Merge branch 'feat/add_relative_action_pi_models' into feat/mirror

fix multi gpu processor bug
revert
2026-05-12 15:19:43 +00:00 · 2026-02-22 16:12:10 +01:00 · 2026-02-22 16:11:52 +01:00 · 2026-02-21 18:48:46 +01:00 · 2026-02-21 18:44:35 +01:00 · 2026-02-21 18:19:51 +01:00
179 changed files with 3174 additions and 7235 deletions
@@ -44,7 +44,7 @@ permissions:
 # Sets up the environment variables
 env:
  UV_VERSION: "0.8.0"
-  PYTHON_VERSION: "3.12"
+  PYTHON_VERSION: "3.10"

 # Ensures that only the latest commit for a PR or branch is built, canceling older runs.
 concurrency:
@@ -61,7 +61,6 @@ 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:
@@ -90,11 +89,5 @@ 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
@@ -37,7 +37,7 @@ permissions:
 # Sets up the environment variables
 env:
  UV_VERSION: "0.8.0"
-  PYTHON_VERSION: "3.12"
+  PYTHON_VERSION: "3.10"
  DOCKER_IMAGE_NAME: huggingface/lerobot-gpu

 # Ensures that only the latest action is built, canceling older runs.
@@ -60,7 +60,6 @@ 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:
@@ -88,12 +87,6 @@ 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

@@ -169,7 +162,6 @@ 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: ${{ needs.build-and-push-docker.outputs.image_tag }} # zizmor: ignore[unpinned-images]
      options: --gpus all --shm-size "16gb"
@@ -181,13 +173,6 @@ jobs:
        shell: bash
        working-directory: /lerobot
    steps:
-      - name: Login to Hugging Face
-        if: env.HF_USER_TOKEN != ''
-        run: |
-          hf auth login --token "$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
@@ -28,7 +28,7 @@ on:
 # Sets up the environment variables
 env:
  UV_VERSION: "0.8.0"
-  PYTHON_VERSION: "3.12"
+  PYTHON_VERSION: "3.10"
  DOCKER_IMAGE_NAME_CPU: huggingface/lerobot-cpu:latest
  DOCKER_IMAGE_NAME_GPU: huggingface/lerobot-gpu:latest

@@ -119,7 +119,6 @@ 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: ${{ needs.build-docker-cpu-nightly.outputs.image_tag }} # zizmor: ignore[unpinned-images]
      options: --shm-size "16gb"
@@ -131,11 +130,6 @@ jobs:
        shell: bash
        working-directory: /lerobot
    steps:
-      - name: Login to Hugging Face
-        if: env.HF_USER_TOKEN != ''
-        run: |
-          hf auth login --token "$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
@@ -152,7 +146,6 @@ 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: ${{ needs.build-docker-gpu-nightly.outputs.image_tag }} # zizmor: ignore[unpinned-images]
      options: --gpus all --shm-size "16gb"
@@ -164,11 +157,6 @@ jobs:
        shell: bash
        working-directory: /lerobot
    steps:
-      - name: Login to Hugging Face
-        if: env.HF_USER_TOKEN != ''
-        run: |
-          hf auth login --token "$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
@@ -186,7 +174,6 @@ 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: ${{ needs.build-docker-gpu-nightly.outputs.image_tag }} # zizmor: ignore[unpinned-images]
      options: --gpus all --shm-size "16gb"
@@ -198,15 +185,12 @@ jobs:
        shell: bash
        working-directory: /lerobot
    steps:
-      - name: Login to Hugging Face
-        if: env.HF_USER_TOKEN != ''
-        run: |
-          hf auth login --token "$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
-        run: pytest -vv tests/training/
+      # TODO(Steven): Investigate why motors tests are failing in multi-GPU setup
+        run: pytest tests -vv --maxfail=10 --ignore=tests/motors/
+        timeout-minutes: 10
@@ -50,7 +50,7 @@ jobs:
      - name: Set up Python
        uses: actions/setup-python@v6
        with:
-          python-version: '3.12'
+          python-version: '3.10'

      - name: Run pre-commit hooks
        uses: pre-commit/action@v3.0.1 # zizmor: ignore[unpinned-uses]
@@ -22,7 +22,7 @@ on:
 # Sets up the environment variables
 env:
  UV_VERSION: "0.8.0"
-  PYTHON_VERSION: "3.12"
+  PYTHON_VERSION: "3.10"

 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: '3.12'
+          python-version: '3.10'

      - name: Extract Version
        id: extract_info
@@ -83,6 +83,14 @@ 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

@@ -29,7 +29,7 @@ permissions:
 # Sets up the environment variables
 env:
  UV_VERSION: "0.8.0"
-  PYTHON_VERSION: "3.12"
+  PYTHON_VERSION: "3.10"
  DOCKER_IMAGE_NAME: huggingface/lerobot-gpu:unbound

 # Ensures that only the latest action is built, canceling older runs.
@@ -48,7 +48,6 @@ 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:
@@ -80,11 +79,7 @@ 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

@@ -142,7 +137,6 @@ 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: ${{ needs.build-and-push-docker.outputs.image_tag }} # zizmor: ignore[unpinned-images]
      options: --gpus all --shm-size "16gb"
@@ -154,11 +148,6 @@ jobs:
        shell: bash
        working-directory: /lerobot
    steps:
-      - name: Login to Hugging Face
-        if: env.HF_USER_TOKEN != ''
-        run: |
-          hf auth login --token "$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
@@ -13,7 +13,7 @@
 # limitations under the License.

 default_language_version:
-    python: python3.12
+    python: python3.10

 exclude: "tests/artifacts/.*\\.safetensors$"

@@ -55,7 +55,7 @@ repos:
    rev: v3.21.0
    hooks:
    -   id: pyupgrade
-        args: [--py312-plus]
+        args: [--py310-plus]

  ##### Markdown Quality #####
  - repo: https://github.com/rbubley/mirrors-prettier
@@ -1,25 +0,0 @@
-# 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.
@@ -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) and our [AI policy](./AI_POLICY.md).
+Whichever way you choose to contribute, please be mindful to respect our [code of conduct](./CODE_OF_CONDUCT.md).

 ## Ways to Contribute

@@ -1,3 +1,2 @@
 include src/lerobot/templates/lerobot_modelcard_template.md
 include src/lerobot/datasets/card_template.md
-include src/lerobot/envs/metaworld_config.json
@@ -135,7 +135,7 @@ Learn how to implement your own simulation environment or benchmark and distribu

 ## Citation

-If you use LeRobot in your project, please cite the GitHub repository to acknowledge the ongoing development and contributors:
+If you use LeRobot in your research, please cite:

 ```bibtex
@misc{cadene2024lerobot,
@@ -146,23 +146,6 @@ If you use LeRobot in your project, please cite the GitHub repository to acknowl
 }
 ```

-If you are referencing our research or the academic paper, please also cite our ICLR publication:
-
-<details>
-<summary><b>ICLR 2026 Paper</b></summary>
-
-```bibtex
-@inproceedings{cadenelerobot,
-  title={LeRobot: An Open-Source Library for End-to-End Robot Learning},
-  author={Cadene, Remi and Alibert, Simon and Capuano, Francesco and Aractingi, Michel and Zouitine, Adil and Kooijmans, Pepijn and Choghari, Jade and Russi, Martino and Pascal, Caroline and Palma, Steven and Shukor, Mustafa and Moss, Jess and Soare, Alexander and Aubakirova, Dana and Lhoest, Quentin and Gallou\'edec, Quentin and Wolf, Thomas},
-  booktitle={The Fourteenth International Conference on Learning Representations},
-  year={2026},
-  url={https://arxiv.org/abs/2602.22818}
-}
-```
-
-</details>
-
 ## Contribute

 We welcome contributions from everyone in the community! To get started, please read our [CONTRIBUTING.md](./CONTRIBUTING.md) guide. Whether you're adding a new feature, improving documentation, or fixing a bug, your help and feedback are invaluable. We're incredibly excited about the future of open-source robotics and can't wait to work with you on what's next—thank you for your support!
@@ -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.
- `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.
+- `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.

 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 \
-        lerobot/aloha_mobile_shrimp_image \
+        aliberts/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 \
-        lerobot/aloha_mobile_shrimp_image \
-        lerobot/paris_street \
-        lerobot/kitchen \
+        aliberts/aloha_mobile_shrimp_image \
+        aliberts/paris_street \
+        aliberts/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 \
-        lerobot/aloha_mobile_shrimp_image \
-        lerobot/paris_street \
-        lerobot/kitchen \
+        aliberts/aloha_mobile_shrimp_image \
+        aliberts/paris_street \
+        aliberts/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%    |
-| 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_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_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**  |
+| 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**  |

-|                                   |          | 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%**   |
+|                                    |          | 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%**   |
@@ -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=3.12
+ARG PYTHON_VERSION=3.10

 # Configure environment variables
 ENV DEBIAN_FRONTEND=noninteractive \
@@ -85,8 +85,6 @@ 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 . .
@@ -19,7 +19,7 @@
 # docker run -it --rm lerobot-user

 # Configure the base image
-ARG PYTHON_VERSION=3.12
+ARG PYTHON_VERSION=3.10
 FROM python:${PYTHON_VERSION}-slim

 # Configure environment variables
@@ -29,8 +29,6 @@
    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
@@ -88,8 +88,5 @@ 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
 ```
@@ -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 (cuda, mps, xpu, cpu)
+    --policy_device=mps \ # POLICY: the device to run the policy on, on the server
    --actions_per_chunk=50 \ # POLICY: the number of actions to output at once
    --chunk_size_threshold=0.5 \ # CLIENT: the threshold for the chunk size before sending a new observation to the server
    --aggregate_fn_name=weighted_average \ # CLIENT: the function to aggregate actions on overlapping portions
@@ -32,7 +32,7 @@ version = "0.1.0"
 dependencies = [
    # your policy-specific dependencies
 ]
-requires-python = ">= 3.12"
+requires-python = ">= 3.11"

 [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 Any
+from typing import Dict, 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 Any
+from typing import Dict, Any
 import torch


@@ -13,7 +13,7 @@ The EarthRover Mini Plus is a fully open source mobile robot that connects throu
 ### Hardware

 - EarthRover Mini robot
- Computer with Python 3.12 or newer
+- Computer with Python 3.10 or newer
 - Internet connection

 ### Setting Up the Frodobots SDK
@@ -170,13 +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
-hf auth login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
 ```

 Store your Hugging Face username:

 ```bash
-HF_USER=$(hf auth whoami | awk -F': *' 'NR==1 {print $2}')
+HF_USER=$(huggingface-cli whoami | head -n 1)
 echo $HF_USER
 ```

@@ -185,16 +185,13 @@ echo $HF_USER
 Use the standard recording command:

 ```bash
-lerobot-record \
+python src/lerobot/scripts/lerobot_record.py \
    --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
 ```

@@ -155,10 +155,10 @@ Upload your repository to Hugging Face:
 pip install huggingface_hub

 # Login to Hugging Face
-hf auth login
+huggingface-cli login

 # Create a new repository
-hf repo create my-org/my-custom-env
+huggingface-cli repo create my-custom-env --type space --org my-org

 # Initialize git and push
 git init
@@ -120,12 +120,9 @@ 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" \
-  --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.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.reset_time_s=10
 ```

@@ -224,15 +224,12 @@ lerobot-record \
    --teleop.port=/dev/tty.usbmodem1201 \
    --teleop.id=right \
    --teleop.side=right \
-    --dataset.repo_id=<USER>/hand_record_test_with_video_data \
+    --dataset.repo_id=nepyope/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
 ```

@@ -244,7 +241,7 @@ lerobot-replay \
    --robot.port=/dev/tty.usbmodem58760432281 \
    --robot.id=right \
    --robot.side=right \
-    --dataset.repo_id=<USER>/hand_record_test_with_camera \
+    --dataset.repo_id=nepyope/hand_record_test_with_camera \
    --dataset.episode=0
 ```

@@ -252,13 +249,13 @@ lerobot-replay \

 ```bash
 lerobot-train \
-  --dataset.repo_id=<USER>/hand_record_test_with_video_data \
+  --dataset.repo_id=nepyope/hand_record_test_with_video_data \
  --policy.type=act \
  --output_dir=outputs/train/hopejr_hand \
  --job_name=hopejr \
  --policy.device=mps \
  --wandb.enable=true \
-  --policy.repo_id=<USER>/hand_test_policy
+  --policy.repo_id=nepyope/hand_test_policy
 ```

 ### Evaluate
@@ -273,11 +270,8 @@ 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=<USER>/eval_hopejr \
+  --dataset.repo_id=nepyope/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
 ```
@@ -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
-hf auth login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
 ```

 Then store your Hugging Face repository name in a variable:

 ```bash
-HF_USER=$(hf auth whoami | awk -F': *' 'NR==1 {print $2}')
+HF_USER=$(hf auth whoami | head -n 1)
 echo $HF_USER
 ```

@@ -185,10 +185,7 @@ lerobot-record \
    --display_data=true \
    --dataset.repo_id=${HF_USER}/record-test \
    --dataset.num_episodes=5 \
-    --dataset.single_task="Grab the black cube" \
-    --dataset.streaming_encoding=true \
-    # --dataset.vcodec=auto \
-    --dataset.encoder_threads=2
+    --dataset.single_task="Grab the black cube"
 ```
 </hfoption>
 <hfoption id="API example">
@@ -327,7 +324,7 @@ 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
-hf upload ${HF_USER}/record-test ~/.cache/huggingface/lerobot/{repo-id} --repo-type dataset
+huggingface-cli upload ${HF_USER}/record-test ~/.cache/huggingface/lerobot/{repo-id} --repo-type dataset
 ```

 #### Record function
@@ -491,7 +488,7 @@ 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
-hf upload ${HF_USER}/act_so101_test \
+huggingface-cli upload ${HF_USER}/act_so101_test \
  outputs/train/act_so101_test/checkpoints/last/pretrained_model
 ```

@@ -499,7 +496,7 @@ You can also upload intermediate checkpoints with:

 ```bash
 CKPT=010000
-hf upload ${HF_USER}/act_so101_test${CKPT} \
+huggingface-cli upload ${HF_USER}/act_so101_test${CKPT} \
  outputs/train/act_so101_test/checkpoints/${CKPT}/pretrained_model
 ```

@@ -518,9 +515,6 @@ 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 \
@@ -1,8 +1,8 @@
 # Installation

-This guide uses `conda` (via miniforge) to manage environments (recommended). If you prefer another environment manager (e.g. `uv`, `venv`), ensure you have Python >=3.12 and `ffmpeg` installed with the `libsvtav1` encoder, then skip ahead to [Environment Setup](#step-2-environment-setup).
+This guide uses conda (via miniforge) to manage environments. If you prefer another environment manager (e.g. `uv`, `venv`), ensure you have Python >=3.10 and ffmpeg installed with the `libsvtav1` encoder, then skip ahead to [Install LeRobot](#step-3-install-lerobot-).

-## Step 1 (`conda` only): Install [`miniforge`](https://conda-forge.org/download/)
+## 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"
@@ -11,47 +11,22 @@ bash Miniforge3-$(uname)-$(uname -m).sh

 ## Step 2: Environment Setup

-Create a virtual environment with Python 3.12:
+Create a virtual environment with Python 3.10, using conda:

-<!-- prettier-ignore-start -->
-<hfoptions id="create_venv">
-<hfoption id="conda">
 ```bash
-conda create -y -n lerobot python=3.12
+conda create -y -n lerobot python=3.10
 ```
-</hfoption>
-<hfoption id="uv">
+
+Then activate your conda environment, you have to do this each time you open a shell to use lerobot:
+
 ```bash
-uv python install 3.12
-uv venv --python 3.12
-```
-</hfoption>
-</hfoptions>
-<!-- prettier-ignore-end -->
-
-Then activate your virtual environment, you have to do this each time you open a shell to use lerobot:
-
-<!-- prettier-ignore-start -->
-<hfoptions id="activate_venv">
-<hfoption id="conda">```bash
 conda activate lerobot
-```</hfoption>
-<hfoption id="uv">
-```bash
-# Linux/macOSsource
-source .venv/bin/activate
-# Windows PowerShell
-source .venv\Scripts\Activate.ps1
 ```
-</hfoption>
-</hfoptions>
-<!-- prettier-ignore-end -->

 When using `conda`, install `ffmpeg` in your environment:

 ```bash
 conda install ffmpeg -c conda-forge
-ffmpeg -version  # ffmpeg 8.X is not yet supported !
 ```

 > [!TIP]
@@ -65,16 +40,6 @@ ffmpeg -version  # ffmpeg 8.X is not yet supported !
 >
 > - _[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`.

-> [!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
-> ```
-
-> [!IMPORTANT]
-> If you are using `uv` you will have to install `ffmpeg` system-wide (outside of the virtual environment). You rely on `uv` and `torchcodec` ability to dynamically link to the system `ffmpeg`.
-
 ## Step 3: Install LeRobot 🤗

 ### From Source
@@ -88,45 +53,23 @@ cd lerobot

 Then, install the library in editable mode. This is useful if you plan to contribute to the code.

-<!-- prettier-ignore-start -->
-<hfoptions id="install_lerobot_src">
-<hfoption id="conda">
 ```bash
 pip install -e .
 ```
-</hfoption>
-<hfoption id="uv">
-```bash
-uv pip install -e .
-```
-</hfoption>
-</hfoptions>
-<!-- prettier-ignore-end -->

 ### Installation from PyPI

 **Core Library:**
 Install the base package with:

-<!-- prettier-ignore-start -->
-<hfoptions id="install_lerobot_pypi">
-<hfoption id="conda">
 ```bash
 pip install lerobot
 ```
-</hfoption>
-<hfoption id="uv">
-```bash
-uv pip install lerobot
-```
-</hfoption>
-</hfoptions>
-<!-- prettier-ignore-end -->

 _This installs only the default dependencies._

 **Extra Features:**
-To install additional functionality, use one of the following (If you are using `uv`, replace `pip install` with `uv pip install` in the commands below.):
+To install additional functionality, use one of the following:

 ```bash
 pip install 'lerobot[all]'          # All available features
@@ -140,10 +83,13 @@ _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`.
-To install these for Linux run:
+To install these for linux run:

 ```bash
 sudo apt-get install cmake build-essential python3-dev pkg-config libavformat-dev libavcodec-dev libavdevice-dev libavutil-dev libswscale-dev libswresample-dev libavfilter-dev
@@ -153,7 +99,7 @@ For other systems, see: [Compiling PyAV](https://pyav.org/docs/develop/overview/

 ## Optional dependencies

-LeRobot provides optional extras for specific functionalities. Multiple extras can be combined (e.g., `.[aloha,feetech]`). For all available extras, refer to `pyproject.toml`. If you are using `uv`, replace `pip install` with `uv pip install` in the commands below.
+LeRobot provides optional extras for specific functionalities. Multiple extras can be combined (e.g., `.[aloha,feetech]`). For all available extras, refer to `pyproject.toml`.

 ### Simulations

@@ -279,13 +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
-hf auth login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
 ```

 Then store your Hugging Face repository name in a variable:

 ```bash
-HF_USER=$(hf auth whoami | awk -F': *' 'NR==1 {print $2}')
+HF_USER=$(huggingface-cli whoami | head -n 1)
 echo $HF_USER
 ```

@@ -41,10 +41,7 @@ lerobot-record \
  --display_data=true \
  --dataset.repo_id=${HF_USER}/record-test \
  --dataset.num_episodes=5 \
-  --dataset.single_task="Grab the black cube" \
-  --dataset.streaming_encoding=true \
-  # --dataset.vcodec=auto \
-  --dataset.encoder_threads=2
+  --dataset.single_task="Grab the black cube"
 ```

 See the [recording guide](./il_robots#record-a-dataset) for more details.
@@ -66,13 +66,12 @@ 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 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`.
+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)

 - Run this example to teleoperate:

  ```bash
-  cd examples/phone_to_so100
-  python teleoperate.py
+  python examples/phone_to_so100/teleoperate.py
  ```

 After running the example:
@@ -85,22 +84,19 @@ 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
-  cd examples/phone_to_so100
-  python record.py
+  python examples/phone_to_so100/record.py
  ```

 - Run this example to replay recorded episodes:

  ```bash
-  cd examples/phone_to_so100
-  python replay.py
+  python examples/phone_to_so100/replay.py
  ```

 - Run this example to evaluate a pretrained policy:

  ```bash
-  cd examples/phone_to_so100
-  python evaluate.py
+  python examples/phone_to_so100/evaluate.py
  ```

 ### Important pipeline steps and options
@@ -34,6 +34,11 @@ 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:
@@ -55,7 +60,7 @@ policy.type=pi0
 For training π₀, you can use the standard LeRobot training script with the appropriate configuration:

 ```bash
-lerobot-train \
+python src/lerobot/scripts/lerobot_train.py \
    --dataset.repo_id=your_dataset \
    --policy.type=pi0 \
    --output_dir=./outputs/pi0_training \
@@ -36,6 +36,11 @@ 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:
@@ -51,7 +56,7 @@ policy.type=pi05
 Here's a complete training command for finetuning the base π₀.₅ model on your own dataset:

 ```bash
-lerobot-train \
+python src/lerobot/scripts/lerobot_train.py\
    --dataset.repo_id=your_dataset \
    --policy.type=pi05 \
    --output_dir=./outputs/pi05_training \
@@ -43,11 +43,16 @@ This approach can transform **any existing VLM** into a VLA by training it to pr
   pip install -e ".[pi]"
   ```

+   > [!NOTE]
+   > For lerobot 0.4.0, if you want to install the pi tag, you will have to do: `pip install "lerobot[pi]@git+https://github.com/huggingface/lerobot.git"`.
+   >
+   > This will be solved in the next patch release
+
 ## Training a Custom FAST Tokenizer

 You have two options for the FAST tokenizer:

-1. **Use the pre-trained tokenizer**: The `lerobot/fast-action-tokenizer` tokenizer was trained on 1M+ real robot action sequences and works as a general-purpose tokenizer.
+1. **Use the pre-trained tokenizer**: The `physical-intelligence/fast` tokenizer was trained on 1M+ real robot action sequences and works as a general-purpose tokenizer.

 2. **Train your own tokenizer**: For maximum performance on your specific dataset, you can finetune the tokenizer on your own data.

@@ -109,15 +114,15 @@ lerobot-train \

 ### 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`                         |
+| Parameter                              | Description                                        | Default                      |
+| -------------------------------------- | -------------------------------------------------- | ---------------------------- |
+| `--policy.gradient_checkpointing=true` | Reduces memory usage significantly during training | `false`                      |
+| `--policy.dtype=bfloat16`              | Use mixed precision training for efficiency        | `float32`                    |
+| `--policy.chunk_size`                  | Number of action steps to predict (action horizon) | `50`                         |
+| `--policy.n_action_steps`              | Number of action steps to execute                  | `50`                         |
+| `--policy.max_action_tokens`           | Maximum number of FAST tokens per action chunk     | `256`                        |
+| `--policy.action_tokenizer_name`       | FAST tokenizer to use                              | `physical-intelligence/fast` |
+| `--policy.compile_model=true`          | Enable torch.compile for faster training           | `false`                      |

 ## Inference

@@ -159,9 +159,6 @@ 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
 ```

@@ -201,9 +198,6 @@ 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
 ```

@@ -269,7 +269,7 @@ This generates visualizations showing video frames with subtask boundaries overl
 Train with **no annotations** - uses linear progress from 0 to 1:

 ```bash
-lerobot-train \
+python src/lerobot/scripts/lerobot_train.py \
  --dataset.repo_id=your-username/your-dataset \
  --policy.type=sarm \
  --policy.annotation_mode=single_stage \
@@ -288,7 +288,7 @@ lerobot-train \
 Train with **dense annotations only** (sparse auto-generated):

 ```bash
-lerobot-train \
+python src/lerobot/scripts/lerobot_train.py \
  --dataset.repo_id=your-username/your-dataset \
  --policy.type=sarm \
  --policy.annotation_mode=dense_only \
@@ -307,7 +307,7 @@ lerobot-train \
 Train with **both sparse and dense annotations**:

 ```bash
-lerobot-train \
+python src/lerobot/scripts/lerobot_train.py \
  --dataset.repo_id=your-username/your-dataset \
  --policy.type=sarm \
  --policy.annotation_mode=dual \
@@ -468,7 +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
-lerobot-train \
+python src/lerobot/scripts/lerobot_train.py \
  --dataset.repo_id=your-username/your-dataset \
  --policy.type=pi0 \
  --use_rabc=true \
@@ -106,9 +106,6 @@ 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 \
@@ -1,155 +0,0 @@
-# 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.
@@ -1,49 +1,23 @@
 # Unitree G1

-<img
-  src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/unitree_thumbnail.jpg"
-  alt="Unitree G1 locomanipulation demo"
-  style={{ width: "100%" }}
-/>
+This guide covers the complete setup process for the Unitree G1 humanoid, from initial connection to running gr00t_wbc locomotion.

-The Unitree G1 humanoid is now supported in LeRobot! You can teleoperate, train locomanipulation policies, test in sim, and more. Both 29 and 23 DoF variants are supported.
+## About
+
+We support both 29 and 23 DOF G1 EDU version. We introduce:
+
+- **`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: Getting Started
+## Connection guide

-### Install LeRobot on Your Machine
+### Step 1: Configure Ethernet Interface

-```bash
-conda create -y -n lerobot python=3.12
-conda activate lerobot
-git clone https://github.com/unitreerobotics/unitree_sdk2_python.git
-cd unitree_sdk2_python && pip install -e .
-git clone https://github.com/huggingface/lerobot.git
-cd lerobot
-pip install -e '.[unitree_g1]'
-```
-
-### Test the Installation (Simulation)
-
-```bash
-lerobot-teleoperate \
-  --robot.type=unitree_g1 \
-  --robot.is_simulation=true \
-  --teleop.type=unitree_g1 \
-  --teleop.id=wbc_unitree \
-  --robot.cameras='{"global_view": {"type": "zmq", "server_address": "localhost", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
-  --display_data=true
-```
-
-This will launch a [MuJoCo sim instance](https://huggingface.co/lerobot/unitree-g1-mujoco/tree/main) for the G1.
-
- Press `9` to release the robot
- Press `7` / `8` to increase / decrease waist height
-
-### Connect to the Robot
-
-The G1's Ethernet IP is fixed at `192.168.123.164`. Your machine must have a static IP on the same subnet: `192.168.123.x` where `x ≠ 164`.
+Set a static IP on the same subnet as the robot:

 ```bash
 # Replace 'enp131s0' with your ethernet interface name (check with `ip a`)
@@ -52,200 +26,266 @@ sudo ip addr add 192.168.123.200/24 dev enp131s0
 sudo ip link set enp131s0 up
 ```

-### SSH into the Robot
+**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

 ```bash
 ssh unitree@192.168.123.164
 # Password: 123
 ```

-### Install LeRobot on the G1
-
-From the robot:
-
-```bash
-conda create -y -n lerobot python=3.12
-conda activate lerobot
-git clone https://github.com/unitreerobotics/unitree_sdk2_python.git
-cd unitree_sdk2_python && pip install -e .
-git clone https://github.com/huggingface/lerobot.git
-cd lerobot
-pip install -e '.[unitree_g1]'
-```
-
-> **Note:** The Unitree SDK requires CycloneDDS v0.10.2. See the [Unitree SDK docs](https://github.com/unitreerobotics/unitree_sdk2_python) for details.
+You should now be connected to the G1's Orin.

 ---

 ## Part 2: Enable WiFi on the Robot

-Wi-Fi connectivity is blocked by default on the G1. To activate:
+Wlan0 is disabled by default on the G1. To enable it:
+
+### Step 1: Enable WiFi Hardware

 ```bash
+sudo rfkill unblock wifi
 sudo rfkill unblock all
+
+# Bring up wlan0
 sudo ip link set wlan0 up
+
+# Enable NetworkManager control of wlan0
 sudo nmcli radio wifi on
 sudo nmcli device set wlan0 managed yes
 sudo systemctl restart NetworkManager
 ```

-**On your laptop** (share internet via Ethernet):
+### Step 2: Enable Internet Forwarding
+
+**On your laptop:**

 ```bash
+# Enable IP forwarding
 sudo sysctl -w net.ipv4.ip_forward=1

-# Replace wlp132s0f0 with your WiFi interface name
+# Set up NAT (replace wlp132s0f0 with your WiFi interface)
 sudo iptables -t nat -A POSTROUTING -o wlp132s0f0 -s 192.168.123.0/24 -j MASQUERADE
 sudo iptables -A FORWARD -i wlp132s0f0 -o enp131s0 -m state --state RELATED,ESTABLISHED -j ACCEPT
 sudo iptables -A FORWARD -i enp131s0 -o wlp132s0f0 -j ACCEPT
 ```

-**On the G1** (set default route through your laptop):
+**On the G1:**

 ```bash
+# Add laptop as default gateway
 sudo ip route del default 2>/dev/null || true
 sudo ip route add default via 192.168.123.200 dev eth0
 echo "nameserver 8.8.8.8" | sudo tee /etc/resolv.conf

-# Verify
+# Test connection
 ping -c 3 8.8.8.8
 ```

-**Connect to a WiFi network:**
+### Step 3: Connect to WiFi Network

 ```bash
+# List available networks
 nmcli device wifi list

+# Connect to your WiFi (example)
 sudo nmcli connection add type wifi ifname wlan0 con-name "YourNetwork" ssid "YourNetwork"
 sudo nmcli connection modify "YourNetwork" wifi-sec.key-mgmt wpa-psk
 sudo nmcli connection modify "YourNetwork" wifi-sec.psk "YourPassword"
 sudo nmcli connection modify "YourNetwork" connection.autoconnect yes
 sudo nmcli connection up "YourNetwork"

+# Check WiFi IP address
 ip a show wlan0
 ```

-You can now SSH over WiFi:
+### Step 4: SSH Over WiFi
+
+Once connected to WiFi, note the robot's IP address and disconnect the Ethernet cable. You can now SSH over WiFi:

 ```bash
-ssh unitree@<ROBOT_WIFI_IP>
+ssh unitree@<YOUR_ROBOT_IP>
 # Password: 123
 ```

+Replace `<YOUR_ROBOT_IP>` with your robot's actual WiFi IP address.
+
 ---

-## Part 3: Teleoperation & Locomotion
+## Part 3: Robot Server Setup

-### Run the Robot Server
+### Step 1: Install LeRobot on the Orin
+
+SSH into the robot and install LeRobot:
+
+```bash
+ssh unitree@<YOUR_ROBOT_IP>
+
+conda create -y -n lerobot python=3.10
+conda activate lerobot
+git clone https://github.com/huggingface/lerobot.git
+cd lerobot
+pip install -e '.[unitree_g1]'
+git clone https://github.com/unitreerobotics/unitree_sdk2_python.git
+cd unitree_sdk2_python  && pip install -e .
+```
+
+**Note**: The Unitree SDK requires CycloneDDS v0.10.2 to be installed. See the [Unitree SDK documentation](https://github.com/unitreerobotics/unitree_sdk2_python) for details.
+
+### Step 2: Run the Robot Server

 On the robot:

 ```bash
-python src/lerobot/robots/unitree_g1/run_g1_server.py --camera
+python src/lerobot/robots/unitree_g1/run_g1_server.py
 ```

-### Run the Locomotion Policy
-
-```bash
-lerobot-teleoperate \
-  --robot.type=unitree_g1 \
-  --robot.is_simulation=false \
-  --robot.robot_ip=<ROBOT_IP> \
-  --teleop.type=unitree_g1 \
-  --teleop.id=wbc_unitree \
-  --robot.cameras='{"global_view": {"type": "zmq", "server_address": "<ROBOT_IP>", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
-  --display_data=true \
-  --robot.controller=HolosomaLocomotionController
-```
-
-We support both [HolosomaLocomotionController](https://github.com/amazon-far/holosoma) and [GrootLocomotionController](https://github.com/NVlabs/GR00T-WholeBodyControl).
+**Important**: Keep this terminal running. The server must be active for remote control.

 ---

-## Part 4: Loco-Manipulation with the Homunculus Exoskeleton
+## Part 4: Controlling the robot

-We provide a loco-manipulation solution via the Homunculus Exoskeleton — an open-source 7 DoF exoskeleton for whole-body control. Assembly instructions [here](https://github.com/nepyope/hmc_exo).
+With the robot server running, you can now control the robot remotely. Let's launch a locomotion policy

-### Calibrate
+### Step 1: Install LeRobot on your machine
+
+```bash
+conda create -y -n lerobot python=3.10
+conda activate lerobot
+git clone https://github.com/huggingface/lerobot.git
+cd lerobot
+pip install -e '.[unitree_g1]'
+git clone https://github.com/unitreerobotics/unitree_sdk2_python.git
+cd unitree_sdk2_python  && pip install -e .
+```
+
+### Step 2: Update Robot IP in Config
+
+Edit the config file to match your robot's WiFi IP:
+
+```python
+# In src/lerobot/robots/unitree_g1/config_unitree_g1.py
+robot_ip: str = "<YOUR_ROBOT_IP>"  # Replace with your robot's WiFi IP.
+```
+
+### 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
+
+```
+
+Press `Ctrl+C` to stop the policy.
+
+---
+
+## Running in Simulation Mode (MuJoCo)
+
+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
+    --teleop.type=unitree_g1 \
+    --teleop.left_arm_config.port=/dev/ttyACM1 \
+    --teleop.right_arm_config.port=/dev/ttyACM0 \
+    --teleop.id=exo
 ```

-During calibration move each joint through its entire range. After fitting, move the joint in a neutral position and press `n` to advance.
-
-### Record a Dataset
+### Teleoperate 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.encoder_threads=2
+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
 ```

-> **Note:** Omit `--teleop.left_arm_config.port` and `--teleop.right_arm_config.port` if you're only using the joystick.
+### Record Dataset in Simulation

-Example dataset: [nepyope/unitree_box_move_blue_full](https://huggingface.co/datasets/nepyope/unitree_box_move_blue_full)
+```bash
+python -m lerobot.scripts.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
+```
+
+Example simulation dataset: [nepyope/teleop_test_sim](https://huggingface.co/datasets/nepyope/teleop_test_sim)

 ---

-## Part 5: Training & Inference
+## Running on Real Robot

-### Train
+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/scripts/lerobot_train.py \
-  --dataset.repo_id=your-username/dataset-name  \
-  --policy.type=pi05 \
-  --output_dir=./outputs/pi05_training \
-  --job_name=pi05_training \
-  --policy.repo_id=your-username/your-repo-id \
-  --policy.pretrained_path=lerobot/pi05_base \
-  --policy.compile_model=true \
-  --policy.gradient_checkpointing=true \
-  --wandb.enable=true \
-  --policy.dtype=bfloat16 \
-  --policy.freeze_vision_encoder=false \
-  --policy.train_expert_only=false \
-  --steps=3000 \
-  --policy.device=cuda \
-  --batch_size=32
+python src/lerobot/cameras/zmq/image_server.py
 ```

-### Inference with RTC
+Keep this running in a separate terminal for camera streaming during recording.

-Once trained, we recommend deploying policies using inference-time RTC:
+### Teleoperate Real Robot

 ```bash
-python examples/rtc/eval_with_real_robot.py \
-  --policy.path=your-username/your-repo-id \
-  --policy.device=cuda \
-  --robot.type=unitree_g1 \
-  --robot.is_simulation=false \
-  --robot.controller=HolosomaLocomotionController \
-  --robot.cameras='{"global_view": {"type": "zmq", "server_address": "<ROBOT_IP>", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
-  --task="task_description" \
-  --duration=1000 \
-  --fps=30 \
-  --rtc.enabled=true
+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
+python -m lerobot.scripts.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
+```
+
+**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
@@ -254,8 +294,8 @@ python examples/rtc/eval_with_real_robot.py \
 - [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)
+- [Unitree_IL_Lerobot](https://github.com/unitreerobotics/unitree_IL_lerobot)

 ---

-_Last updated: March 2026_
+_Last updated: December 2025_
@@ -12,7 +12,6 @@ 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`.
@@ -157,30 +156,6 @@ 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:
@@ -45,7 +45,7 @@ policy.type=wall_x
 For training WallX, you can use the standard LeRobot training script with the appropriate configuration:

 ```bash
-lerobot-train \
+python src/lerobot/scripts/lerobot_train.py \
    --dataset.repo_id=your_dataset \
    --policy.type=wall_x \
    --output_dir=./outputs/wallx_training \
@@ -154,7 +154,7 @@ lerobot-train \

 ```bash
 lerobot-train \
-  --dataset.repo_id=<USER>/bimanual-so100-handover-cube \
+  --dataset.repo_id=pepijn223/bimanual-so100-handover-cube \
  --output_dir=./outputs/xvla_bimanual \
  --job_name=xvla_so101_training \
  --policy.path="lerobot/xvla-base" \
@@ -22,7 +22,7 @@ lerobot-replay \
    --robot.type=so100_follower \
    --robot.port=/dev/tty.usbmodem58760431541 \
    --robot.id=black \
-    --dataset.repo_id=<USER>/record-test \
+    --dataset.repo_id=aliberts/record-test \
    --dataset.episode=2
 ```
 """
@@ -57,7 +57,7 @@ class DatasetReplayConfig:
    repo_id: str
    # Episode to replay.
    episode: int
-    # Root directory where the dataset will be stored (e.g. 'dataset/path'). If None, defaults to $HF_LEROBOT_HOME/repo_id.
+    # Root directory where the dataset will be stored (e.g. 'dataset/path').
    root: str | Path | None = None
    # Limit the frames per second. By default, uses the policy fps.
    fps: int = 30
@@ -1,490 +0,0 @@
-#!/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()
@@ -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=<USER>/smolvla_check_rtc_last3 \
-        --dataset.repo_id=<USER>/check_rtc \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --dataset.repo_id=helper2424/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=<USER>/reuben_pi0 \
-        --dataset.repo_id=<USER>/so101_cube_in_cup \
+        --policy.path=lipsop/reuben_pi0 \
+        --dataset.repo_id=ReubenLim/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=<USER>/smolvla_check_rtc_last3 \
-        --dataset.repo_id=<USER>/check_rtc \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --dataset.repo_id=helper2424/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=<USER>/smolvla_check_rtc_last3 \
-        --dataset.repo_id=<USER>/check_rtc \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --dataset.repo_id=helper2424/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=<USER>/smolvla_check_rtc_last3 \
-        --dataset.repo_id=<USER>/check_rtc \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --dataset.repo_id=helper2424/check_rtc \
        --use_torch_compile=true \
        --torch_compile_backend=inductor \
        --torch_compile_mode=max-autotune \
@@ -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=<USER>/smolvla_check_rtc_last3 \
+        --policy.path=helper2424/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=<USER>/smolvla_check_rtc_last3 \
+        --policy.path=helper2424/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=<USER>/pi05_check_rtc \
+        --policy.path=helper2424/pi05_check_rtc \
        --policy.device=mps \
        --rtc.enabled=true \
        --rtc.execution_horizon=20 \
@@ -78,7 +78,6 @@ from torch import Tensor

 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig  # noqa: F401
 from lerobot.cameras.realsense.configuration_realsense import RealSenseCameraConfig  # noqa: F401
-from lerobot.cameras.zmq.configuration_zmq import ZMQCameraConfig  # noqa: F401
 from lerobot.configs import parser
 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.configs.types import RTCAttentionSchedule
@@ -98,7 +97,6 @@ from lerobot.robots import (  # noqa: F401
    bi_so_follower,
    koch_follower,
    so_follower,
-    unitree_g1,
 )
 from lerobot.robots.utils import make_robot_from_config
 from lerobot.utils.constants import OBS_IMAGES
@@ -14,20 +14,20 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.

+import argparse
 import logging
+import time
 from collections import deque

 import numpy as np
 import onnxruntime as ort
 from huggingface_hub import hf_hub_download

-from lerobot.robots.unitree_g1.g1_utils import (
-    REMOTE_AXES,
-    REMOTE_BUTTONS,
-    G1_29_JointIndex,
-    get_gravity_orientation,
-)
+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__)


@@ -36,13 +36,18 @@ 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"
+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: float = 0.25
 DOF_POS_SCALE: float = 1.0
 DOF_VEL_SCALE: float = 0.05
-CMD_SCALE: list[float] = [2.0, 2.0, 0.25]
+CMD_SCALE: list = [2.0, 2.0, 0.25]


 DEFAULT_GROOT_REPO_ID = "nepyope/GR00T-WholeBodyControl_g1"
@@ -80,11 +85,11 @@ def load_groot_policies(
 class GrootLocomotionController:
    """GR00T lower-body locomotion controller for the Unitree G1."""

-    control_dt = CONTROL_DT  # Expose for unitree_g1.py
-
-    def __init__(self):
-        # Load policies
-        self.policy_balance, self.policy_walk = load_groot_policies()
+    def __init__(self, policy_balance, policy_walk, robot, config):
+        self.policy_balance = policy_balance
+        self.policy_walk = policy_walk
+        self.robot = robot
+        self.config = config

        self.cmd = np.array([0.0, 0.0, 0.0], dtype=np.float32)  # vx, vy, theta_dot

@@ -104,60 +109,45 @@ class GrootLocomotionController:

        logger.info("GrootLocomotionController initialized")

-    def reset(self) -> None:
-        """Reset internal state for a new episode."""
-        self.cmd[:] = 0.0
-        self.groot_qj_all[:] = 0.0
-        self.groot_dqj_all[:] = 0.0
-        self.groot_action[:] = 0.0
-        self.groot_obs_single[:] = 0.0
-        self.groot_obs_stacked[:] = 0.0
-        self.groot_height_cmd = 0.74
-        self.groot_orientation_cmd[:] = 0.0
-        self.groot_obs_history.clear()
-        for _ in range(6):
-            self.groot_obs_history.append(np.zeros(86, dtype=np.float32))
+    def run_step(self):
+        # Get current observation
+        obs = self.robot.get_observation()

-    def run_step(self, action: dict, lowstate) -> dict:
-        """Run one step of the locomotion controller.
+        if not obs:
+            return

-        Args:
-            action: Action dict containing remote.lx/ly/rx/ry and buttons
-            lowstate: Robot lowstate containing motor positions/velocities and IMU
-
-        Returns:
-            Action dict for lower body joints (0-14)
-        """
-        if lowstate is None:
-            return {}
-
-        buttons = [int(action.get(k, 0)) for k in REMOTE_BUTTONS]
-        if buttons[0]:  # R1 - raise waist
+        # 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 buttons[4]:  # R2 - lower waist
+        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)

-        lx, ly, rx, _ry = (action.get(k, 0.0) for k in REMOTE_AXES)
-        self.cmd[0] = ly  # Forward/backward
-        self.cmd[1] = -lx  # Left/right (negated)
-        self.cmd[2] = -rx  # Rotation rate (negated)
+        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

-        # Get joint positions and velocities from lowstate
+        # 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] = lowstate.motor_state[idx].q
-            self.groot_dqj_all[idx] = lowstate.motor_state[idx].dq
+            self.groot_qj_all[idx] = obs[f"{name}.q"]
+            self.groot_dqj_all[idx] = obs[f"{name}.dq"]
+
+        # Adapt observation for g1_23dof
+        for idx in MISSING_JOINTS:
+            self.groot_qj_all[idx] = 0.0
+            self.groot_dqj_all[idx] = 0.0

        # Scale joint positions and velocities
        qj_obs = self.groot_qj_all.copy()
        dqj_obs = self.groot_dqj_all.copy()

        # Express IMU data in gravity frame of reference
-        quat = lowstate.imu_state.quaternion
-        ang_vel = np.array(lowstate.imu_state.gyroscope, dtype=np.float32)
-        gravity_orientation = get_gravity_orientation(quat)
+        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
        qj_obs = (qj_obs - GROOT_DEFAULT_ANGLES) * DOF_POS_SCALE
@@ -196,10 +186,73 @@ class GrootLocomotionController:
        # Transform action back to target joint positions
        target_dof_pos_15 = GROOT_DEFAULT_ANGLES[:15] + self.groot_action * ACTION_SCALE

-        # Build action dict
+        # Build action dict (only first 15 joints for GR00T)
        action_dict = {}
        for i in range(15):
            motor_name = G1_29_JointIndex(i).name
            action_dict[f"{motor_name}.q"] = float(target_dof_pos_15[i])

-        return action_dict
+        # 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_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()
+            groot_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="GR00T Locomotion Controller for Unitree G1")
+    parser.add_argument(
+        "--repo-id",
+        type=str,
+        default=DEFAULT_GROOT_REPO_ID,
+        help=f"Hugging Face Hub repo ID for GR00T policies (default: {DEFAULT_GROOT_REPO_ID})",
+    )
+    args = parser.parse_args()
+
+    run(repo_id=args.repo_id)
@@ -14,21 +14,21 @@
 # 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.g1_utils import (
-    REMOTE_AXES,
-    G1_29_JointArmIndex,
-    G1_29_JointIndex,
-    get_gravity_orientation,
-)
+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)
@@ -40,13 +40,18 @@ 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 = 0.005  # 200Hz
+CONTROL_DT = 0.02  # 50Hz
 ANG_VEL_SCALE = 0.25
 DOF_POS_SCALE = 1.0
 DOF_VEL_SCALE = 0.05
-GAIT_PERIOD = 0.5
+GAIT_PERIOD = 1.0


 DEFAULT_HOLOSOMA_REPO_ID = "nepyope/holosoma_locomotion"
@@ -82,7 +87,7 @@ def load_policy(
    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, load_external_data=False)
+    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:
@@ -96,13 +101,15 @@ def load_policy(


 class HolosomaLocomotionController:
-    """Holosoma lower-body locomotion controller for Unitree G1."""
+    """Holosoma whole-body locomotion controller for Unitree G1."""

-    control_dt = CONTROL_DT  # Expose for unitree_g1.py
+    def __init__(self, policy, robot, kp: np.ndarray, kd: np.ndarray):
+        self.policy = policy
+        self.robot = robot

-    def __init__(self):
-        # Load policy and gains
-        self.policy, self.kp, self.kd = load_policy()
+        # Override robot's PD gains with policy gains
+        self.robot.kp = kp
+        self.robot.kd = kd

        self.cmd = np.zeros(3, dtype=np.float32)

@@ -117,55 +124,35 @@ class HolosomaLocomotionController:
        self.phase_dt = 2 * np.pi / ((1.0 / CONTROL_DT) * GAIT_PERIOD)
        self.is_standing = True

-        logger.info("HolosomaLocomotionController initialized")
+    def run_step(self):
+        # Get current observation
+        obs = self.robot.get_observation()

-    def reset(self) -> None:
-        """Reset internal state for a new episode."""
-        self.cmd[:] = 0.0
-        self.qj[:] = 0.0
-        self.dqj[:] = 0.0
-        self.obs[:] = 0.0
-        self.last_action[:] = 0.0
-        self.phase = np.array([[0.0, np.pi]], dtype=np.float32)
-        self.is_standing = True
+        if not obs:
+            return

-    def run_step(self, action: dict, lowstate) -> dict:
-        """Run one step of the locomotion controller.
-
-        Args:
-            action: Action dict containing remote.lx/ly/rx/ry
-            lowstate: Robot lowstate containing motor positions/velocities and IMU
-
-        Returns:
-            Action dict for lower body joints (0-14)
-        """
-        if lowstate is None:
-            return {}
-
-        lx, ly, rx, _ry = (action.get(k, 0.0) for k in REMOTE_AXES)
-        ly = ly if abs(ly) > 0.1 else 0.0
-        lx = lx if abs(lx) > 0.1 else 0.0
-        rx = rx if abs(rx) > 0.1 else 0.0
-        ly = np.clip(ly, -0.3, 0.3)
-        lx = np.clip(lx, -0.3, 0.3)
+        # 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 from lowstate
+        # Get joint positions and velocities
        for motor in G1_29_JointIndex:
+            name = motor.name
            idx = motor.value
-            self.qj[idx] = lowstate.motor_state[idx].q
-            self.dqj[idx] = lowstate.motor_state[idx].dq
+            self.qj[idx] = obs[f"{name}.q"]
+            self.dqj[idx] = obs[f"{name}.dq"]

-        # Hide arm positions from policy (show DEFAULT_ANGLES instead)
-        # This prevents policy from reacting to teleop arm movements
-        for arm_joint in G1_29_JointArmIndex:
-            self.qj[arm_joint.value] = DEFAULT_ANGLES[arm_joint.value]
-            self.dqj[arm_joint.value] = 0.0
+        # 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 = lowstate.imu_state.quaternion
-        ang_vel = np.array(lowstate.imu_state.gyroscope, dtype=np.float32)
-        gravity = get_gravity_orientation(quat)
+        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
@@ -199,16 +186,79 @@ class HolosomaLocomotionController:
        # 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()
-        policy_action = np.clip(raw_action, -100.0, 100.0)
-        self.last_action = policy_action.copy()
+        action = np.clip(raw_action, -100.0, 100.0)
+        self.last_action = action.copy()

        # Transform action back to target joint positions
-        target = DEFAULT_ANGLES + policy_action * ACTION_SCALE
+        target = DEFAULT_ANGLES + action * ACTION_SCALE

-        # Build action dict (first 15 joints only)
+        # Build action dict
        action_dict = {}
-        for i in range(15):
-            motor_name = G1_29_JointIndex(i).name
-            action_dict[f"{motor_name}.q"] = float(target[i])
+        for motor in G1_29_JointIndex:
+            action_dict[f"{motor.name}.q"] = float(target[motor.value])

-        return action_dict
+        # 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)
@@ -25,11 +25,11 @@ discord = "https://discord.gg/s3KuuzsPFb"

 [project]
 name = "lerobot"
-version = "0.5.0"
+version = "0.4.4"
 description = "🤗 LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch"
 dynamic = ["readme"]
 license = { text = "Apache-2.0" }
-requires-python = ">=3.12"
+requires-python = ">=3.10"
 authors = [
    { name = "Rémi Cadène", email = "re.cadene@gmail.com" },
    { name = "Simon Alibert", email = "alibert.sim@gmail.com" },
@@ -50,8 +50,7 @@ classifiers = [
    "Intended Audience :: Education",
    "Intended Audience :: Science/Research",
    "License :: OSI Approved :: Apache Software License",
-    "Programming Language :: Python :: 3.12",
-    "Programming Language :: Python :: 3.13",
+    "Programming Language :: Python :: 3.10",
    "Topic :: Software Development :: Build Tools",
    "Topic :: Scientific/Engineering :: Artificial Intelligence",
 ]
@@ -60,30 +59,28 @@ keywords = ["lerobot", "huggingface", "robotics",  "machine learning", "artifici
 dependencies = [

    # Hugging Face dependencies
-    "datasets>=4.0.0,<5.0.0",
+    "datasets>=4.0.0,<4.2.0",
    "diffusers>=0.27.2,<0.36.0",
-    "huggingface-hub>=1.0.0,<2.0.0",
+    "huggingface-hub[hf-transfer,cli]>=0.34.2,<0.36.0",
    "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",
-    "pynput>=1.7.8,<1.9.0",
+    "packaging>=24.2,<26.0",
+    "pynput>=1.7.7,<1.9.0",
    "pyserial>=3.5,<4.0",
-
    "wandb>=0.24.0,<0.25.0",
-    "draccus==0.10.0", # TODO: Relax version constraint
+
+    "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 ==
    "gymnasium>=1.1.1,<2.0.0",
    "rerun-sdk>=0.24.0,<0.27.0",

@@ -98,20 +95,14 @@ dependencies = [

 # Common
 pygame-dep = ["pygame>=2.5.1,<2.7.0"]
-placo-dep = ["placo>=0.9.6,<0.9.17"]
-transformers-dep = ["transformers>=5.3.0,<6.0.0"]
+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>=6.31.1,<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]"]
+damiao = ["python-can>=4.2.0,<5.0.0"]

 # Robots
 openarms = ["lerobot[damiao]"]
@@ -119,36 +110,34 @@ 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 = [
-    "unitree-sdk2==1.0.1",
    "pyzmq>=26.2.1,<28.0.0",
    "onnxruntime>=1.16.0,<2.0.0",
    "pin>=3.0.0,<4.0.0",
    "meshcat>=0.3.0,<0.4.0",
-    "lerobot[matplotlib-dep]",
-    "lerobot[pygame-dep]",
+    "matplotlib>=3.9.0,<4.0.0",
    "casadi>=3.6.0,<4.0.0",
 ]
 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,<2.57.0 ; sys_platform == 'darwin'",
+    "pyrealsense2-macosx>=2.54,<2.55.0 ; sys_platform == 'darwin'",
 ]
-phone = ["hebi-py>=2.8.0,<2.12.0", "teleop>=0.1.0,<0.2.0", "fastapi<1.0", "lerobot[scipy-dep]"]
+phone = ["hebi-py>=2.8.0,<2.12.0", "teleop>=0.1.0,<0.2.0", "fastapi<1.0"]

 # Policies
 wallx = [
-    "lerobot[transformers-dep]",
-    "lerobot[peft]",
-    "lerobot[scipy-dep]",
-    "torchdiffeq>=0.2.4,<0.3.0",
-    "lerobot[qwen-vl-utils-dep]",
+    "transformers==4.49.0",
+    "peft==0.17.1",
+    "scipy==1.15.3",
+    "torchdiffeq==0.2.5",
+    "qwen_vl_utils==0.0.11"
 ]
-pi = ["lerobot[transformers-dep]", "lerobot[scipy-dep]"]
+pi = ["transformers @ git+https://github.com/huggingface/transformers.git@fix/lerobot_openpi", "scipy>=1.10.1,<1.15"]
 smolvla = ["lerobot[transformers-dep]", "num2words>=0.5.14,<0.6.0", "accelerate>=1.7.0,<2.0.0", "safetensors>=0.4.3,<1.0.0"]
 groot = [
    "lerobot[transformers-dep]",
-    "lerobot[peft]",
+    "peft>=0.13.0,<1.0.0",
    "dm-tree>=0.1.8,<1.0.0",
    "timm>=1.0.0,<1.1.0",
    "safetensors>=0.4.3,<1.0.0",
@@ -157,13 +146,13 @@ 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", "lerobot[matplotlib-dep]", "lerobot[qwen-vl-utils-dep]"]
+sarm = ["lerobot[transformers-dep]", "faker>=33.0.0,<35.0.0", "matplotlib>=3.10.3,<4.0.0", "qwen-vl-utils>=0.0.14,<0.1.0"]
 xvla = ["lerobot[transformers-dep]"]
 hilserl = ["lerobot[transformers-dep]", "gym-hil>=0.1.13,<0.2.0", "lerobot[grpcio-dep]", "lerobot[placo-dep]"]

 # Features
-async = ["lerobot[grpcio-dep]", "lerobot[matplotlib-dep]"]
-peft = ["lerobot[transformers-dep]", "lerobot[peft-dep]"]
+async = ["lerobot[grpcio-dep]", "matplotlib>=3.10.3,<4.0.0"]
+peft = ["lerobot[transformers-dep]", "peft>=0.18.0,<1.0.0"]

 # Development
 dev = ["pre-commit>=3.7.0,<5.0.0", "debugpy>=1.8.1,<1.9.0", "lerobot[grpcio-dep]", "grpcio-tools==1.73.1", "mypy>=1.19.1"]
@@ -171,19 +160,13 @@ 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
-# NOTE: Explicitly listing scipy helps flatten the dependecy tree.
-aloha = ["gym-aloha>=0.1.2,<0.2.0", "lerobot[scipy-dep]"]
+aloha = ["gym-aloha>=0.1.2,<0.2.0"]
 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; sys_platform == 'linux'", "lerobot[scipy-dep]"]
-metaworld = ["metaworld==3.0.0", "lerobot[scipy-dep]"]
+libero = ["lerobot[transformers-dep]", "hf-libero>=0.1.3,<0.2.0"]
+metaworld = ["metaworld==3.0.0"]

 # 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]",
@@ -191,8 +174,8 @@ all = [
    "lerobot[reachy2]",
    "lerobot[kinematics]",
    "lerobot[intelrealsense]",
-    "lerobot[wallx]",
-    "lerobot[pi]",
+    # "lerobot[wallx]",
+    # "lerobot[pi]", TODO(Pepijn): Update pi to transformers v5
    "lerobot[smolvla]",
    # "lerobot[groot]", TODO(Steven): Gr00t requires specific installation instructions for flash-attn
    "lerobot[xvla]",
@@ -204,11 +187,10 @@ all = [
    "lerobot[aloha]",
    "lerobot[pusht]",
    "lerobot[phone]",
-    "lerobot[libero]; sys_platform == 'linux'",
+    "lerobot[libero]",
    "lerobot[metaworld]",
    "lerobot[sarm]",
    "lerobot[peft]",
-    # "lerobot[unitree_g1]", TODO: Unitree requires specific installation instructions for unitree_sdk2
 ]

 [project.scripts]
@@ -230,14 +212,11 @@ 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 = "py312"
+target-version = "py310"
 line-length = 110
 exclude = ["tests/artifacts/**/*.safetensors", "*_pb2.py", "*_pb2_grpc.py"]

@@ -329,7 +308,7 @@ default.extend-ignore-identifiers-re = [
 # Uncomment [tool.mypy] first, then uncomment individual module overrides as they get proper type annotations

 [tool.mypy]
-python_version = "3.12"
+python_version = "3.10"
 ignore_missing_imports = true
 follow_imports = "skip"
 # warn_return_any = true
@@ -413,3 +392,85 @@ 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 = "peft" },
+    ],
+    [
+        { 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 = "peft" },
+    ],
+    [
+        { extra = "pi" },
+        { extra = "all" },
+    ],
+]
@@ -1,73 +1,76 @@
 #
-# This file is autogenerated by pip-compile with Python 3.12
+# This file is autogenerated by pip-compile with Python 3.10
 # by the following command:
 #
 #    pip-compile --output-file=requirements-macos.txt requirements.in
 #
 -e .[all]
    # via -[all]
-absl-py==2.4.0
+absl-py==2.3.1
    # via
    #   dm-control
    #   dm-env
    #   dm-tree
    #   labmaze
    #   mujoco
-accelerate==1.13.0
+    #   tensorboard
+accelerate==1.11.0
    # via
    #   lerobot
    #   peft
 aiohappyeyeballs==2.6.1
    # via aiohttp
-aiohttp==3.13.3
+aiohttp==3.13.1
    # via fsspec
 aiosignal==1.4.0
    # via aiohttp
-annotated-doc==0.0.4
-    # via
-    #   fastapi
-    #   typer
 annotated-types==0.7.0
    # via pydantic
-anyio==4.12.1
+antlr4-python3-runtime==4.9.3
+    # via
+    #   hydra-core
+    #   omegaconf
+anyio==4.11.0
    # via
-    #   httpx
    #   starlette
    #   watchfiles
-asttokens==3.0.1
+asttokens==3.0.0
    # via stack-data
+async-timeout==5.0.1
+    # via aiohttp
 attrs==25.4.0
    # via
    #   aiohttp
    #   dm-tree
    #   jsonlines
+    #   jsonschema
+    #   referencing
    #   rerun-sdk
 av==15.1.0
+    # via lerobot
+bddl==1.0.1
+    # via libero
+certifi==2025.10.5
    # via
-    #   lerobot
-    #   qwen-vl-utils
-certifi==2026.2.25
-    # via
-    #   httpcore
-    #   httpx
    #   requests
    #   sentry-sdk
 cffi==2.0.0
    # via pymunk
-cfgv==3.5.0
+cfgv==3.4.0
    # via pre-commit
-charset-normalizer==3.4.5
+charset-normalizer==3.4.4
    # via requests
-click==8.3.1
+click==8.3.0
    # via
-    #   typer
    #   uvicorn
    #   wandb
-cloudpickle==3.1.2
-    # via gymnasium
-cmake==4.1.3
+cloudpickle==3.1.1
+    # via
+    #   gymnasium
+    #   libero
+cmake==4.1.0
    # via lerobot
-cmeel==0.59.0
+cmeel==0.57.3
    # via
    #   cmeel-assimp
    #   cmeel-boost
@@ -105,17 +108,15 @@ cmeel-zlib==1.3.1
    # via cmeel-assimp
 coal-library==3.0.1
    # via pin
-contourpy==1.3.3
-    # via
-    #   lerobot
-    #   matplotlib
-coverage[toml]==7.13.4
+contourpy==1.3.2
+    # via matplotlib
+coverage[toml]==7.11.0
    # via pytest-cov
 cycler==0.12.1
    # via matplotlib
-datasets==4.6.1
+datasets==4.1.1
    # via lerobot
-debugpy==1.8.20
+debugpy==1.8.17
    # via lerobot
 decorator==5.2.1
    # via ipython
@@ -129,7 +130,7 @@ dill==0.4.0
    #   multiprocess
 distlib==0.4.0
    # via virtualenv
-dm-control==1.0.37
+dm-control==1.0.34
    # via gym-aloha
 dm-env==1.6
    # via dm-control
@@ -137,55 +138,69 @@ dm-tree==0.1.9
    # via
    #   dm-control
    #   dm-env
+    #   lerobot
 docopt==0.6.2
    # via num2words
 draccus==0.10.0
    # via lerobot
 dynamixel-sdk==3.8.4
    # via lerobot
+easydict==1.13
+    # via libero
+egl-probe @ git+https://github.com/huggingface/egl_probe.git
+    # via
+    #   libero
+    #   robomimic
 eigenpy==3.10.3
    # via coal-library
-einops==0.8.2
-    # via lerobot
-eiquadprog==1.2.9
-    # via placo
-etils[epath,epy]==1.14.0
-    # via mujoco
-executing==2.2.1
-    # via stack-data
-faker==34.0.2
-    # via lerobot
-farama-notifications==0.0.4
-    # via gymnasium
-fastapi==0.135.1
+einops==0.8.1
    # via
    #   lerobot
-    #   teleop
+    #   libero
+eiquadprog==1.2.9
+    # via placo
+etils[epath,epy]==1.13.0
+    # via mujoco
+exceptiongroup==1.3.0
+    # via
+    #   anyio
+    #   ipython
+    #   pytest
+executing==2.2.1
+    # via stack-data
+farama-notifications==0.0.4
+    # via gymnasium
+fastapi==0.119.1
+    # via teleop
+fastjsonschema==2.21.2
+    # via nbformat
 feetech-servo-sdk==1.0.0
    # via lerobot
-filelock==3.25.0
+filelock==3.20.0
    # via
    #   datasets
    #   diffusers
    #   huggingface-hub
-    #   python-discovery
    #   torch
+    #   transformers
    #   virtualenv
-fonttools==4.61.1
+fonttools==4.60.1
    # via matplotlib
 frozenlist==1.8.0
    # via
    #   aiohttp
    #   aiosignal
-fsspec[http]==2026.2.0
+fsspec[http]==2025.9.0
    # via
    #   datasets
    #   etils
    #   huggingface-hub
    #   torch
+future==1.0.0
+    # via libero
 gitdb==4.0.12
    # via gitpython
-gitpython==3.1.46
+gitpython==3.1.45
    # via wandb
 glfw==2.10.0
    # via
@@ -197,6 +212,7 @@ grpcio==1.73.1
    #   lerobot
    #   reachy2-sdk
    #   reachy2-sdk-api
+    #   tensorboard
 grpcio-tools==1.73.1
    # via
    #   lerobot
@@ -207,67 +223,71 @@ gym-hil==0.1.13
    # via lerobot
 gym-pusht==0.1.6
    # via lerobot
-gymnasium==1.2.3
+gymnasium==1.2.1
    # via
    #   gym-aloha
    #   gym-hil
    #   gym-pusht
    #   lerobot
+    #   libero
    #   metaworld
 h11==0.16.0
-    # via
-    #   httpcore
-    #   uvicorn
+    # via uvicorn
+h5py==3.15.1
+    # via robomimic
 hebi-py==2.11.0
    # via lerobot
-hf-xet==1.3.2
+hf-transfer==0.1.9
+    # via huggingface-hub
+hf-xet==1.1.10
    # via huggingface-hub
 hidapi==0.14.0.post4
    # via
    #   gym-hil
    #   lerobot
-httpcore==1.0.9
-    # via httpx
 httptools==0.7.1
    # via uvicorn
-httpx==0.28.1
-    # via
-    #   datasets
-    #   huggingface-hub
-huggingface-hub==1.6.0
+huggingface-hub[cli,hf-transfer]==0.35.3
    # via
    #   accelerate
    #   datasets
    #   diffusers
    #   lerobot
    #   peft
+    #   timm
    #   tokenizers
    #   transformers
-identify==2.6.17
+hydra-core==1.3.2
+    # via libero
+identify==2.6.15
    # via pre-commit
 idna==3.11
    # via
    #   anyio
-    #   httpx
    #   requests
    #   yarl
-imageio[ffmpeg]==2.37.2
+imageio[ffmpeg]==2.37.0
    # via
    #   gym-aloha
    #   gym-hil
    #   lerobot
    #   metaworld
+    #   robomimic
    #   scikit-image
 imageio-ffmpeg==0.6.0
-    # via imageio
-importlib-metadata==8.7.1
+    # via
+    #   imageio
+    #   robomimic
+importlib-metadata==8.7.0
    # via diffusers
+importlib-resources==6.5.2
+    # via etils
 iniconfig==2.3.0
    # via pytest
-ipython==9.11.0
+inquirerpy==0.3.4
+    # via huggingface-hub
+ipython==8.37.0
    # via meshcat
-ipython-pygments-lexers==1.1.1
-    # via ipython
 ischedule==1.2.7
    # via placo
 jedi==0.19.2
@@ -276,24 +296,44 @@ jinja2==3.1.6
    # via torch
 jsonlines==4.0.0
    # via lerobot
+jsonschema==4.25.1
+    # via nbformat
+jsonschema-specifications==2025.9.1
+    # via jsonschema
+jupyter-core==5.9.1
+    # via nbformat
+jupytext==1.18.1
+    # via bddl
 kiwisolver==1.4.9
    # via matplotlib
 labmaze==1.0.6
    # via dm-control
-lazy-loader==0.5
+lazy-loader==0.4
    # via scikit-image
-librt==0.8.1
-    # via mypy
+libero @ git+https://github.com/huggingface/lerobot-libero.git@main
+    # via lerobot
+llvmlite==0.45.1
+    # via numba
 lxml==6.0.2
    # via dm-control
+markdown==3.9
+    # via tensorboard
 markdown-it-py==4.0.0
-    # via rich
+    # via
+    #   jupytext
+    #   mdit-py-plugins
 markupsafe==3.0.3
-    # via jinja2
-matplotlib==3.10.8
-    # via lerobot
+    # via
+    #   jinja2
+    #   werkzeug
+matplotlib==3.10.7
+    # via
+    #   lerobot
+    #   libero
 matplotlib-inline==0.2.1
    # via ipython
+mdit-py-plugins==0.5.0
+    # via jupytext
 mdurl==0.1.2
    # via markdown-it-py
 mergedeep==1.3.4
@@ -306,35 +346,41 @@ mock-serial==0.0.1
    # via lerobot
 mpmath==1.3.0
    # via sympy
-mujoco==3.5.0
+mujoco==3.3.7
    # via
    #   dm-control
    #   gym-aloha
    #   gym-hil
+    #   libero
    #   metaworld
-multidict==6.7.1
+    #   robosuite
+multidict==6.7.0
    # via
    #   aiohttp
    #   yarl
-multiprocess==0.70.18
+multiprocess==0.70.16
    # via datasets
-mypy==1.19.1
-    # via lerobot
 mypy-extensions==1.1.0
+    # via typing-inspect
+nbformat==5.10.4
+    # via jupytext
+networkx==3.4.2
    # via
-    #   mypy
-    #   typing-inspect
-networkx==3.6.1
-    # via
+    #   bddl
    #   scikit-image
    #   torch
-nodeenv==1.10.0
+ninja==1.13.0
+    # via lerobot
+nodeenv==1.9.1
    # via pre-commit
 num2words==0.5.14
    # via lerobot
+numba==0.62.1
+    # via robosuite
 numpy==2.2.6
    # via
    #   accelerate
+    #   bddl
    #   cmeel-boost
    #   contourpy
    #   datasets
@@ -343,14 +389,16 @@ numpy==2.2.6
    #   dm-env
    #   dm-tree
    #   gymnasium
+    #   h5py
    #   hebi-py
    #   imageio
    #   labmaze
-    #   lerobot
+    #   libero
    #   matplotlib
    #   meshcat
    #   metaworld
    #   mujoco
+    #   numba
    #   opencv-python
    #   opencv-python-headless
    #   pandas
@@ -358,18 +406,26 @@ numpy==2.2.6
    #   pyquaternion
    #   reachy2-sdk
    #   rerun-sdk
+    #   robomimic
+    #   robosuite
    #   scikit-image
    #   scipy
    #   shapely
    #   teleop
+    #   tensorboard
+    #   tensorboardx
    #   tifffile
    #   torchvision
    #   transformers
    #   transforms3d
-opencv-python==4.13.0.92
+omegaconf==2.3.0
+    # via hydra-core
+opencv-python==4.12.0.88
    # via
    #   gym-pusht
+    #   libero
    #   reachy2-sdk
+    #   robosuite
 opencv-python-headless==4.12.0.88
    # via lerobot
 orderly-set==5.5.0
@@ -379,87 +435,97 @@ packaging==25.0
    #   accelerate
    #   datasets
    #   huggingface-hub
+    #   hydra-core
+    #   jupytext
    #   lazy-loader
    #   lerobot
    #   matplotlib
    #   peft
    #   pytest
-    #   qwen-vl-utils
    #   reachy2-sdk
    #   scikit-image
+    #   tensorboard
+    #   tensorboardx
    #   transformers
    #   wandb
 pandas==2.3.3
    # via
    #   datasets
    #   lerobot
-parso==0.8.6
+parso==0.8.5
    # via jedi
-pathspec==1.0.4
-    # via mypy
-peft==0.18.1
+peft==0.17.1
    # via lerobot
 pexpect==4.9.0
    # via ipython
-pillow==12.1.1
+pfzy==0.3.4
+    # via inquirerpy
+pillow==12.0.0
    # via
    #   diffusers
    #   imageio
+    #   lerobot
    #   matplotlib
    #   meshcat
-    #   qwen-vl-utils
    #   rerun-sdk
+    #   robosuite
    #   scikit-image
+    #   tensorboard
    #   torchvision
 pin==3.4.0
    # via placo
-placo==0.9.16
+placo==0.9.14
    # via lerobot
-platformdirs==4.9.4
+platformdirs==4.5.0
    # via
-    #   python-discovery
+    #   jupyter-core
    #   virtualenv
    #   wandb
 pluggy==1.6.0
    # via
    #   pytest
    #   pytest-cov
-pre-commit==4.5.1
+pre-commit==4.3.0
    # via lerobot
 prompt-toolkit==3.0.52
-    # via ipython
+    # via
+    #   inquirerpy
+    #   ipython
 propcache==0.4.1
    # via
    #   aiohttp
    #   yarl
-protobuf==6.31.1
+protobuf==6.31.0
    # via
    #   dm-control
    #   grpcio-tools
    #   lerobot
    #   reachy2-sdk
    #   reachy2-sdk-api
+    #   tensorboard
+    #   tensorboardx
    #   wandb
-psutil==7.2.2
+psutil==7.1.1
    # via
    #   accelerate
    #   imageio
    #   peft
+    #   robomimic
 ptyprocess==0.7.0
    # via pexpect
 pure-eval==0.2.3
    # via stack-data
-pyarrow==23.0.1
+pyarrow==21.0.0
    # via
    #   datasets
    #   rerun-sdk
-pycparser==3.0
+pycparser==2.23
    # via cffi
-pydantic==2.12.5
+pydantic==2.12.3
    # via
    #   fastapi
    #   wandb
-pydantic-core==2.41.5
+pydantic-core==2.41.4
    # via pydantic
 pygame==2.6.1
    # via
@@ -469,35 +535,33 @@ pygame==2.6.1
 pygments==2.19.2
    # via
    #   ipython
-    #   ipython-pygments-lexers
    #   pytest
-    #   rich
 pymunk==6.11.1
    # via
    #   gym-pusht
    #   lerobot
-pyngrok==7.5.1
+pyngrok==7.4.1
    # via meshcat
 pynput==1.8.1
    # via
    #   gym-hil
    #   lerobot
-pyobjc-core==12.1
+pyobjc-core==12.0
    # via
    #   pyobjc-framework-applicationservices
    #   pyobjc-framework-cocoa
    #   pyobjc-framework-coretext
    #   pyobjc-framework-quartz
-pyobjc-framework-applicationservices==12.1
+pyobjc-framework-applicationservices==12.0
    # via pynput
-pyobjc-framework-cocoa==12.1
+pyobjc-framework-cocoa==12.0
    # via
    #   pyobjc-framework-applicationservices
    #   pyobjc-framework-coretext
    #   pyobjc-framework-quartz
-pyobjc-framework-coretext==12.1
+pyobjc-framework-coretext==12.0
    # via pyobjc-framework-applicationservices
-pyobjc-framework-quartz==12.1
+pyobjc-framework-quartz==12.0
    # via
    #   pynput
    #   pyobjc-framework-applicationservices
@@ -506,13 +570,13 @@ pyopengl==3.1.10
    # via
    #   dm-control
    #   mujoco
-pyparsing==3.3.2
+pyparsing==3.2.5
    # via
    #   dm-control
    #   matplotlib
 pyquaternion==0.9.9
    # via reachy2-sdk
-pyrealsense2-macosx==2.56.5
+pyrealsense2-macosx==2.54.2
    # via lerobot
 pyserial==3.5
    # via
@@ -521,6 +585,7 @@ pyserial==3.5
    #   lerobot
 pytest==8.4.2
    # via
+    #   bddl
    #   lerobot
    #   pytest-cov
    #   pytest-timeout
@@ -531,14 +596,11 @@ pytest-timeout==2.4.0
    # via lerobot
 python-dateutil==2.9.0.post0
    # via
-    #   faker
    #   matplotlib
    #   pandas
-python-discovery==1.1.1
-    # via virtualenv
-python-dotenv==1.2.2
+python-dotenv==1.1.1
    # via uvicorn
-pytz==2026.1.post1
+pytz==2025.2
    # via pandas
 pyyaml==6.0.3
    # via
@@ -547,10 +609,13 @@ pyyaml==6.0.3
    #   draccus
    #   hebi-py
    #   huggingface-hub
+    #   jupytext
+    #   omegaconf
    #   peft
    #   pre-commit
    #   pyngrok
    #   pyyaml-include
+    #   timm
    #   transformers
    #   uvicorn
    #   wandb
@@ -560,13 +625,15 @@ pyzmq==27.1.0
    # via
    #   lerobot
    #   meshcat
-qwen-vl-utils==0.0.14
-    # via lerobot
-reachy2-sdk==1.0.15
+reachy2-sdk==1.0.14
    # via lerobot
 reachy2-sdk-api==1.0.21
    # via reachy2-sdk
-regex==2026.2.28
+referencing==0.37.0
+    # via
+    #   jsonschema
+    #   jsonschema-specifications
+regex==2025.10.23
    # via
    #   diffusers
    #   transformers
@@ -575,150 +642,184 @@ requests==2.32.5
    #   datasets
    #   diffusers
    #   dm-control
-    #   qwen-vl-utils
+    #   huggingface-hub
    #   teleop
+    #   transformers
    #   wandb
-rerun-sdk==0.26.2
+rerun-sdk==0.26.1
    # via lerobot
 rhoban-cmeel-jsoncpp==1.9.4.9
    # via placo
-rich==14.3.3
-    # via typer
-safetensors==0.7.0
+robomimic==0.2.0
+    # via libero
+robosuite==1.4.0
+    # via libero
+rpds-py==0.28.0
+    # via
+    #   jsonschema
+    #   referencing
+safetensors==0.6.2
    # via
    #   accelerate
    #   diffusers
    #   lerobot
    #   peft
+    #   timm
    #   transformers
 scikit-image==0.25.2
    # via
    #   gym-pusht
    #   lerobot
-scipy==1.17.1
+scipy==1.15.3
    # via
    #   dm-control
-    #   lerobot
    #   metaworld
+    #   robosuite
    #   scikit-image
-    #   torchdiffeq
-sentry-sdk==2.54.0
+sentry-sdk==2.42.1
    # via wandb
 shapely==2.1.2
    # via gym-pusht
-shellingham==1.5.4
-    # via typer
 six==1.17.0
    # via
    #   pynput
    #   python-dateutil
-smmap==5.0.3
+smmap==5.0.2
    # via gitdb
+sniffio==1.3.1
+    # via anyio
 stack-data==0.6.3
    # via ipython
-starlette==0.52.1
+starlette==0.48.0
    # via fastapi
 sympy==1.14.0
    # via torch
-teleop==0.1.4
+teleop==0.1.2
    # via lerobot
-termcolor==3.3.0
-    # via lerobot
-tifffile==2026.3.3
+tensorboard==2.20.0
+    # via robomimic
+tensorboard-data-server==0.7.2
+    # via tensorboard
+tensorboardx==2.6.4
+    # via robomimic
+termcolor==3.1.0
+    # via
+    #   lerobot
+    #   robomimic
+thop==0.1.1.post2209072238
+    # via libero
+tifffile==2025.5.10
    # via scikit-image
-tokenizers==0.22.2
+timm==1.0.20
+    # via lerobot
+tokenizers==0.22.1
    # via transformers
 toml==0.10.2
    # via draccus
-torch==2.10.0
+tomli==2.3.0
+    # via
+    #   cmeel
+    #   coverage
+    #   jupytext
+    #   pytest
+torch==2.7.1
    # via
    #   accelerate
    #   lerobot
    #   peft
-    #   torchdiffeq
+    #   robomimic
+    #   thop
+    #   timm
    #   torchvision
-torchcodec==0.10.0
+torchcodec==0.5
    # via lerobot
-torchdiffeq==0.2.5
-    # via lerobot
-torchvision==0.25.0
-    # via lerobot
-tornado==6.5.4
+torchvision==0.22.1
+    # via
+    #   lerobot
+    #   robomimic
+    #   timm
+tornado==6.5.2
    # via meshcat
-tqdm==4.67.3
+tqdm==4.67.1
    # via
    #   datasets
    #   dm-control
    #   huggingface-hub
    #   peft
+    #   robomimic
    #   transformers
 traitlets==5.14.3
    # via
    #   ipython
+    #   jupyter-core
    #   matplotlib-inline
-transformers==5.3.0
+    #   nbformat
+transformers==4.57.1
    # via
    #   lerobot
+    #   libero
    #   peft
 transforms3d==0.4.2
    # via teleop
-typer==0.24.1
-    # via
-    #   huggingface-hub
-    #   transformers
 typing-extensions==4.15.0
    # via
    #   aiosignal
    #   anyio
    #   etils
-    #   faker
+    #   exceptiongroup
    #   fastapi
    #   gymnasium
    #   huggingface-hub
-    #   mypy
+    #   ipython
+    #   multidict
    #   pydantic
    #   pydantic-core
+    #   referencing
    #   rerun-sdk
    #   starlette
    #   torch
    #   typing-inspect
    #   typing-inspection
+    #   uvicorn
+    #   virtualenv
    #   wandb
 typing-inspect==0.9.0
    # via draccus
 typing-inspection==0.4.2
-    # via
-    #   fastapi
-    #   pydantic
-tzdata==2025.3
+    # via pydantic
+tzdata==2025.2
    # via pandas
 u-msgpack-python==2.8.0
    # via meshcat
-urllib3==2.6.3
+urllib3==2.5.0
    # via
    #   requests
    #   sentry-sdk
-uvicorn[standard]==0.41.0
+uvicorn[standard]==0.38.0
    # via teleop
 uvloop==0.22.1
    # via uvicorn
-virtualenv==21.1.0
+virtualenv==20.35.3
    # via pre-commit
-wandb==0.24.2
-    # via lerobot
+wandb==0.21.4
+    # via
+    #   lerobot
+    #   libero
 watchfiles==1.1.1
    # via uvicorn
-wcwidth==0.6.0
+wcwidth==0.2.14
    # via prompt-toolkit
 websocket-client==1.9.0
    # via teleop
-websockets==16.0
+websockets==15.0.1
    # via uvicorn
-wrapt==2.1.2
+werkzeug==3.1.3
+    # via tensorboard
+wrapt==2.0.0
    # via dm-tree
 xxhash==3.6.0
    # via datasets
-yarl==1.23.0
+yarl==1.22.0
    # via aiohttp
 zipp==3.23.0
    # via
@@ -1,12 +1,12 @@
 #
-# This file is autogenerated by pip-compile with Python 3.12
+# This file is autogenerated by pip-compile with Python 3.10
 # by the following command:
 #
 #    pip-compile --output-file=requirements-ubuntu.txt requirements.in
 #
 -e .[all]
    # via -[all]
-absl-py==2.4.0
+absl-py==2.3.1
    # via
    #   dm-control
    #   dm-env
@@ -14,33 +14,30 @@ absl-py==2.4.0
    #   labmaze
    #   mujoco
    #   tensorboard
-accelerate==1.13.0
+accelerate==1.11.0
    # via
    #   lerobot
    #   peft
 aiohappyeyeballs==2.6.1
    # via aiohttp
-aiohttp==3.13.3
+aiohttp==3.13.1
    # via fsspec
 aiosignal==1.4.0
    # via aiohttp
-annotated-doc==0.0.4
-    # via
-    #   fastapi
-    #   typer
 annotated-types==0.7.0
    # via pydantic
 antlr4-python3-runtime==4.9.3
    # via
    #   hydra-core
    #   omegaconf
-anyio==4.12.1
+anyio==4.11.0
    # via
-    #   httpx
    #   starlette
    #   watchfiles
-asttokens==3.0.1
+asttokens==3.0.0
    # via stack-data
+async-timeout==5.0.1
+    # via aiohttp
 attrs==25.4.0
    # via
    #   aiohttp
@@ -50,35 +47,30 @@ attrs==25.4.0
    #   referencing
    #   rerun-sdk
 av==15.1.0
-    # via
-    #   lerobot
-    #   qwen-vl-utils
+    # via lerobot
 bddl==1.0.1
-    # via hf-libero
-certifi==2026.2.25
+    # via libero
+certifi==2025.10.5
    # via
-    #   httpcore
-    #   httpx
    #   requests
    #   sentry-sdk
 cffi==2.0.0
    # via pymunk
-cfgv==3.5.0
+cfgv==3.4.0
    # via pre-commit
-charset-normalizer==3.4.5
+charset-normalizer==3.4.4
    # via requests
-click==8.3.1
+click==8.3.0
    # via
-    #   typer
    #   uvicorn
    #   wandb
-cloudpickle==3.1.2
+cloudpickle==3.1.1
    # via
    #   gymnasium
-    #   hf-libero
-cmake==4.1.3
+    #   libero
+cmake==4.1.0
    # via lerobot
-cmeel==0.59.0
+cmeel==0.57.3
    # via
    #   cmeel-assimp
    #   cmeel-boost
@@ -116,24 +108,20 @@ cmeel-zlib==1.3.1
    # via cmeel-assimp
 coal-library==3.0.1
    # via pin
-contourpy==1.3.3
-    # via
-    #   lerobot
-    #   matplotlib
-coverage[toml]==7.13.4
+contourpy==1.3.2
+    # via matplotlib
+coverage[toml]==7.11.0
    # via pytest-cov
-cuda-bindings==12.9.4
-    # via torch
-cuda-pathfinder==1.4.1
-    # via cuda-bindings
 cycler==0.12.1
    # via matplotlib
-datasets==4.6.1
+datasets==4.1.1
    # via lerobot
-debugpy==1.8.20
+debugpy==1.8.17
    # via lerobot
 decorator==5.2.1
    # via ipython
+decord==0.6.0
+    # via lerobot
 deepdiff==8.6.1
    # via lerobot
 diffusers==0.35.2
@@ -144,7 +132,7 @@ dill==0.4.0
    #   multiprocess
 distlib==0.4.0
    # via virtualenv
-dm-control==1.0.37
+dm-control==1.0.34
    # via gym-aloha
 dm-env==1.6
    # via dm-control
@@ -152,6 +140,7 @@ dm-tree==0.1.9
    # via
    #   dm-control
    #   dm-env
+    #   lerobot
 docopt==0.6.2
    # via num2words
 draccus==0.10.0
@@ -159,60 +148,66 @@ draccus==0.10.0
 dynamixel-sdk==3.8.4
    # via lerobot
 easydict==1.13
-    # via hf-libero
-egl-probe==1.0.2
-    # via robomimic
+    # via libero
+egl-probe @ git+https://github.com/huggingface/egl_probe.git
+    # via
+    #   libero
+    #   robomimic
 eigenpy==3.10.3
    # via coal-library
-einops==0.8.2
+einops==0.8.1
    # via
-    #   hf-libero
+    #   flash-attn
    #   lerobot
+    #   libero
 eiquadprog==1.2.9
    # via placo
-etils[epath,epy]==1.14.0
+etils[epath,epy]==1.13.0
    # via mujoco
-evdev==1.9.3
+evdev==1.9.2
    # via pynput
+exceptiongroup==1.3.0
+    # via
+    #   anyio
+    #   ipython
+    #   pytest
 executing==2.2.1
    # via stack-data
-faker==34.0.2
-    # via lerobot
 farama-notifications==0.0.4
    # via gymnasium
-fastapi==0.135.1
-    # via
-    #   lerobot
-    #   teleop
+fastapi==0.119.1
+    # via teleop
 fastjsonschema==2.21.2
    # via nbformat
 feetech-servo-sdk==1.0.0
    # via lerobot
-filelock==3.25.0
+filelock==3.20.0
    # via
    #   datasets
    #   diffusers
    #   huggingface-hub
-    #   python-discovery
    #   torch
+    #   transformers
    #   virtualenv
-fonttools==4.61.1
+flash-attn==2.8.3
+    # via lerobot
+fonttools==4.60.1
    # via matplotlib
 frozenlist==1.8.0
    # via
    #   aiohttp
    #   aiosignal
-fsspec[http]==2026.2.0
+fsspec[http]==2025.9.0
    # via
    #   datasets
    #   etils
    #   huggingface-hub
    #   torch
 future==1.0.0
-    # via hf-libero
+    # via libero
 gitdb==4.0.12
    # via gitpython
-gitpython==3.1.46
+gitpython==3.1.45
    # via wandb
 glfw==2.10.0
    # via
@@ -235,60 +230,50 @@ gym-hil==0.1.13
    # via lerobot
 gym-pusht==0.1.6
    # via lerobot
-gymnasium==1.2.3
+gymnasium==1.2.1
    # via
    #   gym-aloha
    #   gym-hil
    #   gym-pusht
-    #   hf-libero
    #   lerobot
+    #   libero
    #   metaworld
 h11==0.16.0
-    # via
-    #   httpcore
-    #   uvicorn
-h5py==3.16.0
+    # via uvicorn
+h5py==3.15.1
    # via robomimic
 hebi-py==2.11.0
    # via lerobot
-hf-egl-probe==1.0.2
-    # via hf-libero
-hf-libero==0.1.3
-    # via lerobot
-hf-xet==1.3.2
+hf-transfer==0.1.9
+    # via huggingface-hub
+hf-xet==1.1.10
    # via huggingface-hub
 hidapi==0.14.0.post4
    # via
    #   gym-hil
    #   lerobot
-httpcore==1.0.9
-    # via httpx
 httptools==0.7.1
    # via uvicorn
-httpx==0.28.1
-    # via
-    #   datasets
-    #   huggingface-hub
-huggingface-hub==1.6.0
+huggingface-hub[cli,hf-transfer]==0.35.3
    # via
    #   accelerate
    #   datasets
    #   diffusers
    #   lerobot
    #   peft
+    #   timm
    #   tokenizers
    #   transformers
 hydra-core==1.3.2
-    # via hf-libero
-identify==2.6.17
+    # via libero
+identify==2.6.15
    # via pre-commit
 idna==3.11
    # via
    #   anyio
-    #   httpx
    #   requests
    #   yarl
-imageio[ffmpeg]==2.37.2
+imageio[ffmpeg]==2.37.0
    # via
    #   gym-aloha
    #   gym-hil
@@ -300,14 +285,16 @@ imageio-ffmpeg==0.6.0
    # via
    #   imageio
    #   robomimic
-importlib-metadata==8.7.1
+importlib-metadata==8.7.0
    # via diffusers
+importlib-resources==6.5.2
+    # via etils
 iniconfig==2.3.0
    # via pytest
-ipython==9.11.0
+inquirerpy==0.3.4
+    # via huggingface-hub
+ipython==8.37.0
    # via meshcat
-ipython-pygments-lexers==1.1.1
-    # via ipython
 ischedule==1.2.7
    # via placo
 jedi==0.19.2
@@ -316,41 +303,40 @@ jinja2==3.1.6
    # via torch
 jsonlines==4.0.0
    # via lerobot
-jsonschema==4.26.0
+jsonschema==4.25.1
    # via nbformat
 jsonschema-specifications==2025.9.1
    # via jsonschema
 jupyter-core==5.9.1
    # via nbformat
-jupytext==1.19.1
+jupytext==1.18.1
    # via bddl
 kiwisolver==1.4.9
    # via matplotlib
 labmaze==1.0.6
    # via dm-control
-lazy-loader==0.5
+lazy-loader==0.4
    # via scikit-image
-librt==0.8.1
-    # via mypy
-llvmlite==0.46.0
+libero @ git+https://github.com/huggingface/lerobot-libero.git@main
+    # via lerobot
+llvmlite==0.45.1
    # via numba
 lxml==6.0.2
    # via dm-control
-markdown==3.10.2
+markdown==3.9
    # via tensorboard
 markdown-it-py==4.0.0
    # via
    #   jupytext
    #   mdit-py-plugins
-    #   rich
 markupsafe==3.0.3
    # via
    #   jinja2
    #   werkzeug
-matplotlib==3.10.8
+matplotlib==3.10.7
    # via
-    #   hf-libero
    #   lerobot
+    #   libero
 matplotlib-inline==0.2.1
    # via ipython
 mdit-py-plugins==0.5.0
@@ -367,38 +353,36 @@ mock-serial==0.0.1
    # via lerobot
 mpmath==1.3.0
    # via sympy
-mujoco==3.5.0
+mujoco==3.3.7
    # via
    #   dm-control
    #   gym-aloha
    #   gym-hil
-    #   hf-libero
+    #   libero
    #   metaworld
    #   robosuite
-multidict==6.7.1
+multidict==6.7.0
    # via
    #   aiohttp
    #   yarl
-multiprocess==0.70.18
+multiprocess==0.70.16
    # via datasets
-mypy==1.19.1
-    # via lerobot
 mypy-extensions==1.1.0
-    # via
-    #   mypy
-    #   typing-inspect
+    # via typing-inspect
 nbformat==5.10.4
    # via jupytext
-networkx==3.6.1
+networkx==3.4.2
    # via
    #   bddl
    #   scikit-image
    #   torch
-nodeenv==1.10.0
+ninja==1.13.0
+    # via lerobot
+nodeenv==1.9.1
    # via pre-commit
 num2words==0.5.14
    # via lerobot
-numba==0.64.0
+numba==0.62.1
    # via robosuite
 numpy==2.2.6
    # via
@@ -407,6 +391,7 @@ numpy==2.2.6
    #   cmeel-boost
    #   contourpy
    #   datasets
+    #   decord
    #   diffusers
    #   dm-control
    #   dm-env
@@ -414,10 +399,9 @@ numpy==2.2.6
    #   gymnasium
    #   h5py
    #   hebi-py
-    #   hf-libero
    #   imageio
    #   labmaze
-    #   lerobot
+    #   libero
    #   matplotlib
    #   meshcat
    #   metaworld
@@ -442,51 +426,49 @@ numpy==2.2.6
    #   torchvision
    #   transformers
    #   transforms3d
-nvidia-cublas-cu12==12.8.4.1
+nvidia-cublas-cu12==12.6.4.1
    # via
    #   nvidia-cudnn-cu12
    #   nvidia-cusolver-cu12
    #   torch
-nvidia-cuda-cupti-cu12==12.8.90
+nvidia-cuda-cupti-cu12==12.6.80
    # via torch
-nvidia-cuda-nvrtc-cu12==12.8.93
+nvidia-cuda-nvrtc-cu12==12.6.77
    # via torch
-nvidia-cuda-runtime-cu12==12.8.90
+nvidia-cuda-runtime-cu12==12.6.77
    # via torch
-nvidia-cudnn-cu12==9.10.2.21
+nvidia-cudnn-cu12==9.5.1.17
    # via torch
-nvidia-cufft-cu12==11.3.3.83
+nvidia-cufft-cu12==11.3.0.4
    # via torch
-nvidia-cufile-cu12==1.13.1.3
+nvidia-cufile-cu12==1.11.1.6
    # via torch
-nvidia-curand-cu12==10.3.9.90
+nvidia-curand-cu12==10.3.7.77
    # via torch
-nvidia-cusolver-cu12==11.7.3.90
+nvidia-cusolver-cu12==11.7.1.2
    # via torch
-nvidia-cusparse-cu12==12.5.8.93
+nvidia-cusparse-cu12==12.5.4.2
    # via
    #   nvidia-cusolver-cu12
    #   torch
-nvidia-cusparselt-cu12==0.7.1
+nvidia-cusparselt-cu12==0.6.3
    # via torch
-nvidia-nccl-cu12==2.27.5
+nvidia-nccl-cu12==2.26.2
    # via torch
-nvidia-nvjitlink-cu12==12.8.93
+nvidia-nvjitlink-cu12==12.6.85
    # via
    #   nvidia-cufft-cu12
    #   nvidia-cusolver-cu12
    #   nvidia-cusparse-cu12
    #   torch
-nvidia-nvshmem-cu12==3.4.5
-    # via torch
-nvidia-nvtx-cu12==12.8.90
+nvidia-nvtx-cu12==12.6.77
    # via torch
 omegaconf==2.3.0
    # via hydra-core
-opencv-python==4.13.0.92
+opencv-python==4.12.0.88
    # via
    #   gym-pusht
-    #   hf-libero
+    #   libero
    #   reachy2-sdk
    #   robosuite
 opencv-python-headless==4.12.0.88
@@ -505,7 +487,6 @@ packaging==25.0
    #   matplotlib
    #   peft
    #   pytest
-    #   qwen-vl-utils
    #   reachy2-sdk
    #   scikit-image
    #   tensorboard
@@ -516,21 +497,21 @@ pandas==2.3.3
    # via
    #   datasets
    #   lerobot
-parso==0.8.6
+parso==0.8.5
    # via jedi
-pathspec==1.0.4
-    # via mypy
-peft==0.18.1
+peft==0.17.1
    # via lerobot
 pexpect==4.9.0
    # via ipython
-pillow==12.1.1
+pfzy==0.3.4
+    # via inquirerpy
+pillow==12.0.0
    # via
    #   diffusers
    #   imageio
+    #   lerobot
    #   matplotlib
    #   meshcat
-    #   qwen-vl-utils
    #   rerun-sdk
    #   robosuite
    #   scikit-image
@@ -538,27 +519,28 @@ pillow==12.1.1
    #   torchvision
 pin==3.4.0
    # via placo
-placo==0.9.16
+placo==0.9.14
    # via lerobot
-platformdirs==4.9.4
+platformdirs==4.5.0
    # via
    #   jupyter-core
-    #   python-discovery
    #   virtualenv
    #   wandb
 pluggy==1.6.0
    # via
    #   pytest
    #   pytest-cov
-pre-commit==4.5.1
+pre-commit==4.3.0
    # via lerobot
 prompt-toolkit==3.0.52
-    # via ipython
+    # via
+    #   inquirerpy
+    #   ipython
 propcache==0.4.1
    # via
    #   aiohttp
    #   yarl
-protobuf==6.31.1
+protobuf==6.31.0
    # via
    #   dm-control
    #   grpcio-tools
@@ -568,7 +550,7 @@ protobuf==6.31.1
    #   tensorboard
    #   tensorboardx
    #   wandb
-psutil==7.2.2
+psutil==7.1.1
    # via
    #   accelerate
    #   imageio
@@ -578,17 +560,17 @@ ptyprocess==0.7.0
    # via pexpect
 pure-eval==0.2.3
    # via stack-data
-pyarrow==23.0.1
+pyarrow==21.0.0
    # via
    #   datasets
    #   rerun-sdk
-pycparser==3.0
+pycparser==2.23
    # via cffi
-pydantic==2.12.5
+pydantic==2.12.3
    # via
    #   fastapi
    #   wandb
-pydantic-core==2.41.5
+pydantic-core==2.41.4
    # via pydantic
 pygame==2.6.1
    # via
@@ -598,14 +580,12 @@ pygame==2.6.1
 pygments==2.19.2
    # via
    #   ipython
-    #   ipython-pygments-lexers
    #   pytest
-    #   rich
 pymunk==6.11.1
    # via
    #   gym-pusht
    #   lerobot
-pyngrok==7.5.1
+pyngrok==7.4.1
    # via meshcat
 pynput==1.8.1
    # via
@@ -615,7 +595,7 @@ pyopengl==3.1.10
    # via
    #   dm-control
    #   mujoco
-pyparsing==3.3.2
+pyparsing==3.2.5
    # via
    #   dm-control
    #   matplotlib
@@ -641,16 +621,13 @@ pytest-timeout==2.4.0
    # via lerobot
 python-dateutil==2.9.0.post0
    # via
-    #   faker
    #   matplotlib
    #   pandas
-python-discovery==1.1.1
-    # via virtualenv
-python-dotenv==1.2.2
+python-dotenv==1.1.1
    # via uvicorn
 python-xlib==0.33
    # via pynput
-pytz==2026.1.post1
+pytz==2025.2
    # via pandas
 pyyaml==6.0.3
    # via
@@ -665,6 +642,7 @@ pyyaml==6.0.3
    #   pre-commit
    #   pyngrok
    #   pyyaml-include
+    #   timm
    #   transformers
    #   uvicorn
    #   wandb
@@ -674,9 +652,7 @@ pyzmq==27.1.0
    # via
    #   lerobot
    #   meshcat
-qwen-vl-utils==0.0.14
-    # via lerobot
-reachy2-sdk==1.0.15
+reachy2-sdk==1.0.14
    # via lerobot
 reachy2-sdk-api==1.0.21
    # via reachy2-sdk
@@ -684,7 +660,7 @@ referencing==0.37.0
    # via
    #   jsonschema
    #   jsonschema-specifications
-regex==2026.2.28
+regex==2025.10.23
    # via
    #   diffusers
    #   transformers
@@ -693,62 +669,60 @@ requests==2.32.5
    #   datasets
    #   diffusers
    #   dm-control
-    #   qwen-vl-utils
+    #   huggingface-hub
    #   teleop
+    #   transformers
    #   wandb
-rerun-sdk==0.26.2
+rerun-sdk==0.26.1
    # via lerobot
 rhoban-cmeel-jsoncpp==1.9.4.9
    # via placo
-rich==14.3.3
-    # via typer
 robomimic==0.2.0
-    # via hf-libero
+    # via libero
 robosuite==1.4.0
-    # via hf-libero
-rpds-py==0.30.0
+    # via libero
+rpds-py==0.28.0
    # via
    #   jsonschema
    #   referencing
-safetensors==0.7.0
+safetensors==0.6.2
    # via
    #   accelerate
    #   diffusers
    #   lerobot
    #   peft
+    #   timm
    #   transformers
 scikit-image==0.25.2
    # via
    #   gym-pusht
    #   lerobot
-scipy==1.17.1
+scipy==1.15.3
    # via
    #   dm-control
-    #   lerobot
    #   metaworld
    #   robosuite
    #   scikit-image
-    #   torchdiffeq
-sentry-sdk==2.54.0
+sentry-sdk==2.42.1
    # via wandb
 shapely==2.1.2
    # via gym-pusht
-shellingham==1.5.4
-    # via typer
 six==1.17.0
    # via
    #   pynput
    #   python-dateutil
    #   python-xlib
-smmap==5.0.3
+smmap==5.0.2
    # via gitdb
+sniffio==1.3.1
+    # via anyio
 stack-data==0.6.3
    # via ipython
-starlette==0.52.1
+starlette==0.48.0
    # via fastapi
 sympy==1.14.0
    # via torch
-teleop==0.1.4
+teleop==0.1.2
    # via lerobot
 tensorboard==2.20.0
    # via robomimic
@@ -756,38 +730,46 @@ tensorboard-data-server==0.7.2
    # via tensorboard
 tensorboardx==2.6.4
    # via robomimic
-termcolor==3.3.0
+termcolor==3.1.0
    # via
    #   lerobot
    #   robomimic
 thop==0.1.1.post2209072238
-    # via hf-libero
-tifffile==2026.3.3
+    # via libero
+tifffile==2025.5.10
    # via scikit-image
-tokenizers==0.22.2
+timm==1.0.20
+    # via lerobot
+tokenizers==0.22.1
    # via transformers
 toml==0.10.2
    # via draccus
-torch==2.10.0
+tomli==2.3.0
+    # via
+    #   cmeel
+    #   coverage
+    #   jupytext
+    #   pytest
+torch==2.7.1
    # via
    #   accelerate
+    #   flash-attn
    #   lerobot
    #   peft
    #   robomimic
    #   thop
-    #   torchdiffeq
+    #   timm
    #   torchvision
-torchcodec==0.10.0
+torchcodec==0.5
    # via lerobot
-torchdiffeq==0.2.5
-    # via lerobot
-torchvision==0.25.0
+torchvision==0.22.1
    # via
    #   lerobot
    #   robomimic
-tornado==6.5.4
+    #   timm
+tornado==6.5.2
    # via meshcat
-tqdm==4.67.3
+tqdm==4.67.1
    # via
    #   datasets
    #   dm-control
@@ -801,29 +783,26 @@ traitlets==5.14.3
    #   jupyter-core
    #   matplotlib-inline
    #   nbformat
-transformers==5.3.0
+transformers==4.57.1
    # via
-    #   hf-libero
    #   lerobot
+    #   libero
    #   peft
 transforms3d==0.4.2
    # via teleop
-triton==3.6.0
+triton==3.3.1
    # via torch
-typer==0.24.1
-    # via
-    #   huggingface-hub
-    #   transformers
 typing-extensions==4.15.0
    # via
    #   aiosignal
    #   anyio
    #   etils
-    #   faker
+    #   exceptiongroup
    #   fastapi
    #   gymnasium
    #   huggingface-hub
-    #   mypy
+    #   ipython
+    #   multidict
    #   pydantic
    #   pydantic-core
    #   referencing
@@ -832,46 +811,46 @@ typing-extensions==4.15.0
    #   torch
    #   typing-inspect
    #   typing-inspection
+    #   uvicorn
+    #   virtualenv
    #   wandb
 typing-inspect==0.9.0
    # via draccus
 typing-inspection==0.4.2
-    # via
-    #   fastapi
-    #   pydantic
-tzdata==2025.3
+    # via pydantic
+tzdata==2025.2
    # via pandas
 u-msgpack-python==2.8.0
    # via meshcat
-urllib3==2.6.3
+urllib3==2.5.0
    # via
    #   requests
    #   sentry-sdk
-uvicorn[standard]==0.41.0
+uvicorn[standard]==0.38.0
    # via teleop
 uvloop==0.22.1
    # via uvicorn
-virtualenv==21.1.0
+virtualenv==20.35.3
    # via pre-commit
-wandb==0.24.2
+wandb==0.21.4
    # via
-    #   hf-libero
    #   lerobot
+    #   libero
 watchfiles==1.1.1
    # via uvicorn
-wcwidth==0.6.0
+wcwidth==0.2.14
    # via prompt-toolkit
 websocket-client==1.9.0
    # via teleop
-websockets==16.0
+websockets==15.0.1
    # via uvicorn
-werkzeug==3.1.6
+werkzeug==3.1.3
    # via tensorboard
-wrapt==2.1.2
+wrapt==2.0.0
    # via dm-tree
 xxhash==3.6.0
    # via datasets
-yarl==1.23.0
+yarl==1.22.0
    # via aiohttp
 zipp==3.23.0
    # via
@@ -1,9 +1,9 @@
 # requirements.in

-# requirements-macos.txt was generated on macOS and is platform-specific (macOS 26.3.1 25D2128 arm64).
-# Darwin MacBook-Pro.local 25.3.0 Darwin Kernel Version 25.3.0: Wed Jan 28 20:54:55 PST 2026; root:xnu-12377.91.3~2/RELEASE_ARM64_T8132 arm64
+# requirements-macos.txt was generated on macOS and is platform-specific (macOS 26.0.1 25A362 arm64).
+# Darwin MacBook-Pro.local 25.0.0 Darwin Kernel Version 25.0.0: Wed Sep 17 21:42:08 PDT 2025; root:xnu-12377.1.9~141/RELEASE_ARM64_T8132 arm64

-# requirements-ubuntu.txt was generated on Linux and is platform-specific (Ubuntu 24.04.4 LTS x86_64).
-# Linux lerobot-linux 6.17.0-14-generic #14~24.04.1-Ubuntu SMP PREEMPT_DYNAMIC Thu Jan 15 15:52:10 UTC 2 x86_64 x86_64 x86_64 GNU/Linux
+# requirements-ubuntu.txt was generated on Linux and is platform-specific (Ubuntu 24.04.3 LTS x86_64).
+# Linux mlerobot-linux 6.14.0-33-generic #33~24.04.1-Ubuntu SMP PREEMPT_DYNAMIC Fri Sep 19 17:02:30 UTC 2 x86_64 x86_64 x86_64 GNU/Linux

 -e .[all]
@@ -63,9 +63,9 @@ from lerobot.transport import (
    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,
@@ -485,9 +485,8 @@ class RobotClient:
 def async_client(cfg: RobotClientConfig):
    logging.info(pformat(asdict(cfg)))

-    # 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!")
+    if cfg.robot.type not in SUPPORTED_ROBOTS:
+        raise ValueError(f"Robot {cfg.robot.type} not yet supported!")

    client = RobotClient(cfg)

@@ -513,5 +512,4 @@ def async_client(cfg: RobotClientConfig):


 if __name__ == "__main__":
-    register_third_party_plugins()
    async_client()  # run the client
@@ -150,7 +150,7 @@ class Camera(abc.ABC):
        """
        pass

-    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
+    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
@@ -530,7 +530,7 @@ class OpenCVCamera(Camera):
        return frame

    @check_if_not_connected
-    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
+    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
@@ -201,7 +201,7 @@ class Reachy2Camera(Camera):
        return self.read()

    @check_if_not_connected
-    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
+    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
@@ -573,7 +573,7 @@ class RealSenseCamera(Camera):

    # NOTE(Steven): Missing implementation for depth for now
    @check_if_not_connected
-    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
+    def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
        """Return the most recent (color) frame captured immediately (Peeking).

        This method is non-blocking and returns whatever is currently in the
@@ -181,7 +181,7 @@ class ZMQCamera(Camera):
        try:
            message = self.socket.recv_string()
        except Exception as e:
-            # zmq is lazy-imported in connect(), so check by name to avoid a top-level import
+            # 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
@@ -23,7 +23,6 @@ import base64
 import contextlib
 import json
 import logging
-import threading
 import time
 from collections import deque

@@ -43,57 +42,10 @@ def encode_image(image: np.ndarray, quality: int = 80) -> str:
    return base64.b64encode(buffer).decode("utf-8")


-class CameraCaptureThread:
-    """Background thread that continuously captures and encodes frames from a camera."""
-
-    def __init__(self, camera: OpenCVCamera, name: str):
-        self.camera = camera
-        self.name = name
-        self.latest_encoded: str | None = None  # Pre-encoded JPEG as base64
-        self.latest_timestamp: float = 0.0
-        self.frame_lock = threading.Lock()
-        self.running = False
-        self.thread: threading.Thread | None = None
-
-    def start(self):
-        """Start the capture thread."""
-        self.running = True
-        self.thread = threading.Thread(target=self._capture_loop, daemon=True)
-        self.thread.start()
-
-    def stop(self):
-        """Stop the capture thread."""
-        self.running = False
-        if self.thread:
-            self.thread.join(timeout=1.0)
-
-    def _capture_loop(self):
-        """Continuously capture and encode frames at the camera's native rate."""
-        while self.running:
-            try:
-                frame = self.camera.read()  # Blocks at camera's native rate
-                timestamp = time.time()
-                # Encode immediately in capture thread (this is the slow part)
-                encoded = encode_image(frame)
-                with self.frame_lock:
-                    self.latest_encoded = encoded
-                    self.latest_timestamp = timestamp
-            except Exception as e:
-                logger.warning(f"Camera {self.name} capture error: {e}")
-                time.sleep(0.01)
-
-    def get_latest(self) -> tuple[str | None, float]:
-        """Get the latest encoded frame and its timestamp."""
-        with self.frame_lock:
-            return self.latest_encoded, self.latest_timestamp
-
-
 class ImageServer:
    def __init__(self, config: dict, port: int = 5555):
-        # fps controls the publish loop rate (how often frames are sent over ZMQ), not the camera capture rate
        self.fps = config.get("fps", 30)
        self.cameras: dict[str, OpenCVCamera] = {}
-        self.capture_threads: dict[str, CameraCaptureThread] = {}

        for name, cfg in config.get("cameras", {}).items():
            shape = cfg.get("shape", [480, 640])
@@ -109,10 +61,6 @@ class ImageServer:
            self.cameras[name] = camera
            logger.info(f"Camera {name}: {shape[1]}x{shape[0]}")

-            # Create capture thread for this camera
-            capture_thread = CameraCaptureThread(camera, name)
-            self.capture_threads[name] = capture_thread
-
        # ZMQ PUB socket
        self.context = zmq.Context()
        self.socket = self.context.socket(zmq.PUB)
@@ -125,18 +73,6 @@ class ImageServer:
    def run(self):
        frame_count = 0
        frame_times = deque(maxlen=60)
-        last_published_ts: dict[str, float] = {}
-
-        # Start all capture threads
-        for capture_thread in self.capture_threads.values():
-            capture_thread.start()
-
-        # Wait for first frames to be captured and encoded
-        logger.info("Waiting for cameras to start capturing...")
-        for name, capture_thread in self.capture_threads.items():
-            while capture_thread.get_latest()[0] is None:
-                time.sleep(0.01)
-            logger.info(f"Camera {name} ready (capture + encode in background)")

        try:
            while True:
@@ -144,12 +80,10 @@ class ImageServer:

                # Build message
                message = {"timestamps": {}, "images": {}}
-                for name, capture_thread in self.capture_threads.items():
-                    encoded, timestamp = capture_thread.get_latest()
-                    if encoded is not None and timestamp > last_published_ts.get(name, 0.0):
-                        message["timestamps"][name] = timestamp
-                        message["images"][name] = encoded
-                        last_published_ts[name] = timestamp
+                for name, cam in self.cameras.items():
+                    frame = cam.read()  # Returns RGB
+                    message["timestamps"][name] = time.time()
+                    message["images"][name] = encode_image(frame)

                # Send as JSON string (suppress if buffer full)
                with contextlib.suppress(zmq.Again):
@@ -168,8 +102,6 @@ class ImageServer:
        except KeyboardInterrupt:
            pass
        finally:
-            for capture_thread in self.capture_threads.values():
-                capture_thread.stop()
            for cam in self.cameras.values():
                cam.disconnect()
            self.socket.close()
@@ -27,7 +27,7 @@ class DatasetConfig:
    # "dataset_index" into the returned item. The index mapping is made according to the order in which the
    # datasets are provided.
    repo_id: str
-    # Root directory where the dataset will be stored (e.g. 'dataset/path'). If None, defaults to $HF_LEROBOT_HOME/repo_id.
+    # Root directory where the dataset will be stored (e.g. 'dataset/path').
    root: str | None = None
    episodes: list[int] | None = None
    image_transforms: ImageTransformsConfig = field(default_factory=ImageTransformsConfig)
@@ -50,9 +50,6 @@ class TrainPipelineConfig(HubMixin):
    # `seed` is used for training (eg: model initialization, dataset shuffling)
    # AND for the evaluation environments.
    seed: int | None = 1000
-    # Set to True to use deterministic cuDNN algorithms for reproducibility.
-    # This disables cudnn.benchmark and may reduce training speed by ~10-20%.
-    cudnn_deterministic: bool = False
    # Number of workers for the dataloader.
    num_workers: int = 4
    batch_size: int = 8
@@ -289,9 +289,7 @@ def aggregate_datasets(

    logging.info("Find all tasks")
    unique_tasks = pd.concat([m.tasks for m in all_metadata]).index.unique()
-    dst_meta.tasks = pd.DataFrame(
-        {"task_index": range(len(unique_tasks))}, index=pd.Index(unique_tasks, name="task")
-    )
+    dst_meta.tasks = pd.DataFrame({"task_index": range(len(unique_tasks))}, index=unique_tasks)

    meta_idx = {"chunk": 0, "file": 0}
    data_idx = {"chunk": 0, "file": 0}
@@ -7,13 +7,6 @@

 This dataset was created using [LeRobot](https://github.com/huggingface/lerobot).

-{% if repo_id is defined and repo_id %}
-<a class="flex" href="https://huggingface.co/spaces/lerobot/visualize_dataset?path={{ repo_id }}">
-<img class="block dark:hidden" src="https://huggingface.co/datasets/huggingface/badges/resolve/main/visualize-this-dataset-xl.svg"/>
-<img class="hidden dark:block" src="https://huggingface.co/datasets/huggingface/badges/resolve/main/visualize-this-dataset-xl-dark.svg"/>
-</a>
-{% endif %}
-
 ## Dataset Description

 {{ dataset_description | default("", true) }}
@@ -37,7 +37,7 @@ import torch
 from tqdm import tqdm

 from lerobot.datasets.aggregate import aggregate_datasets
-from lerobot.datasets.compute_stats import aggregate_stats
+from lerobot.datasets.compute_stats import aggregate_stats, compute_episode_stats, get_feature_stats
 from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
 from lerobot.datasets.utils import (
    DATA_DIR,
@@ -89,8 +89,8 @@ def delete_episodes(
    Args:
        dataset: The source LeRobotDataset.
        episode_indices: List of episode indices to delete.
-        output_dir: Root directory where the edited dataset will be stored. If not specified, defaults to $HF_LEROBOT_HOME/repo_id. Equivalent to new_root in EditDatasetConfig.
-        repo_id: Edited dataset identifier. Equivalent to new_repo_id in EditDatasetConfig.
+        output_dir: Directory to save the new dataset. If None, uses default location.
+        repo_id: Repository ID for the new dataset. If None, appends "_modified" to original.
    """
    if not episode_indices:
        raise ValueError("No episodes to delete")
@@ -152,7 +152,7 @@ def split_dataset(
        dataset: The source LeRobotDataset to split.
        splits: Either a dict mapping split names to episode indices, or a dict mapping
                split names to fractions (must sum to <= 1.0).
-        output_dir: Root directory where the split datasets will be stored. If not specified, defaults to $HF_LEROBOT_HOME/repo_id.
+        output_dir: Base directory for output datasets. If None, uses default location.

    Examples:
      Split by specific episodes
@@ -243,8 +243,8 @@ def merge_datasets(

    Args:
        datasets: List of LeRobotDatasets to merge.
-        output_repo_id: Merged dataset identifier.
-        output_dir: Root directory where the merged dataset will be stored. If not specified, defaults to $HF_LEROBOT_HOME/output_repo_id.
+        output_repo_id: Repository ID for the merged dataset.
+        output_dir: Directory to save the merged dataset. If None, uses default location.
    """
    if not datasets:
        raise ValueError("No datasets to merge")
@@ -288,8 +288,8 @@ def modify_features(
        dataset: The source LeRobotDataset.
        add_features: Optional dict mapping feature names to (feature_values, feature_info) tuples.
        remove_features: Optional feature name(s) to remove. Can be a single string or list.
-        output_dir: Root directory where the edited dataset will be stored. If not specified, defaults to $HF_LEROBOT_HOME/repo_id. Equivalent to new_root in EditDatasetConfig.
-        repo_id: Edited dataset identifier. Equivalent to new_repo_id in EditDatasetConfig.
+        output_dir: Directory to save the new dataset. If None, uses default location.
+        repo_id: Repository ID for the new dataset. If None, appends "_modified" to original.

    Returns:
        New dataset with features modified.
@@ -390,8 +390,8 @@ def add_features(
    Args:
        dataset: The source LeRobotDataset.
        features: Dictionary mapping feature names to (feature_values, feature_info) tuples.
-        output_dir: Root directory where the edited dataset will be stored. If not specified, defaults to $HF_LEROBOT_HOME/repo_id. Equivalent to new_root in EditDatasetConfig.
-        repo_id: Edited dataset identifier. Equivalent to new_repo_id in EditDatasetConfig.
+        output_dir: Directory to save the new dataset. If None, uses default location.
+        repo_id: Repository ID for the new dataset. If None, appends "_modified" to original.

    Returns:
        New dataset with all features added.
@@ -427,8 +427,8 @@ def remove_feature(
    Args:
        dataset: The source LeRobotDataset.
        feature_names: Name(s) of features to remove. Can be a single string or list.
-        output_dir: Root directory where the edited dataset will be stored. If not specified, defaults to $HF_LEROBOT_HOME/repo_id. Equivalent to new_root in EditDatasetConfig.
-        repo_id: Edited dataset identifier. Equivalent to new_repo_id in EditDatasetConfig.
+        output_dir: Directory to save the new dataset. If None, uses default location.
+        repo_id: Repository ID for the new dataset. If None, appends "_modified" to original.

    Returns:
        New dataset with features removed.
@@ -567,22 +567,20 @@ def _copy_and_reindex_data(
 def _keep_episodes_from_video_with_av(
    input_path: Path,
    output_path: Path,
-    episodes_to_keep: list[tuple[int, int]],
+    episodes_to_keep: list[tuple[float, float]],
    fps: float,
    vcodec: str = "libsvtav1",
    pix_fmt: str = "yuv420p",
 ) -> None:
    """Keep only specified episodes from a video file using PyAV.

-    This function decodes frames from specified frame ranges and re-encodes them with
+    This function decodes frames from specified time ranges and re-encodes them with
    properly reset timestamps to ensure monotonic progression.

    Args:
        input_path: Source video file path.
        output_path: Destination video file path.
-        episodes_to_keep: List of (start_frame, end_frame) tuples for episodes to keep.
-            Ranges are half-open intervals: [start_frame, end_frame), where start_frame
-            is inclusive and end_frame is exclusive.
+        episodes_to_keep: List of (start_time, end_time) tuples for episodes to keep.
        fps: Frame rate of the video.
        vcodec: Video codec to use for encoding.
        pix_fmt: Pixel format for output video.
@@ -624,10 +622,9 @@ def _keep_episodes_from_video_with_av(

    # Create set of (start, end) ranges for fast lookup.
    # Convert to a sorted list for efficient checking.
-    frame_ranges = sorted(episodes_to_keep)
+    time_ranges = sorted(episodes_to_keep)

    # Track frame index for setting PTS and current range being processed.
-    src_frame_count = 0
    frame_count = 0
    range_idx = 0

@@ -637,20 +634,21 @@ def _keep_episodes_from_video_with_av(
            if frame is None:
                continue

-            # Check if frame is in any of our desired frame ranges.
+            # Get frame timestamp.
+            frame_time = float(frame.pts * frame.time_base) if frame.pts is not None else 0.0
+
+            # Check if frame is in any of our desired time ranges.
            # Skip ranges that have already passed.
-            while range_idx < len(frame_ranges) and src_frame_count >= frame_ranges[range_idx][1]:
+            while range_idx < len(time_ranges) and frame_time >= time_ranges[range_idx][1]:
                range_idx += 1

            # If we've passed all ranges, stop processing.
-            if range_idx >= len(frame_ranges):
+            if range_idx >= len(time_ranges):
                break

            # Check if frame is in current range.
-            start_frame = frame_ranges[range_idx][0]
-
-            if src_frame_count < start_frame:
-                src_frame_count += 1
+            start_ts, end_ts = time_ranges[range_idx]
+            if frame_time < start_ts:
                continue

            # Frame is in range - create a new frame with reset timestamps.
@@ -663,7 +661,6 @@ def _keep_episodes_from_video_with_av(
            for pkt in v_out.encode(new_frame):
                out.mux(pkt)

-            src_frame_count += 1
            frame_count += 1

    # Flush encoder.
@@ -752,17 +749,15 @@ def _copy_and_reindex_videos(
                        f"videos/{video_key}/to_timestamp"
                    ]
            else:
-                # Build list of frame ranges to keep, in sorted order.
+                # Build list of time ranges to keep, in sorted order.
                sorted_keep_episodes = sorted(episodes_in_file, key=lambda x: episode_mapping[x])
-                episodes_to_keep_ranges: list[tuple[int, int]] = []
+                episodes_to_keep_ranges: list[tuple[float, float]] = []
+
                for old_idx in sorted_keep_episodes:
                    src_ep = src_dataset.meta.episodes[old_idx]
-                    from_frame = round(src_ep[f"videos/{video_key}/from_timestamp"] * src_dataset.meta.fps)
-                    to_frame = round(src_ep[f"videos/{video_key}/to_timestamp"] * src_dataset.meta.fps)
-                    assert src_ep["length"] == to_frame - from_frame, (
-                        f"Episode length mismatch: {src_ep['length']} vs {to_frame - from_frame}"
-                    )
-                    episodes_to_keep_ranges.append((from_frame, to_frame))
+                    from_ts = src_ep[f"videos/{video_key}/from_timestamp"]
+                    to_ts = src_ep[f"videos/{video_key}/to_timestamp"]
+                    episodes_to_keep_ranges.append((from_ts, to_ts))

                # Use PyAV filters to efficiently re-encode only the desired segments.
                assert src_dataset.meta.video_path is not None
@@ -1475,9 +1470,7 @@ def modify_tasks(

    # Collect all unique tasks and create new task mapping
    unique_tasks = sorted(set(episode_to_task.values()))
-    new_task_df = pd.DataFrame(
-        {"task_index": list(range(len(unique_tasks)))}, index=pd.Index(unique_tasks, name="task")
-    )
+    new_task_df = pd.DataFrame({"task_index": list(range(len(unique_tasks)))}, index=unique_tasks)
    task_to_index = {task: idx for idx, task in enumerate(unique_tasks)}

    logging.info(f"Modifying tasks in {dataset.repo_id}")
@@ -1529,9 +1522,125 @@ def modify_tasks(
    return dataset


+def recompute_stats(
+    dataset: LeRobotDataset,
+    skip_image_video: bool = True,
+    delta_action: bool = False,
+    delta_exclude_joints: list[str] | None = None,
+) -> LeRobotDataset:
+    """Recompute stats.json from scratch by iterating all episodes.
+
+    Args:
+        dataset: The LeRobotDataset to recompute stats for.
+        skip_image_video: If True (default), only recompute stats for numeric features
+            (action, state, etc.) and keep existing image/video stats unchanged.
+        delta_action: If True, compute action stats as delta (action - state).
+            Useful when training with use_delta_actions=True so normalization matches.
+        delta_exclude_joints: Joint names to exclude from delta conversion when
+            delta_action=True. These dims keep absolute stats. Uses dataset's
+            action feature names to build the mask. Default: ["gripper"].
+
+    Returns:
+        The same dataset with updated stats.
+    """
+    features = dataset.meta.features
+    numeric_features = {
+        k: v for k, v in features.items()
+        if v["dtype"] not in ["image", "video", "string"]
+        and k not in ["index", "episode_index", "task_index", "frame_index", "timestamp"]
+    }
+
+    if skip_image_video:
+        features_to_compute = numeric_features
+    else:
+        features_to_compute = {
+            k: v for k, v in features.items()
+            if v["dtype"] != "string"
+            and k not in ["index", "episode_index", "task_index", "frame_index", "timestamp"]
+        }
+
+    # Build delta mask if delta_action is enabled
+    delta_mask = None
+    if delta_action and "action" in features and "observation.state" in features:
+        if delta_exclude_joints is None:
+            delta_exclude_joints = ["gripper"]
+        action_names = features["action"].get("names")
+        if action_names is not None:
+            exclude = set(delta_exclude_joints)
+            delta_mask = [n not in exclude for n in action_names]
+        else:
+            action_dim = features["action"]["shape"][0]
+            delta_mask = [True] * action_dim
+        # Only recompute action stats when delta is enabled — state stays unchanged
+        features_to_compute = {"action": features["action"]}
+        logging.info(f"Recomputing action stats as delta (exclude: {delta_exclude_joints})")
+    else:
+        logging.info(f"Recomputing stats for features: {list(features_to_compute.keys())}")
+
+    data_dir = dataset.root / DATA_DIR
+    parquet_files = sorted(data_dir.glob("*/*.parquet"))
+    if not parquet_files:
+        raise ValueError(f"No parquet files found in {data_dir}")
+
+    all_episode_stats = []
+    numeric_keys = [k for k, v in features_to_compute.items() if v["dtype"] not in ["image", "video"]]
+    # Also need state for delta computation even though we don't recompute state stats
+    needs_state = delta_mask is not None
+
+    for parquet_path in tqdm(parquet_files, desc="Computing stats from data files"):
+        df = pd.read_parquet(parquet_path)
+
+        for ep_idx in sorted(df["episode_index"].unique()):
+            ep_df = df[df["episode_index"] == ep_idx]
+            episode_data = {}
+            for key in numeric_keys:
+                if key in ep_df.columns:
+                    values = ep_df[key].values
+                    if hasattr(values[0], "__len__"):
+                        episode_data[key] = np.stack(values)
+                    else:
+                        episode_data[key] = np.array(values)
+
+            # Apply delta conversion to actions before computing stats
+            if delta_mask is not None and "action" in episode_data:
+                from lerobot.processor.delta_action_processor import to_delta_actions
+
+                # Load state for delta even if we're not computing state stats
+                if needs_state and "observation.state" in ep_df.columns:
+                    state_values = ep_df["observation.state"].values
+                    if hasattr(state_values[0], "__len__"):
+                        states = np.stack(state_values)
+                    else:
+                        states = np.array(state_values)
+                    actions_t = torch.from_numpy(episode_data["action"]).float()
+                    states_t = torch.from_numpy(states).float()
+                    episode_data["action"] = to_delta_actions(actions_t, states_t, delta_mask).numpy()
+
+            ep_stats = compute_episode_stats(episode_data, features_to_compute)
+            all_episode_stats.append(ep_stats)
+
+    if not all_episode_stats:
+        logging.warning("No episode stats computed")
+        return dataset
+
+    new_stats = aggregate_stats(all_episode_stats)
+
+    # Merge: keep existing stats for features we didn't recompute
+    if dataset.meta.stats:
+        for key, value in dataset.meta.stats.items():
+            if key not in new_stats:
+                new_stats[key] = value
+
+    write_stats(new_stats, dataset.root)
+    dataset.meta.stats = new_stats
+
+    logging.info(f"Stats recomputed for {len(all_episode_stats)} episodes")
+    return dataset
+
+
 def convert_image_to_video_dataset(
    dataset: LeRobotDataset,
-    output_dir: Path | None = None,
+    output_dir: Path,
    repo_id: str | None = None,
    vcodec: str = "libsvtav1",
    pix_fmt: str = "yuv420p",
@@ -1550,8 +1659,8 @@ def convert_image_to_video_dataset(

    Args:
        dataset: The source LeRobot dataset with images
-        output_dir: Root directory where the edited dataset will be stored. If not specified, defaults to $HF_LEROBOT_HOME/repo_id. Equivalent to new_root in EditDatasetConfig.
-        repo_id: Edited dataset identifier. Equivalent to new_repo_id in EditDatasetConfig.
+        output_dir: Directory to save the new video dataset
+        repo_id: Repository ID for the new dataset (default: original_id + "_video")
        vcodec: Video codec (default: libsvtav1)
        pix_fmt: Pixel format (default: yuv420p)
        g: Group of pictures size (default: 2)
@@ -1602,7 +1711,6 @@ def convert_image_to_video_dataset(
            # Video info will be updated after episodes are encoded

    # Create new metadata for video dataset
-    output_dir = Path(output_dir) if output_dir is not None else HF_LEROBOT_HOME / repo_id
    new_meta = LeRobotDatasetMetadata.create(
        repo_id=repo_id,
        fps=dataset.meta.fps,
@@ -68,7 +68,6 @@ from lerobot.datasets.utils import (
    write_tasks,
 )
 from lerobot.datasets.video_utils import (
-    StreamingVideoEncoder,
    VideoFrame,
    concatenate_video_files,
    decode_video_frames,
@@ -76,11 +75,11 @@ from lerobot.datasets.video_utils import (
    get_safe_default_codec,
    get_video_duration_in_s,
    get_video_info,
-    resolve_vcodec,
 )
 from lerobot.utils.constants import HF_LEROBOT_HOME

 CODEBASE_VERSION = "v3.0"
+VALID_VIDEO_CODECS = {"h264", "hevc", "libsvtav1"}


 class LeRobotDatasetMetadata:
@@ -314,7 +313,7 @@ class LeRobotDatasetMetadata:
        if self.tasks is None:
            new_tasks = tasks
            task_indices = range(len(tasks))
-            self.tasks = pd.DataFrame({"task_index": task_indices}, index=pd.Index(tasks, name="task"))
+            self.tasks = pd.DataFrame({"task_index": task_indices}, index=tasks)
        else:
            new_tasks = [task for task in tasks if task not in self.tasks.index]
            new_task_indices = range(len(self.tasks), len(self.tasks) + len(new_tasks))
@@ -546,19 +545,12 @@ class LeRobotDatasetMetadata:


 def _encode_video_worker(
-    video_key: str,
-    episode_index: int,
-    root: Path,
-    fps: int,
-    vcodec: str = "libsvtav1",
-    encoder_threads: int | None = None,
+    video_key: str, episode_index: int, root: Path, fps: int, vcodec: str = "libsvtav1"
 ) -> Path:
    temp_path = Path(tempfile.mkdtemp(dir=root)) / f"{video_key}_{episode_index:03d}.mp4"
    fpath = DEFAULT_IMAGE_PATH.format(image_key=video_key, episode_index=episode_index, frame_index=0)
    img_dir = (root / fpath).parent
-    encode_video_frames(
-        img_dir, temp_path, fps, vcodec=vcodec, overwrite=True, encoder_threads=encoder_threads
-    )
+    encode_video_frames(img_dir, temp_path, fps, vcodec=vcodec, overwrite=True)
    shutil.rmtree(img_dir)
    return temp_path

@@ -578,9 +570,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
        video_backend: str | None = None,
        batch_encoding_size: int = 1,
        vcodec: str = "libsvtav1",
-        streaming_encoding: bool = False,
-        encoder_queue_maxsize: int = 30,
-        encoder_threads: int | None = None,
    ):
        """
        2 modes are available for instantiating this class, depending on 2 different use cases:
@@ -664,11 +653,11 @@ class LeRobotDataset(torch.utils.data.Dataset):
        for the README).

        Args:
-            repo_id (str): This is the repo id that will be used to fetch the dataset.
-            root (Path | None, optional): Local directory where the dataset will be downloaded and
-                stored. If set, all dataset files will be stored directly under this path. If not set, the
-                dataset files will be stored under $HF_LEROBOT_HOME/repo_id (configurable via the
-                HF_LEROBOT_HOME environment variable).
+            repo_id (str): This is the repo id that will be used to fetch the dataset. Locally, the dataset
+                will be stored under root/repo_id.
+            root (Path | None, optional): Local directory to use for downloading/writing files. You can also
+                set the LEROBOT_HOME environment variable to point to a different location. Defaults to
+                '~/.cache/huggingface/lerobot'.
            episodes (list[int] | None, optional): If specified, this will only load episodes specified by
                their episode_index in this list. Defaults to None.
            image_transforms (Callable | None, optional): You can pass standard v2 image transforms from
@@ -694,17 +683,12 @@ class LeRobotDataset(torch.utils.data.Dataset):
            batch_encoding_size (int, optional): Number of episodes to accumulate before batch encoding videos.
                Set to 1 for immediate encoding (default), or higher for batched encoding. Defaults to 1.
            vcodec (str, optional): Video codec for encoding videos during recording. Options: 'h264', 'hevc',
-                'libsvtav1', 'auto', or hardware-specific codecs like 'h264_videotoolbox', 'h264_nvenc'.
-                Defaults to 'libsvtav1'. Use 'auto' to auto-detect the best available hardware encoder.
-            streaming_encoding (bool, optional): If True, encode video frames in real-time during capture
-                instead of writing PNG images first. This makes save_episode() near-instant. Defaults to False.
-            encoder_queue_maxsize (int, optional): Maximum number of frames to buffer per camera when using
-                streaming encoding. Defaults to 30 (~1s at 30fps).
-            encoder_threads (int | None, optional): Number of threads per encoder instance. None lets the
-                codec auto-detect (default). Lower values reduce CPU usage per encoder. Maps to 'lp' (via svtav1-params) for
-                libsvtav1 and 'threads' for h264/hevc.
+                'libsvtav1'. Defaults to 'libsvtav1'. Use 'h264' for faster encoding on systems where AV1
+                encoding is CPU-heavy.
        """
        super().__init__()
+        if vcodec not in VALID_VIDEO_CODECS:
+            raise ValueError(f"Invalid vcodec '{vcodec}'. Must be one of: {sorted(VALID_VIDEO_CODECS)}")
        self.repo_id = repo_id
        self.root = Path(root) if root else HF_LEROBOT_HOME / repo_id
        self.image_transforms = image_transforms
@@ -716,8 +700,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
        self.delta_indices = None
        self.batch_encoding_size = batch_encoding_size
        self.episodes_since_last_encoding = 0
-        self.vcodec = resolve_vcodec(vcodec)
-        self._encoder_threads = encoder_threads
+        self.vcodec = vcodec

        # Unused attributes
        self.image_writer = None
@@ -725,7 +708,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
        self.writer = None
        self.latest_episode = None
        self._current_file_start_frame = None  # Track the starting frame index of the current parquet file
-        self._streaming_encoder = None

        self.root.mkdir(exist_ok=True, parents=True)

@@ -747,7 +729,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
            # Check if cached dataset contains all requested episodes
            if not self._check_cached_episodes_sufficient():
                raise FileNotFoundError("Cached dataset doesn't contain all requested episodes")
-        except (FileNotFoundError, NotADirectoryError):
+        except (AssertionError, FileNotFoundError, NotADirectoryError):
            if is_valid_version(self.revision):
                self.revision = get_safe_version(self.repo_id, self.revision)
            self.download(download_videos)
@@ -767,19 +749,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
            check_delta_timestamps(self.delta_timestamps, self.fps, self.tolerance_s)
            self.delta_indices = get_delta_indices(self.delta_timestamps, self.fps)

-        # Initialize streaming encoder for resumed recording
-        if streaming_encoding and len(self.meta.video_keys) > 0:
-            self._streaming_encoder = StreamingVideoEncoder(
-                fps=self.meta.fps,
-                vcodec=self.vcodec,
-                pix_fmt="yuv420p",
-                g=2,
-                crf=30,
-                preset=None,
-                queue_maxsize=encoder_queue_maxsize,
-                encoder_threads=encoder_threads,
-            )
-
    def _close_writer(self) -> None:
        """Close and cleanup the parquet writer if it exists."""
        writer = getattr(self, "writer", None)
@@ -839,7 +808,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
            hub_api.upload_folder(**upload_kwargs)

        card = create_lerobot_dataset_card(
-            tags=tags, dataset_info=self.meta.info, license=license, repo_id=self.repo_id, **card_kwargs
+            tags=tags, dataset_info=self.meta.info, license=license, **card_kwargs
        )
        card.push_to_hub(repo_id=self.repo_id, repo_type="dataset", revision=branch)

@@ -1135,8 +1104,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
        """
        self._close_writer()
        self.meta._close_writer()
-        if self._streaming_encoder is not None:
-            self._streaming_encoder.close()

    def create_episode_buffer(self, episode_index: int | None = None) -> dict:
        current_ep_idx = self.meta.total_episodes if episode_index is None else episode_index
@@ -1191,13 +1158,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
        self.episode_buffer["timestamp"].append(timestamp)
        self.episode_buffer["task"].append(frame.pop("task"))  # Remove task from frame after processing

-        # Start streaming encoder on first frame of episode (once, before iterating keys)
-        if frame_index == 0 and self._streaming_encoder is not None:
-            self._streaming_encoder.start_episode(
-                video_keys=list(self.meta.video_keys),
-                temp_dir=self.root,
-            )
-
        # Add frame features to episode_buffer
        for key in frame:
            if key not in self.features:
@@ -1205,10 +1165,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
                    f"An element of the frame is not in the features. '{key}' not in '{self.features.keys()}'."
                )

-            if self.features[key]["dtype"] == "video" and self._streaming_encoder is not None:
-                self._streaming_encoder.feed_frame(key, frame[key])
-                self.episode_buffer[key].append(None)  # Placeholder (video keys are skipped in parquet)
-            elif self.features[key]["dtype"] in ["image", "video"]:
+            if self.features[key]["dtype"] in ["image", "video"]:
                img_path = self._get_image_file_path(
                    episode_index=self.episode_buffer["episode_index"], image_key=key, frame_index=frame_index
                )
@@ -1269,38 +1226,13 @@ class LeRobotDataset(torch.utils.data.Dataset):

        # Wait for image writer to end, so that episode stats over images can be computed
        self._wait_image_writer()
-
-        has_video_keys = len(self.meta.video_keys) > 0
-        use_streaming = self._streaming_encoder is not None and has_video_keys
-        use_batched_encoding = self.batch_encoding_size > 1
-
-        if use_streaming:
-            # Compute stats for non-video features only (video stats come from encoder)
-            non_video_buffer = {
-                k: v
-                for k, v in episode_buffer.items()
-                if self.features.get(k, {}).get("dtype") not in ("video",)
-            }
-            non_video_features = {k: v for k, v in self.features.items() if v["dtype"] != "video"}
-            ep_stats = compute_episode_stats(non_video_buffer, non_video_features)
-        else:
-            ep_stats = compute_episode_stats(episode_buffer, self.features)
+        ep_stats = compute_episode_stats(episode_buffer, self.features)

        ep_metadata = self._save_episode_data(episode_buffer)
+        has_video_keys = len(self.meta.video_keys) > 0
+        use_batched_encoding = self.batch_encoding_size > 1

-        if use_streaming:
-            # Finish streaming encoding and collect results
-            streaming_results = self._streaming_encoder.finish_episode()
-            for video_key in self.meta.video_keys:
-                temp_path, video_stats = streaming_results[video_key]
-                if video_stats is not None:
-                    # Format stats same as compute_episode_stats: normalize to [0,1], reshape to (C,1,1)
-                    ep_stats[video_key] = {
-                        k: v if k == "count" else np.squeeze(v.reshape(1, -1, 1, 1) / 255.0, axis=0)
-                        for k, v in video_stats.items()
-                    }
-                ep_metadata.update(self._save_episode_video(video_key, episode_index, temp_path=temp_path))
-        elif has_video_keys and not use_batched_encoding:
+        if has_video_keys and not use_batched_encoding:
            num_cameras = len(self.meta.video_keys)
            if parallel_encoding and num_cameras > 1:
                # TODO(Steven): Ideally we would like to control the number of threads per encoding such that:
@@ -1314,7 +1246,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
                            self.root,
                            self.fps,
                            self.vcodec,
-                            self._encoder_threads,
                        ): video_key
                        for video_key in self.meta.video_keys
                    }
@@ -1583,10 +1514,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
        return metadata

    def clear_episode_buffer(self, delete_images: bool = True) -> None:
-        # Cancel streaming encoder if active
-        if self._streaming_encoder is not None:
-            self._streaming_encoder.cancel_episode()
-
        # Clean up image files for the current episode buffer
        if delete_images:
            # Wait for the async image writer to finish
@@ -1634,9 +1561,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
        Note: `encode_video_frames` is a blocking call. Making it asynchronous shouldn't speedup encoding,
        since video encoding with ffmpeg is already using multithreading.
        """
-        return _encode_video_worker(
-            video_key, episode_index, self.root, self.fps, self.vcodec, self._encoder_threads
-        )
+        return _encode_video_worker(video_key, episode_index, self.root, self.fps, self.vcodec)

    @classmethod
    def create(
@@ -1653,13 +1578,10 @@ class LeRobotDataset(torch.utils.data.Dataset):
        video_backend: str | None = None,
        batch_encoding_size: int = 1,
        vcodec: str = "libsvtav1",
-        metadata_buffer_size: int = 10,
-        streaming_encoding: bool = False,
-        encoder_queue_maxsize: int = 30,
-        encoder_threads: int | None = None,
    ) -> "LeRobotDataset":
        """Create a LeRobot Dataset from scratch in order to record data."""
-        vcodec = resolve_vcodec(vcodec)
+        if vcodec not in VALID_VIDEO_CODECS:
+            raise ValueError(f"Invalid vcodec '{vcodec}'. Must be one of: {sorted(VALID_VIDEO_CODECS)}")
        obj = cls.__new__(cls)
        obj.meta = LeRobotDatasetMetadata.create(
            repo_id=repo_id,
@@ -1668,7 +1590,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
            features=features,
            root=root,
            use_videos=use_videos,
-            metadata_buffer_size=metadata_buffer_size,
        )
        obj.repo_id = obj.meta.repo_id
        obj.root = obj.meta.root
@@ -1678,7 +1599,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
        obj.batch_encoding_size = batch_encoding_size
        obj.episodes_since_last_encoding = 0
        obj.vcodec = vcodec
-        obj._encoder_threads = encoder_threads

        if image_writer_processes or image_writer_threads:
            obj.start_image_writer(image_writer_processes, image_writer_threads)
@@ -1700,22 +1620,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
        obj._lazy_loading = False
        obj._recorded_frames = 0
        obj._writer_closed_for_reading = False
-
-        # Initialize streaming encoder
-        if streaming_encoding and len(obj.meta.video_keys) > 0:
-            obj._streaming_encoder = StreamingVideoEncoder(
-                fps=fps,
-                vcodec=vcodec,
-                pix_fmt="yuv420p",
-                g=2,
-                crf=30,
-                preset=None,
-                queue_maxsize=encoder_queue_maxsize,
-                encoder_threads=encoder_threads,
-            )
-        else:
-            obj._streaming_encoder = None
-
        return obj


@@ -1771,12 +1675,11 @@ class MultiLeRobotDataset(torch.utils.data.Dataset):
            )
        for repo_id, ds in zip(self.repo_ids, self._datasets, strict=True):
            extra_keys = set(ds.features).difference(intersection_features)
-            if extra_keys:
-                logging.warning(
-                    f"keys {extra_keys} of {repo_id} were disabled as they are not contained in all the "
-                    "other datasets."
-                )
-                self.disabled_features.update(extra_keys)
+            logging.warning(
+                f"keys {extra_keys} of {repo_id} were disabled as they are not contained in all the "
+                "other datasets."
+            )
+            self.disabled_features.update(extra_keys)

        self.image_transforms = image_transforms
        self.delta_timestamps = delta_timestamps
@@ -21,7 +21,7 @@ from collections import deque
 from collections.abc import Iterable, Iterator
 from pathlib import Path
 from pprint import pformat
-from typing import Any
+from typing import Any, Generic, TypeVar

 import datasets
 import numpy as np
@@ -78,6 +78,8 @@ DEFAULT_FEATURES = {
    "task_index": {"dtype": "int64", "shape": (1,), "names": None},
 }

+T = TypeVar("T")
+

 def get_parquet_file_size_in_mb(parquet_path: str | Path) -> float:
    metadata = pq.read_metadata(parquet_path)
@@ -120,9 +122,19 @@ def load_nested_dataset(
        raise FileNotFoundError(f"Provided directory does not contain any parquet file: {pq_dir}")

    with SuppressProgressBars():
-        # We use .from_parquet() memory-mapped loading for efficiency
-        filters = pa_ds.field("episode_index").isin(episodes) if episodes is not None else None
-        return Dataset.from_parquet([str(path) for path in paths], filters=filters, features=features)
+        # When no filtering needed, Dataset uses memory-mapped loading for efficiency
+        # PyArrow loads the entire dataset into memory
+        if episodes is None:
+            return Dataset.from_parquet([str(path) for path in paths], features=features)
+
+        arrow_dataset = pa_ds.dataset(paths, format="parquet")
+        filter_expr = pa_ds.field("episode_index").isin(episodes)
+        table = arrow_dataset.to_table(filter=filter_expr)
+
+        if features is not None:
+            table = table.cast(features.arrow_schema)
+
+        return Dataset(table)


 def get_parquet_num_frames(parquet_path: str | Path) -> int:
@@ -339,7 +351,6 @@ def write_tasks(tasks: pandas.DataFrame, local_dir: Path) -> None:

 def load_tasks(local_dir: Path) -> pandas.DataFrame:
    tasks = pd.read_parquet(local_dir / DEFAULT_TASKS_PATH)
-    tasks.index.name = "task"
    return tasks


@@ -1232,7 +1243,7 @@ class LookAheadError(Exception):
    pass


-class Backtrackable[T]:
+class Backtrackable(Generic[T]):
    """
    Wrap any iterator/iterable so you can step back up to `history` items
    and look ahead up to `lookahead` items.
@@ -36,11 +36,8 @@ Convert a local dataset (works in place):
 ```bash
 python src/lerobot/datasets/v30/convert_dataset_v21_to_v30.py \
    --repo-id=lerobot/pusht \
-    --root=/path/to/local/dataset/directory \
+    --root=/path/to/local/dataset/directory
    --push-to-hub=false
-
-N.B. Path semantics (v2): --root is the exact dataset folder containing
-meta/, data/, videos/. When omitted, defaults to $HF_LEROBOT_HOME/{repo_id}.
 ```

 """
@@ -108,7 +105,7 @@ episodes.jsonl
 {"episode_index": 1, "tasks": ["Put the blue block in the green bowl"], "length": 266}

 NEW
-meta/episodes/chunk-000/file_000.parquet
+meta/episodes/chunk-000/episodes_000.parquet
 episode_index | video_chunk_index | video_file_index | data_chunk_index | data_file_index | tasks | length
 -------------------------
 OLD
@@ -116,16 +113,15 @@ tasks.jsonl
 {"task_index": 1, "task": "Put the blue block in the green bowl"}

 NEW
-meta/tasks.parquet
+meta/tasks/chunk-000/file_000.parquet
 task_index | task
 -------------------------
 OLD
 episodes_stats.jsonl
-{"episode_index": 1, "stats": {"feature_name": {"min": ..., "max": ..., "mean": ..., "std": ..., "count": ...}}}

 NEW
-meta/episodes/chunk-000/file_000.parquet
-episode_index | feature_name/min | feature_name/max | feature_name/mean | feature_name/std | feature_name/count
+meta/episodes_stats/chunk-000/file_000.parquet
+episode_index | mean | std | min | max
 -------------------------
 UPDATE
 meta/info.json
@@ -174,7 +170,7 @@ def convert_tasks(root, new_root):
    tasks, _ = legacy_load_tasks(root)
    task_indices = tasks.keys()
    task_strings = tasks.values()
-    df_tasks = pd.DataFrame({"task_index": task_indices}, index=pd.Index(task_strings, name="task"))
+    df_tasks = pd.DataFrame({"task_index": task_indices}, index=task_strings)
    write_tasks(df_tasks, new_root)


@@ -205,6 +201,7 @@ def convert_data(root: Path, new_root: Path, data_file_size_in_mb: int):

    image_keys = get_image_keys(root)

+    ep_idx = 0
    chunk_idx = 0
    file_idx = 0
    size_in_mb = 0
@@ -214,23 +211,9 @@ def convert_data(root: Path, new_root: Path, data_file_size_in_mb: int):

    logging.info(f"Converting data files from {len(ep_paths)} episodes")

-    for ep_idx, ep_path in enumerate(tqdm.tqdm(ep_paths, desc="convert data files")):
+    for ep_path in tqdm.tqdm(ep_paths, desc="convert data files"):
        ep_size_in_mb = get_parquet_file_size_in_mb(ep_path)
        ep_num_frames = get_parquet_num_frames(ep_path)
-
-        # Check if we need to start a new file BEFORE creating metadata
-        if size_in_mb + ep_size_in_mb >= data_file_size_in_mb and len(paths_to_cat) > 0:
-            # Write the accumulated data files
-            concat_data_files(paths_to_cat, new_root, chunk_idx, file_idx, image_keys)
-
-            # Move to next file
-            chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, DEFAULT_CHUNK_SIZE)
-
-            # Reset for the next file
-            size_in_mb = 0
-            paths_to_cat = []
-
-        # Now create metadata with correct chunk/file indices
        ep_metadata = {
            "episode_index": ep_idx,
            "data/chunk_index": chunk_idx,
@@ -241,7 +224,20 @@ def convert_data(root: Path, new_root: Path, data_file_size_in_mb: int):
        size_in_mb += ep_size_in_mb
        num_frames += ep_num_frames
        episodes_metadata.append(ep_metadata)
-        paths_to_cat.append(ep_path)
+        ep_idx += 1
+
+        if size_in_mb < data_file_size_in_mb:
+            paths_to_cat.append(ep_path)
+            continue
+
+        if paths_to_cat:
+            concat_data_files(paths_to_cat, new_root, chunk_idx, file_idx, image_keys)
+
+        # Reset for the next file
+        size_in_mb = ep_size_in_mb
+        paths_to_cat = [ep_path]
+
+        chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, DEFAULT_CHUNK_SIZE)

    # Write remaining data if any
    if paths_to_cat:
@@ -473,7 +469,7 @@ def convert_dataset(

    # Set root based on whether local dataset path is provided
    use_local_dataset = False
-    root = HF_LEROBOT_HOME / repo_id if root is None else Path(root)
+    root = HF_LEROBOT_HOME / repo_id if root is None else Path(root) / repo_id
    if root.exists():
        validate_local_dataset_version(root)
        use_local_dataset = True
@@ -533,7 +529,7 @@ if __name__ == "__main__":
        type=str,
        required=True,
        help="Repository identifier on Hugging Face: a community or a user name `/` the name of the dataset "
-        "(e.g. `lerobot/pusht`, `<USER>/aloha_sim_insertion_human`).",
+        "(e.g. `lerobot/pusht`, `cadene/aloha_sim_insertion_human`).",
    )
    parser.add_argument(
        "--branch",
@@ -557,7 +553,7 @@ if __name__ == "__main__":
        "--root",
        type=str,
        default=None,
-        help="Local directory to use for downloading/writing the dataset. Defaults to $HF_LEROBOT_HOME/repo_id.",
+        help="Local directory to use for downloading/writing the dataset.",
    )
    parser.add_argument(
        "--push-to-hub",
@@ -13,106 +13,25 @@
 # 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 contextlib
 import glob
 import importlib
 import logging
-import queue
 import shutil
 import tempfile
-import threading
 import warnings
 from dataclasses import dataclass, field
-from fractions import Fraction
 from pathlib import Path
 from threading import Lock
 from typing import Any, ClassVar

 import av
 import fsspec
-import numpy as np
 import pyarrow as pa
 import torch
 import torchvision
 from datasets.features.features import register_feature
 from PIL import Image

-# List of hardware encoders to probe for auto-selection. Availability depends on the platform and FFmpeg build.
-# Determines the order of preference for auto-selection when vcodec="auto" is used.
-HW_ENCODERS = [
-    "h264_videotoolbox",  # macOS
-    "hevc_videotoolbox",  # macOS
-    "h264_nvenc",  # NVIDIA GPU
-    "hevc_nvenc",  # NVIDIA GPU
-    "h264_vaapi",  # Linux Intel/AMD
-    "h264_qsv",  # Intel Quick Sync
-]
-
-VALID_VIDEO_CODECS = {"h264", "hevc", "libsvtav1", "auto"} | set(HW_ENCODERS)
-
-
-def _get_codec_options(
-    vcodec: str,
-    g: int | None = 2,
-    crf: int | None = 30,
-    preset: int | None = None,
-) -> dict:
-    """Build codec-specific options dict for video encoding."""
-    options = {}
-
-    # GOP size (keyframe interval) - supported by VideoToolbox and software encoders
-    if g is not None and (vcodec in ("h264_videotoolbox", "hevc_videotoolbox") or vcodec not in HW_ENCODERS):
-        options["g"] = str(g)
-
-    # Quality control (codec-specific parameter names)
-    if crf is not None:
-        if vcodec in ("h264", "hevc", "libsvtav1"):
-            options["crf"] = str(crf)
-        elif vcodec in ("h264_videotoolbox", "hevc_videotoolbox"):
-            quality = max(1, min(100, int(100 - crf * 2)))
-            options["q:v"] = str(quality)
-        elif vcodec in ("h264_nvenc", "hevc_nvenc"):
-            options["rc"] = "constqp"
-            options["qp"] = str(crf)
-        elif vcodec in ("h264_vaapi",):
-            options["qp"] = str(crf)
-        elif vcodec in ("h264_qsv",):
-            options["global_quality"] = str(crf)
-
-    # Preset (only for libsvtav1)
-    if vcodec == "libsvtav1":
-        options["preset"] = str(preset) if preset is not None else "12"
-
-    return options
-
-
-def detect_available_hw_encoders() -> list[str]:
-    """Probe PyAV/FFmpeg for available hardware video encoders."""
-    available = []
-    for codec_name in HW_ENCODERS:
-        try:
-            av.codec.Codec(codec_name, "w")
-            available.append(codec_name)
-        except Exception:  # nosec B110
-            pass  # nosec B110
-    return available
-
-
-def resolve_vcodec(vcodec: str) -> str:
-    """Validate vcodec and resolve 'auto' to best available HW encoder, fallback to libsvtav1."""
-    if vcodec not in VALID_VIDEO_CODECS:
-        raise ValueError(f"Invalid vcodec '{vcodec}'. Must be one of: {sorted(VALID_VIDEO_CODECS)}")
-    if vcodec != "auto":
-        logging.info(f"Using video codec: {vcodec}")
-        return vcodec
-    available = detect_available_hw_encoders()
-    for encoder in HW_ENCODERS:
-        if encoder in available:
-            logging.info(f"Auto-selected video codec: {encoder}")
-            return encoder
-    logging.info("No hardware encoder available, falling back to software encoder 'libsvtav1'")
-    return "libsvtav1"
-

 def get_safe_default_codec():
    if importlib.util.find_spec("torchcodec"):
@@ -227,17 +146,16 @@ def decode_video_frames_torchvision(
    min_, argmin_ = dist.min(1)

    is_within_tol = min_ < tolerance_s
-    if not is_within_tol.all():
-        raise FrameTimestampError(
-            f"One or several query timestamps unexpectedly violate the tolerance ({min_[~is_within_tol]} > {tolerance_s=})."
-            " It means that the closest frame that can be loaded from the video is too far away in time."
-            " This might be due to synchronization issues with timestamps during data collection."
-            " To be safe, we advise to ignore this item during training."
-            f"\nqueried timestamps: {query_ts}"
-            f"\nloaded timestamps: {loaded_ts}"
-            f"\nvideo: {video_path}"
-            f"\nbackend: {backend}"
-        )
+    assert is_within_tol.all(), (
+        f"One or several query timestamps unexpectedly violate the tolerance ({min_[~is_within_tol]} > {tolerance_s=})."
+        "It means that the closest frame that can be loaded from the video is too far away in time."
+        "This might be due to synchronization issues with timestamps during data collection."
+        "To be safe, we advise to ignore this item during training."
+        f"\nqueried timestamps: {query_ts}"
+        f"\nloaded timestamps: {loaded_ts}"
+        f"\nvideo: {video_path}"
+        f"\nbackend: {backend}"
+    )

    # get closest frames to the query timestamps
    closest_frames = torch.stack([loaded_frames[idx] for idx in argmin_])
@@ -249,11 +167,7 @@ def decode_video_frames_torchvision(
    # convert to the pytorch format which is float32 in [0,1] range (and channel first)
    closest_frames = closest_frames.type(torch.float32) / 255

-    if len(timestamps) != len(closest_frames):
-        raise FrameTimestampError(
-            f"Number of retrieved frames ({len(closest_frames)}) does not match "
-            f"number of queried timestamps ({len(timestamps)})"
-        )
+    assert len(timestamps) == len(closest_frames)
    return closest_frames


@@ -358,16 +272,15 @@ def decode_video_frames_torchcodec(
    min_, argmin_ = dist.min(1)

    is_within_tol = min_ < tolerance_s
-    if not is_within_tol.all():
-        raise FrameTimestampError(
-            f"One or several query timestamps unexpectedly violate the tolerance ({min_[~is_within_tol]} > {tolerance_s=})."
-            " It means that the closest frame that can be loaded from the video is too far away in time."
-            " This might be due to synchronization issues with timestamps during data collection."
-            " To be safe, we advise to ignore this item during training."
-            f"\nqueried timestamps: {query_ts}"
-            f"\nloaded timestamps: {loaded_ts}"
-            f"\nvideo: {video_path}"
-        )
+    assert is_within_tol.all(), (
+        f"One or several query timestamps unexpectedly violate the tolerance ({min_[~is_within_tol]} > {tolerance_s=})."
+        "It means that the closest frame that can be loaded from the video is too far away in time."
+        "This might be due to synchronization issues with timestamps during data collection."
+        "To be safe, we advise to ignore this item during training."
+        f"\nqueried timestamps: {query_ts}"
+        f"\nloaded timestamps: {loaded_ts}"
+        f"\nvideo: {video_path}"
+    )

    # get closest frames to the query timestamps
    closest_frames = torch.stack([loaded_frames[idx] for idx in argmin_])
@@ -396,13 +309,14 @@ def encode_video_frames(
    g: int | None = 2,
    crf: int | None = 30,
    fast_decode: int = 0,
-    log_level: int | None = av.logging.WARNING,
+    log_level: int | None = av.logging.ERROR,
    overwrite: bool = False,
    preset: int | None = None,
-    encoder_threads: int | None = None,
 ) -> None:
    """More info on ffmpeg arguments tuning on `benchmark/video/README.md`"""
-    vcodec = resolve_vcodec(vcodec)
+    # Check encoder availability
+    if vcodec not in ["h264", "hevc", "libsvtav1"]:
+        raise ValueError(f"Unsupported video codec: {vcodec}. Supported codecs are: h264, hevc, libsvtav1.")

    video_path = Path(video_path)
    imgs_dir = Path(imgs_dir)
@@ -433,22 +347,21 @@ def encode_video_frames(
        width, height = dummy_image.size

    # Define video codec options
-    video_options = _get_codec_options(vcodec, g, crf, preset)
+    video_options = {}
+
+    if g is not None:
+        video_options["g"] = str(g)
+
+    if crf is not None:
+        video_options["crf"] = str(crf)

    if fast_decode:
        key = "svtav1-params" if vcodec == "libsvtav1" else "tune"
        value = f"fast-decode={fast_decode}" if vcodec == "libsvtav1" else "fastdecode"
        video_options[key] = value

-    if encoder_threads is not None:
-        if vcodec == "libsvtav1":
-            lp_param = f"lp={encoder_threads}"
-            if "svtav1-params" in video_options:
-                video_options["svtav1-params"] += f":{lp_param}"
-            else:
-                video_options["svtav1-params"] = lp_param
-        else:
-            video_options["threads"] = str(encoder_threads)
+    if vcodec == "libsvtav1":
+        video_options["preset"] = str(preset) if preset is not None else "12"

    # Set logging level
    if log_level is not None:
@@ -567,348 +480,6 @@ def concatenate_video_files(
    Path(tmp_concatenate_path).unlink()


-class _CameraEncoderThread(threading.Thread):
-    """A thread that encodes video frames streamed via a queue into an MP4 file.
-
-    One instance is created per camera per episode. Frames are received as numpy arrays
-    from the main thread, encoded in real-time using PyAV (which releases the GIL during
-    encoding), and written to disk. Stats are computed incrementally using
-    RunningQuantileStats and returned via result_queue.
-    """
-
-    def __init__(
-        self,
-        video_path: Path,
-        fps: int,
-        vcodec: str,
-        pix_fmt: str,
-        g: int | None,
-        crf: int | None,
-        preset: int | None,
-        frame_queue: queue.Queue,
-        result_queue: queue.Queue,
-        stop_event: threading.Event,
-        encoder_threads: int | None = None,
-    ):
-        super().__init__(daemon=True)
-        self.video_path = video_path
-        self.fps = fps
-        self.vcodec = vcodec
-        self.pix_fmt = pix_fmt
-        self.g = g
-        self.crf = crf
-        self.preset = preset
-        self.frame_queue = frame_queue
-        self.result_queue = result_queue
-        self.stop_event = stop_event
-        self.encoder_threads = encoder_threads
-
-    def run(self) -> None:
-        from lerobot.datasets.compute_stats import RunningQuantileStats, auto_downsample_height_width
-
-        container = None
-        output_stream = None
-        stats_tracker = RunningQuantileStats()
-        frame_count = 0
-
-        try:
-            logging.getLogger("libav").setLevel(av.logging.WARNING)
-
-            while True:
-                try:
-                    frame_data = self.frame_queue.get(timeout=1)
-                except queue.Empty:
-                    if self.stop_event.is_set():
-                        break
-                    continue
-
-                if frame_data is None:
-                    # Sentinel: flush and close
-                    break
-
-                # Ensure HWC uint8 numpy array
-                if isinstance(frame_data, np.ndarray):
-                    if frame_data.ndim == 3 and frame_data.shape[0] == 3:
-                        # CHW -> HWC
-                        frame_data = frame_data.transpose(1, 2, 0)
-                    if frame_data.dtype != np.uint8:
-                        frame_data = (frame_data * 255).astype(np.uint8)
-
-                # Open container on first frame (to get width/height)
-                if container is None:
-                    height, width = frame_data.shape[:2]
-                    video_options = _get_codec_options(self.vcodec, self.g, self.crf, self.preset)
-                    if self.encoder_threads is not None:
-                        if self.vcodec == "libsvtav1":
-                            lp_param = f"lp={self.encoder_threads}"
-                            if "svtav1-params" in video_options:
-                                video_options["svtav1-params"] += f":{lp_param}"
-                            else:
-                                video_options["svtav1-params"] = lp_param
-                        else:
-                            video_options["threads"] = str(self.encoder_threads)
-                    Path(self.video_path).parent.mkdir(parents=True, exist_ok=True)
-                    container = av.open(str(self.video_path), "w")
-                    output_stream = container.add_stream(self.vcodec, self.fps, options=video_options)
-                    output_stream.pix_fmt = self.pix_fmt
-                    output_stream.width = width
-                    output_stream.height = height
-                    output_stream.time_base = Fraction(1, self.fps)
-
-                # Encode frame with explicit timestamps
-                pil_img = Image.fromarray(frame_data)
-                video_frame = av.VideoFrame.from_image(pil_img)
-                video_frame.pts = frame_count
-                video_frame.time_base = Fraction(1, self.fps)
-                packet = output_stream.encode(video_frame)
-                if packet:
-                    container.mux(packet)
-
-                # Update stats with downsampled frame (per-channel stats like compute_episode_stats)
-                img_chw = frame_data.transpose(2, 0, 1)  # HWC -> CHW
-                img_downsampled = auto_downsample_height_width(img_chw)
-                # Reshape CHW to (H*W, C) for per-channel stats
-                channels = img_downsampled.shape[0]
-                img_for_stats = img_downsampled.transpose(1, 2, 0).reshape(-1, channels)
-                stats_tracker.update(img_for_stats)
-
-                frame_count += 1
-
-            # Flush encoder
-            if output_stream is not None:
-                packet = output_stream.encode()
-                if packet:
-                    container.mux(packet)
-
-            if container is not None:
-                container.close()
-
-            av.logging.restore_default_callback()
-
-            # Get stats and put on result queue
-            if frame_count >= 2:
-                stats = stats_tracker.get_statistics()
-                self.result_queue.put(("ok", stats))
-            else:
-                self.result_queue.put(("ok", None))
-
-        except Exception as e:
-            logging.error(f"Encoder thread error: {e}")
-            if container is not None:
-                with contextlib.suppress(Exception):
-                    container.close()
-            self.result_queue.put(("error", str(e)))
-
-
-class StreamingVideoEncoder:
-    """Manages per-camera encoder threads for real-time video encoding during recording.
-
-    Instead of writing frames as PNG images and then encoding to MP4 at episode end,
-    this class streams frames directly to encoder threads, eliminating the
-    PNG round-trip and making save_episode() near-instant.
-
-    Uses threading instead of multiprocessing to avoid the overhead of pickling large
-    numpy arrays through multiprocessing.Queue. PyAV's encode() releases the GIL,
-    so encoding runs in parallel with the main recording loop.
-    """
-
-    def __init__(
-        self,
-        fps: int,
-        vcodec: str = "libsvtav1",
-        pix_fmt: str = "yuv420p",
-        g: int | None = 2,
-        crf: int | None = 30,
-        preset: int | None = None,
-        queue_maxsize: int = 30,
-        encoder_threads: int | None = None,
-    ):
-        self.fps = fps
-        self.vcodec = resolve_vcodec(vcodec)
-        self.pix_fmt = pix_fmt
-        self.g = g
-        self.crf = crf
-        self.preset = preset
-        self.queue_maxsize = queue_maxsize
-        self.encoder_threads = encoder_threads
-
-        self._frame_queues: dict[str, queue.Queue] = {}
-        self._result_queues: dict[str, queue.Queue] = {}
-        self._threads: dict[str, _CameraEncoderThread] = {}
-        self._stop_events: dict[str, threading.Event] = {}
-        self._video_paths: dict[str, Path] = {}
-        self._dropped_frames: dict[str, int] = {}
-        self._episode_active = False
-
-    def start_episode(self, video_keys: list[str], temp_dir: Path) -> None:
-        """Start encoder threads for a new episode.
-
-        Args:
-            video_keys: List of video feature keys (e.g. ["observation.images.laptop"])
-            temp_dir: Base directory for temporary MP4 files
-        """
-        if self._episode_active:
-            self.cancel_episode()
-
-        self._dropped_frames.clear()
-
-        for video_key in video_keys:
-            frame_queue: queue.Queue = queue.Queue(maxsize=self.queue_maxsize)
-            result_queue: queue.Queue = queue.Queue(maxsize=1)
-            stop_event = threading.Event()
-
-            temp_video_dir = Path(tempfile.mkdtemp(dir=temp_dir))
-            video_path = temp_video_dir / f"{video_key.replace('/', '_')}_streaming.mp4"
-
-            encoder_thread = _CameraEncoderThread(
-                video_path=video_path,
-                fps=self.fps,
-                vcodec=self.vcodec,
-                pix_fmt=self.pix_fmt,
-                g=self.g,
-                crf=self.crf,
-                preset=self.preset,
-                frame_queue=frame_queue,
-                result_queue=result_queue,
-                stop_event=stop_event,
-                encoder_threads=self.encoder_threads,
-            )
-            encoder_thread.start()
-
-            self._frame_queues[video_key] = frame_queue
-            self._result_queues[video_key] = result_queue
-            self._threads[video_key] = encoder_thread
-            self._stop_events[video_key] = stop_event
-            self._video_paths[video_key] = video_path
-
-        self._episode_active = True
-
-    def feed_frame(self, video_key: str, image: np.ndarray) -> None:
-        """Feed a frame to the encoder for a specific camera.
-
-        A copy of the image is made before enqueueing to prevent race conditions
-        with camera drivers that may reuse buffers. If the encoder queue is full
-        (encoder can't keep up), the frame is dropped with a warning instead of
-        crashing the recording session.
-
-        Args:
-            video_key: The video feature key
-            image: numpy array in (H,W,C) or (C,H,W) format, uint8 or float
-
-        Raises:
-            RuntimeError: If the encoder thread has crashed
-        """
-        if not self._episode_active:
-            raise RuntimeError("No active episode. Call start_episode() first.")
-
-        thread = self._threads[video_key]
-        if not thread.is_alive():
-            # Check for error
-            try:
-                status, msg = self._result_queues[video_key].get_nowait()
-                if status == "error":
-                    raise RuntimeError(f"Encoder thread for {video_key} crashed: {msg}")
-            except queue.Empty:
-                pass
-            raise RuntimeError(f"Encoder thread for {video_key} is not alive")
-
-        try:
-            self._frame_queues[video_key].put(image.copy(), timeout=0.1)
-        except queue.Full:
-            self._dropped_frames[video_key] = self._dropped_frames.get(video_key, 0) + 1
-            count = self._dropped_frames[video_key]
-            # Log periodically to avoid spam (1st, then every 10th)
-            if count == 1 or count % 10 == 0:
-                logging.warning(
-                    f"Encoder queue full for {video_key}, dropped {count} frame(s). "
-                    f"Consider using vcodec='auto' for hardware encoding or increasing encoder_queue_maxsize."
-                )
-
-    def finish_episode(self) -> dict[str, tuple[Path, dict | None]]:
-        """Finish encoding the current episode.
-
-        Sends sentinel values, waits for encoder threads to complete,
-        and collects results.
-
-        Returns:
-            Dict mapping video_key to (mp4_path, stats_dict_or_None)
-        """
-        if not self._episode_active:
-            raise RuntimeError("No active episode to finish.")
-
-        results = {}
-
-        # Report dropped frames
-        for video_key, count in self._dropped_frames.items():
-            if count > 0:
-                logging.warning(f"Episode finished with {count} dropped frame(s) for {video_key}.")
-
-        # Send sentinel to all queues
-        for video_key in self._frame_queues:
-            self._frame_queues[video_key].put(None)
-
-        # Wait for all threads and collect results
-        for video_key in self._threads:
-            self._threads[video_key].join(timeout=120)
-            if self._threads[video_key].is_alive():
-                logging.error(f"Encoder thread for {video_key} did not finish in time")
-                self._stop_events[video_key].set()
-                self._threads[video_key].join(timeout=5)
-                results[video_key] = (self._video_paths[video_key], None)
-                continue
-
-            try:
-                status, data = self._result_queues[video_key].get(timeout=5)
-                if status == "error":
-                    raise RuntimeError(f"Encoder thread for {video_key} failed: {data}")
-                results[video_key] = (self._video_paths[video_key], data)
-            except queue.Empty:
-                logging.error(f"No result from encoder thread for {video_key}")
-                results[video_key] = (self._video_paths[video_key], None)
-
-        self._cleanup()
-        self._episode_active = False
-        return results
-
-    def cancel_episode(self) -> None:
-        """Cancel the current episode, stopping encoder threads and cleaning up."""
-        if not self._episode_active:
-            return
-
-        # Signal all threads to stop
-        for video_key in self._stop_events:
-            self._stop_events[video_key].set()
-
-        # Wait for threads to finish
-        for video_key in self._threads:
-            self._threads[video_key].join(timeout=5)
-
-            # Clean up temp MP4 files
-            video_path = self._video_paths.get(video_key)
-            if video_path is not None and video_path.exists():
-                shutil.rmtree(str(video_path.parent), ignore_errors=True)
-
-        self._cleanup()
-        self._episode_active = False
-
-    def close(self) -> None:
-        """Close the encoder, canceling any in-progress episode."""
-        if self._episode_active:
-            self.cancel_episode()
-
-    def _cleanup(self) -> None:
-        """Clean up queues and thread tracking dicts."""
-        for q in self._frame_queues.values():
-            with contextlib.suppress(Exception):
-                while not q.empty():
-                    q.get_nowait()
-        self._frame_queues.clear()
-        self._result_queues.clear()
-        self._threads.clear()
-        self._stop_events.clear()
-        self._video_paths.clear()
-
-
@dataclass
 class VideoFrame:
    # TODO(rcadene, lhoestq): move to Hugging Face `datasets` repo
@@ -943,7 +514,7 @@ with warnings.catch_warnings():

 def get_audio_info(video_path: Path | str) -> dict:
    # Set logging level
-    logging.getLogger("libav").setLevel(av.logging.WARNING)
+    logging.getLogger("libav").setLevel(av.logging.ERROR)

    # Getting audio stream information
    audio_info = {}
@@ -975,7 +546,7 @@ def get_audio_info(video_path: Path | str) -> dict:

 def get_video_info(video_path: Path | str) -> dict:
    # Set logging level
-    logging.getLogger("libav").setLevel(av.logging.WARNING)
+    logging.getLogger("libav").setLevel(av.logging.ERROR)

    # Getting video stream information
    video_info = {}
@@ -1061,15 +632,8 @@ class VideoEncodingManager:
        return self

    def __exit__(self, exc_type, exc_val, exc_tb):
-        streaming_encoder = getattr(self.dataset, "_streaming_encoder", None)
-
-        if streaming_encoder is not None:
-            # Handle streaming encoder cleanup
-            if exc_type is not None:
-                streaming_encoder.cancel_episode()
-            streaming_encoder.close()
-        elif self.dataset.episodes_since_last_encoding > 0:
-            # Handle any remaining episodes that haven't been batch encoded
+        # Handle any remaining episodes that haven't been batch encoded
+        if self.dataset.episodes_since_last_encoding > 0:
            if exc_type is not None:
                logging.info("Exception occurred. Encoding remaining episodes before exit...")
            else:
@@ -1086,8 +650,8 @@ class VideoEncodingManager:
        # Finalize the dataset to properly close all writers
        self.dataset.finalize()

-        # Clean up episode images if recording was interrupted (only for non-streaming mode)
-        if exc_type is not None and streaming_encoder is None:
+        # Clean up episode images if recording was interrupted
+        if exc_type is not None:
            interrupted_episode_index = self.dataset.num_episodes
            for key in self.dataset.meta.video_keys:
                img_dir = self.dataset._get_image_file_path(
@@ -1101,12 +665,14 @@ class VideoEncodingManager:

        # Clean up any remaining images directory if it's empty
        img_dir = self.dataset.root / "images"
-        if img_dir.exists():
-            png_files = list(img_dir.rglob("*.png"))
-            if len(png_files) == 0:
+        # Check for any remaining PNG files
+        png_files = list(img_dir.rglob("*.png"))
+        if len(png_files) == 0:
+            # Only remove the images directory if no PNG files remain
+            if img_dir.exists():
                shutil.rmtree(img_dir)
                logging.debug("Cleaned up empty images directory")
-            else:
-                logging.debug(f"Images directory is not empty, containing {len(png_files)} PNG files")
+        else:
+            logging.debug(f"Images directory is not empty, containing {len(png_files)} PNG files")

        return False  # Don't suppress the original exception
@@ -29,7 +29,7 @@ from dataclasses import dataclass
 from enum import Enum
 from functools import cached_property
 from pprint import pformat
-from typing import Protocol
+from typing import Protocol, TypeAlias

 import serial
 from deepdiff import DeepDiff
@@ -38,8 +38,8 @@ from tqdm import tqdm
 from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
 from lerobot.utils.utils import enter_pressed, move_cursor_up

-type NameOrID = str | int
-type Value = int | float
+NameOrID: TypeAlias = str | int
+Value: TypeAlias = int | float

 logger = logging.getLogger(__name__)

@@ -1277,4 +1277,4 @@ class SerialMotorsBus(MotorsBusBase):


 # Backward compatibility alias
-MotorsBus = SerialMotorsBus
+MotorsBus: TypeAlias = SerialMotorsBus
@@ -1,18 +0,0 @@
-#!/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 .robstride import RobstrideMotorsBus
-from .tables import *
@@ -1,120 +0,0 @@
-# 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.
-
-"""Configuration tables for Damiao motors."""
-
-from enum import IntEnum
-
-
-# Motor type definitions
-class MotorType(IntEnum):
-    O0 = 0
-    O1 = 1
-    O2 = 2
-    O3 = 3
-    O4 = 4
-    O5 = 5
-    ELO5 = 6
-    O6 = 7
-
-
-class CommMode(IntEnum):
-    PrivateProtocole = 0
-    CANopen = 1
-    MIT = 2
-
-
-# Control modes
-class ControlMode(IntEnum):
-    MIT = 0
-    POS_VEL = 1
-    VEL = 2
-
-
-# Motor limit parameters [PMAX, VMAX, TMAX]
-# PMAX: Maximum position (rad)
-# VMAX: Maximum velocity (rad/s)
-# TMAX: Maximum torque (N·m)
-MOTOR_LIMIT_PARAMS: dict[MotorType, tuple[float, float, float]] = {
-    MotorType.O0: (12.57, 33, 14),
-    MotorType.O1: (12.57, 44, 17),
-    MotorType.O2: (12.57, 33, 20),
-    MotorType.O3: (12.57, 33, 60),
-    MotorType.O4: (12.57, 33, 120),
-    MotorType.O5: (12.57, 50, 5.5),
-    MotorType.ELO5: (12.57, 50, 6),
-    MotorType.O6: (112.5, 50, 36),
-}
-
-# Motor model names
-MODEL_NAMES = {
-    MotorType.O0: "O0",
-    MotorType.O1: "O1",
-    MotorType.O2: "O2",
-    MotorType.O3: "O3",
-    MotorType.O4: "O4",
-    MotorType.O5: "O5",
-    MotorType.ELO5: "ELO5",
-    MotorType.O6: "O6",
-}
-
-# Motor resolution table (encoder counts per revolution)
-MODEL_RESOLUTION = {
-    "O0": 65536,
-    "O1": 65536,
-    "O2": 65536,
-    "O3": 65536,
-    "O4": 65536,
-    "O5": 65536,
-    "ELO5": 65536,
-    "O6": 65536,
-}
-
-# CAN baudrates supported by Robstride motors
-AVAILABLE_BAUDRATES = [
-    1000000,  # 4: 1 mbps (default)
-]
-DEFAULT_BAUDRATE = 1000000
-
-# Default timeout in milliseconds
-DEFAULT_TIMEOUT_MS = 0  # disabled by default, otherwise 20000 is 1s
-
-
-# Data that should be normalized
-NORMALIZED_DATA = ["Present_Position", "Goal_Position"]
-
-
-# MIT control parameter ranges
-MIT_KP_RANGE = (0.0, 500.0)
-MIT_KD_RANGE = (0.0, 5.0)
-
-# CAN frame command IDs
-CAN_CMD_ENABLE = 0xFC
-CAN_CMD_DISABLE = 0xFD
-CAN_CMD_SET_ZERO = 0xFE
-CAN_CMD_CLEAR_FAULT = 0xFB
-
-
-CAN_CMD_QUERY_PARAM = 0x33
-CAN_CMD_WRITE_PARAM = 0x55
-CAN_CMD_SAVE_PARAM = 0xAA
-
-# CAN ID for parameter operations
-CAN_PARAM_ID = 0x7FF
-
-
-RUNNING_TIMEOUT = 0.001
-PARAM_TIMEOUT = 0.01
-
-STATE_CACHE_TTL_S = 0.02
@@ -55,16 +55,10 @@ class DiffusionConfig(PreTrainedConfig):
        normalization_mapping: A dictionary that maps from a str value of FeatureType (e.g., "STATE", "VISUAL") to
            a corresponding NormalizationMode (e.g., NormalizationMode.MIN_MAX)
        vision_backbone: Name of the torchvision resnet backbone to use for encoding images.
-        resize_shape: (H, W) shape to resize images to as a preprocessing step for the vision
-            backbone. If None, no resizing is done and the original image resolution is used.
-        crop_ratio: Ratio in (0, 1] used to derive the crop size from resize_shape
-            (crop_h = int(resize_shape[0] * crop_ratio), likewise for width).
-            Set to 1.0 to disable cropping. Only takes effect when resize_shape is not None.
-        crop_shape: (H, W) shape to crop images to. When resize_shape is set and crop_ratio < 1.0,
-            this is computed automatically. Can also be set directly for legacy configs that use
-            crop-only (without resize). If None and no derivation applies, no cropping is done.
-        crop_is_random: Whether the crop should be random at training time (it's always a center
-            crop in eval mode).
+        crop_shape: (H, W) shape to crop images to as a preprocessing step for the vision backbone. Must fit
+            within the image size. If None, no cropping is done.
+        crop_is_random: Whether the crop should be random at training time (it's always a center crop in eval
+            mode).
        pretrained_backbone_weights: Pretrained weights from torchvision to initialize the backbone.
            `None` means no pretrained weights.
        use_group_norm: Whether to replace batch normalization with group normalization in the backbone.
@@ -120,9 +114,7 @@ class DiffusionConfig(PreTrainedConfig):
    # Architecture / modeling.
    # Vision backbone.
    vision_backbone: str = "resnet18"
-    resize_shape: tuple[int, int] | None = None
-    crop_ratio: float = 1.0
-    crop_shape: tuple[int, int] | None = None
+    crop_shape: tuple[int, int] | None = (84, 84)
    crop_is_random: bool = True
    pretrained_backbone_weights: str | None = None
    use_group_norm: bool = True
@@ -147,10 +139,6 @@ class DiffusionConfig(PreTrainedConfig):
    # Inference
    num_inference_steps: int | None = None

-    # Optimization
-    compile_model: bool = False
-    compile_mode: str = "reduce-overhead"
-
    # Loss computation
    do_mask_loss_for_padding: bool = False

@@ -183,25 +171,6 @@ class DiffusionConfig(PreTrainedConfig):
                f"Got {self.noise_scheduler_type}."
            )

-        if self.resize_shape is not None and (
-            len(self.resize_shape) != 2 or any(d <= 0 for d in self.resize_shape)
-        ):
-            raise ValueError(f"`resize_shape` must be a pair of positive integers. Got {self.resize_shape}.")
-        if not (0 < self.crop_ratio <= 1.0):
-            raise ValueError(f"`crop_ratio` must be in (0, 1]. Got {self.crop_ratio}.")
-
-        if self.resize_shape is not None:
-            if self.crop_ratio < 1.0:
-                self.crop_shape = (
-                    int(self.resize_shape[0] * self.crop_ratio),
-                    int(self.resize_shape[1] * self.crop_ratio),
-                )
-            else:
-                # Explicitly disable cropping for resize+ratio path when crop_ratio == 1.0.
-                self.crop_shape = None
-        if self.crop_shape is not None and (self.crop_shape[0] <= 0 or self.crop_shape[1] <= 0):
-            raise ValueError(f"`crop_shape` must have positive dimensions. Got {self.crop_shape}.")
-
        # Check that the horizon size and U-Net downsampling is compatible.
        # U-Net downsamples by 2 with each stage.
        downsampling_factor = 2 ** len(self.down_dims)
@@ -229,12 +198,13 @@ class DiffusionConfig(PreTrainedConfig):
        if len(self.image_features) == 0 and self.env_state_feature is None:
            raise ValueError("You must provide at least one image or the environment state among the inputs.")

-        if self.resize_shape is None and self.crop_shape is not None:
+        if self.crop_shape is not None:
            for key, image_ft in self.image_features.items():
                if self.crop_shape[0] > image_ft.shape[1] or self.crop_shape[1] > image_ft.shape[2]:
                    raise ValueError(
-                        f"`crop_shape` should fit within the image shapes. Got {self.crop_shape} "
-                        f"for `crop_shape` and {image_ft.shape} for `{key}`."
+                        f"`crop_shape` should fit within the images shapes. Got {self.crop_shape} "
+                        f"for `crop_shape` and {image_ft.shape} for "
+                        f"`{key}`."
                    )

        # Check that all input images have the same shape.
@@ -142,9 +142,6 @@ class DiffusionPolicy(PreTrainedPolicy):
        """Run the batch through the model and compute the loss for training or validation."""
        if self.config.image_features:
            batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
-            for key in self.config.image_features:
-                if self.config.n_obs_steps == 1 and batch[key].ndim == 4:
-                    batch[key] = batch[key].unsqueeze(1)
            batch[OBS_IMAGES] = torch.stack([batch[key] for key in self.config.image_features], dim=-4)
        loss = self.diffusion.compute_loss(batch)
        # no output_dict so returning None
@@ -185,11 +182,6 @@ class DiffusionModel(nn.Module):

        self.unet = DiffusionConditionalUnet1d(config, global_cond_dim=global_cond_dim * config.n_obs_steps)

-        if config.compile_model:
-            # Compile the U-Net. "reduce-overhead" is preferred for the small-batch repetitive loops
-            # common in diffusion inference.
-            self.unet = torch.compile(self.unet, mode=config.compile_mode)
-
        self.noise_scheduler = _make_noise_scheduler(
            config.noise_scheduler_type,
            num_train_timesteps=config.num_train_timesteps,
@@ -454,18 +446,12 @@ class DiffusionRgbEncoder(nn.Module):
    def __init__(self, config: DiffusionConfig):
        super().__init__()
        # Set up optional preprocessing.
-        if config.resize_shape is not None:
-            self.resize = torchvision.transforms.Resize(config.resize_shape)
-        else:
-            self.resize = None
-
-        crop_shape = config.crop_shape
-        if crop_shape is not None:
+        if config.crop_shape is not None:
            self.do_crop = True
            # Always use center crop for eval
-            self.center_crop = torchvision.transforms.CenterCrop(crop_shape)
+            self.center_crop = torchvision.transforms.CenterCrop(config.crop_shape)
            if config.crop_is_random:
-                self.maybe_random_crop = torchvision.transforms.RandomCrop(crop_shape)
+                self.maybe_random_crop = torchvision.transforms.RandomCrop(config.crop_shape)
            else:
                self.maybe_random_crop = self.center_crop
        else:
@@ -491,16 +477,13 @@ class DiffusionRgbEncoder(nn.Module):

        # Set up pooling and final layers.
        # Use a dry run to get the feature map shape.
-        # The dummy shape mirrors the runtime preprocessing order: resize -> crop.
+        # The dummy input should take the number of image channels from `config.image_features` and it should
+        # use the height and width from `config.crop_shape` if it is provided, otherwise it should use the
+        # height and width from `config.image_features`.

        # Note: we have a check in the config class to make sure all images have the same shape.
        images_shape = next(iter(config.image_features.values())).shape
-        if config.crop_shape is not None:
-            dummy_shape_h_w = config.crop_shape
-        elif config.resize_shape is not None:
-            dummy_shape_h_w = config.resize_shape
-        else:
-            dummy_shape_h_w = images_shape[1:]
+        dummy_shape_h_w = config.crop_shape if config.crop_shape is not None else images_shape[1:]
        dummy_shape = (1, images_shape[0], *dummy_shape_h_w)
        feature_map_shape = get_output_shape(self.backbone, dummy_shape)[1:]

@@ -516,10 +499,7 @@ class DiffusionRgbEncoder(nn.Module):
        Returns:
            (B, D) image feature.
        """
-        # Preprocess: resize if configured, then crop if configured.
-
-        if self.resize is not None:
-            x = self.resize(x)
+        # Preprocess: maybe crop (if it was set up in the __init__).
        if self.do_crop:
            if self.training:  # noqa: SIM108
                x = self.maybe_random_crop(x)
@@ -18,9 +18,10 @@ from __future__ import annotations

 import importlib
 import logging
-from typing import Any, TypedDict, Unpack
+from typing import Any, TypedDict

 import torch
+from typing_extensions import Unpack

 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.configs.types import FeatureType
@@ -469,6 +470,13 @@ def make_policy(
    cfg.output_features = {key: ft for key, ft in features.items() if ft.type is FeatureType.ACTION}
    if not cfg.input_features:
        cfg.input_features = {key: ft for key, ft in features.items() if key not in cfg.output_features}
+
+    # Store action feature names for delta_exclude_joints support
+    if ds_meta is not None and hasattr(cfg, "action_feature_names"):
+        action_names = ds_meta.features.get(ACTION, {}).get("names")
+        if action_names is not None:
+            cfg.action_feature_names = list(action_names)
+
    kwargs["config"] = cfg

    # Pass dataset_stats to the policy if available (needed for some policies like SARM)
@@ -4,16 +4,17 @@
 # Licensed under The MIT License [see LICENSE for details]
 # --------------------------------------------------------

-from __future__ import annotations

 # copy from https://github.com/huggingface/transformers/blob/main/src/transformers/models/llava_onevision/image_processing_llava_onevision_fast.py
+from typing import Optional
+
 from transformers.image_processing_utils import (
    BatchFeature,
    get_patch_output_size,
 )
 from transformers.image_processing_utils_fast import (
    BaseImageProcessorFast,
-    ImagesKwargs,
+    DefaultFastImageProcessorKwargs,
    group_images_by_shape,
    reorder_images,
 )
@@ -76,7 +77,7 @@ def crop(img: torch.Tensor, left: int, top: int, right: int, bottom: int) -> tor
    return img[:, top:bottom, left:right]


-class Eagle25VLFastImageProcessorKwargs(ImagesKwargs):
+class Eagle25VLFastImageProcessorKwargs(DefaultFastImageProcessorKwargs):
    max_dynamic_tiles: int | None
    min_dynamic_tiles: int | None
    use_thumbnail: bool | None
@@ -164,11 +165,11 @@ class Eagle25VLImageProcessorFast(BaseImageProcessorFast):

    def _resize_for_patching(
        self,
-        image: torch.Tensor,
+        image: "torch.Tensor",
        target_resolution: tuple,
-        interpolation: F.InterpolationMode,
+        interpolation: "F.InterpolationMode",
        input_data_format: ChannelDimension,
-    ) -> torch.Tensor:
+    ) -> "torch.Tensor":
        """
        Resizes an image to a target resolution while maintaining aspect ratio.

@@ -218,8 +219,8 @@ class Eagle25VLImageProcessorFast(BaseImageProcessorFast):
        return best_ratio

    def _pad_for_patching(
-        self, image: torch.Tensor, target_resolution: tuple, input_data_format: ChannelDimension
-    ) -> torch.Tensor:
+        self, image: "torch.Tensor", target_resolution: tuple, input_data_format: ChannelDimension
+    ) -> "torch.Tensor":
        """
        Pad an image to a target resolution while maintaining aspect ratio.
        """
@@ -235,15 +236,15 @@ class Eagle25VLImageProcessorFast(BaseImageProcessorFast):

    def _get_image_patches(
        self,
-        image: torch.Tensor,
+        image: "torch.Tensor",
        min_num: int,
        max_num: int,
        size: tuple,
        tile_size: int,
        use_thumbnail: bool,
-        interpolation: F.InterpolationMode,
+        interpolation: "F.InterpolationMode",
        pad_during_tiling: bool,
-    ) -> list[torch.Tensor]:
+    ) -> list["torch.Tensor"]:
        image_size = get_image_size(image, channel_dim=ChannelDimension.FIRST)
        orig_height, orig_width = image_size
        aspect_ratio = orig_width / orig_height
@@ -304,8 +305,8 @@ class Eagle25VLImageProcessorFast(BaseImageProcessorFast):

    def _pad_for_batching(
        self,
-        pixel_values: list[torch.Tensor],
-    ) -> list[torch.Tensor]:
+        pixel_values: list["torch.Tensor"],
+    ) -> list["torch.Tensor"]:
        """
        Pads images on the `num_of_patches` dimension with zeros to form a batch of same number of patches.

@@ -326,14 +327,14 @@ class Eagle25VLImageProcessorFast(BaseImageProcessorFast):

    def _preprocess(
        self,
-        images: list[torch.Tensor],
+        images: list["torch.Tensor"],
        do_resize: bool,
        size: SizeDict,
        max_dynamic_tiles: int,
        min_dynamic_tiles: int,
        use_thumbnail: bool,
        pad_during_tiling: bool,
-        interpolation: F.InterpolationMode | None,
+        interpolation: Optional["F.InterpolationMode"],
        do_center_crop: bool,
        crop_size: SizeDict,
        do_rescale: bool,
@@ -50,6 +50,13 @@ class PI0Config(PreTrainedConfig):
    min_period: float = 4e-3
    max_period: float = 4.0

+    # Delta actions: converts absolute actions to delta (relative to state).
+    use_delta_actions: bool = False
+    # Joint names to exclude from delta (kept absolute). Empty list = all dims delta.
+    delta_exclude_joints: list[str] = field(default_factory=lambda: ["gripper"])
+    # Populated at runtime from dataset metadata by make_policy.
+    action_feature_names: list[str] | None = None
+
    # Real-Time Chunking (RTC) configuration
    rtc_config: RTCConfig | None = None

@@ -15,16 +15,16 @@
 # limitations under the License.

 import builtins
-import copy
 import logging
 import math
 from collections import deque
 from pathlib import Path
-from typing import TYPE_CHECKING, Literal, TypedDict, Unpack
+from typing import TYPE_CHECKING, Literal, TypedDict

 import torch
 import torch.nn.functional as F  # noqa: N812
 from torch import Tensor, nn
+from typing_extensions import Unpack

 from lerobot.utils.import_utils import _transformers_available

@@ -32,21 +32,13 @@ from lerobot.utils.import_utils import _transformers_available
 if TYPE_CHECKING or _transformers_available:
    from transformers.models.auto import CONFIG_MAPPING
    from transformers.models.gemma import modeling_gemma
-
-    from lerobot.policies.pi_gemma import (
-        PaliGemmaForConditionalGenerationWithPiGemma,
-        PiGemmaForCausalLM,
-        _gated_residual,
-        layernorm_forward,
-    )
+    from transformers.models.gemma.modeling_gemma import GemmaForCausalLM
+    from transformers.models.paligemma.modeling_paligemma import PaliGemmaForConditionalGeneration
 else:
    CONFIG_MAPPING = None
    modeling_gemma = None
-    PiGemmaForCausalLM = None
-    _gated_residual = None
-    layernorm_forward = None
-    PaliGemmaForConditionalGenerationWithPiGemma = None
-
+    GemmaForCausalLM = None
+    PaliGemmaForConditionalGeneration = None

 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.policies.pi0.configuration_pi0 import DEFAULT_IMAGE_SIZE, PI0Config
@@ -199,7 +191,7 @@ def resize_with_pad_torch(  # see openpi `resize_with_pad_torch` (exact copy)
    if images.dtype == torch.uint8:
        resized_images = torch.round(resized_images).clamp(0, 255).to(torch.uint8)
    elif images.dtype == torch.float32:
-        resized_images = resized_images.clamp(0.0, 1.0)
+        resized_images = resized_images.clamp(-1.0, 1.0)
    else:
        raise ValueError(f"Unsupported image dtype: {images.dtype}")

@@ -210,7 +202,7 @@ def resize_with_pad_torch(  # see openpi `resize_with_pad_torch` (exact copy)
    pad_w1 = pad_w0 + remainder_w

    # Pad
-    constant_value = 0 if images.dtype == torch.uint8 else 0.0
+    constant_value = 0 if images.dtype == torch.uint8 else -1.0
    padded_images = F.pad(
        resized_images,
        (pad_w0, pad_w1, pad_h0, pad_h1),  # left, right, top, bottom
@@ -229,14 +221,14 @@ def resize_with_pad_torch(  # see openpi `resize_with_pad_torch` (exact copy)
 def compute_layer_complete(
    layer_idx, inputs_embeds, attention_mask, position_ids, adarms_cond, paligemma, gemma_expert
 ):
-    models = [paligemma.model.language_model, gemma_expert.model]
+    models = [paligemma.language_model, gemma_expert.model]
    query_states = []
    key_states = []
    value_states = []
    gates = []
    for i, hidden_states in enumerate(inputs_embeds):
        layer = models[i].layers[layer_idx]
-        hidden_states, gate = layernorm_forward(layer.input_layernorm, hidden_states, adarms_cond[i])
+        hidden_states, gate = layer.input_layernorm(hidden_states, cond=adarms_cond[i])  # noqa: PLW2901
        gates.append(gate)
        input_shape = hidden_states.shape[:-1]
        hidden_shape = (*input_shape, -1, layer.self_attn.head_dim)
@@ -262,10 +254,10 @@ def compute_layer_complete(
        query_states, key_states, cos, sin, unsqueeze_dim=1
    )
    batch_size = query_states.shape[0]
-    scaling = paligemma.model.language_model.layers[layer_idx].self_attn.scaling
+    scaling = paligemma.language_model.layers[layer_idx].self_attn.scaling
    # Attention computation
    att_output, _ = modeling_gemma.eager_attention_forward(
-        paligemma.model.language_model.layers[layer_idx].self_attn,
+        paligemma.language_model.layers[layer_idx].self_attn,
        query_states,
        key_states,
        value_states,
@@ -273,7 +265,7 @@ def compute_layer_complete(
        scaling,
    )
    # Get head_dim from the current layer, not from the model
-    head_dim = paligemma.model.language_model.layers[layer_idx].self_attn.head_dim
+    head_dim = paligemma.language_model.layers[layer_idx].self_attn.head_dim
    att_output = att_output.reshape(batch_size, -1, 1 * 8 * head_dim)
    # Process layer outputs
    outputs_embeds = []
@@ -285,15 +277,15 @@ def compute_layer_complete(
            att_output = att_output.to(layer.self_attn.o_proj.weight.dtype)
        out_emb = layer.self_attn.o_proj(att_output[:, start_pos:end_pos])
        # first residual
-        out_emb = _gated_residual(hidden_states, out_emb, gates[i])
+        out_emb = modeling_gemma._gated_residual(hidden_states, out_emb, gates[i])  # noqa: SLF001
        after_first_residual = out_emb.clone()
-        out_emb, gate = layernorm_forward(layer.post_attention_layernorm, out_emb, adarms_cond[i])
+        out_emb, gate = layer.post_attention_layernorm(out_emb, cond=adarms_cond[i])
        # Convert to bfloat16 if the next layer (mlp) uses bfloat16
        if layer.mlp.up_proj.weight.dtype == torch.bfloat16:
            out_emb = out_emb.to(dtype=torch.bfloat16)
        out_emb = layer.mlp(out_emb)
        # second residual
-        out_emb = _gated_residual(after_first_residual, out_emb, gate)
+        out_emb = modeling_gemma._gated_residual(after_first_residual, out_emb, gate)  # noqa: SLF001
        outputs_embeds.append(out_emb)
        start_pos = end_pos
    return outputs_embeds
@@ -366,7 +358,7 @@ class PaliGemmaWithExpertModel(
        vlm_config_hf.text_config.num_hidden_layers = vlm_config.depth
        vlm_config_hf.text_config.num_key_value_heads = vlm_config.num_kv_heads
        vlm_config_hf.text_config.hidden_activation = "gelu_pytorch_tanh"
-        vlm_config_hf.text_config.dtype = "float32"
+        vlm_config_hf.text_config.torch_dtype = "float32"
        vlm_config_hf.text_config.vocab_size = 257152
        vlm_config_hf.text_config.use_adarms = use_adarms[0]
        vlm_config_hf.text_config.adarms_cond_dim = vlm_config.width if use_adarms[0] else None
@@ -374,7 +366,7 @@ class PaliGemmaWithExpertModel(
        vlm_config_hf.vision_config.intermediate_size = 4304
        vlm_config_hf.vision_config.projection_dim = 2048
        vlm_config_hf.vision_config.projector_hidden_act = "gelu_fast"
-        vlm_config_hf.vision_config.dtype = "float32"
+        vlm_config_hf.vision_config.torch_dtype = "float32"

        action_expert_config_hf = CONFIG_MAPPING["gemma"](
            head_dim=action_expert_config.head_dim,
@@ -385,13 +377,13 @@ class PaliGemmaWithExpertModel(
            num_key_value_heads=action_expert_config.num_kv_heads,
            vocab_size=257152,
            hidden_activation="gelu_pytorch_tanh",
-            dtype="float32",
+            torch_dtype="float32",
            use_adarms=use_adarms[1],
            adarms_cond_dim=action_expert_config.width if use_adarms[1] else None,
        )

-        self.paligemma = PaliGemmaForConditionalGenerationWithPiGemma(config=vlm_config_hf)
-        self.gemma_expert = PiGemmaForCausalLM(config=action_expert_config_hf)
+        self.paligemma = PaliGemmaForConditionalGeneration(config=vlm_config_hf)
+        self.gemma_expert = GemmaForCausalLM(config=action_expert_config_hf)
        self.gemma_expert.model.embed_tokens = None

        self.to_bfloat16_for_selected_params(precision)
@@ -406,11 +398,10 @@ class PaliGemmaWithExpertModel(
        else:
            raise ValueError(f"Invalid precision: {precision}")

-        # Keep full vision path in float32 so we never toggle (toggle causes optimizer
-        # "same dtype" error). Align with PI05.
        params_to_keep_float32 = [
-            "vision_tower",
-            "multi_modal_projector",
+            "vision_tower.vision_model.embeddings.patch_embedding.weight",
+            "vision_tower.vision_model.embeddings.patch_embedding.bias",
+            "vision_tower.vision_model.embeddings.position_embedding.weight",
            "input_layernorm",
            "post_attention_layernorm",
            "model.norm",
@@ -422,8 +413,8 @@ class PaliGemmaWithExpertModel(

    def _set_requires_grad(self):
        if self.freeze_vision_encoder:
-            self.paligemma.model.vision_tower.eval()
-            for param in self.paligemma.model.vision_tower.parameters():
+            self.paligemma.vision_tower.eval()
+            for param in self.paligemma.vision_tower.parameters():
                param.requires_grad = False
        if self.train_expert_only:
            self.paligemma.eval()
@@ -433,23 +424,15 @@ class PaliGemmaWithExpertModel(
    def train(self, mode: bool = True):
        super().train(mode)
        if self.freeze_vision_encoder:
-            self.paligemma.model.vision_tower.eval()
+            self.paligemma.vision_tower.eval()
        if self.train_expert_only:
            self.paligemma.eval()

    def embed_image(self, image: torch.Tensor):
-        # Vision tower and multi_modal_projector are kept in float32 (params_to_keep_float32). Align with PI05.
-        out_dtype = image.dtype
-        if image.dtype != torch.float32:
-            image = image.to(torch.float32)
-        image_outputs = self.paligemma.model.get_image_features(image)
-        features = image_outputs.pooler_output * self.paligemma.config.text_config.hidden_size**0.5
-        if features.dtype != out_dtype:
-            features = features.to(out_dtype)
-        return features
+        return self.paligemma.model.get_image_features(image)

    def embed_language_tokens(self, tokens: torch.Tensor):
-        return self.paligemma.model.language_model.embed_tokens(tokens)
+        return self.paligemma.language_model.embed_tokens(tokens)

    def forward(
        self,
@@ -463,7 +446,7 @@ class PaliGemmaWithExpertModel(
        if adarms_cond is None:
            adarms_cond = [None, None]
        if inputs_embeds[1] is None:
-            prefix_output = self.paligemma.model.language_model.forward(
+            prefix_output = self.paligemma.language_model.forward(
                inputs_embeds=inputs_embeds[0],
                attention_mask=attention_mask,
                position_ids=position_ids,
@@ -487,7 +470,7 @@ class PaliGemmaWithExpertModel(
            prefix_output = None
            prefix_past_key_values = None
        else:
-            models = [self.paligemma.model.language_model, self.gemma_expert.model]
+            models = [self.paligemma.language_model, self.gemma_expert.model]
            num_layers = self.paligemma.config.text_config.num_hidden_layers

            # Check if gradient checkpointing is enabled for any of the models
@@ -527,7 +510,7 @@ class PaliGemmaWithExpertModel(
            def compute_final_norms(inputs_embeds, adarms_cond):
                outputs_embeds = []
                for i, hidden_states in enumerate(inputs_embeds):
-                    out_emb, _ = layernorm_forward(models[i].norm, hidden_states, adarms_cond[i])
+                    out_emb, _ = models[i].norm(hidden_states, cond=adarms_cond[i])
                    outputs_embeds.append(out_emb)
                return outputs_embeds

@@ -593,19 +576,29 @@ class PI0Pytorch(nn.Module):  # see openpi `PI0Pytorch`
            # Also compile the main forward pass used during training
            self.forward = torch.compile(self.forward, mode=config.compile_mode)

+        msg = """An incorrect transformer version is used, please create an issue on https://github.com/huggingface/lerobot/issues"""
+
+        try:
+            from transformers.models.siglip import check
+
+            if not check.check_whether_transformers_replace_is_installed_correctly():
+                raise ValueError(msg)
+        except ImportError:
+            raise ValueError(msg) from None
+
    def gradient_checkpointing_enable(self):
        """Enable gradient checkpointing for memory optimization."""
        self.gradient_checkpointing_enabled = True
-        self.paligemma_with_expert.paligemma.model.language_model.gradient_checkpointing = True
-        self.paligemma_with_expert.paligemma.model.vision_tower.gradient_checkpointing = True
+        self.paligemma_with_expert.paligemma.language_model.gradient_checkpointing = True
+        self.paligemma_with_expert.paligemma.vision_tower.gradient_checkpointing = True
        self.paligemma_with_expert.gemma_expert.model.gradient_checkpointing = True
        logging.info("Enabled gradient checkpointing for PI0Pytorch model")

    def gradient_checkpointing_disable(self):
        """Disable gradient checkpointing."""
        self.gradient_checkpointing_enabled = False
-        self.paligemma_with_expert.paligemma.model.language_model.gradient_checkpointing = False
-        self.paligemma_with_expert.paligemma.model.vision_tower.gradient_checkpointing = False
+        self.paligemma_with_expert.paligemma.language_model.gradient_checkpointing = False
+        self.paligemma_with_expert.paligemma.vision_tower.gradient_checkpointing = False
        self.paligemma_with_expert.gemma_expert.model.gradient_checkpointing = False
        logging.info("Disabled gradient checkpointing for PI0Pytorch model")

@@ -767,7 +760,7 @@ class PI0Pytorch(nn.Module):  # see openpi `PI0Pytorch`
        suffix_embs, suffix_pad_masks, suffix_att_masks, adarms_cond = self.embed_suffix(state, x_t, time)

        if (
-            self.paligemma_with_expert.paligemma.model.language_model.layers[0].self_attn.q_proj.weight.dtype
+            self.paligemma_with_expert.paligemma.language_model.layers[0].self_attn.q_proj.weight.dtype
            == torch.bfloat16
        ):
            suffix_embs = suffix_embs.to(dtype=torch.bfloat16)
@@ -841,7 +834,7 @@ class PI0Pytorch(nn.Module):  # see openpi `PI0Pytorch`
        prefix_position_ids = torch.cumsum(prefix_pad_masks, dim=1) - 1

        prefix_att_2d_masks_4d = self._prepare_attention_masks_4d(prefix_att_2d_masks)
-        self.paligemma_with_expert.paligemma.model.language_model.config._attn_implementation = "eager"  # noqa: SLF001
+        self.paligemma_with_expert.paligemma.language_model.config._attn_implementation = "eager"  # noqa: SLF001

        _, past_key_values = self.paligemma_with_expert.forward(
            attention_mask=prefix_att_2d_masks_4d,
@@ -915,7 +908,6 @@ class PI0Pytorch(nn.Module):  # see openpi `PI0Pytorch`
        full_att_2d_masks_4d = self._prepare_attention_masks_4d(full_att_2d_masks)
        self.paligemma_with_expert.gemma_expert.model.config._attn_implementation = "eager"  # noqa: SLF001

-        past_key_values = copy.deepcopy(past_key_values)
        outputs_embeds, _ = self.paligemma_with_expert.forward(
            attention_mask=full_att_2d_masks_4d,
            position_ids=position_ids,
@@ -1005,12 +997,14 @@ class PI0Policy(PreTrainedPolicy):
        # Check if dataset_stats were provided in kwargs
        model = cls(config, **kwargs)

-        # Load state dict (expects keys with "model." prefix)
+        # Now manually load and remap the state dict
        try:
+            # Try to load the pytorch_model.bin or model.safetensors file
            print(f"Loading model from: {pretrained_name_or_path}")
            try:
                from transformers.utils import cached_file

+                # Try safetensors first
                resolved_file = cached_file(
                    pretrained_name_or_path,
                    "model.safetensors",
@@ -1018,7 +1012,7 @@ class PI0Policy(PreTrainedPolicy):
                    force_download=kwargs.get("force_download", False),
                    resume_download=kwargs.get("resume_download"),
                    proxies=kwargs.get("proxies"),
-                    token=kwargs.get("token"),
+                    use_auth_token=kwargs.get("use_auth_token"),
                    revision=kwargs.get("revision"),
                    local_files_only=kwargs.get("local_files_only", False),
                )
@@ -1031,7 +1025,7 @@ class PI0Policy(PreTrainedPolicy):
                print("Returning model without loading pretrained weights")
                return model

-            # First, fix any key differences (see openpi model.py, _fix_pytorch_state_dict_keys)
+            # First, fix any key differences # see openpi `model.py, _fix_pytorch_state_dict_keys`
            fixed_state_dict = model._fix_pytorch_state_dict_keys(original_state_dict, model.config)

            # Then add "model." prefix for all keys that don't already have it
@@ -1076,7 +1070,7 @@ class PI0Policy(PreTrainedPolicy):
                print("All keys loaded successfully!")

        except Exception as e:
-            print(f"Warning: Could not load state dict: {e}")
+            print(f"Warning: Could not remap state dict keys: {e}")

        return model

@@ -1126,14 +1120,6 @@ class PI0Policy(PreTrainedPolicy):
                # Some checkpoints might have this, but current model expects different structure
                logging.warning(f"Vision embedding key might need handling: {key}")

-            if (
-                key == "model.paligemma_with_expert.paligemma.lm_head.weight"
-                or key == "paligemma_with_expert.paligemma.lm_head.weight"
-            ):
-                fixed_state_dict[
-                    "model.paligemma_with_expert.paligemma.model.language_model.embed_tokens.weight"
-                ] = value.clone()
-
            fixed_state_dict[new_key] = value

        return fixed_state_dict
@@ -21,8 +21,10 @@ import torch
 from lerobot.configs.types import PipelineFeatureType, PolicyFeature
 from lerobot.policies.pi0.configuration_pi0 import PI0Config
 from lerobot.processor import (
+    AbsoluteActionsProcessorStep,
    AddBatchDimensionProcessorStep,
    ComplementaryDataProcessorStep,
+    DeltaActionsProcessorStep,
    DeviceProcessorStep,
    NormalizerProcessorStep,
    PolicyAction,
@@ -126,7 +128,13 @@ def make_pi0_pre_post_processors(
        A tuple containing the configured pre-processor and post-processor pipelines.
    """

-    # Add remaining processors
+    delta_step = DeltaActionsProcessorStep(
+        enabled=config.use_delta_actions,
+        exclude_joints=getattr(config, "delta_exclude_joints", []),
+        action_names=getattr(config, "action_feature_names", None),
+    )
+
+    # OpenPI order: raw → delta → normalize → model → unnormalize → absolute
    input_steps: list[ProcessorStep] = [
        RenameObservationsProcessorStep(rename_map={}),  # To mimic the same processor as pretrained one
        AddBatchDimensionProcessorStep(),
@@ -138,6 +146,7 @@ def make_pi0_pre_post_processors(
            padding="max_length",
        ),
        DeviceProcessorStep(device=config.device),
+        delta_step,
        NormalizerProcessorStep(
            features={**config.input_features, **config.output_features},
            norm_map=config.normalization_mapping,
@@ -149,6 +158,7 @@ def make_pi0_pre_post_processors(
        UnnormalizerProcessorStep(
            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
        ),
+        AbsoluteActionsProcessorStep(enabled=config.use_delta_actions, delta_step=delta_step),
        DeviceProcessorStep(device="cpu"),
    ]

@@ -50,6 +50,13 @@ class PI05Config(PreTrainedConfig):
    min_period: float = 4e-3
    max_period: float = 4.0

+    # Delta actions: converts absolute actions to delta (relative to state).
+    use_delta_actions: bool = False
+    # Joint names to exclude from delta (kept absolute). Empty list = all dims delta.
+    delta_exclude_joints: list[str] = field(default_factory=lambda: ["gripper"])
+    # Populated at runtime from dataset metadata by make_policy.
+    action_feature_names: list[str] | None = None
+
    # Real-Time Chunking (RTC) configuration
    rtc_config: RTCConfig | None = None

@@ -15,16 +15,16 @@
 # limitations under the License.

 import builtins
-import copy
 import logging
 import math
 from collections import deque
 from pathlib import Path
-from typing import TYPE_CHECKING, Literal, TypedDict, Unpack
+from typing import TYPE_CHECKING, Literal, TypedDict

 import torch
 import torch.nn.functional as F  # noqa: N812
 from torch import Tensor, nn
+from typing_extensions import Unpack

 from lerobot.utils.import_utils import _transformers_available

@@ -32,20 +32,14 @@ from lerobot.utils.import_utils import _transformers_available
 if TYPE_CHECKING or _transformers_available:
    from transformers.models.auto import CONFIG_MAPPING
    from transformers.models.gemma import modeling_gemma
-
-    from lerobot.policies.pi_gemma import (
-        PaliGemmaForConditionalGenerationWithPiGemma,
-        PiGemmaForCausalLM,
-        _gated_residual,
-        layernorm_forward,
-    )
+    from transformers.models.gemma.modeling_gemma import GemmaForCausalLM
+    from transformers.models.paligemma.modeling_paligemma import PaliGemmaForConditionalGeneration
 else:
    CONFIG_MAPPING = None
    modeling_gemma = None
-    PiGemmaForCausalLM = None
-    _gated_residual = None
-    layernorm_forward = None
-    PaliGemmaForConditionalGenerationWithPiGemma = None
+    GemmaForCausalLM = None
+    PaliGemmaForConditionalGeneration = None
+
 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.policies.pi05.configuration_pi05 import DEFAULT_IMAGE_SIZE, PI05Config
 from lerobot.policies.pretrained import PreTrainedPolicy, T
@@ -98,11 +92,10 @@ def create_sinusoidal_pos_embedding(  # see openpi `create_sinusoidal_pos_embedd


 def sample_beta(alpha, beta, bsize, device):  # see openpi `sample_beta` (exact copy)
-    # Beta sampling uses _sample_dirichlet which isn't implemented for MPS, so sample on CPU
-    alpha_t = torch.tensor(alpha, dtype=torch.float32)
-    beta_t = torch.tensor(beta, dtype=torch.float32)
+    alpha_t = torch.as_tensor(alpha, dtype=torch.float32, device=device)
+    beta_t = torch.as_tensor(beta, dtype=torch.float32, device=device)
    dist = torch.distributions.Beta(alpha_t, beta_t)
-    return dist.sample((bsize,)).to(device)
+    return dist.sample((bsize,))


 def make_att_2d_masks(pad_masks, att_masks):  # see openpi `make_att_2d_masks` (exact copy)
@@ -196,7 +189,7 @@ def resize_with_pad_torch(  # see openpi `resize_with_pad_torch` (exact copy)
    if images.dtype == torch.uint8:
        resized_images = torch.round(resized_images).clamp(0, 255).to(torch.uint8)
    elif images.dtype == torch.float32:
-        resized_images = resized_images.clamp(0.0, 1.0)
+        resized_images = resized_images.clamp(-1.0, 1.0)
    else:
        raise ValueError(f"Unsupported image dtype: {images.dtype}")

@@ -207,7 +200,7 @@ def resize_with_pad_torch(  # see openpi `resize_with_pad_torch` (exact copy)
    pad_w1 = pad_w0 + remainder_w

    # Pad
-    constant_value = 0 if images.dtype == torch.uint8 else 0.0
+    constant_value = 0 if images.dtype == torch.uint8 else -1.0
    padded_images = F.pad(
        resized_images,
        (pad_w0, pad_w1, pad_h0, pad_h1),  # left, right, top, bottom
@@ -226,14 +219,14 @@ def resize_with_pad_torch(  # see openpi `resize_with_pad_torch` (exact copy)
 def compute_layer_complete(
    layer_idx, inputs_embeds, attention_mask, position_ids, adarms_cond, paligemma, gemma_expert
 ):
-    models = [paligemma.model.language_model, gemma_expert.model]
+    models = [paligemma.language_model, gemma_expert.model]
    query_states = []
    key_states = []
    value_states = []
    gates = []
    for i, hidden_states in enumerate(inputs_embeds):
        layer = models[i].layers[layer_idx]
-        hidden_states, gate = layernorm_forward(layer.input_layernorm, hidden_states, adarms_cond[i])
+        hidden_states, gate = layer.input_layernorm(hidden_states, cond=adarms_cond[i])  # noqa: PLW2901
        gates.append(gate)
        input_shape = hidden_states.shape[:-1]
        hidden_shape = (*input_shape, -1, layer.self_attn.head_dim)
@@ -259,10 +252,10 @@ def compute_layer_complete(
        query_states, key_states, cos, sin, unsqueeze_dim=1
    )
    batch_size = query_states.shape[0]
-    scaling = paligemma.model.language_model.layers[layer_idx].self_attn.scaling
+    scaling = paligemma.language_model.layers[layer_idx].self_attn.scaling
    # Attention computation
    att_output, _ = modeling_gemma.eager_attention_forward(
-        paligemma.model.language_model.layers[layer_idx].self_attn,
+        paligemma.language_model.layers[layer_idx].self_attn,
        query_states,
        key_states,
        value_states,
@@ -270,7 +263,7 @@ def compute_layer_complete(
        scaling,
    )
    # Get head_dim from the current layer, not from the model
-    head_dim = paligemma.model.language_model.layers[layer_idx].self_attn.head_dim
+    head_dim = paligemma.language_model.layers[layer_idx].self_attn.head_dim
    att_output = att_output.reshape(batch_size, -1, 1 * 8 * head_dim)
    # Process layer outputs
    outputs_embeds = []
@@ -282,15 +275,15 @@ def compute_layer_complete(
            att_output = att_output.to(layer.self_attn.o_proj.weight.dtype)
        out_emb = layer.self_attn.o_proj(att_output[:, start_pos:end_pos])
        # first residual
-        out_emb = _gated_residual(hidden_states, out_emb, gates[i])
+        out_emb = modeling_gemma._gated_residual(hidden_states, out_emb, gates[i])  # noqa: SLF001
        after_first_residual = out_emb.clone()
-        out_emb, gate = layernorm_forward(layer.post_attention_layernorm, out_emb, adarms_cond[i])
+        out_emb, gate = layer.post_attention_layernorm(out_emb, cond=adarms_cond[i])
        # Convert to bfloat16 if the next layer (mlp) uses bfloat16
        if layer.mlp.up_proj.weight.dtype == torch.bfloat16:
            out_emb = out_emb.to(dtype=torch.bfloat16)
        out_emb = layer.mlp(out_emb)
        # second residual
-        out_emb = _gated_residual(after_first_residual, out_emb, gate)
+        out_emb = modeling_gemma._gated_residual(after_first_residual, out_emb, gate)  # noqa: SLF001
        outputs_embeds.append(out_emb)
        start_pos = end_pos
    return outputs_embeds
@@ -363,7 +356,7 @@ class PaliGemmaWithExpertModel(
        vlm_config_hf.text_config.num_hidden_layers = vlm_config.depth
        vlm_config_hf.text_config.num_key_value_heads = vlm_config.num_kv_heads
        vlm_config_hf.text_config.hidden_activation = "gelu_pytorch_tanh"
-        vlm_config_hf.text_config.dtype = "float32"
+        vlm_config_hf.text_config.torch_dtype = "float32"
        vlm_config_hf.text_config.vocab_size = 257152
        vlm_config_hf.text_config.use_adarms = use_adarms[0]
        vlm_config_hf.text_config.adarms_cond_dim = vlm_config.width if use_adarms[0] else None
@@ -371,7 +364,7 @@ class PaliGemmaWithExpertModel(
        vlm_config_hf.vision_config.intermediate_size = 4304
        vlm_config_hf.vision_config.projection_dim = 2048
        vlm_config_hf.vision_config.projector_hidden_act = "gelu_fast"
-        vlm_config_hf.vision_config.dtype = "float32"
+        vlm_config_hf.vision_config.torch_dtype = "float32"

        action_expert_config_hf = CONFIG_MAPPING["gemma"](
            head_dim=action_expert_config.head_dim,
@@ -382,13 +375,13 @@ class PaliGemmaWithExpertModel(
            num_key_value_heads=action_expert_config.num_kv_heads,
            vocab_size=257152,
            hidden_activation="gelu_pytorch_tanh",
-            dtype="float32",
+            torch_dtype="float32",
            use_adarms=use_adarms[1],
            adarms_cond_dim=action_expert_config.width if use_adarms[1] else None,
        )

-        self.paligemma = PaliGemmaForConditionalGenerationWithPiGemma(config=vlm_config_hf)
-        self.gemma_expert = PiGemmaForCausalLM(config=action_expert_config_hf)
+        self.paligemma = PaliGemmaForConditionalGeneration(config=vlm_config_hf)
+        self.gemma_expert = GemmaForCausalLM(config=action_expert_config_hf)
        self.gemma_expert.model.embed_tokens = None

        self.to_bfloat16_for_selected_params(precision)
@@ -403,11 +396,10 @@ class PaliGemmaWithExpertModel(
        else:
            raise ValueError(f"Invalid precision: {precision}")

-        # Keep full vision path in float32 so we never toggle (toggle causes optimizer
-        # "same dtype" error). Saves memory vs full float32; more memory than only 3 params.
        params_to_keep_float32 = [
-            "vision_tower",
-            "multi_modal_projector",
+            "vision_tower.vision_model.embeddings.patch_embedding.weight",
+            "vision_tower.vision_model.embeddings.patch_embedding.bias",
+            "vision_tower.vision_model.embeddings.position_embedding.weight",
            "input_layernorm",
            "post_attention_layernorm",
            "model.norm",
@@ -419,8 +411,8 @@ class PaliGemmaWithExpertModel(

    def _set_requires_grad(self):
        if self.freeze_vision_encoder:
-            self.paligemma.model.vision_tower.eval()
-            for param in self.paligemma.model.vision_tower.parameters():
+            self.paligemma.vision_tower.eval()
+            for param in self.paligemma.vision_tower.parameters():
                param.requires_grad = False
        if self.train_expert_only:
            self.paligemma.eval()
@@ -430,23 +422,15 @@ class PaliGemmaWithExpertModel(
    def train(self, mode: bool = True):
        super().train(mode)
        if self.freeze_vision_encoder:
-            self.paligemma.model.vision_tower.eval()
+            self.paligemma.vision_tower.eval()
        if self.train_expert_only:
            self.paligemma.eval()

    def embed_image(self, image: torch.Tensor):
-        # Vision tower and multi_modal_projector are kept in float32 (params_to_keep_float32).
-        out_dtype = image.dtype
-        if image.dtype != torch.float32:
-            image = image.to(torch.float32)
-        image_outputs = self.paligemma.model.get_image_features(image)
-        features = image_outputs.pooler_output * self.paligemma.config.text_config.hidden_size**0.5
-        if features.dtype != out_dtype:
-            features = features.to(out_dtype)
-        return features
+        return self.paligemma.model.get_image_features(image)

    def embed_language_tokens(self, tokens: torch.Tensor):
-        return self.paligemma.model.language_model.embed_tokens(tokens)
+        return self.paligemma.language_model.embed_tokens(tokens)

    def forward(
        self,
@@ -460,7 +444,7 @@ class PaliGemmaWithExpertModel(
        if adarms_cond is None:
            adarms_cond = [None, None]
        if inputs_embeds[1] is None:
-            prefix_output = self.paligemma.model.language_model.forward(
+            prefix_output = self.paligemma.language_model.forward(
                inputs_embeds=inputs_embeds[0],
                attention_mask=attention_mask,
                position_ids=position_ids,
@@ -484,7 +468,7 @@ class PaliGemmaWithExpertModel(
            prefix_output = None
            prefix_past_key_values = None
        else:
-            models = [self.paligemma.model.language_model, self.gemma_expert.model]
+            models = [self.paligemma.language_model, self.gemma_expert.model]
            num_layers = self.paligemma.config.text_config.num_hidden_layers

            # Check if gradient checkpointing is enabled for any of the models
@@ -524,7 +508,7 @@ class PaliGemmaWithExpertModel(
            def compute_final_norms(inputs_embeds, adarms_cond):
                outputs_embeds = []
                for i, hidden_states in enumerate(inputs_embeds):
-                    out_emb, _ = layernorm_forward(models[i].norm, hidden_states, adarms_cond[i])
+                    out_emb, _ = models[i].norm(hidden_states, cond=adarms_cond[i])
                    outputs_embeds.append(out_emb)
                return outputs_embeds

@@ -589,19 +573,29 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
            # Also compile the main forward pass used during training
            self.forward = torch.compile(self.forward, mode=config.compile_mode)

+        msg = """An incorrect transformer version is used, please create an issue on https://github.com/huggingface/lerobot/issues"""
+
+        try:
+            from transformers.models.siglip import check
+
+            if not check.check_whether_transformers_replace_is_installed_correctly():
+                raise ValueError(msg)
+        except ImportError:
+            raise ValueError(msg) from None
+
    def gradient_checkpointing_enable(self):
        """Enable gradient checkpointing for memory optimization."""
        self.gradient_checkpointing_enabled = True
-        self.paligemma_with_expert.paligemma.model.language_model.gradient_checkpointing = True
-        self.paligemma_with_expert.paligemma.model.vision_tower.gradient_checkpointing = True
+        self.paligemma_with_expert.paligemma.language_model.gradient_checkpointing = True
+        self.paligemma_with_expert.paligemma.vision_tower.gradient_checkpointing = True
        self.paligemma_with_expert.gemma_expert.model.gradient_checkpointing = True
        logging.info("Enabled gradient checkpointing for PI05Pytorch model")

    def gradient_checkpointing_disable(self):
        """Disable gradient checkpointing."""
        self.gradient_checkpointing_enabled = False
-        self.paligemma_with_expert.paligemma.model.language_model.gradient_checkpointing = False
-        self.paligemma_with_expert.paligemma.model.vision_tower.gradient_checkpointing = False
+        self.paligemma_with_expert.paligemma.language_model.gradient_checkpointing = False
+        self.paligemma_with_expert.paligemma.vision_tower.gradient_checkpointing = False
        self.paligemma_with_expert.gemma_expert.model.gradient_checkpointing = False
        logging.info("Disabled gradient checkpointing for PI05Pytorch model")

@@ -743,7 +737,7 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
        suffix_embs, suffix_pad_masks, suffix_att_masks, adarms_cond = self.embed_suffix(x_t, time)

        if (
-            self.paligemma_with_expert.paligemma.model.language_model.layers[0].self_attn.q_proj.weight.dtype
+            self.paligemma_with_expert.paligemma.language_model.layers[0].self_attn.q_proj.weight.dtype
            == torch.bfloat16
        ):
            suffix_embs = suffix_embs.to(dtype=torch.bfloat16)
@@ -814,7 +808,7 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
        prefix_position_ids = torch.cumsum(prefix_pad_masks, dim=1) - 1

        prefix_att_2d_masks_4d = self._prepare_attention_masks_4d(prefix_att_2d_masks)
-        self.paligemma_with_expert.paligemma.model.language_model.config._attn_implementation = "eager"  # noqa: SLF001
+        self.paligemma_with_expert.paligemma.language_model.config._attn_implementation = "eager"  # noqa: SLF001

        _, past_key_values = self.paligemma_with_expert.forward(
            attention_mask=prefix_att_2d_masks_4d,
@@ -886,7 +880,6 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
        full_att_2d_masks_4d = self._prepare_attention_masks_4d(full_att_2d_masks)
        self.paligemma_with_expert.gemma_expert.model.config._attn_implementation = "eager"  # noqa: SLF001

-        past_key_values = copy.deepcopy(past_key_values)
        outputs_embeds, _ = self.paligemma_with_expert.forward(
            attention_mask=full_att_2d_masks_4d,
            position_ids=position_ids,
@@ -976,12 +969,14 @@ class PI05Policy(PreTrainedPolicy):
        # Check if dataset_stats were provided in kwargs
        model = cls(config, **kwargs)

-        # Load state dict (expects keys with "model." prefix)
+        # Now manually load and remap the state dict
        try:
+            # Try to load the pytorch_model.bin or model.safetensors file
            print(f"Loading model from: {pretrained_name_or_path}")
            try:
                from transformers.utils import cached_file

+                # Try safetensors first
                resolved_file = cached_file(
                    pretrained_name_or_path,
                    "model.safetensors",
@@ -989,7 +984,7 @@ class PI05Policy(PreTrainedPolicy):
                    force_download=kwargs.get("force_download", False),
                    resume_download=kwargs.get("resume_download"),
                    proxies=kwargs.get("proxies"),
-                    token=kwargs.get("token"),
+                    use_auth_token=kwargs.get("use_auth_token"),
                    revision=kwargs.get("revision"),
                    local_files_only=kwargs.get("local_files_only", False),
                )
@@ -1002,7 +997,7 @@ class PI05Policy(PreTrainedPolicy):
                print("Returning model without loading pretrained weights")
                return model

-            # First, fix any key differences (see openpi model.py, _fix_pytorch_state_dict_keys)
+            # First, fix any key differences # see openpi `model.py, _fix_pytorch_state_dict_keys`
            fixed_state_dict = model._fix_pytorch_state_dict_keys(original_state_dict, model.config)

            # Then add "model." prefix for all keys that don't already have it
@@ -1014,6 +1009,8 @@ class PI05Policy(PreTrainedPolicy):
                    new_key = f"model.{key}"
                    remapped_state_dict[new_key] = value
                    remap_count += 1
+                    if remap_count <= 10:  # Only print first 10 to avoid spam
+                        print(f"Remapped: {key} -> {new_key}")
                else:
                    remapped_state_dict[key] = value

@@ -1047,7 +1044,7 @@ class PI05Policy(PreTrainedPolicy):
                print("All keys loaded successfully!")

        except Exception as e:
-            print(f"Warning: Could not load state dict: {e}")
+            print(f"Warning: Could not remap state dict keys: {e}")

        return model

@@ -1101,14 +1098,6 @@ class PI05Policy(PreTrainedPolicy):
                # Some checkpoints might have this, but current model expects different structure
                logging.warning(f"Vision embedding key might need handling: {key}")

-            if (
-                key == "model.paligemma_with_expert.paligemma.lm_head.weight"
-                or key == "paligemma_with_expert.paligemma.lm_head.weight"
-            ):
-                fixed_state_dict[
-                    "model.paligemma_with_expert.paligemma.model.language_model.embed_tokens.weight"
-                ] = value.clone()
-
            fixed_state_dict[new_key] = value

        return fixed_state_dict
@@ -23,8 +23,11 @@ import torch

 from lerobot.configs.types import PipelineFeatureType, PolicyFeature
 from lerobot.policies.pi05.configuration_pi05 import PI05Config
+from lerobot.policies.pi05.modeling_pi05 import pad_vector
 from lerobot.processor import (
+    AbsoluteActionsProcessorStep,
    AddBatchDimensionProcessorStep,
+    DeltaActionsProcessorStep,
    DeviceProcessorStep,
    NormalizerProcessorStep,
    PolicyAction,
@@ -67,6 +70,9 @@ class Pi05PrepareStateTokenizerProcessorStep(ProcessorStep):
        # TODO: check if this necessary
        state = deepcopy(state)

+        # Prepare state (pad to max_state_dim)
+        state = pad_vector(state, self.max_state_dim)
+
        # State should already be normalized to [-1, 1] by the NormalizerProcessorStep that runs before this step
        # Discretize into 256 bins (see openpi `PaligemmaTokenizer.tokenize()`)
        state_np = state.cpu().numpy()
@@ -125,10 +131,19 @@ def make_pi05_pre_post_processors(
        A tuple containing the configured pre-processor and post-processor pipelines.
    """

-    # Add remaining processors
+    delta_step = DeltaActionsProcessorStep(
+        enabled=config.use_delta_actions,
+        exclude_joints=getattr(config, "delta_exclude_joints", []),
+        action_names=getattr(config, "action_feature_names", None),
+    )
+
+    # OpenPI order: raw → delta → normalize → model → unnormalize → absolute
+    # NOTE: NormalizerProcessorStep MUST come before Pi05PrepareStateTokenizerProcessorStep
+    # because the tokenizer step expects normalized state in [-1, 1] range for discretization
    input_steps: list[ProcessorStep] = [
        RenameObservationsProcessorStep(rename_map={}),  # To mimic the same processor as pretrained one
        AddBatchDimensionProcessorStep(),
+        delta_step,
        # NOTE: NormalizerProcessorStep MUST come before Pi05PrepareStateTokenizerProcessorStep
        # because the tokenizer step expects normalized state in [-1, 1] range for discretization
        NormalizerProcessorStep(
@@ -150,6 +165,7 @@ def make_pi05_pre_post_processors(
        UnnormalizerProcessorStep(
            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
        ),
+        AbsoluteActionsProcessorStep(enabled=config.use_delta_actions, delta_step=delta_step),
        DeviceProcessorStep(device="cpu"),
    ]

@@ -41,6 +41,9 @@ class PI0FastConfig(PreTrainedConfig):
    max_action_dim: int = 32
    max_action_tokens: int = 256

+    # Delta actions: converts absolute actions to delta (relative to state).
+    use_delta_actions: bool = False
+
    # Real-Time Chunking (RTC) configuration
    rtc_config: RTCConfig | None = None

@@ -54,7 +57,7 @@ class PI0FastConfig(PreTrainedConfig):

    tokenizer_max_length: int = 200  # see openpi `__post_init__`
    text_tokenizer_name: str = "google/paligemma-3b-pt-224"
-    action_tokenizer_name: str = "lerobot/fast-action-tokenizer"
+    action_tokenizer_name: str = "physical-intelligence/fast"
    temperature: float = 0.0
    max_decoding_steps: int = 256
    fast_skip_tokens: int = 128
@@ -19,12 +19,13 @@ import logging
 import math
 from collections import deque
 from pathlib import Path
-from typing import TYPE_CHECKING, Literal, TypedDict, Unpack
+from typing import TYPE_CHECKING, Literal, TypedDict

 import numpy as np
 import torch
 import torch.nn.functional as F  # noqa: N812
 from torch import Tensor, nn
+from typing_extensions import Unpack

 from lerobot.utils.import_utils import _scipy_available, _transformers_available

@@ -37,27 +38,24 @@ else:
 if TYPE_CHECKING or _transformers_available:
    from transformers import AutoTokenizer
    from transformers.models.auto import CONFIG_MAPPING
-
-    from lerobot.policies.pi_gemma import (
-        PaliGemmaForConditionalGenerationWithPiGemma,
-        PiGemmaModel,
-    )
+    from transformers.models.paligemma.modeling_paligemma import PaliGemmaForConditionalGeneration
 else:
    CONFIG_MAPPING = None
+    PaliGemmaForConditionalGeneration = None
    AutoTokenizer = None
-    PiGemmaModel = None
-    PaliGemmaForConditionalGenerationWithPiGemma = None

 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.policies.pi0_fast.configuration_pi0_fast import PI0FastConfig
 from lerobot.policies.pretrained import PreTrainedPolicy, T
 from lerobot.policies.rtc.modeling_rtc import RTCProcessor
+from lerobot.processor.delta_action_processor import to_absolute_actions
 from lerobot.utils.constants import (
    ACTION,
    ACTION_TOKEN_MASK,
    ACTION_TOKENS,
    OBS_LANGUAGE_ATTENTION_MASK,
    OBS_LANGUAGE_TOKENS,
+    OBS_STATE,
    OPENPI_ATTENTION_MASK_VALUE,
 )

@@ -125,7 +123,7 @@ def resize_with_pad_torch(  # see openpi `resize_with_pad_torch` (exact copy)
    if images.dtype == torch.uint8:
        resized_images = torch.round(resized_images).clamp(0, 255).to(torch.uint8)
    elif images.dtype == torch.float32:
-        resized_images = resized_images.clamp(0.0, 1.0)
+        resized_images = resized_images.clamp(-1.0, 1.0)
    else:
        raise ValueError(f"Unsupported image dtype: {images.dtype}")

@@ -136,7 +134,7 @@ def resize_with_pad_torch(  # see openpi `resize_with_pad_torch` (exact copy)
    pad_w1 = pad_w0 + remainder_w

    # Pad
-    constant_value = 0 if images.dtype == torch.uint8 else 0.0
+    constant_value = 0 if images.dtype == torch.uint8 else -1.0
    padded_images = F.pad(
        resized_images,
        (pad_w0, pad_w1, pad_h0, pad_h1),  # left, right, top, bottom
@@ -210,22 +208,16 @@ class PI0FastPaliGemma(nn.Module):
        vlm_config_hf.text_config.num_hidden_layers = vlm_config.depth
        vlm_config_hf.text_config.num_key_value_heads = vlm_config.num_kv_heads
        vlm_config_hf.text_config.hidden_activation = "gelu_pytorch_tanh"
-        vlm_config_hf.text_config.dtype = "float32"
+        vlm_config_hf.text_config.torch_dtype = "float32"
        vlm_config_hf.text_config.vocab_size = 257152
        vlm_config_hf.text_config.use_adarms = use_adarms[0]
        vlm_config_hf.text_config.adarms_cond_dim = vlm_config.width if use_adarms[0] else None
        vlm_config_hf.vision_config.intermediate_size = 4304
        vlm_config_hf.vision_config.projection_dim = 2048
        vlm_config_hf.vision_config.projector_hidden_act = "gelu_fast"
-        vlm_config_hf.vision_config.dtype = "float32"
+        vlm_config_hf.vision_config.torch_dtype = "float32"

-        self.paligemma = PaliGemmaForConditionalGenerationWithPiGemma(config=vlm_config_hf)
-
-        # Use PI Gemma (AdaRMS) as language model when use_adarms[0] is True so that
-        # forward(..., adarms_cond=...) is supported (same as pi0/pi05).
-        if use_adarms[0]:
-            text_config = self.paligemma.config.text_config
-            self.paligemma.model.language_model = PiGemmaModel(text_config)
+        self.paligemma = PaliGemmaForConditionalGeneration(config=vlm_config_hf)

        self.to_bfloat16_for_selected_params(precision)

@@ -238,11 +230,10 @@ class PI0FastPaliGemma(nn.Module):
        else:
            raise ValueError(f"Invalid precision: {precision}")

-        # Keep full vision path in float32 so we never toggle (toggle causes optimizer
-        # "same dtype" error). Align with PI05.
        params_to_keep_float32 = [
-            "vision_tower",
-            "multi_modal_projector",
+            "vision_tower.vision_model.embeddings.patch_embedding.weight",
+            "vision_tower.vision_model.embeddings.patch_embedding.bias",
+            "vision_tower.vision_model.embeddings.position_embedding.weight",
            "input_layernorm",
            "post_attention_layernorm",
            "model.norm",
@@ -253,18 +244,10 @@ class PI0FastPaliGemma(nn.Module):
                param.data = param.data.to(dtype=torch.float32)

    def embed_image(self, image: torch.Tensor):
-        # Vision tower and multi_modal_projector are kept in float32 (params_to_keep_float32). Align with PI05.
-        out_dtype = image.dtype
-        if image.dtype != torch.float32:
-            image = image.to(torch.float32)
-        image_outputs = self.paligemma.model.get_image_features(image)
-        features = image_outputs.pooler_output * self.paligemma.config.text_config.hidden_size**0.5
-        if features.dtype != out_dtype:
-            features = features.to(out_dtype)
-        return features
+        return self.paligemma.model.get_image_features(image)

    def embed_language_tokens(self, tokens: torch.Tensor):
-        return self.paligemma.model.language_model.embed_tokens(tokens)
+        return self.paligemma.language_model.embed_tokens(tokens)

    def forward(
        self,
@@ -278,7 +261,7 @@ class PI0FastPaliGemma(nn.Module):
        if adarms_cond is None:
            adarms_cond = [None, None]
        if inputs_embeds[1] is None:
-            prefix_output = self.paligemma.model.language_model.forward(
+            prefix_output = self.paligemma.language_model.forward(
                inputs_embeds=inputs_embeds[0],
                attention_mask=attention_mask,
                position_ids=position_ids,
@@ -325,14 +308,24 @@ class PI0FastPytorch(nn.Module):  # see openpi `PI0Pytorch`
            self.sample_actions_fast = torch.compile(self.sample_actions_fast, mode=config.compile_mode)
            self.forward = torch.compile(self.forward, mode=config.compile_mode)

+        msg = """An incorrect transformer version is used, please create an issue on https://github.com/huggingface/lerobot/issues"""
+
+        try:
+            from transformers.models.siglip import check
+
+            if not check.check_whether_transformers_replace_is_installed_correctly():
+                raise ValueError(msg)
+        except ImportError:
+            raise ValueError(msg) from None
+
    def gradient_checkpointing_enable(self):
        """Enable gradient checkpointing for memory optimization."""
        self.gradient_checkpointing_enabled = True
        # Call the proper gradient_checkpointing_enable() method with use_reentrant=False for better memory efficiency
-        self.paligemma_with_expert.paligemma.model.language_model.gradient_checkpointing_enable(
+        self.paligemma_with_expert.paligemma.language_model.gradient_checkpointing_enable(
            gradient_checkpointing_kwargs={"use_reentrant": False}
        )
-        self.paligemma_with_expert.paligemma.model.vision_tower.gradient_checkpointing_enable(
+        self.paligemma_with_expert.paligemma.vision_tower.gradient_checkpointing_enable(
            gradient_checkpointing_kwargs={"use_reentrant": False}
        )
        logging.info("Enabled gradient checkpointing for PI0FastPytorch model")
@@ -341,8 +334,8 @@ class PI0FastPytorch(nn.Module):  # see openpi `PI0Pytorch`
        """Disable gradient checkpointing."""
        self.gradient_checkpointing_enabled = False
        # Call the proper gradient_checkpointing_disable() method
-        self.paligemma_with_expert.paligemma.model.language_model.gradient_checkpointing_disable()
-        self.paligemma_with_expert.paligemma.model.vision_tower.gradient_checkpointing_disable()
+        self.paligemma_with_expert.paligemma.language_model.gradient_checkpointing_disable()
+        self.paligemma_with_expert.paligemma.vision_tower.gradient_checkpointing_disable()
        logging.info("Disabled gradient checkpointing for PI0FastPytorch model")

    def _apply_checkpoint(self, func, *args, **kwargs):
@@ -532,7 +525,7 @@ class PI0FastPytorch(nn.Module):  # see openpi `PI0Pytorch`

        # Convert embeddings to bfloat16 if needed
        if (
-            self.paligemma_with_expert.paligemma.model.language_model.layers[0].self_attn.q_proj.weight.dtype
+            self.paligemma_with_expert.paligemma.language_model.layers[0].self_attn.q_proj.weight.dtype
            == torch.bfloat16
        ):
            prefix_embs = prefix_embs.to(dtype=torch.bfloat16)
@@ -625,7 +618,7 @@ class PI0FastPytorch(nn.Module):  # see openpi `PI0Pytorch`
        )

        if (
-            self.paligemma_with_expert.paligemma.model.language_model.layers[0].self_attn.q_proj.weight.dtype
+            self.paligemma_with_expert.paligemma.language_model.layers[0].self_attn.q_proj.weight.dtype
            == torch.bfloat16
        ):
            prefix_embs = prefix_embs.to(dtype=torch.bfloat16)
@@ -723,7 +716,7 @@ class PI0FastPytorch(nn.Module):  # see openpi `PI0Pytorch`

        # Ensure correct precision (bfloat16/float32)
        if (
-            self.paligemma_with_expert.paligemma.model.language_model.layers[0].self_attn.q_proj.weight.dtype
+            self.paligemma_with_expert.paligemma.language_model.layers[0].self_attn.q_proj.weight.dtype
            == torch.bfloat16
        ):
            prefix_embs = prefix_embs.to(dtype=torch.bfloat16)
@@ -906,12 +899,14 @@ class PI0FastPolicy(PreTrainedPolicy):
        # Check if dataset_stats were provided in kwargs
        model = cls(config, **kwargs)

-        # Load state dict (expects keys with "model." prefix)
+        # Now manually load and remap the state dict
        try:
+            # Try to load the pytorch_model.bin or model.safetensors file
            print(f"Loading model from: {pretrained_name_or_path}")
            try:
                from transformers.utils import cached_file

+                # Try safetensors first
                resolved_file = cached_file(
                    pretrained_name_or_path,
                    "model.safetensors",
@@ -919,7 +914,7 @@ class PI0FastPolicy(PreTrainedPolicy):
                    force_download=kwargs.get("force_download", False),
                    resume_download=kwargs.get("resume_download"),
                    proxies=kwargs.get("proxies"),
-                    token=kwargs.get("token"),
+                    use_auth_token=kwargs.get("use_auth_token"),
                    revision=kwargs.get("revision"),
                    local_files_only=kwargs.get("local_files_only", False),
                )
@@ -932,9 +927,8 @@ class PI0FastPolicy(PreTrainedPolicy):
                print("Returning model without loading pretrained weights")
                return model

-            # First, fix any key differences (see openpi model.py, _fix_pytorch_state_dict_keys)
+            # First, fix any key differences # see openpi `model.py, _fix_pytorch_state_dict_keys`
            fixed_state_dict = model._fix_pytorch_state_dict_keys(original_state_dict, model.config)
-
            # Then add "model." prefix for all keys that don't already have it
            remapped_state_dict = {}
            remap_count = 0
@@ -944,6 +938,8 @@ class PI0FastPolicy(PreTrainedPolicy):
                    new_key = f"model.{key}"
                    remapped_state_dict[new_key] = value
                    remap_count += 1
+                    if remap_count <= 10:  # Only print first 10 to avoid spam
+                        print(f"Remapped: {key} -> {new_key}")
                else:
                    remapped_state_dict[key] = value

@@ -977,7 +973,7 @@ class PI0FastPolicy(PreTrainedPolicy):
                print("All keys loaded successfully!")

        except Exception as e:
-            print(f"Warning: Could not load state dict: {e}")
+            print(f"Warning: Could not remap state dict keys: {e}")

        return model

@@ -1321,6 +1317,12 @@ class PI0FastPolicy(PreTrainedPolicy):
            action_tokens, action_horizon=action_horizon, action_dim=action_dim
        )

+        if self.config.use_delta_actions and OBS_STATE in batch:
+            state = pad_vector(batch[OBS_STATE], self.config.max_state_dim)
+            continuous_actions = to_absolute_actions(
+                continuous_actions, state, [True] * continuous_actions.shape[-1]
+        )
+
        return continuous_actions

    def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict]:
@@ -23,9 +23,11 @@ import torch

 from lerobot.configs.types import PipelineFeatureType, PolicyFeature
 from lerobot.policies.pi0_fast.configuration_pi0_fast import PI0FastConfig
+from lerobot.policies.pi0_fast.modeling_pi0_fast import pad_vector
 from lerobot.processor import (
    ActionTokenizerProcessorStep,
    AddBatchDimensionProcessorStep,
+    DeltaActionsProcessorStep,
    DeviceProcessorStep,
    NormalizerProcessorStep,
    PolicyAction,
@@ -68,6 +70,9 @@ class Pi0FastPrepareStateAndLanguageTokenizerProcessorStep(ProcessorStep):
        # TODO: check if this necessary
        state = deepcopy(state)

+        # Prepare state (pad to max_state_dim)
+        state = pad_vector(state, self.max_state_dim)
+
        # State should already be normalized to [-1, 1] by the NormalizerProcessorStep that runs before this step
        # Discretize into 256 bins (see openpi `PaligemmaTokenizer.tokenize()`)
        state_np = state.cpu().numpy()
@@ -143,6 +148,7 @@ def make_pi0_fast_pre_post_processors(
            padding_side="right",
            padding="max_length",
        ),
+        DeltaActionsProcessorStep(enabled=config.use_delta_actions),
        ActionTokenizerProcessorStep(
            action_tokenizer_name=config.action_tokenizer_name,
            max_action_tokens=config.max_action_tokens,
@@ -1,363 +0,0 @@
-# Copyright 2025 Physical Intelligence and The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-from __future__ import annotations
-
-from typing import TYPE_CHECKING
-
-import torch
-from torch import nn
-
-from lerobot.utils.import_utils import _transformers_available
-
-if TYPE_CHECKING or _transformers_available:
-    from transformers.cache_utils import DynamicCache
-    from transformers.masking_utils import create_causal_mask
-    from transformers.modeling_layers import GradientCheckpointingLayer
-    from transformers.modeling_outputs import BaseModelOutputWithPast
-    from transformers.models.gemma.modeling_gemma import (
-        GemmaAttention,
-        GemmaConfig,
-        GemmaForCausalLM,
-        GemmaMLP,
-        GemmaModel,
-    )
-    from transformers.models.paligemma.modeling_paligemma import (
-        PaliGemmaForConditionalGeneration,
-        PaliGemmaModel,
-    )
-else:
-    GemmaAttention = None
-    GemmaConfig = None
-    GemmaForCausalLM = None
-    GemmaMLP = None
-    GemmaModel = None
-    PaliGemmaModel = None
-    PaliGemmaForConditionalGeneration = None
-    DynamicCache = None
-    GradientCheckpointingLayer = None
-    BaseModelOutputWithPast = None
-    create_causal_mask = None
-
-
-def _gated_residual(
-    x: torch.Tensor | None,
-    y: torch.Tensor | None,
-    gate: torch.Tensor | None,
-) -> torch.Tensor | None:
-    """Gated residual: x + y when gate is None, else x + y * gate."""
-    if x is None and y is None:
-        return None
-    if x is None or y is None:
-        return x if x is not None else y
-    if gate is None:
-        return x + y
-    return x + y * gate
-
-
-def layernorm_forward(
-    layernorm: nn.Module,
-    x: torch.Tensor,
-    cond: torch.Tensor | None = None,
-):
-    """
-    call layernorm and return hidden states and gate
-    if cond is not None, use conditional norm
-    otherwise, use normal gemma norm
-    """
-    if cond is not None:
-        return layernorm(x, cond=cond)
-    else:
-        return layernorm(x)
-
-
-class PiGemmaRMSNorm(nn.Module):
-    """
-    Adaptive RMSNorm for PI Gemma (AdaRMS).
-    When cond_dim is set, uses cond to modulate scale/shift/gate; otherwise behaves like standard GemmaRMSNorm.
-    forward(x, cond=None) returns (output, gate) for use with _gated_residual.
-    """
-
-    def __init__(self, dim: int, eps: float = 1e-6, cond_dim: int | None = None):
-        super().__init__()
-        self.eps = eps
-        self.dim = dim
-        self.cond_dim = cond_dim
-        if cond_dim is not None:
-            self.dense = nn.Linear(cond_dim, dim * 3, bias=True)
-            nn.init.zeros_(self.dense.weight)
-        else:
-            self.weight = nn.Parameter(torch.zeros(dim))
-            self.dense = None
-
-    def _norm(self, x):
-        # Compute variance in float32 (like the source implementation)
-        var = torch.mean(torch.square(x.float()), dim=-1, keepdim=True)
-        # Compute normalization in float32
-        normed_inputs = x * torch.rsqrt(var + self.eps)
-        return normed_inputs
-
-    def forward(
-        self,
-        x: torch.Tensor,
-        cond: torch.Tensor | None = None,
-    ) -> tuple[torch.Tensor, torch.Tensor | None]:
-        dtype = x.dtype
-        normed = self._norm(x)
-        if cond is None or self.dense is None:
-            normed = normed * (1.0 + self.weight.float())
-            return normed.type_as(x), None
-        if cond.shape[-1] != self.cond_dim:
-            raise ValueError(f"Expected cond dim {self.cond_dim}, got {cond.shape[-1]}")
-        modulation = self.dense(cond)
-        if len(x.shape) == 3:
-            modulation = modulation.unsqueeze(1)
-        scale, shift, gate = modulation.chunk(3, dim=-1)
-        normed = normed * (1 + scale.float()) + shift.float()
-        return normed.to(dtype), gate.to(dtype)
-
-    def extra_repr(self) -> str:
-        if self.dense is not None:
-            return f"dim={self.dim}, eps={self.eps}, adaptive=True, cond_dim={self.cond_dim}"
-        return f"dim={self.dim}, eps={self.eps}"
-
-
-def _get_pi_gemma_decoder_layer_base():
-    """base for PiGemmaDecoderLayer"""
-
-    class _PiGemmaDecoderLayerBase(GradientCheckpointingLayer):
-        """Decoder layer that uses PiGemmaRMSNorm and _gated_residual, compatible with v5 Gemma."""
-
-        def __init__(self, config: GemmaConfig, layer_idx: int):
-            super().__init__()
-            self.hidden_size = config.hidden_size
-            self.self_attn = GemmaAttention(config=config, layer_idx=layer_idx)
-            self.mlp = GemmaMLP(config)
-            cond_dim = (
-                getattr(config, "adarms_cond_dim", None) if getattr(config, "use_adarms", False) else None
-            )
-            self.input_layernorm = PiGemmaRMSNorm(
-                config.hidden_size, eps=config.rms_norm_eps, cond_dim=cond_dim
-            )
-            self.post_attention_layernorm = PiGemmaRMSNorm(
-                config.hidden_size, eps=config.rms_norm_eps, cond_dim=cond_dim
-            )
-
-        def forward(
-            self,
-            hidden_states: torch.Tensor,
-            attention_mask: torch.Tensor | None = None,
-            position_ids: torch.LongTensor | None = None,
-            past_key_values=None,
-            use_cache: bool = False,
-            cache_position: torch.LongTensor | None = None,
-            position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
-            adarms_cond: torch.Tensor | None = None,
-            **kwargs,
-        ) -> torch.Tensor:
-            residual = hidden_states
-            hidden_states, gate = self.input_layernorm(hidden_states, cond=adarms_cond)
-            hidden_states, _ = self.self_attn(
-                hidden_states,
-                attention_mask=attention_mask,
-                position_ids=position_ids,
-                past_key_values=past_key_values,
-                use_cache=use_cache,
-                cache_position=cache_position,
-                position_embeddings=position_embeddings,
-                **kwargs,
-            )
-
-            hidden_states = _gated_residual(residual, hidden_states, gate)
-
-            residual = hidden_states
-            hidden_states, gate = self.post_attention_layernorm(hidden_states, cond=adarms_cond)
-            hidden_states = self.mlp(hidden_states)
-            hidden_states = _gated_residual(residual, hidden_states, gate)
-            return hidden_states
-
-    return _PiGemmaDecoderLayerBase
-
-
-class PiGemmaModel(GemmaModel):  # type: ignore[misc]
-    """
-    GemmaModel extended with AdaRMS (adaptive RMSNorm) and gated residuals when config.use_adarms is True.
-    """
-
-    def __init__(self, config: GemmaConfig, **kwargs):
-        super().__init__(config, **kwargs)
-        # if not getattr(config, "use_adarms", False):
-        #     return
-        cond_dim = getattr(config, "adarms_cond_dim", None)
-        pi_gemma_decoder_layer_base = _get_pi_gemma_decoder_layer_base()
-        self.layers = nn.ModuleList(
-            [pi_gemma_decoder_layer_base(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
-        )
-        self.norm = PiGemmaRMSNorm(config.hidden_size, eps=config.rms_norm_eps, cond_dim=cond_dim)
-
-    def forward(
-        self,
-        input_ids: torch.LongTensor | None = None,
-        attention_mask: torch.Tensor | None = None,
-        position_ids: torch.LongTensor | None = None,
-        past_key_values: DynamicCache | None = None,
-        inputs_embeds: torch.FloatTensor | None = None,
-        use_cache: bool | None = None,
-        output_attentions: bool | None = None,
-        output_hidden_states: bool | None = None,
-        cache_position: torch.LongTensor | None = None,
-        adarms_cond: torch.Tensor | None = None,
-        **kwargs,
-    ) -> BaseModelOutputWithPast:
-        """
-        adarms_cond (`torch.Tensor` of shape `(batch_size, cond_dim)`, *optional*):
-            Condition for ADARMS.
-        """
-        output_attentions = (
-            output_attentions if output_attentions is not None else self.config.output_attentions
-        )
-        output_hidden_states = (
-            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
-        )
-        use_cache = use_cache if use_cache is not None else self.config.use_cache
-
-        if (input_ids is None) ^ (inputs_embeds is not None):
-            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
-
-        if self.gradient_checkpointing and self.training and use_cache:
-            import logging
-
-            logging.warning(
-                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
-            )
-            use_cache = False
-
-        if inputs_embeds is None:
-            inputs_embeds = self.embed_tokens(input_ids)
-
-        if use_cache and past_key_values is None:
-            past_key_values = DynamicCache()
-
-        if cache_position is None:
-            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
-            cache_position = torch.arange(
-                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
-            )
-
-        if position_ids is None:
-            position_ids = cache_position.unsqueeze(0)
-
-        causal_mask = create_causal_mask(
-            config=self.config,
-            inputs_embeds=inputs_embeds,
-            attention_mask=attention_mask,
-            cache_position=cache_position,
-            past_key_values=past_key_values,
-            position_ids=position_ids,
-        )
-
-        # embed positions
-        hidden_states = inputs_embeds
-        # Convert to bfloat16 if the first layer uses bfloat16
-        if len(self.layers) > 0 and self.layers[0].self_attn.q_proj.weight.dtype == torch.bfloat16:
-            hidden_states = hidden_states.to(torch.bfloat16)
-
-        # create position embeddings to be shared across the decoder layers
-        position_embeddings = self.rotary_emb(hidden_states, position_ids)
-
-        # normalized
-        # Gemma downcasts the below to float16, causing sqrt(3072)=55.4256 to become 55.5
-        # See https://github.com/huggingface/transformers/pull/29402
-
-        # decoder layers
-        all_hidden_states = () if output_hidden_states else None
-        all_self_attns = () if output_attentions else None
-
-        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
-            if output_hidden_states:
-                all_hidden_states += (hidden_states,)
-
-            layer_outputs = decoder_layer(
-                hidden_states,
-                attention_mask=causal_mask,
-                position_ids=position_ids,
-                past_key_values=past_key_values,
-                output_attentions=output_attentions,
-                use_cache=use_cache,
-                cache_position=cache_position,
-                position_embeddings=position_embeddings,
-                adarms_cond=adarms_cond,
-                **kwargs,
-            )
-
-            hidden_states = layer_outputs
-
-            if output_attentions:
-                all_self_attns += (layer_outputs[1],)
-
-        hidden_states, _ = self.norm(hidden_states, adarms_cond)
-
-        # add hidden states from the last decoder layer
-        if output_hidden_states:
-            all_hidden_states += (hidden_states,)
-
-        return BaseModelOutputWithPast(
-            last_hidden_state=hidden_states,
-            past_key_values=past_key_values if use_cache else None,
-            hidden_states=all_hidden_states,
-            attentions=all_self_attns,
-        )
-
-
-class PiGemmaForCausalLM(GemmaForCausalLM):  # type: ignore[misc]
-    """
-    Causal LM wrapper using PiGemmaModel as the backbone, for consistency with GemmaForCausalLM
-    and the language model used in pi0_fast. Use this for the action expert in pi0/pi05.
-    """
-
-    def __init__(self, config: GemmaConfig, **kwargs):
-        super().__init__(config, **kwargs)
-        self.model = PiGemmaModel(config)
-
-
-class PaliGemmaModelWithPiGemma(PaliGemmaModel):
-    """PaliGemmaModel whose language_model is PiGemmaModel (custom decoder with PiGemmaRMSNorm and gated residuals)."""
-
-    def __init__(self, config):
-        super().__init__(config)
-        self.language_model = PiGemmaModel(config.text_config)
-
-
-class PaliGemmaForConditionalGenerationWithPiGemma(PaliGemmaForConditionalGeneration):
-    """PaliGemmaForConditionalGeneration using PiGemma decoder for the language model."""
-
-    def __init__(self, config):
-        super().__init__(config)
-        self.model = PaliGemmaModelWithPiGemma(config)
-
-    # Make modules available through conditional class for BC
-    @property
-    def language_model(self):
-        return self.model.language_model
-
-
-__all__ = [
-    "PiGemmaModel",
-    "PiGemmaForCausalLM",
-    "PiGemmaRMSNorm",
-    "_gated_residual",
-    "layernorm_forward",
-    "PaliGemmaModelWithPiGemma",
-    "PaliGemmaForConditionalGenerationWithPiGemma",
-]
@@ -19,7 +19,7 @@ import os
 from importlib.resources import files
 from pathlib import Path
 from tempfile import TemporaryDirectory
-from typing import TypedDict, TypeVar, Unpack
+from typing import TypedDict, TypeVar

 import packaging
 import safetensors
@@ -28,6 +28,7 @@ from huggingface_hub.constants import SAFETENSORS_SINGLE_FILE
 from huggingface_hub.errors import HfHubHTTPError
 from safetensors.torch import load_model as load_model_as_safetensor, save_model as save_model_as_safetensor
 from torch import Tensor, nn
+from typing_extensions import Unpack

 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.configs.train import TrainPipelineConfig
@@ -33,7 +33,7 @@ class RewardClassifierConfig(PreTrainedConfig):
    latent_dim: int = 256
    image_embedding_pooling_dim: int = 8
    dropout_rate: float = 0.1
-    model_name: str = "helper2424/resnet10"  # TODO: This needs to be updated. The model on the Hub doesn't call self.post_init() in its __init__, which is required by transformers v5 to set all_tied_weights_keys. The from_pretrained call fails when it tries to access this attribute during _finalize_model_loading.
+    model_name: str = "helper2424/resnet10"
    device: str = "cpu"
    model_type: str = "cnn"  # "transformer" or "cnn"
    num_cameras: int = 2
@@ -27,18 +27,18 @@ Usage:
    # Full RA-BC computation with visualizations
    python src/lerobot/policies/sarm/compute_rabc_weights.py \\
        --dataset-repo-id lerobot/aloha_sim_insertion_human \\
-        --reward-model-path <USER>/sarm_single_uni4
+        --reward-model-path pepijn223/sarm_single_uni4

    # Faster computation with stride (compute every 5 frames, interpolate the rest)
    python src/lerobot/policies/sarm/compute_rabc_weights.py \\
        --dataset-repo-id lerobot/aloha_sim_insertion_human \\
-        --reward-model-path <USER>/sarm_single_uni4 \\
+        --reward-model-path pepijn223/sarm_single_uni4 \\
        --stride 5

    # Visualize predictions only (no RA-BC computation)
    python src/lerobot/policies/sarm/compute_rabc_weights.py \\
        --dataset-repo-id lerobot/aloha_sim_insertion_human \\
-        --reward-model-path <USER>/sarm_single_uni4 \\
+        --reward-model-path pepijn223/sarm_single_uni4 \\
        --visualize-only \\
        --num-visualizations 5

@@ -714,12 +714,12 @@ Examples:
    # Full RA-BC computation with visualizations
    python src/lerobot/policies/sarm/compute_rabc_weights.py \\
        --dataset-repo-id lerobot/aloha_sim_insertion_human \\
-        --reward-model-path <USER>/sarm_single_uni4
+        --reward-model-path pepijn223/sarm_single_uni4

    # Visualize predictions only (no RA-BC computation)
    python src/lerobot/policies/sarm/compute_rabc_weights.py \\
        --dataset-repo-id lerobot/aloha_sim_insertion_human \\
-        --reward-model-path <USER>/sarm_single_uni4 \\
+        --reward-model-path pepijn223/sarm_single_uni4 \\
        --visualize-only \\
        --num-visualizations 10
        """,
@@ -277,7 +277,9 @@ class SARMEncodingProcessorStep(ProcessorStep):

        # When language is perturbed, targets are zero so perturbed samples don't contribute to progress loss
        if self.dataset_meta is not None:
-            episodes_df = self.dataset_meta.episodes.to_pandas()
+            episodes_df = None
+            if self.sparse_subtask_names != ["task"]:
+                episodes_df = self.dataset_meta.episodes.to_pandas()

            # Generate sparse targets
            if self.sparse_temporal_proportions is not None:
@@ -85,7 +85,7 @@ class SmolVLAConfig(PreTrainedConfig):
    scheduler_decay_lr: float = 2.5e-6

    vlm_model_name: str = "HuggingFaceTB/SmolVLM2-500M-Video-Instruct"  # Select the VLM backbone.
-    load_vlm_weights: bool = False  # Set to False in case of training the expert from scratch. True when init from pretrained SmolVLA weights
+    load_vlm_weights: bool = False  # Set to True in case of training the expert from scratch. True when init from pretrained SmolVLA weights

    add_image_special_tokens: bool = False  # Whether to use special image tokens around image features.

@@ -106,9 +106,6 @@ class SmolVLAConfig(PreTrainedConfig):
    # Real-Time Chunking (RTC) configuration
    rtc_config: RTCConfig | None = None

-    compile_model: bool = False  # Whether to use torch.compile for model optimization
-    compile_mode: str = "max-autotune"  # Torch compile mode
-
    def __post_init__(self):
        super().__post_init__()

@@ -30,7 +30,7 @@ Example of finetuning the smolvla pretrained model (`smolvla_base`):
 ```bash
 lerobot-train \
 --policy.path=lerobot/smolvla_base \
--dataset.repo_id=<USER>/svla_so100_task1_v3 \
+--dataset.repo_id=danaaubakirova/svla_so100_task1_v3 \
 --batch_size=64 \
 --steps=200000
 ```
@@ -40,7 +40,7 @@ and an action expert.
 ```bash
 lerobot-train \
 --policy.type=smolvla \
--dataset.repo_id=<USER>/svla_so100_task1_v3 \
+--dataset.repo_id=danaaubakirova/svla_so100_task1_v3 \
 --batch_size=64 \
 --steps=200000
 ```
@@ -54,11 +54,12 @@ policy = SmolVLAPolicy.from_pretrained("lerobot/smolvla_base")

 import math
 from collections import deque
-from typing import TypedDict, Unpack
+from typing import TypedDict

 import torch
 import torch.nn.functional as F  # noqa: N812
 from torch import Tensor, nn
+from typing_extensions import Unpack

 from lerobot.policies.pretrained import PreTrainedPolicy
 from lerobot.policies.rtc.modeling_rtc import RTCProcessor
@@ -592,12 +593,6 @@ class VLAFlowMatching(nn.Module):
        self.prefix_length = self.config.prefix_length
        self.rtc_processor = rtc_processor

-        # Compile model if requested
-        if config.compile_model:
-            torch.set_float32_matmul_precision("high")
-            self.sample_actions = torch.compile(self.sample_actions, mode=config.compile_mode)
-            self.forward = torch.compile(self.forward, mode=config.compile_mode)
-
    def _rtc_enabled(self):
        return self.config.rtc_config is not None and self.config.rtc_config.enabled

@@ -77,6 +77,7 @@ class SmolVLMWithExpertModel(nn.Module):
            print(f"Loading  {model_id} weights ...")
            self.vlm = AutoModelForImageTextToText.from_pretrained(
                model_id,
+                device_map=device,
                torch_dtype="bfloat16",
                low_cpu_mem_usage=True,
            )
@@ -55,7 +55,7 @@ class WallXConfig(PreTrainedConfig):
    pretrained_name_or_path: str = "x-square-robot/wall-oss-flow"

    # Tokenizer settings
-    action_tokenizer_path: str | None = "lerobot/fast-action-tokenizer"
+    action_tokenizer_path: str | None = "physical-intelligence/fast"

    # Action prediction mode: "diffusion" or "fast"
    prediction_mode: str = "diffusion"
@@ -261,15 +261,10 @@ class Qwen2_5_VLMoEForAction(Qwen2_5_VLForConditionalGeneration):
    and optional LoRA fine-tuning support.
    """

-    _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
+    _tied_weights_keys = ["lm_head.weight"]
    config_class = Qwen2_5_VLConfig
    _no_split_modules = ["Qwen2_5_VLDecoderLayer_with_MoE", "Qwen2_5_VLVisionBlock"]

-    def init_weights(self):
-        if getattr(self.model, "language_model", None) is not None:
-            return
-        super().init_weights()
-
    @classmethod
    def from_pretrained(
        cls,
@@ -317,11 +312,6 @@ class Qwen2_5_VLMoEForAction(Qwen2_5_VLForConditionalGeneration):
            processor.action_processor = action_tokenizer
        else:
            action_tokenizer = None
-
-        # add pad_token_id to config
-        config.pad_token_id = processor.tokenizer.pad_token_id
-        config.text_config.pad_token_id = processor.tokenizer.pad_token_id
-
        # Initialize model with configuration and processor
        model = cls(config, processor=processor, action_tokenizer=action_tokenizer, **kwargs)

@@ -341,7 +331,7 @@ class Qwen2_5_VLMoEForAction(Qwen2_5_VLForConditionalGeneration):
                force_download=kwargs.get("force_download", False),
                resume_download=kwargs.get("resume_download"),
                proxies=kwargs.get("proxies"),
-                token=kwargs.get("token"),
+                use_auth_token=kwargs.get("use_auth_token"),
                revision=kwargs.get("revision"),
                local_files_only=kwargs.get("local_files_only", False),
            )
@@ -21,7 +21,6 @@ class Qwen2_5_VLVisionConfig(PretrainedConfig):
        window_size=112,
        out_hidden_size=3584,
        fullatt_block_indexes=[7, 15, 23, 31],
-        initializer_range=0.02,
        **kwargs,
    ):
        super().__init__(**kwargs)
@@ -39,7 +38,6 @@ class Qwen2_5_VLVisionConfig(PretrainedConfig):
        self.window_size = window_size
        self.fullatt_block_indexes = fullatt_block_indexes
        self.out_hidden_size = out_hidden_size
-        self.initializer_range = initializer_range


 class Qwen2_5_VLConfig(PretrainedConfig):
@@ -11,6 +11,7 @@ from transformers.activations import ACT2FN
 from transformers.cache_utils import (
    Cache,
    DynamicCache,
+    SlidingWindowCache,
    StaticCache,
 )
 from transformers.generation import GenerationMixin
@@ -30,15 +31,6 @@ from transformers.utils import (

 from .configuration_qwen2_5_vl import Qwen2_5_VLConfig, Qwen2_5_VLVisionConfig

-
-# TODO(Steven): SlidingWindowCache was removed in transformers v5. Define a placeholder so isinstance checks
-# always return False (which is the correct behavior when no sliding window cache is in use).
-class _SlidingWindowCachePlaceholder:
-    pass
-
-
-SlidingWindowCache = _SlidingWindowCachePlaceholder
-
 if is_flash_attn_2_available():
    from flash_attn import flash_attn_func, flash_attn_varlen_func
    from flash_attn.layers.rotary import apply_rotary_emb
@@ -602,40 +594,19 @@ class Qwen2_5_VisionTransformerPretrainedModel(Qwen2_5_VLPreTrainedModel):
        return hidden_states


-def _compute_default_rope_parameters_qwen2_5_vl(config, device=None):
-    """
-    compute default rope parameters for Qwen2_5_VL
-    """
-    base = config.text_config.rope_parameters["rope_theta"]
-    dim = config.hidden_size // config.num_attention_heads
-    inv_freq = 1.0 / (
-        base ** (torch.arange(0, dim, 2, dtype=torch.int64).to(device=device, dtype=torch.float) / dim)
-    )
-    return inv_freq, 1.0
-
-
 class Qwen2_5_VLRotaryEmbedding(nn.Module):
    def __init__(self, config: Qwen2_5_VLConfig, device=None):
        super().__init__()
        # BC: "rope_type" was originally "type"
        if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
-        elif hasattr(config, "rope_parameters") and config.rope_parameters is not None:
-            self.rope_type = config.rope_parameters.get("rope_type", "default")
        else:
            self.rope_type = "default"
        self.max_seq_len_cached = config.max_position_embeddings
        self.original_max_seq_len = config.max_position_embeddings

        self.config = config
-
-        if self.rope_type == "default":
-            self.rope_init_fn = _compute_default_rope_parameters_qwen2_5_vl
-            self.rope_kwargs = {}
-        else:
-            rope_type_key = "linear" if self.rope_type == "linear" else self.rope_type
-            self.rope_init_fn = ROPE_INIT_FUNCTIONS[rope_type_key]
-            self.rope_kwargs = {}
+        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]

        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
        self.register_buffer("inv_freq", inv_freq, persistent=False)
@@ -1596,7 +1567,7 @@ QWEN2_5_VL_INPUTS_DOCSTRING = r"""


 class Qwen2_5_VLForConditionalGeneration(Qwen2_5_VLPreTrainedModel, GenerationMixin):
-    _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
+    _tied_weights_keys = ["lm_head.weight"]
    config_class = Qwen2_5_VLConfig
    _no_split_modules = ["Qwen2_5_VLDecoderLayer", "Qwen2_5_VLVisionBlock"]

@@ -144,7 +144,7 @@ def preprocesser_call(
    """
    # Process image inputs
    if images is not None and len(images) > 0:
-        image_inputs = processor.image_processor(images=images, return_tensors=return_tensors)
+        image_inputs = processor.image_processor(images=images, videos=None, return_tensors=return_tensors)
        image_grid_thw = image_inputs["image_grid_thw"]
    else:
        image_inputs = {}
@@ -152,7 +152,7 @@ def preprocesser_call(

    # Process video inputs
    if videos is not None:
-        videos_inputs = processor.image_processor(videos=videos, return_tensors=return_tensors)
+        videos_inputs = processor.image_processor(images=None, videos=videos, return_tensors=return_tensors)
        video_grid_thw = videos_inputs["video_grid_thw"]
    else:
        videos_inputs = {}
@@ -276,8 +276,6 @@ class Florence2LanguageConfig(PretrainedConfig):
        )

        # ensure backward compatibility for BART CNN models
-        if not hasattr(self, "forced_bos_token_id"):
-            self.forced_bos_token_id = None
        if self.forced_bos_token_id is None and kwargs.get("force_bos_token_to_be_generated", False):
            self.forced_bos_token_id = self.bos_token_id
            warnings.warn(
@@ -1951,10 +1951,7 @@ class Florence2Decoder(Florence2LanguagePreTrainedModel):


 class Florence2LanguageModel(Florence2LanguagePreTrainedModel):
-    _tied_weights_keys = {
-        "encoder.embed_tokens.weight": "shared.weight",
-        "decoder.embed_tokens.weight": "shared.weight",
-    }
+    _tied_weights_keys = ["encoder.embed_tokens.weight", "decoder.embed_tokens.weight"]

    def __init__(self, config: Florence2LanguageConfig):
        super().__init__(config)
@@ -2079,10 +2076,7 @@ class Florence2LanguageModel(Florence2LanguagePreTrainedModel):

 class Florence2LanguageForConditionalGeneration(Florence2LanguagePreTrainedModel, GenerationMixin):
    base_model_prefix = "model"
-    _tied_weights_keys = {
-        "model.encoder.embed_tokens.weight": "model.shared.weight",
-        "model.decoder.embed_tokens.weight": "model.shared.weight",
-    }
+    _tied_weights_keys = ["encoder.embed_tokens.weight", "decoder.embed_tokens.weight", "lm_head.weight"]
    _keys_to_ignore_on_load_missing = ["final_logits_bias"]

    def __init__(self, config: Florence2LanguageConfig):
@@ -2442,10 +2436,11 @@ FLORENCE2_INPUTS_DOCSTRING = r"""
    FLORENCE2_START_DOCSTRING,
 )
 class Florence2ForConditionalGeneration(Florence2PreTrainedModel):
-    _tied_weights_keys = {
-        "language_model.model.encoder.embed_tokens.weight": "language_model.model.shared.weight",
-        "language_model.model.decoder.embed_tokens.weight": "language_model.model.shared.weight",
-    }
+    _tied_weights_keys = [
+        "language_model.encoder.embed_tokens.weight",
+        "language_model.decoder.embed_tokens.weight",
+        "language_model.lm_head.weight",
+    ]

    def __init__(self, config: Florence2Config):
        super().__init__(config)
@@ -28,7 +28,14 @@ from .core import (
    RobotObservation,
    TransitionKey,
 )
-from .delta_action_processor import MapDeltaActionToRobotActionStep, MapTensorToDeltaActionDictStep
+from .delta_action_processor import (
+    AbsoluteActionsProcessorStep,
+    DeltaActionsProcessorStep,
+    MapDeltaActionToRobotActionStep,
+    MapTensorToDeltaActionDictStep,
+    to_absolute_actions,
+    to_delta_actions,
+)
 from .device_processor import DeviceProcessorStep
 from .factory import (
    make_default_processors,
@@ -44,7 +51,6 @@ from .hil_processor import (
    AddTeleopActionAsComplimentaryDataStep,
    AddTeleopEventsAsInfoStep,
    GripperPenaltyProcessorStep,
-    GymHILAdapterProcessorStep,
    ImageCropResizeProcessorStep,
    InterventionActionProcessorStep,
    RewardClassifierProcessorStep,
@@ -88,7 +94,6 @@ __all__ = [
    "DoneProcessorStep",
    "EnvAction",
    "EnvTransition",
-    "GymHILAdapterProcessorStep",
    "GripperPenaltyProcessorStep",
    "hotswap_stats",
    "IdentityProcessorStep",
@@ -99,6 +104,8 @@ __all__ = [
    "make_default_teleop_action_processor",
    "make_default_robot_action_processor",
    "make_default_robot_observation_processor",
+    "AbsoluteActionsProcessorStep",
+    "DeltaActionsProcessorStep",
    "MapDeltaActionToRobotActionStep",
    "MapTensorToDeltaActionDictStep",
    "NormalizerProcessorStep",
@@ -128,6 +135,8 @@ __all__ = [
    "transition_to_batch",
    "TransitionKey",
    "TruncatedProcessorStep",
+    "to_absolute_actions",
+    "to_delta_actions",
    "UnnormalizerProcessorStep",
    "VanillaObservationProcessorStep",
 ]
@@ -17,7 +17,7 @@
 from __future__ import annotations

 from enum import Enum
-from typing import Any, TypedDict
+from typing import Any, TypeAlias, TypedDict

 import numpy as np
 import torch
@@ -36,10 +36,10 @@ class TransitionKey(str, Enum):
    COMPLEMENTARY_DATA = "complementary_data"


-PolicyAction = torch.Tensor
-RobotAction = dict[str, Any]
-EnvAction = np.ndarray
-RobotObservation = dict[str, Any]
+PolicyAction: TypeAlias = torch.Tensor
+RobotAction: TypeAlias = dict[str, Any]
+EnvAction: TypeAlias = np.ndarray
+RobotObservation: TypeAlias = dict[str, Any]


 EnvTransition = TypedDict(
@@ -14,12 +14,54 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.

-from dataclasses import dataclass
+from collections.abc import Sequence
+from dataclasses import dataclass, field
+from typing import Any
+
+import torch
+from torch import Tensor

 from lerobot.configs.types import FeatureType, PipelineFeatureType, PolicyFeature
+from lerobot.utils.constants import OBS_STATE

-from .core import PolicyAction, RobotAction
-from .pipeline import ActionProcessorStep, ProcessorStepRegistry, RobotActionProcessorStep
+from .core import EnvTransition, PolicyAction, RobotAction, TransitionKey
+from .pipeline import ActionProcessorStep, ProcessorStep, ProcessorStepRegistry, RobotActionProcessorStep
+
+
+def to_delta_actions(actions: Tensor, state: Tensor, mask: Sequence[bool]) -> Tensor:
+    """Convert absolute actions to delta: delta = action - state (for masked dims).
+
+    Args:
+        actions: (B, T, action_dim) or (B, action_dim).
+        state: (B, state_dim). Broadcast across time dimension.
+        mask: Which dims to convert. Can be shorter than action_dim.
+    """
+    mask_t = torch.tensor(mask, dtype=actions.dtype, device=actions.device)
+    dims = mask_t.shape[0]
+    state_offset = state[..., :dims] * mask_t
+    if actions.ndim == 3:
+        state_offset = state_offset.unsqueeze(-2)
+    actions = actions.clone()
+    actions[..., :dims] -= state_offset
+    return actions
+
+
+def to_absolute_actions(actions: Tensor, state: Tensor, mask: Sequence[bool]) -> Tensor:
+    """Convert delta actions back to absolute: absolute = delta + state (for masked dims).
+
+    Args:
+        actions: (B, T, action_dim) or (B, action_dim).
+        state: (B, state_dim). Broadcast across time dimension.
+        mask: Which dims to convert. Can be shorter than action_dim.
+    """
+    mask_t = torch.tensor(mask, dtype=actions.dtype, device=actions.device)
+    dims = mask_t.shape[0]
+    state_offset = state[..., :dims] * mask_t
+    if actions.ndim == 3:
+        state_offset = state_offset.unsqueeze(-2)
+    actions = actions.clone()
+    actions[..., :dims] += state_offset
+    return actions


@ProcessorStepRegistry.register("map_tensor_to_delta_action_dict")
@@ -141,3 +183,126 @@ class MapDeltaActionToRobotActionStep(RobotActionProcessorStep):
            )

        return features
+
+
+@ProcessorStepRegistry.register("delta_actions_processor")
+@dataclass
+class DeltaActionsProcessorStep(ProcessorStep):
+    """Converts absolute actions to delta actions (action -= state) for masked dimensions.
+
+    Mirrors OpenPI's DeltaActions transform. Applied during preprocessing so the model
+    trains on relative offsets instead of absolute positions.
+    Caches the last seen state so a paired AbsoluteActionsProcessorStep can reverse
+    the conversion during postprocessing.
+
+    Attributes:
+        enabled: Whether to apply the delta conversion.
+        exclude_joints: Joint names to keep absolute (not converted to delta).
+        action_names: Action dimension names from dataset metadata, used to build
+            the mask from exclude_joints. If None, all dims are converted.
+    """
+
+    enabled: bool = False
+    exclude_joints: list[str] = field(default_factory=list)
+    action_names: list[str] | None = None
+    _last_state: torch.Tensor | None = field(default=None, init=False, repr=False)
+
+    def _build_mask(self, action_dim: int) -> list[bool]:
+        if not self.exclude_joints or self.action_names is None:
+            return [True] * action_dim
+
+        exclude_tokens = [str(name).lower() for name in self.exclude_joints if name]
+        if not exclude_tokens:
+            return [True] * action_dim
+
+        mask = []
+        for name in self.action_names[:action_dim]:
+            action_name = str(name).lower()
+            is_excluded = any(token == action_name or token in action_name for token in exclude_tokens)
+            mask.append(not is_excluded)
+
+        if len(mask) < action_dim:
+            mask.extend([True] * (action_dim - len(mask)))
+
+        return mask
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        observation = transition.get(TransitionKey.OBSERVATION, {})
+        state = observation.get(OBS_STATE) if observation else None
+
+        # Always cache state for the paired AbsoluteActionsProcessorStep
+        if state is not None:
+            self._last_state = state
+
+        if not self.enabled:
+            return transition
+
+        new_transition = transition.copy()
+        action = new_transition.get(TransitionKey.ACTION)
+        if action is None or state is None:
+            return new_transition
+
+        mask = self._build_mask(action.shape[-1])
+        new_transition[TransitionKey.ACTION] = to_delta_actions(action, state, mask)
+        return new_transition
+
+    def get_config(self) -> dict[str, Any]:
+        return {"enabled": self.enabled, "exclude_joints": self.exclude_joints}
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        return features
+
+
+@ProcessorStepRegistry.register("absolute_actions_processor")
+@dataclass
+class AbsoluteActionsProcessorStep(ProcessorStep):
+    """Converts delta actions back to absolute actions (action += state) for all dimensions.
+
+    Mirrors OpenPI's AbsoluteActions transform. Applied during postprocessing so
+    predicted deltas are converted back to absolute positions for execution.
+    Reads the cached state from its paired DeltaActionsProcessorStep.
+
+    Attributes:
+        enabled: Whether to apply the absolute conversion.
+        delta_step: Reference to the paired DeltaActionsProcessorStep that caches state.
+    """
+
+    enabled: bool = False
+    delta_step: DeltaActionsProcessorStep | None = field(default=None, repr=False)
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        if not self.enabled:
+            return transition
+
+        if self.delta_step is None:
+            raise RuntimeError(
+                "AbsoluteActionsProcessorStep requires a paired DeltaActionsProcessorStep "
+                "but delta_step is None. Ensure delta_step is set when constructing the postprocessor."
+            )
+
+        if self.delta_step._last_state is None:
+            raise RuntimeError(
+                "AbsoluteActionsProcessorStep requires state from DeltaActionsProcessorStep "
+                "but no state has been cached. Ensure the preprocessor runs before the postprocessor."
+            )
+
+        new_transition = transition.copy()
+        action = new_transition.get(TransitionKey.ACTION)
+        if action is None:
+            return new_transition
+
+        mask = self.delta_step._build_mask(action.shape[-1])
+        new_transition[TransitionKey.ACTION] = to_absolute_actions(
+            action, self.delta_step._last_state, mask
+        )
+        return new_transition
+
+    def get_config(self) -> dict[str, Any]:
+        return {"enabled": self.enabled}
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        return features
@@ -20,7 +20,6 @@ from lerobot.configs.types import PipelineFeatureType, PolicyFeature

 from .converters import to_tensor
 from .core import EnvAction, EnvTransition, PolicyAction
-from .hil_processor import TELEOP_ACTION_KEY
 from .pipeline import ActionProcessorStep, ProcessorStep, ProcessorStepRegistry


@@ -90,13 +89,6 @@ class Numpy2TorchActionProcessorStep(ProcessorStep):
            torch_action = to_tensor(action, dtype=None)  # Preserve original dtype
            new_transition[TransitionKey.ACTION] = torch_action

-        complementary_data = new_transition.get(TransitionKey.COMPLEMENTARY_DATA, {})
-        if TELEOP_ACTION_KEY in complementary_data:
-            teleop_action = complementary_data[TELEOP_ACTION_KEY]
-            if isinstance(teleop_action, EnvAction):
-                complementary_data[TELEOP_ACTION_KEY] = to_tensor(teleop_action)
-            new_transition[TransitionKey.COMPLEMENTARY_DATA] = complementary_data
-
        return new_transition

    def transform_features(
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Pepijn	0394fae446	Merge branch 'feat/add_relative_action_pi_models' into feat/mirror	2026-02-22 16:12:10 +01:00
Pepijn	602b8e66a6	fix multi gpu processor bug	2026-02-22 16:11:52 +01:00
Pepijn	ab4dce6fed	revert	2026-02-21 18:48:46 +01:00
Pepijn	40f4386e4a	nccl	2026-02-21 18:44:35 +01:00
Pepijn	87a91b4b08	Merge branch 'feat/add_relative_action_pi_models' into feat/mirror	2026-02-21 18:19:51 +01:00
Pepijn	fadb900c36	compute before dist	2026-02-21 18:19:12 +01:00
Pepijn	de0663226a	max 1m frames	2026-02-21 17:44:12 +01:00
Pepijn	0ca9d66cae	max 1m frames	2026-02-21 17:43:58 +01:00
Pepijn	2222f25da3	Merge branch 'feat/add_relative_action_pi_models' into feat/mirror	2026-02-21 17:28:35 +01:00
Pepijn	acae8417aa	fix	2026-02-21 17:28:26 +01:00
Pepijn	2697f65cf6	stats for entire dataset	2026-02-21 17:15:55 +01:00
Pepijn	74f42f218e	stats for entire dataset	2026-02-21 17:15:45 +01:00
Pepijn	ca9d49e305	Merge branch 'feat/add_relative_action_pi_models' into feat/mirror	2026-02-21 17:12:52 +01:00
Pepijn	6705876d47	use quantiles	2026-02-21 17:12:43 +01:00
Pepijn	aadbd27675	Merge branch 'feat/add_relative_action_pi_models' into feat/mirror	2026-02-21 08:48:39 +01:00
Pepijn	5221647b5e	fix	2026-02-21 08:48:08 +01:00
Pepijn	9c981300dd	stats per chunck	2026-02-21 08:37:38 +01:00
Pepijn	f5b27aad1b	stats per chunck	2026-02-21 08:37:19 +01:00
Pepijn	75f1285507	Merge branch 'feat/add_relative_action_pi_models' into feat/mirror	2026-02-21 08:02:39 +01:00
Pepijn	33cedc2f71	sample 1m	2026-02-21 08:02:25 +01:00
Pepijn	aa32e6c4ab	Merge branch 'feat/add_relative_action_pi_models' into feat/mirror	2026-02-21 07:52:10 +01:00
Pepijn	f906270ec4	load from parquet	2026-02-21 07:51:57 +01:00
Pepijn	733b6d84db	Merge branch 'feat/add_relative_action_pi_models' into feat/mirror	2026-02-21 07:42:00 +01:00
Pepijn	8abc9037a3	sample 100k	2026-02-21 07:41:42 +01:00
Pepijn	e4d4ac0bda	Merge branch 'feat/add_relative_action_pi_models' into feat/mirror	2026-02-21 00:03:37 +01:00
Pepijn	e79b2a439b	calulate chunk based stats	2026-02-21 00:03:21 +01:00
Pepijn	f9ae78ca74	Merge branch 'feat/add_relative_action_pi_models' into feat/mirror	2026-02-20 23:04:36 +01:00
Pepijn	e1ced538e3	only recompute state for stats	2026-02-20 23:04:20 +01:00
Pepijn	2a98602ad6	Merge branch 'feat/add_relative_action_pi_models' into feat/mirror	2026-02-20 22:54:46 +01:00
Pepijn	a2f5b3571e	normalzie after delta conversion	2026-02-20 22:54:29 +01:00
Pepijn	cecf2eff4f	Merge branch 'feat/add_relative_action_pi_models' into feat/mirror	2026-02-20 17:59:19 +01:00
Pepijn	7e6b598a51	add recomputation of stats and option to compute delta stats	2026-02-20 17:59:06 +01:00
Pepijn	4fa41ba806	formatting	2026-02-13 17:46:18 +01:00
Pepijn	1de2b87a92	Add option for pi family models to train with relative actions (relative to state)	2026-02-13 17:45:59 +01:00
Pepijn	e3c511db67	add push to hub	2026-02-05 09:25:49 +01:00
Pepijn	aed4130d39	add swap wrist camera's	2026-02-04 22:47:28 +01:00
Pepijn	d26349c692	add push to hub	2026-02-04 19:17:40 +01:00
Pepijn	a9bce4732b	fix setting metadata	2026-02-04 19:04:43 +01:00
Pepijn	86d69e3c1d	add mirroring	2026-02-04 18:56:51 +01:00