trim episodes

.github/workflows/fast_tests.yml

@@ -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

.github/workflows/full_tests.yml

@@ -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: ${{ needs.build-and-push-docker.outputs.image_tag }} # zizmor: ignore[unpinned-images]
       options: --gpus all --shm-size "16gb"
@@ -181,13 +173,8 @@ jobs:
         shell: bash
         working-directory: /lerobot
     steps:
-      - name: Login to Hugging Face
-        if: env.HF_USER_TOKEN != ''
-        run: |
-          hf auth login --token "$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
+        run: chmod +x /lerobot/.venv/lib/python3.10/site-packages/triton/backends/nvidia/bin/ptxas
       - name: Run pytest on GPU
         run: pytest tests -vv --maxfail=10
       - name: Run end-to-end tests

.github/workflows/nightly.yml

@@ -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

.github/workflows/quality.yml

@@ -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]

.github/workflows/release.yml

@@ -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
 

.github/workflows/unbound_deps_tests.yml

@@ -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

.pre-commit-config.yaml

@@ -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

AI_POLICY.md

@@ -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.

CONTRIBUTING.md

@@ -2,7 +2,7 @@
 
 Everyone is welcome to contribute, and we value everybody's contribution. Code is not the only way to help the community. Answering questions, helping others, reaching out, and improving the documentation are immensely valuable.
 
-Whichever way you choose to contribute, please be mindful to respect our [code of conduct](./CODE_OF_CONDUCT.md) 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
 

README.md

@@ -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!

docker/Dockerfile.internal

@@ -24,7 +24,7 @@ ARG OS_VERSION=22.04
 FROM nvidia/cuda:${CUDA_VERSION}-base-ubuntu${OS_VERSION}
 
 # Define Python version argument
-ARG PYTHON_VERSION=3.12
+ARG PYTHON_VERSION=3.10
 
 # Configure environment variables
 ENV DEBIAN_FRONTEND=noninteractive \

docker/Dockerfile.user

@@ -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

docs/source/async.mdx

@@ -48,7 +48,7 @@ python -m lerobot.async_inference.robot_client \
     --task="dummy" \ # POLICY: The task to run the policy on (`Fold my t-shirt`). Not necessarily defined for all policies, such as `act`
     --policy_type=your_policy_type \ # POLICY: the type of policy to run (smolvla, act, etc)
     --pretrained_name_or_path=user/model \ # POLICY: the model name/path on server to the checkpoint to run (e.g., lerobot/smolvla_base)
-    --policy_device=mps \ # POLICY: the device to run the policy on, on the server (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

docs/source/bring_your_own_policies.mdx

@@ -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
 
 

docs/source/earthrover_mini_plus.mdx

@@ -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
 ```
 

docs/source/envhub.mdx

@@ -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

docs/source/il_robots.mdx

@@ -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=$(NO_COLOR=1 hf auth whoami | awk -F': *' 'NR==1 {print $2}')
+HF_USER=$(hf auth whoami | awk -F': *' 'NR==1 {print $2}')
 echo $HF_USER
 ```
 
@@ -327,7 +327,7 @@ You can look for other LeRobot datasets on the hub by searching for `LeRobot` [t
 You can also push your local dataset to the Hub manually, running:
 
 ```bash
-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 +491,7 @@ If your local computer doesn't have a powerful GPU you could utilize Google Cola
 Once training is done, upload the latest checkpoint with:
 
 ```bash
-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 +499,7 @@ 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
 ```
 

docs/source/installation.mdx

@@ -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]
@@ -72,9 +47,6 @@ ffmpeg -version  # ffmpeg 8.X is not yet supported !
 > 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 +60,23 @@ 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 +90,13 @@ _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 +106,7 @@ 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
 

docs/source/lekiwi.mdx

@@ -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
 ```
 

docs/source/pi0.mdx

@@ -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:

docs/source/pi05.mdx

@@ -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:

docs/source/pi0fast.mdx

@@ -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
 

docs/source/unitree_g1.mdx

@@ -1,72 +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 the Unitree SDK
+### Step 1: Configure Ethernet Interface
 
-Follow the [unitree_sdk2_python installation guide](https://github.com/unitreerobotics/unitree_sdk2_python#installation). Tested with `unitree_sdk2py==1.0.1` and `cyclonedds==0.10.2`:
-
-```bash
-conda create -y -n lerobot python=3.12
-conda activate lerobot
-git clone https://github.com/unitreerobotics/unitree_sdk2_python.git
-cd unitree_sdk2_python
-pip install -e .
-cd ..
-```
-
-### Install LeRobot
-
-```bash
-conda install ffmpeg -c conda-forge
-conda install -c conda-forge "pinocchio>=3.0.0,<4.0.0"
-git clone https://github.com/huggingface/lerobot.git
-cd lerobot
-pip install -e '.[unitree_g1]'
-```
-
-<Tip>
-  For now, pinocchio must be installed from conda-forge (not pip) to include the
-  CasADi bindings needed for arm IK.
-</Tip>
-
-### Test the Installation (Simulation)
-
-The simulation environment has its own dependencies. Check the Simulation environment dependencies: [Unitree G1 Mujoco EnvHub](https://huggingface.co/lerobot/unitree-g1-mujoco/tree/main).
-
-```bash
-pip install mujoco loguru msgpack msgpack-numpy
-```
-
-```bash
-lerobot-teleoperate \
-  --robot.type=unitree_g1 \
-  --robot.is_simulation=true \
-  --teleop.type=unitree_g1 \
-  --teleop.id=wbc_unitree \
-  --robot.cameras='{"global_view": {"type": "zmq", "server_address": "localhost", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30, "warmup_s": 5}}' \
-  --display_data=true \
-  --robot.controller=GrootLocomotionController
-```
-
-This will launch a [MuJoCo sim instance](https://huggingface.co/lerobot/unitree-g1-mujoco/tree/main) for the G1. You can connect a gamepad to your machine before launching in order to control the robot's locomotion in sim. We support both [HolosomaLocomotionController](https://github.com/amazon-far/holosoma) and [GrootLocomotionController](https://github.com/NVlabs/GR00T-WholeBodyControl) via `--robot.controller`.
-
-- Press `9` to release the robot
-- Press `7` / `8` to increase / decrease waist height
-
-### Connect to the Physical Robot
-
-The G1's Ethernet IP is fixed at `192.168.123.164`. Your machine must have a static IP on the same subnet: `192.168.123.x` where `x ≠ 164`.
+Set a static IP on the same subnet as the robot:
 
 ```bash
 # Replace 'enp131s0' with your ethernet interface name (check with `ip a`)
@@ -75,23 +26,47 @@ 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
 ```
 
-### Share Internet via Ethernet
+You should now be connected to the G1's Orin.
 
-The G1 needs internet access to clone repos and install packages. Share your laptop's connection over Ethernet:
+---
+
+## Part 2: Enable WiFi on the Robot
+
+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
+```
+
+### 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
@@ -100,193 +75,223 @@ sudo iptables -A FORWARD -i enp131s0 -o wlp132s0f0 -j ACCEPT
 **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
 ```
 
-### Install the Unitree SDK on the G1
-
-Follow the [unitree_sdk2_python installation guide](https://github.com/unitreerobotics/unitree_sdk2_python#installation):
-
-```bash
-conda create -y -n lerobot python=3.12
-conda activate lerobot
-git clone https://github.com/unitreerobotics/unitree_sdk2_python.git
-cd unitree_sdk2_python
-python -m pip install -e .
-cd ..
-```
-
-### Install LeRobot on the G1
-
-```bash
-git clone https://github.com/huggingface/lerobot.git
-cd lerobot
-conda install -c conda-forge "pinocchio>=3.0.0,<4.0.0"
-python -m pip install -e '.[unitree_g1]'
-```
-
-<Tip>
-  For now, pinocchio must be installed from conda-forge (not pip) to include the
-  CasADi bindings needed for arm IK.
-</Tip>
-
-### (Optional) Enable WiFi on the Robot
-
-For wireless SSH access, you can enable WiFi on the G1 (it's blocked by default):
-
-```bash
-sudo rfkill unblock all
-sudo ip link set wlan0 up
-sudo nmcli radio wifi on
-sudo nmcli device set wlan0 managed yes
-sudo systemctl restart NetworkManager
-```
-
-**Connect to a WiFi network:**
+### 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 then SSH over WiFi instead of Ethernet:
+### 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
 ```
 
----
-
-## Part 2: Teleoperation & Locomotion
-
-### Run the Robot Server
-
-On the robot (from `~/lerobot`):
-
-```bash
-cd ~/lerobot
-python src/lerobot/robots/unitree_g1/run_g1_server.py --camera
-```
-
-### Run the Locomotion Policy
-
-You can run the teleoperation client from your laptop over Ethernet, over WiFi (experimental), or directly on the robot itself. Mind potential latency introduced by your network.
-
-**From your laptop:**
-
-```bash
-lerobot-teleoperate \
-  --robot.type=unitree_g1 \
-  --robot.is_simulation=false \
-  --robot.robot_ip=<ROBOT_IP> \
-  --teleop.type=unitree_g1 \
-  --teleop.id=wbc_unitree \
-  --robot.cameras='{"global_view": {"type": "zmq", "server_address": "<ROBOT_IP>", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
-  --display_data=true \
-  --robot.controller=HolosomaLocomotionController
-```
-
-We support both [GrootLocomotionController](https://github.com/NVlabs/GR00T-WholeBodyControl) and [HolosomaLocomotionController](https://github.com/amazon-far/holosoma) via `--robot.controller`.
+Replace `<YOUR_ROBOT_IP>` with your robot's actual WiFi IP address.
 
 ---
 
-## Part 3: Loco-Manipulation with the Homunculus Exoskeleton
+## Part 3: Robot Server Setup
 
-We provide a loco-manipulation solution via the Homunculus Exoskeleton — an open-source 7 DoF exoskeleton for whole-body control. Check it out [here](https://github.com/nepyope/hmc_exo).
+### Step 1: Install LeRobot on the Orin
 
-### Calibrate
+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
+```
+
+**Important**: Keep this terminal running. The server must be active for remote control.
+
+---
+
+## Part 4: Controlling the robot
+
+With the robot server running, you can now control the robot remotely. Let's launch a locomotion policy
+
+### Step 1: Install LeRobot on your machine
+
+```bash
+conda create -y -n lerobot python=3.10
+conda 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.
+### Teleoperate in Simulation
 
-### Record a Dataset
+```bash
+lerobot-teleoperate \
+    --robot.type=unitree_g1 \
+    --robot.is_simulation=true \
+    --teleop.type=unitree_g1 \
+    --teleop.left_arm_config.port=/dev/ttyACM1 \
+    --teleop.right_arm_config.port=/dev/ttyACM0 \
+    --teleop.id=exo \
+    --fps=100
+```
+
+### Record Dataset in Simulation
 
 ```bash
 lerobot-record \
-  --robot.type=unitree_g1 \
-  --robot.is_simulation=true \
-  --robot.cameras='{"global_view": {"type": "zmq", "server_address": "localhost", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
-  --teleop.type=unitree_g1 \
-  --teleop.left_arm_config.port=/dev/ttyACM1 \
-  --teleop.right_arm_config.port=/dev/ttyACM0 \
-  --teleop.id=exo \
-  --dataset.repo_id=your-username/dataset-name \
-  --dataset.single_task="Test" \
-  --dataset.num_episodes=2 \
-  --dataset.episode_time_s=5 \
-  --dataset.reset_time_s=5 \
-  --dataset.push_to_hub=true \
-  --dataset.streaming_encoding=true \
-  --dataset.encoder_threads=2
+    --robot.type=unitree_g1 \
+    --robot.is_simulation=true \
+    --robot.cameras='{"global_view": {"type": "zmq", "server_address": "localhost", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
+    --teleop.type=unitree_g1 \
+    --teleop.left_arm_config.port=/dev/ttyACM1 \
+    --teleop.right_arm_config.port=/dev/ttyACM0 \
+    --teleop.id=exo \
+    --dataset.repo_id=your-username/dataset-name \
+    --dataset.single_task="Test" \
+    --dataset.num_episodes=2 \
+    --dataset.episode_time_s=5 \
+    --dataset.reset_time_s=5 \
+    --dataset.push_to_hub=true \
+    --dataset.streaming_encoding=true \
+    # --dataset.vcodec=auto \
+    --dataset.encoder_threads=2
 ```
 
-> **Note:** Omit `--teleop.left_arm_config.port` and `--teleop.right_arm_config.port` if you're only using the joystick.
-
-Example dataset: [nepyope/unitree_box_move_blue_full](https://huggingface.co/datasets/nepyope/unitree_box_move_blue_full)
+Example simulation dataset: [nepyope/teleop_test_sim](https://huggingface.co/datasets/nepyope/teleop_test_sim)
 
 ---
 
-## Part 4: 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
+lerobot-record \
+    --robot.type=unitree_g1 \
+    --robot.is_simulation=false \
+    --robot.cameras='{"global_view": {"type": "zmq", "server_address": "172.18.129.215", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
+    --teleop.type=unitree_g1 \
+    --teleop.left_arm_config.port=/dev/ttyACM1 \
+    --teleop.right_arm_config.port=/dev/ttyACM0 \
+    --teleop.id=exo \
+    --dataset.repo_id=your-username/dataset-name \
+    --dataset.single_task="Test" \
+    --dataset.num_episodes=2 \
+    --dataset.episode_time_s=5 \
+    --dataset.reset_time_s=5 \
+    --dataset.push_to_hub=true \
+    --dataset.streaming_encoding=true \
+    # --dataset.vcodec=auto \
+    --dataset.encoder_threads=2
+```
+
+**Note**: Update `server_address` to match your robot's camera server IP.
+
+Example real robot dataset: [nepyope/teleop_test_real](https://huggingface.co/datasets/nepyope/teleop_test_real)
+
 ---
 
 ## Additional Resources
@@ -295,8 +300,8 @@ 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_

examples/backward_compatibility/replay.py

@@ -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

examples/dataset/slurm_compute_rabc.py

@@ -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()

examples/rtc/eval_with_real_robot.py

@@ -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

examples/unitree_g1/gr00t_locomotion.py

@@ -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)

examples/unitree_g1/holosoma_locomotion.py

@@ -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)

pyproject.toml

@@ -25,11 +25,11 @@ discord = "https://discord.gg/s3KuuzsPFb"
 
 [project]
 name = "lerobot"
-version = "0.5.1"
+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",
 ]
@@ -62,28 +61,26 @@ dependencies = [
     # Hugging Face dependencies
     "datasets>=4.0.0,<5.0.0",
     "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,<4.14.0",
+    "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,<2.11.0", # TODO: Bump dependency
+    "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')", # TODO: Bump dependency
+    "torchvision>=0.21.0,<0.26.0", # TODO: Bump dependency
+
+    "draccus==0.10.0", # TODO: Remove ==
     "gymnasium>=1.1.1,<2.0.0",
     "rerun-sdk>=0.24.0,<0.27.0",
 
@@ -98,14 +95,10 @@ 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"]
@@ -119,35 +112,34 @@ 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",
-    "onnx>=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",
@@ -156,13 +148,13 @@ 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"]
@@ -170,19 +162,13 @@ 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]",
@@ -190,8 +176,8 @@ 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]",
@@ -203,11 +189,10 @@ 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]
@@ -236,7 +221,7 @@ lerobot = ["envs/*.json"]
 where = ["src"]
 
 [tool.ruff]
-target-version = "py312"
+target-version = "py310"
 line-length = 110
 exclude = ["tests/artifacts/**/*.safetensors", "*_pb2.py", "*_pb2_grpc.py"]
 
@@ -328,7 +313,7 @@ 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
@@ -412,3 +397,85 @@ 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" },
+    ],
+]

requirements-macos.txt

@@ -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

requirements-ubuntu.txt

@@ -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

requirements.in

@@ -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]

src/lerobot/async_inference/robot_client.py

@@ -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

src/lerobot/cameras/zmq/camera_zmq.py

@@ -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

src/lerobot/cameras/zmq/image_server.py

@@ -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()

src/lerobot/configs/default.py

@@ -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)

src/lerobot/configs/train.py

@@ -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

src/lerobot/datasets/aggregate.py

@@ -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}

src/lerobot/datasets/dataset_tools.py

@@ -47,6 +47,7 @@ from lerobot.datasets.utils import (
     DEFAULT_EPISODES_PATH,
     get_parquet_file_size_in_mb,
     load_episodes,
+    load_info,
     update_chunk_file_indices,
     write_info,
     write_stats,
@@ -89,8 +90,8 @@ 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 +153,7 @@ 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 +244,8 @@ 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 +289,8 @@ 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 +391,8 @@ 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 +428,8 @@ 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.
@@ -1475,9 +1476,7 @@ 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}")
@@ -1531,7 +1530,7 @@ def modify_tasks(
 
 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 +1549,8 @@ 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 +1601,6 @@ 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,
@@ -1777,3 +1775,296 @@ def convert_image_to_video_dataset(
 
     # Return new dataset
     return LeRobotDataset(repo_id=repo_id, root=output_dir)
+
+
+def trim_episodes_by_frames(
+    dataset: LeRobotDataset,
+    episode_frames_to_keep: dict[int, list[int]],
+    output_dir: str | Path | None = None,
+    repo_id: str | None = None,
+) -> LeRobotDataset:
+    """Trim multiple episodes to keep only specific frames.
+
+    This function creates a new dataset where the specified episodes contain only
+    the frames at the given indices. All other episodes are copied as-is.
+
+    Args:
+        dataset: The source LeRobotDataset.
+        episode_frames_to_keep: Dict mapping episode indices to lists of global frame indices to keep.
+        output_dir: Directory to save the new dataset. If None, uses default location.
+        repo_id: Repository ID for the new dataset. If None, appends "_trimmed" to original.
+
+    Returns:
+        A new LeRobotDataset with the trimmed episodes.
+    """
+    if not episode_frames_to_keep:
+        raise ValueError("No episodes to trim")
+
+    for ep_idx in episode_frames_to_keep:
+        if ep_idx >= dataset.meta.total_episodes:
+            raise ValueError(f"Episode {ep_idx} does not exist")
+        if not episode_frames_to_keep[ep_idx]:
+            raise ValueError(f"No frames to keep for episode {ep_idx}")
+
+    if repo_id is None:
+        repo_id = f"{dataset.repo_id}_trimmed"
+    output_dir = Path(output_dir) if output_dir is not None else HF_LEROBOT_HOME / repo_id
+
+    total_trimmed = sum(len(frames) for frames in episode_frames_to_keep.values())
+    logging.info(f"Trimming {len(episode_frames_to_keep)} episodes, keeping {total_trimmed} frames total")
+
+    # Create new metadata
+    new_meta = LeRobotDatasetMetadata.create(
+        repo_id=repo_id,
+        fps=dataset.meta.fps,
+        features=dataset.meta.features,
+        robot_type=dataset.meta.robot_type,
+        root=output_dir,
+        use_videos=len(dataset.meta.video_keys) > 0,
+    )
+
+    # Build set of all frames to keep (for episodes being trimmed)
+    # and compute new frame counts per episode
+    all_keep_frames: set[int] = set()
+    trimmed_frame_counts: dict[int, int] = {}
+    for ep_idx, frames in episode_frames_to_keep.items():
+        all_keep_frames.update(frames)
+        trimmed_frame_counts[ep_idx] = len(frames)
+
+    # Copy and filter data
+    _copy_and_reindex_data_with_multi_frame_filter(
+        dataset, new_meta, episode_frames_to_keep, all_keep_frames
+    )
+
+    # Handle videos if present
+    if dataset.meta.video_keys:
+        _copy_and_reindex_videos_with_multi_frame_filter(
+            dataset, new_meta, episode_frames_to_keep
+        )
+
+    # Copy episode metadata
+    _copy_and_reindex_episodes_metadata_for_multi_trim(
+        dataset, new_meta, trimmed_frame_counts
+    )
+
+    logging.info(f"Created trimmed dataset with {new_meta.total_frames} frames at {output_dir}")
+    
+    # Return the metadata instead of trying to load as LeRobotDataset
+    # This avoids Hub validation issues when the repo doesn't exist yet
+    return new_meta
+
+
+# Keep old function for backward compatibility
+def trim_episode_by_frames(
+    dataset: LeRobotDataset,
+    episode_index: int,
+    keep_frame_indices: list[int],
+    output_dir: str | Path | None = None,
+    repo_id: str | None = None,
+) -> LeRobotDataset:
+    """Trim a single episode. Wrapper around trim_episodes_by_frames."""
+    return trim_episodes_by_frames(
+        dataset,
+        episode_frames_to_keep={episode_index: keep_frame_indices},
+        output_dir=output_dir,
+        repo_id=repo_id,
+    )
+
+
+def _copy_and_reindex_data_with_multi_frame_filter(
+    src_dataset: LeRobotDataset,
+    dst_meta: LeRobotDatasetMetadata,
+    episode_frames_to_keep: dict[int, list[int]],
+    all_keep_frames: set[int],
+) -> None:
+    """Copy data files with frame-level filtering for multiple episodes."""
+    if src_dataset.meta.episodes is None:
+        src_dataset.meta.episodes = load_episodes(src_dataset.meta.root)
+
+    # Copy tasks
+    if dst_meta.tasks is None and src_dataset.meta.tasks is not None:
+        # Tasks are stored with task string as index
+        dst_meta.save_episode_tasks(list(src_dataset.meta.tasks.index))
+
+    # Get all parquet files
+    data_dir = src_dataset.root / "data"
+    parquet_files = sorted(data_dir.glob("chunk-*/file-*.parquet"))
+
+    trim_episode_set = set(episode_frames_to_keep.keys())
+    global_index = 0
+
+    for parquet_path in tqdm(parquet_files, desc="Processing data files"):
+        df = pd.read_parquet(parquet_path)
+
+        # Filter: keep all frames from non-trimmed episodes, 
+        # and only specified frames from trimmed episodes
+        mask = (~df["episode_index"].isin(trim_episode_set)) | (df["index"].isin(all_keep_frames))
+        df = df[mask].copy().reset_index(drop=True)
+
+        if len(df) == 0:
+            continue
+
+        # Reindex
+        df["index"] = range(global_index, global_index + len(df))
+
+        # Recalculate frame_index within each episode
+        for ep_idx in df["episode_index"].unique():
+            ep_mask = df["episode_index"] == ep_idx
+            df.loc[ep_mask, "frame_index"] = range(ep_mask.sum())
+
+        # Recalculate timestamps based on frame_index and fps
+        df["timestamp"] = df["frame_index"] / src_dataset.meta.fps
+
+        # Determine output path (keep same structure)
+        rel_path = parquet_path.relative_to(src_dataset.root)
+        dst_path = dst_meta.root / rel_path
+        dst_path.parent.mkdir(parents=True, exist_ok=True)
+
+        _write_parquet(df, dst_path, dst_meta)
+        global_index += len(df)
+
+
+def _copy_and_reindex_videos_with_multi_frame_filter(
+    src_dataset: LeRobotDataset,
+    dst_meta: LeRobotDatasetMetadata,
+    episode_frames_to_keep: dict[int, list[int]],
+) -> None:
+    """Copy video files for trimmed dataset.
+    
+    In v3.0 datasets, multiple episodes are concatenated into single video files.
+    Each episode has from_timestamp/to_timestamp indicating its portion of the video.
+    
+    For trimming, we copy the original video files as-is and update the metadata
+    timestamps in _copy_and_reindex_episodes_metadata_for_multi_trim.
+    """
+    for video_key in src_dataset.meta.video_keys:
+        video_dir = src_dataset.root / "videos" / video_key
+        dst_video_dir = dst_meta.root / "videos" / video_key
+        
+        if not video_dir.exists():
+            logging.warning(f"Video directory not found: {video_dir}")
+            continue
+        
+        # Copy all video files (they contain concatenated episodes)
+        # The metadata timestamps will handle which portions to use
+        copied_files = set()
+        for chunk_dir in video_dir.glob("chunk-*"):
+            dst_chunk_dir = dst_video_dir / chunk_dir.name
+            dst_chunk_dir.mkdir(parents=True, exist_ok=True)
+            
+            for video_file in chunk_dir.glob("*.mp4"):
+                if video_file.name not in copied_files:
+                    dst_path = dst_chunk_dir / video_file.name
+                    if not dst_path.exists():
+                        shutil.copy(video_file, dst_path)
+                        copied_files.add(video_file.name)
+        
+        logging.info(f"Copied {len(copied_files)} video files for {video_key}")
+
+
+def _trim_video_frames(
+    src_path: Path,
+    dst_path: Path,
+    keep_frame_indices: list[int],
+    fps: float,
+    episode_start_idx: int,
+) -> None:
+    """Trim a video to keep only specific frames using ffmpeg."""
+    import subprocess
+
+    # Convert global indices to local indices within the episode
+    local_indices = sorted([idx - episode_start_idx for idx in keep_frame_indices])
+
+    if not local_indices:
+        logging.warning(f"No frames to keep for video {src_path}")
+        return
+
+    # Calculate start and end times
+    start_frame = local_indices[0]
+    end_frame = local_indices[-1]
+
+    start_time = start_frame / fps
+    duration = (end_frame - start_frame + 1) / fps
+
+    # Use ffmpeg to trim
+    cmd = [
+        "ffmpeg", "-y",
+        "-ss", str(start_time),
+        "-i", str(src_path),
+        "-t", str(duration),
+        "-c", "copy",  # Fast copy without re-encoding
+        str(dst_path)
+    ]
+
+    try:
+        subprocess.run(cmd, check=True, capture_output=True)
+    except subprocess.CalledProcessError as e:
+        logging.error(f"Failed to trim video: {e.stderr.decode()}")
+        # Fallback: copy the whole video
+        shutil.copy(src_path, dst_path)
+
+
+def _copy_and_reindex_episodes_metadata_for_multi_trim(
+    src_dataset: LeRobotDataset,
+    dst_meta: LeRobotDatasetMetadata,
+    trimmed_frame_counts: dict[int, int],
+) -> None:
+    """Copy and update episode metadata for trimmed dataset."""
+    if src_dataset.meta.episodes is None:
+        src_dataset.meta.episodes = load_episodes(src_dataset.meta.root)
+
+    # Calculate new frame counts and indices
+    episodes_data = []
+    global_idx = 0
+
+    for old_ep_idx in range(src_dataset.meta.total_episodes):
+        src_ep = src_dataset.meta.episodes[old_ep_idx]
+
+        if old_ep_idx in trimmed_frame_counts:
+            ep_length = trimmed_frame_counts[old_ep_idx]
+        else:
+            ep_length = src_ep["length"]
+
+        ep_data = {
+            "episode_index": old_ep_idx,
+            "tasks": src_ep["tasks"],
+            "length": ep_length,
+            "data/chunk_index": src_ep["data/chunk_index"],
+            "data/file_index": src_ep["data/file_index"],
+            "dataset_from_index": global_idx,
+            "dataset_to_index": global_idx + ep_length,
+        }
+
+        # Copy video metadata - preserve timestamps for concatenated videos
+        for video_key in src_dataset.meta.video_keys:
+            ep_data[f"videos/{video_key}/chunk_index"] = src_ep[f"videos/{video_key}/chunk_index"]
+            ep_data[f"videos/{video_key}/file_index"] = src_ep[f"videos/{video_key}/file_index"]
+            
+            # Keep original from_timestamp (start position in concatenated video)
+            orig_from_ts = src_ep[f"videos/{video_key}/from_timestamp"]
+            ep_data[f"videos/{video_key}/from_timestamp"] = orig_from_ts
+            
+            # For trimmed episodes, update to_timestamp based on new length
+            # For non-trimmed episodes, keep original to_timestamp
+            if old_ep_idx in trimmed_frame_counts:
+                ep_data[f"videos/{video_key}/to_timestamp"] = orig_from_ts + (ep_length / src_dataset.meta.fps)
+            else:
+                ep_data[f"videos/{video_key}/to_timestamp"] = src_ep[f"videos/{video_key}/to_timestamp"]
+
+        ep_data["meta/episodes/chunk_index"] = 0
+        ep_data["meta/episodes/file_index"] = 0
+
+        episodes_data.append(ep_data)
+        global_idx += ep_length
+
+    # Save episodes metadata
+    df = pd.DataFrame(episodes_data)
+    episodes_path = dst_meta.root / DEFAULT_EPISODES_PATH.format(chunk_index=0, file_index=0)
+    episodes_path.parent.mkdir(parents=True, exist_ok=True)
+    df.to_parquet(episodes_path)
+
+    # Update info.json
+    info = load_info(src_dataset.root)
+    info["total_episodes"] = len(episodes_data)
+    info["total_frames"] = global_idx
+    write_info(info, dst_meta.root)

src/lerobot/datasets/lerobot_dataset.py

@@ -314,7 +314,7 @@ 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))
@@ -664,11 +664,11 @@ 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 HF_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
@@ -1771,12 +1771,11 @@ 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

src/lerobot/datasets/utils.py

@@ -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)
@@ -339,7 +341,6 @@ 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
 
 
@@ -578,7 +579,7 @@ def get_hf_features_from_features(features: dict) -> datasets.Features:
             continue
         elif ft["dtype"] == "image":
             hf_features[key] = datasets.Image()
-        elif ft["shape"] == (1,) and ft["names"] is None:
+        elif ft["shape"] == (1,):
             hf_features[key] = datasets.Value(dtype=ft["dtype"])
         elif len(ft["shape"]) == 1:
             hf_features[key] = datasets.Sequence(
@@ -1232,7 +1233,7 @@ 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.

src/lerobot/datasets/v30/convert_dataset_v21_to_v30.py

@@ -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}.
 ```
 
 """
@@ -57,7 +54,6 @@ import pyarrow as pa
 import tqdm
 from datasets import Dataset, Features, Image
 from huggingface_hub import HfApi, snapshot_download
-from huggingface_hub.errors import RevisionNotFoundError
 from requests import HTTPError
 
 from lerobot.datasets.compute_stats import aggregate_stats
@@ -109,7 +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
@@ -117,16 +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
@@ -175,7 +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)
 
 
@@ -206,6 +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
@@ -215,23 +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,
@@ -242,7 +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:
@@ -474,7 +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
@@ -512,7 +507,7 @@ def convert_dataset(
         hub_api = HfApi()
         try:
             hub_api.delete_tag(repo_id, tag=CODEBASE_VERSION, repo_type="dataset")
-        except(HTTPError,RevisionNotFoundError) as e:
+        except HTTPError as e:
             print(f"tag={CODEBASE_VERSION} probably doesn't exist. Skipping exception ({e})")
             pass
         hub_api.delete_files(
@@ -558,7 +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",

src/lerobot/motors/motors_bus.py

@@ -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

src/lerobot/policies/diffusion/configuration_diffusion.py

@@ -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
@@ -183,25 +175,6 @@ 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 +202,13 @@ 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.

src/lerobot/policies/diffusion/modeling_diffusion.py

@@ -454,18 +454,12 @@ 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 +485,13 @@ 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 +507,7 @@ 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)

src/lerobot/policies/factory.py

@@ -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

src/lerobot/policies/groot/eagle2_hg_model/image_processing_eagle2_5_vl_fast.py

@@ -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,

src/lerobot/policies/pi0/modeling_pi0.py

@@ -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

src/lerobot/policies/pi05/modeling_pi05.py

@@ -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

src/lerobot/policies/pi05/processor_pi05.py

@@ -23,6 +23,7 @@ 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 (
     AddBatchDimensionProcessorStep,
     DeviceProcessorStep,
@@ -67,6 +68,9 @@ 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()

src/lerobot/policies/pi0_fast/configuration_pi0_fast.py

@@ -54,7 +54,7 @@ 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

src/lerobot/policies/pi0_fast/modeling_pi0_fast.py

@@ -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
 
@@ -37,16 +38,11 @@ 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
@@ -125,7 +121,7 @@ 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 +132,7 @@ 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 +206,16 @@ 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 +228,10 @@ 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 +242,10 @@ 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 +259,7 @@ 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 +306,24 @@ 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 +332,8 @@ 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 +523,7 @@ 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 +616,7 @@ 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 +714,7 @@ 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 +897,14 @@ 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 +912,7 @@ 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 +925,8 @@ 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 +936,8 @@ 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 +971,7 @@ 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
 

src/lerobot/policies/pi0_fast/processor_pi0_fast.py

@@ -23,6 +23,7 @@ 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,
@@ -68,6 +69,9 @@ 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()

src/lerobot/policies/pi_gemma.py

@@ -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",
-]

src/lerobot/policies/pretrained.py

@@ -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

src/lerobot/policies/sac/reward_model/configuration_classifier.py

@@ -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

src/lerobot/policies/smolvla/configuration_smolvla.py

@@ -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__()
 

src/lerobot/policies/smolvla/modeling_smolvla.py

@@ -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
 

src/lerobot/policies/smolvla/smolvlm_with_expert.py

@@ -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,
             )

src/lerobot/policies/wall_x/configuration_wall_x.py

@@ -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"

src/lerobot/policies/wall_x/modeling_wall_x.py

@@ -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),
             )

src/lerobot/policies/wall_x/qwen_model/configuration_qwen2_5_vl.py

@@ -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):

src/lerobot/policies/wall_x/qwen_model/qwen2_5_vl_moe.py

@@ -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"]
 

src/lerobot/policies/wall_x/utils.py

@@ -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 = {}

src/lerobot/policies/xvla/configuration_florence2.py

@@ -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(

src/lerobot/policies/xvla/modeling_florence2.py

@@ -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)

src/lerobot/processor/core.py

@@ -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(

src/lerobot/processor/pipeline.py

@@ -39,7 +39,7 @@ from collections.abc import Callable, Iterable, Sequence
 from copy import deepcopy
 from dataclasses import dataclass, field
 from pathlib import Path
-from typing import Any, TypedDict, TypeVar, cast
+from typing import Any, Generic, TypeAlias, TypedDict, TypeVar, cast
 
 import torch
 from huggingface_hub import hf_hub_download
@@ -251,7 +251,7 @@ class ProcessorMigrationError(Exception):
 
 
 @dataclass
-class DataProcessorPipeline[TInput, TOutput](HubMixin):
+class DataProcessorPipeline(HubMixin, Generic[TInput, TOutput]):
     """A sequential pipeline for processing data, integrated with the Hugging Face Hub.
 
     This class chains together multiple `ProcessorStep` instances to form a complete
@@ -1432,8 +1432,8 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
 
 
 # Type aliases for semantic clarity.
-RobotProcessorPipeline = DataProcessorPipeline[TInput, TOutput]
-PolicyProcessorPipeline = DataProcessorPipeline[TInput, TOutput]
+RobotProcessorPipeline: TypeAlias = DataProcessorPipeline[TInput, TOutput]
+PolicyProcessorPipeline: TypeAlias = DataProcessorPipeline[TInput, TOutput]
 
 
 class ObservationProcessorStep(ProcessorStep, ABC):

src/lerobot/processor/tokenizer_processor.py

@@ -336,7 +336,7 @@ class ActionTokenizerProcessorStep(ActionProcessorStep):
     Requires the `transformers` library to be installed.
 
     Attributes:
-        tokenizer_name: The name of a pretrained processor from the Hugging Face Hub (e.g., "lerobot/fast-action-tokenizer").
+        tokenizer_name: The name of a pretrained processor from the Hugging Face Hub (e.g., "physical-intelligence/fast").
         tokenizer: A pre-initialized processor/tokenizer object. If provided, `tokenizer_name` is ignored.
         trust_remote_code: Whether to trust remote code when loading the tokenizer (required for some tokenizers).
         action_tokenizer: The internal tokenizer/processor instance, loaded during initialization.

src/lerobot/robots/so_follower/config_so_follower.py

@@ -15,6 +15,7 @@
 # limitations under the License.
 
 from dataclasses import dataclass, field
+from typing import TypeAlias
 
 from lerobot.cameras import CameraConfig
 
@@ -49,5 +50,5 @@ class SOFollowerRobotConfig(RobotConfig, SOFollowerConfig):
     pass
 
 
-SO100FollowerConfig = SOFollowerRobotConfig
-SO101FollowerConfig = SOFollowerRobotConfig
+SO100FollowerConfig: TypeAlias = SOFollowerRobotConfig
+SO101FollowerConfig: TypeAlias = SOFollowerRobotConfig

src/lerobot/robots/so_follower/so_follower.py

@@ -17,6 +17,7 @@
 import logging
 import time
 from functools import cached_property
+from typing import TypeAlias
 
 from lerobot.cameras.utils import make_cameras_from_configs
 from lerobot.motors import Motor, MotorCalibration, MotorNormMode
@@ -229,5 +230,5 @@ class SOFollower(Robot):
         logger.info(f"{self} disconnected.")
 
 
-SO100Follower = SOFollower
-SO101Follower = SOFollower
+SO100Follower: TypeAlias = SOFollower
+SO101Follower: TypeAlias = SOFollower

src/lerobot/robots/unitree_g1/__init__.py

@@ -16,5 +16,3 @@
 
 from .config_unitree_g1 import UnitreeG1Config
 from .unitree_g1 import UnitreeG1
-
-__all__ = ["UnitreeG1", "UnitreeG1Config"]

src/lerobot/robots/unitree_g1/config_unitree_g1.py

@@ -27,10 +27,11 @@ _GAINS: dict[str, dict[str, list[float]]] = {
     },  # pitch, roll, yaw, knee, ankle_pitch, ankle_roll
     "right_leg": {"kp": [150, 150, 150, 300, 40, 40], "kd": [2, 2, 2, 4, 2, 2]},
     "waist": {"kp": [250, 250, 250], "kd": [5, 5, 5]},  # yaw, roll, pitch
-    "left_arm": {"kp": [50, 50, 80, 80], "kd": [3, 3, 3, 3]},  # shoulder_pitch/roll/yaw, elbow
+    "left_arm": {"kp": [80, 80, 80, 80], "kd": [3, 3, 3, 3]},  # shoulder_pitch/roll/yaw, elbow
     "left_wrist": {"kp": [40, 40, 40], "kd": [1.5, 1.5, 1.5]},  # roll, pitch, yaw
-    "right_arm": {"kp": [50, 50, 80, 80], "kd": [3, 3, 3, 3]},
+    "right_arm": {"kp": [80, 80, 80, 80], "kd": [3, 3, 3, 3]},
     "right_wrist": {"kp": [40, 40, 40], "kd": [1.5, 1.5, 1.5]},
+    "other": {"kp": [80, 80, 80, 80, 80, 80], "kd": [3, 3, 3, 3, 3, 3]},
 }
 
 
@@ -67,7 +68,3 @@ class UnitreeG1Config(RobotConfig):
 
     # Compensates for gravity on the unitree's arms using the arm ik solver
     gravity_compensation: bool = False
-
-    # Lower-body controller class name, e.g. "GrootLocomotionController" or
-    # "HolosomaLocomotionController". None disables it.
-    controller: str | None = None

src/lerobot/robots/unitree_g1/g1_utils.py

@@ -14,34 +14,12 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import importlib
 from enum import IntEnum
 
-import numpy as np
-
 # ruff: noqa: N801, N815
 
 NUM_MOTORS = 29
 
-REMOTE_AXES = ("remote.lx", "remote.ly", "remote.rx", "remote.ry")
-REMOTE_BUTTONS = tuple(f"remote.button.{i}" for i in range(16))
-REMOTE_KEYS = REMOTE_AXES + REMOTE_BUTTONS
-
-
-def default_remote_input() -> dict[str, float]:
-    """Return a zeroed-out remote input dict (axes + buttons)."""
-    return dict.fromkeys(REMOTE_KEYS, 0.0)
-
-
-def get_gravity_orientation(quaternion: list[float] | np.ndarray) -> np.ndarray:
-    """Get gravity orientation from quaternion [w, x, y, z]."""
-    qw, qx, qy, qz = quaternion
-    gravity_orientation = np.zeros(3, dtype=np.float32)
-    gravity_orientation[0] = 2 * (-qz * qx + qw * qy)
-    gravity_orientation[1] = -2 * (qz * qy + qw * qx)
-    gravity_orientation[2] = 1 - 2 * (qw * qw + qz * qz)
-    return gravity_orientation
-
 
 class G1_29_JointArmIndex(IntEnum):
     # Left arm
@@ -51,7 +29,7 @@ class G1_29_JointArmIndex(IntEnum):
     kLeftElbow = 18
     kLeftWristRoll = 19
     kLeftWristPitch = 20
-    kLeftWristYaw = 21
+    kLeftWristyaw = 21
 
     # Right arm
     kRightShoulderPitch = 22
@@ -63,21 +41,6 @@ class G1_29_JointArmIndex(IntEnum):
     kRightWristYaw = 28
 
 
-def make_locomotion_controller(name: str | None):
-    """Instantiate a locomotion controller by class name. Returns None if name is None."""
-    if name is None:
-        return None
-    controllers = {
-        "GrootLocomotionController": "lerobot.robots.unitree_g1.gr00t_locomotion",
-        "HolosomaLocomotionController": "lerobot.robots.unitree_g1.holosoma_locomotion",
-    }
-    module_path = controllers.get(name)
-    if module_path is None:
-        raise ValueError(f"Unknown controller: {name!r}. Available: {list(controllers)}")
-    module = importlib.import_module(module_path)
-    return getattr(module, name)()
-
-
 class G1_29_JointIndex(IntEnum):
     # Left leg
     kLeftHipPitch = 0
@@ -106,7 +69,7 @@ class G1_29_JointIndex(IntEnum):
     kLeftElbow = 18
     kLeftWristRoll = 19
     kLeftWristPitch = 20
-    kLeftWristYaw = 21
+    kLeftWristyaw = 21
 
     # Right arm
     kRightShoulderPitch = 22

src/lerobot/robots/unitree_g1/robot_kinematic_processor.py

@@ -16,11 +16,13 @@
 
 import logging
 import os
-from collections import deque
+import sys
 
 import numpy as np
 
 logger = logging.getLogger(__name__)
+parent2_dir = os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+sys.path.append(parent2_dir)
 
 
 class WeightedMovingFilter:
@@ -29,14 +31,18 @@ class WeightedMovingFilter:
         self._weights = np.array(weights)
         self._data_size = data_size
         self._filtered_data = np.zeros(self._data_size)
-        self._data_queue = deque(maxlen=self._window_size)
+        self._data_queue = []
 
     def _apply_filter(self):
         if len(self._data_queue) < self._window_size:
             return self._data_queue[-1]
 
         data_array = np.array(self._data_queue)
-        return data_array.T @ self._weights
+        temp_filtered_data = np.zeros(self._data_size)
+        for i in range(self._data_size):
+            temp_filtered_data[i] = np.convolve(data_array[:, i], self._weights, mode="valid")[-1]
+
+        return temp_filtered_data
 
     def add_data(self, new_data):
         assert len(new_data) == self._data_size
@@ -46,6 +52,9 @@ class WeightedMovingFilter:
         ):  # skip duplicate data
             return
 
+        if len(self._data_queue) >= self._window_size:
+            self._data_queue.pop(0)
+
         self._data_queue.append(new_data)
         self._filtered_data = self._apply_filter()
 
@@ -62,6 +71,8 @@ class G1_29_ArmIK:  # noqa: N801
         from pinocchio import casadi as cpin
 
         self._pin = pin
+        np.set_printoptions(precision=5, suppress=True, linewidth=200)
+
         self.unit_test = unit_test
 
         self.repo_path = snapshot_download("lerobot/unitree-g1-mujoco")
@@ -238,35 +249,50 @@ class G1_29_ArmIK:  # noqa: N801
         self.opti.set_value(self.param_tf_r, right_wrist)
         self.opti.set_value(self.var_q_last, self.init_data)  # for smooth
 
-        converged = True
         try:
             self.opti.solve()
+
             sol_q = self.opti.value(self.var_q)
+            self.smooth_filter.add_data(sol_q)
+            sol_q = self.smooth_filter.filtered_data
+
+            if current_lr_arm_motor_dq is not None:
+                v = current_lr_arm_motor_dq * 0.0
+            else:
+                v = (sol_q - self.init_data) * 0.0
+
+            self.init_data = sol_q
+
+            sol_tauff = self._pin.rnea(
+                self.reduced_robot.model,
+                self.reduced_robot.data,
+                sol_q,
+                v,
+                np.zeros(self.reduced_robot.model.nv),
+            )
+
+            return sol_q, sol_tauff
+
         except Exception as e:
-            converged = False
-            logger.error(f"IK convergence error: {e}")
+            logger.error(f"ERROR in convergence, plotting debug info.{e}")
+
             sol_q = self.opti.debug.value(self.var_q)
+            self.smooth_filter.add_data(sol_q)
+            sol_q = self.smooth_filter.filtered_data
 
-        self.smooth_filter.add_data(sol_q)
-        sol_q = self.smooth_filter.filtered_data
-        self.init_data = sol_q
+            if current_lr_arm_motor_dq is not None:
+                v = current_lr_arm_motor_dq * 0.0
+            else:
+                v = (sol_q - self.init_data) * 0.0
+
+            self.init_data = sol_q
 
-        if not converged:
             logger.error(
                 f"sol_q:{sol_q} \nmotorstate: \n{current_lr_arm_motor_q} \nleft_pose: \n{left_wrist} \nright_pose: \n{right_wrist}"
             )
+
             return current_lr_arm_motor_q, np.zeros(self.reduced_robot.model.nv)
 
-        sol_tauff = self._pin.rnea(
-            self.reduced_robot.model,
-            self.reduced_robot.data,
-            sol_q,
-            np.zeros(self.reduced_robot.model.nv),
-            np.zeros(self.reduced_robot.model.nv),
-        )
-
-        return sol_q, sol_tauff
-
     def solve_tau(self, current_lr_arm_motor_q=None, current_lr_arm_motor_dq=None):
         try:
             q_g1 = np.array(current_lr_arm_motor_q, dtype=float)

src/lerobot/robots/unitree_g1/run_g1_server.py

@@ -24,7 +24,6 @@ This server runs on the robot and forwards:
 Uses JSON for secure serialization instead of pickle.
 """
 
-import argparse
 import base64
 import contextlib
 import json
@@ -39,8 +38,6 @@ from unitree_sdk2py.idl.default import unitree_hg_msg_dds__LowCmd_
 from unitree_sdk2py.idl.unitree_hg.msg.dds_ import LowCmd_ as hg_LowCmd, LowState_ as hg_LowState
 from unitree_sdk2py.utils.crc import CRC
 
-from lerobot.cameras.zmq.image_server import ImageServer
-
 # DDS topic names follow Unitree SDK naming conventions
 # ruff: noqa: N816
 kTopicLowCommand_Debug = "rt/lowcmd"  # action to robot
@@ -153,32 +150,6 @@ def cmd_forward_loop(
 
 def main() -> None:
     """Main entry point for the robot server bridge."""
-    parser = argparse.ArgumentParser(description="DDS-to-ZMQ bridge server for Unitree G1")
-    parser.add_argument("--camera", action="store_true", help="Also launch camera server")
-    parser.add_argument("--camera-device", type=int, default=4, help="Camera device ID (default: 4)")
-    parser.add_argument("--camera-fps", type=int, default=30, help="Camera FPS (default: 30)")
-    parser.add_argument("--camera-width", type=int, default=640, help="Camera width (default: 640)")
-    parser.add_argument("--camera-height", type=int, default=480, help="Camera height (default: 480)")
-    parser.add_argument("--camera-port", type=int, default=5555, help="Camera ZMQ port (default: 5555)")
-    args = parser.parse_args()
-
-    # Optionally start camera server in background thread
-    camera_thread = None
-    if args.camera:
-        camera_config = {
-            "fps": args.camera_fps,
-            "cameras": {
-                "head_camera": {
-                    "device_id": args.camera_device,
-                    "shape": [args.camera_height, args.camera_width],
-                }
-            },
-        }
-        camera_server = ImageServer(camera_config, port=args.camera_port)
-        camera_thread = threading.Thread(target=camera_server.run, daemon=True)
-        camera_thread.start()
-        print(f"Camera server started on port {args.camera_port} (device {args.camera_device})")
-
     # initialize DDS
     ChannelFactoryInitialize(0)
 
@@ -235,8 +206,6 @@ def main() -> None:
         shutdown_event.set()
         ctx.term()  # terminates blocking zmq.recv() calls
         t_state.join(timeout=2.0)
-        if camera_thread is not None:
-            camera_thread.join(timeout=2.0)
 
 
 if __name__ == "__main__":

src/lerobot/robots/unitree_g1/unitree_g1.py

@@ -14,67 +14,27 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from __future__ import annotations
-
 import logging
+import struct
 import threading
 import time
 from dataclasses import dataclass, field
 from functools import cached_property
-from typing import TYPE_CHECKING, Protocol, runtime_checkable
+from typing import Any
 
 import numpy as np
 
 from lerobot.cameras.utils import make_cameras_from_configs
 from lerobot.envs.factory import make_env
 from lerobot.processor import RobotAction, RobotObservation
-from lerobot.robots.unitree_g1.g1_kinematics import G1_29_ArmIK
-from lerobot.robots.unitree_g1.g1_utils import (
-    REMOTE_AXES,
-    REMOTE_KEYS,
-    G1_29_JointArmIndex,
-    G1_29_JointIndex,
-    default_remote_input,
-    make_locomotion_controller,
-)
-from lerobot.utils.import_utils import _unitree_sdk_available
+from lerobot.robots.unitree_g1.g1_utils import G1_29_JointArmIndex, G1_29_JointIndex
+from lerobot.robots.unitree_g1.robot_kinematic_processor import G1_29_ArmIK
 
 from ..robot import Robot
 from .config_unitree_g1 import UnitreeG1Config
 
-if TYPE_CHECKING or _unitree_sdk_available:
-    from unitree_sdk2py.core.channel import (
-        ChannelFactoryInitialize as _SDKChannelFactoryInitialize,
-        ChannelPublisher as _SDKChannelPublisher,
-        ChannelSubscriber as _SDKChannelSubscriber,
-    )
-    from unitree_sdk2py.idl.default import unitree_hg_msg_dds__LowCmd_
-    from unitree_sdk2py.idl.unitree_hg.msg.dds_ import (
-        LowCmd_ as hg_LowCmd,
-        LowState_ as hg_LowState,
-    )
-    from unitree_sdk2py.utils.crc import CRC
-else:
-    _SDKChannelFactoryInitialize = None
-    _SDKChannelPublisher = None
-    _SDKChannelSubscriber = None
-    unitree_hg_msg_dds__LowCmd_ = None
-    hg_LowCmd = None
-    hg_LowState = None
-    CRC = None
-
 logger = logging.getLogger(__name__)
 
-
-@runtime_checkable
-class LocomotionController(Protocol):
-    control_dt: float
-
-    def run_step(self, action: dict, lowstate) -> dict: ...
-
-    def reset(self) -> None: ...
-
-
 # DDS topic names follow Unitree SDK naming conventions
 # ruff: noqa: N816
 kTopicLowCommand_Debug = "rt/lowcmd"
@@ -103,7 +63,7 @@ class IMUState:
 class G1_29_LowState:  # noqa: N801
     motor_state: list[MotorState] = field(default_factory=lambda: [MotorState() for _ in G1_29_JointIndex])
     imu_state: IMUState = field(default_factory=IMUState)
-    wireless_remote: bytes | None = None  # Raw wireless remote data
+    wireless_remote: Any = None  # Raw wireless remote data
     mode_machine: int = 0  # Robot mode
 
 
@@ -111,6 +71,25 @@ class UnitreeG1(Robot):
     config_class = UnitreeG1Config
     name = "unitree_g1"
 
+    # unitree remote controller
+    class RemoteController:
+        def __init__(self):
+            self.lx = 0
+            self.ly = 0
+            self.rx = 0
+            self.ry = 0
+            self.button = [0] * 16
+
+        def set(self, data):
+            # wireless_remote
+            keys = struct.unpack("H", data[2:4])[0]
+            for i in range(16):
+                self.button[i] = (keys & (1 << i)) >> i
+            self.lx = struct.unpack("f", data[4:8])[0]
+            self.rx = struct.unpack("f", data[8:12])[0]
+            self.ry = struct.unpack("f", data[12:16])[0]
+            self.ly = struct.unpack("f", data[20:24])[0]
+
     def __init__(self, config: UnitreeG1Config):
         super().__init__(config)
 
@@ -124,9 +103,11 @@ class UnitreeG1(Robot):
 
         # Import channel classes based on mode
         if config.is_simulation:
-            self._ChannelFactoryInitialize = _SDKChannelFactoryInitialize
-            self._ChannelPublisher = _SDKChannelPublisher
-            self._ChannelSubscriber = _SDKChannelSubscriber
+            from unitree_sdk2py.core.channel import (
+                ChannelFactoryInitialize,
+                ChannelPublisher,
+                ChannelSubscriber,
+            )
         else:
             from lerobot.robots.unitree_g1.unitree_sdk2_socket import (
                 ChannelFactoryInitialize,
@@ -134,30 +115,22 @@ class UnitreeG1(Robot):
                 ChannelSubscriber,
             )
 
-            self._ChannelFactoryInitialize = ChannelFactoryInitialize
-            self._ChannelPublisher = ChannelPublisher
-            self._ChannelSubscriber = ChannelSubscriber
+        # Store for use in connect()
+        self._ChannelFactoryInitialize = ChannelFactoryInitialize
+        self._ChannelPublisher = ChannelPublisher
+        self._ChannelSubscriber = ChannelSubscriber
 
         # Initialize state variables
         self.sim_env = None
         self._env_wrapper = None
         self._lowstate = None
-        self._lowstate_lock = threading.Lock()
         self._shutdown_event = threading.Event()
         self.subscribe_thread = None
+        self.remote_controller = self.RemoteController()
 
-        self.arm_ik = G1_29_ArmIK() if config.gravity_compensation else None
+        self.arm_ik = G1_29_ArmIK()
 
-        # Lower-body controller loaded dynamically
-        self.controller: LocomotionController | None = make_locomotion_controller(config.controller)
-
-        # Controller thread state
-        self._controller_thread = None
-        self._controller_action_lock = threading.Lock()
-        self.controller_input = default_remote_input()
-        self.controller_output = {}
-
-    def _subscribe_lowstate(self):  # polls robot state @ 250Hz
+    def _subscribe_motor_state(self):  # polls robot state @ 250Hz
         while not self._shutdown_event.is_set():
             start_time = time.time()
 
@@ -170,11 +143,11 @@ class UnitreeG1(Robot):
                 lowstate = G1_29_LowState()
 
                 # Capture motor states using jointindex
-                for joint in G1_29_JointIndex:
-                    lowstate.motor_state[joint].q = msg.motor_state[joint].q
-                    lowstate.motor_state[joint].dq = msg.motor_state[joint].dq
-                    lowstate.motor_state[joint].tau_est = msg.motor_state[joint].tau_est
-                    lowstate.motor_state[joint].temperature = msg.motor_state[joint].temperature
+                for id in G1_29_JointIndex:
+                    lowstate.motor_state[id].q = msg.motor_state[id].q
+                    lowstate.motor_state[id].dq = msg.motor_state[id].dq
+                    lowstate.motor_state[id].tau_est = msg.motor_state[id].tau_est
+                    lowstate.motor_state[id].temperature = msg.motor_state[id].temperature
 
                 # Capture IMU state
                 lowstate.imu_state.quaternion = list(msg.imu_state.quaternion)
@@ -189,106 +162,31 @@ class UnitreeG1(Robot):
                 # Capture mode_machine
                 lowstate.mode_machine = msg.mode_machine
 
-                with self._lowstate_lock:
-                    self._lowstate = lowstate
+                self._lowstate = lowstate
 
             current_time = time.time()
             all_t_elapsed = current_time - start_time
             sleep_time = max(0, (self.control_dt - all_t_elapsed))  # maintain constant control dt
             time.sleep(sleep_time)
 
-    def publish_lowcmd(
-        self,
-        action: RobotAction,
-        kp: np.ndarray | list[float] | None = None,
-        kd: np.ndarray | list[float] | None = None,
-        tau: np.ndarray | list[float] | None = None,
-    ) -> None:  # writes robot command whenever requested
-        for motor in G1_29_JointIndex:
-            key = f"{motor.name}.q"
-            if key in action:
-                self.msg.motor_cmd[motor.value].q = action[key]
-                self.msg.motor_cmd[motor.value].qd = 0
-                self.msg.motor_cmd[motor.value].kp = (
-                    kp[motor.value] if kp is not None else self.kp[motor.value]
-                )
-                self.msg.motor_cmd[motor.value].kd = (
-                    kd[motor.value] if kd is not None else self.kd[motor.value]
-                )
-                self.msg.motor_cmd[motor.value].tau = tau[motor.value] if tau is not None else 0.0
-
-        self.msg.crc = self.crc.Crc(self.msg)
-        self.lowcmd_publisher.Write(self.msg)
-
-    @property
-    def _cameras_ft(self) -> dict[str, tuple]:
-        return {
-            cam: (self.config.cameras[cam].height, self.config.cameras[cam].width, 3) for cam in self.cameras
-        }
-
-    @cached_property
-    def observation_features(self) -> dict[str, type | tuple]:
-        return {**self._motors_ft, **self._cameras_ft}
-
     @cached_property
     def action_features(self) -> dict[str, type]:
-        if self.controller is None:
-            return {f"{G1_29_JointIndex(motor).name}.q": float for motor in G1_29_JointIndex}
+        return {f"{G1_29_JointIndex(motor).name}.q": float for motor in G1_29_JointIndex}
 
-        arm_features = {f"{G1_29_JointArmIndex(motor).name}.q": float for motor in G1_29_JointArmIndex}
-        remote_features = dict.fromkeys(REMOTE_AXES, float)
-        return {**arm_features, **remote_features}
-
-    def _controller_loop(self):
-        """Background thread that runs controller at policy's control_dt."""
-        control_dt = self.controller.control_dt
-        logger.info(f"Controller loop starting with control_dt={control_dt} ({1.0 / control_dt:.1f}Hz)")
-
-        loop_count = 0
-        last_log_time = time.time()
-
-        while not self._shutdown_event.is_set():
-            start_time = time.time()
-
-            with self._lowstate_lock:
-                lowstate = self._lowstate
-
-            if lowstate is not None and self.controller is not None:
-                loop_count += 1
-                if time.time() - last_log_time >= 5.0:  # Log every 5 seconds
-                    actual_hz = loop_count / (time.time() - last_log_time)
-                    logger.info(
-                        f"Controller actual rate: {actual_hz:.1f}Hz (target: {1.0 / control_dt:.1f}Hz)"
-                    )
-                    loop_count = 0
-                    last_log_time = time.time()
-                # Read controller input snapshot
-                with self._controller_action_lock:
-                    controller_input = dict(self.controller_input)
-
-                # Run controller step
-                controller_action = self.controller.run_step(controller_input, lowstate)
-
-                # Write controller output snapshot
-                with self._controller_action_lock:
-                    self.controller_output = dict(controller_action)
-
-                ctrl_kp = self.controller.kp if hasattr(self.controller, "kp") else None
-                ctrl_kd = self.controller.kd if hasattr(self.controller, "kd") else None
-                self.publish_lowcmd(controller_action, kp=ctrl_kp, kd=ctrl_kd)
-
-            elapsed = time.time() - start_time
-            sleep_time = max(0, control_dt - elapsed)
-            time.sleep(sleep_time)
-
-    def calibrate(self) -> None:
-        # TODO: implement g1_29 calibration
+    def calibrate(self) -> None:  # robot is already calibrated
         pass
 
     def configure(self) -> None:
         pass
 
     def connect(self, calibrate: bool = True) -> None:  # connect to DDS
+        from unitree_sdk2py.idl.default import unitree_hg_msg_dds__LowCmd_
+        from unitree_sdk2py.idl.unitree_hg.msg.dds_ import (
+            LowCmd_ as hg_LowCmd,
+            LowState_ as hg_LowState,
+        )
+        from unitree_sdk2py.utils.crc import CRC
+
         # Initialize DDS channel and simulation environment
         if self.config.is_simulation:
             self._ChannelFactoryInitialize(0, "lo")
@@ -296,7 +194,7 @@ class UnitreeG1(Robot):
             # Extract the actual gym env from the dict structure
             self.sim_env = self._env_wrapper["hub_env"][0].envs[0]
         else:
-            self._ChannelFactoryInitialize(0, config=self.config)
+            self._ChannelFactoryInitialize(0)
 
         # Initialize direct motor control interface
         self.lowcmd_publisher = self._ChannelPublisher(kTopicLowCommand_Debug, hg_LowCmd)
@@ -305,7 +203,7 @@ class UnitreeG1(Robot):
         self.lowstate_subscriber.Init()
 
         # Start subscribe thread to read robot state
-        self.subscribe_thread = threading.Thread(target=self._subscribe_lowstate)
+        self.subscribe_thread = threading.Thread(target=self._subscribe_motor_state)
         self.subscribe_thread.start()
 
         # Connect cameras
@@ -322,53 +220,25 @@ class UnitreeG1(Robot):
 
         # Wait for first state message to arrive
         lowstate = None
-        deadline = time.time() + 10.0
         while lowstate is None:
-            with self._lowstate_lock:
-                lowstate = self._lowstate
+            lowstate = self._lowstate
             if lowstate is None:
-                if time.time() > deadline:
-                    raise TimeoutError("Timed out waiting for robot state (10s)")
-                logger.warning("[UnitreeG1] Waiting for robot state...")
                 time.sleep(0.01)
-        logger.info("[UnitreeG1] Connected to robot.")
+            logger.warning("[UnitreeG1] Waiting for robot state...")
+        logger.warning("[UnitreeG1] Connected to robot.")
         self.msg.mode_machine = lowstate.mode_machine
 
+        # Initialize all motors with unified kp/kd from config
         self.kp = np.array(self.config.kp, dtype=np.float32)
         self.kd = np.array(self.config.kd, dtype=np.float32)
 
-        for joint in G1_29_JointIndex:
-            self.msg.motor_cmd[joint].mode = 1
-            self.msg.motor_cmd[joint].kp = self.kp[joint.value]
-            self.msg.motor_cmd[joint].kd = self.kd[joint.value]
-            self.msg.motor_cmd[joint].q = lowstate.motor_state[joint.value].q
-
-        # Start controller thread if enabled
-        if self.controller is not None:
-            self._controller_thread = threading.Thread(target=self._controller_loop, daemon=True)
-            self._controller_thread.start()
-            fps = int(1.0 / self.controller.control_dt)
-            logger.info(f"Controller thread started ({fps}Hz)")
-
-    def _send_zero_torque(self) -> None:
-        """Send a zero-gain command to make joints passive before shutting down."""
-        try:
-            with self._lowstate_lock:
-                lowstate = self._lowstate
-            if lowstate is None:
-                return
-            action = {f"{motor.name}.q": lowstate.motor_state[motor.value].q for motor in G1_29_JointIndex}
-            zero_gains = np.zeros(29, dtype=np.float32)
-            self.publish_lowcmd(action, kp=zero_gains, kd=zero_gains, tau=zero_gains)
-            logger.info("Sent zero-torque command for safe shutdown")
-        except Exception as e:
-            logger.warning(f"Failed to send zero-torque on disconnect: {e}")
+        for id in G1_29_JointIndex:
+            self.msg.motor_cmd[id].mode = 1
+            self.msg.motor_cmd[id].kp = self.kp[id.value]
+            self.msg.motor_cmd[id].kd = self.kd[id.value]
+            self.msg.motor_cmd[id].q = lowstate.motor_state[id.value].q
 
     def disconnect(self):
-        # Put robot in passive mode before stopping threads
-        if not self.config.is_simulation:
-            self._send_zero_torque()
-
         # Signal thread to stop and unblock any waits
         self._shutdown_event.set()
 
@@ -378,12 +248,6 @@ class UnitreeG1(Robot):
             if self.subscribe_thread.is_alive():
                 logger.warning("Subscribe thread did not stop cleanly")
 
-        # Wait for controller thread to finish
-        if self._controller_thread is not None:
-            self._controller_thread.join(timeout=2.0)
-            if self._controller_thread.is_alive():
-                logger.warning("Controller thread did not stop cleanly")
-
         # Close simulation environment
         if self.config.is_simulation and self.sim_env is not None:
             try:
@@ -410,8 +274,7 @@ class UnitreeG1(Robot):
             cam.disconnect()
 
     def get_observation(self) -> RobotObservation:
-        with self._lowstate_lock:
-            lowstate = self._lowstate
+        lowstate = self._lowstate
         if lowstate is None:
             return {}
 
@@ -450,9 +313,14 @@ class UnitreeG1(Robot):
             obs["imu.rpy.pitch"] = lowstate.imu_state.rpy[1]
             obs["imu.rpy.yaw"] = lowstate.imu_state.rpy[2]
 
-        # Wireless remote (raw bytes for teleoperator)
-        if lowstate.wireless_remote:
-            obs["wireless_remote"] = lowstate.wireless_remote
+        # Controller - parse wireless_remote and add to obs
+        if lowstate.wireless_remote and len(lowstate.wireless_remote) >= 24:
+            self.remote_controller.set(lowstate.wireless_remote)
+        obs["remote.buttons"] = self.remote_controller.button.copy()
+        obs["remote.lx"] = self.remote_controller.lx
+        obs["remote.ly"] = self.remote_controller.ly
+        obs["remote.rx"] = self.remote_controller.rx
+        obs["remote.ry"] = self.remote_controller.ry
 
         # Cameras - read images from ZMQ cameras
         for cam_name, cam in self._cameras.items():
@@ -460,63 +328,73 @@ class UnitreeG1(Robot):
 
         return obs
 
-    def send_action(self, action: RobotAction) -> RobotAction:
-        action_to_publish = action
-        if self.controller is not None:
-            # Controller thread owns legs/waist. Here we only update joystick inputs
-            # and publish arm targets from the teleoperator.
-            self._update_controller_action(action)
-            arm_prefixes = tuple(j.name for j in G1_29_JointArmIndex)
-            action_to_publish = {
-                key: value
-                for key, value in action.items()
-                if key.endswith(".q") and key.startswith(arm_prefixes)
-            }
-
-        tau = None
-        if self.config.gravity_compensation and self.arm_ik is not None:
-            tau = np.zeros(29, dtype=np.float32)
-            action_np = np.array(
-                [
-                    action_to_publish.get(f"{joint.name}.q", self.msg.motor_cmd[joint.value].q)
-                    for joint in G1_29_JointArmIndex
-                ],
-                dtype=np.float32,
-            )
-            arm_tau = self.arm_ik.solve_tau(action_np)
-            arm_start_idx = G1_29_JointArmIndex.kLeftShoulderPitch.value
-            for joint in G1_29_JointArmIndex:
-                local_idx = joint.value - arm_start_idx
-                tau[joint.value] = arm_tau[local_idx]
-
-        self.publish_lowcmd(action_to_publish, tau=tau)
-        return action
-
-    def _update_controller_action(self, action: RobotAction) -> None:
-        """Update controller input state from incoming teleop action."""
-        with self._controller_action_lock:
-            for key in REMOTE_KEYS:
-                if key in action:
-                    self.controller_input[key] = action[key]
-
     @property
     def is_calibrated(self) -> bool:
         return True
 
     @property
     def is_connected(self) -> bool:
-        with self._lowstate_lock:
-            return self._lowstate is not None
+        return self._lowstate is not None
 
     @property
     def _motors_ft(self) -> dict[str, type]:
-        """Joint positions for all 29 joints."""
         return {f"{G1_29_JointIndex(motor).name}.q": float for motor in G1_29_JointIndex}
 
     @property
     def cameras(self) -> dict:
         return self._cameras
 
+    @property
+    def _cameras_ft(self) -> dict[str, tuple]:
+        return {
+            cam: (self.config.cameras[cam].height, self.config.cameras[cam].width, 3) for cam in self.cameras
+        }
+
+    @cached_property
+    def observation_features(self) -> dict[str, type | tuple]:
+        return {**self._motors_ft, **self._cameras_ft}
+
+    def send_action(self, action: RobotAction) -> RobotAction:
+        for motor in G1_29_JointIndex:
+            key = f"{motor.name}.q"
+            if key in action:
+                self.msg.motor_cmd[motor.value].q = action[key]
+                self.msg.motor_cmd[motor.value].qd = 0
+                self.msg.motor_cmd[motor.value].kp = self.kp[motor.value]
+                self.msg.motor_cmd[motor.value].kd = self.kd[motor.value]
+                self.msg.motor_cmd[motor.value].tau = 0
+
+        if self.config.gravity_compensation:
+            # Build action_np from motor commands (arm joints are indices 15-28, local indices 0-13)
+            action_np = np.zeros(14)
+            arm_start_idx = G1_29_JointArmIndex.kLeftShoulderPitch.value  # 15
+            for joint in G1_29_JointArmIndex:
+                local_idx = joint.value - arm_start_idx
+                action_np[local_idx] = self.msg.motor_cmd[joint.value].q
+            tau = self.arm_ik.solve_tau(action_np)
+
+            # Apply tau back to motor commands
+            for joint in G1_29_JointArmIndex:
+                local_idx = joint.value - arm_start_idx
+                self.msg.motor_cmd[joint.value].tau = tau[local_idx]
+
+        self.msg.crc = self.crc.Crc(self.msg)
+        self.lowcmd_publisher.Write(self.msg)
+        return action
+
+    def get_gravity_orientation(self, quaternion):  # get gravity orientation from quaternion
+        """Get gravity orientation from quaternion."""
+        qw = quaternion[0]
+        qx = quaternion[1]
+        qy = quaternion[2]
+        qz = quaternion[3]
+
+        gravity_orientation = np.zeros(3)
+        gravity_orientation[0] = 2 * (-qz * qx + qw * qy)
+        gravity_orientation[1] = -2 * (qz * qy + qw * qx)
+        gravity_orientation[2] = 1 - 2 * (qw * qw + qz * qz)
+        return gravity_orientation
+
     def reset(
         self,
         control_dt: float | None = None,
@@ -529,9 +407,15 @@ class UnitreeG1(Robot):
 
         if self.config.is_simulation and self.sim_env is not None:
             self.sim_env.reset()
-            self.publish_lowcmd(
-                {f"{motor.name}.q": float(default_positions[motor.value]) for motor in G1_29_JointIndex}
-            )
+
+            for motor in G1_29_JointIndex:
+                self.msg.motor_cmd[motor.value].q = default_positions[motor.value]
+                self.msg.motor_cmd[motor.value].qd = 0
+                self.msg.motor_cmd[motor.value].kp = self.kp[motor.value]
+                self.msg.motor_cmd[motor.value].kd = self.kd[motor.value]
+                self.msg.motor_cmd[motor.value].tau = 0
+            self.msg.crc = self.crc.Crc(self.msg)
+            self.lowcmd_publisher.Write(self.msg)
         else:
             total_time = 3.0
             num_steps = int(total_time / control_dt)
@@ -562,8 +446,4 @@ class UnitreeG1(Robot):
                 sleep_time = max(0, control_dt - elapsed)
                 time.sleep(sleep_time)
 
-        # Reset controller internal state (gait phase, obs history, etc.)
-        if self.controller is not None and hasattr(self.controller, "reset"):
-            self.controller.reset()
-
         logger.info("Reached default position")

src/lerobot/robots/unitree_g1/unitree_sdk2_socket.py

@@ -22,8 +22,6 @@ import zmq
 
 from lerobot.robots.unitree_g1.config_unitree_g1 import UnitreeG1Config
 
-# Module-level ZMQ state mirrors the Unitree SDK's global ChannelFactory Singleton.
-# Only one robot connection per process is supported.
 _ctx: zmq.Context | None = None
 _lowcmd_sock: zmq.Socket | None = None
 _lowstate_sock: zmq.Socket | None = None
@@ -99,22 +97,17 @@ def lowcmd_to_dict(topic: str, msg: Any) -> dict[str, Any]:
     }
 
 
-def ChannelFactoryInitialize(domain_id: int = 0, config: Any = None) -> None:  # noqa: N802
+def ChannelFactoryInitialize(*args: Any, **kwargs: Any) -> None:  # noqa: N802
     """
     Initialize ZMQ sockets for robot communication.
 
     This function mimics the Unitree SDK's ChannelFactoryInitialize but uses
     ZMQ sockets to connect to the robot server bridge instead of DDS.
-
-    Args:
-        domain_id: Ignored (for API compatibility with Unitree SDK)
-        config: UnitreeG1Config instance with robot_ip
     """
     global _ctx, _lowcmd_sock, _lowstate_sock
 
     # read socket config
-    if config is None:
-        config = UnitreeG1Config()
+    config = UnitreeG1Config()
     robot_ip = config.robot_ip
 
     ctx = zmq.Context.instance()

src/lerobot/scripts/lerobot_calibrate.py

@@ -56,7 +56,6 @@ from lerobot.teleoperators import (  # noqa: F401
     make_teleoperator_from_config,
     omx_leader,
     openarm_leader,
-    openarm_mini,
     so_leader,
     unitree_g1,
 )

src/lerobot/scripts/lerobot_dataset_viz.py

@@ -132,13 +132,10 @@ def visualize_dataset(
 
     logging.info("Logging to Rerun")
 
-    first_index = None
     for batch in tqdm.tqdm(dataloader, total=len(dataloader)):
-        if first_index is None:
-            first_index = batch["index"][0].item()
         # iterate over the batch
         for i in range(len(batch["index"])):
-            rr.set_time("frame_index", sequence=batch["index"][i].item() - first_index)
+            rr.set_time("frame_index", sequence=batch["frame_index"][i].item())
             rr.set_time("timestamp", timestamp=batch["timestamp"][i].item())
 
             # display each camera image

src/lerobot/scripts/lerobot_edit_dataset.py

@@ -21,9 +21,6 @@ This script allows you to delete episodes, split datasets, merge datasets,
 remove features, modify tasks, and convert image datasets to video format.
 When new_repo_id is specified, creates a new dataset.
 
-Path semantics (v2): --root and --new_root are exact dataset folders containing
-meta/, data/, videos/. When omitted, defaults to $HF_LEROBOT_HOME/{repo_id}.
-
 Usage Examples:
 
 Delete episodes 0, 2, and 5 from a dataset:
@@ -32,34 +29,19 @@ Delete episodes 0, 2, and 5 from a dataset:
         --operation.type delete_episodes \
         --operation.episode_indices "[0, 2, 5]"
 
-Delete episodes from a local dataset at a specific path:
-    lerobot-edit-dataset \
-        --repo_id lerobot/pusht \
-        --root /path/to/pusht \
-        --operation.type delete_episodes \
-        --operation.episode_indices "[0, 2, 5]"
-
-Delete episodes and save to a new dataset at a specific path and with a new repo_id:
+Delete episodes and save to a new dataset:
     lerobot-edit-dataset \
         --repo_id lerobot/pusht \
         --new_repo_id lerobot/pusht_filtered \
-        --new_root /path/to/pusht_filtered \
         --operation.type delete_episodes \
         --operation.episode_indices "[0, 2, 5]"
 
-Split dataset by fractions (pusht_train, pusht_val):
+Split dataset by fractions:
     lerobot-edit-dataset \
         --repo_id lerobot/pusht \
         --operation.type split \
         --operation.splits '{"train": 0.8, "val": 0.2}'
 
-Split dataset by fractions and save split datasets to a specific folder (base_folder/train, base_folder/val):
-    lerobot-edit-dataset \
-        --repo_id lerobot/pusht \
-        --new_root /path/to/base_folder \
-        --operation.type split \
-        --operation.splits '{"train": 0.8, "val": 0.2}'
-
 Split dataset by episode indices:
     lerobot-edit-dataset \
         --repo_id lerobot/pusht \
@@ -74,29 +56,15 @@ Split into more than two splits:
 
 Merge multiple datasets:
     lerobot-edit-dataset \
-        --new_repo_id lerobot/pusht_merged \
+        --repo_id lerobot/pusht_merged \
         --operation.type merge \
         --operation.repo_ids "['lerobot/pusht_train', 'lerobot/pusht_val']"
 
-Merge multiple datasets to a specific output path:
-    lerobot-edit-dataset \
-        --new_repo_id lerobot/pusht_merged \
-        --new_root /path/to/pusht_merged \
-        --operation.type merge \
-        --operation.repo_ids "['lerobot/pusht_train', 'lerobot/pusht_val']"
-
-Merge multiple datasets from a list of local dataset paths:
-    lerobot-edit-dataset \
-        --new_repo_id lerobot/pusht_merged \
-        --operation.type merge \
-        --operation.repo_ids "['pusht_train', 'pusht_val']" \
-        --operation.roots "['/path/to/pusht_train', '/path/to/pusht_val']"
-
 Remove camera feature:
     lerobot-edit-dataset \
         --repo_id lerobot/pusht \
         --operation.type remove_feature \
-        --operation.feature_names "['observation.image']"
+        --operation.feature_names "['observation.images.top']"
 
 Modify tasks - set a single task for all episodes (WARNING: modifies in-place):
     lerobot-edit-dataset \
@@ -120,8 +88,8 @@ Modify tasks - set default task with overrides for specific episodes (WARNING: m
 Convert image dataset to video format and save locally:
     lerobot-edit-dataset \
         --repo_id lerobot/pusht_image \
-        --new_root /path/to/output/pusht_video \
-        --operation.type convert_image_to_video
+        --operation.type convert_image_to_video \
+        --operation.output_dir /path/to/output/pusht_video
 
 Convert image dataset to video format and save with new repo_id:
     lerobot-edit-dataset \
@@ -136,6 +104,28 @@ Convert image dataset to video format and push to hub:
         --operation.type convert_image_to_video \
         --push_to_hub true
 
+Trim single episode to keep only frames within timestamp range:
+    python -m lerobot.scripts.lerobot_edit_dataset \
+        --repo_id lerobot/pusht \
+        --new_repo_id lerobot/pusht_trimmed \
+        --operation.type trim_episode \
+        --operation.episode_index 0 \
+        --operation.start_timestamp 10.0 \
+        --operation.end_timestamp 30.0
+
+Trim multiple episodes at once (use null for no limit):
+    python -m lerobot.scripts.lerobot_edit_dataset \
+        --repo_id lerobot/pusht \
+        --operation.type trim_episode \
+        --operation.episode_trims '{"0": [10.0, 30.0], "2": [5.0, null], "3": [null, 20.0]}'
+
+Trim and re-upload to same repo (overwrites original):
+    python -m lerobot.scripts.lerobot_edit_dataset \
+        --repo_id lerobot/pusht \
+        --operation.type trim_episode \
+        --operation.episode_index 0 \
+        --operation.start_timestamp 10.0 \
+        --push_to_hub true
 Show dataset information:
     lerobot-edit-dataset \
         --repo_id lerobot/pusht_image \
@@ -199,7 +189,6 @@ class SplitConfig(OperationConfig):
 @dataclass
 class MergeConfig(OperationConfig):
     repo_ids: list[str] | None = None
-    roots: list[str] | None = None
 
 
 @OperationConfig.register_subclass("remove_feature")
@@ -233,46 +222,59 @@ class ConvertImageToVideoConfig(OperationConfig):
 @OperationConfig.register_subclass("info")
 @dataclass
 class InfoConfig(OperationConfig):
+    type: str = "info"
     show_features: bool = False
 
 
+@dataclass
+class TrimEpisodeConfig:
+    """Trim episodes to keep only frames within timestamp ranges.
+    
+    Supports multiple episodes via episode_trims dict:
+        --operation.episode_trims '{"0": [10.0, 30.0], "2": [5.0, 20.0]}'
+    
+    Or single episode via legacy parameters:
+        --operation.episode_index 0 --operation.start_timestamp 10.0 --operation.end_timestamp 30.0
+    """
+    type: str = "trim_episode"
+    # Multi-episode support: dict mapping episode_index -> [start_timestamp, end_timestamp]
+    # Use null for no limit, e.g. {"0": [10.0, null], "2": [null, 30.0]}
+    episode_trims: dict[str, list[float | None]] | None = None
+    # Legacy single-episode parameters (used if episode_trims is None)
+    episode_index: int | None = None
+    start_timestamp: float | None = None  # Keep frames from this timestamp (inclusive)
+    end_timestamp: float | None = None    # Keep frames until this timestamp (inclusive)
+
+
 @dataclass
 class EditDatasetConfig:
-    # Operation configuration.
+    repo_id: str
+    operation: (
+        DeleteEpisodesConfig | SplitConfig | MergeConfig | RemoveFeatureConfig | ConvertImageToVideoConfig | TrimEpisodeConfig
+    )
     operation: OperationConfig
-    # Input dataset identifier. Always required unless for Merge operation.
-    repo_id: str | None = None
-    # Root directory where the input dataset is stored. If not specified, defaults to $HF_LEROBOT_HOME/repo_id.
     root: str | None = None
-    # Edited dataset identifier. When both new_repo_id (resp. new_root) and repo_id (resp. root) are identical, modifications are applied in-place and a backup of the original dataset is created. Required for Merge operation.
     new_repo_id: str | None = None
-    # Root directory where the edited dataset will be stored. If not specified, defaults to $HF_LEROBOT_HOME/new_repo_id. For Split operation, this is the base directory for the split datasets.
-    new_root: str | None = None
-    # Upload dataset to Hugging Face hub.
     push_to_hub: bool = False
 
 
-def get_output_path(
-    repo_id: str,
-    new_repo_id: str | None,
-    root: Path | str | None,
-    new_root: Path | str | None,
-) -> tuple[str, Path]:
-    input_path = Path(root) if root else HF_LEROBOT_HOME / repo_id
+def get_output_path(repo_id: str, new_repo_id: str | None, root: Path | None) -> tuple[str, Path]:
+    if new_repo_id:
+        output_repo_id = new_repo_id
+        output_dir = root / new_repo_id if root else HF_LEROBOT_HOME / new_repo_id
+    else:
+        output_repo_id = repo_id
+        dataset_path = root / repo_id if root else HF_LEROBOT_HOME / repo_id
+        old_path = Path(str(dataset_path) + "_old")
 
-    output_repo_id = new_repo_id if new_repo_id else repo_id
-    output_path = Path(new_root) if new_root else HF_LEROBOT_HOME / output_repo_id
+        if dataset_path.exists():
+            if old_path.exists():
+                shutil.rmtree(old_path)
+            shutil.move(str(dataset_path), str(old_path))
 
-    # In case of in-place modification, create a backup of the original dataset (if it exists)
-    if output_path == input_path:
-        backup_path = input_path.with_name(input_path.name + "_old")
+        output_dir = dataset_path
 
-        if input_path.exists():
-            if backup_path.exists():
-                shutil.rmtree(backup_path)
-            shutil.move(input_path, backup_path)
-
-    return output_repo_id, output_path
+    return output_repo_id, output_dir
 
 
 def handle_delete_episodes(cfg: EditDatasetConfig) -> None:
@@ -284,15 +286,11 @@ def handle_delete_episodes(cfg: EditDatasetConfig) -> None:
 
     dataset = LeRobotDataset(cfg.repo_id, root=cfg.root)
     output_repo_id, output_dir = get_output_path(
-        cfg.repo_id,
-        new_repo_id=cfg.new_repo_id,
-        root=cfg.root,
-        new_root=cfg.new_root,
+        cfg.repo_id, cfg.new_repo_id, Path(cfg.root) if cfg.root else None
     )
 
-    # In case of in-place modification, make the dataset point to the backup directory
-    if output_dir == dataset.root:
-        dataset.root = dataset.root.with_name(dataset.root.name + "_old")
+    if cfg.new_repo_id is None:
+        dataset.root = Path(str(dataset.root) + "_old")
 
     logging.info(f"Deleting episodes {cfg.operation.episode_indices} from {cfg.repo_id}")
     new_dataset = delete_episodes(
@@ -319,27 +317,19 @@ def handle_split(cfg: EditDatasetConfig) -> None:
             "splits dict must be specified with split names as keys and fractions/episode lists as values"
         )
 
-    if cfg.new_repo_id is not None:
-        logging.warning(
-            "split uses the original dataset identifier --repo_id to generate split names. The --new_repo_id parameter is ignored."
-        )
-
     dataset = LeRobotDataset(cfg.repo_id, root=cfg.root)
 
     logging.info(f"Splitting dataset {cfg.repo_id} with splits: {cfg.operation.splits}")
-    split_datasets = split_dataset(
-        dataset,
-        splits=cfg.operation.splits,
-        output_dir=cfg.new_root,
-    )
+    split_datasets = split_dataset(dataset, splits=cfg.operation.splits)
 
     for split_name, split_ds in split_datasets.items():
+        split_repo_id = f"{cfg.repo_id}_{split_name}"
         logging.info(
             f"{split_name}: {split_ds.meta.total_episodes} episodes, {split_ds.meta.total_frames} frames"
         )
 
         if cfg.push_to_hub:
-            logging.info(f"Pushing {split_name} split to hub as {split_ds.repo_id}")
+            logging.info(f"Pushing {split_name} split to hub as {split_repo_id}")
             LeRobotDataset(split_ds.repo_id, root=split_ds.root).push_to_hub()
 
 
@@ -350,29 +340,18 @@ def handle_merge(cfg: EditDatasetConfig) -> None:
     if not cfg.operation.repo_ids:
         raise ValueError("repo_ids must be specified for merge operation")
 
-    if cfg.repo_id is not None or cfg.root is not None:
-        logging.warning(
-            "merge uses --new_repo_id and --new_root for the merged dataset. The --repo_id and --root parameters are ignored."
-        )
+    if not cfg.repo_id:
+        raise ValueError("repo_id must be specified as the output repository for merged dataset")
 
-    if cfg.operation.roots:
-        if len(cfg.operation.roots) != len(cfg.operation.repo_ids):
-            raise ValueError("repo_ids and roots must have the same length for merge operation")
-        logging.info(f"Loading {len(cfg.operation.roots)} datasets to merge")
-        datasets = [
-            LeRobotDataset(repo_id=repo_id, root=root)
-            for repo_id, root in zip(cfg.operation.repo_ids, cfg.operation.roots, strict=True)
-        ]
-    else:
-        logging.info(f"Loading {len(cfg.operation.repo_ids)} datasets to merge")
-        datasets = [LeRobotDataset(repo_id) for repo_id in cfg.operation.repo_ids]
+    logging.info(f"Loading {len(cfg.operation.repo_ids)} datasets to merge")
+    datasets = [LeRobotDataset(repo_id, root=cfg.root) for repo_id in cfg.operation.repo_ids]
 
-    output_dir = Path(cfg.new_root) if cfg.new_root else HF_LEROBOT_HOME / cfg.new_repo_id
+    output_dir = Path(cfg.root) / cfg.repo_id if cfg.root else HF_LEROBOT_HOME / cfg.repo_id
 
-    logging.info(f"Merging datasets into {cfg.new_repo_id}")
+    logging.info(f"Merging datasets into {cfg.repo_id}")
     merged_dataset = merge_datasets(
         datasets,
-        output_repo_id=cfg.new_repo_id,
+        output_repo_id=cfg.repo_id,
         output_dir=output_dir,
     )
 
@@ -382,7 +361,7 @@ def handle_merge(cfg: EditDatasetConfig) -> None:
     )
 
     if cfg.push_to_hub:
-        logging.info(f"Pushing to hub as {cfg.new_repo_id}")
+        logging.info(f"Pushing to hub as {cfg.repo_id}")
         LeRobotDataset(merged_dataset.repo_id, root=output_dir).push_to_hub()
 
 
@@ -395,15 +374,11 @@ def handle_remove_feature(cfg: EditDatasetConfig) -> None:
 
     dataset = LeRobotDataset(cfg.repo_id, root=cfg.root)
     output_repo_id, output_dir = get_output_path(
-        cfg.repo_id,
-        new_repo_id=cfg.new_repo_id,
-        root=cfg.root,
-        new_root=cfg.new_root,
+        cfg.repo_id, cfg.new_repo_id, Path(cfg.root) if cfg.root else None
     )
 
-    # In case of in-place modification, make the dataset point to the backup directory
-    if output_dir == dataset.root:
-        dataset.root = dataset.root.with_name(dataset.root.name + "_old")
+    if cfg.new_repo_id is None:
+        dataset.root = Path(str(dataset.root) + "_old")
 
     logging.info(f"Removing features {cfg.operation.feature_names} from {cfg.repo_id}")
     new_dataset = remove_feature(
@@ -421,6 +396,92 @@ def handle_remove_feature(cfg: EditDatasetConfig) -> None:
         LeRobotDataset(output_repo_id, root=output_dir).push_to_hub()
 
 
+def handle_trim_episode(cfg: EditDatasetConfig) -> None:
+    """Trim episodes to keep only frames within timestamp ranges."""
+    if not isinstance(cfg.operation, TrimEpisodeConfig):
+        raise ValueError("Operation config must be TrimEpisodeConfig")
+
+    # Parse episode trims - support both multi-episode dict and legacy single episode
+    episode_trims: dict[int, tuple[float | None, float | None]] = {}
+    
+    if cfg.operation.episode_trims is not None:
+        # Multi-episode mode
+        for ep_str, ts_range in cfg.operation.episode_trims.items():
+            ep_idx = int(ep_str)
+            start_ts = ts_range[0] if len(ts_range) > 0 else None
+            end_ts = ts_range[1] if len(ts_range) > 1 else None
+            episode_trims[ep_idx] = (start_ts, end_ts)
+    elif cfg.operation.episode_index is not None:
+        # Legacy single-episode mode
+        if cfg.operation.start_timestamp is None and cfg.operation.end_timestamp is None:
+            raise ValueError("At least one of start_timestamp or end_timestamp must be specified")
+        episode_trims[cfg.operation.episode_index] = (
+            cfg.operation.start_timestamp,
+            cfg.operation.end_timestamp,
+        )
+    else:
+        raise ValueError("Either episode_trims or episode_index must be specified")
+
+    dataset = LeRobotDataset(cfg.repo_id, root=cfg.root)
+    output_repo_id, output_dir = get_output_path(
+        cfg.repo_id, cfg.new_repo_id, Path(cfg.root) if cfg.root else None
+    )
+
+    if cfg.new_repo_id is None:
+        dataset.root = Path(str(dataset.root) + "_old")
+
+    logging.info(f"Trimming {len(episode_trims)} episode(s) from {cfg.repo_id}")
+
+    # Get episode boundaries and find frames to keep for each episode
+    episodes_info = dataset.meta.episodes
+    all_frames_to_keep: dict[int, list[int]] = {}
+    
+    for ep_idx, (start_ts, end_ts) in episode_trims.items():
+        if ep_idx >= len(episodes_info["episode_index"]):
+            raise ValueError(f"Episode {ep_idx} does not exist (dataset has {len(episodes_info['episode_index'])} episodes)")
+
+        from_frame = episodes_info["dataset_from_index"][ep_idx]
+        to_frame = episodes_info["dataset_to_index"][ep_idx]
+        
+        logging.info(f"Episode {ep_idx}: trimming to [{start_ts}, {end_ts}]")
+        logging.info(f"  Original frames: {from_frame} to {to_frame} ({to_frame - from_frame} frames)")
+
+        # Find frames within timestamp range
+        frames_to_keep = []
+        for frame_idx in range(from_frame, to_frame):
+            frame = dataset.hf_dataset[frame_idx]
+            ts = frame["timestamp"]
+            
+            in_range = True
+            if start_ts is not None and ts < start_ts:
+                in_range = False
+            if end_ts is not None and ts > end_ts:
+                in_range = False
+            
+            if in_range:
+                frames_to_keep.append(frame_idx)
+
+        if not frames_to_keep:
+            raise ValueError(f"Episode {ep_idx}: No frames found in timestamp range [{start_ts}, {end_ts}]")
+
+        logging.info(f"  Keeping {len(frames_to_keep)} frames (indices {frames_to_keep[0]} to {frames_to_keep[-1]})")
+        all_frames_to_keep[ep_idx] = frames_to_keep
+
+    from lerobot.datasets.dataset_tools import trim_episodes_by_frames
+    
+    new_dataset = trim_episodes_by_frames(
+        dataset,
+        episode_frames_to_keep=all_frames_to_keep,
+        output_dir=output_dir,
+        repo_id=output_repo_id,
+    )
+
+    logging.info(f"Dataset saved to {output_dir}")
+    logging.info(f"Episodes: {new_dataset.meta.total_episodes}, Frames: {new_dataset.meta.total_frames}")
+
+    if cfg.push_to_hub:
+        logging.info(f"Pushing to hub as {output_repo_id}")
+        LeRobotDataset(output_repo_id, root=output_dir).push_to_hub()
 def handle_modify_tasks(cfg: EditDatasetConfig) -> None:
     if not isinstance(cfg.operation, ModifyTasksConfig):
         raise ValueError("Operation config must be ModifyTasksConfig")
@@ -431,10 +492,9 @@ def handle_modify_tasks(cfg: EditDatasetConfig) -> None:
     if new_task is None and episode_tasks_raw is None:
         raise ValueError("Must specify at least one of new_task or episode_tasks for modify_tasks operation")
 
-    if cfg.new_repo_id is not None or cfg.new_root is not None:
-        logging.warning(
-            "modify_tasks modifies datasets in-place. The --new_repo_id and --new_root parameters are ignored."
-        )
+    # Warn about in-place modification behavior
+    if cfg.new_repo_id is not None:
+        logging.warning("modify_tasks modifies datasets in-place. The --new_repo_id parameter is ignored.")
 
     dataset = LeRobotDataset(cfg.repo_id, root=cfg.root)
     logging.warning(f"Modifying dataset in-place at {dataset.root}. Original data will be overwritten.")
@@ -470,30 +530,32 @@ def handle_convert_image_to_video(cfg: EditDatasetConfig) -> None:
     dataset = LeRobotDataset(cfg.repo_id, root=cfg.root)
 
     # Determine output directory and repo_id
-    # Priority: 1) new_root, 2) new_repo_id, 3) operation.output_dir, 4) auto-generated name
+    # Priority: 1) new_repo_id, 2) operation.output_dir, 3) auto-generated name
     output_dir_config = getattr(cfg.operation, "output_dir", None)
-    if output_dir_config:
-        logging.warning(
-            "--operation.output_dir is deprecated and will be removed in future versions. "
-            "Please use --new_root instead."
-        )
 
-    if cfg.new_root:
-        output_dir = Path(cfg.new_root)
-        output_repo_id = cfg.new_repo_id or f"{cfg.repo_id}_video"
-        logging.info(f"Saving to new_root: {output_dir} as {output_repo_id}")
-    elif cfg.new_repo_id:
+    if cfg.new_repo_id:
+        # Use new_repo_id for both local storage and hub push
         output_repo_id = cfg.new_repo_id
-        output_dir = HF_LEROBOT_HOME / cfg.new_repo_id
+        # Place new dataset as a sibling to the original dataset
+        # Get the parent of the actual dataset root (not cfg.root which might be the lerobot cache dir)
+        # Extract just the dataset name (after last slash) for the local directory
+        local_dir_name = cfg.new_repo_id.split("/")[-1]
+        output_dir = dataset.root.parent / local_dir_name
         logging.info(f"Saving to new dataset: {cfg.new_repo_id} at {output_dir}")
     elif output_dir_config:
+        # Use custom output directory for local-only storage
         output_dir = Path(output_dir_config)
+        # Extract repo name from output_dir for the dataset
         output_repo_id = output_dir.name
-        logging.info(f"Saving to local directory: {output_dir} as {output_repo_id}")
+        logging.info(f"Saving to local directory: {output_dir}")
     else:
+        # Auto-generate name: append "_video" to original repo_id
         output_repo_id = f"{cfg.repo_id}_video"
-        output_dir = HF_LEROBOT_HOME / output_repo_id
-        logging.info(f"Saving to auto-generated location: {output_dir} as {output_repo_id}")
+        # Place new dataset as a sibling to the original dataset
+        # Extract just the dataset name (after last slash) for the local directory
+        local_dir_name = output_repo_id.split("/")[-1]
+        output_dir = dataset.root.parent / local_dir_name
+        logging.info(f"Saving to auto-generated location: {output_dir}")
 
     logging.info(f"Converting dataset {cfg.repo_id} to video format")
 
@@ -568,20 +630,8 @@ def handle_info(cfg: EditDatasetConfig):
         sys.stdout.write(f"{feature_dump_str}\n")
 
 
-def _validate_config(cfg: EditDatasetConfig) -> None:
-    if isinstance(cfg.operation, MergeConfig):
-        if not cfg.new_repo_id:
-            raise ValueError("--new_repo_id is required for merge operation (the merged dataset identifier)")
-    else:
-        if not cfg.repo_id:
-            raise ValueError(
-                f"--repo_id is required for {cfg.operation.type} operation (the input dataset identifier)"
-            )
-
-
 @parser.wrap()
 def edit_dataset(cfg: EditDatasetConfig) -> None:
-    _validate_config(cfg)
     operation_type = cfg.operation.type
 
     if operation_type == "delete_episodes":
@@ -596,6 +646,8 @@ def edit_dataset(cfg: EditDatasetConfig) -> None:
         handle_modify_tasks(cfg)
     elif operation_type == "convert_image_to_video":
         handle_convert_image_to_video(cfg)
+    elif operation_type == "trim_episode":
+        handle_trim_episode(cfg)
     elif operation_type == "info":
         handle_info(cfg)
     else:

src/lerobot/scripts/lerobot_find_joint_limits.py

@@ -61,7 +61,6 @@ from lerobot.teleoperators import (  # noqa: F401
     make_teleoperator_from_config,
     omx_leader,
     openarm_leader,
-    openarm_mini,
     so_leader,
 )
 from lerobot.utils.robot_utils import precise_sleep

src/lerobot/scripts/lerobot_record.py

@@ -125,7 +125,6 @@ from lerobot.teleoperators import (  # noqa: F401
     make_teleoperator_from_config,
     omx_leader,
     openarm_leader,
-    openarm_mini,
     reachy2_teleoperator,
     so_leader,
     unitree_g1,
@@ -155,7 +154,7 @@ class DatasetRecordConfig:
     repo_id: str
     # A short but accurate description of the task performed during the recording (e.g. "Pick the Lego block and drop it in the box on the right.")
     single_task: 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 | Path | None = None
     # Limit the frames per second.
     fps: int = 30
@@ -334,7 +333,6 @@ def record_loop(
         preprocessor.reset()
         postprocessor.reset()
 
-    no_action_count = 0
     timestamp = 0
     start_episode_t = time.perf_counter()
     while timestamp < control_time_s:
@@ -369,8 +367,6 @@ def record_loop(
             act_processed_policy: RobotAction = make_robot_action(action_values, dataset.features)
 
         elif policy is None and isinstance(teleop, Teleoperator):
-            if robot.name == "unitree_g1":
-                teleop.send_feedback(obs)
             act = teleop.get_action()
 
             # Applies a pipeline to the raw teleop action, default is IdentityProcessor
@@ -384,13 +380,11 @@ def record_loop(
             act = {**arm_action, **base_action} if len(base_action) > 0 else arm_action
             act_processed_teleop = teleop_action_processor((act, obs))
         else:
-            no_action_count += 1
-            if no_action_count == 1 or no_action_count % 10 == 0:
-                logging.warning(
-                    "No policy or teleoperator provided, skipping action generation. "
-                    "This is likely to happen when resetting the environment without a teleop device. "
-                    "The robot won't be at its rest position at the start of the next episode."
-                )
+            logging.info(
+                "No policy or teleoperator provided, skipping action generation."
+                "This is likely to happen when resetting the environment without a teleop device."
+                "The robot won't be at its rest position at the start of the next episode."
+            )
             continue
 
         # Applies a pipeline to the action, default is IdentityProcessor
@@ -558,6 +552,10 @@ def record(cfg: RecordConfig) -> LeRobotDataset:
                 ):
                     log_say("Reset the environment", cfg.play_sounds)
 
+                    # reset g1 robot
+                    if robot.name == "unitree_g1":
+                        robot.reset()
+
                     record_loop(
                         robot=robot,
                         events=events,

src/lerobot/scripts/lerobot_replay.py

@@ -80,7 +80,7 @@ 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

src/lerobot/scripts/lerobot_teleoperate.py

@@ -60,7 +60,6 @@ import rerun as rr
 
 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.processor import (
     RobotAction,
@@ -95,7 +94,6 @@ from lerobot.teleoperators import (  # noqa: F401
     make_teleoperator_from_config,
     omx_leader,
     openarm_leader,
-    openarm_mini,
     reachy2_teleoperator,
     so_leader,
     unitree_g1,
@@ -154,6 +152,7 @@ def teleop_loop(
 
     display_len = max(len(key) for key in robot.action_features)
     start = time.perf_counter()
+
     while True:
         loop_start = time.perf_counter()
 
@@ -163,9 +162,6 @@ def teleop_loop(
         # given that it is the identity processor as default
         obs = robot.get_observation()
 
-        if robot.name == "unitree_g1":
-            teleop.send_feedback(obs)
-
         # Get teleop action
         raw_action = teleop.get_action()
 

src/lerobot/scripts/lerobot_train.py

@@ -209,11 +209,7 @@ def train(cfg: TrainPipelineConfig, accelerator: Accelerator | None = None):
 
     # Use accelerator's device
     device = accelerator.device
-    if cfg.cudnn_deterministic:
-        torch.backends.cudnn.deterministic = True
-        torch.backends.cudnn.benchmark = False
-    else:
-        torch.backends.cudnn.benchmark = True
+    torch.backends.cudnn.benchmark = True
     torch.backends.cuda.matmul.allow_tf32 = True
 
     # Dataset loading synchronization: main process downloads first to avoid race conditions
@@ -384,10 +380,10 @@ def train(cfg: TrainPipelineConfig, accelerator: Accelerator | None = None):
         "dataloading_s": AverageMeter("data_s", ":.3f"),
     }
 
-    # Keep global batch size for logging; MetricsTracker handles world size internally.
+    # Use effective batch size for proper epoch calculation in distributed training
     effective_batch_size = cfg.batch_size * accelerator.num_processes
     train_tracker = MetricsTracker(
-        cfg.batch_size,
+        effective_batch_size,
         dataset.num_frames,
         dataset.num_episodes,
         train_metrics,

src/lerobot/scripts/lerobot_train_tokenizer.py

@@ -306,7 +306,7 @@ def train_fast_tokenizer(
 
     # download the tokenizer source code (not pretrained weights)
     # we'll train a new tokenizer on our own data
-    base_tokenizer = AutoProcessor.from_pretrained("lerobot/fast-action-tokenizer", trust_remote_code=True)
+    base_tokenizer = AutoProcessor.from_pretrained("physical-intelligence/fast", trust_remote_code=True)
 
     # convert action_chunks array to list of arrays (expected by .fit())
     action_data_list = [action_chunks[i] for i in range(len(action_chunks))]

src/lerobot/teleoperators/so_leader/config_so_leader.py

@@ -15,6 +15,7 @@
 # limitations under the License.
 
 from dataclasses import dataclass
+from typing import TypeAlias
 
 from ..config import TeleoperatorConfig
 
@@ -37,5 +38,5 @@ class SOLeaderTeleopConfig(TeleoperatorConfig, SOLeaderConfig):
     pass
 
 
-SO100LeaderConfig = SOLeaderTeleopConfig
-SO101LeaderConfig = SOLeaderTeleopConfig
+SO100LeaderConfig: TypeAlias = SOLeaderTeleopConfig
+SO101LeaderConfig: TypeAlias = SOLeaderTeleopConfig

src/lerobot/teleoperators/so_leader/so_leader.py

@@ -16,6 +16,7 @@
 
 import logging
 import time
+from typing import TypeAlias
 
 from lerobot.motors import Motor, MotorCalibration, MotorNormMode
 from lerobot.motors.feetech import (
@@ -155,5 +156,5 @@ class SOLeader(Teleoperator):
         logger.info(f"{self} disconnected.")
 
 
-SO100Leader = SOLeader
-SO101Leader = SOLeader
+SO100Leader: TypeAlias = SOLeader
+SO101Leader: TypeAlias = SOLeader

src/lerobot/teleoperators/unitree_g1/__init__.py

@@ -19,13 +19,3 @@ from .exo_calib import ExoskeletonCalibration, ExoskeletonJointCalibration
 from .exo_ik import ExoskeletonIKHelper
 from .exo_serial import ExoskeletonArm
 from .unitree_g1 import UnitreeG1Teleoperator
-
-__all__ = [
-    "ExoskeletonArmPortConfig",
-    "ExoskeletonCalibration",
-    "ExoskeletonIKHelper",
-    "ExoskeletonJointCalibration",
-    "ExoskeletonArm",
-    "UnitreeG1Teleoperator",
-    "UnitreeG1TeleoperatorConfig",
-]

src/lerobot/teleoperators/unitree_g1/exo_calib.py

@@ -35,9 +35,6 @@ import serial
 logger = logging.getLogger(__name__)
 
 
-ADC_MAX = 2**12 - 1
-ADC_HALF = ADC_MAX / 2
-
 # exoskeleton joint names -> ADC channel pairs. TODO: add wrist pitch and wrist yaw
 JOINTS = {
     "shoulder_pitch": (0, 1),
@@ -62,7 +59,7 @@ class ExoskeletonCalibration:
 
     version: int = 2
     side: str = ""
-    adc_max: int = ADC_MAX
+    adc_max: int = 2**12 - 1
     joints: list[ExoskeletonJointCalibration] = field(default_factory=list)
 
     def to_dict(self) -> dict:
@@ -95,7 +92,7 @@ class ExoskeletonCalibration:
         return cls(
             version=data.get("version", 2),
             side=data.get("side", ""),
-            adc_max=data.get("adc_max", ADC_MAX),
+            adc_max=data.get("adc_max", 2**12 - 1),
             joints=joints,
         )
 
@@ -115,8 +112,11 @@ class CalibParams:
 
 
 def normalize_angle(angle: float) -> float:
-    """Normalize angle to [-pi, pi]."""
-    return float(np.arctan2(np.sin(angle), np.cos(angle)))
+    while angle > np.pi:
+        angle -= 2 * np.pi
+    while angle < -np.pi:
+        angle += 2 * np.pi
+    return angle
 
 
 def joint_z_and_angle(raw16: list[int], j: ExoskeletonJointCalibration) -> tuple[np.ndarray, float]:
@@ -125,7 +125,7 @@ def joint_z_and_angle(raw16: list[int], j: ExoskeletonJointCalibration) -> tuple
     """
     pair = JOINTS[j.name]
     s, c = raw16[pair[0]], raw16[pair[1]]  # get sin and cos
-    p = np.array([float(c) - ADC_HALF, float(s) - ADC_HALF])  # center the raw values
+    p = np.array([float(c) - (2**12 - 1) / 2, float(s) - (2**12 - 1) / 2])  # center the raw values
     z = np.asarray(j.T) @ (
         p - np.asarray(j.center_fit)
     )  # center the ellipse and invert the transformation matrix to get unit circle coords
@@ -167,7 +167,7 @@ def run_exo_calibration(
 
     def read_joint_point(raw16: list[int], pair: tuple[int, int]):
         s, c = raw16[pair[0]], raw16[pair[1]]
-        return float(c) - ADC_HALF, float(s) - ADC_HALF, float(s), float(c)
+        return float(c) - (2**12 - 1) / 2, float(s) - (2**12 - 1) / 2, float(s), float(c)
 
     def select_fit_subset(xs, ys):
         """Select and filter points for ellipse fitting. Trims outliers by radius and downsamples."""
@@ -317,7 +317,7 @@ def run_exo_calibration(
                         calib = ExoskeletonCalibration(
                             version=2,
                             side=side,
-                            adc_max=ADC_MAX,
+                            adc_max=2**12 - 1,
                             joints=[
                                 ExoskeletonJointCalibration(
                                     name=j["name"],
@@ -367,8 +367,8 @@ def run_exo_calibration(
                     state["win_s"].append(s_raw)
                     state["win_c"].append(c_raw)
                     if len(state["win_s"]) >= max(3, params.median_window):
-                        state["ys"].append(running_median(state["win_s"]) - ADC_HALF)
-                        state["xs"].append(running_median(state["win_c"]) - ADC_HALF)
+                        state["ys"].append(running_median(state["win_s"]) - (2**12 - 1) / 2)
+                        state["xs"].append(running_median(state["win_c"]) - (2**12 - 1) / 2)
                 else:
                     jdata = joints_out[-1]
                     z = np.array(jdata["T"]) @ (np.array([x_raw, y_raw]) - np.array(jdata["center_fit"]))

src/lerobot/teleoperators/unitree_g1/exo_ik.py

@@ -25,8 +25,8 @@ from dataclasses import dataclass
 
 import numpy as np
 
-from lerobot.robots.unitree_g1.g1_kinematics import G1_29_ArmIK
 from lerobot.robots.unitree_g1.g1_utils import G1_29_JointArmIndex
+from lerobot.robots.unitree_g1.robot_kinematic_processor import G1_29_ArmIK
 
 from .exo_calib import JOINTS
 

src/lerobot/teleoperators/unitree_g1/exo_serial.py

@@ -32,29 +32,25 @@ def parse_raw16(line: bytes) -> list[int] | None:
         if len(parts) < 16:
             return None
         return [int(x) for x in parts[:16]]
-    except (ValueError, IndexError):
+    except Exception:
         return None
 
 
 def read_raw_from_serial(ser) -> list[int] | None:
     """Read latest sample from serial; if buffer is backed up, keep only the newest."""
-    try:
-        last = None
-        while ser.in_waiting > 0:
-            b = ser.readline()
-            if not b:
-                break
-            raw16 = parse_raw16(b)
-            if raw16 is not None:
-                last = raw16
-        if last is None:
-            b = ser.readline()
-            if b:
-                last = parse_raw16(b)
-        return last
-    except serial.SerialException as e:
-        logger.warning(f"Serial read error: {e}")
-        return None
+    last = None
+    while ser.in_waiting > 0:
+        b = ser.readline()
+        if not b:
+            break
+        raw16 = parse_raw16(b)
+        if raw16 is not None:
+            last = raw16
+    if last is None:
+        b = ser.readline()
+        if b:
+            last = parse_raw16(b)
+    return last
 
 
 @dataclass
@@ -119,6 +115,5 @@ class ExoskeletonArm:
         return {} if raw is None else exo_raw_to_angles(raw, self.calibration)
 
     def calibrate(self) -> None:
-        if not self.is_connected:
-            raise RuntimeError("Cannot calibrate: exoskeleton not connected")
-        self.calibration = run_exo_calibration(self._ser, self.side, self.calibration_fpath)
+        ser = self._ser
+        self.calibration = run_exo_calibration(ser, self.side, self.calibration_fpath)

src/lerobot/teleoperators/unitree_g1/unitree_g1.py

@@ -17,22 +17,9 @@
 import logging
 import time
 from functools import cached_property
-from typing import TYPE_CHECKING, Any
 
-from lerobot.robots.unitree_g1.g1_utils import REMOTE_AXES, G1_29_JointArmIndex
+from lerobot.robots.unitree_g1.g1_utils import G1_29_JointIndex
 from lerobot.utils.constants import HF_LEROBOT_CALIBRATION, TELEOPERATORS
-from lerobot.utils.import_utils import _unitree_sdk_available
-
-if TYPE_CHECKING or _unitree_sdk_available:
-    from unitree_sdk2py.utils.joystick import Joystick
-else:
-
-    class Joystick:
-        def __init__(self):
-            raise ImportError(
-                "unitree_sdk2py is required for RemoteController. Install with: pip install unitree_sdk2py"
-            )
-
 
 from ..teleoperator import Teleoperator
 from .config_unitree_g1 import UnitreeG1TeleoperatorConfig
@@ -42,120 +29,6 @@ from .exo_serial import ExoskeletonArm
 logger = logging.getLogger(__name__)
 
 
-class RemoteController:
-    """Unitree remote controller data parser for joystick and button state."""
-
-    # ADC parameters for exoskeleton joystick (12-bit ADC)
-    ADC_MAX = 4095
-    ADC_HALF = ADC_MAX / 2
-    JOYSTICK_X_IDX = 11  # X axis in raw ADC array
-    JOYSTICK_BTN_IDX = 12  # Button in raw ADC array
-    JOYSTICK_Y_IDX = 13  # Y axis in raw ADC array
-
-    # Map SDK named buttons to positional indices matching the wireless_remote
-    # byte layout (little-endian uint16 from bytes 2-3).
-    _BUTTON_MAP: list[str] = [
-        "RB",
-        "LB",
-        "start",
-        "back",
-        "RT",
-        "LT",
-        "",
-        "",
-        "A",
-        "B",
-        "X",
-        "Y",
-        "up",
-        "right",
-        "down",
-        "left",
-    ]
-
-    def __init__(self):
-        self.lx = 0.0
-        self.ly = 0.0
-        self.rx = 0.0
-        self.ry = 0.0
-        self.button = [0] * 16
-        self.remote_action = dict.fromkeys(REMOTE_AXES, 0.0)
-
-        # SDK joystick parser for wireless remote bytes
-        self._joystick = Joystick()
-        # Disable axis smoothing and deadzone to preserve raw values
-        for axis in (self._joystick.lx, self._joystick.ly, self._joystick.rx, self._joystick.ry):
-            axis.smooth = 1.0
-            axis.deadzone = 0.0
-
-        # Joystick center calibration (read at connect time)
-        self.left_center_x = self.ADC_HALF
-        self.left_center_y = self.ADC_HALF
-        self.right_center_x = self.ADC_HALF
-        self.right_center_y = self.ADC_HALF
-
-        # Whether to use exo joystick (detected at connect time)
-        self.use_left_exo_joystick = False
-        self.use_right_exo_joystick = False
-
-    def _sync_remote_action(self) -> None:
-        self.remote_action.update(zip(REMOTE_AXES, (self.lx, self.ly, self.rx, self.ry), strict=True))
-
-    def calibrate_center(self, raw16: list[int] | None, side: str) -> None:
-        if raw16 is None or len(raw16) < 16:
-            logger.info(f"{side.capitalize()} exo joystick: no data available")
-            return
-
-        btn_val = raw16[self.JOYSTICK_BTN_IDX]
-        logger.info(f"{side.capitalize()} exo joystick button ADC: {btn_val} (threshold: {self.ADC_HALF})")
-        if btn_val <= self.ADC_HALF:
-            logger.info(f"{side.capitalize()} exo joystick not detected (button below threshold)")
-            return
-
-        x = raw16[self.JOYSTICK_X_IDX]
-        y = raw16[self.JOYSTICK_Y_IDX]
-        if side == "left":
-            self.use_left_exo_joystick = True
-            self.left_center_x, self.left_center_y = x, y
-        else:
-            self.use_right_exo_joystick = True
-            self.right_center_x, self.right_center_y = x, y
-        logger.info(f"{side.capitalize()} exo joystick enabled, center: x={x}, y={y}")
-
-    def set_from_exo(self, raw16: list[int] | None, side: str) -> None:
-        if raw16 is None or len(raw16) < 16:
-            return
-
-        if side == "left":
-            if not self.use_left_exo_joystick:
-                return
-            self.lx = (raw16[self.JOYSTICK_X_IDX] - self.left_center_x) / self.ADC_HALF
-            self.ly = (raw16[self.JOYSTICK_Y_IDX] - self.left_center_y) / self.ADC_HALF
-            self.button[4] = 1 if raw16[self.JOYSTICK_BTN_IDX] < self.ADC_HALF else 0
-            return
-
-        if not self.use_right_exo_joystick:
-            return
-        self.rx = (raw16[self.JOYSTICK_X_IDX] - self.right_center_x) / self.ADC_HALF
-        self.ry = (raw16[self.JOYSTICK_Y_IDX] - self.right_center_y) / self.ADC_HALF
-        self.button[0] = 1 if raw16[self.JOYSTICK_BTN_IDX] < self.ADC_HALF else 0
-
-    def set_from_wireless(self, wireless_remote: bytes) -> None:
-        """Parse Unitree wireless remote raw bytes into joystick + button state."""
-        if len(wireless_remote) < 24:
-            return
-        self._joystick.extract(wireless_remote)
-
-        self.lx = self._joystick.lx.data
-        self.ly = self._joystick.ly.data
-        self.rx = self._joystick.rx.data
-        self.ry = self._joystick.ry.data
-
-        for i, name in enumerate(self._BUTTON_MAP):
-            if name:
-                self.button[i] = getattr(self._joystick, name).data
-
-
 class UnitreeG1Teleoperator(Teleoperator):
     """
     Bimanual exoskeleton arms teleoperator for Unitree G1 arms.
@@ -170,13 +43,6 @@ class UnitreeG1Teleoperator(Teleoperator):
     def __init__(self, config: UnitreeG1TeleoperatorConfig):
         super().__init__(config)
         self.config = config
-        left_exo_enabled = bool(config.left_arm_config.port.strip())
-        right_exo_enabled = bool(config.right_arm_config.port.strip())
-        if left_exo_enabled != right_exo_enabled:
-            raise ValueError(
-                "Invalid exo config: set both left/right exo ports, or leave both empty for remote-only mode."
-            )
-        self._arm_control_enabled = left_exo_enabled and right_exo_enabled
 
         # Setup calibration directory
         self.calibration_dir = (
@@ -204,37 +70,24 @@ class UnitreeG1Teleoperator(Teleoperator):
         )
 
         self.ik_helper: ExoskeletonIKHelper | None = None
-        self.remote_controller = RemoteController()
 
     @cached_property
     def action_features(self) -> dict[str, type]:
-        remote_features = dict.fromkeys(self.remote_controller.remote_action, float)
-        if not self._arm_control_enabled:
-            return remote_features
-        joint_features = {f"{name}.q": float for name in self._g1_arm_joint_names}
-        return {**joint_features, **remote_features}
+        return {f"{name}.q": float for name in self._g1_joint_names}
 
     @cached_property
     def feedback_features(self) -> dict[str, type]:
-        return {"wireless_remote": bytes}
+        return {}
 
     @property
     def is_connected(self) -> bool:
-        if not self._arm_control_enabled:
-            return True
         return self.left_arm.is_connected and self.right_arm.is_connected
 
     @property
     def is_calibrated(self) -> bool:
-        if not self._arm_control_enabled:
-            return True
         return self.left_arm.is_calibrated and self.right_arm.is_calibrated
 
     def connect(self, calibrate: bool = True) -> None:
-        if not self._arm_control_enabled:
-            logger.warning("Exo ports not fully configured; teleop will send joystick only (no arm actions)")
-            return
-
         self.left_arm.connect(calibrate)
         self.right_arm.connect(calibrate)
 
@@ -242,13 +95,6 @@ class UnitreeG1Teleoperator(Teleoperator):
         self.ik_helper = ExoskeletonIKHelper(frozen_joints=frozen_joints)
         logger.info("IK helper initialized")
 
-        time.sleep(0.1)  # Give serial time to populate buffer
-
-        left_raw = self.left_arm.read_raw()
-        right_raw = self.right_arm.read_raw()
-        self.remote_controller.calibrate_center(left_raw, "left")
-        self.remote_controller.calibrate_center(right_raw, "right")
-
     def calibrate(self) -> None:
         if not self.left_arm.is_calibrated:
             logger.info("Starting calibration for left arm...")
@@ -269,33 +115,12 @@ class UnitreeG1Teleoperator(Teleoperator):
         pass
 
     def get_action(self) -> dict[str, float]:
-        joint_action = {}
-        left_raw = None
-        right_raw = None
-        if self._arm_control_enabled:
-            left_raw = self.left_arm.read_raw()
-            right_raw = self.right_arm.read_raw()
+        left_angles = self.left_arm.get_angles()
+        right_angles = self.right_arm.get_angles()
+        return self.ik_helper.compute_g1_joints_from_exo(left_angles, right_angles)
 
-            left_angles = self.left_arm.get_angles()
-            right_angles = self.right_arm.get_angles()
-            joint_action = self.ik_helper.compute_g1_joints_from_exo(left_angles, right_angles)
-
-        # Wireless remote has priority when non-zero; otherwise, use exo joystick.
-        rc = self.remote_controller
-        wireless_active = (
-            abs(rc.lx) > 1e-3 or abs(rc.ly) > 1e-3 or abs(rc.rx) > 1e-3 or abs(rc.ry) > 1e-3
-        ) or any(rc.button)
-        if self._arm_control_enabled and not wireless_active:
-            rc.set_from_exo(left_raw, "left")
-            rc.set_from_exo(right_raw, "right")
-
-        rc._sync_remote_action()
-        return {**joint_action, **rc.remote_action}
-
-    def send_feedback(self, feedback: dict[str, Any]) -> None:
-        wireless_remote = feedback.get("wireless_remote")
-        if wireless_remote is not None:
-            self.remote_controller.set_from_wireless(wireless_remote)
+    def send_feedback(self, feedback: dict[str, float]) -> None:
+        raise NotImplementedError("Exoskeleton arms do not support feedback")
 
     def disconnect(self) -> None:
         self.left_arm.disconnect()
@@ -328,5 +153,5 @@ class UnitreeG1Teleoperator(Teleoperator):
             print("\n\nVisualization stopped.")
 
     @cached_property
-    def _g1_arm_joint_names(self) -> list[str]:
-        return [joint.name for joint in G1_29_JointArmIndex]
+    def _g1_joint_names(self) -> list[str]:
+        return [joint.name for joint in G1_29_JointIndex]

src/lerobot/utils/import_utils.py

@@ -74,8 +74,6 @@ _peft_available = is_package_available("peft")
 _scipy_available = is_package_available("scipy")
 _reachy2_sdk_available = is_package_available("reachy2_sdk")
 _can_available = is_package_available("python-can", "can")
-_unitree_sdk_available = is_package_available("unitree-sdk2py", "unitree_sdk2py")
-_pygame_available = is_package_available("pygame")
 
 
 def make_device_from_device_class(config: ChoiceRegistry) -> Any:

src/lerobot/utils/io_utils.py

@@ -16,10 +16,12 @@
 import json
 import warnings
 from pathlib import Path
+from typing import TypeVar
 
 import imageio
 
 JsonLike = str | int | float | bool | None | list["JsonLike"] | dict[str, "JsonLike"] | tuple["JsonLike", ...]
+T = TypeVar("T", bound=JsonLike)
 
 
 def write_video(video_path, stacked_frames, fps):
@@ -31,7 +33,7 @@ def write_video(video_path, stacked_frames, fps):
         imageio.mimsave(video_path, stacked_frames, fps=fps)
 
 
-def deserialize_json_into_object[T: JsonLike](fpath: Path, obj: T) -> T:
+def deserialize_json_into_object(fpath: Path, obj: T) -> T:
     """
     Loads the JSON data from `fpath` and recursively fills `obj` with the
     corresponding values (strictly matching structure and types).

src/lerobot/utils/logging_utils.py

@@ -104,10 +104,9 @@ class MetricsTracker:
         self.metrics = metrics
 
         self.steps = initial_step
-        world_size = accelerator.num_processes if accelerator else 1
         # A sample is an (observation,action) pair, where observation and action
         # can be on multiple timestamps. In a batch, we have `batch_size` number of samples.
-        self.samples = self.steps * self._batch_size * world_size
+        self.samples = self.steps * self._batch_size
         self.episodes = self.samples / self._avg_samples_per_ep
         self.epochs = self.samples / self._num_frames
         self.accelerator = accelerator
@@ -133,8 +132,7 @@ class MetricsTracker:
         Updates metrics that depend on 'step' for one step.
         """
         self.steps += 1
-        world_size = self.accelerator.num_processes if self.accelerator else 1
-        self.samples += self._batch_size * world_size
+        self.samples += self._batch_size * (self.accelerator.num_processes if self.accelerator else 1)
         self.episodes = self.samples / self._avg_samples_per_ep
         self.epochs = self.samples / self._num_frames
 

tests/artifacts/policies/pusht_diffusion_/actions.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:54aecbc1af72a4cd5e9261492f5e7601890517516257aacdf2a0ffb3ce281f1b
+oid sha256:19eaaa85f66ba4aa6388dbb83819ffad6ea4363247208f871a8dc385689f6fc8
 size 992

tests/artifacts/policies/pusht_diffusion_/grad_stats.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:88a9c3775a2aa1e90a08850521970070a4fcf0f6b82aab43cd8ccc5cf77e0013
+oid sha256:227296eaeeb54acdc3dae2eb8af3d4d08fb87e245337624447140b1e91cfd002
 size 47424

tests/artifacts/policies/pusht_diffusion_/output_dict.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:91a2635e05a75fe187a5081504c5f35ce3417378813fa2deaf9ca4e8200e1819
+oid sha256:271b00cb2f0cd5fd26b1d53463638e3d1a6e92692ec625fcffb420ca190869e5
 size 68