Merge branch 'main' into feature/add-multitask-dit

* fix(ci): skip HF log in (and tests) in forks and community PRs

* chore(test): remove comment about test meant to be only run locally

* fix(tests): no hf log in decorator for xvla

* fix(test): no decorator in yield

.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/documentation.yml

@@ -18,6 +18,11 @@ name: Documentation
 on:
   # Allows running this workflow manually from the Actions tab
   workflow_dispatch:
+    inputs:
+      version:
+        description: 'Version tag (e.g. v0.1.2) - Leave empty for standard main build'
+        required: false
+        type: string
 
   # Triggers the workflow on push events to main for the docs folder
   push:
@@ -54,7 +59,13 @@ jobs:
     with:
       commit_sha: ${{ github.sha }}
       package: lerobot
-      additional_args: --not_python_module ${{ github.event_name == 'release' && format('--version {0}', github.event.release.tag_name) || '' }}
+      additional_args: >-
+        --not_python_module
+        ${{
+          (github.event_name == 'release' && format('--version {0}', github.event.release.tag_name)) ||
+          (inputs.version != '' && format('--version {0}', inputs.version)) ||
+          ''
+        }}
     secrets:
       token: ${{ secrets.HUGGINGFACE_PUSH }}
       hf_token: ${{ secrets.HF_DOC_BUILD_PUSH }}

.github/workflows/fast_tests.yml

@@ -44,7 +44,7 @@ permissions:
 # Sets up the environment variables
 env:
   UV_VERSION: "0.8.0"
-  PYTHON_VERSION: "3.10"
+  PYTHON_VERSION: "3.12"
 
 # Ensures that only the latest commit for a PR or branch is built, canceling older runs.
 concurrency:
@@ -61,6 +61,7 @@ jobs:
       MUJOCO_GL: egl
       HF_HOME: /mnt/cache/.cache/huggingface
       HF_LEROBOT_HOME: /mnt/cache/.cache/huggingface/lerobot
+      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     steps:
       - uses: actions/checkout@v6
         with:
@@ -89,5 +90,11 @@ jobs:
       - name: Install lerobot with test extras
         run: uv sync --extra "test"
 
+      - name: Login to Hugging Face
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          uv run hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
+          uv run hf auth whoami
+
       - name: Run pytest
         run: uv run pytest tests -vv --maxfail=10

.github/workflows/full_tests.yml

@@ -37,7 +37,7 @@ permissions:
 # Sets up the environment variables
 env:
   UV_VERSION: "0.8.0"
-  PYTHON_VERSION: "3.10"
+  PYTHON_VERSION: "3.12"
   DOCKER_IMAGE_NAME: huggingface/lerobot-gpu
 
 # Ensures that only the latest action is built, canceling older runs.
@@ -60,6 +60,7 @@ jobs:
       MUJOCO_GL: egl
       HF_HOME: /mnt/cache/.cache/huggingface
       HF_LEROBOT_HOME: /mnt/cache/.cache/huggingface/lerobot
+      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     steps:
       - uses: actions/checkout@v6
         with:
@@ -87,6 +88,12 @@ jobs:
       - name: Install lerobot with all extras
         run: uv sync --extra all # TODO(Steven): Make flash-attn optional
 
+      - name: Login to Hugging Face
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          uv run hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
+          uv run hf auth whoami
+
       - name: Run pytest (all extras)
         run: uv run pytest tests -vv --maxfail=10
 
@@ -101,9 +108,11 @@ jobs:
     runs-on:
       group: aws-general-8-plus
     if: |
-      (github.event_name == 'pull_request_review' && github.event.review.state == 'approved' && github.event.pull_request.head.repo.fork == false) ||
-      github.event_name == 'push' ||
-      github.event_name == 'workflow_dispatch'
+      github.repository == 'huggingface/lerobot' && (
+        (github.event_name == 'pull_request_review' && github.event.review.state == 'approved' && github.event.pull_request.head.repo.fork == false) ||
+        github.event_name == 'push' ||
+        github.event_name == 'workflow_dispatch'
+      )
     outputs:
       image_tag: ${{ steps.set_tag.outputs.image_tag }}
     env:
@@ -160,6 +169,7 @@ jobs:
       HF_LEROBOT_HOME: /home/user_lerobot/.cache/huggingface/lerobot
       TORCH_HOME: /home/user_lerobot/.cache/torch
       TRITON_CACHE_DIR: /home/user_lerobot/.cache/triton
+      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     container:
       image: ${{ needs.build-and-push-docker.outputs.image_tag }} # zizmor: ignore[unpinned-images]
       options: --gpus all --shm-size "16gb"
@@ -171,6 +181,13 @@ jobs:
         shell: bash
         working-directory: /lerobot
     steps:
+      - name: Login to Hugging Face
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
+          hf auth whoami
+      - name: Fix ptxas permissions
+        run: chmod +x /lerobot/.venv/lib/python3.12/site-packages/triton/backends/nvidia/bin/ptxas
       - name: Run pytest on GPU
         run: pytest tests -vv --maxfail=10
       - name: Run end-to-end tests
@@ -186,15 +203,18 @@ 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 +225,7 @@ 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/nightly.yml

@@ -28,7 +28,7 @@ on:
 # Sets up the environment variables
 env:
   UV_VERSION: "0.8.0"
-  PYTHON_VERSION: "3.10"
+  PYTHON_VERSION: "3.12"
   DOCKER_IMAGE_NAME_CPU: huggingface/lerobot-cpu:latest
   DOCKER_IMAGE_NAME_GPU: huggingface/lerobot-gpu:latest
 
@@ -119,6 +119,7 @@ jobs:
       HF_LEROBOT_HOME: /home/user_lerobot/.cache/huggingface/lerobot
       TORCH_HOME: /home/user_lerobot/.cache/torch
       TRITON_CACHE_DIR: /home/user_lerobot/.cache/triton
+      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     container:
       image: ${{ needs.build-docker-cpu-nightly.outputs.image_tag }} # zizmor: ignore[unpinned-images]
       options: --shm-size "16gb"
@@ -130,6 +131,11 @@ jobs:
         shell: bash
         working-directory: /lerobot
     steps:
+      - name: Login to Hugging Face
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
+          hf auth whoami
       - name: Run pytest on CPU
         run: pytest tests -vv --maxfail=10
       - name: Run end-to-end tests
@@ -146,6 +152,7 @@ jobs:
       HF_LEROBOT_HOME: /home/user_lerobot/.cache/huggingface/lerobot
       TORCH_HOME: /home/user_lerobot/.cache/torch
       TRITON_CACHE_DIR: /home/user_lerobot/.cache/triton
+      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     container:
       image: ${{ needs.build-docker-gpu-nightly.outputs.image_tag }} # zizmor: ignore[unpinned-images]
       options: --gpus all --shm-size "16gb"
@@ -157,6 +164,11 @@ jobs:
         shell: bash
         working-directory: /lerobot
     steps:
+      - name: Login to Hugging Face
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
+          hf auth whoami
       - name: Run pytest on GPU
         run: pytest tests -vv --maxfail=10
       - name: Run end-to-end tests
@@ -174,6 +186,7 @@ jobs:
       TORCH_HOME: /home/user_lerobot/.cache/torch
       TRITON_CACHE_DIR: /home/user_lerobot/.cache/triton
       CUDA_VISIBLE_DEVICES: "0,1,2,3"
+      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     container:
       image: ${{ needs.build-docker-gpu-nightly.outputs.image_tag }} # zizmor: ignore[unpinned-images]
       options: --gpus all --shm-size "16gb"
@@ -185,12 +198,15 @@ jobs:
         shell: bash
         working-directory: /lerobot
     steps:
+      - name: Login to Hugging Face
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
+          hf auth whoami
       - name: Verify GPU availability
         run: |
           nvidia-smi
           python -c "import torch; print(f'PyTorch CUDA available: {torch.cuda.is_available()}'); print(f'Number of GPUs: {torch.cuda.device_count()}')"
 
       - name: Run multi-GPU training tests
-      # TODO(Steven): Investigate why motors tests are failing in multi-GPU setup
-        run: pytest tests -vv --maxfail=10 --ignore=tests/motors/
-        timeout-minutes: 10
+        run: pytest -vv tests/training/

.github/workflows/quality.yml

@@ -50,7 +50,7 @@ jobs:
       - name: Set up Python
         uses: actions/setup-python@v6
         with:
-          python-version: '3.10'
+          python-version: '3.12'
 
       - name: Run pre-commit hooks
         uses: pre-commit/action@v3.0.1 # zizmor: ignore[unpinned-uses]

.github/workflows/release.yml

@@ -22,7 +22,7 @@ on:
 # Sets up the environment variables
 env:
   UV_VERSION: "0.8.0"
-  PYTHON_VERSION: "3.10"
+  PYTHON_VERSION: "3.12"
 
 jobs:
   # This job builds the Python package and publishes it to PyPI
@@ -45,7 +45,7 @@ jobs:
       - name: Set up Python
         uses: actions/setup-python@v6
         with:
-          python-version: '3.10'
+          python-version: '3.12'
 
       - name: Extract Version
         id: extract_info
@@ -83,14 +83,6 @@ jobs:
             exit 1
           fi
 
-      - name: Remove Tags with Git dependencies
-        # TODO(Steven): Temporary patch to remove pi from PyPi 0.4.0 release due to its reliance on git dependencies.
-        run: |
-          echo "::info:: Checking for Git dependencies to remove from pyproject.toml..."
-          grep -E '@ git\+https|lerobot\[pi\]' pyproject.toml | sed 's/^/::warning:: Removing line: /' || true
-          sed -E -i '/@ git\+https|lerobot\[pi\]/d' pyproject.toml
-          echo "::info:: Git dependencies removed. Proceeding with build."
-
       - name: Install build dependencies
         run: python -m pip install build
 

.github/workflows/unbound_deps_tests.yml

@@ -20,8 +20,8 @@ on:
   workflow_dispatch:
 
   # Run on the 1st and 15th of every month at 09:00 UTC
-  schedule:
-    - cron: '0 2 1,15 * *'
+  # schedule:
+  #  - cron: '0 2 1,15 * *'
 
 permissions:
   contents: read
@@ -29,7 +29,7 @@ permissions:
 # Sets up the environment variables
 env:
   UV_VERSION: "0.8.0"
-  PYTHON_VERSION: "3.10"
+  PYTHON_VERSION: "3.12"
   DOCKER_IMAGE_NAME: huggingface/lerobot-gpu:unbound
 
 # Ensures that only the latest action is built, canceling older runs.
@@ -48,6 +48,7 @@ jobs:
       MUJOCO_GL: egl
       HF_HOME: /mnt/cache/.cache/huggingface
       HF_LEROBOT_HOME: /mnt/cache/.cache/huggingface/lerobot
+      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     steps:
       - uses: actions/checkout@v6
         with:
@@ -79,7 +80,11 @@ jobs:
 
       - name: Install lerobot with all extras
         run: uv sync --extra all # TODO(Steven): Make flash-attn optional
-
+      - name: Login to Hugging Face
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          uv run hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
+          uv run hf auth whoami
       - name: Run pytest (all extras)
         run: uv run pytest tests -vv
 
@@ -91,6 +96,7 @@ jobs:
     name: Build and Push Docker
     runs-on:
       group: aws-general-8-plus
+    if: github.repository == 'huggingface/lerobot'
     outputs:
       image_tag: ${{ env.DOCKER_IMAGE_NAME }}
     env:
@@ -136,6 +142,7 @@ jobs:
       HF_LEROBOT_HOME: /home/user_lerobot/.cache/huggingface/lerobot
       TORCH_HOME: /home/user_lerobot/.cache/torch
       TRITON_CACHE_DIR: /home/user_lerobot/.cache/triton
+      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     container:
       image: ${{ needs.build-and-push-docker.outputs.image_tag }} # zizmor: ignore[unpinned-images]
       options: --gpus all --shm-size "16gb"
@@ -147,6 +154,11 @@ jobs:
         shell: bash
         working-directory: /lerobot
     steps:
+      - name: Login to Hugging Face
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
+          hf auth whoami
       - name: Run pytest on GPU
         run: pytest tests -vv
       - name: Run end-to-end tests
@@ -162,15 +174,19 @@ 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 +197,7 @@ 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"

.pre-commit-config.yaml

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 default_language_version:
-    python: python3.10
+    python: python3.12
 
 exclude: "tests/artifacts/.*\\.safetensors$"
 
@@ -55,7 +55,7 @@ repos:
     rev: v3.21.0
     hooks:
     -   id: pyupgrade
-        args: [--py310-plus]
+        args: [--py312-plus]
 
   ##### Markdown Quality #####
   - repo: https://github.com/rbubley/mirrors-prettier

AI_POLICY.md

@@ -0,0 +1,25 @@
+# AI Usage Policy
+
+The LeRobot project welcomes contributions from everyone, and we have a few guidelines regarding AI usage to ensure high code quality, clear communication, and a healthy open-source ecosystem:
+
+- **Please disclose significant AI assistance.** If you used AI tools (e.g., Copilot, Claude, Cursor, ChatGPT) to generate a substantial portion of your code or text, let us know in your PR description. Transparency helps us review your changes more effectively.
+- **Own your code (The Human-in-the-Loop).** You must fully understand all the changes you are proposing. If you cannot explain what your AI-assisted code does or how it interacts with LeRobot's broader architecture, please take the time to learn and test it before submitting.
+- **Keep issues and discussions focused.** You are welcome to use AI to help draft issues or PR descriptions, but please review and edit them carefully before posting. AI can often be overly verbose; trimming the noise and getting straight to the point helps our maintainers address your needs faster.
+
+Our core maintainers also use AI tools to aid their workflows, but they do so while bringing deep contextual knowledge of the LeRobot codebase to validate the output. We ask all contributors to apply that same level of rigor.
+
+## Remember the Human Maintainers
+
+Please remember that LeRobot is maintained by a dedicated team of humans.
+
+Every discussion, issue, and pull request is read and reviewed by real people. While AI tools can generate thousands of lines of code in seconds, reviewing that code still takes human time and energy. Submitting unverified or low-effort AI output puts an unfair burden on our maintainers.
+
+Today, the quality of the AI output still heavily depends on the developer driving the tool. We ask that you respect our maintainers' time by thoroughly vetting, testing, and refining your submissions.
+
+## AI is Welcome Here
+
+LeRobot operates at the cutting edge of AI and robotics, and many of our maintainers actively embrace AI coding assistants as valuable productivity tools. We are a pro-AI project!
+
+Our reason for having an AI policy is not an anti-AI stance. Rather, it exists to ensure that AI is used to enhance human contributions, not replace them with unverified noise. It's about how the tools are used, not the tools themselves.
+
+We value the unique human insight you bring to the LeRobot community. Let AI empower your workflow, but always let your own judgment take the wheel.

CONTRIBUTING.md

@@ -2,7 +2,7 @@
 
 Everyone is welcome to contribute, and we value everybody's contribution. Code is not the only way to help the community. Answering questions, helping others, reaching out, and improving the documentation are immensely valuable.
 
-Whichever way you choose to contribute, please be mindful to respect our [code of conduct](./CODE_OF_CONDUCT.md).
+Whichever way you choose to contribute, please be mindful to respect our [code of conduct](./CODE_OF_CONDUCT.md) and our [AI policy](./AI_POLICY.md).
 
 ## Ways to Contribute
 
@@ -14,7 +14,7 @@ You can contribute in many ways:
 - **Documentation:** Improve examples, guides, and docstrings.
 - **Feedback:** Submit tickets related to bugs or desired new features.
 
-If you are unsure where to start, join our [Discord Channel](https://discord.gg/JkrYNdmw).
+If you are unsure where to start, join our [Discord Channel](https://discord.gg/q8Dzzpym3f).
 
 ## Development Setup
 

MANIFEST.in

@@ -1,2 +1,3 @@
 include src/lerobot/templates/lerobot_modelcard_template.md
 include src/lerobot/datasets/card_template.md
+include src/lerobot/envs/metaworld_config.json

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), [Multitask DiT Policy](./docs/source/policy_multi_task_dit_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 +128,8 @@ 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.
+- **[Chinese Tutorials: LeRobot+SO-ARM101中文教程-同济子豪兄](https://zihao-ai.feishu.cn/wiki/space/7589642043471924447)** Detailed doc for assembling, teleoperate, dataset, train, deploy. Verified by Seed Studio and 5 global hackathon players.
+- **[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.
 

SECURITY.md

@@ -0,0 +1,48 @@
+# Security Policy
+
+## Project Status & Philosophy
+
+`lerobot` has so far been primarily a research and prototyping tool, which is why deployment security hasn’t been a strong focus until now. As `lerobot` continues to be adopted and deployed in production, we are paying much closer attention to these kinds of issues.
+
+Fortunately, being an open-source project, the community can also help by reporting and fixing vulnerabilities. We appreciate your efforts to responsibly disclose your findings and will make every effort to acknowledge your contributions.
+
+## Reporting a Vulnerability
+
+To report a security issue, please use the GitHub Security Advisory ["Report a Vulnerability"](https://github.com/huggingface/lerobot/security/advisories/new) tab.
+
+The `lerobot` team will send a response indicating the next steps in handling your report. After the initial reply to your report, the security team will keep you informed of the progress towards a fix and full announcement, and may ask for additional information or guidance.
+
+#### Hugging Face Security Team
+
+Since this project is part of the Hugging Face ecosystem, feel free to submit vulnerability reports directly to: **[security@huggingface.co](mailto:security@huggingface.co)**. Someone from the HF security team will review the report and recommend next steps.
+
+#### Open Source Disclosures
+
+If reporting a vulnerability specific to the open-source codebase (and not the underlying Hub infrastructure), you may also use [Huntr](https://huntr.com), a vulnerability disclosure program for open source software.
+
+## Supported Versions
+
+Currently, we treat `lerobot` as a rolling release. We prioritize security updates for the latest available version (`main` branch).
+
+| Version  | Supported |
+| -------- | --------- |
+| Latest   | ✅        |
+| < Latest | ❌        |
+
+## Secure Usage Guidelines
+
+`lerobot` is tightly coupled to the Hugging Face Hub for sharing data and pretrained policies. When downloading artifacts uploaded by others, you expose yourself to risks. Please read below for recommendations to keep your runtime and robot environment safe.
+
+### Remote Artefacts (Weights & Policies)
+
+Models and policies uploaded to the Hugging Face Hub come in different formats. We heavily recommend uploading and downloading models in the [`safetensors`](https://github.com/huggingface/safetensors) format.
+
+`safetensors` was developed specifically to prevent arbitrary code execution on your system, which is critical when running software on physical hardware/robots.
+
+To avoid loading models from unsafe formats (e.g., `pickle`), you should ensure you are prioritizing `safetensors` files.
+
+### Remote Code
+
+Some models or environments on the Hub may require `trust_remote_code=True` to run custom architecture code.
+
+Please **always** verify the content of the modeling files when using this argument. We recommend setting a specific `revision` (commit hash) when loading remote code to ensure you protect yourself from unverified updates to the repository.

benchmarks/video/README.md

@@ -28,9 +28,9 @@ We don't expect the same optimal settings for a dataset of images from a simulat
 For these reasons, we run this benchmark on four representative datasets:
 
 - `lerobot/pusht_image`: (96 x 96 pixels) simulation with simple geometric shapes, fixed camera.
-- `aliberts/aloha_mobile_shrimp_image`: (480 x 640 pixels) real-world indoor, moving camera.
-- `aliberts/paris_street`: (720 x 1280 pixels) real-world outdoor, moving camera.
-- `aliberts/kitchen`: (1080 x 1920 pixels) real-world indoor, fixed camera.
+- `lerobot/aloha_mobile_shrimp_image`: (480 x 640 pixels) real-world indoor, moving camera.
+- `lerobot/paris_street`: (720 x 1280 pixels) real-world outdoor, moving camera.
+- `lerobot/kitchen`: (1080 x 1920 pixels) real-world indoor, fixed camera.
 
 Note: The datasets used for this benchmark need to be image datasets, not video datasets.
 
@@ -179,7 +179,7 @@ python benchmark/video/run_video_benchmark.py \
     --output-dir outputs/video_benchmark \
     --repo-ids \
         lerobot/pusht_image \
-        aliberts/aloha_mobile_shrimp_image \
+        lerobot/aloha_mobile_shrimp_image \
     --vcodec libx264 libx265 \
     --pix-fmt yuv444p yuv420p \
     --g 2 20 None \
@@ -203,9 +203,9 @@ python benchmark/video/run_video_benchmark.py \
     --output-dir outputs/video_benchmark \
     --repo-ids \
         lerobot/pusht_image \
-        aliberts/aloha_mobile_shrimp_image \
-        aliberts/paris_street \
-        aliberts/kitchen \
+        lerobot/aloha_mobile_shrimp_image \
+        lerobot/paris_street \
+        lerobot/kitchen \
     --vcodec libx264 libx265 \
     --pix-fmt yuv444p yuv420p \
     --g 1 2 3 4 5 6 10 15 20 40 None \
@@ -221,9 +221,9 @@ python benchmark/video/run_video_benchmark.py \
     --output-dir outputs/video_benchmark \
     --repo-ids \
         lerobot/pusht_image \
-        aliberts/aloha_mobile_shrimp_image \
-        aliberts/paris_street \
-        aliberts/kitchen \
+        lerobot/aloha_mobile_shrimp_image \
+        lerobot/paris_street \
+        lerobot/kitchen \
     --vcodec libsvtav1 \
     --pix-fmt yuv420p \
     --g 1 2 3 4 5 6 10 15 20 40 None \
@@ -252,37 +252,37 @@ Since we're using av1 encoding, we're choosing the `pyav` decoder as `video_read
 
 These tables show the results for `g=2` and `crf=30`, using `timestamps-modes=6_frames` and `backend=pyav`
 
-| video_images_size_ratio            | vcodec     | pix_fmt |           |           |           |
-| ---------------------------------- | ---------- | ------- | --------- | --------- | --------- |
-|                                    | libx264    |         | libx265   |           | libsvtav1 |
-| repo_id                            | yuv420p    | yuv444p | yuv420p   | yuv444p   | yuv420p   |
-| lerobot/pusht_image                | **16.97%** | 17.58%  | 18.57%    | 18.86%    | 22.06%    |
-| aliberts/aloha_mobile_shrimp_image | 2.14%      | 2.11%   | 1.38%     | **1.37%** | 5.59%     |
-| aliberts/paris_street              | 2.12%      | 2.13%   | **1.54%** | **1.54%** | 4.43%     |
-| aliberts/kitchen                   | 1.40%      | 1.39%   | **1.00%** | **1.00%** | 2.52%     |
+| video_images_size_ratio           | vcodec     | pix_fmt |           |           |           |
+| --------------------------------- | ---------- | ------- | --------- | --------- | --------- |
+|                                   | libx264    |         | libx265   |           | libsvtav1 |
+| repo_id                           | yuv420p    | yuv444p | yuv420p   | yuv444p   | yuv420p   |
+| lerobot/pusht_image               | **16.97%** | 17.58%  | 18.57%    | 18.86%    | 22.06%    |
+| lerobot/aloha_mobile_shrimp_image | 2.14%      | 2.11%   | 1.38%     | **1.37%** | 5.59%     |
+| lerobot/paris_street              | 2.12%      | 2.13%   | **1.54%** | **1.54%** | 4.43%     |
+| lerobot/kitchen                   | 1.40%      | 1.39%   | **1.00%** | **1.00%** | 2.52%     |
 
-| video_images_load_time_ratio       | vcodec  | pix_fmt |          |         |           |
-| ---------------------------------- | ------- | ------- | -------- | ------- | --------- |
-|                                    | libx264 |         | libx265  |         | libsvtav1 |
-| repo_id                            | yuv420p | yuv444p | yuv420p  | yuv444p | yuv420p   |
-| lerobot/pusht_image                | 6.45    | 5.19    | **1.90** | 2.12    | 2.47      |
-| aliberts/aloha_mobile_shrimp_image | 11.80   | 7.92    | 0.71     | 0.85    | **0.48**  |
-| aliberts/paris_street              | 2.21    | 2.05    | 0.36     | 0.49    | **0.30**  |
-| aliberts/kitchen                   | 1.46    | 1.46    | 0.28     | 0.51    | **0.26**  |
+| video_images_load_time_ratio      | vcodec  | pix_fmt |          |         |           |
+| --------------------------------- | ------- | ------- | -------- | ------- | --------- |
+|                                   | libx264 |         | libx265  |         | libsvtav1 |
+| repo_id                           | yuv420p | yuv444p | yuv420p  | yuv444p | yuv420p   |
+| lerobot/pusht_image               | 6.45    | 5.19    | **1.90** | 2.12    | 2.47      |
+| lerobot/aloha_mobile_shrimp_image | 11.80   | 7.92    | 0.71     | 0.85    | **0.48**  |
+| lerobot/paris_street              | 2.21    | 2.05    | 0.36     | 0.49    | **0.30**  |
+| lerobot/kitchen                   | 1.46    | 1.46    | 0.28     | 0.51    | **0.26**  |
 
-|                                    |          | vcodec   | pix_fmt      |          |           |              |
-| ---------------------------------- | -------- | -------- | ------------ | -------- | --------- | ------------ |
-|                                    |          | libx264  |              | libx265  |           | libsvtav1    |
-| repo_id                            | metric   | yuv420p  | yuv444p      | yuv420p  | yuv444p   | yuv420p      |
-| lerobot/pusht_image                | avg_mse  | 2.90E-04 | **2.03E-04** | 3.13E-04 | 2.29E-04  | 2.19E-04     |
-|                                    | avg_psnr | 35.44    | 37.07        | 35.49    | **37.30** | 37.20        |
-|                                    | avg_ssim | 98.28%   | **98.85%**   | 98.31%   | 98.84%    | 98.72%       |
-| aliberts/aloha_mobile_shrimp_image | avg_mse  | 2.76E-04 | 2.59E-04     | 3.17E-04 | 3.06E-04  | **1.30E-04** |
-|                                    | avg_psnr | 35.91    | 36.21        | 35.88    | 36.09     | **40.17**    |
-|                                    | avg_ssim | 95.19%   | 95.18%       | 95.00%   | 95.05%    | **97.73%**   |
-| aliberts/paris_street              | avg_mse  | 6.89E-04 | 6.70E-04     | 4.03E-03 | 4.02E-03  | **3.09E-04** |
-|                                    | avg_psnr | 33.48    | 33.68        | 32.05    | 32.15     | **35.40**    |
-|                                    | avg_ssim | 93.76%   | 93.75%       | 89.46%   | 89.46%    | **95.46%**   |
-| aliberts/kitchen                   | avg_mse  | 2.50E-04 | 2.24E-04     | 4.28E-04 | 4.18E-04  | **1.53E-04** |
-|                                    | avg_psnr | 36.73    | 37.33        | 36.56    | 36.75     | **39.12**    |
-|                                    | avg_ssim | 95.47%   | 95.58%       | 95.52%   | 95.53%    | **96.82%**   |
+|                                   |          | vcodec   | pix_fmt      |          |           |              |
+| --------------------------------- | -------- | -------- | ------------ | -------- | --------- | ------------ |
+|                                   |          | libx264  |              | libx265  |           | libsvtav1    |
+| repo_id                           | metric   | yuv420p  | yuv444p      | yuv420p  | yuv444p   | yuv420p      |
+| lerobot/pusht_image               | avg_mse  | 2.90E-04 | **2.03E-04** | 3.13E-04 | 2.29E-04  | 2.19E-04     |
+|                                   | avg_psnr | 35.44    | 37.07        | 35.49    | **37.30** | 37.20        |
+|                                   | avg_ssim | 98.28%   | **98.85%**   | 98.31%   | 98.84%    | 98.72%       |
+| lerobot/aloha_mobile_shrimp_image | avg_mse  | 2.76E-04 | 2.59E-04     | 3.17E-04 | 3.06E-04  | **1.30E-04** |
+|                                   | avg_psnr | 35.91    | 36.21        | 35.88    | 36.09     | **40.17**    |
+|                                   | avg_ssim | 95.19%   | 95.18%       | 95.00%   | 95.05%    | **97.73%**   |
+| lerobot/paris_street              | avg_mse  | 6.89E-04 | 6.70E-04     | 4.03E-03 | 4.02E-03  | **3.09E-04** |
+|                                   | avg_psnr | 33.48    | 33.68        | 32.05    | 32.15     | **35.40**    |
+|                                   | avg_ssim | 93.76%   | 93.75%       | 89.46%   | 89.46%    | **95.46%**   |
+| lerobot/kitchen                   | avg_mse  | 2.50E-04 | 2.24E-04     | 4.28E-04 | 4.18E-04  | **1.53E-04** |
+|                                   | avg_psnr | 36.73    | 37.33        | 36.56    | 36.75     | **39.12**    |
+|                                   | avg_ssim | 95.47%   | 95.58%       | 95.52%   | 95.53%    | **96.82%**   |

docker/Dockerfile.internal

@@ -24,7 +24,7 @@ ARG OS_VERSION=22.04
 FROM nvidia/cuda:${CUDA_VERSION}-base-ubuntu${OS_VERSION}
 
 # Define Python version argument
-ARG PYTHON_VERSION=3.10
+ARG PYTHON_VERSION=3.12
 
 # Configure environment variables
 ENV DEBIAN_FRONTEND=noninteractive \
@@ -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
@@ -85,6 +85,8 @@ RUN if [ "$UNBOUND_DEPS" = "true" ]; then \
 
 RUN uv pip install --no-cache ".[all]"
 
+RUN chmod +x /lerobot/.venv/lib/python${PYTHON_VERSION}/site-packages/triton/backends/nvidia/bin/ptxas
+
 # Copy the rest of the application source code
 # Make sure to have the git-LFS files for testing
 COPY --chown=user_lerobot:user_lerobot . .

docker/Dockerfile.user

@@ -19,7 +19,7 @@
 # docker run -it --rm lerobot-user
 
 # Configure the base image
-ARG PYTHON_VERSION=3.10
+ARG PYTHON_VERSION=3.12
 FROM python:${PYTHON_VERSION}-slim
 
 # Configure environment variables
@@ -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

@@ -7,8 +7,6 @@
 - sections:
   - local: il_robots
     title: Imitation Learning for Robots
-  - local: cameras
-    title: Cameras
   - local: bring_your_own_policies
     title: Bring Your Own Policies
   - local: integrate_hardware
@@ -19,6 +17,8 @@
     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
@@ -27,6 +27,10 @@
     title: Porting Large Datasets
   - local: using_dataset_tools
     title: Using the Dataset Tools
+  - local: dataset_subtask
+    title: Using Subtasks in the Dataset
+  - local: streaming_video_encoding
+    title: Streaming Video Encoding
   title: "Datasets"
 - sections:
   - local: act
@@ -35,12 +39,16 @@
     title: SmolVLA
   - local: pi0
     title: π₀ (Pi0)
+  - local: pi0fast
+    title: π₀-FAST (Pi0Fast)
   - local: pi05
     title: π₀.₅ (Pi05)
   - local: groot
     title: NVIDIA GR00T N1.5
   - local: xvla
     title: X-VLA
+  - local: multi_task_dit
+    title: Multitask DiT Policy
   - local: walloss
     title: WALL-OSS
   title: "Policies"
@@ -59,6 +67,8 @@
     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
@@ -93,11 +103,19 @@
     title: Unitree G1
   - local: earthrover_mini_plus
     title: Earth Rover Mini
+  - local: omx
+    title: OMX
+  - local: openarm
+    title: OpenArm
   title: "Robots"
 - sections:
   - local: phone_teleop
     title: Phone
   title: "Teleoperators"
+- sections:
+  - local: cameras
+    title: Cameras
+  title: "Sensors"
 - sections:
   - local: torch_accelerators
     title: PyTorch accelerators
@@ -107,6 +125,8 @@
     title: Notebooks
   - local: feetech
     title: Updating Feetech Firmware
+  - local: damiao
+    title: Damiao Motors and CAN Bus
   title: "Resources"
 - sections:
   - local: contributing

docs/source/act.mdx

@@ -88,5 +88,8 @@ lerobot-record \
   --dataset.repo_id=${HF_USER}/eval_act_your_dataset \
   --dataset.num_episodes=10 \
   --dataset.single_task="Your task description" \
+  --dataset.streaming_encoding=true \
+  --dataset.encoder_threads=2 \
+  # --dataset.vcodec=auto \
   --policy.path=${HF_USER}/act_policy
 ```

docs/source/async.mdx

@@ -48,7 +48,7 @@ python -m lerobot.async_inference.robot_client \
     --task="dummy" \ # POLICY: The task to run the policy on (`Fold my t-shirt`). Not necessarily defined for all policies, such as `act`
     --policy_type=your_policy_type \ # POLICY: the type of policy to run (smolvla, act, etc)
     --pretrained_name_or_path=user/model \ # POLICY: the model name/path on server to the checkpoint to run (e.g., lerobot/smolvla_base)
-    --policy_device=mps \ # POLICY: the device to run the policy on, on the server
+    --policy_device=mps \ # POLICY: the device to run the policy on, on the server (cuda, mps, xpu, cpu)
     --actions_per_chunk=50 \ # POLICY: the number of actions to output at once
     --chunk_size_threshold=0.5 \ # CLIENT: the threshold for the chunk size before sending a new observation to the server
     --aggregate_fn_name=weighted_average \ # CLIENT: the function to aggregate actions on overlapping portions
@@ -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
@@ -195,6 +195,7 @@ client_cfg = RobotClientConfig(
     robot=robot_cfg,
     server_address="localhost:8080",
     policy_device="mps",
+    client_device="cpu",
     policy_type="smolvla",
     pretrained_name_or_path="<user>/smolvla_async",
     chunk_size_threshold=0.5,

docs/source/bring_your_own_policies.mdx

@@ -32,7 +32,7 @@ version = "0.1.0"
 dependencies = [
     # your policy-specific dependencies
 ]
-requires-python = ">= 3.11"
+requires-python = ">= 3.12"
 
 [build-system]
 build-backend = # your-build-backend
@@ -82,7 +82,7 @@ Create your policy implementation by inheriting from LeRobot's base `PreTrainedP
 # modeling_my_custom_policy.py
 import torch
 import torch.nn as nn
-from typing import Dict, Any
+from typing import Any
 
 from lerobot.policies.pretrained import PreTrainedPolicy
 from .configuration_my_custom_policy import MyCustomPolicyConfig
@@ -91,7 +91,7 @@ class MyCustomPolicy(PreTrainedPolicy):
     config_class = MyCustomPolicyConfig
     name = "my_custom_policy"
 
-    def __init__(self, config: MyCustomPolicyConfig, dataset_stats: Dict[str, Any] = None):
+    def __init__(self, config: MyCustomPolicyConfig, dataset_stats: dict[str, Any] = None):
         super().__init__(config, dataset_stats)
         ...
 ```
@@ -102,7 +102,7 @@ Create processor functions:
 
 ```python
 # processor_my_custom_policy.py
-from typing import Dict, Any
+from typing import Any
 import torch
 
 

docs/source/cameras.mdx

@@ -1,12 +1,22 @@
 # Cameras
 
-LeRobot offers multiple options for video capture, including phone cameras, built-in laptop cameras, external webcams, and Intel RealSense cameras. To efficiently record frames from most cameras, you can use either the `OpenCVCamera` or `RealSenseCamera` class. For additional compatibility details on the `OpenCVCamera` class, refer to the [Video I/O with OpenCV Overview](https://docs.opencv.org/4.x/d0/da7/videoio_overview.html).
+LeRobot offers multiple options for video capture:
 
-### Finding your camera
+| Class             | Supported Cameras                   |
+| ----------------- | ----------------------------------- |
+| `OpenCVCamera`    | Phone, built-in laptop, USB webcams |
+| `ZMQCamera`       | Network-connected cameras           |
+| `RealSenseCamera` | Intel RealSense (with depth)        |
+| `Reachy2Camera`   | Reachy 2 robot cameras              |
 
-To instantiate a camera, you need a camera identifier. This identifier might change if you reboot your computer or re-plug your camera, a behavior mostly dependant on your operating system.
+> [!TIP]
+> For `OpenCVCamera` compatibility details, see the [Video I/O with OpenCV Overview](https://docs.opencv.org/4.x/d0/da7/videoio_overview.html).
 
-To find the camera indices of the cameras plugged into your system, run the following script:
+### Find your camera
+
+Every camera requires a unique identifier to be instantiated, allowing you to distinguish between multiple connected devices.
+
+`OpenCVCamera` and `RealSenseCamera` support auto-discovery. Run the command below to list available devices and their identifiers. Note that these identifiers may change after rebooting your computer or re-plugging the camera, depending on your operating system.
 
 ```bash
 lerobot-find-cameras opencv # or realsense for Intel Realsense cameras
@@ -14,7 +24,7 @@ lerobot-find-cameras opencv # or realsense for Intel Realsense cameras
 
 The output will look something like this if you have two cameras connected:
 
-```
+```bash
 --- Detected Cameras ---
 Camera #0:
   Name: OpenCV Camera @ 0
@@ -33,13 +43,37 @@ Camera #0:
 > [!WARNING]
 > When using Intel RealSense cameras in `macOS`, you could get this [error](https://github.com/IntelRealSense/librealsense/issues/12307): `Error finding RealSense cameras: failed to set power state`, this can be solved by running the same command with `sudo` permissions. Note that using RealSense cameras in `macOS` is unstable.
 
-## Use Cameras
+`ZMQCamera` and `Reachy2Camera` do not support auto-discovery. They must be configured manually by providing their network address and port or robot SDK settings.
 
-Below are two examples, demonstrating how to work with the API.
+## Use cameras
 
-- **Asynchronous frame capture** using an OpenCV-based camera
+### Frame access modes
+
+All camera classes implement three access modes for capturing frames:
+
+| Method                    | Behavior                                                                                                                                                   | Blocks?        | Best For                                 |
+| ------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------- | ---------------------------------------- |
+| `read()`                  | Waits for the camera hardware to return a frame. May block for a long time depending on the camera and SDK.                                                | Yes            | Simple scripts, sequential capture       |
+| `async_read(timeout_ms)`  | Returns the latest unconsumed frame from background thread. Blocks only if buffer is empty, up to `timeout_ms`. Raises `TimeoutError` if no frame arrives. | With a timeout | Control loops synchronized to camera FPS |
+| `read_latest(max_age_ms)` | Peeks at the most recent frame in buffer (may be stale). Raises `TimeoutError` if frame is older than `max_age_ms`.                                        | No             | UI visualization, logging, monitoring    |
+
+### Usage examples
+
+The following examples show how to use the camera API to configure and capture frames from different camera types.
+
+- **Blocking and non-blocking frame capture** using an OpenCV-based camera
 - **Color and depth capture** using an Intel RealSense camera
 
+> [!WARNING]
+> Failing to cleanly disconnect cameras can cause resource leaks. Use the context manager protocol to ensure automatic cleanup:
+>
+> ```python
+> with OpenCVCamera(config) as camera:
+>     ...
+> ```
+>
+> You can also call `connect()` and `disconnect()` manually, but always use a `finally` block for the latter.
+
 <hfoptions id="shell_restart">
 <hfoption id="Open CV Camera">
 
@@ -60,16 +94,30 @@ config = OpenCVCameraConfig(
 )
 
 # Instantiate and connect an `OpenCVCamera`, performing a warm-up read (default).
-camera = OpenCVCamera(config)
-camera.connect()
+with OpenCVCamera(config) as camera:
+
+    # Read a frame synchronously — blocks until hardware delivers a new frame
+    frame = camera.read()
+    print(f"read() call returned frame with shape:", frame.shape)
+
+    # Read a frame asynchronously with a timeout — returns the latest unconsumed frame or waits up to timeout_ms for a new one
+    try:
+        for i in range(10):
+            frame = camera.async_read(timeout_ms=200)
+            print(f"async_read call returned frame {i} with shape:", frame.shape)
+    except TimeoutError as e:
+        print(f"No frame received within timeout: {e}")
+
+    # Instantly return a frame - returns the most recent frame captured by the camera
+    try:
+        initial_frame = camera.read_latest(max_age_ms=1000)
+        for i in range(10):
+            frame = camera.read_latest(max_age_ms=1000)
+            print(f"read_latest call returned frame {i} with shape:", frame.shape)
+            print(f"Was a new frame received by the camera? {not (initial_frame == frame).any()}")
+    except TimeoutError as e:
+        print(f"Frame too old: {e}")
 
-# Read frames asynchronously in a loop via `async_read(timeout_ms)`
-try:
-    for i in range(10):
-        frame = camera.async_read(timeout_ms=200)
-        print(f"Async frame {i} shape:", frame.shape)
-finally:
-    camera.disconnect()
 ```
 <!-- prettier-ignore-end -->
 
@@ -111,10 +159,10 @@ finally:
 </hfoption>
 </hfoptions>
 
-## Use your phone
+## Use your phone's camera
 
 <hfoptions id="use phone">
-<hfoption id="Mac">
+<hfoption id="iPhone & macOS">
 
 To use your iPhone as a camera on macOS, enable the Continuity Camera feature:
 
@@ -124,83 +172,49 @@ To use your iPhone as a camera on macOS, enable the Continuity Camera feature:
 
 For more details, visit [Apple support](https://support.apple.com/en-gb/guide/mac-help/mchl77879b8a/mac).
 
-Your iPhone should be detected automatically when running the camera setup script in the next section.
-
 </hfoption>
-<hfoption id="Linux">
+<hfoption id="OBS virtual camera">
 
-If you want to use your phone as a camera on Linux, follow these steps to set up a virtual camera
+If you want to use your phone as a camera using OBS, follow these steps to set up a virtual camera.
 
-1. _Install `v4l2loopback-dkms` and `v4l-utils`_. Those packages are required to create virtual camera devices (`v4l2loopback`) and verify their settings with the `v4l2-ctl` utility from `v4l-utils`. Install them using:
+1. _(Linux only) Install `v4l2loopback-dkms` and `v4l-utils`_. These packages create virtual camera devices and verify their settings. Install with:
 
-<!-- prettier-ignore-start -->
-```python
+```bash
 sudo apt install v4l2loopback-dkms v4l-utils
 ```
-<!-- prettier-ignore-end -->
 
-2. _Install [DroidCam](https://droidcam.app) on your phone_. This app is available for both iOS and Android.
-3. _Install [OBS Studio](https://obsproject.com)_. This software will help you manage the camera feed. Install it using [Flatpak](https://flatpak.org):
+2. _Install the [DroidCam app](https://droidcam.app) on your phone_. This app is available for both iOS and Android.
+3. _Download and install [OBS Studio](https://obsproject.com)_.
+4. _Download and install the [DroidCam OBS plugin](https://droidcam.app/obs)_.
+5. _Start OBS Studio_.
 
-<!-- prettier-ignore-start -->
-```python
-flatpak install flathub com.obsproject.Studio
-```
-<!-- prettier-ignore-end -->
-
-4. _Install the DroidCam OBS plugin_. This plugin integrates DroidCam with OBS Studio. Install it with:
-
-<!-- prettier-ignore-start -->
-```python
-flatpak install flathub com.obsproject.Studio.Plugin.DroidCam
-```
-<!-- prettier-ignore-end -->
-
-5. _Start OBS Studio_. Launch with:
-
-<!-- prettier-ignore-start -->
-```python
-flatpak run com.obsproject.Studio
-```
-<!-- prettier-ignore-end -->
-
-6. _Add your phone as a source_. Follow the instructions [here](https://droidcam.app/obs/usage). Be sure to set the resolution to `640x480`.
-7. _Adjust resolution settings_. In OBS Studio, go to `File > Settings > Video`. Change the `Base(Canvas) Resolution` and the `Output(Scaled) Resolution` to `640x480` by manually typing it in.
+6. _Add your phone as a source_. Follow the instructions [here](https://droidcam.app/obs/usage). Be sure to set the resolution to `640x480` to avoid the watermarks.
+7. _Adjust resolution settings_. In OBS Studio, go to `File > Settings > Video` or `OBS > Preferences... > Video`. Change the `Base(Canvas) Resolution` and the `Output(Scaled) Resolution` to `640x480` by manually typing it.
 8. _Start virtual camera_. In OBS Studio, follow the instructions [here](https://obsproject.com/kb/virtual-camera-guide).
-9. _Verify the virtual camera setup_. Use `v4l2-ctl` to list the devices:
+9. _Verify the virtual camera setup and resolution_.
+   - **Linux**: Use `v4l2-ctl` to list devices and check resolution:
+     ```bash
+     v4l2-ctl --list-devices  # find VirtualCam and note its /dev/videoX path
+     v4l2-ctl -d /dev/videoX --get-fmt-video  # replace with your VirtualCam path
+     ```
+     You should see `VirtualCam` listed and resolution `640x480`.
+   - **macOS**: Open Photo Booth or FaceTime and select "OBS Virtual Camera" as the input.
+   - **Windows**: The native Camera app doesn't support virtual cameras. Use a video conferencing app (Zoom, Teams) or run `lerobot-find-cameras opencv` directly to verify.
 
-<!-- prettier-ignore-start -->
-```python
-v4l2-ctl --list-devices
-```
-<!-- prettier-ignore-end -->
+<details>
+<summary><strong>Troubleshooting</strong></summary>
 
-You should see an entry like:
+> The virtual camera resolution is incorrect.
 
-```
-VirtualCam (platform:v4l2loopback-000):
-/dev/video1
-```
+Delete the virtual camera source and recreate it. The resolution cannot be changed after creation.
 
-10. _Check the camera resolution_. Use `v4l2-ctl` to ensure that the virtual camera output resolution is `640x480`. Change `/dev/video1` to the port of your virtual camera from the output of `v4l2-ctl --list-devices`.
+> Error reading frame in background thread for OpenCVCamera(X): OpenCVCamera(X) frame width=640 or height=480 do not match configured width=1920 or height=1080.
 
-<!-- prettier-ignore-start -->
-```python
-v4l2-ctl -d /dev/video1 --get-fmt-video
-```
-<!-- prettier-ignore-end -->
+This error is caused by OBS Virtual Camera advertising a `1920x1080` resolution despite rescaling. The only fix for now is to comment out the width and height check in `_postprocess_image()`.
 
-You should see an entry like:
-
-```
->>> Format Video Capture:
->>>	Width/Height      : 640/480
->>>	Pixel Format      : 'YUYV' (YUYV 4:2:2)
-```
-
-Troubleshooting: If the resolution is not correct you will have to delete the Virtual Camera port and try again as it cannot be changed.
-
-If everything is set up correctly, you can proceed with the rest of the tutorial.
+</details>
 
 </hfoption>
 </hfoptions>
+
+If everything is set up correctly, your phone will appear as a standard OpenCV camera and can be used with `OpenCVCamera`.

docs/source/damiao.mdx

@@ -0,0 +1,165 @@
+# Damiao Motors and CAN Bus
+
+This guide covers setup and usage of Damiao motors with LeRobot via CAN bus communication.
+
+Currently, only Linux is supported, as the OpenArms CAN adapter only has drivers for Linux.
+
+## Linux CAN Setup
+
+Before using Damiao motors, you need to set up the CAN interface on your Linux system.
+
+### Install CAN Utilities
+
+```bash
+sudo apt-get install can-utils
+```
+
+### Configure CAN Interface (Manual)
+
+For standard CAN FD (recommended for OpenArms):
+
+```bash
+sudo ip link set can0 down
+sudo ip link set can0 type can bitrate 1000000 dbitrate 5000000 fd on
+sudo ip link set can0 up
+```
+
+For standard CAN (without FD):
+
+```bash
+sudo ip link set can0 down
+sudo ip link set can0 type can bitrate 1000000
+sudo ip link set can0 up
+```
+
+### Configure CAN Interface (Using LeRobot)
+
+LeRobot provides a utility script to setup and test CAN interfaces:
+
+```bash
+# Setup multiple interfaces (e.g., OpenArms Followers with 2 CAN buses)
+lerobot-setup-can --mode=setup --interfaces=can0,can1
+```
+
+## Debugging CAN Communication
+
+Use the built-in debug tools to test motor communication:
+
+```bash
+# Test motors on all interfaces
+lerobot-setup-can --mode=test --interfaces=can0,can1
+
+# Run speed/latency test
+lerobot-setup-can --mode=speed --interfaces=can0
+```
+
+The test mode will scan for motors (IDs 0x01-0x08) and report which ones respond. Example output:
+
+```
+can0: UP (CAN FD)
+  Motor 0x01 (joint_1): ✓ FOUND
+    → Response 0x11 [FD]: 00112233...
+  Motor 0x02 (joint_2): ✓ FOUND
+  Motor 0x03 (joint_3): ✗ No response
+  ...
+  Summary: 2/8 motors found
+```
+
+## Usage
+
+### Basic Setup
+
+```python
+from lerobot.motors import Motor
+from lerobot.motors.damiao import DamiaoMotorsBus
+
+# Define your motors with send/receive CAN IDs
+motors = {
+    "joint_1": Motor(id=0x01, motor_type_str="dm8009", recv_id=0x11),
+    "joint_2": Motor(id=0x02, motor_type_str="dm4340", recv_id=0x12),
+    "joint_3": Motor(id=0x03, motor_type_str="dm4310", recv_id=0x13),
+}
+
+# Create the bus
+bus = DamiaoMotorsBus(
+    port="can0",  # Linux socketcan interface
+    motors=motors,
+)
+
+# Connect
+bus.connect()
+```
+
+### Reading Motor States
+
+```python
+# Read single motor position (degrees)
+position = bus.read("Present_Position", "joint_1")
+
+# Read from multiple motors
+positions = bus.sync_read("Present_Position")  # All motors
+positions = bus.sync_read("Present_Position", ["joint_1", "joint_2"])
+
+# Read all states at once (position, velocity, torque)
+states = bus.sync_read_all_states()
+# Returns: {'joint_1': {'position': 45.2, 'velocity': 1.3, 'torque': 0.5}, ...}
+```
+
+### Writing Motor Commands
+
+```python
+# Enable torque
+bus.enable_torque()
+
+# Set goal position (degrees)
+bus.write("Goal_Position", "joint_1", 45.0)
+
+# Set positions for multiple motors
+bus.sync_write("Goal_Position", {
+    "joint_1": 45.0,
+    "joint_2": -30.0,
+    "joint_3": 90.0,
+})
+
+# Disable torque
+bus.disable_torque()
+```
+
+## Configuration Options
+
+| Parameter      | Default   | Description                                                 |
+| -------------- | --------- | ----------------------------------------------------------- |
+| `port`         | -         | CAN interface (`can0`) or serial port (`/dev/cu.usbmodem*`) |
+| `use_can_fd`   | `True`    | Enable CAN FD for higher data rates                         |
+| `bitrate`      | `1000000` | Nominal bitrate (1 Mbps)                                    |
+| `data_bitrate` | `5000000` | CAN FD data bitrate (5 Mbps)                                |
+
+## Motor Configuration
+
+Each motor requires:
+
+- `id`: CAN ID for sending commands
+- `motor_type`: One of the supported motor types (e.g., `"dm8009"`, `"dm4340"`)
+- `recv_id`: CAN ID for receiving responses
+
+OpenArms default IDs follow the pattern: send ID `0x0N`, receive ID `0x1N` where N is the joint number.
+
+## Troubleshooting
+
+### No Response from Motors
+
+1. **Check power**
+2. **Verify CAN wiring**: Check CAN-H, CAN-L, and GND connections
+3. **Check motor IDs**: Use Damiao Debugging Tools to verify/configure IDs
+4. **Test CAN interface**: Run `candump can0` to see if messages are being received
+5. **Run diagnostics**: `lerobot-setup-can --mode=test --interfaces=can0`
+
+### Motor Timeout Parameter
+
+If motors were configured with timeout=0, they won't respond to commands. Use Damiao Debugging Tools to set a non-zero timeout value.
+
+### Verify CAN FD Status
+
+```bash
+ip -d link show can0 | grep fd
+```

docs/source/dataset_subtask.mdx

@@ -0,0 +1,278 @@
+# Using Subtasks in LeRobot Datasets
+
+Subtask support in robotics datasets has proven effective in improving robot reasoning and understanding. Subtasks are particularly useful for:
+
+- **Hierarchical policies**: Building policies that include subtask predictions to visualize robot reasoning in real time
+- **Reward modeling**: Helping reward models understand task progression (e.g., SARM-style stage-aware reward models)
+- **Task decomposition**: Breaking down complex manipulation tasks into atomic, interpretable steps
+
+LeRobotDataset now supports subtasks as part of its dataset structure, alongside tasks.
+
+## What are Subtasks?
+
+While a **task** describes the overall goal (e.g., "Pick up the apple and place it in the basket"), **subtasks** break down the execution into finer-grained steps:
+
+1. "Approach the apple"
+2. "Grasp the apple"
+3. "Lift the apple"
+4. "Move to basket"
+5. "Release the apple"
+
+Each frame in the dataset can be annotated with its corresponding subtask, enabling models to learn and predict these intermediate stages.
+
+<img
+  src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/subtask-asset.png"
+  alt="An overview of subtask annotation showing how frames are labeled with intermediate subtask stages"
+  width="80%"
+/>
+
+<p>
+  <em>Figure: Overview of subtask annotation.</em>
+</p>
+
+**Reference:** _Subtask-learning based for robot self-assembly in flexible collaborative assembly in manufacturing_, Original Article, Published: 19 April 2022.
+
+## Dataset Structure
+
+Subtask information is stored in the dataset metadata:
+
+```
+my-dataset/
+├── data/
+│   └── ...
+├── meta/
+│   ├── info.json
+│   ├── stats.json
+│   ├── tasks.parquet
+│   ├── subtasks.parquet      # Subtask index → subtask string mapping
+│   └── episodes/
+│       └── ...
+└── videos/
+    └── ...
+```
+
+### Subtasks Parquet File
+
+The `meta/subtasks.parquet` file maps subtask indices to their natural language descriptions:
+
+| subtask_index | subtask (index column) |
+| ------------- | ---------------------- |
+| 0             | "Approach the apple"   |
+| 1             | "Grasp the apple"      |
+| 2             | "Lift the apple"       |
+| ...           | ...                    |
+
+### Frame-Level Annotations
+
+Each frame in the dataset can include a `subtask_index` field that references the subtasks parquet file:
+
+```python
+# Example frame data in the parquet file
+{
+    "index": 42,
+    "timestamp": 1.4,
+    "episode_index": 0,
+    "task_index": 0,
+    "subtask_index": 2,  # References "Lift the apple"
+    "observation.state": [...],
+    "action": [...],
+}
+```
+
+## Annotating Datasets with Subtasks
+
+We provide a HuggingFace Space for easily annotating any LeRobotDataset with subtasks:
+
+**[https://huggingface.co/spaces/lerobot/annotate](https://huggingface.co/spaces/lerobot/annotate)**
+
+After completing your annotation:
+
+1. Click "Push to Hub" to upload your annotated dataset
+2. You can also run the annotation space locally by following the instructions at [github.com/huggingface/lerobot-annotate](https://github.com/huggingface/lerobot-annotate)
+
+## Loading Datasets with Subtasks
+
+When you load a dataset with subtask annotations, the subtask information is automatically available:
+
+```python
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+
+# Load a dataset with subtask annotations
+dataset = LeRobotDataset("jadechoghari/collect-fruit-annotated")
+
+# Access a sample
+sample = dataset[100]
+
+# The sample includes both task and subtask information
+print(sample["task"])        # "Collect the fruit"
+print(sample["subtask"])     # "Grasp the apple"
+print(sample["task_index"])  # tensor(0)
+print(sample["subtask_index"])  # tensor(2)
+```
+
+### Checking for Subtask Support
+
+You can check if a dataset has subtask annotations:
+
+```python
+# Check if subtasks are available
+has_subtasks = (
+    "subtask_index" in dataset.features
+    and dataset.meta.subtasks is not None
+)
+
+if has_subtasks:
+    print(f"Dataset has {len(dataset.meta.subtasks)} unique subtasks")
+    print("Subtasks:", list(dataset.meta.subtasks.index))
+```
+
+## Using Subtasks for Training
+
+### With the Tokenizer Processor
+
+The `TokenizerProcessor` automatically handles subtask tokenization for Vision-Language Action (VLA) models:
+
+```python
+from lerobot.processor.tokenizer_processor import TokenizerProcessor
+from lerobot.processor.pipeline import ProcessorPipeline
+
+# Create a tokenizer processor
+tokenizer_processor = TokenizerProcessor(
+    tokenizer_name_or_path="google/paligemma-3b-pt-224",
+    padding="max_length",
+    max_length=64,
+)
+
+# The processor will automatically tokenize subtasks if present in the batch
+# and add them to the observation under:
+# - "observation.subtask.tokens"
+# - "observation.subtask.attention_mask"
+```
+
+When subtasks are available in the batch, the tokenizer processor adds:
+
+- `observation.subtask.tokens`: Tokenized subtask text
+- `observation.subtask.attention_mask`: Attention mask for the subtask tokens
+
+### DataLoader with Subtasks
+
+```python
+import torch
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+
+dataset = LeRobotDataset("jadechoghari/collect-fruit-annotated")
+
+dataloader = torch.utils.data.DataLoader(
+    dataset,
+    batch_size=16,
+    shuffle=True,
+)
+
+for batch in dataloader:
+    # Access subtask information in the batch
+    subtasks = batch["subtask"]  # List of subtask strings
+    subtask_indices = batch["subtask_index"]  # Tensor of subtask indices
+
+    # Use for training hierarchical policies or reward models
+    print(f"Batch subtasks: {set(subtasks)}")
+```
+
+## Example Datasets with Subtask Annotations
+
+Try loading a dataset with subtask annotations:
+
+```python
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+
+# Example dataset with subtask annotations
+dataset = LeRobotDataset("jadechoghari/collect-fruit-annotated")
+
+# Explore the subtasks
+print("Available subtasks:")
+for subtask_name in dataset.meta.subtasks.index:
+    print(f"  - {subtask_name}")
+
+# Get subtask distribution
+subtask_counts = {}
+for i in range(len(dataset)):
+    sample = dataset[i]
+    subtask = sample["subtask"]
+    subtask_counts[subtask] = subtask_counts.get(subtask, 0) + 1
+
+print("\nSubtask distribution:")
+for subtask, count in sorted(subtask_counts.items(), key=lambda x: -x[1]):
+    print(f"  {subtask}: {count} frames")
+```
+
+## Use Cases
+
+### 1. Hierarchical Policy Training
+
+Train policies that predict both actions and current subtask:
+
+```python
+class HierarchicalPolicy(nn.Module):
+    def __init__(self, num_subtasks):
+        super().__init__()
+        self.action_head = nn.Linear(hidden_dim, action_dim)
+        self.subtask_head = nn.Linear(hidden_dim, num_subtasks)
+
+    def forward(self, observations):
+        features = self.encoder(observations)
+        actions = self.action_head(features)
+        subtask_logits = self.subtask_head(features)
+        return actions, subtask_logits
+```
+
+### 2. Stage-Aware Reward Modeling (SARM)
+
+Build reward models that understand task progression:
+
+```python
+# SARM predicts:
+# - Stage: Which subtask is being executed (discrete)
+# - Progress: How far along the subtask (continuous 0-1)
+
+class SARMRewardModel(nn.Module):
+    def forward(self, observations):
+        features = self.encoder(observations)
+        stage_logits = self.stage_classifier(features)
+        progress = self.progress_regressor(features)
+        return stage_logits, progress
+```
+
+### 3. Progress Visualization
+
+Monitor robot execution by tracking subtask progression:
+
+```python
+def visualize_execution(model, observations):
+    for t, obs in enumerate(observations):
+        action, subtask_logits = model(obs)
+        predicted_subtask = subtask_names[subtask_logits.argmax()]
+        print(f"t={t}: Executing '{predicted_subtask}'")
+```
+
+## API Reference
+
+### LeRobotDataset Properties
+
+| Property                    | Type                   | Description                                |
+| --------------------------- | ---------------------- | ------------------------------------------ |
+| `meta.subtasks`             | `pd.DataFrame \| None` | DataFrame mapping subtask names to indices |
+| `features["subtask_index"]` | `dict`                 | Feature spec for subtask_index if present  |
+
+### Sample Keys
+
+When subtasks are available, each sample includes:
+
+| Key             | Type           | Description                          |
+| --------------- | -------------- | ------------------------------------ |
+| `subtask_index` | `torch.Tensor` | Integer index of the current subtask |
+| `subtask`       | `str`          | Natural language subtask description |
+
+## Related Resources
+
+- [SARM Paper](https://arxiv.org/pdf/2509.25358) - Stage-Aware Reward Modeling for Long Horizon Robot Manipulation
+- [LeRobot Annotate Space](https://huggingface.co/spaces/lerobot/annotate) - Interactive annotation tool
+- [LeRobotDataset v3.0](./lerobot-dataset-v3) - Dataset format documentation

docs/source/earthrover_mini_plus.mdx

@@ -1,5 +1,11 @@
 # EarthRover Mini Plus
 
+<img
+  src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Earth_Rover_Mini_5_240c9adc-4f9e-44b7-982f-5d1dc24af1d8.png.webp"
+  alt="EarthRover Mini Plus"
+  width="70%"
+/>
+
 The EarthRover Mini Plus is a fully open source mobile robot that connects through the cloud using the Frodobots SDK. This lets you control the robot and record datasets for training AI models.
 
 ## What You Need
@@ -7,28 +13,47 @@ The EarthRover Mini Plus is a fully open source mobile robot that connects throu
 ### Hardware
 
 - EarthRover Mini robot
-- Computer with Python 3.10 or newer
+- Computer with Python 3.12 or newer
 - Internet connection
 
 ### 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:
 
@@ -145,13 +170,13 @@ Once you can drive the robot well, you can start recording data to train AI mode
 We use Hugging Face to store your data online. First, log in with your token from [Hugging Face settings](https://huggingface.co/settings/tokens):
 
 ```bash
-huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+hf auth login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
 ```
 
 Store your Hugging Face username:
 
 ```bash
-HF_USER=$(huggingface-cli whoami | head -n 1)
+HF_USER=$(hf auth whoami | awk -F': *' 'NR==1 {print $2}')
 echo $HF_USER
 ```
 
@@ -160,13 +185,16 @@ echo $HF_USER
 Use the standard recording command:
 
 ```bash
-python src/lerobot/scripts/lerobot_record.py \
+lerobot-record \
     --robot.type=earthrover_mini_plus \
     --teleop.type=keyboard_rover \
     --dataset.repo_id=your_username/dataset_name \
     --dataset.num_episodes=2 \
     --dataset.fps=10 \
     --dataset.single_task="Navigate around obstacles" \
+    --dataset.streaming_encoding=true \
+    --dataset.encoder_threads=2 \
+    # --dataset.vcodec=auto \
     --display_data=true
 ```
 

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:
@@ -148,10 +155,10 @@ Upload your repository to Hugging Face:
 pip install huggingface_hub
 
 # Login to Hugging Face
-huggingface-cli login
+hf auth login
 
 # Create a new repository
-huggingface-cli repo create my-custom-env --type space --org my-org
+hf repo create my-org/my-custom-env
 
 # Initialize git and push
 git init

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 \
@@ -114,9 +120,12 @@ lerobot-record \
   --display_data=true \
   --dataset.repo_id=<user>/eval_groot-bimanual  \
   --dataset.num_episodes=10 \
-  --dataset.single_task="Grab and handover the red cube to the other arm"
-  --policy.path=<user>/groot-bimanual # your trained model
-  --dataset.episode_time_s=30
+  --dataset.single_task="Grab and handover the red cube to the other arm" \
+  --dataset.streaming_encoding=true \
+  --dataset.encoder_threads=2 \
+  # --dataset.vcodec=auto \
+  --policy.path=<user>/groot-bimanual \ # your trained model
+  --dataset.episode_time_s=30 \
   --dataset.reset_time_s=10
 ```
 

docs/source/hope_jr.mdx

@@ -224,12 +224,15 @@ lerobot-record \
     --teleop.port=/dev/tty.usbmodem1201 \
     --teleop.id=right \
     --teleop.side=right \
-    --dataset.repo_id=nepyope/hand_record_test_with_video_data \
+    --dataset.repo_id=<USER>/hand_record_test_with_video_data \
     --dataset.single_task="Hand recording test with video data" \
     --dataset.num_episodes=1 \
     --dataset.episode_time_s=5 \
     --dataset.push_to_hub=true \
     --dataset.private=true \
+    --dataset.streaming_encoding=true \
+    --dataset.encoder_threads=2 \
+    # --dataset.vcodec=auto \
     --display_data=true
 ```
 
@@ -241,7 +244,7 @@ lerobot-replay \
     --robot.port=/dev/tty.usbmodem58760432281 \
     --robot.id=right \
     --robot.side=right \
-    --dataset.repo_id=nepyope/hand_record_test_with_camera \
+    --dataset.repo_id=<USER>/hand_record_test_with_camera \
     --dataset.episode=0
 ```
 
@@ -249,13 +252,13 @@ lerobot-replay \
 
 ```bash
 lerobot-train \
-  --dataset.repo_id=nepyope/hand_record_test_with_video_data \
+  --dataset.repo_id=<USER>/hand_record_test_with_video_data \
   --policy.type=act \
   --output_dir=outputs/train/hopejr_hand \
   --job_name=hopejr \
   --policy.device=mps \
   --wandb.enable=true \
-  --policy.repo_id=nepyope/hand_test_policy
+  --policy.repo_id=<USER>/hand_test_policy
 ```
 
 ### Evaluate
@@ -270,8 +273,11 @@ lerobot-record \
   --robot.side=right \
   --robot.cameras='{"main": {"type": "opencv", "index_or_path": 0, "width": 640, "height": 480, "fps": 30}}' \
   --display_data=false \
-  --dataset.repo_id=nepyope/eval_hopejr \
+  --dataset.repo_id=<USER>/eval_hopejr \
   --dataset.single_task="Evaluate hopejr hand policy" \
   --dataset.num_episodes=10 \
+  --dataset.streaming_encoding=true \
+  --dataset.encoder_threads=2 \
+  # --dataset.vcodec=auto \
   --policy.path=outputs/train/hopejr_hand/checkpoints/last/pretrained_model
 ```

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",
@@ -159,13 +159,13 @@ We use the Hugging Face hub features for uploading your dataset. If you haven't
 Add your token to the CLI by running this command:
 
 ```bash
-huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+hf auth login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
 ```
 
 Then store your Hugging Face repository name in a variable:
 
 ```bash
-HF_USER=$(hf auth whoami | head -n 1)
+HF_USER=$(hf auth whoami | awk -F': *' 'NR==1 {print $2}')
 echo $HF_USER
 ```
 
@@ -185,7 +185,10 @@ lerobot-record \
     --display_data=true \
     --dataset.repo_id=${HF_USER}/record-test \
     --dataset.num_episodes=5 \
-    --dataset.single_task="Grab the black cube"
+    --dataset.single_task="Grab the black cube" \
+    --dataset.streaming_encoding=true \
+    # --dataset.vcodec=auto \
+    --dataset.encoder_threads=2
 ```
 </hfoption>
 <hfoption id="API example">
@@ -195,9 +198,9 @@ 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
@@ -324,7 +327,7 @@ You can look for other LeRobot datasets on the hub by searching for `LeRobot` [t
 You can also push your local dataset to the Hub manually, running:
 
 ```bash
-huggingface-cli upload ${HF_USER}/record-test ~/.cache/huggingface/lerobot/{repo-id} --repo-type dataset
+hf upload ${HF_USER}/record-test ~/.cache/huggingface/lerobot/{repo-id} --repo-type dataset
 ```
 
 #### Record function
@@ -408,8 +411,8 @@ 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 +435,7 @@ 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()
 ```
@@ -488,7 +491,7 @@ If your local computer doesn't have a powerful GPU you could utilize Google Cola
 Once training is done, upload the latest checkpoint with:
 
 ```bash
-huggingface-cli upload ${HF_USER}/act_so101_test \
+hf upload ${HF_USER}/act_so101_test \
   outputs/train/act_so101_test/checkpoints/last/pretrained_model
 ```
 
@@ -496,7 +499,7 @@ You can also upload intermediate checkpoints with:
 
 ```bash
 CKPT=010000
-huggingface-cli upload ${HF_USER}/act_so101_test${CKPT} \
+hf upload ${HF_USER}/act_so101_test${CKPT} \
   outputs/train/act_so101_test/checkpoints/${CKPT}/pretrained_model
 ```
 
@@ -515,6 +518,9 @@ lerobot-record  \
   --display_data=false \
   --dataset.repo_id=${HF_USER}/eval_so100 \
   --dataset.single_task="Put lego brick into the transparent box" \
+  --dataset.streaming_encoding=true \
+  --dataset.encoder_threads=2 \
+  # --dataset.vcodec=auto \
   # <- Teleop optional if you want to teleoperate in between episodes \
   # --teleop.type=so100_leader \
   # --teleop.port=/dev/ttyACM0 \
@@ -531,8 +537,8 @@ 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/installation.mdx

@@ -1,18 +1,20 @@
 # Installation
 
-## Install [`miniforge`](https://conda-forge.org/download/)
+This guide uses conda (via miniforge) to manage environments. If you prefer another environment manager (e.g. `uv`, `venv`), ensure you have Python >=3.12 and ffmpeg installed with the `libsvtav1` encoder, then skip ahead to [Install LeRobot](#step-3-install-lerobot-).
+
+## Step 1: Install [`miniforge`](https://conda-forge.org/download/)
 
 ```bash
 wget "https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-$(uname)-$(uname -m).sh"
 bash Miniforge3-$(uname)-$(uname -m).sh
 ```
 
-## Environment Setup
+## Step 2: Environment Setup
 
-Create a virtual environment with Python 3.10, using conda:
+Create a virtual environment with Python 3.12, using conda:
 
 ```bash
-conda create -y -n lerobot python=3.10
+conda create -y -n lerobot python=3.12
 ```
 
 Then activate your conda environment, you have to do this each time you open a shell to use lerobot:
@@ -38,7 +40,14 @@ conda install ffmpeg -c conda-forge
 >
 > - _[On Linux only]_ If you want to bring your own ffmpeg: Install [ffmpeg build dependencies](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu#GettheDependencies) and [compile ffmpeg from source with libsvtav1](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu#libsvtav1), and make sure you use the corresponding ffmpeg binary to your install with `which ffmpeg`.
 
-## Install LeRobot 🤗
+> [!NOTE]
+> When installing LeRobot inside WSL (Windows Subsystem for Linux), make sure to install `evdev` with the following command:
+>
+> ```bash
+> conda install evdev -c conda-forge
+> ```
+
+## Step 3: Install LeRobot 🤗
 
 ### From Source
 
@@ -81,9 +90,6 @@ _Replace `[...]` with your desired features._
 For a full list of optional dependencies, see:
 https://pypi.org/project/lerobot/
 
-> [!NOTE]
-> For lerobot 0.4.0, if you want to install pi, you will have to do: `pip install "lerobot[pi]@git+https://github.com/huggingface/lerobot.git"`
-
 ### Troubleshooting
 
 If you encounter build errors, you may need to install additional dependencies: `cmake`, `build-essential`, and `ffmpeg libs`.

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

@@ -1,5 +1,11 @@
 # LeKiwi
 
+<img
+  src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/1740517739083.jpeg"
+  alt="LeKiwi"
+  width="70%"
+/>
+
 In the steps below, we explain how to assemble the LeKiwi mobile robot.
 
 ## Source the parts
@@ -204,7 +210,7 @@ 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",
@@ -273,13 +279,13 @@ We use the Hugging Face hub features for uploading your dataset. If you haven't
 Add your token to the CLI by running this command:
 
 ```bash
-huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+hf auth login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
 ```
 
 Then store your Hugging Face repository name in a variable:
 
 ```bash
-HF_USER=$(huggingface-cli whoami | head -n 1)
+HF_USER=$(hf auth whoami | awk -F': *' 'NR==1 {print $2}')
 echo $HF_USER
 ```
 

docs/source/lerobot-dataset-v3.mdx

@@ -41,7 +41,10 @@ lerobot-record \
   --display_data=true \
   --dataset.repo_id=${HF_USER}/record-test \
   --dataset.num_episodes=5 \
-  --dataset.single_task="Grab the black cube"
+  --dataset.single_task="Grab the black cube" \
+  --dataset.streaming_encoding=true \
+  # --dataset.vcodec=auto \
+  --dataset.encoder_threads=2
 ```
 
 See the [recording guide](./il_robots#record-a-dataset) for more details.

docs/source/libero.mdx

@@ -42,6 +42,7 @@ lerobot-eval \
 ```
 
 - `--env.task` picks the suite (`libero_object`, `libero_spatial`, etc.).
+- `--env.task_ids` picks task ids to run (`[0]`, `[1,2,3]`, etc.). Omit this flag (or set it to `null`) to run all tasks in the suite.
 - `--eval.batch_size` controls how many environments run in parallel.
 - `--eval.n_episodes` sets how many episodes to run in total.
 

docs/source/multi_task_dit.mdx

@@ -0,0 +1,340 @@
+# Multitask DiT Policy
+
+Multitask Diffusion Transformer (DiT) Policy is an evolution of the original Diffusion Policy architecture, which leverages a large DiT with text and vision conditioning for multitask robot learning. This implementation supports both diffusion and flow matching objectives for action generation, enabling robots to perform diverse manipulation tasks conditioned on language instructions.
+
+## Model Overview
+
+The model uses:
+
+- **CLIP Vision Encoder**: Processes RGB images from multiple camera views
+- **CLIP Text Encoder**: Encodes language task instructions (frozen weights with learnable projection)
+- **Diffusion Transformer**: Predicts action sequences conditioned on observations and language
+- **Two Objectives**: Supports both diffusion (DDPM/DDIM) and flow matching for action generation
+
+This model is exciting because you can achieve extremely high dexterity, competitive with multi-billion parameter
+VLAs, with only ~450M parameters and significantly less training.
+
+## Installation Requirements
+
+Multitask DiT Policy has additional dependencies. Install it with:
+
+```bash
+pip install lerobot[multi_task_dit]
+```
+
+This will install all necessary dependencies including the HuggingFace Transformers library for CLIP models.
+
+## Usage
+
+To use Multitask DiT in your LeRobot configuration, specify the policy type as:
+
+```python
+policy.type=multi_task_dit
+```
+
+## Training
+
+### Basic Training Command
+
+Here's a complete training command for training Multitask DiT on your dataset:
+
+```bash
+lerobot-train \
+  --dataset.repo_id=YOUR_DATASET \
+  --output_dir=./outputs/multitask_dit_training \
+  --batch_size=32 \
+  --steps=5000 \
+  --save_freq=500 \
+  --log_freq=100 \
+  --policy.type=multi_task_dit \
+  --policy.device=cuda \
+  --policy.repo_id="HF_USER/multitask-dit-your-robot" \
+  --wandb.enable=true
+```
+
+### Recommended Hyperparameters and Dataset Details (30Hz Control Frequency)
+
+For reliable performance, start with these suggested default hyperparameters:
+
+```bash
+lerobot-train \
+  --dataset.repo_id=YOUR_DATASET \
+  --output_dir=./outputs/mutitask_dit_training \
+  --batch_size=320 \
+  --steps=30000 \
+  --policy.type=multi_task_dit \
+  --policy.device=cuda \
+  --policy.horizon=32 \
+  --policy.n_action_steps=24 \
+  --policy.objective=diffusion \
+  --policy.noise_scheduler_type=DDPM \
+  --policy.num_train_timesteps=100 \
+  --policy.repo_id="HF_USER/multitask-dit-your-robot" \
+  --wandb.enable=true
+```
+
+**Key Parameters:**
+
+- **Batch Size**: 192-320 - If you have access to a GPU that can support this, you will get the best training dynamics
+- **Horizon**: 32 - number of action steps to predict, ~1.0 sec at 30Hz
+- **n_action_steps**: 24 - ~0.8 seconds at 30Hz
+- **Objective**: `diffusion` - start with diffusion and experiment with flow matching if generation quality is poor
+- **Training Steps**: >30k steps recommended for a single task
+
+### Training Configuration Parameters
+
+#### Objective Selection
+
+Choose between diffusion and flow matching:
+
+```bash
+# Diffusion objective (default)
+--policy.objective=diffusion \
+--policy.noise_scheduler_type=DDPM \  # or "DDIM"
+--policy.num_train_timesteps=100 \
+--policy.num_inference_steps=10 \  # For faster inference
+--policy.beta_schedule=squaredcos_cap_v2 \  # Noise schedule type
+--policy.prediction_type=epsilon \  # "epsilon" (predict noise) or "sample" (predict clean)
+--policy.clip_sample=true \  # Clip samples during denoising
+--policy.clip_sample_range=1.0  # Clipping range [-x, x]
+
+# Flow matching objective
+--policy.objective=flow_matching \
+--policy.timestep_sampling_strategy=beta \  # or "uniform" | the beta sampling strategy performance appears much better in practice
+--policy.num_integration_steps=100 \
+--policy.integration_method=euler \  # or "rk4"
+--policy.sigma_min=0.0  # Minimum noise in flow interpolation path
+```
+
+#### Transformer Architecture
+
+Adjust model capacity based on dataset size:
+
+```bash
+# Small datasets (< 100 examples)
+--policy.num_layers=4 \
+--policy.hidden_dim=512 \
+--policy.num_heads=8  # should ideally be hidden_dim // 64
+
+# Medium datasets (100-5k examples) - default
+--policy.num_layers=6 \
+--policy.hidden_dim=512 \
+--policy.num_heads=8  # should ideally be hidden_dim // 64
+
+# Large datasets (> 5k examples)
+--policy.num_layers=8 \
+--policy.hidden_dim=512 \
+--policy.num_heads=8   # should ideally be hidden_dim // 64
+```
+
+**Positional Encoding Options:**
+
+The model supports two positional encoding methods for action sequences:
+
+```bash
+# Rotary Position Embedding (RoPE) - default, recommended
+--policy.use_rope=true \
+--policy.rope_base=10000.0  # Base frequency for RoPE
+
+# Absolute positional encoding
+--policy.use_positional_encoding=true  # Disables RoPE when true
+```
+
+**Other Transformer Parameters:**
+
+```bash
+--policy.dropout=0.1  # Dropout rate for DiT blocks (0.0-1.0)
+--policy.timestep_embed_dim=256  # Timestep embedding dimension
+```
+
+#### Vision Encoder Configuration
+
+```bash
+# Use different CLIP model for more expressivity at the cost of inference time
+# experiment with larger or smaller models depending on the complexity of your tasks and size of dataset
+--policy.vision_encoder_name=openai/clip-vit-large-patch14
+
+# Use separate vision encoder per camera
+# This may be useful when cameras have significantly different characteristics, but
+# be wary of increased VRAM footprint.
+--policy.use_separate_rgb_encoder_per_camera=true
+
+# Image preprocessing
+--policy.image_resize_shape=[XXX,YYY] \ # you may need to resize your images for inference speed ups
+--policy.image_crop_shape=[224,224] \
+--policy.image_crop_is_random=true  # Random during training, center at inference
+```
+
+#### Text Encoder Configuration
+
+```bash
+# Use different CLIP text encoder model
+# same as vision: experiment with larger or smaller models depending on the
+# complexity of your tasks and size of dataset
+--policy.text_encoder_name=openai/clip-vit-large-patch14
+```
+
+#### Learning Rate Configuration
+
+The vision encoder uses a separate learning rate multiplier, where 1/10th is suggested to be the ideal staritng point:
+
+```bash
+--policy.optimizer_lr=2e-5 \
+--policy.vision_encoder_lr_multiplier=0.1  # Vision encoder LR = 0.1 * optimizer_lr
+```
+
+### Training Tuning Guidelines
+
+#### 1. Flow Matching with Beta Sampling
+
+The original diffusion implementation here is based on the work described in [TRI's LBM paper](https://arxiv.org/abs/2507.05331)
+
+Additionally, we have implemented a flow-matching objective, which is described at a high-level in [Boston Dynamics blog post](https://bostondynamics.com/blog/large-behavior-models-atlas-find-new-footing/).
+
+Consider testing the flow-matching objective and evaluating performance differences for your task:
+
+```bash
+--policy.objective=flow_matching \
+--policy.timestep_sampling_strategy=beta \
+--policy.timestep_sampling_alpha=1.5 \
+--policy.timestep_sampling_beta=1.0 \
+--policy.timestep_sampling_s=0.999
+```
+
+This hasn't been shown to be a silver bullet across every user case, but it occasionally results in smoother and more consistent actions.
+
+#### 2. Number of Transformer Layers
+
+Match model capacity to your dataset size:
+
+- **Small datasets** (< 100 examples): Reduce to 4 layers
+- **Large datasets** (> 5k examples): Increase to 8 layers
+
+#### 3. `horizon` Tuning
+
+The model can be sensitive to the horizon you choose. Start with around a 1 second horizon based on your control frequency:
+
+- **30 Hz frequency**: `horizon=30`
+- **10 Hz frequency**: `horizon=10`
+
+Then experiment with increasing from there. The horizon determines how far into the future the model predicts actions.
+
+#### 4. `n_action_steps` Sensitivity
+
+The model can also be very sensitive to `n_action_steps`. Start with it being around 0.8 seconds based on your control frequency and tune from there:
+
+- **Lower values**: More reactive but potentially less stable for long-horizon tasks
+- **Higher values**: Better for long-horizon execution but open-loop failures are limited in their recovery
+
+### Inference Tuning
+
+For faster inference, use DDIM with fewer sampling steps:
+
+```bash
+--policy.noise_scheduler_type=DDIM \
+--policy.num_inference_steps=10
+```
+
+### Resuming Training
+
+To resume training from a checkpoint:
+
+```bash
+lerobot-train \
+  --config_path=./outputs/mutitask_dit_training/checkpoints/last/pretrained_model/train_config.json \
+  --resume=true
+```
+
+The checkpoint directory should contain `model.safetensors` and `config.json` files (saved automatically during training). When resuming, the configuration is loaded from the checkpoint, so you don't need to specify other parameters.
+
+## Common Failure Modes and Debugging
+
+Training these models can be finicky. Here are common failure modes and debugging approaches:
+
+### Idling / No Motion
+
+The model may "collapse" during inference, resulting in static or no motion. This can occur when:
+
+1. **Insufficient training data**: If you only have 20-50 examples, try to roughly double your dataset size. Once you have above 300 examples, if you're still seeing this, the task may be too complex.
+
+2. **Multiple similar tasks**: When your dataset contains multiple similar tasks (e.g., picking up 2 different objects), the model may rely too heavily on language conditioning which might not be rich enough.
+
+**Debugging tips:**
+
+- Increase dataset size (double until you get to over 300 examples)
+- Train for longer, up to 100k steps, even when the loss flatlines
+- Check if the model is receiving proper language instructions or increase diversity of instruction
+
+### Executing the Wrong Task
+
+Sometimes the robot will completely ignore your instruction and perform some other task. This generally only happens if you have trained on multiple tasks.
+
+**Potential causes:**
+
+- Language instruction ambiguity
+- Insufficient task-specific training data
+- Model confusion between similar tasks in the multitask dataset
+
+**Debugging tips:**
+
+- Verify language instruction specificity, especially if descriptions are similar between multiple tasks
+- Check task distribution in your training dataset and add weighting to the failing/ignored task
+- Consider task-specific fine-tuning
+
+### Training Instability
+
+If training loss is unstable or diverging:
+
+- Try adjusting learning rate between `1e-5` and `3e-4`
+- Increase batch size if possible
+- Check that your dataset normalization is correct
+- Verify image preprocessing is working correctly
+
+## Performance Considerations
+
+### GPU Requirements
+
+- **Inference**: At least an RTX 5070 Ti (or equivalent GPU) is recommended for reasonable speed performance
+- **Training**: A GPU with enough VRAM to load batch sizes of >64 is ideal, which will vary depending on the number of image observations, etc
+
+### Batch Size Recommendations
+
+- **Minimum**: 64 (less than this may result in unstable training)
+- **Recommended**: 256-320 (best performance, requires larger GPU)
+
+## Example: Training on Custom Dataset
+
+Here's a complete example training on a custom dataset:
+
+```bash
+lerobot-train \
+  --dataset.repo_id=YOUR_DATASET \
+  --output_dir=./outputs/mutitask_dit_training \
+  --batch_size=320 \
+  --steps=30000 \
+  --save_freq=1000 \
+  --log_freq=100 \
+  --eval_freq=1000 \
+  --policy.type=multi_task_dit \
+  --policy.device=cuda \
+  --policy.horizon=32 \
+  --policy.n_action_steps=24 \
+  --policy.objective=diffusion \
+  --policy.noise_scheduler_type=DDPM \
+  --policy.num_layers=6 \
+  --policy.hidden_dim=512 \
+  --policy.vision_encoder_name=openai/clip-vit-base-patch16 \
+  --policy.image_resize_shape=[320,240] \
+  --policy.image_crop_shape=[224,224] \
+  --policy.repo_id="HF_USER/multitask-dit-your-robot" \
+  --wandb.enable=true \
+  --wandb.project=multitask_dit
+```
+
+## References
+
+For more details on the technical implementation and architecture, see:
+
+- [A Careful Examination of Large Behavior Models for Multitask Dexterous Manipulation](https://arxiv.org/abs/2507.05331)
+- [Large Behavior Models and Atlas Find New Footing](https://bostondynamics.com/blog/large-behavior-models-atlas-find-new-footing/)
+- [Dissecting and Open-Sourcing Multitask Diffusion Transformer Policy](https://brysonkjones.substack.com/p/dissecting-and-open-sourcing-multitask-diffusion-transformer-policy)

docs/source/omx.mdx

@@ -0,0 +1,197 @@
+## Order and Assemble the parts
+
+First, assemble the OMX hardware following the official assembly guide.
+
+OMX Assembly Guide: https://ai.robotis.com/omx/assembly_guide_omx.html
+
+OMX robots are shipped preconfigured from the factory. Motor IDs, communication parameters, and joint offsets are already set, so no additional motor setup or calibration is required before using LeRobot.
+
+## Install LeRobot 🤗
+
+To install LeRobot, follow our [Installation Guide](./installation)
+
+In addition to these instructions, you need to install the Dynamixel SDK:
+
+```bash
+pip install -e ".[dynamixel]"
+```
+
+## Connect the robot
+
+To find the port for each bus servo adapter, run this script:
+
+```bash
+lerobot-find-port
+```
+
+This command runs and when prompted, disconnect the USB cable from either the leader or follower arm and press Enter. The output will show 'The port of this MotorsBus is [port]'. This identifies the port for the disconnected arm. Repeat for the other arm to identify both ports.
+
+<hfoptions id="find_port">
+<hfoption id="Mac">
+
+Example output on macOS:
+
+```
+Finding all available ports for the MotorBus.
+['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
+Remove the USB cable from your MotorsBus and press Enter when done.
+
+[...Disconnect corresponding leader or follower arm and press Enter...]
+
+The port of this MotorsBus is /dev/tty.usbmodem575E0032081
+Reconnect the USB cable.
+```
+
+Where the found port is: `/dev/tty.usbmodem575E0032081` corresponding to your leader or follower arm.
+
+</hfoption>
+<hfoption id="Linux">
+
+On Linux, we strongly recommend using udev rules to assign persistent and human-readable device names to the OMX leader and follower arms. This avoids issues where device names such as ttyACM0 and ttyACM1 change when the robot is unplugged, replugged, or when the system is rebooted.
+
+#### 1. Find your device serial numbers
+
+You should have obtained the port numbers like ../../ttyACM? for the leader and follower using `lerobot-find-port`. You can match those results with the serial numbers using the `ls -l /dev/serial/by-id/` command.
+To create udev rules, you need the unique serial number for each OMX device. The easiest way is to list devices under:
+
+```bash
+ls -l /dev/serial/by-id/
+```
+
+You will see output similar to:
+
+```bash
+usb-ROBOTIS_OpenRB-150_228BDD7B503059384C2E3120FF0A2B19-if00 -> ../../ttyACM0
+usb-ROBOTIS_OpenRB-150_67E1ED68503059384C2E3120FF092234-if00 -> ../../ttyACM1
+```
+
+In each line, the serial number is the long string after `usb-ROBOTIS_OpenRB-150_` and before `-if00`.
+
+Follower serial: `228BDD7B503059384C2E3120FF0A2B19`
+
+Leader serial: `67E1ED68503059384C2E3120FF092234`
+
+#### 2. Create the udev rule
+
+Create a new udev rule file:
+
+```bash
+sudo nano /etc/udev/rules.d/99-omx.rules
+```
+
+Paste the following lines, replacing the serial numbers with the values you found above:
+
+```bash
+SUBSYSTEM=="tty", ATTRS{idVendor}=="0403", ATTRS{serial}=="228BDD7B503059384C2E3120FF0A2B19", SYMLINK+="omx_follower"
+SUBSYSTEM=="tty", ATTRS{idVendor}=="0403", ATTRS{serial}=="67E1ED68503059384C2E3120FF092234", SYMLINK+="omx_leader"
+```
+
+Save the file and reload udev rules:
+
+```bash
+sudo udevadm control --reload-rules
+sudo udevadm trigger
+```
+
+Now unplug and replug both devices once.
+
+#### 3. Verify the symlinks
+
+Check that the persistent device names exist:
+
+```bash
+ls -l /dev/omx_follower /dev/omx_leader
+```
+
+You should see them pointing to ttyACM\* devices:
+
+```bash
+/dev/omx_follower -> ttyACM*
+/dev/omx_leader   -> ttyACM*
+```
+
+These names remain stable across reboots and reconnections.
+
+</hfoption>
+</hfoptions>
+
+## Teleoperate
+
+After identifying the correct ports, you can directly teleoperate the follower arm using the leader arm.
+
+<hfoptions id="teleoperate">
+<hfoption id="Mac">
+
+### Teleoperate without camera
+
+```bash
+lerobot-teleoperate \
+  --robot.type=omx_follower \
+  --robot.port=<your_follower_port> \
+  --robot.id=omx_follower_arm \
+  --teleop.type=omx_leader \
+  --teleop.port=<your_leader_port> \
+  --teleop.id=omx_leader_arm
+```
+
+During teleoperation, motions of the leader arm are mirrored in real time by the follower arm. OMX is already preconfigured, teleoperation can begin immediately without any calibration steps.
+
+### Teleoperate with camera
+
+You can also enable camera input during teleoperation by providing a camera configuration for the follower arm.
+
+```bash
+lerobot-teleoperate \
+  --robot.type=omx_follower \
+  --robot.port=<your_follower_port> \
+  --robot.id=omx_follower_arm \
+  --robot.cameras="{front: {type: opencv, index_or_path: '/dev/video0', width: 640, height: 480, fps: 30}}" \
+  --teleop.type=omx_leader \
+  --teleop.port=<your_leader_port> \
+  --teleop.id=omx_leader_arm \
+  --display_data=true
+```
+
+When the camera is enabled, the camera stream is displayed in real time and synchronized with the robot state. This setup is useful for visual monitoring and can be reused later for demonstration recording and imitation learning.
+
+</hfoption>
+<hfoption id="Linux">
+
+### Teleoperate without camera
+
+```bash
+lerobot-teleoperate \
+  --robot.type=omx_follower \
+  --robot.port=/dev/omx_follower \
+  --robot.id=omx_follower_arm \
+  --teleop.type=omx_leader \
+  --teleop.port=/dev/omx_leader \
+  --teleop.id=omx_leader_arm
+```
+
+During teleoperation, motions of the leader arm are mirrored in real time by the follower arm. OMX is already preconfigured, teleoperation can begin immediately without any calibration steps.
+
+### Teleoperate with camera
+
+You can also enable camera input during teleoperation by providing a camera configuration for the follower arm.
+
+```bash
+lerobot-teleoperate \
+  --robot.type=omx_follower \
+  --robot.port=/dev/omx_follower \
+  --robot.id=omx_follower_arm \
+  --robot.cameras="{front: {type: opencv, index_or_path: '/dev/video0', width: 640, height: 480, fps: 30}}" \
+  --teleop.type=omx_leader \
+  --teleop.port=/dev/omx_leader \
+  --teleop.id=omx_leader_arm \
+  --display_data=true
+```
+
+When the camera is enabled, the camera stream is displayed in real time and synchronized with the robot state. This setup is useful for visual monitoring and can be reused later for demonstration recording and imitation learning.
+
+</hfoption>
+</hfoptions>
+
+Congrats 🎉, your robot is all set to learn a task on its own.
+
+> If you have any questions or need help, please reach out on [Discord](https://discord.com/invite/robotis).

docs/source/openarm.mdx

@@ -0,0 +1,276 @@
+# OpenArm
+
+[OpenArm](https://openarm.dev) is an open-source 7DOF humanoid arm designed for physical AI research and deployment.
+
+To get your OpenArm, assembled or DIY, and join the global community, browse verified and certified manufacturers worldwide at [openarm.dev](https://openarm.dev).
+
+## What's Unique?
+
+- **Human-Scale Design**: OpenArm is designed with human-like proportions, scaled for a person around 160-165cm tall. This provides an optimal balance between practical reach and manageable inertia for safe, responsive operation.
+
+- **Safety-First Architecture**: Built with QDD backdrivable motors and high compliance, OpenArm prioritizes safe human-robot interaction while maintaining practical payload capabilities (6.0kg peak / 4.1kg nominal) for real-world tasks.
+
+- **Built for Durability**: Critical structural components use aluminum and stainless steel construction, ensuring robust performance for repetitive data collection and continuous research use.
+
+- **Fully Accessible & Buildable**: Every component, from CNC parts and 3D-printed casings to electrical wiring is designed to be purchasable and buildable by individual researchers and labs, with complete fabrication data provided.
+
+- **Practical & Affordable**: At $6,500 USD for a complete bimanual system, OpenArm delivers research-grade capabilities at a fraction of traditional humanoid robot costs.
+
+## Platform Requirements
+
+<Tip warning={true}>
+  **Linux Only**: OpenArm currently only works on Linux. The CAN bus USB adapter
+  does not have macOS drivers and has not been tested on Windows.
+</Tip>
+
+## Safety Guide
+
+Before operating OpenArm, please read the [official safety guide](https://docs.openarm.dev/getting-started/safety-guide). Key points:
+
+- **Secure installation**: Fasten the arm to a flat, stable surface with screws or clamps
+- **Safe distance**: Keep body parts and objects outside the range of motion during operation
+- **Protective equipment**: Always wear safety goggles; use additional PPE as needed
+- **Payload limits**: Do not exceed specified payload limits (6.0kg peak / 4.1kg nominal per arm)
+- **Emergency stop**: Know the location and operation of the emergency stop device
+- **Regular inspection**: Check for loose screws, damaged mechanical limits, unusual noises, and wiring damage
+
+## Hardware Setup
+
+Follow the official [OpenArm hardware documentation](https://docs.openarm.dev) for:
+
+- Bill of materials and sourcing
+- 3D printing instructions
+- Mechanical assembly
+- Electrical wiring
+
+The hardware repositories are available at [github.com/enactic/openarm](https://github.com/enactic/openarm).
+
+## CAN Bus Setup
+
+OpenArm uses CAN bus communication with Damiao motors. Once you have the CAN bus USB adapter plugged into your Linux PC, follow the [Damiao Motors and CAN Bus guide](./damiao) to configure the interface.
+
+Quick setup:
+
+```bash
+# Setup CAN interfaces
+lerobot-setup-can --mode=setup --interfaces=can0,can1
+
+# Test motor communication
+lerobot-setup-can --mode=test --interfaces=can0,can1
+```
+
+## Install LeRobot 🤗
+
+Follow our [Installation Guide](./installation), then install the Damiao motor support:
+
+```bash
+pip install -e ".[damiao]"
+```
+
+## Usage
+
+### Follower Arm (Robot)
+
+<hfoptions id="follower">
+<hfoption id="Command">
+
+```bash
+lerobot-calibrate \
+    --robot.type=openarm_follower \
+    --robot.port=can0 \
+    --robot.side=right \
+    --robot.id=my_openarm_follower
+```
+
+</hfoption>
+<hfoption id="API example">
+
+```python
+from lerobot.robots.openarm_follower import OpenArmFollower, OpenArmFollowerConfig
+
+config = OpenArmFollowerConfig(
+    port="can0",
+    side="right",  # or "left" for left arm
+    id="my_openarm_follower",
+)
+
+follower = OpenArmFollower(config)
+follower.connect()
+
+# Read current state
+obs = follower.get_observation()
+print(obs)
+
+# Send action (position in degrees)
+action = {
+    "joint_1.pos": 0.0,
+    "joint_2.pos": 0.0,
+    "joint_3.pos": 0.0,
+    "joint_4.pos": 45.0,
+    "joint_5.pos": 0.0,
+    "joint_6.pos": 0.0,
+    "joint_7.pos": 0.0,
+    "gripper.pos": 0.0,
+}
+follower.send_action(action)
+
+follower.disconnect()
+```
+
+</hfoption>
+</hfoptions>
+
+### Leader Arm (Teleoperator)
+
+The leader arm is used for teleoperation - manually moving it to control the follower arm.
+
+<hfoptions id="leader">
+<hfoption id="Command">
+
+```bash
+lerobot-calibrate \
+    --teleop.type=openarm_leader \
+    --teleop.port=can1 \
+    --teleop.id=my_openarm_leader
+```
+
+</hfoption>
+<hfoption id="API example">
+
+```python
+from lerobot.teleoperators.openarm_leader import OpenArmLeader, OpenArmLeaderConfig
+
+config = OpenArmLeaderConfig(
+    port="can1",
+    id="my_openarm_leader",
+    manual_control=True,  # Disable torque for manual movement
+)
+
+leader = OpenArmLeader(config)
+leader.connect()
+
+# Read current position (as action to send to follower)
+action = leader.get_action()
+print(action)
+
+leader.disconnect()
+```
+
+</hfoption>
+</hfoptions>
+
+### Teleoperation
+
+To teleoperate OpenArm with leader-follower control:
+
+```bash
+lerobot-teleoperate \
+    --robot.type=openarm_follower \
+    --robot.port=can0 \
+    --robot.side=right \
+    --robot.id=my_follower \
+    --teleop.type=openarm_leader \
+    --teleop.port=can1 \
+    --teleop.id=my_leader
+```
+
+### Bimanual Teleoperation
+
+To teleoperate a bimanual OpenArm setup with two leader and two follower arms:
+
+```bash
+lerobot-teleoperate \
+    --robot.type=bi_openarm_follower \
+    --robot.left_arm_config.port=can0 \
+    --robot.left_arm_config.side=left \
+    --robot.right_arm_config.port=can1 \
+    --robot.right_arm_config.side=right \
+    --robot.id=my_bimanual_follower \
+    --teleop.type=bi_openarm_leader \
+    --teleop.left_arm_config.port=can2 \
+    --teleop.right_arm_config.port=can3 \
+    --teleop.id=my_bimanual_leader
+```
+
+### Recording Data
+
+To record a dataset during teleoperation:
+
+```bash
+lerobot-record \
+    --robot.type=openarm_follower \
+    --robot.port=can0 \
+    --robot.side=right \
+    --robot.id=my_follower \
+    --teleop.type=openarm_leader \
+    --teleop.port=can1 \
+    --teleop.id=my_leader \
+    --repo-id=my_hf_username/my_openarm_dataset \
+    --fps=30 \
+    --num-episodes=10
+```
+
+## Configuration Options
+
+### Follower Configuration
+
+| Parameter             | Default   | Description                                                |
+| --------------------- | --------- | ---------------------------------------------------------- |
+| `port`                | -         | CAN interface (e.g., `can0`)                               |
+| `side`                | `None`    | Arm side: `"left"`, `"right"`, or `None` for custom limits |
+| `use_can_fd`          | `True`    | Enable CAN FD for higher data rates                        |
+| `can_bitrate`         | `1000000` | Nominal bitrate (1 Mbps)                                   |
+| `can_data_bitrate`    | `5000000` | CAN FD data bitrate (5 Mbps)                               |
+| `max_relative_target` | `None`    | Safety limit for relative target positions                 |
+| `position_kp`         | Per-joint | Position control proportional gains                        |
+| `position_kd`         | Per-joint | Position control derivative gains                          |
+
+### Leader Configuration
+
+| Parameter          | Default   | Description                         |
+| ------------------ | --------- | ----------------------------------- |
+| `port`             | -         | CAN interface (e.g., `can1`)        |
+| `manual_control`   | `True`    | Disable torque for manual movement  |
+| `use_can_fd`       | `True`    | Enable CAN FD for higher data rates |
+| `can_bitrate`      | `1000000` | Nominal bitrate (1 Mbps)            |
+| `can_data_bitrate` | `5000000` | CAN FD data bitrate (5 Mbps)        |
+
+## Motor Configuration
+
+OpenArm uses Damiao motors with the following default configuration:
+
+| Joint                       | Motor Type | Send ID | Recv ID |
+| --------------------------- | ---------- | ------- | ------- |
+| joint_1 (Shoulder pan)      | DM8009     | 0x01    | 0x11    |
+| joint_2 (Shoulder lift)     | DM8009     | 0x02    | 0x12    |
+| joint_3 (Shoulder rotation) | DM4340     | 0x03    | 0x13    |
+| joint_4 (Elbow flex)        | DM4340     | 0x04    | 0x14    |
+| joint_5 (Wrist roll)        | DM4310     | 0x05    | 0x15    |
+| joint_6 (Wrist pitch)       | DM4310     | 0x06    | 0x16    |
+| joint_7 (Wrist rotation)    | DM4310     | 0x07    | 0x17    |
+| gripper                     | DM4310     | 0x08    | 0x18    |
+
+## Troubleshooting
+
+### No Response from Motors
+
+1. Check power supply connections
+2. Verify CAN wiring (CAN-H, CAN-L, GND)
+3. Run diagnostics: `lerobot-setup-can --mode=test --interfaces=can0`
+4. See the [Damiao troubleshooting guide](./damiao#troubleshooting) for more details
+
+### CAN Interface Not Found
+
+Ensure the CAN interface is configured:
+
+```bash
+ip link show can0
+```
+
+## Resources
+
+- [OpenArm Website](https://openarm.dev)
+- [OpenArm Documentation](https://docs.openarm.dev)
+- [OpenArm GitHub](https://github.com/enactic/openarm)
+- [Safety Guide](https://docs.openarm.dev/getting-started/safety-guide)
+- [Damiao Motors and CAN Bus](./damiao)

docs/source/peft_training.mdx

@@ -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
@@ -66,12 +66,13 @@ Run on of the examples scripts to teleoperate, record a dataset, replay a datase
 
 All scripts assume you configured your robot (e.g., SO-100 follower) and set the correct serial port.
 
-Additionally you need to **copy the urdf of the robot to the examples folder**. For the examples in this tutorial (Using SO100/SO101) it is highly recommended to use the urdf in the [SO-ARM100 repo](https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf)
+Additionally you need to **copy the URDF of the robot into the examples folder**. For the examples in this tutorial (using SO100/SO101), copy the `SO101` folder from the [SO-ARM100 repo](https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101) into the `examples/phone_to_so100/` directory, so that the URDF file path becomes `examples/phone_to_so100/SO101/so101_new_calib.urdf`.
 
 - Run this example to teleoperate:
 
   ```bash
-  python examples/phone_to_so100/teleoperate.py
+  cd examples/phone_to_so100
+  python teleoperate.py
   ```
 
 After running the example:
@@ -84,26 +85,29 @@ Additionally you can customize mapping or safety limits by editing the processor
 - Run this example to record a dataset, which saves absolute end effector observations and actions:
 
   ```bash
-  python examples/phone_to_so100/record.py
+  cd examples/phone_to_so100
+  python record.py
   ```
 
 - Run this example to replay recorded episodes:
 
   ```bash
-  python examples/phone_to_so100/replay.py
+  cd examples/phone_to_so100
+  python replay.py
   ```
 
 - Run this example to evaluate a pretrained policy:
 
   ```bash
-  python examples/phone_to_so100/evaluate.py
+  cd examples/phone_to_so100
+  python evaluate.py
   ```
 
 ### Important pipeline steps and options
 
 - Kinematics are used in multiple steps. We use [Placo](https://github.com/Rhoban/placo) which is a wrapper around Pinocchio for handling our kinematics. We construct the kinematics object by passing the robot's URDF and target frame. We set `target_frame_name` to the gripper frame.
 
-  ```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 +118,7 @@ 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 +131,7 @@ 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 +142,7 @@ 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 +151,7 @@ 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 +161,7 @@ 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 +171,7 @@ 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.
@@ -28,11 +34,6 @@ As described by Physical Intelligence, while AI has achieved remarkable success
    pip install -e ".[pi]"
    ```
 
-   > [!NOTE]
-   > For lerobot 0.4.0, if you want to install 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 Data and Capabilities
 
 π₀ is trained on the largest robot interaction dataset to date, combining three key data sources:
@@ -54,7 +55,7 @@ policy.type=pi0
 For training π₀, you can use the standard LeRobot training script with the appropriate configuration:
 
 ```bash
-python src/lerobot/scripts/lerobot_train.py \
+lerobot-train \
     --dataset.repo_id=your_dataset \
     --policy.type=pi0 \
     --output_dir=./outputs/pi0_training \
@@ -64,6 +65,8 @@ 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 +82,15 @@ 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

@@ -36,11 +36,6 @@ This diverse training mixture creates a "curriculum" that enables generalization
    pip install -e ".[pi]"
    ```
 
-   > [!NOTE]
-   > For lerobot 0.4.0, if you want to install 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
-
 ## Usage
 
 To use π₀.₅ in your LeRobot configuration, specify the policy type as:
@@ -56,7 +51,7 @@ policy.type=pi05
 Here's a complete training command for finetuning the base π₀.₅ model on your own dataset:
 
 ```bash
-python src/lerobot/scripts/lerobot_train.py\
+lerobot-train \
     --dataset.repo_id=your_dataset \
     --policy.type=pi05 \
     --output_dir=./outputs/pi05_training \
@@ -67,6 +62,8 @@ 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 +79,15 @@ 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 +1,241 @@
+# π₀-FAST (Pi0-FAST)
+
+π₀-FAST is a **Vision-Language-Action model for general robot control** that uses autoregressive next-token prediction to model continuous robot actions.
+
+## Model Overview
+
+π₀-FAST combines the power of Vision-Language Models with a novel action tokenization approach called **FAST (Frequency-space Action Sequence Tokenization)**. This enables training autoregressive VLAs on highly dexterous tasks that are impossible with standard binning-based discretization, while training **up to 5x faster** than diffusion-based approaches like π₀.
+
+<img
+  src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/lerobot-pifast.png"
+  alt="An overview of Pi0-FAST"
+  width="85%"
+/>
+
+### Why FAST?
+
+Standard approaches for robot action tokenization use simple per-dimension, per-timestep binning schemes. While passable for simple behaviors, this rapidly breaks down for complex and dexterous skills that require precision and high-frequency control.
+
+FAST solves this by compressing action sequences using signal processing techniques, resulting in a dense sequence of action tokens that can be predicted autoregressively—just like language tokens.
+
+### How FAST Tokenization Works
+
+The FAST tokenizer compresses action sequences through the following steps:
+
+1. **Normalize**: Take a continuous action chunk of shape `(H, D)` where `H` is the horizon and `D` is the action dimension. Normalize using one of the supported normalization methods (Quantiles recommended to handle outliers).
+
+2. **Discrete Cosine Transform (DCT)**: Apply DCT (via scipy) to each action dimension separately. DCT is a compression algorithm commonly used in image and audio codecs (JPEG, MP3).
+
+3. **Quantization**: Round and remove insignificant coefficients for each action dimension, producing a sparse frequency matrix.
+
+4. **Flatten**: Flatten the matrix into a 1D vector, with low-frequency components first.
+
+5. **Byte Pair Encoding (BPE)**: Train a BPE tokenizer to compress the DCT coefficients into dense action tokens, typically achieving **10x compression** over prior tokenization approaches.
+
+This approach can transform **any existing VLM** into a VLA by training it to predict these FAST tokens.
+
+## Installation Requirements
+
+1. Install LeRobot by following our [Installation Guide](./installation).
+2. Install π₀-FAST dependencies by running:
+
+   ```bash
+   pip install -e ".[pi]"
+   ```
+
+## Training a Custom FAST Tokenizer
+
+You have two options for the FAST tokenizer:
+
+1. **Use the pre-trained tokenizer**: The `lerobot/fast-action-tokenizer` tokenizer was trained on 1M+ real robot action sequences and works as a general-purpose tokenizer.
+
+2. **Train your own tokenizer**: For maximum performance on your specific dataset, you can finetune the tokenizer on your own data.
+
+### Training Your Own Tokenizer
+
+```bash
+lerobot-train-tokenizer \
+    --repo_id "user/my-lerobot-dataset" \
+    --action_horizon 10 \
+    --encoded_dims "0:6" \
+    --vocab_size 1024 \
+    --scale 10.0 \
+    --normalization_mode QUANTILES \
+    --output_dir "./my_fast_tokenizer" \
+    --push_to_hub \
+    --hub_repo_id "username/my-action-tokenizer"
+```
+
+### Key Tokenizer Parameters
+
+| Parameter              | Description                                                                       | Default      |
+| ---------------------- | --------------------------------------------------------------------------------- | ------------ |
+| `--repo_id`            | LeRobot dataset repository ID                                                     | Required     |
+| `--action_horizon`     | Number of future actions in each chunk                                            | `10`         |
+| `--encoded_dims`       | Comma-separated dimension ranges to encode (e.g., `"0:6,7:23"`)                   | `"0:6,7:23"` |
+| `--vocab_size`         | BPE vocabulary size                                                               | `1024`       |
+| `--scale`              | DCT scaling factor for quantization                                               | `10.0`       |
+| `--normalization_mode` | Normalization mode (`MEAN_STD`, `MIN_MAX`, `QUANTILES`, `QUANTILE10`, `IDENTITY`) | `QUANTILES`  |
+| `--sample_fraction`    | Fraction of chunks to sample per episode                                          | `0.1`        |
+
+## Usage
+
+To use π₀-FAST in LeRobot, specify the policy type as:
+
+```python
+policy.type=pi0_fast
+```
+
+## Training
+
+For training π₀-FAST, you can use the LeRobot training script:
+
+```bash
+lerobot-train \
+    --dataset.repo_id=your_dataset \
+    --policy.type=pi0_fast \
+    --output_dir=./outputs/pi0fast_training \
+    --job_name=pi0fast_training \
+    --policy.pretrained_path=lerobot/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                              | `lerobot/fast-action-tokenizer` |
+| `--policy.compile_model=true`          | Enable torch.compile for faster training           | `false`                         |
+
+## Inference
+
+### KV-Caching for Fast Inference
+
+π₀-FAST supports **KV-caching**, a widely used optimization in LLM inference. This caches the key-value pairs from the attention mechanism, avoiding redundant computation during autoregressive decoding.
+
+```python
+# KV-caching is enabled by default
+policy.use_kv_cache=true
+```
+
+### Inference Example
+
+```python
+from lerobot.policies.pi0_fast import PI0FastPolicy, PI0FastConfig
+
+# Load the policy
+policy = PI0FastPolicy.from_pretrained("your-model-path")
+
+# During inference
+actions = policy.predict_action_chunk(batch)
+```
+
+## Model Architecture
+
+π₀-FAST uses a PaliGemma-based architecture:
+
+- **Vision Encoder**: SigLIP vision tower for image understanding
+- **Language Model**: Gemma 2B for processing language instructions and predicting action tokens
+
+The model takes images, text instructions, and robot state as input, and outputs discrete FAST tokens that are decoded back to continuous actions.
+
+## Configuration Options
+
+| Parameter            | Description                                     | Default    |
+| -------------------- | ----------------------------------------------- | ---------- |
+| `paligemma_variant`  | VLM backbone variant (`gemma_300m`, `gemma_2b`) | `gemma_2b` |
+| `max_state_dim`      | Maximum state vector dimension (padded)         | `32`       |
+| `max_action_dim`     | Maximum action vector dimension (padded)        | `32`       |
+| `temperature`        | Sampling temperature (0.0 for greedy)           | `0.0`      |
+| `max_decoding_steps` | Maximum decoding steps                          | `256`      |
+| `use_kv_cache`       | Enable KV caching for faster inference          | `true`     |
+
+## Comparison with π₀
+
+| Feature               | π₀                        | π₀-FAST                      |
+| --------------------- | ------------------------- | ---------------------------- |
+| Action Representation | Flow Matching (Diffusion) | Autoregressive Tokens (FAST) |
+| Training Speed        | 1x                        | **5x faster**                |
+| Dexterity             | High                      | High                         |
+| Inference Method      | Iterative Denoising       | Autoregressive Decoding      |
+| KV-Caching            | N/A                       | Supported                    |
+
+## Reproducing π₀Fast results
+
+We reproduce the results of π₀Fast on the LIBERO benchmark using the LeRobot implementation. We take the LeRobot PiFast base model [lerobot/pi0fast-base](https://huggingface.co/lerobot/pi0fast-base) and finetune for an additional 40kk steps in bfloat16, with batch size of 256 on 8 H100 GPUs using the [HuggingFace LIBERO dataset](https://huggingface.co/datasets/HuggingFaceVLA/libero).
+
+The finetuned model can be found here:
+
+- **π₀Fast LIBERO**: [lerobot/pi0fast-libero](https://huggingface.co/lerobot/pi0fast-libero)
+
+With the following training command:
+
+```bash
+lerobot-train \
+  --dataset.repo_id=lerobot/libero \
+  --output_dir=outputs/libero_pi0fast \
+  --job_name=libero_pi0fast \
+  --policy.path=lerobot/pi0fast_base \
+  --policy.dtype=bfloat16 \
+  --steps=100000 \
+  --save_freq=20000 \
+  --batch_size=4 \
+  --policy.device=cuda \
+  --policy.scheduler_warmup_steps=4000 \
+  --policy.scheduler_decay_steps=100000 \
+  --policy.scheduler_decay_lr=1e-5 \
+  --policy.gradient_checkpointing=true \
+  --policy.chunk_size=10 \
+  --policy.n_action_steps=10 \
+  --policy.max_action_tokens=256 \
+  --policy.empty_cameras=1 \
+```
+
+We then evaluate the finetuned model using the LeRobot LIBERO implementation, by running the following command:
+
+```bash
+tasks="libero_object,libero_spatial,libero_goal,libero_10"
+lerobot-eval \
+  --policy.path=lerobot/pi0fast-libero \
+  --policy.max_action_tokens=256 \
+  --env.type=libero \
+  --policy.gradient_checkpointing=false \
+  --env.task=${tasks} \
+  --eval.batch_size=1 \
+  --eval.n_episodes=1 \
+  --rename_map='{"observation.images.image":"observation.images.base_0_rgb","observation.images.image2":"observation.images.left_wrist_0_rgb"}'
+```
+
+**Note:** We set `n_action_steps=10`, similar to the original OpenPI implementation.
+
+### Results
+
+We obtain the following results on the LIBERO benchmark:
+
+| Model       | LIBERO Spatial | LIBERO Object | LIBERO Goal | LIBERO 10 | Average  |
+| ----------- | -------------- | ------------- | ----------- | --------- | -------- |
+| **π₀-fast** | 70.0           | 100.0         | 100.0       | 60.0      | **82.5** |
+
+The full evaluation output folder, including videos, is available [here](https://drive.google.com/drive/folders/1HXpwPTRm4hx6g1sF2P7OOqGG0TwPU7LQ?usp=sharing)
+
+## License
+
+This model follows the **Apache 2.0 License**, consistent with the original [OpenPI repository](https://github.com/Physical-Intelligence/openpi).
+
+## References
+
+- [FAST: Efficient Robot Action Tokenization](https://www.physicalintelligence.company/research/fast) - Physical Intelligence Blog
+- [OpenPI Repository](https://github.com/Physical-Intelligence/openpi) - Original implementation
+- [FAST Tokenizer on Hugging Face](https://huggingface.co/physical-intelligence/fast) - Pre-trained tokenizer

docs/source/policy_multi_task_dit_README.md

@@ -0,0 +1,37 @@
+# Multitask DiT Policy
+
+## Citation
+
+If you use this work, please cite the following works:
+
+```bibtex
+@misc{jones2025multitaskditpolicy,
+  author = {Bryson Jones},
+  title = {Dissecting and Open-Sourcing Multitask Diffusion Transformer Policy},
+  year = {2025},
+  url = {https://brysonkjones.substack.com/p/dissecting-and-open-sourcing-multitask-diffusion-transformer-policy},
+  note = {Blog post}
+}
+```
+
+```bibtex
+@misc{trilbmteam2025carefulexaminationlargebehaviormodels,
+  author       = {TRI LBM Team},
+  title        = {A Careful Examination of Large Behavior Models for Multitask Dexterous Manipulation},
+  year         = {2025},
+  eprint       = {arXiv:2507.05331},
+  archivePrefix = {arXiv},
+  primaryClass = {cs.RO},
+  url          = {https://arxiv.org/abs/2507.05331}
+}
+```
+
+```bibtex
+@misc{bostondynamics2025largebehaviormodelsatlas,
+  author       = {Boston Dynamics and TRI Research Team},
+  title        = {Large Behavior Models and Atlas Find New Footing},
+  year         = {2025},
+  url          = {https://bostondynamics.com/blog/large-behavior-models-atlas-find-new-footing/},
+  note         = {Blog post}
+}
+```

docs/source/policy_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 \
@@ -157,6 +159,9 @@ python -m lerobot.record \
     --dataset.fps=15 \
     --dataset.push_to_hub=true \
     --dataset.private=true \
+    --dataset.streaming_encoding=true \
+    --dataset.encoder_threads=2 \
+    # --dataset.vcodec=auto \
     --display_data=true
 ```
 
@@ -165,7 +170,7 @@ 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 +182,8 @@ 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 \
@@ -194,6 +201,9 @@ python -m lerobot.record \
     --dataset.fps=15 \
     --dataset.push_to_hub=true \
     --dataset.private=true \
+    --dataset.streaming_encoding=true \
+    --dataset.encoder_threads=2 \
+    # --dataset.vcodec=auto \
     --display_data=true
 ```
 
@@ -212,9 +222,10 @@ 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 +233,60 @@ 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 +299,9 @@ 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.
@@ -263,7 +269,7 @@ This generates visualizations showing video frames with subtask boundaries overl
 Train with **no annotations** - uses linear progress from 0 to 1:
 
 ```bash
-python src/lerobot/scripts/lerobot_train.py \
+lerobot-train \
   --dataset.repo_id=your-username/your-dataset \
   --policy.type=sarm \
   --policy.annotation_mode=single_stage \
@@ -282,7 +288,7 @@ python src/lerobot/scripts/lerobot_train.py \
 Train with **dense annotations only** (sparse auto-generated):
 
 ```bash
-python src/lerobot/scripts/lerobot_train.py \
+lerobot-train \
   --dataset.repo_id=your-username/your-dataset \
   --policy.type=sarm \
   --policy.annotation_mode=dense_only \
@@ -301,7 +307,7 @@ python src/lerobot/scripts/lerobot_train.py \
 Train with **both sparse and dense annotations**:
 
 ```bash
-python src/lerobot/scripts/lerobot_train.py \
+lerobot-train \
   --dataset.repo_id=your-username/your-dataset \
   --policy.type=sarm \
   --policy.annotation_mode=dual \
@@ -462,7 +468,7 @@ This script:
 Once you have the progress file, train your policy with RA-BC weighting. The progress file is auto-detected from the dataset path (`sarm_progress.parquet`). Currently PI0, PI0.5 and SmolVLA are supported with RA-BC:
 
 ```bash
-python src/lerobot/scripts/lerobot_train.py \
+lerobot-train \
   --dataset.repo_id=your-username/your-dataset \
   --policy.type=pi0 \
   --use_rabc=true \

docs/source/smolvla.mdx

@@ -106,6 +106,9 @@ lerobot-record \
   --dataset.repo_id=${HF_USER}/eval_DATASET_NAME_test \  # <- This will be the dataset name on HF Hub
   --dataset.episode_time_s=50 \
   --dataset.num_episodes=10 \
+  --dataset.streaming_encoding=true \
+  --dataset.encoder_threads=2 \
+  # --dataset.vcodec=auto \
   # <- Teleop optional if you want to teleoperate in between episodes \
   # --teleop.type=so100_leader \
   # --teleop.port=/dev/ttyACM0 \

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

@@ -1,5 +1,18 @@
 # SO-101
 
+<div style="display: flex; align-items: center; gap: 10px;">
+  <img
+    src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/SO101_Follower.webp"
+    alt="SO-101"
+    width="60%"
+  />
+  <img
+    src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/SO101_Leader.webp"
+    alt="SO-101"
+    width="60%"
+  />
+</div>
+
 In the steps below, we explain how to assemble our flagship robot, the SO-101.
 
 ## Source the parts
@@ -125,7 +138,7 @@ 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 +214,7 @@ 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 +377,7 @@ 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 +426,7 @@ 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/streaming_video_encoding.mdx

@@ -0,0 +1,155 @@
+# Streaming Video Encoding Guide
+
+## 1. Overview
+
+Streaming video encoding eliminates the traditional PNG round-trip during video dataset recording. Instead of:
+
+1. Capture frame -> write PNG to disk -> (at episode end) read PNG's -> encode to MP4 -> delete PNG's
+
+Frames can be encoded in real-time during capture:
+
+1. Capture frame -> queue to encoder thread -> encode to MP4 directly
+
+This makes `save_episode()` near-instant (the video is already encoded by the time the episode ends) and removes the blocking wait that previously occurred between episodes, especially with multiple cameras in long episodes.
+
+## 2. Tuning Parameters
+
+| Parameter               | CLI Flag                          | Type          | Default       | Description                                                       |
+| ----------------------- | --------------------------------- | ------------- | ------------- | ----------------------------------------------------------------- |
+| `streaming_encoding`    | `--dataset.streaming_encoding`    | `bool`        | `True`        | Enable real-time encoding during capture                          |
+| `vcodec`                | `--dataset.vcodec`                | `str`         | `"libsvtav1"` | Video codec. `"auto"` detects best HW encoder                     |
+| `encoder_threads`       | `--dataset.encoder_threads`       | `int \| None` | `None` (auto) | Threads per encoder instance. `None` will leave the vcoded decide |
+| `encoder_queue_maxsize` | `--dataset.encoder_queue_maxsize` | `int`         | `60`          | Max buffered frames per camera (~2s at 30fps). Consumes RAM       |
+
+## 3. Performance Considerations
+
+Streaming encoding means the CPU is encoding video **during** the capture loop, not after. This creates a CPU budget that must be shared between:
+
+- **Control loop** (reading cameras, control the robot, writing non-video data)
+- **Encoder threads** (one pool per camera)
+- **Rerun visualization** (if enabled)
+- **OS and other processes**
+
+### Resolution & Number of Cameras Impact
+
+| Setup                     | Throughput (px/sec) | CPU Encoding Load | Notes                          |
+| ------------------------- | ------------------- | ----------------- | ------------------------------ |
+| 2camsx 640x480x3 @30fps   | 55M                 | Low               | Works on most systems          |
+| 2camsx 1280x720x3 @30fps  | 165M                | Moderate          | Comfortable on modern systems  |
+| 2camsx 1920x1080x3 @30fps | 373M                | High              | Requires powerful high-end CPU |
+
+### `encoder_threads` Tuning
+
+This parameter controls how many threads each encoder instance uses internally:
+
+- **Higher values** (e.g., 4-5): Faster encoding, but uses more CPU cores per camera. Good for high-end systems with many cores.
+- **Lower values** (e.g., 1-2): Less CPU per camera, freeing cores for capture and visualization. Good for low-res images and capable CPUs.
+- **`None` (default)**: Lets the codec decide. Information available in the codec logs.
+
+### Backpressure and Frame Dropping
+
+Each camera has a bounded queue (`encoder_queue_maxsize`, default 60 frames). When the encoder can't keep up:
+
+1. The queue fills up (consuming RAM)
+2. New frames are **dropped** (not blocked) — the capture loop continues uninterrupted
+3. A warning is logged: `"Encoder queue full for {camera}, dropped N frame(s)"`
+4. At episode end, total dropped frames per camera are reported
+
+### Symptoms of Encoder Falling Behind
+
+- **System feels laggy and freezes**: all CPUs are at 100%
+- **Dropped frame warnings** in the log or lower frames/FPS than expected in the recorded dataset
+- **Choppy robot movement**: If CPU is severely overloaded, even the capture loop may be affected
+- **Accumulated rerun lag**: Visualization falls behind real-time
+
+## 4. Hardware-Accelerated Encoding
+
+### When to Use
+
+Use HW encoding when:
+
+- CPU is the bottleneck (dropped frames, choppy robot, rerun lag)
+- You have compatible hardware (GPU or dedicated encoder)
+- You're recording at high throughput (high resolution or with many cameras)
+
+### Choosing a Codec
+
+| Codec                 | CPU Usage | File Size      | Quality | Notes                                                            |
+| --------------------- | --------- | -------------- | ------- | ---------------------------------------------------------------- |
+| `libsvtav1` (default) | High      | Smallest       | Best    | Default. Best compression but most CPU-intensive                 |
+| `h264`                | Medium    | ~30-50% larger | Good    | Software H.264. Lower CPU                                        |
+| HW encoders           | Very Low  | Largest        | Good    | Offloads to dedicated hardware. Best for CPU-constrained systems |
+
+### Available HW Encoders
+
+| Encoder             | Platform      | Hardware                                                                                         | CLI Value                            |
+| ------------------- | ------------- | ------------------------------------------------------------------------------------------------ | ------------------------------------ |
+| `h264_videotoolbox` | macOS         | Apple Silicon / Intel                                                                            | `--dataset.vcodec=h264_videotoolbox` |
+| `hevc_videotoolbox` | macOS         | Apple Silicon / Intel                                                                            | `--dataset.vcodec=hevc_videotoolbox` |
+| `h264_nvenc`        | Linux/Windows | NVIDIA GPU                                                                                       | `--dataset.vcodec=h264_nvenc`        |
+| `hevc_nvenc`        | Linux/Windows | NVIDIA GPU                                                                                       | `--dataset.vcodec=hevc_nvenc`        |
+| `h264_vaapi`        | Linux         | Intel/AMD GPU                                                                                    | `--dataset.vcodec=h264_vaapi`        |
+| `h264_qsv`          | Linux/Windows | Intel Quick Sync                                                                                 | `--dataset.vcodec=h264_qsv`          |
+| `auto`              | Any           | Probes the system for available HW encoders. Falls back to `libsvtav1` if no HW encoder is found | `--dataset.vcodec=auto`              |
+
+> [!NOTE]
+> In order to use the HW accelerated encoders you might need to upgrade your GPU drivers.
+
+> [!NOTE]
+> `libsvtav1` is the default because it provides the best training performance; other vcodecs can reduce CPU usage and be faster, but they typically produce larger files and may affect training time.
+
+## 5. Troubleshooting
+
+| Symptom                                                            | Likely Cause                                 | Fix                                                                                                                                                                                                                                                                                  |
+| ------------------------------------------------------------------ | -------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
+| System freezes or choppy robot movement or Rerun visualization lag | CPU starved (100% load usage)                | Close other apps, reduce encoding throughput, lower `encoder_threads`, use `h264`, use `display_data=False`. If the CPU continues to be at 100% then it might be insufficient for your setup, consider `--dataset.streaming_encoding=false` or HW encoding (`--dataset.vcodec=auto`) |
+| "Encoder queue full" warnings or dropped frames in dataset         | Encoder can't keep up (Queue overflow)       | If CPU is not at 100%: Increase `encoder_threads`, increase `encoder_queue_maxsize` or use HW encoding (`--dataset.vcodec=auto`).                                                                                                                                                    |
+| High RAM usage                                                     | Queue filling faster than encoding           | `encoder_threads` too low or CPU insufficient. Reduce `encoder_queue_maxsize` or use HW encoding                                                                                                                                                                                     |
+| Large video files                                                  | Using HW encoder or H.264                    | Expected trade-off. Switch to `libsvtav1` if CPU allows                                                                                                                                                                                                                              |
+| `save_episode()` still slow                                        | `streaming_encoding` is `False`              | Set `--dataset.streaming_encoding=true`                                                                                                                                                                                                                                              |
+| Encoder thread crash                                               | Codec not available or invalid settings      | Check `vcodec` is installed, try `--dataset.vcodec=auto`                                                                                                                                                                                                                             |
+| Recorded dataset is missing frames                                 | CPU/GPU starvation or occasional load spikes | If ~5% of frames are missing, your system is likely overloaded — follow the recommendations above. If fewer frames are missing (~2%), they are probably due to occasional transient load spikes (often at startup) and can be considered expected.                                   |
+
+## 6. Recommended Configurations
+
+These estimates are conservative; we recommend testing them on your setup—start with a low load and increase it gradually.
+
+### High-End Systems: modern 12+ cores (24+ threads)
+
+A throughput between ~250-500M px/sec should be comfortable in CPU. For even better results try HW encoding if available.
+
+```bash
+# 3camsx 1280x720x3 @30fps: Defaults work well. Optionally increase encoder parallelism.
+# 2camsx 1920x1080x3 @30fps: Defaults work well. Optionally increase encoder parallelism.
+lerobot-record --dataset.encoder_threads=5 ...
+
+# 3camsx 1920x1080x3 @30fps: Might require some tuning.
+```
+
+### Mid-Range Systems: modern 8+ cores (16+ threads) or Apple Silicon
+
+A throughput between ~80-300M px/sec should be possible in CPU.
+
+```bash
+# 3camsx 640x480x3 @30fps: Defaults work well. Optionally decrease encoder parallelism.
+# 2camsx 1280x720x3 @30fps: Defaults work well. Optionally decrease encoder parallelism.
+lerobot-record --dataset.encoder_threads=2 ...
+
+# 2camsx 1920x1080x3 @30fps: Might require some tuning.
+```
+
+### Low-Resource Systems: modern 4+ cores (8+ threads) or Raspberry Pi 5
+
+On very constrained systems, streaming encoding may compete too heavily with the capture loop. Disabling it falls back to the PNG-based approach where encoding happens between episodes (blocking, but doesn't interfere with capture). Alternatively, record at a lower throughput to reduce both capture and encoding load. Consider also changing codec to `h264` and using batch encoding.
+
+```bash
+# 2camsx 640x480x3 @30fps: Requires some tuning.
+
+# Use H.264, disable streaming, consider batching encoding
+lerobot-record --dataset.vcodec=h264 --dataset.streaming_encoding=false ...
+```
+
+## 7. Closing note
+
+Performance ultimately depends on your exact setup — frames-per-second, resolution, CPU cores and load, available memory, episode length, and the encoder you choose. Always test with your target workload, be mindful about your CPU & system capabilities and tune `encoder_threads`, `encoder_queue_maxsize`, and
+`vcodec` reasonably. That said, a common practical configuration (for many applications) is three cameras at 640×480x3 @30fps; this usually runs fine with the default streaming video encoding settings in modern systems. Always verify your recorded dataset is healthy by comparing the video duration to the CLI episode duration and confirming the row count equals FPS × CLI duration.

docs/source/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.
 
 ---
 
@@ -124,7 +123,7 @@ SSH into the robot and install LeRobot:
 ```bash
 ssh unitree@<YOUR_ROBOT_IP>
 
-conda create -y -n lerobot python=3.10
+conda create -y -n lerobot python=3.12
 conda activate lerobot
 git clone https://github.com/huggingface/lerobot.git
 cd lerobot
@@ -147,14 +146,14 @@ 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
 
 ```bash
-conda create -y -n lerobot python=3.10
+conda create -y -n lerobot python=3.12
 conda activate lerobot
 git clone https://github.com/huggingface/lerobot.git
 cd lerobot
@@ -172,34 +171,134 @@ 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 test policies before deploying on the physical robot using MuJoCo simulation. Set `is_simulation=True` in config or pass `--robot.is_simulation=true` via CLI.
+
+### Calibrate Exoskeleton Teleoperator
+
+```bash
+lerobot-calibrate \
+    --teleop.type=unitree_g1 \
+    --teleop.left_arm_config.port=/dev/ttyACM1 \
+    --teleop.right_arm_config.port=/dev/ttyACM0 \
+    --teleop.id=exo
+```
+
+### Teleoperate in Simulation
+
+```bash
+lerobot-teleoperate \
+    --robot.type=unitree_g1 \
+    --robot.is_simulation=true \
+    --teleop.type=unitree_g1 \
+    --teleop.left_arm_config.port=/dev/ttyACM1 \
+    --teleop.right_arm_config.port=/dev/ttyACM0 \
+    --teleop.id=exo \
+    --fps=100
+```
+
+### Record Dataset in Simulation
+
+```bash
+lerobot-record \
+    --robot.type=unitree_g1 \
+    --robot.is_simulation=true \
+    --robot.cameras='{"global_view": {"type": "zmq", "server_address": "localhost", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
+    --teleop.type=unitree_g1 \
+    --teleop.left_arm_config.port=/dev/ttyACM1 \
+    --teleop.right_arm_config.port=/dev/ttyACM0 \
+    --teleop.id=exo \
+    --dataset.repo_id=your-username/dataset-name \
+    --dataset.single_task="Test" \
+    --dataset.num_episodes=2 \
+    --dataset.episode_time_s=5 \
+    --dataset.reset_time_s=5 \
+    --dataset.push_to_hub=true \
+    --dataset.streaming_encoding=true \
+    # --dataset.vcodec=auto \
+    --dataset.encoder_threads=2
+```
+
+Example simulation dataset: [nepyope/teleop_test_sim](https://huggingface.co/datasets/nepyope/teleop_test_sim)
+
+---
+
+## Running on Real Robot
+
+Once the robot server is running on the G1 (see Part 3), you can teleoperate and record on the real robot.
+
+### Start the Camera Server
+
+On the robot, start the ZMQ image server:
+
+```bash
+python src/lerobot/cameras/zmq/image_server.py
+```
+
+Keep this running in a separate terminal for camera streaming during recording.
+
+### Teleoperate Real Robot
+
+```bash
+lerobot-teleoperate \
+    --robot.type=unitree_g1 \
+    --robot.is_simulation=false \
+    --teleop.type=unitree_g1 \
+    --teleop.left_arm_config.port=/dev/ttyACM1 \
+    --teleop.right_arm_config.port=/dev/ttyACM0 \
+    --teleop.id=exo \
+    --fps=100
+```
+
+### Record Dataset on Real Robot
+
+```bash
+lerobot-record \
+    --robot.type=unitree_g1 \
+    --robot.is_simulation=false \
+    --robot.cameras='{"global_view": {"type": "zmq", "server_address": "172.18.129.215", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
+    --teleop.type=unitree_g1 \
+    --teleop.left_arm_config.port=/dev/ttyACM1 \
+    --teleop.right_arm_config.port=/dev/ttyACM0 \
+    --teleop.id=exo \
+    --dataset.repo_id=your-username/dataset-name \
+    --dataset.single_task="Test" \
+    --dataset.num_episodes=2 \
+    --dataset.episode_time_s=5 \
+    --dataset.reset_time_s=5 \
+    --dataset.push_to_hub=true \
+    --dataset.streaming_encoding=true \
+    # --dataset.vcodec=auto \
+    --dataset.encoder_threads=2
+```
+
+**Note**: Update `server_address` to match your robot's camera server IP.
+
+Example real robot dataset: [nepyope/teleop_test_real](https://huggingface.co/datasets/nepyope/teleop_test_real)
+
+---
 
 ## 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/using_dataset_tools.mdx

@@ -12,6 +12,7 @@ LeRobot provides several utilities for manipulating datasets:
 4. **Add Features** - Add new features to a dataset
 5. **Remove Features** - Remove features from a dataset
 6. **Convert to Video** - Convert image-based datasets to video format for efficient storage
+7. **Show the Info of Datasets** - Show the summary of datasets information such as number of episode etc.
 
 The core implementation is in `lerobot.datasets.dataset_tools`.
 An example script detailing how to use the tools API is available in `examples/dataset/use_dataset_tools.py`.
@@ -95,26 +96,26 @@ Convert an image-based dataset to video format, creating a new LeRobotDataset wh
 # Local-only: Save to a custom output directory (no hub push)
 lerobot-edit-dataset \
     --repo_id lerobot/pusht_image \
-    --operation.type convert_to_video \
+    --operation.type convert_image_to_video \
     --operation.output_dir /path/to/output/pusht_video
 
 # Save with new repo_id (local storage)
 lerobot-edit-dataset \
     --repo_id lerobot/pusht_image \
     --new_repo_id lerobot/pusht_video \
-    --operation.type convert_to_video
+    --operation.type convert_image_to_video
 
 # Convert and push to Hugging Face Hub
 lerobot-edit-dataset \
     --repo_id lerobot/pusht_image \
     --new_repo_id lerobot/pusht_video \
-    --operation.type convert_to_video \
+    --operation.type convert_image_to_video \
     --push_to_hub true
 
 # Convert with custom video codec and quality settings
 lerobot-edit-dataset \
     --repo_id lerobot/pusht_image \
-    --operation.type convert_to_video \
+    --operation.type convert_image_to_video \
     --operation.output_dir outputs/pusht_video \
     --operation.vcodec libsvtav1 \
     --operation.pix_fmt yuv420p \
@@ -124,16 +125,23 @@ lerobot-edit-dataset \
 # Convert only specific episodes
 lerobot-edit-dataset \
     --repo_id lerobot/pusht_image \
-    --operation.type convert_to_video \
+    --operation.type convert_image_to_video \
     --operation.output_dir outputs/pusht_video \
     --operation.episode_indices "[0, 1, 2, 5, 10]"
 
 # Convert with multiple workers for parallel processing
 lerobot-edit-dataset \
     --repo_id lerobot/pusht_image \
-    --operation.type convert_to_video \
+    --operation.type convert_image_to_video \
     --operation.output_dir outputs/pusht_video \
     --operation.num_workers 8
+
+# For memory-constrained systems, users can now specify limits:
+lerobot-edit-dataset \
+    --repo_id lerobot/pusht_image \
+    --operation.type convert_to_video \
+    --operation.max_episodes_per_batch 50 \
+    --operation.max_frames_per_batch 10000
 ```
 
 **Parameters:**
@@ -149,6 +157,30 @@ lerobot-edit-dataset \
 
 **Note:** The resulting dataset will be a proper LeRobotDataset with all cameras encoded as videos in the `videos/` directory, with parquet files containing only metadata (no raw image data). All episodes, stats, and tasks are preserved.
 
+### Show the information of datasets
+
+Show the information of datasets such as number of episode, number of frame, File size and so on.
+No change will be made to the dataset
+
+```bash
+
+# Show dataset information without feature details
+lerobot-edit-dataset \
+    --repo_id lerobot/pusht_image \
+    --operation.type info \
+
+# Show dataset information with feature details
+lerobot-edit-dataset \
+    --repo_id lerobot/pusht_image \
+    --operation.type info \
+    --operation.show_features true
+
+```
+
+**Parameters:**
+
+- `parameters`: The flag to control show or no show dataset information with feature details.(default=false)
+
 ### Push to Hub
 
 Add the `--push_to_hub true` flag to any command to automatically upload the resulting dataset to the Hugging Face Hub:

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.
@@ -39,7 +45,7 @@ policy.type=wall_x
 For training WallX, you can use the standard LeRobot training script with the appropriate configuration:
 
 ```bash
-python src/lerobot/scripts/lerobot_train.py \
+lerobot-train \
     --dataset.repo_id=your_dataset \
     --policy.type=wall_x \
     --output_dir=./outputs/wallx_training \

docs/source/xvla.mdx

@@ -154,7 +154,7 @@ lerobot-train \
 
 ```bash
 lerobot-train \
-  --dataset.repo_id=pepijn223/bimanual-so100-handover-cube \
+  --dataset.repo_id=<USER>/bimanual-so100-handover-cube \
   --output_dir=./outputs/xvla_bimanual \
   --job_name=xvla_so101_training \
   --policy.path="lerobot/xvla-base" \

examples/backward_compatibility/replay.py

@@ -22,7 +22,7 @@ lerobot-replay \
     --robot.type=so100_follower \
     --robot.port=/dev/tty.usbmodem58760431541 \
     --robot.id=black \
-    --dataset.repo_id=aliberts/record-test \
+    --dataset.repo_id=<USER>/record-test \
     --dataset.episode=2
 ```
 """
@@ -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
@@ -58,7 +57,7 @@ class DatasetReplayConfig:
     repo_id: str
     # Episode to replay.
     episode: int
-    # Root directory where the dataset will be stored (e.g. 'dataset/path').
+    # Root directory where the dataset will be stored (e.g. 'dataset/path'). If None, defaults to $HF_LEROBOT_HOME/repo_id.
     root: str | Path | None = None
     # Limit the frames per second. By default, uses the policy fps.
     fps: int = 30
@@ -82,24 +81,25 @@ def replay(cfg: ReplayConfig):
     actions = dataset.hf_dataset.select_columns(ACTION)
     robot.connect()
 
-    log_say("Replaying episode", cfg.play_sounds, blocking=True)
-    for idx in range(dataset.num_frames):
-        start_episode_t = time.perf_counter()
+    try:
+        log_say("Replaying episode", cfg.play_sounds, blocking=True)
+        for idx in range(dataset.num_frames):
+            start_episode_t = time.perf_counter()
 
-        action_array = actions[idx][ACTION]
-        action = {}
-        for i, name in enumerate(dataset.features[ACTION]["names"]):
-            key = f"{name.removeprefix('main_')}.pos"
-            action[key] = action_array[i].item()
+            action_array = actions[idx][ACTION]
+            action = {}
+            for i, name in enumerate(dataset.features[ACTION]["names"]):
+                key = f"{name.removeprefix('main_')}.pos"
+                action[key] = action_array[i].item()
 
-        action["shoulder_lift.pos"] = -(action["shoulder_lift.pos"] - 90)
-        action["elbow_flex.pos"] -= 90
-        robot.send_action(action)
+            action["shoulder_lift.pos"] = -(action["shoulder_lift.pos"] - 90)
+            action["elbow_flex.pos"] -= 90
+            robot.send_action(action)
 
-        dt_s = time.perf_counter() - start_episode_t
-        precise_sleep(1 / dataset.fps - dt_s)
-
-    robot.disconnect()
+            dt_s = time.perf_counter() - start_episode_t
+            precise_sleep(max(1 / dataset.fps - dt_s, 0.0))
+    finally:
+        robot.disconnect()
 
 
 if __name__ == "__main__":

examples/dataset/slurm_compute_rabc.py

@@ -0,0 +1,490 @@
+#!/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.
+
+"""
+SLURM-distributed SARM RA-BC annotation pipeline.
+
+Computes SARM progress values for all frames in a dataset, distributed across
+SLURM workers, then merges the shards into a single sarm_progress.parquet.
+
+Two subcommands, each a separate SLURM submission:
+
+  compute    – N workers, each computes progress for a subset of episodes
+  aggregate  – 1 worker, merges N shards into sarm_progress.parquet, pushes to hub
+
+Usage:
+    python slurm_compute_rabc.py compute \\
+        --repo-id user/dataset --reward-model-path user/sarm_model \\
+        --stride 10 --device cpu --workers 50 --partition cpu
+
+    python slurm_compute_rabc.py aggregate \\
+        --repo-id user/dataset --reward-model-path user/sarm_model \\
+        --partition cpu --push-to-hub
+"""
+
+import argparse
+from pathlib import Path
+
+from datatrove.executor import LocalPipelineExecutor
+from datatrove.executor.slurm import SlurmPipelineExecutor
+from datatrove.pipeline.base import PipelineStep
+
+
+class ComputeProgressShards(PipelineStep):
+    """Each worker computes SARM progress for its assigned episodes."""
+
+    def __init__(
+        self, repo_id, reward_model_path, stride=1, head_mode="sparse", device="cpu", shard_dir="rabc_shards"
+    ):
+        super().__init__()
+        if stride < 1:
+            raise ValueError(f"stride must be >= 1, got {stride}")
+        self.repo_id = repo_id
+        self.reward_model_path = reward_model_path
+        self.stride = stride
+        self.head_mode = head_mode
+        self.device = device
+        self.shard_dir = shard_dir
+
+    def run(self, data=None, rank: int = 0, world_size: int = 1):
+        import logging
+        from pathlib import Path
+
+        import numpy as np
+        import pyarrow as pa
+        import pyarrow.parquet as pq
+        import torch
+        from tqdm import tqdm
+
+        from lerobot.policies.sarm.compute_rabc_weights import (
+            generate_all_frame_indices,
+            interpolate_progress,
+            load_sarm_resources,
+        )
+        from lerobot.utils.utils import init_logging
+
+        init_logging()
+
+        dataset, reward_model, preprocess = load_sarm_resources(
+            self.repo_id,
+            self.reward_model_path,
+            self.device,
+        )
+
+        if hasattr(preprocess, "eval"):
+            preprocess.eval()
+        for step in preprocess.steps:
+            if hasattr(step, "eval"):
+                step.eval()
+
+        image_key = reward_model.config.image_key
+        state_key = reward_model.config.state_key
+        frame_gap = reward_model.config.frame_gap
+        center_idx = reward_model.config.n_obs_steps // 2
+
+        dual_mode = reward_model.config.uses_dual_heads
+        compute_sparse = self.head_mode in ("sparse", "both") or not dual_mode
+        compute_dense = self.head_mode in ("dense", "both") and dual_mode
+
+        my_episodes = list(range(dataset.num_episodes))[rank::world_size]
+        if not my_episodes:
+            logging.info(f"Rank {rank}: no episodes assigned")
+            return
+        logging.info(f"Rank {rank}: {len(my_episodes)} / {dataset.num_episodes} episodes")
+
+        all_rows = []
+
+        for ep_idx in tqdm(my_episodes, desc=f"Rank {rank}"):
+            ep = dataset.meta.episodes[ep_idx]
+            ep_start, ep_end = ep["dataset_from_index"], ep["dataset_to_index"]
+            task = dataset[ep_start].get("task", "perform the task")
+
+            all_ep_indices = generate_all_frame_indices(ep_start, ep_end, frame_gap)
+            if self.stride > 1:
+                compute_indices = [i for i in all_ep_indices if (i - ep_start) % self.stride == 0]
+                if (ep_end - 1) not in compute_indices:
+                    compute_indices.append(ep_end - 1)
+                compute_indices = sorted(set(compute_indices))
+            else:
+                compute_indices = all_ep_indices
+
+            frame_results = {}
+            for qi in tqdm(compute_indices, desc=f"  Ep {ep_idx}", leave=False):
+                try:
+                    sample = dataset[qi]
+                    batch = {
+                        image_key: sample[image_key],
+                        "task": task,
+                        "index": qi,
+                        "episode_index": ep_idx,
+                    }
+                    if state_key in sample:
+                        batch[state_key] = sample[state_key]
+
+                    with torch.no_grad():
+                        processed = preprocess(batch)
+                        vf = processed["video_features"].to(self.device)
+                        tf = processed["text_features"].to(self.device)
+                        sf = processed.get("state_features")
+                        if sf is not None:
+                            sf = sf.to(self.device)
+                        lengths = processed.get("lengths")
+
+                        sparse_val = dense_val = np.nan
+                        if compute_sparse:
+                            r = reward_model.calculate_rewards(
+                                text_embeddings=tf,
+                                video_embeddings=vf,
+                                state_features=sf,
+                                lengths=lengths,
+                                return_all_frames=True,
+                                head_mode="sparse",
+                            )
+                            sparse_val = float(r[0, center_idx] if r.ndim == 2 else r[center_idx])
+                        if compute_dense:
+                            r = reward_model.calculate_rewards(
+                                text_embeddings=tf,
+                                video_embeddings=vf,
+                                state_features=sf,
+                                lengths=lengths,
+                                return_all_frames=True,
+                                head_mode="dense",
+                            )
+                            dense_val = float(r[0, center_idx] if r.ndim == 2 else r[center_idx])
+
+                        frame_results[qi] = (sparse_val, dense_val)
+                except Exception as e:
+                    logging.warning(f"Failed frame {qi}: {e}")
+
+            if not frame_results:
+                logging.warning(f"Episode {ep_idx}: all frames failed, skipping")
+                continue
+
+            # Interpolate to all frames in this episode
+            computed_idx = np.array(sorted(frame_results.keys()))
+            all_frame_arr = np.arange(ep_start, ep_end)
+
+            sparse_vals = np.array([frame_results[i][0] for i in computed_idx]) if compute_sparse else None
+            dense_vals = np.array([frame_results[i][1] for i in computed_idx]) if compute_dense else None
+
+            if self.stride > 1 and len(computed_idx) > 1:
+                if compute_sparse:
+                    sparse_vals = interpolate_progress(computed_idx, sparse_vals, all_frame_arr)
+                if compute_dense:
+                    dense_vals = interpolate_progress(computed_idx, dense_vals, all_frame_arr)
+                output_frames = all_frame_arr
+            else:
+                # Use only successfully computed frames to avoid indexing mismatch on failures
+                output_frames = computed_idx
+
+            for i, fi in enumerate(output_frames):
+                row = {"index": int(fi), "episode_index": ep_idx, "frame_index": int(fi - ep_start)}
+                if compute_sparse:
+                    row["progress_sparse"] = float(sparse_vals[i])
+                if compute_dense:
+                    row["progress_dense"] = float(dense_vals[i])
+                all_rows.append(row)
+
+        if all_rows:
+            import pandas as pd
+
+            df = pd.DataFrame(all_rows).sort_values("index").reset_index(drop=True)
+            table = pa.Table.from_pandas(df, preserve_index=False)
+            table = table.replace_schema_metadata({b"reward_model_path": self.reward_model_path.encode()})
+            shard_dir = Path(self.shard_dir)
+            shard_dir.mkdir(parents=True, exist_ok=True)
+            out = shard_dir / f"shard_{rank:05d}.parquet"
+            pq.write_table(table, out)
+            logging.info(f"Rank {rank}: saved {len(df)} rows to {out}")
+
+
+class AggregateProgress(PipelineStep):
+    """Merge all shard parquets into final sarm_progress.parquet."""
+
+    def __init__(self, repo_id, reward_model_path, shard_dir="rabc_shards", push_to_hub=False):
+        super().__init__()
+        self.repo_id = repo_id
+        self.reward_model_path = reward_model_path
+        self.shard_dir = shard_dir
+        self.push_to_hub = push_to_hub
+
+    def run(self, data=None, rank: int = 0, world_size: int = 1):
+        import datetime
+        import logging
+        import os
+        from pathlib import Path
+
+        import pandas as pd
+        import pyarrow as pa
+        import pyarrow.parquet as pq
+
+        from lerobot.datasets.lerobot_dataset import LeRobotDataset
+        from lerobot.utils.utils import init_logging
+
+        init_logging()
+        if rank != 0:
+            return
+
+        shard_dir = Path(self.shard_dir)
+        shards = sorted(shard_dir.glob("shard_*.parquet"))
+        if not shards:
+            raise FileNotFoundError(f"No shards found in {shard_dir}")
+
+        # Log shard modification time range to help detect stale files
+        mtimes = [os.path.getmtime(s) for s in shards]
+        oldest = datetime.datetime.fromtimestamp(min(mtimes)).isoformat(timespec="seconds")
+        newest = datetime.datetime.fromtimestamp(max(mtimes)).isoformat(timespec="seconds")
+        logging.info(f"Aggregating {len(shards)} shards (oldest: {oldest}, newest: {newest})")
+
+        df = pd.concat([pd.read_parquet(s) for s in shards], ignore_index=True)
+        df = df.sort_values("index").reset_index(drop=True)
+
+        table = pa.Table.from_pandas(df, preserve_index=False)
+        table = table.replace_schema_metadata({b"reward_model_path": self.reward_model_path.encode()})
+
+        temp_ds = LeRobotDataset(self.repo_id, download_videos=False)
+        out_path = Path(temp_ds.root) / "sarm_progress.parquet"
+        out_path.parent.mkdir(parents=True, exist_ok=True)
+        pq.write_table(table, out_path)
+        logging.info(f"Saved {len(df)} rows to {out_path}")
+
+        for col in ["progress_sparse", "progress_dense"]:
+            if col in df.columns:
+                v = df[col].dropna()
+                logging.info(
+                    f"{col}: mean={v.mean():.4f} std={v.std():.4f} min={v.min():.4f} max={v.max():.4f}"
+                )
+
+        if self.push_to_hub:
+            from huggingface_hub import HfApi
+
+            api = HfApi()
+            hub_path = "sarm_progress.parquet"
+            logging.info(f"Uploading to {self.repo_id}/{hub_path}")
+            api.upload_file(
+                path_or_fileobj=str(out_path),
+                path_in_repo=hub_path,
+                repo_id=self.repo_id,
+                repo_type="dataset",
+            )
+            logging.info(f"Uploaded: https://huggingface.co/datasets/{self.repo_id}/blob/main/{hub_path}")
+
+
+def make_compute_executor(
+    repo_id,
+    reward_model_path,
+    stride,
+    head_mode,
+    device,
+    shard_dir,
+    logs_dir,
+    job_name,
+    slurm,
+    workers,
+    partition,
+    cpus_per_task,
+    mem_per_cpu,
+):
+    kwargs = {
+        "pipeline": [
+            ComputeProgressShards(repo_id, reward_model_path, stride, head_mode, device, str(shard_dir)),
+        ],
+        "logging_dir": str(logs_dir / job_name),
+    }
+
+    if slurm:
+        kwargs.update(
+            {
+                "job_name": job_name,
+                "tasks": workers,
+                "workers": workers,
+                "time": "24:00:00",
+                "partition": partition,
+                "cpus_per_task": cpus_per_task,
+                "sbatch_args": {"mem-per-cpu": mem_per_cpu},
+            }
+        )
+        return SlurmPipelineExecutor(**kwargs)
+
+    kwargs.update({"tasks": workers, "workers": 1})
+    return LocalPipelineExecutor(**kwargs)
+
+
+def make_aggregate_executor(
+    repo_id,
+    reward_model_path,
+    shard_dir,
+    logs_dir,
+    job_name,
+    slurm,
+    partition,
+    cpus_per_task,
+    mem_per_cpu,
+    push_to_hub,
+):
+    kwargs = {
+        "pipeline": [
+            AggregateProgress(repo_id, reward_model_path, str(shard_dir), push_to_hub),
+        ],
+        "logging_dir": str(logs_dir / job_name),
+    }
+
+    if slurm:
+        kwargs.update(
+            {
+                "job_name": job_name,
+                "tasks": 1,
+                "workers": 1,
+                "time": "02:00:00",
+                "partition": partition,
+                "cpus_per_task": cpus_per_task,
+                "sbatch_args": {"mem-per-cpu": mem_per_cpu},
+            }
+        )
+        return SlurmPipelineExecutor(**kwargs)
+
+    kwargs.update({"tasks": 1, "workers": 1})
+    return LocalPipelineExecutor(**kwargs)
+
+
+def _add_shared_args(p):
+    p.add_argument(
+        "--repo-id",
+        type=str,
+        required=True,
+        help="Hugging Face repository identifier, e.g. 'user/dataset'.",
+    )
+    p.add_argument(
+        "--shard-dir",
+        type=Path,
+        default=Path("rabc_shards"),
+        help="Directory to read/write per-rank parquet shards.",
+    )
+    p.add_argument(
+        "--logs-dir",
+        type=Path,
+        default=Path("logs"),
+        help="Directory for datatrove logs.",
+    )
+    p.add_argument(
+        "--job-name",
+        type=str,
+        default=None,
+        help="SLURM job name (defaults to rabc_<subcommand>).",
+    )
+    p.add_argument(
+        "--slurm",
+        type=int,
+        default=1,
+        help="1 = submit via SLURM; 0 = run locally (useful for debugging).",
+    )
+    p.add_argument(
+        "--partition",
+        type=str,
+        default=None,
+        help="SLURM partition to submit to.",
+    )
+    p.add_argument(
+        "--cpus-per-task",
+        type=int,
+        default=4,
+        help="Number of CPUs per SLURM task.",
+    )
+    p.add_argument(
+        "--mem-per-cpu",
+        type=str,
+        default="4G",
+        help="Memory per CPU, e.g. '4G' or '1950M'.",
+    )
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="SLURM-distributed SARM RA-BC annotation pipeline",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+    )
+    sub = parser.add_subparsers(dest="command", required=True)
+
+    # compute subcommand
+    cp = sub.add_parser(
+        "compute",
+        help="Distribute progress computation across SLURM workers.",
+    )
+    _add_shared_args(cp)
+    cp.add_argument(
+        "--reward-model-path",
+        type=str,
+        required=True,
+        help="Path or HF repo id of the SARM reward model.",
+    )
+    cp.add_argument(
+        "--stride",
+        type=int,
+        default=1,
+        help="Compute every Nth frame; intermediate frames are interpolated (must be >= 1).",
+    )
+    cp.add_argument(
+        "--head-mode",
+        type=str,
+        default="sparse",
+        choices=["sparse", "dense", "both"],
+        help="Which reward head(s) to compute.",
+    )
+    cp.add_argument(
+        "--device",
+        type=str,
+        default="cpu",
+        help="Device for reward model inference, e.g. 'cpu' or 'cuda'.",
+    )
+    cp.add_argument(
+        "--workers",
+        type=int,
+        default=50,
+        help="Number of parallel SLURM tasks (one shard per worker).",
+    )
+
+    # aggregate subcommand
+    ap = sub.add_parser(
+        "aggregate",
+        help="Merge per-rank shards into a single sarm_progress.parquet.",
+    )
+    _add_shared_args(ap)
+    ap.add_argument(
+        "--reward-model-path",
+        type=str,
+        required=True,
+        help="Path or HF repo id of the SARM reward model (stored in parquet metadata).",
+    )
+    ap.add_argument(
+        "--push-to-hub",
+        action="store_true",
+        help="Upload sarm_progress.parquet to the Hugging Face Hub after aggregation.",
+    )
+
+    args = parser.parse_args()
+    job_name = args.job_name or f"rabc_{args.command}"
+    kwargs = vars(args)
+    kwargs["slurm"] = kwargs.pop("slurm") == 1
+    kwargs["job_name"] = job_name
+    command = kwargs.pop("command")
+
+    executor = make_compute_executor(**kwargs) if command == "compute" else make_aggregate_executor(**kwargs)
+
+    executor.run()
+
+
+if __name__ == "__main__":
+    main()

examples/lekiwi/evaluate.py

@@ -78,40 +78,24 @@ def main():
     listener, events = init_keyboard_listener()
     init_rerun(session_name="lekiwi_evaluate")
 
-    if not robot.is_connected:
-        raise ValueError("Robot is not connected!")
+    try:
+        if not robot.is_connected:
+            raise ValueError("Robot is not connected!")
 
-    print("Starting evaluate loop...")
-    recorded_episodes = 0
-    while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
-        log_say(f"Running inference, recording eval episode {recorded_episodes} of {NUM_EPISODES}")
+        print("Starting evaluate loop...")
+        recorded_episodes = 0
+        while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
+            log_say(f"Running inference, recording eval episode {recorded_episodes} of {NUM_EPISODES}")
 
-        # Main record loop
-        record_loop(
-            robot=robot,
-            events=events,
-            fps=FPS,
-            policy=policy,
-            preprocessor=preprocessor,  # Pass the pre and post policy processors
-            postprocessor=postprocessor,
-            dataset=dataset,
-            control_time_s=EPISODE_TIME_SEC,
-            single_task=TASK_DESCRIPTION,
-            display_data=True,
-            teleop_action_processor=teleop_action_processor,
-            robot_action_processor=robot_action_processor,
-            robot_observation_processor=robot_observation_processor,
-        )
-
-        # Reset the environment if not stopping or re-recording
-        if not events["stop_recording"] and (
-            (recorded_episodes < NUM_EPISODES - 1) or events["rerecord_episode"]
-        ):
-            log_say("Reset the environment")
+            # Main record loop
             record_loop(
                 robot=robot,
                 events=events,
                 fps=FPS,
+                policy=policy,
+                preprocessor=preprocessor,  # Pass the pre and post policy processors
+                postprocessor=postprocessor,
+                dataset=dataset,
                 control_time_s=EPISODE_TIME_SEC,
                 single_task=TASK_DESCRIPTION,
                 display_data=True,
@@ -120,24 +104,42 @@ def main():
                 robot_observation_processor=robot_observation_processor,
             )
 
-        if events["rerecord_episode"]:
-            log_say("Re-record episode")
-            events["rerecord_episode"] = False
-            events["exit_early"] = False
-            dataset.clear_episode_buffer()
-            continue
+            # Reset the environment if not stopping or re-recording
+            if not events["stop_recording"] and (
+                (recorded_episodes < NUM_EPISODES - 1) or events["rerecord_episode"]
+            ):
+                log_say("Reset the environment")
+                record_loop(
+                    robot=robot,
+                    events=events,
+                    fps=FPS,
+                    control_time_s=EPISODE_TIME_SEC,
+                    single_task=TASK_DESCRIPTION,
+                    display_data=True,
+                    teleop_action_processor=teleop_action_processor,
+                    robot_action_processor=robot_action_processor,
+                    robot_observation_processor=robot_observation_processor,
+                )
 
-        # Save episode
-        dataset.save_episode()
-        recorded_episodes += 1
+            if events["rerecord_episode"]:
+                log_say("Re-record episode")
+                events["rerecord_episode"] = False
+                events["exit_early"] = False
+                dataset.clear_episode_buffer()
+                continue
 
-    # Clean up
-    log_say("Stop recording")
-    robot.disconnect()
-    listener.stop()
+            # Save episode
+            dataset.save_episode()
+            recorded_episodes += 1
 
-    dataset.finalize()
-    dataset.push_to_hub()
+    finally:
+        # Clean up
+        log_say("Stop recording")
+        robot.disconnect()
+        listener.stop()
+
+        dataset.finalize()
+        dataset.push_to_hub()
 
 
 if __name__ == "__main__":

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
@@ -74,40 +74,23 @@ def main():
     listener, events = init_keyboard_listener()
     init_rerun(session_name="lekiwi_record")
 
-    if not robot.is_connected or not leader_arm.is_connected or not keyboard.is_connected:
-        raise ValueError("Robot or teleop is not connected!")
+    try:
+        if not robot.is_connected or not leader_arm.is_connected or not keyboard.is_connected:
+            raise ValueError("Robot or teleop is not connected!")
 
-    print("Starting record loop...")
-    recorded_episodes = 0
-    while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
-        log_say(f"Recording episode {recorded_episodes}")
+        print("Starting record loop...")
+        recorded_episodes = 0
+        while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
+            log_say(f"Recording episode {recorded_episodes}")
 
-        # Main record loop
-        record_loop(
-            robot=robot,
-            events=events,
-            fps=FPS,
-            dataset=dataset,
-            teleop=[leader_arm, keyboard],
-            control_time_s=EPISODE_TIME_SEC,
-            single_task=TASK_DESCRIPTION,
-            display_data=True,
-            teleop_action_processor=teleop_action_processor,
-            robot_action_processor=robot_action_processor,
-            robot_observation_processor=robot_observation_processor,
-        )
-
-        # Reset the environment if not stopping or re-recording
-        if not events["stop_recording"] and (
-            (recorded_episodes < NUM_EPISODES - 1) or events["rerecord_episode"]
-        ):
-            log_say("Reset the environment")
+            # Main record loop
             record_loop(
                 robot=robot,
                 events=events,
                 fps=FPS,
+                dataset=dataset,
                 teleop=[leader_arm, keyboard],
-                control_time_s=RESET_TIME_SEC,
+                control_time_s=EPISODE_TIME_SEC,
                 single_task=TASK_DESCRIPTION,
                 display_data=True,
                 teleop_action_processor=teleop_action_processor,
@@ -115,26 +98,44 @@ def main():
                 robot_observation_processor=robot_observation_processor,
             )
 
-        if events["rerecord_episode"]:
-            log_say("Re-record episode")
-            events["rerecord_episode"] = False
-            events["exit_early"] = False
-            dataset.clear_episode_buffer()
-            continue
+            # Reset the environment if not stopping or re-recording
+            if not events["stop_recording"] and (
+                (recorded_episodes < NUM_EPISODES - 1) or events["rerecord_episode"]
+            ):
+                log_say("Reset the environment")
+                record_loop(
+                    robot=robot,
+                    events=events,
+                    fps=FPS,
+                    teleop=[leader_arm, keyboard],
+                    control_time_s=RESET_TIME_SEC,
+                    single_task=TASK_DESCRIPTION,
+                    display_data=True,
+                    teleop_action_processor=teleop_action_processor,
+                    robot_action_processor=robot_action_processor,
+                    robot_observation_processor=robot_observation_processor,
+                )
 
-        # Save episode
-        dataset.save_episode()
-        recorded_episodes += 1
+            if events["rerecord_episode"]:
+                log_say("Re-record episode")
+                events["rerecord_episode"] = False
+                events["exit_early"] = False
+                dataset.clear_episode_buffer()
+                continue
 
-    # Clean up
-    log_say("Stop recording")
-    robot.disconnect()
-    leader_arm.disconnect()
-    keyboard.disconnect()
-    listener.stop()
+            # Save episode
+            dataset.save_episode()
+            recorded_episodes += 1
+    finally:
+        # Clean up
+        log_say("Stop recording")
+        robot.disconnect()
+        leader_arm.disconnect()
+        keyboard.disconnect()
+        listener.stop()
 
-    dataset.finalize()
-    dataset.push_to_hub()
+        dataset.finalize()
+        dataset.push_to_hub()
 
 
 if __name__ == "__main__":

examples/lekiwi/replay.py

@@ -42,25 +42,27 @@ def main():
     # Connect to the robot
     robot.connect()
 
-    if not robot.is_connected:
-        raise ValueError("Robot is not connected!")
+    try:
+        if not robot.is_connected:
+            raise ValueError("Robot is not connected!")
 
-    print("Starting replay loop...")
-    log_say(f"Replaying episode {EPISODE_IDX}")
-    for idx in range(len(episode_frames)):
-        t0 = time.perf_counter()
+        print("Starting replay loop...")
+        log_say(f"Replaying episode {EPISODE_IDX}")
+        for idx in range(len(episode_frames)):
+            t0 = time.perf_counter()
 
-        # Get recorded action from dataset
-        action = {
-            name: float(actions[idx][ACTION][i]) for i, name in enumerate(dataset.features[ACTION]["names"])
-        }
+            # Get recorded action from dataset
+            action = {
+                name: float(actions[idx][ACTION][i])
+                for i, name in enumerate(dataset.features[ACTION]["names"])
+            }
 
-        # Send action to robot
-        _ = robot.send_action(action)
+            # Send action to robot
+            _ = robot.send_action(action)
 
-        precise_sleep(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
-
-    robot.disconnect()
+            precise_sleep(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
+    finally:
+        robot.disconnect()
 
 
 if __name__ == "__main__":

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
@@ -143,38 +142,24 @@ def main():
     listener, events = init_keyboard_listener()
     init_rerun(session_name="phone_so100_evaluate")
 
-    if not robot.is_connected:
-        raise ValueError("Robot is not connected!")
+    try:
+        if not robot.is_connected:
+            raise ValueError("Robot is not connected!")
 
-    print("Starting evaluate loop...")
-    episode_idx = 0
-    for episode_idx in range(NUM_EPISODES):
-        log_say(f"Running inference, recording eval episode {episode_idx + 1} of {NUM_EPISODES}")
+        print("Starting evaluate loop...")
+        episode_idx = 0
+        for episode_idx in range(NUM_EPISODES):
+            log_say(f"Running inference, recording eval episode {episode_idx + 1} of {NUM_EPISODES}")
 
-        # Main record loop
-        record_loop(
-            robot=robot,
-            events=events,
-            fps=FPS,
-            policy=policy,
-            preprocessor=preprocessor,  # Pass the pre and post policy processors
-            postprocessor=postprocessor,
-            dataset=dataset,
-            control_time_s=EPISODE_TIME_SEC,
-            single_task=TASK_DESCRIPTION,
-            display_data=True,
-            teleop_action_processor=make_default_teleop_action_processor(),
-            robot_action_processor=robot_ee_to_joints_processor,
-            robot_observation_processor=robot_joints_to_ee_pose_processor,
-        )
-
-        # Reset the environment if not stopping or re-recording
-        if not events["stop_recording"] and ((episode_idx < NUM_EPISODES - 1) or events["rerecord_episode"]):
-            log_say("Reset the environment")
+            # Main record loop
             record_loop(
                 robot=robot,
                 events=events,
                 fps=FPS,
+                policy=policy,
+                preprocessor=preprocessor,  # Pass the pre and post policy processors
+                postprocessor=postprocessor,
+                dataset=dataset,
                 control_time_s=EPISODE_TIME_SEC,
                 single_task=TASK_DESCRIPTION,
                 display_data=True,
@@ -183,24 +168,41 @@ def main():
                 robot_observation_processor=robot_joints_to_ee_pose_processor,
             )
 
-        if events["rerecord_episode"]:
-            log_say("Re-record episode")
-            events["rerecord_episode"] = False
-            events["exit_early"] = False
-            dataset.clear_episode_buffer()
-            continue
+            # Reset the environment if not stopping or re-recording
+            if not events["stop_recording"] and (
+                (episode_idx < NUM_EPISODES - 1) or events["rerecord_episode"]
+            ):
+                log_say("Reset the environment")
+                record_loop(
+                    robot=robot,
+                    events=events,
+                    fps=FPS,
+                    control_time_s=EPISODE_TIME_SEC,
+                    single_task=TASK_DESCRIPTION,
+                    display_data=True,
+                    teleop_action_processor=make_default_teleop_action_processor(),
+                    robot_action_processor=robot_ee_to_joints_processor,
+                    robot_observation_processor=robot_joints_to_ee_pose_processor,
+                )
 
-        # Save episode
-        dataset.save_episode()
-        episode_idx += 1
+            if events["rerecord_episode"]:
+                log_say("Re-record episode")
+                events["rerecord_episode"] = False
+                events["exit_early"] = False
+                dataset.clear_episode_buffer()
+                continue
 
-    # Clean up
-    log_say("Stop recording")
-    robot.disconnect()
-    listener.stop()
+            # Save episode
+            dataset.save_episode()
+            episode_idx += 1
+    finally:
+        # Clean up
+        log_say("Stop recording")
+        robot.disconnect()
+        listener.stop()
 
-    dataset.finalize()
-    dataset.push_to_hub()
+        dataset.finalize()
+        dataset.push_to_hub()
 
 
 if __name__ == "__main__":

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
@@ -150,38 +149,23 @@ def main():
     listener, events = init_keyboard_listener()
     init_rerun(session_name="phone_so100_record")
 
-    if not robot.is_connected or not phone.is_connected:
-        raise ValueError("Robot or teleop is not connected!")
+    try:
+        if not robot.is_connected or not phone.is_connected:
+            raise ValueError("Robot or teleop is not connected!")
 
-    print("Starting record loop. Move your phone to teleoperate the robot...")
-    episode_idx = 0
-    while episode_idx < NUM_EPISODES and not events["stop_recording"]:
-        log_say(f"Recording episode {episode_idx + 1} of {NUM_EPISODES}")
+        print("Starting record loop. Move your phone to teleoperate the robot...")
+        episode_idx = 0
+        while episode_idx < NUM_EPISODES and not events["stop_recording"]:
+            log_say(f"Recording episode {episode_idx + 1} of {NUM_EPISODES}")
 
-        # Main record loop
-        record_loop(
-            robot=robot,
-            events=events,
-            fps=FPS,
-            teleop=phone,
-            dataset=dataset,
-            control_time_s=EPISODE_TIME_SEC,
-            single_task=TASK_DESCRIPTION,
-            display_data=True,
-            teleop_action_processor=phone_to_robot_ee_pose_processor,
-            robot_action_processor=robot_ee_to_joints_processor,
-            robot_observation_processor=robot_joints_to_ee_pose,
-        )
-
-        # Reset the environment if not stopping or re-recording
-        if not events["stop_recording"] and (episode_idx < NUM_EPISODES - 1 or events["rerecord_episode"]):
-            log_say("Reset the environment")
+            # Main record loop
             record_loop(
                 robot=robot,
                 events=events,
                 fps=FPS,
                 teleop=phone,
-                control_time_s=RESET_TIME_SEC,
+                dataset=dataset,
+                control_time_s=EPISODE_TIME_SEC,
                 single_task=TASK_DESCRIPTION,
                 display_data=True,
                 teleop_action_processor=phone_to_robot_ee_pose_processor,
@@ -189,25 +173,43 @@ def main():
                 robot_observation_processor=robot_joints_to_ee_pose,
             )
 
-        if events["rerecord_episode"]:
-            log_say("Re-recording episode")
-            events["rerecord_episode"] = False
-            events["exit_early"] = False
-            dataset.clear_episode_buffer()
-            continue
+            # Reset the environment if not stopping or re-recording
+            if not events["stop_recording"] and (
+                episode_idx < NUM_EPISODES - 1 or events["rerecord_episode"]
+            ):
+                log_say("Reset the environment")
+                record_loop(
+                    robot=robot,
+                    events=events,
+                    fps=FPS,
+                    teleop=phone,
+                    control_time_s=RESET_TIME_SEC,
+                    single_task=TASK_DESCRIPTION,
+                    display_data=True,
+                    teleop_action_processor=phone_to_robot_ee_pose_processor,
+                    robot_action_processor=robot_ee_to_joints_processor,
+                    robot_observation_processor=robot_joints_to_ee_pose,
+                )
 
-        # Save episode
-        dataset.save_episode()
-        episode_idx += 1
+            if events["rerecord_episode"]:
+                log_say("Re-recording episode")
+                events["rerecord_episode"] = False
+                events["exit_early"] = False
+                dataset.clear_episode_buffer()
+                continue
 
-    # Clean up
-    log_say("Stop recording")
-    robot.disconnect()
-    phone.disconnect()
-    listener.stop()
+            # Save episode
+            dataset.save_episode()
+            episode_idx += 1
+    finally:
+        # Clean up
+        log_say("Stop recording")
+        robot.disconnect()
+        phone.disconnect()
+        listener.stop()
 
-    dataset.finalize()
-    dataset.push_to_hub()
+        dataset.finalize()
+        dataset.push_to_hub()
 
 
 if __name__ == "__main__":

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
@@ -74,32 +73,34 @@ def main():
     # Connect to the robot
     robot.connect()
 
-    if not robot.is_connected:
-        raise ValueError("Robot is not connected!")
+    try:
+        if not robot.is_connected:
+            raise ValueError("Robot is not connected!")
 
-    print("Starting replay loop...")
-    log_say(f"Replaying episode {EPISODE_IDX}")
-    for idx in range(len(episode_frames)):
-        t0 = time.perf_counter()
+        print("Starting replay loop...")
+        log_say(f"Replaying episode {EPISODE_IDX}")
+        for idx in range(len(episode_frames)):
+            t0 = time.perf_counter()
 
-        # Get recorded action from dataset
-        ee_action = {
-            name: float(actions[idx][ACTION][i]) for i, name in enumerate(dataset.features[ACTION]["names"])
-        }
+            # Get recorded action from dataset
+            ee_action = {
+                name: float(actions[idx][ACTION][i])
+                for i, name in enumerate(dataset.features[ACTION]["names"])
+            }
 
-        # Get robot observation
-        robot_obs = robot.get_observation()
+            # Get robot observation
+            robot_obs = robot.get_observation()
 
-        # Dataset EE -> robot joints
-        joint_action = robot_ee_to_joints_processor((ee_action, robot_obs))
+            # Dataset EE -> robot joints
+            joint_action = robot_ee_to_joints_processor((ee_action, robot_obs))
 
-        # Send action to robot
-        _ = robot.send_action(joint_action)
+            # Send action to robot
+            _ = robot.send_action(joint_action)
 
-        precise_sleep(1.0 / dataset.fps - (time.perf_counter() - t0))
-
-    # Clean up
-    robot.disconnect()
+            precise_sleep(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
+    finally:
+        # Clean up
+        robot.disconnect()
 
 
 if __name__ == "__main__":

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_dataset.py

@@ -27,8 +27,8 @@ measuring consistency and ground truth alignment.
 Usage:
     # Basic usage with smolvla policy
     uv run python examples/rtc/eval_dataset.py \
-        --policy.path=helper2424/smolvla_check_rtc_last3 \
-        --dataset.repo_id=helper2424/check_rtc \
+        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --dataset.repo_id=<USER>/check_rtc \
         --rtc.execution_horizon=8 \
         --device=mps \
         --rtc.max_guidance_weight=10.0 \
@@ -58,16 +58,16 @@ Usage:
         --device=cuda
 
     uv run python examples/rtc/eval_dataset.py \
-        --policy.path=lipsop/reuben_pi0 \
-        --dataset.repo_id=ReubenLim/so101_cube_in_cup \
+        --policy.path=<USER>/reuben_pi0 \
+        --dataset.repo_id=<USER>/so101_cube_in_cup \
         --rtc.execution_horizon=8 \
         --device=cuda
 
     # With torch.compile for faster inference (PyTorch 2.0+)
     # Note: CUDA graphs disabled by default due to in-place ops in denoising loop
     uv run python examples/rtc/eval_dataset.py \
-        --policy.path=helper2424/smolvla_check_rtc_last3 \
-        --dataset.repo_id=helper2424/check_rtc \
+        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --dataset.repo_id=<USER>/check_rtc \
         --rtc.execution_horizon=8 \
         --device=mps \
         --use_torch_compile=true \
@@ -75,8 +75,8 @@ Usage:
 
     # With torch.compile on CUDA (CUDA graphs disabled by default)
     uv run python examples/rtc/eval_dataset.py \
-        --policy.path=helper2424/smolvla_check_rtc_last3 \
-        --dataset.repo_id=helper2424/check_rtc \
+        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --dataset.repo_id=<USER>/check_rtc \
         --rtc.execution_horizon=8 \
         --device=cuda \
         --use_torch_compile=true \
@@ -84,8 +84,8 @@ Usage:
 
     # Enable CUDA graphs (advanced - may cause tensor aliasing errors)
     uv run python examples/rtc/eval_dataset.py \
-        --policy.path=helper2424/smolvla_check_rtc_last3 \
-        --dataset.repo_id=helper2424/check_rtc \
+        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --dataset.repo_id=<USER>/check_rtc \
         --use_torch_compile=true \
         --torch_compile_backend=inductor \
         --torch_compile_mode=max-autotune \

examples/rtc/eval_with_real_robot.py

@@ -28,7 +28,7 @@ For simulation environments, see eval_with_simulation.py
 Usage:
     # Run RTC with Real robot with RTC
     uv run examples/rtc/eval_with_real_robot.py \
-        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --policy.path=<USER>/smolvla_check_rtc_last3 \
         --policy.device=mps \
         --rtc.enabled=true \
         --rtc.execution_horizon=20 \
@@ -41,7 +41,7 @@ Usage:
 
     # Run RTC with Real robot without RTC
     uv run examples/rtc/eval_with_real_robot.py \
-        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --policy.path=<USER>/smolvla_check_rtc_last3 \
         --policy.device=mps \
         --rtc.enabled=false \
         --robot.type=so100_follower \
@@ -53,7 +53,7 @@ Usage:
 
     # Run RTC with Real robot with pi0.5 policy
     uv run examples/rtc/eval_with_real_robot.py \
-        --policy.path=helper2424/pi05_check_rtc \
+        --policy.path=<USER>/pi05_check_rtc \
         --policy.device=mps \
         --rtc.enabled=true \
         --rtc.execution_horizon=20 \
@@ -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
@@ -143,38 +142,24 @@ def main():
     listener, events = init_keyboard_listener()
     init_rerun(session_name="so100_so100_evaluate")
 
-    if not robot.is_connected:
-        raise ValueError("Robot is not connected!")
+    try:
+        if not robot.is_connected:
+            raise ValueError("Robot is not connected!")
 
-    print("Starting evaluate loop...")
-    episode_idx = 0
-    for episode_idx in range(NUM_EPISODES):
-        log_say(f"Running inference, recording eval episode {episode_idx + 1} of {NUM_EPISODES}")
+        print("Starting evaluate loop...")
+        episode_idx = 0
+        for episode_idx in range(NUM_EPISODES):
+            log_say(f"Running inference, recording eval episode {episode_idx + 1} of {NUM_EPISODES}")
 
-        # Main record loop
-        record_loop(
-            robot=robot,
-            events=events,
-            fps=FPS,
-            policy=policy,
-            preprocessor=preprocessor,  # Pass the pre and post policy processors
-            postprocessor=postprocessor,
-            dataset=dataset,
-            control_time_s=EPISODE_TIME_SEC,
-            single_task=TASK_DESCRIPTION,
-            display_data=True,
-            teleop_action_processor=make_default_teleop_action_processor(),
-            robot_action_processor=robot_ee_to_joints_processor,
-            robot_observation_processor=robot_joints_to_ee_pose_processor,
-        )
-
-        # Reset the environment if not stopping or re-recording
-        if not events["stop_recording"] and ((episode_idx < NUM_EPISODES - 1) or events["rerecord_episode"]):
-            log_say("Reset the environment")
+            # Main record loop
             record_loop(
                 robot=robot,
                 events=events,
                 fps=FPS,
+                policy=policy,
+                preprocessor=preprocessor,  # Pass the pre and post policy processors
+                postprocessor=postprocessor,
+                dataset=dataset,
                 control_time_s=EPISODE_TIME_SEC,
                 single_task=TASK_DESCRIPTION,
                 display_data=True,
@@ -183,24 +168,41 @@ def main():
                 robot_observation_processor=robot_joints_to_ee_pose_processor,
             )
 
-        if events["rerecord_episode"]:
-            log_say("Re-record episode")
-            events["rerecord_episode"] = False
-            events["exit_early"] = False
-            dataset.clear_episode_buffer()
-            continue
+            # Reset the environment if not stopping or re-recording
+            if not events["stop_recording"] and (
+                (episode_idx < NUM_EPISODES - 1) or events["rerecord_episode"]
+            ):
+                log_say("Reset the environment")
+                record_loop(
+                    robot=robot,
+                    events=events,
+                    fps=FPS,
+                    control_time_s=EPISODE_TIME_SEC,
+                    single_task=TASK_DESCRIPTION,
+                    display_data=True,
+                    teleop_action_processor=make_default_teleop_action_processor(),
+                    robot_action_processor=robot_ee_to_joints_processor,
+                    robot_observation_processor=robot_joints_to_ee_pose_processor,
+                )
 
-        # Save episode
-        dataset.save_episode()
-        episode_idx += 1
+            if events["rerecord_episode"]:
+                log_say("Re-record episode")
+                events["rerecord_episode"] = False
+                events["exit_early"] = False
+                dataset.clear_episode_buffer()
+                continue
 
-    # Clean up
-    log_say("Stop recording")
-    robot.disconnect()
-    listener.stop()
+            # Save episode
+            dataset.save_episode()
+            episode_idx += 1
+    finally:
+        # Clean up
+        log_say("Stop recording")
+        robot.disconnect()
+        listener.stop()
 
-    dataset.finalize()
-    dataset.push_to_hub()
+        dataset.finalize()
+        dataset.push_to_hub()
 
 
 if __name__ == "__main__":

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
@@ -148,38 +146,23 @@ def main():
     listener, events = init_keyboard_listener()
     init_rerun(session_name="recording_phone")
 
-    if not leader.is_connected or not follower.is_connected:
-        raise ValueError("Robot or teleop is not connected!")
+    try:
+        if not leader.is_connected or not follower.is_connected:
+            raise ValueError("Robot or teleop is not connected!")
 
-    print("Starting record loop...")
-    episode_idx = 0
-    while episode_idx < NUM_EPISODES and not events["stop_recording"]:
-        log_say(f"Recording episode {episode_idx + 1} of {NUM_EPISODES}")
+        print("Starting record loop...")
+        episode_idx = 0
+        while episode_idx < NUM_EPISODES and not events["stop_recording"]:
+            log_say(f"Recording episode {episode_idx + 1} of {NUM_EPISODES}")
 
-        # Main record loop
-        record_loop(
-            robot=follower,
-            events=events,
-            fps=FPS,
-            teleop=leader,
-            dataset=dataset,
-            control_time_s=EPISODE_TIME_SEC,
-            single_task=TASK_DESCRIPTION,
-            display_data=True,
-            teleop_action_processor=leader_joints_to_ee,
-            robot_action_processor=ee_to_follower_joints,
-            robot_observation_processor=follower_joints_to_ee,
-        )
-
-        # Reset the environment if not stopping or re-recording
-        if not events["stop_recording"] and (episode_idx < NUM_EPISODES - 1 or events["rerecord_episode"]):
-            log_say("Reset the environment")
+            # Main record loop
             record_loop(
                 robot=follower,
                 events=events,
                 fps=FPS,
                 teleop=leader,
-                control_time_s=RESET_TIME_SEC,
+                dataset=dataset,
+                control_time_s=EPISODE_TIME_SEC,
                 single_task=TASK_DESCRIPTION,
                 display_data=True,
                 teleop_action_processor=leader_joints_to_ee,
@@ -187,25 +170,44 @@ def main():
                 robot_observation_processor=follower_joints_to_ee,
             )
 
-        if events["rerecord_episode"]:
-            log_say("Re-recording episode")
-            events["rerecord_episode"] = False
-            events["exit_early"] = False
-            dataset.clear_episode_buffer()
-            continue
+            # Reset the environment if not stopping or re-recording
+            if not events["stop_recording"] and (
+                episode_idx < NUM_EPISODES - 1 or events["rerecord_episode"]
+            ):
+                log_say("Reset the environment")
+                record_loop(
+                    robot=follower,
+                    events=events,
+                    fps=FPS,
+                    teleop=leader,
+                    control_time_s=RESET_TIME_SEC,
+                    single_task=TASK_DESCRIPTION,
+                    display_data=True,
+                    teleop_action_processor=leader_joints_to_ee,
+                    robot_action_processor=ee_to_follower_joints,
+                    robot_observation_processor=follower_joints_to_ee,
+                )
 
-        # Save episode
-        dataset.save_episode()
-        episode_idx += 1
+            if events["rerecord_episode"]:
+                log_say("Re-recording episode")
+                events["rerecord_episode"] = False
+                events["exit_early"] = False
+                dataset.clear_episode_buffer()
+                continue
 
-    # Clean up
-    log_say("Stop recording")
-    leader.disconnect()
-    follower.disconnect()
-    listener.stop()
+            # Save episode
+            dataset.save_episode()
+            episode_idx += 1
 
-    dataset.finalize()
-    dataset.push_to_hub()
+    finally:
+        # Clean up
+        log_say("Stop recording")
+        leader.disconnect()
+        follower.disconnect()
+        listener.stop()
+
+        dataset.finalize()
+        dataset.push_to_hub()
 
 
 if __name__ == "__main__":

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
@@ -75,32 +74,35 @@ def main():
     # Connect to the robot
     robot.connect()
 
-    if not robot.is_connected:
-        raise ValueError("Robot is not connected!")
+    try:
+        if not robot.is_connected:
+            raise ValueError("Robot is not connected!")
 
-    print("Starting replay loop...")
-    log_say(f"Replaying episode {EPISODE_IDX}")
-    for idx in range(len(episode_frames)):
-        t0 = time.perf_counter()
+        print("Starting replay loop...")
+        log_say(f"Replaying episode {EPISODE_IDX}")
+        for idx in range(len(episode_frames)):
+            t0 = time.perf_counter()
 
-        # Get recorded action from dataset
-        ee_action = {
-            name: float(actions[idx][ACTION][i]) for i, name in enumerate(dataset.features[ACTION]["names"])
-        }
+            # Get recorded action from dataset
+            ee_action = {
+                name: float(actions[idx][ACTION][i])
+                for i, name in enumerate(dataset.features[ACTION]["names"])
+            }
 
-        # Get robot observation
-        robot_obs = robot.get_observation()
+            # Get robot observation
+            robot_obs = robot.get_observation()
 
-        # Dataset EE -> robot joints
-        joint_action = robot_ee_to_joints_processor((ee_action, robot_obs))
+            # Dataset EE -> robot joints
+            joint_action = robot_ee_to_joints_processor((ee_action, robot_obs))
 
-        # Send action to robot
-        _ = robot.send_action(joint_action)
+            # 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()
+    finally:
+        # Clean up
+        robot.disconnect()
 
 
 if __name__ == "__main__":

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():
@@ -30,6 +30,7 @@ def main():
         robot=robot_cfg,
         server_address=server_address,
         policy_device="mps",
+        client_device="cpu",
         policy_type="act",
         pretrained_name_or_path="<user>/robot_learning_tutorial_act",
         chunk_size_threshold=0.5,  # g

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

@@ -25,11 +25,11 @@ discord = "https://discord.gg/s3KuuzsPFb"
 
 [project]
 name = "lerobot"
-version = "0.4.3"
+version = "0.4.5"
 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"
+requires-python = ">=3.12"
 authors = [
     { name = "Rémi Cadène", email = "re.cadene@gmail.com" },
     { name = "Simon Alibert", email = "alibert.sim@gmail.com" },
@@ -50,7 +50,8 @@ classifiers = [
     "Intended Audience :: Education",
     "Intended Audience :: Science/Research",
     "License :: OSI Approved :: Apache Software License",
-    "Programming Language :: Python :: 3.10",
+    "Programming Language :: Python :: 3.12",
+    "Programming Language :: Python :: 3.13",
     "Topic :: Software Development :: Build Tools",
     "Topic :: Scientific/Engineering :: Artificial Intelligence",
 ]
@@ -59,28 +60,30 @@ keywords = ["lerobot", "huggingface", "robotics",  "machine learning", "artifici
 dependencies = [
 
     # Hugging Face dependencies
-    "datasets>=4.0.0,<4.2.0",
+    "datasets>=4.0.0,<5.0.0",
     "diffusers>=0.27.2,<0.36.0",
-    "huggingface-hub[hf-transfer,cli]>=0.34.2,<0.36.0",
+    "huggingface-hub>=1.0.0,<2.0.0",
     "accelerate>=1.10.0,<2.0.0",
 
     # Core dependencies
+    "numpy>=2.0.0,<2.3.0", # NOTE: Explicitly listing numpy helps the resolver converge faster. Upper bound imposed by opencv-python-headless.
     "setuptools>=71.0.0,<81.0.0",
     "cmake>=3.29.0.1,<4.2.0",
+    "packaging>=24.2,<26.0",
+
+    "torch>=2.2.1,<2.11.0",
+    "torchcodec>=0.2.1,<0.11.0; 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')",
+    "torchvision>=0.21.0,<0.26.0",
+
     "einops>=0.8.0,<0.9.0",
     "opencv-python-headless>=4.9.0,<4.13.0",
     "av>=15.0.0,<16.0.0",
     "jsonlines>=4.0.0,<5.0.0",
-    "packaging>=24.2,<26.0",
-    "pynput>=1.7.7,<1.9.0",
+    "pynput>=1.7.8,<1.9.0",
     "pyserial>=3.5,<4.0",
-    "wandb>=0.20.0,<0.22.0", # TODO: Bumb dependency (compatible with protobuf)
 
-    "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
-    "torchvision>=0.21.0,<0.23.0", # TODO: Bumb dependency
-
-    "draccus==0.10.0", # TODO: Remove ==
+    "wandb>=0.24.0,<0.25.0",
+    "draccus==0.10.0", # TODO: Relax version constraint
     "gymnasium>=1.1.1,<2.0.0",
     "rerun-sdk>=0.24.0,<0.27.0",
 
@@ -95,43 +98,56 @@ dependencies = [
 
 # Common
 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)
+placo-dep = ["placo>=0.9.6,<0.9.17"]
+transformers-dep = ["transformers>=5.3.0,<6.0.0"]
+grpcio-dep = ["grpcio==1.73.1", "protobuf>=6.31.1,<6.32.0"]
+can-dep = ["python-can>=4.2.0,<5.0.0"]
+peft-dep = ["peft>=0.18.0,<1.0.0"]
+scipy-dep = ["scipy>=1.14.0,<2.0.0"]
+qwen-vl-utils-dep = ["qwen-vl-utils>=0.0.11,<0.1.0"]
+matplotlib-dep = ["matplotlib>=3.10.3,<4.0.0", "contourpy>=1.3.0,<2.0.0"] # NOTE: Explicitly listing contourpy helps the resolver converge faster.
 
 # Motors
 feetech = ["feetech-servo-sdk>=1.0.0,<2.0.0"]
 dynamixel = ["dynamixel-sdk>=3.7.31,<3.9.0"]
+damiao = ["lerobot[can-dep]"]
+robstride = ["lerobot[can-dep]"]
 
 # Robots
+openarms = ["lerobot[damiao]"]
 gamepad = ["lerobot[pygame-dep]", "hidapi>=0.14.0,<0.15.0"]
 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",
+    "pin>=3.0.0,<4.0.0",
+    "meshcat>=0.3.0,<0.4.0",
+    "lerobot[matplotlib-dep]",
+    "casadi>=3.6.0,<4.0.0",
 ]
-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'",
-    "pyrealsense2-macosx>=2.54,<2.55.0 ; sys_platform == 'darwin'",
+    "pyrealsense2-macosx>=2.54,<2.57.0 ; sys_platform == 'darwin'",
 ]
-phone = ["hebi-py>=2.8.0,<2.12.0", "teleop>=0.1.0,<0.2.0", "fastapi<1.0"]
+phone = ["hebi-py>=2.8.0,<2.12.0", "teleop>=0.1.0,<0.2.0", "fastapi<1.0", "lerobot[scipy-dep]"]
 
 # Policies
 wallx = [
-    "transformers==4.49.0",
-    "peft==0.17.1",
-    "scipy==1.15.3",
-    "torchdiffeq==0.2.5",
-    "qwen_vl_utils==0.0.11"
+    "lerobot[transformers-dep]",
+    "lerobot[peft]",
+    "lerobot[scipy-dep]",
+    "torchdiffeq>=0.2.4,<0.3.0",
+    "lerobot[qwen-vl-utils-dep]",
 ]
-pi = ["transformers @ git+https://github.com/huggingface/transformers.git@fix/lerobot_openpi"]
+pi = ["lerobot[transformers-dep]", "lerobot[scipy-dep]"]
 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"]
+multi_task_dit = ["lerobot[transformers-dep]"]
 groot = [
     "lerobot[transformers-dep]",
-    "peft>=0.13.0,<1.0.0",
+    "lerobot[peft]",
     "dm-tree>=0.1.8,<1.0.0",
     "timm>=1.0.0,<1.1.0",
     "safetensors>=0.4.3,<1.0.0",
@@ -140,12 +156,13 @@ 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", "lerobot[matplotlib-dep]", "lerobot[qwen-vl-utils-dep]"]
 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"]
+async = ["lerobot[grpcio-dep]", "lerobot[matplotlib-dep]"]
+peft = ["lerobot[transformers-dep]", "lerobot[peft-dep]"]
 
 # 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"]
@@ -153,13 +170,19 @@ test = ["pytest>=8.1.0,<9.0.0", "pytest-timeout>=2.4.0,<3.0.0", "pytest-cov>=5.0
 video_benchmark = ["scikit-image>=0.23.2,<0.26.0", "pandas>=2.2.2,<2.4.0"]
 
 # Simulation
-aloha = ["gym-aloha>=0.1.2,<0.2.0"]
+# NOTE: Explicitly listing scipy helps flatten the dependecy tree.
+aloha = ["gym-aloha>=0.1.2,<0.2.0", "lerobot[scipy-dep]"]
 pusht = ["gym-pusht>=0.1.5,<0.2.0", "pymunk>=6.6.0,<7.0.0"] # TODO: Fix pymunk version in gym-pusht instead
-libero = ["lerobot[transformers-dep]", "hf-libero>=0.1.3,<0.2.0"]
-metaworld = ["metaworld==3.0.0"]
+libero = ["lerobot[transformers-dep]", "hf-libero>=0.1.3,<0.2.0; sys_platform == 'linux'", "lerobot[scipy-dep]"]
+metaworld = ["metaworld==3.0.0", "lerobot[scipy-dep]"]
 
 # All
 all = [
+    # NOTE(resolver hint): scipy is pulled in transitively via lerobot[scipy-dep] through
+    # multiple extras (aloha, metaworld, pi, wallx, phone). Listing it explicitly
+    # helps pip's resolver converge by constraining scipy early, before it encounters
+    # the loose scipy requirements from transitive deps like dm-control and metaworld.
+    "scipy>=1.14.0,<2.0.0",
     "lerobot[dynamixel]",
     "lerobot[gamepad]",
     "lerobot[hopejr]",
@@ -167,8 +190,8 @@ all = [
     "lerobot[reachy2]",
     "lerobot[kinematics]",
     "lerobot[intelrealsense]",
-    # "lerobot[wallx]",
-    # "lerobot[pi]", TODO(Pepijn): Update pi to transformers v5
+    "lerobot[wallx]",
+    "lerobot[pi]",
     "lerobot[smolvla]",
     # "lerobot[groot]", TODO(Steven): Gr00t requires specific installation instructions for flash-attn
     "lerobot[xvla]",
@@ -180,9 +203,10 @@ all = [
     "lerobot[aloha]",
     "lerobot[pusht]",
     "lerobot[phone]",
-    "lerobot[libero]",
+    "lerobot[libero]; sys_platform == 'linux'",
     "lerobot[metaworld]",
-    "lerobot[sarm]"
+    "lerobot[sarm]",
+    "lerobot[peft]",
 ]
 
 [project.scripts]
@@ -195,18 +219,23 @@ 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"
 lerobot-imgtransform-viz="lerobot.scripts.lerobot_imgtransform_viz:main"
 lerobot-edit-dataset="lerobot.scripts.lerobot_edit_dataset:main"
+lerobot-setup-can="lerobot.scripts.lerobot_setup_can:main"
 
 # ---------------- Tool Configurations ----------------
+[tool.setuptools.package-data]
+lerobot = ["envs/*.json"]
+
 [tool.setuptools.packages.find]
 where = ["src"]
 
 [tool.ruff]
-target-version = "py310"
+target-version = "py312"
 line-length = 110
 exclude = ["tests/artifacts/**/*.safetensors", "*_pb2.py", "*_pb2_grpc.py"]
 
@@ -275,6 +304,7 @@ default.extend-ignore-identifiers-re = [
     "thw",
     "inpt",
     "ROBOTIS",
+    "OT_VALUE"
 ]
 
 # TODO: Uncomment when ready to use
@@ -297,7 +327,7 @@ default.extend-ignore-identifiers-re = [
 # Uncomment [tool.mypy] first, then uncomment individual module overrides as they get proper type annotations
 
 [tool.mypy]
-python_version = "3.10"
+python_version = "3.12"
 ignore_missing_imports = true
 follow_imports = "skip"
 # warn_return_any = true
@@ -349,9 +379,9 @@ ignore_errors = false
 module = "lerobot.cameras.*"
 ignore_errors = false
 
-# [[tool.mypy.overrides]]
-# module = "lerobot.motors.*"
-# ignore_errors = false
+[[tool.mypy.overrides]]
+module = "lerobot.motors.*"
+ignore_errors = false
 
 # [[tool.mypy.overrides]]
 # module = "lerobot.robots.*"
@@ -381,77 +411,3 @@ ignore_errors = false
 # [[tool.mypy.overrides]]
 # module = "lerobot.scripts.*"
 # ignore_errors = false
-
-[tool.uv]
-# wallx requires transformers==4.49.0 which conflicts with other extras that need >=4.53.0
-conflicts = [
-    [
-        { extra = "wallx" },
-        { extra = "transformers-dep" },
-    ],
-    [
-        { extra = "wallx" },
-        { extra = "pi" },
-    ],
-    [
-        { extra = "wallx" },
-        { extra = "smolvla" },
-    ],
-    [
-        { extra = "wallx" },
-        { extra = "groot" },
-    ],
-    [
-        { extra = "wallx" },
-        { extra = "xvla" },
-    ],
-    [
-        { extra = "wallx" },
-        { extra = "sarm" },
-    ],
-    [
-        { extra = "wallx" },
-        { extra = "hilserl" },
-    ],
-    [
-        { extra = "wallx" },
-        { extra = "libero" },
-    ],
-    [
-        { extra = "wallx" },
-        { extra = "all" },
-    ],
-    # pi uses custom branch which conflicts with transformers-dep
-    [
-        { extra = "pi" },
-        { extra = "transformers-dep" },
-    ],
-    [
-        { extra = "pi" },
-        { extra = "smolvla" },
-    ],
-    [
-        { extra = "pi" },
-        { extra = "groot" },
-    ],
-    [
-        { extra = "pi" },
-        { extra = "xvla" },
-    ],
-    [
-        { extra = "pi" },
-        { extra = "sarm" },
-    ],
-    [
-        { extra = "pi" },
-        { extra = "hilserl" },
-    ],
-    [
-        { extra = "pi" },
-        { extra = "libero" },
-    ],
-    [
-        { 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/configs.py

@@ -126,6 +126,12 @@ class RobotClientConfig:
 
     # Device configuration
     policy_device: str = field(default="cpu", metadata={"help": "Device for policy inference"})
+    client_device: str = field(
+        default="cpu",
+        metadata={
+            "help": "Device to move actions to after receiving from server (e.g., for downstream planners)"
+        },
+    )
 
     # Control behavior configuration
     chunk_size_threshold: float = field(default=0.5, metadata={"help": "Threshold for chunk size control"})
@@ -161,6 +167,9 @@ class RobotClientConfig:
         if not self.policy_device:
             raise ValueError("policy_device cannot be empty")
 
+        if not self.client_device:
+            raise ValueError("client_device cannot be empty")
+
         if self.chunk_size_threshold < 0 or self.chunk_size_threshold > 1:
             raise ValueError(f"chunk_size_threshold must be between 0 and 1, got {self.chunk_size_threshold}")
 
@@ -184,6 +193,7 @@ class RobotClientConfig:
             "policy_type": self.policy_type,
             "pretrained_name_or_path": self.pretrained_name_or_path,
             "policy_device": self.policy_device,
+            "client_device": self.client_device,
             "chunk_size_threshold": self.chunk_size_threshold,
             "fps": self.fps,
             "actions_per_chunk": self.actions_per_chunk,

src/lerobot/async_inference/constants.py

@@ -23,7 +23,7 @@ DEFAULT_INFERENCE_LATENCY = 1 / DEFAULT_FPS
 DEFAULT_OBS_QUEUE_TIMEOUT = 2
 
 # All action chunking policies
-SUPPORTED_POLICIES = ["act", "smolvla", "diffusion", "tdmpc", "vqbet", "pi0", "pi05"]
+SUPPORTED_POLICIES = ["act", "smolvla", "diffusion", "tdmpc", "vqbet", "pi0", "pi05", "groot"]
 
 # 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/helpers.py

@@ -18,6 +18,7 @@ import os
 import time
 from dataclasses import dataclass, field
 from pathlib import Path
+from typing import Any
 
 import torch
 
@@ -39,8 +40,8 @@ from lerobot.utils.utils import init_logging
 
 Action = torch.Tensor
 
-# observation as received from the robot
-RawObservation = dict[str, torch.Tensor]
+# observation as received from the robot (can be numpy arrays, floats, etc.)
+RawObservation = dict[str, Any]
 
 # observation as those recorded in LeRobot dataset (keys are different)
 LeRobotObservation = dict[str, torch.Tensor]

src/lerobot/async_inference/policy_server.py

@@ -381,6 +381,8 @@ class PolicyServer(services_pb2_grpc.AsyncInferenceServicer):
         action_tensor = torch.stack(processed_actions, dim=1).squeeze(0)
         self.logger.debug(f"Postprocessed action shape: {action_tensor.shape}")
 
+        action_tensor = action_tensor.detach().cpu()
+
         """5. Convert to TimedAction list"""
         action_chunk = self._time_action_chunk(
             observation_t.get_timestamp(), list(action_tensor), observation_t.get_timestep()

src/lerobot/async_inference/robot_client.py

@@ -25,6 +25,7 @@ python src/lerobot/async_inference/robot_client.py \
     --policy_type=act \
     --pretrained_name_or_path=user/model \
     --policy_device=mps \
+    --client_device=cpu \
     --actions_per_chunk=50 \
     --chunk_size_threshold=0.5 \
     --aggregate_fn_name=weighted_average \
@@ -51,21 +52,20 @@ from lerobot.cameras.realsense.configuration_realsense import RealSenseCameraCon
 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
     services_pb2_grpc,  # type: ignore
 )
 from lerobot.transport.utils import grpc_channel_options, send_bytes_in_chunks
+from lerobot.utils.import_utils import register_third_party_plugins
 
 from .configs import RobotClientConfig
-from .constants import SUPPORTED_ROBOTS
 from .helpers import (
     Action,
     FPSTracker,
@@ -286,6 +286,21 @@ class RobotClient:
                 timed_actions = pickle.loads(actions_chunk.data)  # nosec
                 deserialize_time = time.perf_counter() - deserialize_start
 
+                # Log device type of received actions
+                if len(timed_actions) > 0:
+                    received_device = timed_actions[0].get_action().device.type
+                    self.logger.debug(f"Received actions on device: {received_device}")
+
+                # Move actions to client_device (e.g., for downstream planners that need GPU)
+                client_device = self.config.client_device
+                if client_device != "cpu":
+                    for timed_action in timed_actions:
+                        if timed_action.get_action().device.type != client_device:
+                            timed_action.action = timed_action.get_action().to(client_device)
+                    self.logger.debug(f"Converted actions to device: {client_device}")
+                else:
+                    self.logger.debug(f"Actions kept on device: {client_device}")
+
                 self.action_chunk_size = max(self.action_chunk_size, len(timed_actions))
 
                 # Calculate network latency if we have matching observations
@@ -470,8 +485,9 @@ class RobotClient:
 def async_client(cfg: RobotClientConfig):
     logging.info(pformat(asdict(cfg)))
 
-    if cfg.robot.type not in SUPPORTED_ROBOTS:
-        raise ValueError(f"Robot {cfg.robot.type} not yet supported!")
+    # TODO: Assert if checking robot support is still needed with the plugin system
+    # if cfg.robot.type not in SUPPORTED_ROBOTS:
+    #     raise ValueError(f"Robot {cfg.robot.type} not yet supported!")
 
     client = RobotClient(cfg)
 
@@ -497,4 +513,5 @@ def async_client(cfg: RobotClientConfig):
 
 
 if __name__ == "__main__":
+    register_third_party_plugins()
     async_client()  # run the client

src/lerobot/cameras/__init__.py

@@ -13,5 +13,5 @@
 # limitations under the License.
 
 from .camera import Camera
-from .configs import CameraConfig, ColorMode, Cv2Rotation
+from .configs import CameraConfig, ColorMode, Cv2Backends, Cv2Rotation
 from .utils import make_cameras_from_configs

src/lerobot/cameras/camera.py

@@ -15,11 +15,12 @@
 # limitations under the License.
 
 import abc
+import warnings
 from typing import Any
 
 from numpy.typing import NDArray  # type: ignore  # TODO: add type stubs for numpy.typing
 
-from .configs import CameraConfig, ColorMode
+from .configs import CameraConfig
 
 
 class Camera(abc.ABC):
@@ -30,20 +31,12 @@ class Camera(abc.ABC):
 
     Manages basic camera properties (FPS, resolution) and core operations:
     - Connection/disconnection
-    - Frame capture (sync/async)
+    - Frame capture (sync/async/latest)
 
     Attributes:
         fps (int | None): Configured frames per second
         width (int | None): Frame width in pixels
         height (int | None): Frame height in pixels
-
-    Example:
-        class MyCamera(Camera):
-            def __init__(self, config): ...
-            @property
-            def is_connected(self) -> bool: ...
-            def connect(self, warmup=True): ...
-            # Plus other required methods
     """
 
     def __init__(self, config: CameraConfig):
@@ -56,6 +49,32 @@ class Camera(abc.ABC):
         self.width: int | None = config.width
         self.height: int | None = config.height
 
+    def __enter__(self):
+        """
+        Context manager entry.
+        Automatically connects to the camera.
+        """
+        self.connect()
+        return self
+
+    def __exit__(self, exc_type, exc_value, traceback) -> None:
+        """
+        Context manager exit.
+        Automatically disconnects, ensuring resources are released even on error.
+        """
+        self.disconnect()
+
+    def __del__(self) -> None:
+        """
+        Destructor safety net.
+        Attempts to disconnect if the object is garbage collected without cleanup.
+        """
+        try:
+            if self.is_connected:
+                self.disconnect()
+        except Exception:  # nosec B110
+            pass
+
     @property
     @abc.abstractmethod
     def is_connected(self) -> bool:
@@ -89,12 +108,10 @@ class Camera(abc.ABC):
         pass
 
     @abc.abstractmethod
-    def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
-        """Capture and return a single frame from the camera.
+    def read(self) -> NDArray[Any]:
+        """Capture and return a single frame from the camera synchronously.
 
-        Args:
-            color_mode: Desired color mode for the output frame. If None,
-                        uses the camera's default color mode.
+        This is a blocking call that will wait for the hardware and its SDK.
 
         Returns:
             np.ndarray: Captured frame as a numpy array.
@@ -103,17 +120,64 @@ class Camera(abc.ABC):
 
     @abc.abstractmethod
     def async_read(self, timeout_ms: float = ...) -> NDArray[Any]:
-        """Asynchronously capture and return a single frame from the camera.
+        """Return the most recent new frame.
+
+        This method retrieves the latest frame captured by the background thread.
+        If a new frame is already available in the buffer (captured since the last call),
+        it returns it immediately.
+
+        It blocks up to `timeout_ms` only if the buffer is empty or if the latest frame
+        was already consumed by a previous `async_read` call.
+
+        Essentially, this method return the latest unconsumed frame, waiting if necessary
+        for a new one to arrive within the specified timeout.
+
+        Usage:
+            - Ideal for control loops where you want to ensure every processed frame
+            is fresh, effectively synchronizing your loop to the camera's FPS.
+            - Causes of a timeout usually include: very low camera FPS, heavy processing load,
+            or if the camera is disconnected.
 
         Args:
-            timeout_ms: Maximum time to wait for a frame in milliseconds.
-                        Defaults to implementation-specific timeout.
+            timeout_ms: Maximum time to wait for a new frame in milliseconds.
+                        Defaults to 200ms (0.2s).
 
         Returns:
             np.ndarray: Captured frame as a numpy array.
+
+        Raises:
+            TimeoutError: If no new frame arrives within `timeout_ms`.
         """
         pass
 
+    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
+        """Return the most recent frame captured immediately (Peeking).
+
+        This method is non-blocking and returns whatever is currently in the
+        memory buffer. The frame may be stale,
+        meaning it could have been captured a while ago (hanging camera scenario e.g.).
+
+        Usage:
+            Ideal for scenarios requiring zero latency or decoupled frequencies & when
+            we want a guaranteed frame, such as UI visualization, logging, or
+            non-critical monitoring.
+
+        Returns:
+            NDArray[Any]: The frame image (numpy array).
+
+        Raises:
+            TimeoutError: If the latest frame is older than `max_age_ms`.
+            NotConnectedError: If the camera is not connected.
+            RuntimeError: If the camera is connected but has not captured any frames yet.
+        """
+        warnings.warn(
+            f"{self.__class__.__name__}.read_latest() is not implemented. "
+            "Please override read_latest(); it will be required in future releases.",
+            FutureWarning,
+            stacklevel=2,
+        )
+        return self.async_read()
+
     @abc.abstractmethod
     def disconnect(self) -> None:
         """Disconnect from the camera and release resources."""

src/lerobot/cameras/configs.py

@@ -25,6 +25,10 @@ class ColorMode(str, Enum):
     RGB = "rgb"
     BGR = "bgr"
 
+    @classmethod
+    def _missing_(cls, value: object) -> None:
+        raise ValueError(f"`color_mode` is expected to be in {list(cls)}, but {value} is provided.")
+
 
 class Cv2Rotation(int, Enum):
     NO_ROTATION = 0
@@ -32,6 +36,25 @@ class Cv2Rotation(int, Enum):
     ROTATE_180 = 180
     ROTATE_270 = -90
 
+    @classmethod
+    def _missing_(cls, value: object) -> None:
+        raise ValueError(f"`rotation` is expected to be in {list(cls)}, but {value} is provided.")
+
+
+# Subset from https://docs.opencv.org/3.4/d4/d15/group__videoio__flags__base.html
+class Cv2Backends(int, Enum):
+    ANY = 0
+    V4L2 = 200
+    DSHOW = 700
+    PVAPI = 800
+    ANDROID = 1000
+    AVFOUNDATION = 1200
+    MSMF = 1400
+
+    @classmethod
+    def _missing_(cls, value: object) -> None:
+        raise ValueError(f"`backend` is expected to be in {list(cls)}, but {value} is provided.")
+
 
 @dataclass(kw_only=True)
 class CameraConfig(draccus.ChoiceRegistry, abc.ABC):  # type: ignore  # TODO: add type stubs for draccus

src/lerobot/cameras/opencv/camera_opencv.py

@@ -32,10 +32,11 @@ 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
 
-from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
+from lerobot.utils.errors import DeviceNotConnectedError
 
 from ..camera import Camera
-from ..utils import get_cv2_backend, get_cv2_rotation
+from ..utils import get_cv2_rotation
 from .configuration_opencv import ColorMode, OpenCVCameraConfig
 
 # NOTE(Steven): The maximum opencv device index depends on your operating system. For instance,
@@ -70,34 +71,24 @@ class OpenCVCamera(Camera):
     Example:
         ```python
         from lerobot.cameras.opencv import OpenCVCamera
-        from lerobot.cameras.configuration_opencv import OpenCVCameraConfig, ColorMode, Cv2Rotation
+        from lerobot.cameras.configuration_opencv import OpenCVCameraConfig
 
         # Basic usage with camera index 0
         config = OpenCVCameraConfig(index_or_path=0)
         camera = OpenCVCamera(config)
         camera.connect()
 
-        # Read 1 frame synchronously
+        # Read 1 frame synchronously (blocking)
         color_image = camera.read()
-        print(color_image.shape)
 
-        # Read 1 frame asynchronously
+        # Read 1 frame asynchronously (waits for new frame with a timeout)
         async_image = camera.async_read()
 
+        # Get the latest frame immediately (no wait, returns timestamp)
+        latest_image, timestamp = camera.read_latest()
+
         # When done, properly disconnect the camera using
         camera.disconnect()
-
-        # Example with custom settings
-        custom_config = OpenCVCameraConfig(
-            index_or_path='/dev/video0', # Or use an index
-            fps=30,
-            width=1280,
-            height=720,
-            color_mode=ColorMode.RGB,
-            rotation=Cv2Rotation.ROTATE_90
-        )
-        custom_camera = OpenCVCamera(custom_config)
-        # ... connect, read, disconnect ...
         ```
     """
 
@@ -123,10 +114,11 @@ class OpenCVCamera(Camera):
         self.stop_event: Event | None = None
         self.frame_lock: Lock = Lock()
         self.latest_frame: NDArray[Any] | None = None
+        self.latest_timestamp: float | None = None
         self.new_frame_event: Event = Event()
 
         self.rotation: int | None = get_cv2_rotation(config.rotation)
-        self.backend: int = get_cv2_backend()
+        self.backend: int = config.backend
 
         if self.height and self.width:
             self.capture_width, self.capture_height = self.width, self.height
@@ -141,20 +133,23 @@ class OpenCVCamera(Camera):
         """Checks if the camera is currently connected and opened."""
         return isinstance(self.videocapture, cv2.VideoCapture) and self.videocapture.isOpened()
 
+    @check_if_already_connected
     def connect(self, warmup: bool = True) -> None:
         """
         Connects to the OpenCV camera specified in the configuration.
 
         Initializes the OpenCV VideoCapture object, sets desired camera properties
-        (FPS, width, height), and performs initial checks.
+        (FPS, width, height), starts the background reading thread and performs initial checks.
+
+        Args:
+            warmup (bool): If True, waits at connect() time until at least one valid frame
+                           has been captured by the background thread. Defaults to True.
 
         Raises:
             DeviceAlreadyConnectedError: If the camera is already connected.
-            ConnectionError: If the specified camera index/path is not found or the camera is found but fails to open.
-            RuntimeError: If the camera opens but fails to apply requested FPS/resolution settings.
+            ConnectionError: If the specified camera index/path is not found or fails to open.
+            RuntimeError: If the camera opens but fails to apply requested settings.
         """
-        if self.is_connected:
-            raise DeviceAlreadyConnectedError(f"{self} is already connected.")
 
         # Use 1 thread for OpenCV operations to avoid potential conflicts or
         # blocking in multi-threaded applications, especially during data collection.
@@ -170,15 +165,20 @@ class OpenCVCamera(Camera):
             )
 
         self._configure_capture_settings()
+        self._start_read_thread()
 
-        if warmup:
+        if warmup and self.warmup_s > 0:
             start_time = time.time()
             while time.time() - start_time < self.warmup_s:
-                self.read()
+                self.async_read(timeout_ms=self.warmup_s * 1000)
                 time.sleep(0.1)
+            with self.frame_lock:
+                if self.latest_frame is None:
+                    raise ConnectionError(f"{self} failed to capture frames during warmup.")
 
         logger.info(f"{self} connected.")
 
+    @check_if_not_connected
     def _configure_capture_settings(self) -> None:
         """
         Applies the specified FOURCC, FPS, width, and height settings to the connected camera.
@@ -196,11 +196,8 @@ class OpenCVCamera(Camera):
         Raises:
             RuntimeError: If the camera fails to set any of the specified properties
                           to the requested value.
-            DeviceNotConnectedError: If the camera is not connected when attempting
-                                     to configure settings.
+            DeviceNotConnectedError: If the camera is not connected.
         """
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"Cannot configure settings for {self} as it is not connected.")
 
         # Set FOURCC first (if specified) as it can affect available FPS/resolution options
         if self.config.fourcc is not None:
@@ -339,6 +336,18 @@ class OpenCVCamera(Camera):
 
         return found_cameras_info
 
+    def _read_from_hardware(self) -> NDArray[Any]:
+        if self.videocapture is None:
+            raise DeviceNotConnectedError(f"{self} videocapture is not initialized")
+
+        ret, frame = self.videocapture.read()
+
+        if not ret:
+            raise RuntimeError(f"{self} read failed (status={ret}).")
+
+        return frame
+
+    @check_if_not_connected
     def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
         """
         Reads a single frame synchronously from the camera.
@@ -346,11 +355,6 @@ class OpenCVCamera(Camera):
         This is a blocking call. It waits for the next available frame from the
         camera hardware via OpenCV.
 
-        Args:
-            color_mode (Optional[ColorMode]): If specified, overrides the default
-                color mode (`self.color_mode`) for this read operation (e.g.,
-                request RGB even if default is BGR).
-
         Returns:
             np.ndarray: The captured frame as a NumPy array in the format
                        (height, width, channels), using the specified or default
@@ -362,34 +366,31 @@ class OpenCVCamera(Camera):
                           received frame dimensions don't match expectations before rotation.
             ValueError: If an invalid `color_mode` is requested.
         """
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")
 
         start_time = time.perf_counter()
 
-        if self.videocapture is None:
-            raise DeviceNotConnectedError(f"{self} videocapture is not initialized")
+        if color_mode is not None:
+            logger.warning(
+                f"{self} read() color_mode parameter is deprecated and will be removed in future versions."
+            )
 
-        ret, frame = self.videocapture.read()
+        if self.thread is None or not self.thread.is_alive():
+            raise RuntimeError(f"{self} read thread is not running.")
 
-        if not ret or frame is None:
-            raise RuntimeError(f"{self} read failed (status={ret}).")
-
-        processed_frame = self._postprocess_image(frame, color_mode)
+        self.new_frame_event.clear()
+        frame = self.async_read(timeout_ms=10000)
 
         read_duration_ms = (time.perf_counter() - start_time) * 1e3
         logger.debug(f"{self} read took: {read_duration_ms:.1f}ms")
 
-        return processed_frame
+        return frame
 
-    def _postprocess_image(self, image: NDArray[Any], color_mode: ColorMode | None = None) -> NDArray[Any]:
+    def _postprocess_image(self, image: NDArray[Any]) -> NDArray[Any]:
         """
         Applies color conversion, dimension validation, and rotation to a raw frame.
 
         Args:
             image (np.ndarray): The raw image frame (expected BGR format from OpenCV).
-            color_mode (Optional[ColorMode]): The target color mode (RGB or BGR). If None,
-                                             uses the instance's default `self.color_mode`.
 
         Returns:
             np.ndarray: The processed image frame.
@@ -399,11 +400,10 @@ class OpenCVCamera(Camera):
             RuntimeError: If the raw frame dimensions do not match the configured
                           `width` and `height`.
         """
-        requested_color_mode = self.color_mode if color_mode is None else color_mode
 
-        if requested_color_mode not in (ColorMode.RGB, ColorMode.BGR):
+        if self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
             raise ValueError(
-                f"Invalid color mode '{requested_color_mode}'. Expected {ColorMode.RGB} or {ColorMode.BGR}."
+                f"Invalid color mode '{self.color_mode}'. Expected {ColorMode.RGB} or {ColorMode.BGR}."
             )
 
         h, w, c = image.shape
@@ -417,7 +417,7 @@ class OpenCVCamera(Camera):
             raise RuntimeError(f"{self} frame channels={c} do not match expected 3 channels (RGB/BGR).")
 
         processed_image = image
-        if requested_color_mode == ColorMode.RGB:
+        if self.color_mode == ColorMode.RGB:
             processed_image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
 
         if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE, cv2.ROTATE_180]:
@@ -431,7 +431,7 @@ class OpenCVCamera(Camera):
 
         On each iteration:
         1. Reads a color frame
-        2. Stores result in latest_frame (thread-safe)
+        2. Stores result in latest_frame and updates timestamp (thread-safe)
         3. Sets new_frame_event to notify listeners
 
         Stops on DeviceNotConnectedError, logs other errors and continues.
@@ -439,30 +439,37 @@ class OpenCVCamera(Camera):
         if self.stop_event is None:
             raise RuntimeError(f"{self}: stop_event is not initialized before starting read loop.")
 
+        failure_count = 0
         while not self.stop_event.is_set():
             try:
-                color_image = self.read()
+                raw_frame = self._read_from_hardware()
+                processed_frame = self._postprocess_image(raw_frame)
+                capture_time = time.perf_counter()
 
                 with self.frame_lock:
-                    self.latest_frame = color_image
+                    self.latest_frame = processed_frame
+                    self.latest_timestamp = capture_time
                 self.new_frame_event.set()
+                failure_count = 0
 
             except DeviceNotConnectedError:
                 break
             except Exception as e:
-                logger.warning(f"Error reading frame in background thread for {self}: {e}")
+                if failure_count <= 10:
+                    failure_count += 1
+                    logger.warning(f"Error reading frame in background thread for {self}: {e}")
+                else:
+                    raise RuntimeError(f"{self} exceeded maximum consecutive read failures.") from 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_read_thread()
 
         self.stop_event = Event()
         self.thread = Thread(target=self._read_loop, args=(), name=f"{self}_read_loop")
         self.thread.daemon = True
         self.thread.start()
+        time.sleep(0.1)
 
     def _stop_read_thread(self) -> None:
         """Signals the background read thread to stop and waits for it to join."""
@@ -475,6 +482,12 @@ class OpenCVCamera(Camera):
         self.thread = None
         self.stop_event = None
 
+        with self.frame_lock:
+            self.latest_frame = None
+            self.latest_timestamp = None
+            self.new_frame_event.clear()
+
+    @check_if_not_connected
     def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
         """
         Reads the latest available frame asynchronously.
@@ -482,6 +495,7 @@ class OpenCVCamera(Camera):
         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.
+        It is “best effort” under high FPS.
 
         Args:
             timeout_ms (float): Maximum time in milliseconds to wait for a frame
@@ -496,17 +510,14 @@ class OpenCVCamera(Camera):
             TimeoutError: If no frame becomes available within the specified timeout.
             RuntimeError: If an unexpected error occurs.
         """
-        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()
+            raise RuntimeError(f"{self} read thread is not running.")
 
         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}."
+                f"Read thread alive: {self.thread.is_alive()}."
             )
 
         with self.frame_lock:
@@ -518,6 +529,41 @@ class OpenCVCamera(Camera):
 
         return frame
 
+    @check_if_not_connected
+    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
+        """Return the most recent frame captured immediately (Peeking).
+
+        This method is non-blocking and returns whatever is currently in the
+        memory buffer. The frame may be stale,
+        meaning it could have been captured a while ago (hanging camera scenario e.g.).
+
+        Returns:
+            NDArray[Any]: The frame image (numpy array).
+
+        Raises:
+            TimeoutError: If the latest frame is older than `max_age_ms`.
+            DeviceNotConnectedError: If the camera is not connected.
+            RuntimeError: If the camera is connected but has not captured any frames yet.
+        """
+
+        if self.thread is None or not self.thread.is_alive():
+            raise RuntimeError(f"{self} read thread is not running.")
+
+        with self.frame_lock:
+            frame = self.latest_frame
+            timestamp = self.latest_timestamp
+
+        if frame is None or timestamp is None:
+            raise RuntimeError(f"{self} has not captured any frames yet.")
+
+        age_ms = (time.perf_counter() - timestamp) * 1e3
+        if age_ms > max_age_ms:
+            raise TimeoutError(
+                f"{self} latest frame is too old: {age_ms:.1f} ms (max allowed: {max_age_ms} ms)."
+            )
+
+        return frame
+
     def disconnect(self) -> None:
         """
         Disconnects from the camera and cleans up resources.
@@ -538,4 +584,9 @@ class OpenCVCamera(Camera):
             self.videocapture.release()
             self.videocapture = None
 
+        with self.frame_lock:
+            self.latest_frame = None
+            self.latest_timestamp = None
+            self.new_frame_event.clear()
+
         logger.info(f"{self} disconnected.")

src/lerobot/cameras/opencv/configuration_opencv.py

@@ -15,9 +15,9 @@
 from dataclasses import dataclass
 from pathlib import Path
 
-from ..configs import CameraConfig, ColorMode, Cv2Rotation
+from ..configs import CameraConfig, ColorMode, Cv2Backends, Cv2Rotation
 
-__all__ = ["OpenCVCameraConfig", "ColorMode", "Cv2Rotation"]
+__all__ = ["OpenCVCameraConfig", "ColorMode", "Cv2Rotation", "Cv2Backends"]
 
 
 @CameraConfig.register_subclass("opencv")
@@ -50,6 +50,7 @@ class OpenCVCameraConfig(CameraConfig):
         rotation: Image rotation setting (0°, 90°, 180°, or 270°). Defaults to no rotation.
         warmup_s: Time reading frames before returning from connect (in seconds)
         fourcc: FOURCC code for video format (e.g., "MJPG", "YUYV", "I420"). Defaults to None (auto-detect).
+        backend: OpenCV backend identifier (https://docs.opencv.org/3.4/d4/d15/group__videoio__flags__base.html). Defaults to ANY.
 
     Note:
         - Only 3-channel color output (RGB/BGR) is currently supported.
@@ -62,22 +63,12 @@ class OpenCVCameraConfig(CameraConfig):
     rotation: Cv2Rotation = Cv2Rotation.NO_ROTATION
     warmup_s: int = 1
     fourcc: str | None = None
+    backend: Cv2Backends = Cv2Backends.ANY
 
     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.rotation not in (
-            Cv2Rotation.NO_ROTATION,
-            Cv2Rotation.ROTATE_90,
-            Cv2Rotation.ROTATE_180,
-            Cv2Rotation.ROTATE_270,
-        ):
-            raise ValueError(
-                f"`rotation` is expected to be in {(Cv2Rotation.NO_ROTATION, Cv2Rotation.ROTATE_90, Cv2Rotation.ROTATE_180, Cv2Rotation.ROTATE_270)}, but {self.rotation} is provided."
-            )
+        self.color_mode = ColorMode(self.color_mode)
+        self.rotation = Cv2Rotation(self.rotation)
+        self.backend = Cv2Backends(self.backend)
 
         if self.fourcc is not None and (not isinstance(self.fourcc, str) or len(self.fourcc) != 4):
             raise ValueError(

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"]:
@@ -74,7 +74,4 @@ class Reachy2CameraConfig(CameraConfig):
                 f"`image_type` is expected to be 'left' or 'right' for teleop camera, and 'rgb' or 'depth' for depth camera, but {self.image_type} is provided."
             )
 
-        if self.color_mode not in ["rgb", "bgr"]:
-            raise ValueError(
-                f"`color_mode` is expected to be 'rgb' or 'bgr', but {self.color_mode} is provided."
-            )
+        self.color_mode = ColorMode(self.color_mode)

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 +31,20 @@ 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.decorators import check_if_not_connected
+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 +80,12 @@ class Reachy2Camera(Camera):
 
         self.config = config
 
-        self.fps = config.fps
         self.color_mode = config.color_mode
+        self.latest_frame: NDArray[Any] | None = None
+        self.latest_timestamp: float | None = None
 
         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})"
 
@@ -100,154 +106,87 @@ 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.")
 
+    @check_if_not_connected
     def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
         """
         Reads a single frame synchronously from the camera.
 
-        This is a blocking call.
-
-        Args:
-            color_mode (Optional[ColorMode]): If specified, overrides the default
-                color mode (`self.color_mode`) for this read operation (e.g.,
-                request RGB even if default is BGR).
+        This method retrieves the most recent frame available in Reachy 2's low-level software.
 
         Returns:
             np.ndarray: The captured frame as a NumPy array in the format
                        (height, width, channels), using the specified or default
                        color mode and applying any configured rotation.
         """
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")
-
         start_time = time.perf_counter()
 
-        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 color_mode is not None:
+            logger.warning(
+                f"{self} read() color_mode parameter is deprecated and will be removed in future versions."
+            )
+
+        frame: NDArray[Any] = np.empty((0, 0, 3), dtype=np.uint8)
+
+        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:
+            raise RuntimeError(f"Internal error: No frame available for {self}.")
 
-            if self.config.color_mode == "rgb":
-                frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+        if self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
+            raise ValueError(
+                f"Invalid color mode '{self.color_mode}'. Expected {ColorMode.RGB} or {ColorMode.BGR}."
+            )
+        if self.color_mode == ColorMode.RGB:
+            frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+
+        self.latest_frame = frame
+        self.latest_timestamp = time.perf_counter()
 
         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
-
+    @check_if_not_connected
     def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
         """
-        Reads the latest available frame asynchronously.
-
-        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.
-
-        Args:
-            timeout_ms (float): Maximum time in milliseconds to wait for a frame
-                to become available. Defaults to 200ms (0.2 seconds).
+        Same as read()
 
         Returns:
             np.ndarray: The latest captured frame as a NumPy array in the format
@@ -258,28 +197,40 @@ class Reachy2Camera(Camera):
             TimeoutError: If no frame becomes available within the specified timeout.
             RuntimeError: If an unexpected error occurs.
         """
-        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()
+        return self.read()
 
-        if not self.new_frame_event.wait(timeout=timeout_ms / 1000.0):
-            thread_alive = self.thread is not None and self.thread.is_alive()
+    @check_if_not_connected
+    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
+        """Return the most recent frame captured immediately (Peeking).
+
+        This method is non-blocking and returns whatever is currently in the
+        memory buffer. The frame may be stale,
+        meaning it could have been captured a while ago (hanging camera scenario e.g.).
+
+        Returns:
+            tuple[NDArray, float]:
+                - The frame image (numpy array).
+                - The timestamp (time.perf_counter) when this frame was captured.
+
+        Raises:
+            TimeoutError: If the latest frame is older than `max_age_ms`.
+            DeviceNotConnectedError: If the camera is not connected.
+            RuntimeError: If the camera is connected but has not captured any frames yet.
+        """
+
+        if self.latest_frame is None or self.latest_timestamp is None:
+            raise RuntimeError(f"{self} has not captured any frames yet.")
+
+        age_ms = (time.perf_counter() - self.latest_timestamp) * 1e3
+        if age_ms > max_age_ms:
             raise TimeoutError(
-                f"Timed out waiting for frame from camera {self} after {timeout_ms} ms. "
-                f"Read thread alive: {thread_alive}."
+                f"{self} latest frame is too old: {age_ms:.1f} ms (max allowed: {max_age_ms} ms)."
             )
 
-        with self.frame_lock:
-            frame = self.latest_frame
-            self.new_frame_event.clear()
-
-        if frame is None:
-            raise RuntimeError(f"Internal error: Event set but no frame available for {self}.")
-
-        return frame
+        return self.latest_frame
 
+    @check_if_not_connected
     def disconnect(self) -> None:
         """
         Stops the background read thread (if running).
@@ -287,11 +238,6 @@ class Reachy2Camera(Camera):
         Raises:
             DeviceNotConnectedError: If the camera is already disconnected.
         """
-        if not self.is_connected and self.thread is None:
-            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/realsense/camera_realsense.py

@@ -30,7 +30,8 @@ try:
 except Exception as e:
     logging.info(f"Could not import realsense: {e}")
 
-from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
+from lerobot.utils.errors import DeviceNotConnectedError
 
 from ..camera import Camera
 from ..configs import ColorMode
@@ -72,15 +73,14 @@ class RealSenseCamera(Camera):
         camera = RealSenseCamera(config)
         camera.connect()
 
-        # Read 1 frame synchronously
+        # Read 1 frame synchronously (blocking)
         color_image = camera.read()
-        print(color_image.shape)
 
-        # Read 1 frame asynchronously
+        # Read 1 frame asynchronously (waits for new frame with a timeout)
         async_image = camera.async_read()
 
-        # When done, properly disconnect the camera using
-        camera.disconnect()
+        # Get the latest frame immediately (no wait, returns timestamp)
+        latest_image, timestamp = camera.read_latest()
 
         # Example with depth capture and custom settings
         custom_config = RealSenseCameraConfig(
@@ -133,7 +133,9 @@ class RealSenseCamera(Camera):
         self.thread: Thread | None = None
         self.stop_event: Event | None = None
         self.frame_lock: Lock = Lock()
-        self.latest_frame: NDArray[Any] | None = None
+        self.latest_color_frame: NDArray[Any] | None = None
+        self.latest_depth_frame: NDArray[Any] | None = None
+        self.latest_timestamp: float | None = None
         self.new_frame_event: Event = Event()
 
         self.rotation: int | None = get_cv2_rotation(config.rotation)
@@ -151,6 +153,7 @@ class RealSenseCamera(Camera):
         """Checks if the camera pipeline is started and streams are active."""
         return self.rs_pipeline is not None and self.rs_profile is not None
 
+    @check_if_already_connected
     def connect(self, warmup: bool = True) -> None:
         """
         Connects to the RealSense camera specified in the configuration.
@@ -158,14 +161,16 @@ class RealSenseCamera(Camera):
         Initializes the RealSense pipeline, configures the required streams (color
         and optionally depth), starts the pipeline, and validates the actual stream settings.
 
+        Args:
+            warmup (bool): If True, waits at connect() time until at least one valid frame
+                           has been captured by the background thread. Defaults to True.
+
         Raises:
             DeviceAlreadyConnectedError: If the camera is already connected.
             ValueError: If the configuration is invalid (e.g., missing serial/name, name not unique).
             ConnectionError: If the camera is found but fails to start the pipeline or no RealSense devices are detected at all.
             RuntimeError: If the pipeline starts but fails to apply requested settings.
         """
-        if self.is_connected:
-            raise DeviceAlreadyConnectedError(f"{self} is already connected.")
 
         self.rs_pipeline = rs.pipeline()
         rs_config = rs.config()
@@ -181,15 +186,18 @@ class RealSenseCamera(Camera):
             ) from e
 
         self._configure_capture_settings()
+        self._start_read_thread()
 
-        if warmup:
-            time.sleep(
-                1
-            )  # NOTE(Steven): RS cameras need a bit of time to warm up before the first read. If we don't wait, the first read from the warmup will raise.
-            start_time = time.time()
-            while time.time() - start_time < self.warmup_s:
-                self.read()
-                time.sleep(0.1)
+        # NOTE(Steven/Caroline): Enforcing at least one second of warmup as RS cameras need a bit of time before the first read. If we don't wait, the first read from the warmup will raise.
+        self.warmup_s = max(self.warmup_s, 1)
+
+        start_time = time.time()
+        while time.time() - start_time < self.warmup_s:
+            self.async_read(timeout_ms=self.warmup_s * 1000)
+            time.sleep(0.1)
+        with self.frame_lock:
+            if self.latest_color_frame is None or self.use_depth and self.latest_depth_frame is None:
+                raise ConnectionError(f"{self} failed to capture frames during warmup.")
 
         logger.info(f"{self} connected.")
 
@@ -282,6 +290,7 @@ class RealSenseCamera(Camera):
             if self.use_depth:
                 rs_config.enable_stream(rs.stream.depth)
 
+    @check_if_not_connected
     def _configure_capture_settings(self) -> None:
         """Sets fps, width, and height from device stream if not already configured.
 
@@ -291,8 +300,6 @@ class RealSenseCamera(Camera):
         Raises:
             DeviceNotConnectedError: If device is not connected.
         """
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"Cannot validate settings for {self} as it is not connected.")
 
         if self.rs_profile is None:
             raise RuntimeError(f"{self}: rs_profile must be initialized before use.")
@@ -312,6 +319,7 @@ class RealSenseCamera(Camera):
                 self.width, self.height = actual_width, actual_height
                 self.capture_width, self.capture_height = actual_width, actual_height
 
+    @check_if_not_connected
     def read_depth(self, timeout_ms: int = 200) -> NDArray[Any]:
         """
         Reads a single frame (depth) synchronously from the camera.
@@ -319,9 +327,6 @@ class RealSenseCamera(Camera):
         This is a blocking call. It waits for a coherent set of frames (depth)
         from the camera hardware via the RealSense pipeline.
 
-        Args:
-            timeout_ms (int): Maximum time in milliseconds to wait for a frame. Defaults to 200ms.
-
         Returns:
             np.ndarray: The depth map as a NumPy array (height, width)
                   of type `np.uint16` (raw depth values in millimeters) and rotation.
@@ -330,44 +335,50 @@ class RealSenseCamera(Camera):
             DeviceNotConnectedError: If the camera is not connected.
             RuntimeError: If reading frames from the pipeline fails or frames are invalid.
         """
+        if timeout_ms:
+            logger.warning(
+                f"{self} read() timeout_ms parameter is deprecated and will be removed in future versions."
+            )
 
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")
         if not self.use_depth:
             raise RuntimeError(
                 f"Failed to capture depth frame '.read_depth()'. Depth stream is not enabled for {self}."
             )
 
-        start_time = time.perf_counter()
+        if self.thread is None or not self.thread.is_alive():
+            raise RuntimeError(f"{self} read thread is not running.")
 
+        self.new_frame_event.clear()
+
+        _ = self.async_read(timeout_ms=10000)
+
+        with self.frame_lock:
+            depth_map = self.latest_depth_frame
+
+        if depth_map is None:
+            raise RuntimeError("No depth frame available. Ensure camera is streaming.")
+
+        return depth_map
+
+    def _read_from_hardware(self):
         if self.rs_pipeline is None:
             raise RuntimeError(f"{self}: rs_pipeline must be initialized before use.")
 
-        ret, frame = self.rs_pipeline.try_wait_for_frames(timeout_ms=timeout_ms)
+        ret, frame = self.rs_pipeline.try_wait_for_frames(timeout_ms=10000)
 
         if not ret or frame is None:
-            raise RuntimeError(f"{self} read_depth failed (status={ret}).")
+            raise RuntimeError(f"{self} read failed (status={ret}).")
 
-        depth_frame = frame.get_depth_frame()
-        depth_map = np.asanyarray(depth_frame.get_data())
+        return frame
 
-        depth_map_processed = self._postprocess_image(depth_map, depth_frame=True)
-
-        read_duration_ms = (time.perf_counter() - start_time) * 1e3
-        logger.debug(f"{self} read took: {read_duration_ms:.1f}ms")
-
-        return depth_map_processed
-
-    def read(self, color_mode: ColorMode | None = None, timeout_ms: int = 200) -> NDArray[Any]:
+    @check_if_not_connected
+    def read(self, color_mode: ColorMode | None = None, timeout_ms: int = 0) -> NDArray[Any]:
         """
         Reads a single frame (color) synchronously from the camera.
 
         This is a blocking call. It waits for a coherent set of frames (color)
         from the camera hardware via the RealSense pipeline.
 
-        Args:
-            timeout_ms (int): Maximum time in milliseconds to wait for a frame. Defaults to 200ms.
-
         Returns:
             np.ndarray: The captured color frame as a NumPy array
               (height, width, channels), processed according to `color_mode` and rotation.
@@ -378,39 +389,36 @@ class RealSenseCamera(Camera):
             ValueError: If an invalid `color_mode` is requested.
         """
 
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")
-
         start_time = time.perf_counter()
 
-        if self.rs_pipeline is None:
-            raise RuntimeError(f"{self}: rs_pipeline must be initialized before use.")
+        if color_mode is not None:
+            logger.warning(
+                f"{self} read() color_mode parameter is deprecated and will be removed in future versions."
+            )
 
-        ret, frame = self.rs_pipeline.try_wait_for_frames(timeout_ms=timeout_ms)
+        if timeout_ms:
+            logger.warning(
+                f"{self} read() timeout_ms parameter is deprecated and will be removed in future versions."
+            )
 
-        if not ret or frame is None:
-            raise RuntimeError(f"{self} read failed (status={ret}).")
+        if self.thread is None or not self.thread.is_alive():
+            raise RuntimeError(f"{self} read thread is not running.")
 
-        color_frame = frame.get_color_frame()
-        color_image_raw = np.asanyarray(color_frame.get_data())
+        self.new_frame_event.clear()
 
-        color_image_processed = self._postprocess_image(color_image_raw, color_mode)
+        frame = self.async_read(timeout_ms=10000)
 
         read_duration_ms = (time.perf_counter() - start_time) * 1e3
         logger.debug(f"{self} read took: {read_duration_ms:.1f}ms")
 
-        return color_image_processed
+        return frame
 
-    def _postprocess_image(
-        self, image: NDArray[Any], color_mode: ColorMode | None = None, depth_frame: bool = False
-    ) -> NDArray[Any]:
+    def _postprocess_image(self, image: NDArray[Any], depth_frame: bool = False) -> NDArray[Any]:
         """
         Applies color conversion, dimension validation, and rotation to a raw color frame.
 
         Args:
             image (np.ndarray): The raw image frame (expected RGB format from RealSense).
-            color_mode (Optional[ColorMode]): The target color mode (RGB or BGR). If None,
-                                             uses the instance's default `self.color_mode`.
 
         Returns:
             np.ndarray: The processed image frame according to `self.color_mode` and `self.rotation`.
@@ -421,9 +429,9 @@ class RealSenseCamera(Camera):
                           `width` and `height`.
         """
 
-        if color_mode and color_mode not in (ColorMode.RGB, ColorMode.BGR):
+        if self.color_mode and self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
             raise ValueError(
-                f"Invalid requested color mode '{color_mode}'. Expected {ColorMode.RGB} or {ColorMode.BGR}."
+                f"Invalid requested color mode '{self.color_mode}'. Expected {ColorMode.RGB} or {ColorMode.BGR}."
             )
 
         if depth_frame:
@@ -454,7 +462,7 @@ class RealSenseCamera(Camera):
 
         On each iteration:
         1. Reads a color frame with 500ms timeout
-        2. Stores result in latest_frame (thread-safe)
+        2. Stores result in latest_frame and updates timestamp (thread-safe)
         3. Sets new_frame_event to notify listeners
 
         Stops on DeviceNotConnectedError, logs other errors and continues.
@@ -462,25 +470,41 @@ class RealSenseCamera(Camera):
         if self.stop_event is None:
             raise RuntimeError(f"{self}: stop_event is not initialized before starting read loop.")
 
+        failure_count = 0
         while not self.stop_event.is_set():
             try:
-                color_image = self.read(timeout_ms=500)
+                frame = self._read_from_hardware()
+                color_frame_raw = frame.get_color_frame()
+                color_frame = np.asanyarray(color_frame_raw.get_data())
+                processed_color_frame = self._postprocess_image(color_frame)
+
+                if self.use_depth:
+                    depth_frame_raw = frame.get_depth_frame()
+                    depth_frame = np.asanyarray(depth_frame_raw.get_data())
+                    processed_depth_frame = self._postprocess_image(depth_frame, depth_frame=True)
+
+                capture_time = time.perf_counter()
 
                 with self.frame_lock:
-                    self.latest_frame = color_image
+                    self.latest_color_frame = processed_color_frame
+                    if self.use_depth:
+                        self.latest_depth_frame = processed_depth_frame
+                    self.latest_timestamp = capture_time
                 self.new_frame_event.set()
+                failure_count = 0
 
             except DeviceNotConnectedError:
                 break
             except Exception as e:
-                logger.warning(f"Error reading frame in background thread for {self}: {e}")
+                if failure_count <= 10:
+                    failure_count += 1
+                    logger.warning(f"Error reading frame in background thread for {self}: {e}")
+                else:
+                    raise RuntimeError(f"{self} exceeded maximum consecutive read failures.") from 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_read_thread()
 
         self.stop_event = Event()
         self.thread = Thread(target=self._read_loop, args=(), name=f"{self}_read_loop")
@@ -498,7 +522,14 @@ class RealSenseCamera(Camera):
         self.thread = None
         self.stop_event = None
 
+        with self.frame_lock:
+            self.latest_color_frame = None
+            self.latest_depth_frame = None
+            self.latest_timestamp = None
+            self.new_frame_event.clear()
+
     # NOTE(Steven): Missing implementation for depth for now
+    @check_if_not_connected
     def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
         """
         Reads the latest available frame data (color) asynchronously.
@@ -506,6 +537,7 @@ class RealSenseCamera(Camera):
         This method retrieves the most recent color 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.
+        It is “best effort” under high FPS.
 
         Args:
             timeout_ms (float): Maximum time in milliseconds to wait for a frame
@@ -520,21 +552,18 @@ class RealSenseCamera(Camera):
             TimeoutError: If no frame data becomes available within the specified timeout.
             RuntimeError: If the background thread died unexpectedly or another error occurs.
         """
-        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()
+            raise RuntimeError(f"{self} read thread is not running.")
 
         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}."
+                f"Read thread alive: {self.thread.is_alive()}."
             )
 
         with self.frame_lock:
-            frame = self.latest_frame
+            frame = self.latest_color_frame
             self.new_frame_event.clear()
 
         if frame is None:
@@ -542,6 +571,42 @@ class RealSenseCamera(Camera):
 
         return frame
 
+    # NOTE(Steven): Missing implementation for depth for now
+    @check_if_not_connected
+    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
+        """Return the most recent (color) frame captured immediately (Peeking).
+
+        This method is non-blocking and returns whatever is currently in the
+        memory buffer. The frame may be stale,
+        meaning it could have been captured a while ago (hanging camera scenario e.g.).
+
+        Returns:
+            NDArray[Any]: The frame image (numpy array).
+
+        Raises:
+            TimeoutError: If the latest frame is older than `max_age_ms`.
+            DeviceNotConnectedError: If the camera is not connected.
+            RuntimeError: If the camera is connected but has not captured any frames yet.
+        """
+
+        if self.thread is None or not self.thread.is_alive():
+            raise RuntimeError(f"{self} read thread is not running.")
+
+        with self.frame_lock:
+            frame = self.latest_color_frame
+            timestamp = self.latest_timestamp
+
+        if frame is None or timestamp is None:
+            raise RuntimeError(f"{self} has not captured any frames yet.")
+
+        age_ms = (time.perf_counter() - timestamp) * 1e3
+        if age_ms > max_age_ms:
+            raise TimeoutError(
+                f"{self} latest frame is too old: {age_ms:.1f} ms (max allowed: {max_age_ms} ms)."
+            )
+
+        return frame
+
     def disconnect(self) -> None:
         """
         Disconnects from the camera, stops the pipeline, and cleans up resources.
@@ -565,4 +630,10 @@ class RealSenseCamera(Camera):
             self.rs_pipeline = None
             self.rs_profile = None
 
+        with self.frame_lock:
+            self.latest_color_frame = None
+            self.latest_depth_frame = None
+            self.latest_timestamp = None
+            self.new_frame_event.clear()
+
         logger.info(f"{self} disconnected.")

src/lerobot/cameras/realsense/configuration_realsense.py

@@ -60,20 +60,8 @@ class RealSenseCameraConfig(CameraConfig):
     warmup_s: int = 1
 
     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.rotation not in (
-            Cv2Rotation.NO_ROTATION,
-            Cv2Rotation.ROTATE_90,
-            Cv2Rotation.ROTATE_180,
-            Cv2Rotation.ROTATE_270,
-        ):
-            raise ValueError(
-                f"`rotation` is expected to be in {(Cv2Rotation.NO_ROTATION, Cv2Rotation.ROTATE_90, Cv2Rotation.ROTATE_180, Cv2Rotation.ROTATE_270)}, but {self.rotation} is provided."
-            )
+        self.color_mode = ColorMode(self.color_mode)
+        self.rotation = Cv2Rotation(self.rotation)
 
         values = (self.fps, self.width, self.height)
         if any(v is not None for v in values) and any(v is None for v in values):

src/lerobot/cameras/utils.py

@@ -14,7 +14,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import platform
 from typing import cast
 
 from lerobot.utils.import_utils import make_device_from_device_class
@@ -43,6 +42,11 @@ 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))
@@ -63,14 +67,3 @@ def get_cv2_rotation(rotation: Cv2Rotation) -> int | None:
         return int(cv2.ROTATE_90_COUNTERCLOCKWISE)
     else:
         return None
-
-
-def get_cv2_backend() -> int:
-    import cv2
-
-    if platform.system() == "Windows":
-        return int(cv2.CAP_MSMF)  # Use MSMF for Windows instead of AVFOUNDATION
-    # elif platform.system() == "Darwin":  # macOS
-    #     return cv2.CAP_AVFOUNDATION
-    else:  # Linux and others
-        return int(cv2.CAP_ANY)

src/lerobot/cameras/zmq/__init__.py

@@ -0,0 +1,20 @@
+#!/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 .camera_zmq import ZMQCamera
+from .configuration_zmq import ZMQCameraConfig
+
+__all__ = ["ZMQCamera", "ZMQCameraConfig"]

src/lerobot/cameras/zmq/camera_zmq.py

@@ -0,0 +1,385 @@
+#!/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.decorators import check_if_already_connected, check_if_not_connected
+from lerobot.utils.errors import DeviceNotConnectedError
+
+from ..camera import Camera
+from ..configs import ColorMode
+from .configuration_zmq import ZMQCameraConfig
+
+logger = logging.getLogger(__name__)
+
+
+class ZMQCamera(Camera):
+    """
+    Manages camera interactions via ZeroMQ for receiving frames from a remote server.
+
+    This class connects to a ZMQ Publisher, subscribes to frame topics, and decodes
+    incoming JSON messages containing Base64 encoded images. It supports both
+    synchronous and asynchronous frame reading patterns.
+
+    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()
+
+        # Read 1 frame synchronously (blocking)
+        color_image = camera.read()
+
+        # Read 1 frame asynchronously (waits for new frame with a timeout)
+        async_image = camera.async_read()
+
+        # Get the latest frame immediately (no wait, returns timestamp)
+        latest_image, timestamp = camera.read_latest()
+
+        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
+
+        # ZMQ Context and Socket
+        self.context: zmq.Context | None = None
+        self.socket: zmq.Socket | None = None
+        self._connected = False
+
+        # Threading resources
+        self.thread: Thread | None = None
+        self.stop_event: Event | None = None
+        self.frame_lock: Lock = Lock()
+        self.latest_frame: NDArray[Any] | None = None
+        self.latest_timestamp: float | 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:
+        """Checks if the ZMQ socket is initialized and connected."""
+        return self._connected and self.context is not None and self.socket is not None
+
+    @check_if_already_connected
+    def connect(self, warmup: bool = True) -> None:
+        """Connect to ZMQ camera server.
+
+        Args:
+            warmup (bool): If True, waits for the camera to provide at least one
+                           valid frame before returning. Defaults to True.
+        """
+
+        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 not provided
+            if self.width is None or self.height is None:
+                # Read directly from hardware because the thread isn't running yet
+                temp_frame = self._read_from_hardware()
+                h, w = temp_frame.shape[:2]
+                self.height = h
+                self.width = w
+                logger.info(f"{self} resolution detected: {w}x{h}")
+
+            self._start_read_thread()
+            logger.info(f"{self} connected.")
+
+            if warmup:
+                # Ensure we have captured at least one frame via the thread
+                start_time = time.time()
+                while time.time() - start_time < (self.config.warmup_s):  # Wait a bit more than timeout
+                    self.async_read(timeout_ms=self.config.warmup_s * 1000)
+                    time.sleep(0.1)
+
+                with self.frame_lock:
+                    if self.latest_frame is None:
+                        raise ConnectionError(f"{self} failed to capture frames during warmup.")
+
+        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]]:
+        """
+        Detection not implemented for ZMQ cameras. These cameras require manual configuration (server address/port).
+        """
+        raise NotImplementedError("Camera detection is not implemented for ZMQ cameras.")
+
+    def _read_from_hardware(self) -> NDArray[Any]:
+        """
+        Reads a single frame directly from the ZMQ socket.
+        """
+        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:
+            # Check for ZMQ timeout (EAGAIN/Again) without requiring global zmq import
+            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
+
+    @check_if_not_connected
+    def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
+        """
+        Reads a single frame synchronously from the camera.
+
+        This is a blocking call. It waits for the next available frame from the
+        camera background thread.
+
+        Returns:
+            np.ndarray: Decoded frame (height, width, 3)
+        """
+        start_time = time.perf_counter()
+
+        if color_mode is not None:
+            logger.warning(
+                f"{self} read() color_mode parameter is deprecated and will be removed in future versions."
+            )
+
+        if self.thread is None or not self.thread.is_alive():
+            raise RuntimeError(f"{self} read thread is not running.")
+
+        self.new_frame_event.clear()
+        frame = self.async_read(timeout_ms=10000)
+
+        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.
+        """
+        if self.stop_event is None:
+            raise RuntimeError(f"{self}: stop_event is not initialized.")
+
+        failure_count = 0
+        while not self.stop_event.is_set():
+            try:
+                frame = self._read_from_hardware()
+                capture_time = time.perf_counter()
+
+                with self.frame_lock:
+                    self.latest_frame = frame
+                    self.latest_timestamp = capture_time
+                self.new_frame_event.set()
+                failure_count = 0
+
+            except DeviceNotConnectedError:
+                break
+            except (TimeoutError, Exception) as e:
+                if failure_count <= 10:
+                    failure_count += 1
+                    logger.warning(f"Read error: {e}")
+                else:
+                    raise RuntimeError(f"{self} exceeded maximum consecutive read failures.") from e
+
+    def _start_read_thread(self) -> None:
+        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)
+
+        with self.frame_lock:
+            self.latest_frame = None
+            self.latest_timestamp = None
+            self.new_frame_event.clear()
+
+        self.stop_event = Event()
+        self.thread = Thread(target=self._read_loop, daemon=True, name=f"{self}_read_loop")
+        self.thread.start()
+        time.sleep(0.1)
+
+    def _stop_read_thread(self) -> None:
+        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
+
+        with self.frame_lock:
+            self.latest_frame = None
+            self.latest_timestamp = None
+            self.new_frame_event.clear()
+
+    @check_if_not_connected
+    def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
+        """
+        Reads the latest available frame asynchronously.
+
+        Args:
+            timeout_ms (float): Maximum time in milliseconds to wait for a frame
+                to become available. Defaults to 200ms.
+
+        Returns:
+            np.ndarray: The latest captured frame.
+
+        Raises:
+            DeviceNotConnectedError: If the camera is not connected.
+            TimeoutError: If no frame data becomes available within the specified timeout.
+            RuntimeError: If the background thread is not running.
+        """
+
+        if self.thread is None or not self.thread.is_alive():
+            raise RuntimeError(f"{self} read thread is not running.")
+
+        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
+
+    @check_if_not_connected
+    def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
+        """Return the most recent frame captured immediately (Peeking).
+
+        This method is non-blocking and returns whatever is currently in the
+        memory buffer. The frame may be stale,
+        meaning it could have been captured a while ago (hanging camera scenario e.g.).
+
+        Returns:
+            NDArray[Any]: The frame image (numpy array).
+
+        Raises:
+            TimeoutError: If the latest frame is older than `max_age_ms`.
+            DeviceNotConnectedError: If the camera is not connected.
+            RuntimeError: If the camera is connected but has not captured any frames yet.
+        """
+
+        if self.thread is None or not self.thread.is_alive():
+            raise RuntimeError(f"{self} read thread is not running.")
+
+        with self.frame_lock:
+            frame = self.latest_frame
+            timestamp = self.latest_timestamp
+
+        if frame is None or timestamp is None:
+            raise RuntimeError(f"{self} has not captured any frames yet.")
+
+        age_ms = (time.perf_counter() - timestamp) * 1e3
+        if age_ms > max_age_ms:
+            raise TimeoutError(
+                f"{self} latest frame is too old: {age_ms:.1f} ms (max allowed: {max_age_ms} ms)."
+            )
+
+        return frame
+
+    def disconnect(self) -> None:
+        """Disconnect from ZMQ camera."""
+        if not self.is_connected and self.thread is None:
+            raise DeviceNotConnectedError(f"{self} not connected.")
+
+        if self.thread is not None:
+            self._stop_read_thread()
+
+        self._cleanup()
+
+        with self.frame_lock:
+            self.latest_frame = None
+            self.latest_timestamp = None
+            self.new_frame_event.clear()
+
+        logger.info(f"{self} disconnected.")