCurrentProcessorStep",
     "NormalizerProcessorStep",
     "Numpy2TorchActionProcessorStep",
     "ObservationProcessorStep",
@@ -122,6 +119,7 @@ __all__ = [
     "AddBatchDimensionProcessorStep",
     "RobotProcessorPipeline",
     "TokenizerProcessorStep",
+    "ActionTokenizerProcessorStep",
     "Torch2NumpyActionProcessorStep",
     "RobotActionToPolicyActionProcessorStep",
     "PolicyActionToRobotActionProcessorStep",

src/lerobot/processor/converters.py

@@ -23,7 +23,7 @@ from typing import Any
 import numpy as np
 import torch
 
-from lerobot.utils.constants import ACTION, DONE, OBS_PREFIX, REWARD, TRUNCATED
+from lerobot.utils.constants import ACTION, DONE, INFO, OBS_PREFIX, REWARD, TRUNCATED
 
 from .core import EnvTransition, PolicyAction, RobotAction, RobotObservation, TransitionKey
 
@@ -176,7 +176,7 @@ def _extract_complementary_data(batch: dict[str, Any]) -> dict[str, Any]:
 
 
 def create_transition(
-    observation: dict[str, Any] | None = None,
+    observation: RobotObservation | None = None,
     action: PolicyAction | RobotAction | None = None,
     reward: float = 0.0,
     done: bool = False,
@@ -384,7 +384,7 @@ def transition_to_batch(transition: EnvTransition) -> dict[str, Any]:
         REWARD: transition.get(TransitionKey.REWARD, 0.0),
         DONE: transition.get(TransitionKey.DONE, False),
         TRUNCATED: transition.get(TransitionKey.TRUNCATED, False),
-        "info": transition.get(TransitionKey.INFO, {}),
+        INFO: transition.get(TransitionKey.INFO, {}),
     }
 
     # Add complementary data.

src/lerobot/processor/core.py

@@ -45,7 +45,7 @@ RobotObservation: TypeAlias = dict[str, Any]
 EnvTransition = TypedDict(
     "EnvTransition",
     {
-        TransitionKey.OBSERVATION.value: dict[str, Any] | None,
+        TransitionKey.OBSERVATION.value: RobotObservation | None,
         TransitionKey.ACTION.value: PolicyAction | RobotAction | EnvAction | None,
         TransitionKey.REWARD.value: float | torch.Tensor | None,
         TransitionKey.DONE.value: bool | torch.Tensor | None,

src/lerobot/processor/env_processor.py

@@ -18,7 +18,7 @@ from dataclasses import dataclass
 import torch
 
 from lerobot.configs.types import PipelineFeatureType, PolicyFeature
-from lerobot.utils.constants import OBS_IMAGES, OBS_STATE
+from lerobot.utils.constants import OBS_IMAGES, OBS_PREFIX, OBS_STATE, OBS_STR
 
 from .pipeline import ObservationProcessorStep, ProcessorStepRegistry
 
@@ -60,8 +60,9 @@ class LiberoProcessorStep(ObservationProcessorStep):
 
                 processed_obs[key] = img
         # Process robot_state into a flat state vector
-        if "observation.robot_state" in processed_obs:
-            robot_state = processed_obs.pop("observation.robot_state")
+        observation_robot_state_str = OBS_PREFIX + "robot_state"
+        if observation_robot_state_str in processed_obs:
+            robot_state = processed_obs.pop(observation_robot_state_str)
 
             # Extract components
             eef_pos = robot_state["eef"]["pos"]  # (B, 3,)
@@ -98,8 +99,8 @@ class LiberoProcessorStep(ObservationProcessorStep):
         state_feats = {}
 
         # add our new flattened state
-        state_feats["observation.state"] = PolicyFeature(
-            key="observation.state",
+        state_feats[OBS_STATE] = PolicyFeature(
+            key=OBS_STATE,
             shape=(8,),  # [eef_pos(3), axis_angle(3), gripper(2)]
             dtype="float32",
             description=("Concatenated end-effector position (3), axis-angle (3), and gripper qpos (2)."),
@@ -152,3 +153,78 @@ class LiberoProcessorStep(ObservationProcessorStep):
             result[mask] = axis * angle.unsqueeze(1)
 
         return result
+
+
+@dataclass
+@ProcessorStepRegistry.register(name="isaaclab_arena_processor")
+class IsaaclabArenaProcessorStep(ObservationProcessorStep):
+    """
+    Processes IsaacLab Arena observations into LeRobot format.
+
+    **State Processing:**
+    - Extracts state components from obs["policy"] based on `state_keys`.
+    - Concatenates into a flat vector mapped to "observation.state".
+
+    **Image Processing:**
+    - Extracts images from obs["camera_obs"] based on `camera_keys`.
+    - Converts from (B, H, W, C) uint8 to (B, C, H, W) float32 [0, 1].
+    - Maps to "observation.images.<camera_name>".
+    """
+
+    # Configurable from IsaacLabEnv config / cli args: --env.state_keys="robot_joint_pos,left_eef_pos"
+    state_keys: tuple[str, ...]
+
+    # Configurable from IsaacLabEnv config / cli args: --env.camera_keys="robot_pov_cam_rgb"
+    camera_keys: tuple[str, ...]
+
+    def _process_observation(self, observation):
+        """
+        Processes both image and policy state observations from IsaacLab Arena.
+        """
+        processed_obs = {}
+
+        if f"{OBS_STR}.camera_obs" in observation:
+            camera_obs = observation[f"{OBS_STR}.camera_obs"]
+
+            for cam_name, img in camera_obs.items():
+                if cam_name not in self.camera_keys:
+                    continue
+
+                img = img.permute(0, 3, 1, 2).contiguous()
+                if img.dtype == torch.uint8:
+                    img = img.float() / 255.0
+                elif img.dtype != torch.float32:
+                    img = img.float()
+
+                processed_obs[f"{OBS_IMAGES}.{cam_name}"] = img
+
+        # Process policy state -> observation.state
+        if f"{OBS_STR}.policy" in observation:
+            policy_obs = observation[f"{OBS_STR}.policy"]
+
+            # Collect state components in order
+            state_components = []
+            for key in self.state_keys:
+                if key in policy_obs:
+                    component = policy_obs[key]
+                    # Flatten extra dims: (B, N, M) -> (B, N*M)
+                    if component.dim() > 2:
+                        batch_size = component.shape[0]
+                        component = component.view(batch_size, -1)
+                    state_components.append(component)
+
+            if state_components:
+                state = torch.cat(state_components, dim=-1)
+                state = state.float()
+                processed_obs[OBS_STATE] = state
+
+        return processed_obs
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        """Not used for policy evaluation."""
+        return features
+
+    def observation(self, observation):
+        return self._process_observation(observation)

src/lerobot/processor/normalize_processor.py

@@ -30,7 +30,7 @@ from lerobot.utils.constants import ACTION
 
 from .converters import from_tensor_to_numpy, to_tensor
 from .core import EnvTransition, PolicyAction, TransitionKey
-from .pipeline import PolicyProcessorPipeline, ProcessorStep, ProcessorStepRegistry
+from .pipeline import PolicyProcessorPipeline, ProcessorStep, ProcessorStepRegistry, RobotObservation
 
 
 @dataclass
@@ -239,7 +239,7 @@ class _NormalizationMixin:
             config["normalize_observation_keys"] = sorted(self.normalize_observation_keys)
         return config
 
-    def _normalize_observation(self, observation: dict[str, Any], inverse: bool) -> dict[str, Tensor]:
+    def _normalize_observation(self, observation: RobotObservation, inverse: bool) -> dict[str, Tensor]:
         """
         Applies (un)normalization to all relevant features in an observation dictionary.
 

src/lerobot/processor/pipeline.py

@@ -49,7 +49,7 @@ from lerobot.configs.types import PipelineFeatureType, PolicyFeature
 from lerobot.utils.hub import HubMixin
 
 from .converters import batch_to_transition, create_transition, transition_to_batch
-from .core import EnvAction, EnvTransition, PolicyAction, RobotAction, TransitionKey
+from .core import EnvAction, EnvTransition, PolicyAction, RobotAction, RobotObservation, TransitionKey
 
 # Generic type variables for pipeline input and output.
 TInput = TypeVar("TInput")
@@ -1337,7 +1337,7 @@ class DataProcessorPipeline(HubMixin, Generic[TInput, TOutput]):
         return features
 
     # Convenience methods for processing individual parts of a transition.
-    def process_observation(self, observation: dict[str, Any]) -> dict[str, Any]:
+    def process_observation(self, observation: RobotObservation) -> RobotObservation:
         """Processes only the observation part of a transition through the pipeline.
 
         Args:
@@ -1440,7 +1440,7 @@ class ObservationProcessorStep(ProcessorStep, ABC):
     """An abstract `ProcessorStep` that specifically targets the observation in a transition."""
 
     @abstractmethod
-    def observation(self, observation: dict[str, Any]) -> dict[str, Any]:
+    def observation(self, observation: RobotObservation) -> RobotObservation:
         """Processes an observation dictionary. Subclasses must implement this method.
 
         Args:

src/lerobot/processor/tokenizer_processor.py

@@ -23,22 +23,29 @@ token IDs and attention masks, which are then added to the observation dictionar
 
 from __future__ import annotations
 
+import logging
 from dataclasses import dataclass, field
 from typing import TYPE_CHECKING, Any
 
 import torch
 
 from lerobot.configs.types import FeatureType, PipelineFeatureType, PolicyFeature
-from lerobot.utils.constants import OBS_LANGUAGE_ATTENTION_MASK, OBS_LANGUAGE_TOKENS
+from lerobot.utils.constants import (
+    ACTION_TOKEN_MASK,
+    ACTION_TOKENS,
+    OBS_LANGUAGE_ATTENTION_MASK,
+    OBS_LANGUAGE_TOKENS,
+)
 from lerobot.utils.import_utils import _transformers_available
 
-from .core import EnvTransition, TransitionKey
-from .pipeline import ObservationProcessorStep, ProcessorStepRegistry
+from .core import EnvTransition, RobotObservation, TransitionKey
+from .pipeline import ActionProcessorStep, ObservationProcessorStep, ProcessorStepRegistry
 
 # Conditional import for type checking and lazy loading
 if TYPE_CHECKING or _transformers_available:
-    from transformers import AutoTokenizer
+    from transformers import AutoProcessor, AutoTokenizer
 else:
+    AutoProcessor = None
     AutoTokenizer = None
 
 
@@ -132,7 +139,7 @@ class TokenizerProcessorStep(ObservationProcessorStep):
 
         return None
 
-    def observation(self, observation: dict[str, Any]) -> dict[str, Any]:
+    def observation(self, observation: RobotObservation) -> RobotObservation:
         """
         Tokenizes the task description and adds it to the observation dictionary.
 
@@ -268,3 +275,256 @@ class TokenizerProcessorStep(ObservationProcessorStep):
             )
 
         return features
+
+
+@dataclass
+@ProcessorStepRegistry.register(name="action_tokenizer_processor")
+class ActionTokenizerProcessorStep(ActionProcessorStep):
+    """
+    Processor step to tokenize action data using a fast action tokenizer.
+
+    This step takes action tensors from an `EnvTransition`, tokenizes them using
+    a Hugging Face `transformers` AutoProcessor (such as the Physical Intelligence "fast" tokenizer),
+    and returns the tokenized action.
+
+    Requires the `transformers` library to be installed.
+
+    Attributes:
+        tokenizer_name: The name of a pretrained processor from the Hugging Face Hub (e.g., "physical-intelligence/fast").
+        tokenizer: A pre-initialized processor/tokenizer object. If provided, `tokenizer_name` is ignored.
+        trust_remote_code: Whether to trust remote code when loading the tokenizer (required for some tokenizers).
+        action_tokenizer: The internal tokenizer/processor instance, loaded during initialization.
+        paligemma_tokenizer_name: The name of a pretrained PaliGemma tokenizer from the Hugging Face Hub (e.g., "google/paligemma-3b-pt-224").
+    """
+
+    action_tokenizer_name: str | None = None
+    action_tokenizer_input_object: Any | None = None
+    trust_remote_code: bool = True
+    max_action_tokens: int = 256
+    fast_skip_tokens: int = 128
+    paligemma_tokenizer_name: str = "google/paligemma-3b-pt-224"
+    # Internal tokenizer instance (not part of the config)
+    action_tokenizer: Any = field(default=None, init=False, repr=False)
+    _paligemma_tokenizer: Any = field(default=None, init=False, repr=False)
+
+    def __post_init__(self):
+        """
+        Initializes the action tokenizer after the dataclass is created.
+
+        It checks for the availability of the `transformers` library and loads the tokenizer
+        either from a provided object or by name from the Hugging Face Hub.
+
+        Raises:
+            ImportError: If the `transformers` library is not installed.
+            ValueError: If neither `tokenizer` nor `tokenizer_name` is provided.
+        """
+        if not _transformers_available:
+            raise ImportError(
+                "The 'transformers' library is not installed. "
+                "Please install it with `pip install 'lerobot[transformers-dep]'` to use ActionTokenizerProcessorStep."
+            )
+
+        if self.action_tokenizer_input_object is not None:
+            self.action_tokenizer = self.action_tokenizer_input_object
+
+        elif self.action_tokenizer_name is not None:
+            if AutoProcessor is None:
+                raise ImportError("AutoProcessor is not available")
+            self.action_tokenizer = AutoProcessor.from_pretrained(
+                self.action_tokenizer_name, trust_remote_code=self.trust_remote_code
+            )
+        else:
+            raise ValueError(
+                "Either 'action_tokenizer' or 'action_tokenizer_name' must be provided. "
+                "Pass a tokenizer object directly or a tokenizer name to auto-load."
+            )
+
+        self._paligemma_tokenizer = AutoTokenizer.from_pretrained(
+            self.paligemma_tokenizer_name,
+            trust_remote_code=self.trust_remote_code,
+            add_eos_token=True,
+            add_bos_token=False,
+        )
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        """
+        Applies action tokenization to the transition.
+
+        This overrides the base class to handle both tokens and mask.
+
+        Args:
+            transition: The input transition with action data.
+
+        Returns:
+            The processed transition with tokenized actions and mask in complementary data.
+        """
+        self._current_transition = transition.copy()
+        new_transition = self._current_transition
+
+        action = new_transition.get(TransitionKey.ACTION)
+        if action is None:
+            # During inference, no action is available, skip tokenization
+            return new_transition
+
+        # Tokenize and get both tokens and mask
+        tokens, mask = self._tokenize_action(action)
+
+        # Store mask in complementary data
+        complementary_data = new_transition.get(TransitionKey.COMPLEMENTARY_DATA, {})
+        if complementary_data is None:
+            complementary_data = {}
+        complementary_data[ACTION_TOKEN_MASK] = mask
+        complementary_data[ACTION_TOKENS] = tokens
+        new_transition[TransitionKey.COMPLEMENTARY_DATA] = complementary_data
+        return new_transition
+
+    def _act_tokens_to_paligemma_tokens(self, tokens: torch.Tensor) -> torch.Tensor:
+        """
+        Converts action tokens to PaliGemma tokens.
+        """
+        return self._paligemma_tokenizer.vocab_size - 1 - self.fast_skip_tokens - tokens
+
+    def _tokenize_action(self, action: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
+        """
+        Tokenizes the action tensor and creates a mask.
+
+        Args:
+            action: The input action tensor to tokenize. Shape: (B, H, action_dim) or (H, action_dim,)
+
+        Returns:
+            A tuple of (tokens, mask) where:
+            - tokens: Tensor of token IDs with shape (B, max_action_tokens)
+            - mask: Boolean mask with shape (B, max_action_tokens), True for real tokens, False for padding
+        """
+        if action is None:
+            raise ValueError("Action cannot be None")
+
+        # Get the device and dtype of the input action
+        device = action.device if isinstance(action, torch.Tensor) else None
+
+        # Handle single sample (add batch dimension)
+        single_sample = action.dim() == 1
+        if single_sample:
+            action = action.unsqueeze(0)
+
+        batch_size = action.shape[0]
+
+        # Tokenize the action batch
+        # The fast tokenizer expects action data and returns token IDs
+        tokens_list = []
+        masks_list = []
+
+        for i in range(batch_size):
+            # Tokenize single action (move to CPU first as tokenizer uses scipy which requires numpy)
+            action_cpu = action[i : i + 1].cpu()
+            tokens = self.action_tokenizer(action_cpu)
+
+            # Convert to numpy array if it's a list
+            if isinstance(tokens, list) or not isinstance(tokens, torch.Tensor):
+                tokens = torch.tensor(tokens, dtype=torch.long, device=action.device)
+            else:
+                # Move tokens back to the same device as input action
+                tokens = tokens.to(device=action.device)
+
+            # Flatten to 1D if needed
+            if tokens.dim() > 1:
+                tokens = tokens.flatten()
+
+            bos_id = self._paligemma_tokenizer.bos_token_id
+            # add bos
+            tokens = torch.cat(
+                [
+                    torch.tensor([bos_id], device=action.device),
+                    torch.tensor(
+                        self._paligemma_tokenizer.encode("Action: ", add_special_tokens=False),
+                        device=action.device,
+                    ),
+                    self._act_tokens_to_paligemma_tokens(tokens),
+                    torch.tensor(self._paligemma_tokenizer.encode("|"), device=action.device),
+                ]
+            )
+
+            # Truncate or pad to max_action_tokens
+            if len(tokens) > self.max_action_tokens:
+                logging.warning(
+                    f"Token length ({len(tokens)}) exceeds max length ({self.max_action_tokens}), truncating. "
+                    "Consider increasing the `max_action_tokens` in your model config if this happens frequently."
+                )
+                tokens = tokens[: self.max_action_tokens]
+                mask = torch.ones(self.max_action_tokens, dtype=torch.bool, device=action.device)
+            else:
+                mask = torch.cat(
+                    [
+                        torch.ones(len(tokens), dtype=torch.bool, device=action.device),
+                        torch.zeros(
+                            self.max_action_tokens - len(tokens), dtype=torch.bool, device=action.device
+                        ),
+                    ]
+                )
+                # Pad tokens with zeros
+                tokens = torch.nn.functional.pad(tokens, (0, self.max_action_tokens - len(tokens)), value=0)
+
+            tokens_list.append(tokens)
+            masks_list.append(mask)
+
+        # Stack into batched tensors
+        tokens_batch = torch.stack(tokens_list, dim=0)  # (B, max_action_tokens)
+        masks_batch = torch.stack(masks_list, dim=0)  # (B, max_action_tokens)
+
+        # Remove batch dimension if input was single sample
+        if single_sample:
+            tokens_batch = tokens_batch.squeeze(0)
+            masks_batch = masks_batch.squeeze(0)
+
+        # Move to the same device as the input
+        if device is not None:
+            tokens_batch = tokens_batch.to(device)
+            masks_batch = masks_batch.to(device)
+
+        return tokens_batch, masks_batch
+
+    def action(self, action: torch.Tensor) -> torch.Tensor:
+        """
+        This method is not used since we override __call__.
+        Required by ActionProcessorStep ABC.
+        """
+        tokens, _ = self._tokenize_action(action)
+        return tokens
+
+    def get_config(self) -> dict[str, Any]:
+        """
+        Returns the serializable configuration of the processor.
+
+        Note: The tokenizer object itself is not serialized. If the processor was initialized
+        with a tokenizer name, that name will be included in the config.
+
+        Returns:
+            A dictionary with the processor's configuration parameters.
+        """
+        config = {
+            "trust_remote_code": self.trust_remote_code,
+            "max_action_tokens": self.max_action_tokens,
+        }
+
+        # Only save tokenizer_name if it was used to create the tokenizer
+        if self.action_tokenizer_name is not None and self.action_tokenizer_input_object is None:
+            config["action_tokenizer_name"] = self.action_tokenizer_name
+
+        return config
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        """
+        Updates feature definitions to reflect tokenized actions.
+
+        This updates the policy features dictionary to indicate that the action
+        has been tokenized into a sequence of token IDs with shape (max_action_tokens,).
+
+        Args:
+            features: The dictionary of existing policy features.
+
+        Returns:
+            The updated dictionary of policy features.
+        """
+        return features