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

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

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

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

* fix(test): no decorator in yield

.github/workflows/fast_tests.yml

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

.github/workflows/full_tests.yml

@@ -37,7 +37,7 @@ permissions:
 # Sets up the environment variables
 env:
   UV_VERSION: "0.8.0"
-  PYTHON_VERSION: "3.10"
+  PYTHON_VERSION: "3.12"
   DOCKER_IMAGE_NAME: huggingface/lerobot-gpu
 
 # Ensures that only the latest action is built, canceling older runs.
@@ -60,6 +60,7 @@ jobs:
       MUJOCO_GL: egl
       HF_HOME: /mnt/cache/.cache/huggingface
       HF_LEROBOT_HOME: /mnt/cache/.cache/huggingface/lerobot
+      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     steps:
       - uses: actions/checkout@v6
         with:
@@ -87,6 +88,12 @@ jobs:
       - name: Install lerobot with all extras
         run: uv sync --extra all # TODO(Steven): Make flash-attn optional
 
+      - name: Login to Hugging Face
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          uv run hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
+          uv run hf auth whoami
+
       - name: Run pytest (all extras)
         run: uv run pytest tests -vv --maxfail=10
 
@@ -162,6 +169,7 @@ jobs:
       HF_LEROBOT_HOME: /home/user_lerobot/.cache/huggingface/lerobot
       TORCH_HOME: /home/user_lerobot/.cache/torch
       TRITON_CACHE_DIR: /home/user_lerobot/.cache/triton
+      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     container:
       image: ${{ needs.build-and-push-docker.outputs.image_tag }} # zizmor: ignore[unpinned-images]
       options: --gpus all --shm-size "16gb"
@@ -173,8 +181,13 @@ jobs:
         shell: bash
         working-directory: /lerobot
     steps:
+      - name: Login to Hugging Face
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
+          hf auth whoami
       - name: Fix ptxas permissions
-        run: chmod +x /lerobot/.venv/lib/python3.10/site-packages/triton/backends/nvidia/bin/ptxas
+        run: chmod +x /lerobot/.venv/lib/python3.12/site-packages/triton/backends/nvidia/bin/ptxas
       - name: Run pytest on GPU
         run: pytest tests -vv --maxfail=10
       - name: Run end-to-end tests

.github/workflows/nightly.yml

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

.github/workflows/quality.yml

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

.github/workflows/release.yml

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

.github/workflows/unbound_deps_tests.yml

@@ -29,7 +29,7 @@ permissions:
 # Sets up the environment variables
 env:
   UV_VERSION: "0.8.0"
-  PYTHON_VERSION: "3.10"
+  PYTHON_VERSION: "3.12"
   DOCKER_IMAGE_NAME: huggingface/lerobot-gpu:unbound
 
 # Ensures that only the latest action is built, canceling older runs.
@@ -48,6 +48,7 @@ jobs:
       MUJOCO_GL: egl
       HF_HOME: /mnt/cache/.cache/huggingface
       HF_LEROBOT_HOME: /mnt/cache/.cache/huggingface/lerobot
+      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     steps:
       - uses: actions/checkout@v6
         with:
@@ -79,7 +80,11 @@ jobs:
 
       - name: Install lerobot with all extras
         run: uv sync --extra all # TODO(Steven): Make flash-attn optional
-
+      - name: Login to Hugging Face
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          uv run hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
+          uv run hf auth whoami
       - name: Run pytest (all extras)
         run: uv run pytest tests -vv
 
@@ -137,6 +142,7 @@ jobs:
       HF_LEROBOT_HOME: /home/user_lerobot/.cache/huggingface/lerobot
       TORCH_HOME: /home/user_lerobot/.cache/torch
       TRITON_CACHE_DIR: /home/user_lerobot/.cache/triton
+      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     container:
       image: ${{ needs.build-and-push-docker.outputs.image_tag }} # zizmor: ignore[unpinned-images]
       options: --gpus all --shm-size "16gb"
@@ -148,6 +154,11 @@ jobs:
         shell: bash
         working-directory: /lerobot
     steps:
+      - name: Login to Hugging Face
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
+          hf auth whoami
       - name: Run pytest on GPU
         run: pytest tests -vv
       - name: Run end-to-end tests

.pre-commit-config.yaml

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

AI_POLICY.md

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

CONTRIBUTING.md

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

README.md

@@ -100,11 +100,11 @@ lerobot-train \
   --dataset.repo_id=lerobot/aloha_mobile_cabinet
 ```
 
-| Category                   | Models                                                                                                                                                                                                       |
-| -------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
-| **Imitation Learning**     | [ACT](./docs/source/policy_act_README.md), [Diffusion](./docs/source/policy_diffusion_README.md), [VQ-BeT](./docs/source/policy_vqbet_README.md)                                                             |
-| **Reinforcement Learning** | [HIL-SERL](./docs/source/hilserl.mdx), [TDMPC](./docs/source/policy_tdmpc_README.md) & QC-FQL (coming soon)                                                                                                  |
-| **VLAs Models**            | [Pi0Fast](./docs/source/pi0fast.mdx), [Pi0.5](./docs/source/pi05.mdx), [GR00T N1.5](./docs/source/policy_groot_README.md), [SmolVLA](./docs/source/policy_smolvla_README.md), [XVLA](./docs/source/xvla.mdx) |
+| Category                   | Models                                                                                                                                                                                                                  |
+| -------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| **Imitation Learning**     | [ACT](./docs/source/policy_act_README.md), [Diffusion](./docs/source/policy_diffusion_README.md), [VQ-BeT](./docs/source/policy_vqbet_README.md), [Multitask DiT Policy](./docs/source/policy_multi_task_dit_README.md) |
+| **Reinforcement Learning** | [HIL-SERL](./docs/source/hilserl.mdx), [TDMPC](./docs/source/policy_tdmpc_README.md) & QC-FQL (coming soon)                                                                                                             |
+| **VLAs Models**            | [Pi0Fast](./docs/source/pi0fast.mdx), [Pi0.5](./docs/source/pi05.mdx), [GR00T N1.5](./docs/source/policy_groot_README.md), [SmolVLA](./docs/source/policy_smolvla_README.md), [XVLA](./docs/source/xvla.mdx)            |
 
 Similarly to the hardware, you can easily implement your own policy & leverage LeRobot's data collection, training, and visualization tools, and share your model to the HF Hub
 

docker/Dockerfile.internal

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

docker/Dockerfile.user

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

docs/source/_toctree.yml

@@ -47,6 +47,8 @@
     title: NVIDIA GR00T N1.5
   - local: xvla
     title: X-VLA
+  - local: multi_task_dit
+    title: Multitask DiT Policy
   - local: walloss
     title: WALL-OSS
   title: "Policies"

docs/source/async.mdx

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

docs/source/bring_your_own_policies.mdx

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

docs/source/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 3.10 or newer
+- Computer with Python 3.12 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
-huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+hf auth login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
 ```
 
 Store your Hugging Face username:
 
 ```bash
-HF_USER=$(huggingface-cli whoami | head -n 1)
+HF_USER=$(hf auth whoami | awk -F': *' 'NR==1 {print $2}')
 echo $HF_USER
 ```
 

docs/source/envhub.mdx

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

docs/source/il_robots.mdx

@@ -159,7 +159,7 @@ We use the Hugging Face hub features for uploading your dataset. If you haven't
 Add your token to the CLI by running this command:
 
 ```bash
-huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+hf auth login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
 ```
 
 Then store your Hugging Face repository name in a variable:
@@ -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
-huggingface-cli upload ${HF_USER}/record-test ~/.cache/huggingface/lerobot/{repo-id} --repo-type dataset
+hf upload ${HF_USER}/record-test ~/.cache/huggingface/lerobot/{repo-id} --repo-type dataset
 ```
 
 #### Record function
@@ -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
-huggingface-cli upload ${HF_USER}/act_so101_test \
+hf 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
-huggingface-cli upload ${HF_USER}/act_so101_test${CKPT} \
+hf upload ${HF_USER}/act_so101_test${CKPT} \
   outputs/train/act_so101_test/checkpoints/${CKPT}/pretrained_model
 ```
 

docs/source/installation.mdx

@@ -1,6 +1,6 @@
 # Installation
 
-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-).
+This guide uses conda (via miniforge) to manage environments. If you prefer another environment manager (e.g. `uv`, `venv`), ensure you have Python >=3.12 and ffmpeg installed with the `libsvtav1` encoder, then skip ahead to [Install LeRobot](#step-3-install-lerobot-).
 
 ## Step 1: Install [`miniforge`](https://conda-forge.org/download/)
 
@@ -11,10 +11,10 @@ bash Miniforge3-$(uname)-$(uname -m).sh
 
 ## Step 2: Environment Setup
 
-Create a virtual environment with Python 3.10, using conda:
+Create a virtual environment with Python 3.12, using conda:
 
 ```bash
-conda create -y -n lerobot python=3.10
+conda create -y -n lerobot python=3.12
 ```
 
 Then activate your conda environment, you have to do this each time you open a shell to use lerobot:
@@ -90,9 +90,6 @@ _Replace `[...]` with your desired features._
 For a full list of optional dependencies, see:
 https://pypi.org/project/lerobot/
 
-> [!NOTE]
-> For lerobot 0.4.0, if you want to install pi, you will have to do: `pip install "lerobot[pi]@git+https://github.com/huggingface/lerobot.git"`
-
 ### Troubleshooting
 
 If you encounter build errors, you may need to install additional dependencies: `cmake`, `build-essential`, and `ffmpeg libs`.

docs/source/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
-huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+hf auth login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
 ```
 
 Then store your Hugging Face repository name in a variable:
 
 ```bash
-HF_USER=$(huggingface-cli whoami | head -n 1)
+HF_USER=$(hf auth whoami | awk -F': *' 'NR==1 {print $2}')
 echo $HF_USER
 ```
 

docs/source/multi_task_dit.mdx

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

docs/source/pi0.mdx

@@ -34,11 +34,6 @@ As described by Physical Intelligence, while AI has achieved remarkable success
    pip install -e ".[pi]"
    ```
 
-   > [!NOTE]
-   > For lerobot 0.4.0, if you want to install pi tag, you will have to do: `pip install "lerobot[pi]@git+https://github.com/huggingface/lerobot.git"`.
-   >
-   > This will be solved in the next patch release
-
 ## Training Data and Capabilities
 
 π₀ is trained on the largest robot interaction dataset to date, combining three key data sources:

docs/source/pi05.mdx

@@ -36,11 +36,6 @@ This diverse training mixture creates a "curriculum" that enables generalization
    pip install -e ".[pi]"
    ```
 
-   > [!NOTE]
-   > For lerobot 0.4.0, if you want to install pi tag, you will have to do: `pip install "lerobot[pi]@git+https://github.com/huggingface/lerobot.git"`.
-   >
-   > This will be solved in the next patch release
-
 ## Usage
 
 To use π₀.₅ in your LeRobot configuration, specify the policy type as:

docs/source/pi0fast.mdx

@@ -43,16 +43,11 @@ 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 `physical-intelligence/fast` tokenizer was trained on 1M+ real robot action sequences and works as a general-purpose tokenizer.
+1. **Use the pre-trained tokenizer**: The `lerobot/fast-action-tokenizer` tokenizer was trained on 1M+ real robot action sequences and works as a general-purpose tokenizer.
 
 2. **Train your own tokenizer**: For maximum performance on your specific dataset, you can finetune the tokenizer on your own data.
 
@@ -114,15 +109,15 @@ 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                              | `physical-intelligence/fast` |
-| `--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                              | `lerobot/fast-action-tokenizer` |
+| `--policy.compile_model=true`          | Enable torch.compile for faster training           | `false`                         |
 
 ## Inference
 

docs/source/policy_multi_task_dit_README.md

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

docs/source/unitree_g1.mdx

@@ -123,7 +123,7 @@ SSH into the robot and install LeRobot:
 ```bash
 ssh unitree@<YOUR_ROBOT_IP>
 
-conda create -y -n lerobot python=3.10
+conda create -y -n lerobot python=3.12
 conda activate lerobot
 git clone https://github.com/huggingface/lerobot.git
 cd lerobot
@@ -153,7 +153,7 @@ With the robot server running, you can now control the robot remotely. Let's lau
 ### Step 1: Install LeRobot on your machine
 
 ```bash
-conda create -y -n lerobot python=3.10
+conda create -y -n lerobot python=3.12
 conda activate lerobot
 git clone https://github.com/huggingface/lerobot.git
 cd lerobot

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').
+    # Root directory where the dataset will be stored (e.g. 'dataset/path'). If None, defaults to $HF_LEROBOT_HOME/repo_id.
     root: str | Path | None = None
     # Limit the frames per second. By default, uses the policy fps.
     fps: int = 30

examples/dataset/slurm_compute_rabc.py

@@ -0,0 +1,490 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+SLURM-distributed SARM RA-BC annotation pipeline.
+
+Computes SARM progress values for all frames in a dataset, distributed across
+SLURM workers, then merges the shards into a single sarm_progress.parquet.
+
+Two subcommands, each a separate SLURM submission:
+
+  compute    – N workers, each computes progress for a subset of episodes
+  aggregate  – 1 worker, merges N shards into sarm_progress.parquet, pushes to hub
+
+Usage:
+    python slurm_compute_rabc.py compute \\
+        --repo-id user/dataset --reward-model-path user/sarm_model \\
+        --stride 10 --device cpu --workers 50 --partition cpu
+
+    python slurm_compute_rabc.py aggregate \\
+        --repo-id user/dataset --reward-model-path user/sarm_model \\
+        --partition cpu --push-to-hub
+"""
+
+import argparse
+from pathlib import Path
+
+from datatrove.executor import LocalPipelineExecutor
+from datatrove.executor.slurm import SlurmPipelineExecutor
+from datatrove.pipeline.base import PipelineStep
+
+
+class ComputeProgressShards(PipelineStep):
+    """Each worker computes SARM progress for its assigned episodes."""
+
+    def __init__(
+        self, repo_id, reward_model_path, stride=1, head_mode="sparse", device="cpu", shard_dir="rabc_shards"
+    ):
+        super().__init__()
+        if stride < 1:
+            raise ValueError(f"stride must be >= 1, got {stride}")
+        self.repo_id = repo_id
+        self.reward_model_path = reward_model_path
+        self.stride = stride
+        self.head_mode = head_mode
+        self.device = device
+        self.shard_dir = shard_dir
+
+    def run(self, data=None, rank: int = 0, world_size: int = 1):
+        import logging
+        from pathlib import Path
+
+        import numpy as np
+        import pyarrow as pa
+        import pyarrow.parquet as pq
+        import torch
+        from tqdm import tqdm
+
+        from lerobot.policies.sarm.compute_rabc_weights import (
+            generate_all_frame_indices,
+            interpolate_progress,
+            load_sarm_resources,
+        )
+        from lerobot.utils.utils import init_logging
+
+        init_logging()
+
+        dataset, reward_model, preprocess = load_sarm_resources(
+            self.repo_id,
+            self.reward_model_path,
+            self.device,
+        )
+
+        if hasattr(preprocess, "eval"):
+            preprocess.eval()
+        for step in preprocess.steps:
+            if hasattr(step, "eval"):
+                step.eval()
+
+        image_key = reward_model.config.image_key
+        state_key = reward_model.config.state_key
+        frame_gap = reward_model.config.frame_gap
+        center_idx = reward_model.config.n_obs_steps // 2
+
+        dual_mode = reward_model.config.uses_dual_heads
+        compute_sparse = self.head_mode in ("sparse", "both") or not dual_mode
+        compute_dense = self.head_mode in ("dense", "both") and dual_mode
+
+        my_episodes = list(range(dataset.num_episodes))[rank::world_size]
+        if not my_episodes:
+            logging.info(f"Rank {rank}: no episodes assigned")
+            return
+        logging.info(f"Rank {rank}: {len(my_episodes)} / {dataset.num_episodes} episodes")
+
+        all_rows = []
+
+        for ep_idx in tqdm(my_episodes, desc=f"Rank {rank}"):
+            ep = dataset.meta.episodes[ep_idx]
+            ep_start, ep_end = ep["dataset_from_index"], ep["dataset_to_index"]
+            task = dataset[ep_start].get("task", "perform the task")
+
+            all_ep_indices = generate_all_frame_indices(ep_start, ep_end, frame_gap)
+            if self.stride > 1:
+                compute_indices = [i for i in all_ep_indices if (i - ep_start) % self.stride == 0]
+                if (ep_end - 1) not in compute_indices:
+                    compute_indices.append(ep_end - 1)
+                compute_indices = sorted(set(compute_indices))
+            else:
+                compute_indices = all_ep_indices
+
+            frame_results = {}
+            for qi in tqdm(compute_indices, desc=f"  Ep {ep_idx}", leave=False):
+                try:
+                    sample = dataset[qi]
+                    batch = {
+                        image_key: sample[image_key],
+                        "task": task,
+                        "index": qi,
+                        "episode_index": ep_idx,
+                    }
+                    if state_key in sample:
+                        batch[state_key] = sample[state_key]
+
+                    with torch.no_grad():
+                        processed = preprocess(batch)
+                        vf = processed["video_features"].to(self.device)
+                        tf = processed["text_features"].to(self.device)
+                        sf = processed.get("state_features")
+                        if sf is not None:
+                            sf = sf.to(self.device)
+                        lengths = processed.get("lengths")
+
+                        sparse_val = dense_val = np.nan
+                        if compute_sparse:
+                            r = reward_model.calculate_rewards(
+                                text_embeddings=tf,
+                                video_embeddings=vf,
+                                state_features=sf,
+                                lengths=lengths,
+                                return_all_frames=True,
+                                head_mode="sparse",
+                            )
+                            sparse_val = float(r[0, center_idx] if r.ndim == 2 else r[center_idx])
+                        if compute_dense:
+                            r = reward_model.calculate_rewards(
+                                text_embeddings=tf,
+                                video_embeddings=vf,
+                                state_features=sf,
+                                lengths=lengths,
+                                return_all_frames=True,
+                                head_mode="dense",
+                            )
+                            dense_val = float(r[0, center_idx] if r.ndim == 2 else r[center_idx])
+
+                        frame_results[qi] = (sparse_val, dense_val)
+                except Exception as e:
+                    logging.warning(f"Failed frame {qi}: {e}")
+
+            if not frame_results:
+                logging.warning(f"Episode {ep_idx}: all frames failed, skipping")
+                continue
+
+            # Interpolate to all frames in this episode
+            computed_idx = np.array(sorted(frame_results.keys()))
+            all_frame_arr = np.arange(ep_start, ep_end)
+
+            sparse_vals = np.array([frame_results[i][0] for i in computed_idx]) if compute_sparse else None
+            dense_vals = np.array([frame_results[i][1] for i in computed_idx]) if compute_dense else None
+
+            if self.stride > 1 and len(computed_idx) > 1:
+                if compute_sparse:
+                    sparse_vals = interpolate_progress(computed_idx, sparse_vals, all_frame_arr)
+                if compute_dense:
+                    dense_vals = interpolate_progress(computed_idx, dense_vals, all_frame_arr)
+                output_frames = all_frame_arr
+            else:
+                # Use only successfully computed frames to avoid indexing mismatch on failures
+                output_frames = computed_idx
+
+            for i, fi in enumerate(output_frames):
+                row = {"index": int(fi), "episode_index": ep_idx, "frame_index": int(fi - ep_start)}
+                if compute_sparse:
+                    row["progress_sparse"] = float(sparse_vals[i])
+                if compute_dense:
+                    row["progress_dense"] = float(dense_vals[i])
+                all_rows.append(row)
+
+        if all_rows:
+            import pandas as pd
+
+            df = pd.DataFrame(all_rows).sort_values("index").reset_index(drop=True)
+            table = pa.Table.from_pandas(df, preserve_index=False)
+            table = table.replace_schema_metadata({b"reward_model_path": self.reward_model_path.encode()})
+            shard_dir = Path(self.shard_dir)
+            shard_dir.mkdir(parents=True, exist_ok=True)
+            out = shard_dir / f"shard_{rank:05d}.parquet"
+            pq.write_table(table, out)
+            logging.info(f"Rank {rank}: saved {len(df)} rows to {out}")
+
+
+class AggregateProgress(PipelineStep):
+    """Merge all shard parquets into final sarm_progress.parquet."""
+
+    def __init__(self, repo_id, reward_model_path, shard_dir="rabc_shards", push_to_hub=False):
+        super().__init__()
+        self.repo_id = repo_id
+        self.reward_model_path = reward_model_path
+        self.shard_dir = shard_dir
+        self.push_to_hub = push_to_hub
+
+    def run(self, data=None, rank: int = 0, world_size: int = 1):
+        import datetime
+        import logging
+        import os
+        from pathlib import Path
+
+        import pandas as pd
+        import pyarrow as pa
+        import pyarrow.parquet as pq
+
+        from lerobot.datasets.lerobot_dataset import LeRobotDataset
+        from lerobot.utils.utils import init_logging
+
+        init_logging()
+        if rank != 0:
+            return
+
+        shard_dir = Path(self.shard_dir)
+        shards = sorted(shard_dir.glob("shard_*.parquet"))
+        if not shards:
+            raise FileNotFoundError(f"No shards found in {shard_dir}")
+
+        # Log shard modification time range to help detect stale files
+        mtimes = [os.path.getmtime(s) for s in shards]
+        oldest = datetime.datetime.fromtimestamp(min(mtimes)).isoformat(timespec="seconds")
+        newest = datetime.datetime.fromtimestamp(max(mtimes)).isoformat(timespec="seconds")
+        logging.info(f"Aggregating {len(shards)} shards (oldest: {oldest}, newest: {newest})")
+
+        df = pd.concat([pd.read_parquet(s) for s in shards], ignore_index=True)
+        df = df.sort_values("index").reset_index(drop=True)
+
+        table = pa.Table.from_pandas(df, preserve_index=False)
+        table = table.replace_schema_metadata({b"reward_model_path": self.reward_model_path.encode()})
+
+        temp_ds = LeRobotDataset(self.repo_id, download_videos=False)
+        out_path = Path(temp_ds.root) / "sarm_progress.parquet"
+        out_path.parent.mkdir(parents=True, exist_ok=True)
+        pq.write_table(table, out_path)
+        logging.info(f"Saved {len(df)} rows to {out_path}")
+
+        for col in ["progress_sparse", "progress_dense"]:
+            if col in df.columns:
+                v = df[col].dropna()
+                logging.info(
+                    f"{col}: mean={v.mean():.4f} std={v.std():.4f} min={v.min():.4f} max={v.max():.4f}"
+                )
+
+        if self.push_to_hub:
+            from huggingface_hub import HfApi
+
+            api = HfApi()
+            hub_path = "sarm_progress.parquet"
+            logging.info(f"Uploading to {self.repo_id}/{hub_path}")
+            api.upload_file(
+                path_or_fileobj=str(out_path),
+                path_in_repo=hub_path,
+                repo_id=self.repo_id,
+                repo_type="dataset",
+            )
+            logging.info(f"Uploaded: https://huggingface.co/datasets/{self.repo_id}/blob/main/{hub_path}")
+
+
+def make_compute_executor(
+    repo_id,
+    reward_model_path,
+    stride,
+    head_mode,
+    device,
+    shard_dir,
+    logs_dir,
+    job_name,
+    slurm,
+    workers,
+    partition,
+    cpus_per_task,
+    mem_per_cpu,
+):
+    kwargs = {
+        "pipeline": [
+            ComputeProgressShards(repo_id, reward_model_path, stride, head_mode, device, str(shard_dir)),
+        ],
+        "logging_dir": str(logs_dir / job_name),
+    }
+
+    if slurm:
+        kwargs.update(
+            {
+                "job_name": job_name,
+                "tasks": workers,
+                "workers": workers,
+                "time": "24:00:00",
+                "partition": partition,
+                "cpus_per_task": cpus_per_task,
+                "sbatch_args": {"mem-per-cpu": mem_per_cpu},
+            }
+        )
+        return SlurmPipelineExecutor(**kwargs)
+
+    kwargs.update({"tasks": workers, "workers": 1})
+    return LocalPipelineExecutor(**kwargs)
+
+
+def make_aggregate_executor(
+    repo_id,
+    reward_model_path,
+    shard_dir,
+    logs_dir,
+    job_name,
+    slurm,
+    partition,
+    cpus_per_task,
+    mem_per_cpu,
+    push_to_hub,
+):
+    kwargs = {
+        "pipeline": [
+            AggregateProgress(repo_id, reward_model_path, str(shard_dir), push_to_hub),
+        ],
+        "logging_dir": str(logs_dir / job_name),
+    }
+
+    if slurm:
+        kwargs.update(
+            {
+                "job_name": job_name,
+                "tasks": 1,
+                "workers": 1,
+                "time": "02:00:00",
+                "partition": partition,
+                "cpus_per_task": cpus_per_task,
+                "sbatch_args": {"mem-per-cpu": mem_per_cpu},
+            }
+        )
+        return SlurmPipelineExecutor(**kwargs)
+
+    kwargs.update({"tasks": 1, "workers": 1})
+    return LocalPipelineExecutor(**kwargs)
+
+
+def _add_shared_args(p):
+    p.add_argument(
+        "--repo-id",
+        type=str,
+        required=True,
+        help="Hugging Face repository identifier, e.g. 'user/dataset'.",
+    )
+    p.add_argument(
+        "--shard-dir",
+        type=Path,
+        default=Path("rabc_shards"),
+        help="Directory to read/write per-rank parquet shards.",
+    )
+    p.add_argument(
+        "--logs-dir",
+        type=Path,
+        default=Path("logs"),
+        help="Directory for datatrove logs.",
+    )
+    p.add_argument(
+        "--job-name",
+        type=str,
+        default=None,
+        help="SLURM job name (defaults to rabc_<subcommand>).",
+    )
+    p.add_argument(
+        "--slurm",
+        type=int,
+        default=1,
+        help="1 = submit via SLURM; 0 = run locally (useful for debugging).",
+    )
+    p.add_argument(
+        "--partition",
+        type=str,
+        default=None,
+        help="SLURM partition to submit to.",
+    )
+    p.add_argument(
+        "--cpus-per-task",
+        type=int,
+        default=4,
+        help="Number of CPUs per SLURM task.",
+    )
+    p.add_argument(
+        "--mem-per-cpu",
+        type=str,
+        default="4G",
+        help="Memory per CPU, e.g. '4G' or '1950M'.",
+    )
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="SLURM-distributed SARM RA-BC annotation pipeline",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+    )
+    sub = parser.add_subparsers(dest="command", required=True)
+
+    # compute subcommand
+    cp = sub.add_parser(
+        "compute",
+        help="Distribute progress computation across SLURM workers.",
+    )
+    _add_shared_args(cp)
+    cp.add_argument(
+        "--reward-model-path",
+        type=str,
+        required=True,
+        help="Path or HF repo id of the SARM reward model.",
+    )
+    cp.add_argument(
+        "--stride",
+        type=int,
+        default=1,
+        help="Compute every Nth frame; intermediate frames are interpolated (must be >= 1).",
+    )
+    cp.add_argument(
+        "--head-mode",
+        type=str,
+        default="sparse",
+        choices=["sparse", "dense", "both"],
+        help="Which reward head(s) to compute.",
+    )
+    cp.add_argument(
+        "--device",
+        type=str,
+        default="cpu",
+        help="Device for reward model inference, e.g. 'cpu' or 'cuda'.",
+    )
+    cp.add_argument(
+        "--workers",
+        type=int,
+        default=50,
+        help="Number of parallel SLURM tasks (one shard per worker).",
+    )
+
+    # aggregate subcommand
+    ap = sub.add_parser(
+        "aggregate",
+        help="Merge per-rank shards into a single sarm_progress.parquet.",
+    )
+    _add_shared_args(ap)
+    ap.add_argument(
+        "--reward-model-path",
+        type=str,
+        required=True,
+        help="Path or HF repo id of the SARM reward model (stored in parquet metadata).",
+    )
+    ap.add_argument(
+        "--push-to-hub",
+        action="store_true",
+        help="Upload sarm_progress.parquet to the Hugging Face Hub after aggregation.",
+    )
+
+    args = parser.parse_args()
+    job_name = args.job_name or f"rabc_{args.command}"
+    kwargs = vars(args)
+    kwargs["slurm"] = kwargs.pop("slurm") == 1
+    kwargs["job_name"] = job_name
+    command = kwargs.pop("command")
+
+    executor = make_compute_executor(**kwargs) if command == "compute" else make_aggregate_executor(**kwargs)
+
+    executor.run()
+
+
+if __name__ == "__main__":
+    main()

pyproject.toml

@@ -25,11 +25,11 @@ discord = "https://discord.gg/s3KuuzsPFb"
 
 [project]
 name = "lerobot"
-version = "0.4.4"
+version = "0.4.5"
 description = "🤗 LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch"
 dynamic = ["readme"]
 license = { text = "Apache-2.0" }
-requires-python = ">=3.10"
+requires-python = ">=3.12"
 authors = [
     { name = "Rémi Cadène", email = "re.cadene@gmail.com" },
     { name = "Simon Alibert", email = "alibert.sim@gmail.com" },
@@ -50,7 +50,8 @@ classifiers = [
     "Intended Audience :: Education",
     "Intended Audience :: Science/Research",
     "License :: OSI Approved :: Apache Software License",
-    "Programming Language :: Python :: 3.10",
+    "Programming Language :: Python :: 3.12",
+    "Programming Language :: Python :: 3.13",
     "Topic :: Software Development :: Build Tools",
     "Topic :: Scientific/Engineering :: Artificial Intelligence",
 ]
@@ -61,26 +62,28 @@ dependencies = [
     # Hugging Face dependencies
     "datasets>=4.0.0,<5.0.0",
     "diffusers>=0.27.2,<0.36.0",
-    "huggingface-hub[hf-transfer,cli]>=0.34.2,<0.36.0",
+    "huggingface-hub>=1.0.0,<2.0.0",
     "accelerate>=1.10.0,<2.0.0",
 
     # Core dependencies
+    "numpy>=2.0.0,<2.3.0", # NOTE: Explicitly listing numpy helps the resolver converge faster. Upper bound imposed by opencv-python-headless.
     "setuptools>=71.0.0,<81.0.0",
     "cmake>=3.29.0.1,<4.2.0",
+    "packaging>=24.2,<26.0",
+
+    "torch>=2.2.1,<2.11.0",
+    "torchcodec>=0.2.1,<0.11.0; sys_platform != 'win32' and (sys_platform != 'linux' or (platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')) and (sys_platform != 'darwin' or platform_machine != 'x86_64')",
+    "torchvision>=0.21.0,<0.26.0",
+
     "einops>=0.8.0,<0.9.0",
     "opencv-python-headless>=4.9.0,<4.13.0",
     "av>=15.0.0,<16.0.0",
     "jsonlines>=4.0.0,<5.0.0",
-    "packaging>=24.2,<26.0",
-    "pynput>=1.7.7,<1.9.0",
+    "pynput>=1.7.8,<1.9.0",
     "pyserial>=3.5,<4.0",
+
     "wandb>=0.24.0,<0.25.0",
-
-    "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 ==
+    "draccus==0.10.0", # TODO: Relax version constraint
     "gymnasium>=1.1.1,<2.0.0",
     "rerun-sdk>=0.24.0,<0.27.0",
 
@@ -95,10 +98,14 @@ dependencies = [
 
 # Common
 pygame-dep = ["pygame>=2.5.1,<2.7.0"]
-placo-dep = ["placo>=0.9.6,<0.10.0"]
-transformers-dep = ["transformers>=4.57.1,<5.0.0"]
+placo-dep = ["placo>=0.9.6,<0.9.17"]
+transformers-dep = ["transformers>=5.3.0,<6.0.0"]
 grpcio-dep = ["grpcio==1.73.1", "protobuf>=6.31.1,<6.32.0"]
 can-dep = ["python-can>=4.2.0,<5.0.0"]
+peft-dep = ["peft>=0.18.0,<1.0.0"]
+scipy-dep = ["scipy>=1.14.0,<2.0.0"]
+qwen-vl-utils-dep = ["qwen-vl-utils>=0.0.11,<0.1.0"]
+matplotlib-dep = ["matplotlib>=3.10.3,<4.0.0", "contourpy>=1.3.0,<2.0.0"] # NOTE: Explicitly listing contourpy helps the resolver converge faster.
 
 # Motors
 feetech = ["feetech-servo-sdk>=1.0.0,<2.0.0"]
@@ -116,30 +123,31 @@ unitree_g1 = [
     "onnxruntime>=1.16.0,<2.0.0",
     "pin>=3.0.0,<4.0.0",
     "meshcat>=0.3.0,<0.4.0",
-    "matplotlib>=3.9.0,<4.0.0",
+    "lerobot[matplotlib-dep]",
     "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,<2.55.0 ; sys_platform == 'darwin'",
+    "pyrealsense2-macosx>=2.54,<2.57.0 ; sys_platform == 'darwin'",
 ]
-phone = ["hebi-py>=2.8.0,<2.12.0", "teleop>=0.1.0,<0.2.0", "fastapi<1.0"]
+phone = ["hebi-py>=2.8.0,<2.12.0", "teleop>=0.1.0,<0.2.0", "fastapi<1.0", "lerobot[scipy-dep]"]
 
 # Policies
 wallx = [
-    "transformers==4.49.0",
-    "peft==0.17.1",
-    "scipy==1.15.3",
-    "torchdiffeq==0.2.5",
-    "qwen_vl_utils==0.0.11"
+    "lerobot[transformers-dep]",
+    "lerobot[peft]",
+    "lerobot[scipy-dep]",
+    "torchdiffeq>=0.2.4,<0.3.0",
+    "lerobot[qwen-vl-utils-dep]",
 ]
-pi = ["transformers @ git+https://github.com/huggingface/transformers.git@fix/lerobot_openpi", "scipy>=1.10.1,<1.15"]
+pi = ["lerobot[transformers-dep]", "lerobot[scipy-dep]"]
 smolvla = ["lerobot[transformers-dep]", "num2words>=0.5.14,<0.6.0", "accelerate>=1.7.0,<2.0.0", "safetensors>=0.4.3,<1.0.0"]
+multi_task_dit = ["lerobot[transformers-dep]"]
 groot = [
     "lerobot[transformers-dep]",
-    "peft>=0.13.0,<1.0.0",
+    "lerobot[peft]",
     "dm-tree>=0.1.8,<1.0.0",
     "timm>=1.0.0,<1.1.0",
     "safetensors>=0.4.3,<1.0.0",
@@ -148,13 +156,13 @@ groot = [
     "ninja>=1.11.1,<2.0.0",
     "flash-attn>=2.5.9,<3.0.0 ; sys_platform != 'darwin'"
 ]
-sarm = ["lerobot[transformers-dep]", "faker>=33.0.0,<35.0.0", "matplotlib>=3.10.3,<4.0.0", "qwen-vl-utils>=0.0.14,<0.1.0"]
+sarm = ["lerobot[transformers-dep]", "faker>=33.0.0,<35.0.0", "lerobot[matplotlib-dep]", "lerobot[qwen-vl-utils-dep]"]
 xvla = ["lerobot[transformers-dep]"]
 hilserl = ["lerobot[transformers-dep]", "gym-hil>=0.1.13,<0.2.0", "lerobot[grpcio-dep]", "lerobot[placo-dep]"]
 
 # Features
-async = ["lerobot[grpcio-dep]", "matplotlib>=3.10.3,<4.0.0"]
-peft = ["lerobot[transformers-dep]", "peft>=0.18.0,<1.0.0"]
+async = ["lerobot[grpcio-dep]", "lerobot[matplotlib-dep]"]
+peft = ["lerobot[transformers-dep]", "lerobot[peft-dep]"]
 
 # Development
 dev = ["pre-commit>=3.7.0,<5.0.0", "debugpy>=1.8.1,<1.9.0", "lerobot[grpcio-dep]", "grpcio-tools==1.73.1", "mypy>=1.19.1"]
@@ -162,13 +170,19 @@ test = ["pytest>=8.1.0,<9.0.0", "pytest-timeout>=2.4.0,<3.0.0", "pytest-cov>=5.0
 video_benchmark = ["scikit-image>=0.23.2,<0.26.0", "pandas>=2.2.2,<2.4.0"]
 
 # Simulation
-aloha = ["gym-aloha>=0.1.2,<0.2.0"]
+# NOTE: Explicitly listing scipy helps flatten the dependecy tree.
+aloha = ["gym-aloha>=0.1.2,<0.2.0", "lerobot[scipy-dep]"]
 pusht = ["gym-pusht>=0.1.5,<0.2.0", "pymunk>=6.6.0,<7.0.0"] # TODO: Fix pymunk version in gym-pusht instead
-libero = ["lerobot[transformers-dep]", "hf-libero>=0.1.3,<0.2.0"]
-metaworld = ["metaworld==3.0.0"]
+libero = ["lerobot[transformers-dep]", "hf-libero>=0.1.3,<0.2.0; sys_platform == 'linux'", "lerobot[scipy-dep]"]
+metaworld = ["metaworld==3.0.0", "lerobot[scipy-dep]"]
 
 # All
 all = [
+    # NOTE(resolver hint): scipy is pulled in transitively via lerobot[scipy-dep] through
+    # multiple extras (aloha, metaworld, pi, wallx, phone). Listing it explicitly
+    # helps pip's resolver converge by constraining scipy early, before it encounters
+    # the loose scipy requirements from transitive deps like dm-control and metaworld.
+    "scipy>=1.14.0,<2.0.0",
     "lerobot[dynamixel]",
     "lerobot[gamepad]",
     "lerobot[hopejr]",
@@ -176,8 +190,8 @@ all = [
     "lerobot[reachy2]",
     "lerobot[kinematics]",
     "lerobot[intelrealsense]",
-    # "lerobot[wallx]",
-    # "lerobot[pi]", TODO(Pepijn): Update pi to transformers v5
+    "lerobot[wallx]",
+    "lerobot[pi]",
     "lerobot[smolvla]",
     # "lerobot[groot]", TODO(Steven): Gr00t requires specific installation instructions for flash-attn
     "lerobot[xvla]",
@@ -189,7 +203,7 @@ all = [
     "lerobot[aloha]",
     "lerobot[pusht]",
     "lerobot[phone]",
-    "lerobot[libero]",
+    "lerobot[libero]; sys_platform == 'linux'",
     "lerobot[metaworld]",
     "lerobot[sarm]",
     "lerobot[peft]",
@@ -221,7 +235,7 @@ lerobot = ["envs/*.json"]
 where = ["src"]
 
 [tool.ruff]
-target-version = "py310"
+target-version = "py312"
 line-length = 110
 exclude = ["tests/artifacts/**/*.safetensors", "*_pb2.py", "*_pb2_grpc.py"]
 
@@ -313,7 +327,7 @@ default.extend-ignore-identifiers-re = [
 # Uncomment [tool.mypy] first, then uncomment individual module overrides as they get proper type annotations
 
 [tool.mypy]
-python_version = "3.10"
+python_version = "3.12"
 ignore_missing_imports = true
 follow_imports = "skip"
 # warn_return_any = true
@@ -397,85 +411,3 @@ ignore_errors = false
 # [[tool.mypy.overrides]]
 # module = "lerobot.scripts.*"
 # ignore_errors = false
-
-[tool.uv]
-# wallx requires transformers==4.49.0 which conflicts with other extras that need >=4.53.0
-conflicts = [
-    [
-        { extra = "wallx" },
-        { extra = "transformers-dep" },
-    ],
-    [
-        { extra = "wallx" },
-        { extra = "pi" },
-    ],
-    [
-        { extra = "wallx" },
-        { extra = "smolvla" },
-    ],
-    [
-        { extra = "wallx" },
-        { extra = "groot" },
-    ],
-    [
-        { extra = "wallx" },
-        { extra = "xvla" },
-    ],
-    [
-        { extra = "wallx" },
-        { extra = "sarm" },
-    ],
-    [
-        { extra = "wallx" },
-        { extra = "hilserl" },
-    ],
-    [
-        { extra = "wallx" },
-        { extra = "libero" },
-    ],
-    [
-        { extra = "wallx" },
-        { extra = "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" },
-    ],
-]

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,
@@ -485,8 +485,9 @@ class RobotClient:
 def async_client(cfg: RobotClientConfig):
     logging.info(pformat(asdict(cfg)))
 
-    if cfg.robot.type not in SUPPORTED_ROBOTS:
-        raise ValueError(f"Robot {cfg.robot.type} not yet supported!")
+    # TODO: Assert if checking robot support is still needed with the plugin system
+    # if cfg.robot.type not in SUPPORTED_ROBOTS:
+    #     raise ValueError(f"Robot {cfg.robot.type} not yet supported!")
 
     client = RobotClient(cfg)
 
@@ -512,4 +513,5 @@ def async_client(cfg: RobotClientConfig):
 
 
 if __name__ == "__main__":
+    register_third_party_plugins()
     async_client()  # run the client

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').
+    # Root directory where the dataset will be stored (e.g. 'dataset/path'). If None, defaults to $HF_LEROBOT_HOME/repo_id.
     root: str | None = None
     episodes: list[int] | None = None
     image_transforms: ImageTransformsConfig = field(default_factory=ImageTransformsConfig)

src/lerobot/datasets/aggregate.py

@@ -289,7 +289,9 @@ 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=unique_tasks)
+    dst_meta.tasks = pd.DataFrame(
+        {"task_index": range(len(unique_tasks))}, index=pd.Index(unique_tasks, name="task")
+    )
 
     meta_idx = {"chunk": 0, "file": 0}
     data_idx = {"chunk": 0, "file": 0}

src/lerobot/datasets/dataset_tools.py

@@ -47,7 +47,6 @@ 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,
@@ -90,8 +89,8 @@ def delete_episodes(
     Args:
         dataset: The source LeRobotDataset.
         episode_indices: List of episode indices to delete.
-        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.
+        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.
     """
     if not episode_indices:
         raise ValueError("No episodes to delete")
@@ -153,7 +152,7 @@ def split_dataset(
         dataset: The source LeRobotDataset to split.
         splits: Either a dict mapping split names to episode indices, or a dict mapping
                 split names to fractions (must sum to <= 1.0).
-        output_dir: Base directory for output datasets. If None, uses default location.
+        output_dir: Root directory where the split datasets will be stored. If not specified, defaults to $HF_LEROBOT_HOME/repo_id.
 
     Examples:
       Split by specific episodes
@@ -244,8 +243,8 @@ def merge_datasets(
 
     Args:
         datasets: List of LeRobotDatasets to merge.
-        output_repo_id: Repository ID for the merged dataset.
-        output_dir: Directory to save the merged dataset. If None, uses default location.
+        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.
     """
     if not datasets:
         raise ValueError("No datasets to merge")
@@ -289,8 +288,8 @@ def modify_features(
         dataset: The source LeRobotDataset.
         add_features: Optional dict mapping feature names to (feature_values, feature_info) tuples.
         remove_features: Optional feature name(s) to remove. Can be a single string or list.
-        output_dir: 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.
+        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.
 
     Returns:
         New dataset with features modified.
@@ -391,8 +390,8 @@ def add_features(
     Args:
         dataset: The source LeRobotDataset.
         features: Dictionary mapping feature names to (feature_values, feature_info) tuples.
-        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.
+        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.
 
     Returns:
         New dataset with all features added.
@@ -428,8 +427,8 @@ def remove_feature(
     Args:
         dataset: The source LeRobotDataset.
         feature_names: Name(s) of features to remove. Can be a single string or list.
-        output_dir: 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.
+        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.
 
     Returns:
         New dataset with features removed.
@@ -1476,7 +1475,9 @@ 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=unique_tasks)
+    new_task_df = pd.DataFrame(
+        {"task_index": list(range(len(unique_tasks)))}, index=pd.Index(unique_tasks, name="task")
+    )
     task_to_index = {task: idx for idx, task in enumerate(unique_tasks)}
 
     logging.info(f"Modifying tasks in {dataset.repo_id}")
@@ -1530,7 +1531,7 @@ def modify_tasks(
 
 def convert_image_to_video_dataset(
     dataset: LeRobotDataset,
-    output_dir: Path,
+    output_dir: Path | None = None,
     repo_id: str | None = None,
     vcodec: str = "libsvtav1",
     pix_fmt: str = "yuv420p",
@@ -1549,8 +1550,8 @@ def convert_image_to_video_dataset(
 
     Args:
         dataset: The source LeRobot dataset with images
-        output_dir: Directory to save the new video dataset
-        repo_id: Repository ID for the new dataset (default: original_id + "_video")
+        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.
         vcodec: Video codec (default: libsvtav1)
         pix_fmt: Pixel format (default: yuv420p)
         g: Group of pictures size (default: 2)
@@ -1601,6 +1602,7 @@ 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,
@@ -1775,296 +1777,3 @@ 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=tasks)
+            self.tasks = pd.DataFrame({"task_index": task_indices}, index=pd.Index(tasks, name="task"))
         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. 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'.
+            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).
             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
@@ -1771,11 +1771,12 @@ 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)
-            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)
+            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)
 
         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, Generic, TypeVar
+from typing import Any
 
 import datasets
 import numpy as np
@@ -78,8 +78,6 @@ 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)
@@ -341,6 +339,7 @@ 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
 
 
@@ -1233,7 +1232,7 @@ class LookAheadError(Exception):
     pass
 
 
-class Backtrackable(Generic[T]):
+class Backtrackable[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

@@ -36,8 +36,11 @@ 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}.
 ```
 
 """
@@ -105,7 +108,7 @@ episodes.jsonl
 {"episode_index": 1, "tasks": ["Put the blue block in the green bowl"], "length": 266}
 
 NEW
-meta/episodes/chunk-000/episodes_000.parquet
+meta/episodes/chunk-000/file_000.parquet
 episode_index | video_chunk_index | video_file_index | data_chunk_index | data_file_index | tasks | length
 -------------------------
 OLD
@@ -113,15 +116,16 @@ tasks.jsonl
 {"task_index": 1, "task": "Put the blue block in the green bowl"}
 
 NEW
-meta/tasks/chunk-000/file_000.parquet
+meta/tasks.parquet
 task_index | task
 -------------------------
 OLD
 episodes_stats.jsonl
+{"episode_index": 1, "stats": {"feature_name": {"min": ..., "max": ..., "mean": ..., "std": ..., "count": ...}}}
 
 NEW
-meta/episodes_stats/chunk-000/file_000.parquet
-episode_index | mean | std | min | max
+meta/episodes/chunk-000/file_000.parquet
+episode_index | feature_name/min | feature_name/max | feature_name/mean | feature_name/std | feature_name/count
 -------------------------
 UPDATE
 meta/info.json
@@ -170,7 +174,7 @@ 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=task_strings)
+    df_tasks = pd.DataFrame({"task_index": task_indices}, index=pd.Index(task_strings, name="task"))
     write_tasks(df_tasks, new_root)
 
 
@@ -201,7 +205,6 @@ 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
@@ -211,9 +214,23 @@ 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_path in tqdm.tqdm(ep_paths, desc="convert data files"):
+    for ep_idx, ep_path in enumerate(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,
@@ -224,20 +241,7 @@ 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)
-        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)
+        paths_to_cat.append(ep_path)
 
     # Write remaining data if any
     if paths_to_cat:
@@ -469,7 +473,7 @@ 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) / repo_id
+    root = HF_LEROBOT_HOME / repo_id if root is None else Path(root)
     if root.exists():
         validate_local_dataset_version(root)
         use_local_dataset = True
@@ -553,7 +557,7 @@ if __name__ == "__main__":
         "--root",
         type=str,
         default=None,
-        help="Local directory to use for downloading/writing the dataset.",
+        help="Local directory to use for downloading/writing the dataset. Defaults to $HF_LEROBOT_HOME/repo_id.",
     )
     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, TypeAlias
+from typing import Protocol
 
 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
 
-NameOrID: TypeAlias = str | int
-Value: TypeAlias = int | float
+type NameOrID = str | int
+type Value = int | float
 
 logger = logging.getLogger(__name__)
 
@@ -1277,4 +1277,4 @@ class SerialMotorsBus(MotorsBusBase):
 
 
 # Backward compatibility alias
-MotorsBus: TypeAlias = SerialMotorsBus
+MotorsBus = SerialMotorsBus

src/lerobot/policies/__init__.py

@@ -15,6 +15,7 @@
 from .act.configuration_act import ACTConfig as ACTConfig
 from .diffusion.configuration_diffusion import DiffusionConfig as DiffusionConfig
 from .groot.configuration_groot import GrootConfig as GrootConfig
+from .multi_task_dit.configuration_multi_task_dit import MultiTaskDiTConfig as MultiTaskDiTConfig
 from .pi0.configuration_pi0 import PI0Config as PI0Config
 from .pi0_fast.configuration_pi0_fast import PI0FastConfig as PI0FastConfig
 from .pi05.configuration_pi05 import PI05Config as PI05Config
@@ -28,6 +29,7 @@ from .xvla.configuration_xvla import XVLAConfig as XVLAConfig
 __all__ = [
     "ACTConfig",
     "DiffusionConfig",
+    "MultiTaskDiTConfig",
     "PI0Config",
     "PI05Config",
     "PI0FastConfig",

src/lerobot/policies/diffusion/configuration_diffusion.py

@@ -55,10 +55,16 @@ 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.
-        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).
+        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).
         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.
@@ -114,7 +120,9 @@ class DiffusionConfig(PreTrainedConfig):
     # Architecture / modeling.
     # Vision backbone.
     vision_backbone: str = "resnet18"
-    crop_shape: tuple[int, int] | None = (84, 84)
+    resize_shape: tuple[int, int] | None = None
+    crop_ratio: float = 1.0
+    crop_shape: tuple[int, int] | None = None
     crop_is_random: bool = True
     pretrained_backbone_weights: str | None = None
     use_group_norm: bool = True
@@ -175,6 +183,25 @@ 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)
@@ -202,13 +229,12 @@ 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.crop_shape is not None:
+        if self.resize_shape is None and 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 images shapes. Got {self.crop_shape} "
-                        f"for `crop_shape` and {image_ft.shape} for "
-                        f"`{key}`."
+                        f"`crop_shape` should fit within the image shapes. Got {self.crop_shape} "
+                        f"for `crop_shape` and {image_ft.shape} for `{key}`."
                     )
 
         # Check that all input images have the same shape.

src/lerobot/policies/diffusion/modeling_diffusion.py

@@ -454,12 +454,18 @@ class DiffusionRgbEncoder(nn.Module):
     def __init__(self, config: DiffusionConfig):
         super().__init__()
         # Set up optional preprocessing.
-        if config.crop_shape is not None:
+        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:
             self.do_crop = True
             # Always use center crop for eval
-            self.center_crop = torchvision.transforms.CenterCrop(config.crop_shape)
+            self.center_crop = torchvision.transforms.CenterCrop(crop_shape)
             if config.crop_is_random:
-                self.maybe_random_crop = torchvision.transforms.RandomCrop(config.crop_shape)
+                self.maybe_random_crop = torchvision.transforms.RandomCrop(crop_shape)
             else:
                 self.maybe_random_crop = self.center_crop
         else:
@@ -485,13 +491,16 @@ class DiffusionRgbEncoder(nn.Module):
 
         # Set up pooling and final layers.
         # Use a dry run to get the feature map shape.
-        # 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`.
+        # The dummy shape mirrors the runtime preprocessing order: resize -> crop.
 
         # 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
-        dummy_shape_h_w = config.crop_shape if config.crop_shape is not None else images_shape[1:]
+        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 = (1, images_shape[0], *dummy_shape_h_w)
         feature_map_shape = get_output_shape(self.backbone, dummy_shape)[1:]
 
@@ -507,7 +516,10 @@ class DiffusionRgbEncoder(nn.Module):
         Returns:
             (B, D) image feature.
         """
-        # Preprocess: maybe crop (if it was set up in the __init__).
+        # Preprocess: resize if configured, then crop if configured.
+
+        if self.resize is not None:
+            x = self.resize(x)
         if self.do_crop:
             if self.training:  # noqa: SIM108
                 x = self.maybe_random_crop(x)

src/lerobot/policies/factory.py

@@ -18,10 +18,9 @@ from __future__ import annotations
 
 import importlib
 import logging
-from typing import Any, TypedDict
+from typing import Any, TypedDict, Unpack
 
 import torch
-from typing_extensions import Unpack
 
 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.configs.types import FeatureType
@@ -32,6 +31,7 @@ from lerobot.envs.utils import env_to_policy_features
 from lerobot.policies.act.configuration_act import ACTConfig
 from lerobot.policies.diffusion.configuration_diffusion import DiffusionConfig
 from lerobot.policies.groot.configuration_groot import GrootConfig
+from lerobot.policies.multi_task_dit.configuration_multi_task_dit import MultiTaskDiTConfig
 from lerobot.policies.pi0.configuration_pi0 import PI0Config
 from lerobot.policies.pi05.configuration_pi05 import PI05Config
 from lerobot.policies.pretrained import PreTrainedPolicy
@@ -67,8 +67,7 @@ def get_policy_class(name: str) -> type[PreTrainedPolicy]:
 
     Args:
         name: The name of the policy. Supported names are "tdmpc", "diffusion", "act",
-              "vqbet", "pi0", "pi05", "sac", "reward_classifier", "smolvla", "wall_x".
-
+            "multi_task_dit", "vqbet", "pi0", "pi05", "sac", "reward_classifier", "smolvla", "wall_x".
     Returns:
         The policy class corresponding to the given name.
 
@@ -87,6 +86,10 @@ def get_policy_class(name: str) -> type[PreTrainedPolicy]:
         from lerobot.policies.act.modeling_act import ACTPolicy
 
         return ACTPolicy
+    elif name == "multi_task_dit":
+        from lerobot.policies.multi_task_dit.modeling_multi_task_dit import MultiTaskDiTPolicy
+
+        return MultiTaskDiTPolicy
     elif name == "vqbet":
         from lerobot.policies.vqbet.modeling_vqbet import VQBeTPolicy
 
@@ -147,8 +150,8 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
 
     Args:
         policy_type: The type of the policy. Supported types include "tdmpc",
-                     "diffusion", "act", "vqbet", "pi0", "pi05", "sac", "smolvla",
-                     "reward_classifier", "wall_x".
+                     "multi_task_dit", "diffusion", "act", "vqbet", "pi0", "pi05", "sac",
+                     "smolvla", "reward_classifier", "wall_x".
         **kwargs: Keyword arguments to be passed to the configuration class constructor.
 
     Returns:
@@ -163,6 +166,8 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
         return DiffusionConfig(**kwargs)
     elif policy_type == "act":
         return ACTConfig(**kwargs)
+    elif policy_type == "multi_task_dit":
+        return MultiTaskDiTConfig(**kwargs)
     elif policy_type == "vqbet":
         return VQBeTConfig(**kwargs)
     elif policy_type == "pi0":
@@ -309,6 +314,16 @@ def make_pre_post_processors(
             dataset_stats=kwargs.get("dataset_stats"),
         )
 
+    elif isinstance(policy_cfg, MultiTaskDiTConfig):
+        from lerobot.policies.multi_task_dit.processor_multi_task_dit import (
+            make_multi_task_dit_pre_post_processors,
+        )
+
+        processors = make_multi_task_dit_pre_post_processors(
+            config=policy_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+        )
+
     elif isinstance(policy_cfg, VQBeTConfig):
         from lerobot.policies.vqbet.processor_vqbet import make_vqbet_pre_post_processors
 

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

@@ -4,17 +4,16 @@
 # 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,
-    DefaultFastImageProcessorKwargs,
+    ImagesKwargs,
     group_images_by_shape,
     reorder_images,
 )
@@ -77,7 +76,7 @@ def crop(img: torch.Tensor, left: int, top: int, right: int, bottom: int) -> tor
     return img[:, top:bottom, left:right]
 
 
-class Eagle25VLFastImageProcessorKwargs(DefaultFastImageProcessorKwargs):
+class Eagle25VLFastImageProcessorKwargs(ImagesKwargs):
     max_dynamic_tiles: int | None
     min_dynamic_tiles: int | None
     use_thumbnail: bool | None
@@ -165,11 +164,11 @@ 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.
 
@@ -219,8 +218,8 @@ 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.
         """
@@ -236,15 +235,15 @@ 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
@@ -305,8 +304,8 @@ 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.
 
@@ -327,14 +326,14 @@ 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: Optional["F.InterpolationMode"],
+        interpolation: F.InterpolationMode | None,
         do_center_crop: bool,
         crop_size: SizeDict,
         do_rescale: bool,

src/lerobot/policies/multi_task_dit/README.md

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

src/lerobot/policies/multi_task_dit/__init__.py

@@ -0,0 +1,21 @@
+#!/usr/bin/env python
+
+# Copyright 2025 Bryson Jones 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 .configuration_multi_task_dit import MultiTaskDiTConfig
+from .modeling_multi_task_dit import MultiTaskDiTPolicy
+from .processor_multi_task_dit import make_multi_task_dit_pre_post_processors
+
+__all__ = ["MultiTaskDiTConfig", "MultiTaskDiTPolicy", "make_multi_task_dit_pre_post_processors"]

src/lerobot/policies/multi_task_dit/configuration_multi_task_dit.py

@@ -0,0 +1,256 @@
+#!/usr/bin/env python
+
+# Copyright 2025 Bryson Jones 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.
+
+import logging
+from dataclasses import dataclass, field
+
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import NormalizationMode
+from lerobot.optim.optimizers import AdamConfig
+from lerobot.optim.schedulers import DiffuserSchedulerConfig
+
+
+@PreTrainedConfig.register_subclass("multi_task_dit")
+@dataclass
+class MultiTaskDiTConfig(PreTrainedConfig):
+    """Configuration for the Multi-Task Diffusion Transformer (DiT) policy.
+
+    A transformer-based policy that supports both diffusion and flow matching objectives
+    for multi-task robot learning with text and vision conditioning.
+    """
+
+    n_obs_steps: int = 2  # Number of observation steps for temporal context
+    horizon: int = 32  # Number of action steps to predict
+    n_action_steps: int = 24  # Actions executed per policy call (~0.8s at 30Hz)
+
+    # Objective Selection
+    objective: str = "diffusion"  # "diffusion" or "flow_matching"
+
+    # --- Diffusion-specific (used when objective="diffusion") ---
+    noise_scheduler_type: str = "DDPM"  # "DDPM" or "DDIM"
+    num_train_timesteps: int = 100  # Number of diffusion timesteps
+    beta_schedule: str = "squaredcos_cap_v2"  # Noise schedule type
+    beta_start: float = 0.0001  # Starting noise level
+    beta_end: float = 0.02  # Ending noise level
+    prediction_type: str = "epsilon"  # "epsilon" (predict noise) or "sample" (predict clean)
+    clip_sample: bool = True  # Clip samples during denoising
+    clip_sample_range: float = 1.0  # Clipping range [-x, x]
+    num_inference_steps: int | None = None  # Denoising steps at inference (defaults to num_train_timesteps)
+
+    # --- Flow Matching-specific (used when objective="flow_matching") ---
+    sigma_min: float = 0.0  # Minimum noise in flow interpolation path
+    num_integration_steps: int = 100  # ODE integration steps at inference
+    integration_method: str = "euler"  # ODE solver: "euler" or "rk4"
+    timestep_sampling_strategy: str = "beta"  # "uniform" or "beta"
+
+    timestep_sampling_s: float = 0.999  # (beta only) Max timestep threshold
+    timestep_sampling_alpha: float = 1.5  # (beta only) Beta distribution alpha
+    timestep_sampling_beta: float = 1.0  # (beta only) Beta distribution beta
+
+    # Transformer Architecture
+    hidden_dim: int = 512  # Transformer hidden dimension
+    num_layers: int = 6  # Number of transformer layers
+    num_heads: int = 8  # Number of attention heads
+    dropout: float = 0.1  # Dropout rate
+    use_positional_encoding: bool = False  # Use absolute positional encoding
+    timestep_embed_dim: int = 256  # Timestep embedding dimension
+    use_rope: bool = True  # Use Rotary Position Embedding
+    rope_base: float = 10000.0  # RoPE base frequency
+
+    # Vision Encoder (CLIP)
+    vision_encoder_name: str = "openai/clip-vit-base-patch16"  # HuggingFace CLIP model
+    use_separate_rgb_encoder_per_camera: bool = False  # Separate encoder per camera view
+    vision_encoder_lr_multiplier: float = 0.1  # LR multiplier for vision encoder
+    image_resize_shape: tuple[int, int] | None = None  # Resize images before crop
+    image_crop_shape: tuple[int, int] | None = (224, 224)  # Crop shape (CLIP default)
+    image_crop_is_random: bool = True  # Random crop during training, center at inference
+
+    # Text Encoder (CLIP)
+    text_encoder_name: str = "openai/clip-vit-base-patch16"  # HuggingFace CLIP model
+    tokenizer_max_length: int = 77  # Max length for tokenized text (CLIP default is 77)
+    tokenizer_padding: str = "max_length"  # Padding strategy: "max_length" or "longest"
+    tokenizer_padding_side: str = "right"  # Padding side: "left" or "right"
+    tokenizer_truncation: bool = True  # Whether to truncate sequences longer than max_length
+
+    # Normalization
+    normalization_mapping: dict[str, NormalizationMode] = field(
+        default_factory=lambda: {
+            "VISUAL": NormalizationMode.MEAN_STD,
+            "STATE": NormalizationMode.MIN_MAX,
+            "ACTION": NormalizationMode.MIN_MAX,
+        }
+    )
+
+    # Training/Optimizer
+    optimizer_lr: float = 2e-5
+    optimizer_betas: tuple = (0.95, 0.999)
+    optimizer_eps: float = 1e-8
+    optimizer_weight_decay: float = 0.0
+    scheduler_name: str = "cosine"
+    scheduler_warmup_steps: int = 0
+    do_mask_loss_for_padding: bool = False
+
+    # Auto-calculated
+    drop_n_last_frames: int | None = None
+
+    def __post_init__(self):
+        super().__post_init__()
+
+        if self.drop_n_last_frames is None:
+            self.drop_n_last_frames = self.horizon - self.n_action_steps - self.n_obs_steps + 1
+
+        self._validate()
+
+    def _validate(self):
+        """Validate configuration parameters."""
+        # Objective validation
+        if self.objective not in ["diffusion", "flow_matching"]:
+            raise ValueError(f"objective must be 'diffusion' or 'flow_matching', got '{self.objective}'")
+
+        # Transformer validation
+        if self.hidden_dim <= 0:
+            raise ValueError("hidden_dim must be positive")
+        if self.num_layers <= 0:
+            raise ValueError("num_layers must be positive")
+        if self.num_heads <= 0:
+            raise ValueError("num_heads must be positive")
+        if self.hidden_dim % self.num_heads != 0:
+            raise ValueError("hidden_dim must be divisible by num_heads")
+        if not (0.0 <= self.dropout <= 1.0):
+            raise ValueError("dropout must be between 0.0 and 1.0")
+
+        # Vision encoder validation
+        if "clip" not in self.vision_encoder_name.lower():
+            raise ValueError(
+                f"vision_encoder_name must be a CLIP model (contain 'clip'), got '{self.vision_encoder_name}'"
+            )
+        if (
+            self.image_resize_shape
+            and self.image_crop_shape
+            and (
+                self.image_crop_shape[0] > self.image_resize_shape[0]
+                or self.image_crop_shape[1] > self.image_resize_shape[1]
+            )
+        ):
+            logging.warning(
+                "image_crop_shape %s must be <= image_resize_shape %s; disabling cropping.",
+                self.image_crop_shape,
+                self.image_resize_shape,
+            )
+            self.image_crop_shape = None
+
+        # Text encoder validation
+        if "clip" not in self.text_encoder_name.lower():
+            raise ValueError(
+                f"text_encoder_name must be a CLIP model (contain 'clip'), got '{self.text_encoder_name}'"
+            )
+
+        # Objective-specific validation
+        if self.objective == "diffusion":
+            if self.noise_scheduler_type not in ["DDPM", "DDIM"]:
+                raise ValueError(
+                    f"noise_scheduler_type must be 'DDPM' or 'DDIM', got {self.noise_scheduler_type}"
+                )
+            if self.prediction_type not in ["epsilon", "sample"]:
+                raise ValueError(f"prediction_type must be 'epsilon' or 'sample', got {self.prediction_type}")
+            if self.num_train_timesteps <= 0:
+                raise ValueError(f"num_train_timesteps must be positive, got {self.num_train_timesteps}")
+            if not (0.0 <= self.beta_start <= self.beta_end <= 1.0):
+                raise ValueError(f"Invalid beta values: {self.beta_start}, {self.beta_end}")
+
+        elif self.objective == "flow_matching":
+            if not (0.0 <= self.sigma_min <= 1.0):
+                raise ValueError(f"sigma_min must be in [0, 1], got {self.sigma_min}")
+            if self.num_integration_steps <= 0:
+                raise ValueError(f"num_integration_steps must be positive, got {self.num_integration_steps}")
+            if self.integration_method not in ["euler", "rk4"]:
+                raise ValueError(
+                    f"integration_method must be 'euler' or 'rk4', got {self.integration_method}"
+                )
+            if self.timestep_sampling_strategy not in ["uniform", "beta"]:
+                raise ValueError("timestep_sampling_strategy must be 'uniform' or 'beta'")
+            if self.timestep_sampling_strategy == "beta":
+                if not (0.0 < self.timestep_sampling_s <= 1.0):
+                    raise ValueError(f"timestep_sampling_s must be in (0, 1], got {self.timestep_sampling_s}")
+                if self.timestep_sampling_alpha <= 0:
+                    raise ValueError("timestep_sampling_alpha must be positive")
+                if self.timestep_sampling_beta <= 0:
+                    raise ValueError("timestep_sampling_beta must be positive")
+
+    def get_optimizer_preset(self) -> AdamConfig:
+        return AdamConfig(
+            lr=self.optimizer_lr,
+            betas=self.optimizer_betas,
+            eps=self.optimizer_eps,
+            weight_decay=self.optimizer_weight_decay,
+        )
+
+    def get_scheduler_preset(self) -> DiffuserSchedulerConfig:
+        return DiffuserSchedulerConfig(
+            name=self.scheduler_name,
+            num_warmup_steps=self.scheduler_warmup_steps,
+        )
+
+    def validate_features(self) -> None:
+        """Validate that required input features are present and properly configured."""
+        # If the configured crop doesn't fit, disable cropping instead of erroring.
+        # Note: if image_resize_shape is set, cropping is applied *after* resizing.
+        if self.image_crop_shape is not None:
+            for key, image_ft in self.image_features.items():
+                # image_ft.shape is (C, H, W)
+                effective_h, effective_w = (
+                    self.image_resize_shape
+                    if self.image_resize_shape is not None
+                    else (image_ft.shape[1], image_ft.shape[2])
+                )
+                if self.image_crop_shape[0] > effective_h or self.image_crop_shape[1] > effective_w:
+                    logging.warning(
+                        "image_crop_shape %s doesn't fit within effective image shape (%s, %s) for '%s'; disabling cropping.",
+                        self.image_crop_shape,
+                        effective_h,
+                        effective_w,
+                        key,
+                    )
+                    self.image_crop_shape = None
+                    break
+
+        if len(self.image_features) > 0:
+            first_key, first_ft = next(iter(self.image_features.items()))
+            for key, image_ft in self.image_features.items():
+                if image_ft.shape != first_ft.shape:
+                    raise ValueError(
+                        f"Image '{key}' shape {image_ft.shape} != '{first_key}' shape {first_ft.shape}"
+                    )
+
+    @property
+    def is_diffusion(self) -> bool:
+        return self.objective == "diffusion"
+
+    @property
+    def is_flow_matching(self) -> bool:
+        return self.objective == "flow_matching"
+
+    @property
+    def observation_delta_indices(self) -> list:
+        return list(range(1 - self.n_obs_steps, 1))
+
+    @property
+    def action_delta_indices(self) -> list:
+        return list(range(1 - self.n_obs_steps, 1 - self.n_obs_steps + self.horizon))
+
+    @property
+    def reward_delta_indices(self) -> None:
+        return None

src/lerobot/policies/multi_task_dit/modeling_multi_task_dit.py

@@ -0,0 +1,803 @@
+#!/usr/bin/env python
+
+# Copyright 2025 Bryson Jones 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.
+
+"""Multi-Task Diffusion Transformer (DiT) Policy
+
+Transformer-based diffusion policy for multi-task robot learning with text and vision conditioning.
+Supports both diffusion and flow matching objectives for action generation.
+
+References:
+- https://arxiv.org/abs/2507.05331
+- https://bostondynamics.com/blog/large-behavior-models-atlas-find-new-footing/
+- https://brysonkjones.substack.com/p/dissecting-and-open-sourcing-multitask-diffusion-transformer-policy
+"""
+
+import math
+from collections import deque
+from typing import TYPE_CHECKING
+
+import einops
+import torch
+import torch.nn as nn
+import torch.nn.functional as F  # noqa: N812
+import torchvision
+from diffusers.schedulers.scheduling_ddim import DDIMScheduler
+from diffusers.schedulers.scheduling_ddpm import DDPMScheduler
+from torch import Tensor
+
+from lerobot.policies.multi_task_dit.configuration_multi_task_dit import MultiTaskDiTConfig
+from lerobot.utils.import_utils import _transformers_available
+
+# Conditional import for type checking and lazy loading
+if TYPE_CHECKING or _transformers_available:
+    from transformers import CLIPTextModel, CLIPVisionModel
+else:
+    CLIPTextModel = None
+    CLIPVisionModel = None
+from lerobot.policies.pretrained import PreTrainedPolicy
+from lerobot.policies.utils import populate_queues
+from lerobot.utils.constants import (
+    ACTION,
+    OBS_IMAGES,
+    OBS_LANGUAGE_ATTENTION_MASK,
+    OBS_LANGUAGE_TOKENS,
+    OBS_STATE,
+)
+
+# -- Policy --
+
+
+class MultiTaskDiTPolicy(PreTrainedPolicy):
+    config_class = MultiTaskDiTConfig
+    name = "multi_task_dit"
+
+    def __init__(self, config: MultiTaskDiTConfig, **kwargs):
+        super().__init__(config)
+        config.validate_features()
+        self.config = config
+
+        self._queues = None
+
+        self.observation_encoder = ObservationEncoder(config)
+        conditioning_dim = self.observation_encoder.conditioning_dim
+        self.noise_predictor = DiffusionTransformer(config, conditioning_dim=conditioning_dim)
+
+        action_dim = config.action_feature.shape[0]
+        horizon = config.horizon
+
+        if config.is_diffusion:
+            self.objective = DiffusionObjective(
+                config,
+                action_dim=action_dim,
+                horizon=horizon,
+                do_mask_loss_for_padding=config.do_mask_loss_for_padding,
+            )
+        elif config.is_flow_matching:
+            self.objective = FlowMatchingObjective(
+                config,
+                action_dim=action_dim,
+                horizon=horizon,
+                do_mask_loss_for_padding=config.do_mask_loss_for_padding,
+            )
+        else:
+            raise ValueError(f"Unsupported objective: {config.objective}")
+
+        self.reset()
+
+    def get_optim_params(self) -> list:
+        """Returns parameter groups with different learning rates for vision vs non-vision parameters"""
+        non_vision_params = []
+        vision_encoder_params = []
+
+        for name, param in self.named_parameters():
+            if not param.requires_grad:
+                continue
+
+            if "observation_encoder.vision_encoder" in name:
+                vision_encoder_params.append(param)
+            else:
+                non_vision_params.append(param)
+
+        return [
+            {"params": non_vision_params},
+            {
+                "params": vision_encoder_params,
+                "lr": self.config.optimizer_lr * self.config.vision_encoder_lr_multiplier,
+            },
+        ]
+
+    def _generate_actions(self, batch: dict[str, Tensor]) -> Tensor:
+        batch_size, n_obs_steps = batch[OBS_STATE].shape[:2]
+        assert n_obs_steps == self.config.n_obs_steps
+
+        conditioning_vec = self.observation_encoder.encode(batch)
+        actions = self.objective.conditional_sample(self.noise_predictor, batch_size, conditioning_vec)
+
+        start = n_obs_steps - 1
+        end = start + self.config.n_action_steps
+        actions = actions[:, start:end]
+        return actions
+
+    def reset(self):
+        """Clear observation and action queues. Should be called on `env.reset()`"""
+        self._queues = {
+            OBS_STATE: deque(maxlen=self.config.n_obs_steps),
+            ACTION: deque(maxlen=self.config.n_action_steps),
+        }
+
+        if self.config.image_features:
+            self._queues[OBS_IMAGES] = deque(maxlen=self.config.n_obs_steps)
+
+    @torch.no_grad()
+    def predict_action_chunk(self, batch: dict[str, Tensor]) -> Tensor:
+        """Predict a chunk of actions given environment observations"""
+        self.eval()
+
+        for k in batch:
+            if k in self._queues:
+                batch[k] = torch.stack(list(self._queues[k]), dim=1)
+
+        actions = self._generate_actions(batch)
+        return actions
+
+    def _prepare_batch(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
+        """Prepare batch by stacking image features if needed."""
+        if self.config.image_features:
+            batch = dict(batch)  # shallow copy to avoid modifying original
+            batch[OBS_IMAGES] = torch.stack([batch[key] for key in self.config.image_features], dim=-4)
+
+        return batch
+
+    @torch.no_grad()
+    def select_action(self, batch: dict[str, Tensor]) -> Tensor:
+        """Select a single action given environment observations"""
+        if ACTION in batch:
+            batch = dict(batch)  # shallow copy to avoid modifying original
+            batch.pop(ACTION)
+
+        batch = self._prepare_batch(batch)
+
+        self._queues = populate_queues(self._queues, batch)
+
+        if len(self._queues[ACTION]) == 0:
+            actions = self.predict_action_chunk(batch)
+            self._queues[ACTION].extend(actions.transpose(0, 1))
+
+        action = self._queues[ACTION].popleft()
+        return action
+
+    def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict | None]:
+        """Run the batch through the model and compute the loss for training"""
+        batch = self._prepare_batch(batch)
+
+        conditioning_vec = self.observation_encoder.encode(batch)
+        loss = self.objective.compute_loss(self.noise_predictor, batch, conditioning_vec)
+
+        return loss, None
+
+
+# -- Observation Encoders --
+
+
+class CLIPVisionEncoder(nn.Module):
+    """CLIP vision encoder using the CLS token for global image representation."""
+
+    def __init__(self, model_name: str):
+        super().__init__()
+        self.model_name = model_name
+        self.model = CLIPVisionModel.from_pretrained(self.model_name)
+        self.num_non_spatial_tokens = 1
+        self.embed_dim = self.model.config.hidden_size
+
+    def forward(self, x: Tensor) -> Tensor:
+        """Encode RGB image to CLS token."""
+        outputs = self.model(pixel_values=x, output_hidden_states=False)
+        cls_token = outputs.last_hidden_state[:, 0]
+        b, embed_dim = cls_token.shape
+        return cls_token.reshape(b, embed_dim, 1, 1)
+
+    def get_output_shape(self) -> tuple:
+        return (self.embed_dim, 1, 1)
+
+
+class CLIPTextEncoder(nn.Module):
+    """CLIP text encoder with frozen weights and a learnable projection layer.
+
+    Accepts pre-tokenized inputs (input_ids and attention_mask) from the processor pipeline. See the processor
+    pipeline to see how the tokenization is handled.
+    """
+
+    def __init__(self, model_name: str = "openai/clip-vit-base-patch16", projection_dim: int = 512):
+        super().__init__()
+        self.model_name = model_name
+        self.projection_dim = projection_dim
+        self.text_encoder = CLIPTextModel.from_pretrained(model_name)
+
+        for param in self.text_encoder.parameters():
+            param.requires_grad = False
+
+        self.text_embed_dim = self.text_encoder.config.hidden_size
+        self.projection = nn.Linear(self.text_embed_dim, projection_dim)
+
+    def forward(self, input_ids: Tensor, attention_mask: Tensor) -> Tensor:
+        """Encode pre-tokenized text to feature vectors."""
+        # Ensure inputs are on the same device as the model
+        device = next(self.parameters()).device
+        input_ids = input_ids.to(device)
+        attention_mask = attention_mask.to(device)
+
+        with torch.no_grad():
+            outputs = self.text_encoder(input_ids=input_ids, attention_mask=attention_mask)
+            clip_features = outputs.pooler_output
+
+        return self.projection(clip_features)
+
+
+class ObservationEncoder(nn.Module):
+    """Handles all observation processing for the conditioning vector."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self._setup_preprocessing(config)
+
+        if config.image_features:
+            self.num_cameras = len(config.image_features)
+            self.camera_names = list(config.image_features.keys())
+
+            if config.use_separate_rgb_encoder_per_camera:
+                self.vision_encoders = nn.ModuleList(
+                    [CLIPVisionEncoder(model_name=config.vision_encoder_name) for _ in self.camera_names]
+                )
+                self.vision_encoder = None
+            else:
+                self.vision_encoder = CLIPVisionEncoder(model_name=config.vision_encoder_name)
+                self.vision_encoders = None
+        else:
+            self.vision_encoder = None
+            self.vision_encoders = None
+            self.camera_names = []
+            self.num_cameras = 0
+
+        if hasattr(config, "robot_state_feature") and config.robot_state_feature:
+            self.robot_state_dim = config.robot_state_feature.shape[0]
+        else:
+            self.robot_state_dim = 0
+
+        self.text_dim = config.hidden_dim
+        self.text_encoder = CLIPTextEncoder(model_name=config.text_encoder_name, projection_dim=self.text_dim)
+
+        self._setup_vector_output()
+
+    def _apply_preprocessing(self, images: Tensor) -> Tensor:
+        if self.do_resize:
+            images = self.resize(images)
+        if self.do_crop:
+            images = self.maybe_random_crop(images) if self.training else self.center_crop(images)
+        return images
+
+    def _setup_preprocessing(self, config):
+        if config.image_resize_shape is not None:
+            self.do_resize = True
+            self.resize = torchvision.transforms.Resize(
+                size=config.image_resize_shape,
+                interpolation=torchvision.transforms.InterpolationMode.BILINEAR,
+                antialias=True,
+            )
+        else:
+            self.do_resize = False
+
+        if config.image_crop_shape is not None:
+            self.do_crop = True
+            self.center_crop = torchvision.transforms.CenterCrop(config.image_crop_shape)
+            if config.image_crop_is_random:
+                self.maybe_random_crop = torchvision.transforms.RandomCrop(config.image_crop_shape)
+            else:
+                self.maybe_random_crop = self.center_crop
+        else:
+            self.do_crop = False
+
+    def _setup_vector_output(self):
+        total_dim = 0
+
+        if self.vision_encoder is not None or self.vision_encoders is not None:
+            encoder_to_check = self.vision_encoder or next(iter(self.vision_encoders))
+            feature_map_shape = encoder_to_check.get_output_shape()
+            c, h, w = feature_map_shape
+            spatial_feature_dim = c * h * w
+            total_dim += spatial_feature_dim * self.num_cameras
+
+        total_dim += self.robot_state_dim
+        total_dim += self.text_dim
+
+        self.conditioning_dim = total_dim * self.config.n_obs_steps
+
+    def encode(self, batch: dict) -> Tensor:
+        """Encode observations to vector format."""
+        batch_size, n_obs_steps = batch[OBS_STATE].shape[:2]
+        conditioning_feats = []
+
+        conditioning_feats.append(batch[OBS_STATE])
+
+        if self.vision_encoder is not None or self.vision_encoders is not None:
+            images = batch[OBS_IMAGES]
+
+            if len(images.shape) == 5:
+                images = images.unsqueeze(1)
+
+            if self.config.use_separate_rgb_encoder_per_camera:
+                camera_features = []
+                for cam_idx in range(self.num_cameras):
+                    cam_images = images[:, :, cam_idx]
+                    cam_images_flat = einops.rearrange(cam_images, "b s c h w -> (b s) c h w")
+                    cam_images_flat = self._apply_preprocessing(cam_images_flat)
+                    cam_features = self.vision_encoders[cam_idx](cam_images_flat)
+                    cam_visual_features = cam_features.flatten(start_dim=1)
+                    cam_features_reshaped = einops.rearrange(
+                        cam_visual_features, "(b s) f -> b s f", b=batch_size, s=n_obs_steps
+                    )
+                    camera_features.append(cam_features_reshaped)
+                img_features = torch.cat(camera_features, dim=-1)
+                conditioning_feats.append(img_features)
+            else:
+                images_flat = einops.rearrange(images, "b s n c h w -> (b s n) c h w")
+                images_flat = self._apply_preprocessing(images_flat)
+                visual_features = self.vision_encoder(images_flat).flatten(start_dim=1)
+                img_features = einops.rearrange(
+                    visual_features, "(b s n) f -> b s (n f)", b=batch_size, s=n_obs_steps, n=self.num_cameras
+                )
+                conditioning_feats.append(img_features)
+
+        if self.text_encoder is not None and OBS_LANGUAGE_TOKENS in batch:
+            input_ids = batch[OBS_LANGUAGE_TOKENS]  # [batch_size, seq_length]
+            attention_mask = batch[OBS_LANGUAGE_ATTENTION_MASK]  # [batch_size, seq_length]
+
+            text_features = self.text_encoder(input_ids, attention_mask)
+
+            text_features = text_features.unsqueeze(1).expand(-1, n_obs_steps, -1)
+            conditioning_feats.append(text_features)
+
+        combined_features = torch.cat(conditioning_feats, dim=-1)
+        return combined_features.flatten(start_dim=1)
+
+
+# -- Transformer Components --
+
+
+def modulate(x: Tensor, shift: Tensor, scale: Tensor) -> Tensor:
+    """Modulate input with shift and scale for AdaLN-Zero."""
+    return x * (1 + scale) + shift
+
+
+class SinusoidalPosEmb(nn.Module):
+    """Sinusoidal positional embeddings for timesteps."""
+
+    def __init__(self, dim: int):
+        super().__init__()
+        self.dim = dim
+
+    def forward(self, x: Tensor) -> Tensor:
+        device = x.device
+        half_dim = self.dim // 2
+        emb = math.log(10000) / (half_dim - 1)
+        emb = torch.exp(torch.arange(half_dim, device=device) * -emb)
+        emb = x[:, None] * emb[None, :]
+        emb = torch.cat((emb.sin(), emb.cos()), dim=-1)
+        return emb
+
+
+class RotaryPositionalEmbedding(nn.Module):
+    """Rotary Position Embedding (RoPE) for transformers."""
+
+    def __init__(self, head_dim: int, max_seq_len: int = 512, base: float = 10000.0):
+        super().__init__()
+        assert head_dim % 2 == 0, "head_dim must be even for RoPE"
+
+        self.head_dim = head_dim
+        self.max_seq_len = max_seq_len
+        self.base = base
+
+        inv_freq = 1.0 / (base ** (torch.arange(0, head_dim, 2).float() / head_dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self._precompute_cache(max_seq_len)
+
+    def _precompute_cache(self, seq_len: int):
+        t = torch.arange(seq_len, dtype=self.inv_freq.dtype)
+        freqs = torch.outer(t, self.inv_freq)
+        emb = torch.cat((freqs, freqs), dim=-1)
+        self.register_buffer("_cos_cached", emb.cos()[None, None, :, :], persistent=False)
+        self.register_buffer("_sin_cached", emb.sin()[None, None, :, :], persistent=False)
+
+    def _rotate_half(self, x: Tensor) -> Tensor:
+        x1 = x[..., : x.shape[-1] // 2]
+        x2 = x[..., x.shape[-1] // 2 :]
+        return torch.cat((-x2, x1), dim=-1)
+
+    def forward(self, q: Tensor, k: Tensor) -> tuple[Tensor, Tensor]:
+        seq_len = q.shape[2]
+        if seq_len > self.max_seq_len:
+            raise ValueError(f"Sequence length {seq_len} exceeds max_seq_len {self.max_seq_len}.")
+
+        cos = self._cos_cached[:, :, :seq_len, :].to(q.dtype)
+        sin = self._sin_cached[:, :, :seq_len, :].to(q.dtype)
+
+        q_rotated = (q * cos) + (self._rotate_half(q) * sin)
+        k_rotated = (k * cos) + (self._rotate_half(k) * sin)
+        return q_rotated, k_rotated
+
+
+class RoPEAttention(nn.Module):
+    """Multi-head self-attention with Rotary Position Embedding (RoPE)."""
+
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        dropout: float = 0.0,
+        max_seq_len: int = 512,
+        rope_base: float = 10000.0,
+    ):
+        super().__init__()
+        assert hidden_size % num_heads == 0, "hidden_size must be divisible by num_heads"
+
+        self.hidden_size = hidden_size
+        self.num_heads = num_heads
+        self.head_dim = hidden_size // num_heads
+        self.scale = self.head_dim**-0.5
+
+        self.qkv_proj = nn.Linear(hidden_size, 3 * hidden_size, bias=True)
+        self.out_proj = nn.Linear(hidden_size, hidden_size, bias=True)
+        self.dropout = nn.Dropout(dropout) if dropout > 0 else nn.Identity()
+        self.rope = RotaryPositionalEmbedding(head_dim=self.head_dim, max_seq_len=max_seq_len, base=rope_base)
+
+    def forward(self, x: Tensor) -> Tensor:
+        B, T, _ = x.shape  # noqa: N806
+
+        qkv = self.qkv_proj(x)
+        qkv = qkv.reshape(B, T, 3, self.num_heads, self.head_dim)
+        qkv = qkv.permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]
+
+        q, k = self.rope(q, k)
+
+        attn_out = torch.nn.functional.scaled_dot_product_attention(
+            q,
+            k,
+            v,
+            dropout_p=self.dropout.p if isinstance(self.dropout, nn.Dropout) and self.training else 0.0,
+        )
+
+        attn_out = attn_out.transpose(1, 2).reshape(B, T, self.hidden_size)
+        return self.out_proj(attn_out)
+
+
+class TransformerBlock(nn.Module):
+    """DiT-style transformer block with AdaLN-Zero."""
+
+    def __init__(
+        self,
+        hidden_size: int = 128,
+        num_heads: int = 4,
+        num_features: int = 128,
+        dropout: float = 0.0,
+        use_rope: bool = False,
+        max_seq_len: int = 512,
+        rope_base: float = 10000.0,
+    ):
+        super().__init__()
+        self.use_rope = use_rope
+
+        if use_rope:
+            self.attn = RoPEAttention(
+                hidden_size=hidden_size,
+                num_heads=num_heads,
+                dropout=dropout,
+                max_seq_len=max_seq_len,
+                rope_base=rope_base,
+            )
+        else:
+            self.multihead_attn = nn.MultiheadAttention(
+                hidden_size, num_heads=num_heads, batch_first=True, dropout=dropout
+            )
+
+        self.norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+
+        self.mlp = nn.Sequential(
+            nn.Linear(hidden_size, hidden_size * 4),
+            nn.GELU(approximate="tanh"),
+            nn.Linear(hidden_size * 4, hidden_size),
+        )
+
+        self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(num_features, 6 * hidden_size, bias=True))
+
+    def forward(self, x: Tensor, features: Tensor) -> Tensor:
+        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(
+            features
+        ).chunk(6, dim=1)
+
+        attn_input = modulate(self.norm1(x), shift_msa.unsqueeze(1), scale_msa.unsqueeze(1))
+
+        if self.use_rope:
+            attn_out = self.attn(attn_input)
+        else:
+            attn_out, _ = self.multihead_attn(attn_input, attn_input, attn_input)
+
+        x = x + gate_msa.unsqueeze(1) * attn_out
+
+        mlp_input = modulate(self.norm2(x), shift_mlp.unsqueeze(1), scale_mlp.unsqueeze(1))
+        mlp_out = self.mlp(mlp_input)
+        x = x + gate_mlp.unsqueeze(1) * mlp_out
+
+        return x
+
+
+class DiffusionTransformer(nn.Module):
+    """Transformer-based diffusion noise prediction model."""
+
+    def __init__(self, config, conditioning_dim: int):
+        super().__init__()
+        self.config = config
+        self.conditioning_dim = conditioning_dim
+
+        self.action_dim = config.action_feature.shape[0]
+        self.horizon = config.horizon
+        self.hidden_size = config.hidden_dim
+        self.num_layers = config.num_layers
+        self.num_heads = config.num_heads
+        self.dropout = config.dropout
+        self.use_rope = config.use_rope
+
+        self.timestep_embed_dim = config.timestep_embed_dim
+        self.time_mlp = nn.Sequential(
+            SinusoidalPosEmb(self.timestep_embed_dim),
+            nn.Linear(self.timestep_embed_dim, 2 * self.timestep_embed_dim),
+            nn.GELU(),
+            nn.Linear(2 * self.timestep_embed_dim, self.timestep_embed_dim),
+            nn.GELU(),
+        )
+
+        self.cond_dim = self.timestep_embed_dim + conditioning_dim
+        self.input_proj = nn.Linear(self.action_dim, self.hidden_size)
+
+        if config.use_positional_encoding:
+            self.pos_embedding = nn.Parameter(
+                torch.empty(1, self.horizon, self.hidden_size).normal_(std=0.02)
+            )
+        else:
+            self.pos_embedding = None
+
+        self.transformer_blocks = nn.ModuleList(
+            [
+                TransformerBlock(
+                    hidden_size=self.hidden_size,
+                    num_heads=self.num_heads,
+                    num_features=self.cond_dim,
+                    dropout=self.dropout,
+                    use_rope=self.use_rope,
+                    max_seq_len=self.horizon,
+                    rope_base=config.rope_base,
+                )
+                for _ in range(self.num_layers)
+            ]
+        )
+
+        self.output_proj = nn.Linear(self.hidden_size, self.action_dim)
+        self._initialize_weights()
+
+    def _initialize_weights(self):
+        for block in self.transformer_blocks:
+            nn.init.constant_(block.adaLN_modulation[-1].weight, 0)
+            nn.init.constant_(block.adaLN_modulation[-1].bias, 0)
+
+    def forward(self, x: Tensor, timestep: Tensor, conditioning_vec: Tensor) -> Tensor:
+        _, seq_len, _ = x.shape
+
+        timestep_features = self.time_mlp(timestep)
+        cond_features = torch.cat([timestep_features, conditioning_vec], dim=-1)
+
+        hidden_seq = self.input_proj(x)
+
+        if self.pos_embedding is not None:
+            hidden_seq = hidden_seq + self.pos_embedding[:, :seq_len, :]
+
+        for block in self.transformer_blocks:
+            hidden_seq = block(hidden_seq, cond_features)
+
+        return self.output_proj(hidden_seq)
+
+
+# -- Objectives --
+
+
+class DiffusionObjective(nn.Module):
+    """Standard diffusion (DDPM/DDIM) objective implementation."""
+
+    def __init__(self, config, action_dim: int, horizon: int, do_mask_loss_for_padding: bool = False):
+        super().__init__()
+        self.config = config
+        self.action_dim = action_dim
+        self.horizon = horizon
+        self.do_mask_loss_for_padding = do_mask_loss_for_padding
+
+        scheduler_kwargs = {
+            "num_train_timesteps": config.num_train_timesteps,
+            "beta_start": config.beta_start,
+            "beta_end": config.beta_end,
+            "beta_schedule": config.beta_schedule,
+            "clip_sample": config.clip_sample,
+            "clip_sample_range": config.clip_sample_range,
+            "prediction_type": config.prediction_type,
+        }
+
+        if config.noise_scheduler_type == "DDPM":
+            self.noise_scheduler: DDPMScheduler | DDIMScheduler = DDPMScheduler(**scheduler_kwargs)
+        elif config.noise_scheduler_type == "DDIM":
+            self.noise_scheduler = DDIMScheduler(**scheduler_kwargs)
+        else:
+            raise ValueError(f"Unsupported noise scheduler type {config.noise_scheduler_type}")
+
+        self.num_inference_steps = (
+            config.num_inference_steps
+            if config.num_inference_steps is not None
+            else self.noise_scheduler.config.num_train_timesteps
+        )
+
+    def compute_loss(self, model: nn.Module, batch: dict[str, Tensor], conditioning_vec: Tensor) -> Tensor:
+        clean_actions = batch[ACTION]
+        noise = torch.randn_like(clean_actions)
+        timesteps = torch.randint(
+            low=0,
+            high=self.noise_scheduler.config.num_train_timesteps,
+            size=(clean_actions.shape[0],),
+            device=clean_actions.device,
+        ).long()
+        noisy_actions = self.noise_scheduler.add_noise(clean_actions, noise, timesteps)
+
+        prediction_type = self.noise_scheduler.config.prediction_type
+        if prediction_type == "epsilon":
+            target = noise
+        elif prediction_type == "sample":
+            target = clean_actions
+        else:
+            raise ValueError(f"Unsupported prediction type: {prediction_type}")
+
+        predicted = model(noisy_actions, timesteps, conditioning_vec=conditioning_vec)
+        loss = F.mse_loss(predicted, target, reduction="none")
+
+        if self.do_mask_loss_for_padding and "action_is_pad" in batch:
+            valid_actions = ~batch["action_is_pad"]
+            loss = loss * valid_actions.unsqueeze(-1)
+
+        return loss.mean()
+
+    def conditional_sample(self, model: nn.Module, batch_size: int, conditioning_vec: Tensor) -> Tensor:
+        device = next(model.parameters()).device
+        dtype = next(model.parameters()).dtype
+
+        sample = torch.randn(
+            size=(batch_size, self.horizon, self.action_dim),
+            dtype=dtype,
+            device=device,
+        )
+
+        self.noise_scheduler.set_timesteps(self.num_inference_steps)
+        for t in self.noise_scheduler.timesteps:
+            model_output = model(
+                sample,
+                torch.full(sample.shape[:1], t, dtype=torch.long, device=sample.device),
+                conditioning_vec=conditioning_vec,
+            )
+            sample = self.noise_scheduler.step(model_output, t, sample).prev_sample
+
+        return sample
+
+
+class FlowMatchingObjective(nn.Module):
+    """Flow matching objective: trains a model to predict velocity fields."""
+
+    def __init__(self, config, action_dim: int, horizon: int, do_mask_loss_for_padding: bool = False):
+        super().__init__()
+        self.config = config
+        self.action_dim = action_dim
+        self.horizon = horizon
+        self.do_mask_loss_for_padding = do_mask_loss_for_padding
+
+    def _sample_timesteps(self, batch_size: int, device: torch.device) -> Tensor:
+        if self.config.timestep_sampling_strategy == "uniform":
+            return torch.rand(batch_size, device=device)
+        elif self.config.timestep_sampling_strategy == "beta":
+            beta_dist = torch.distributions.Beta(
+                self.config.timestep_sampling_alpha, self.config.timestep_sampling_beta
+            )
+            u = beta_dist.sample((batch_size,)).to(device)
+            return self.config.timestep_sampling_s * (1.0 - u)
+        else:
+            raise ValueError(f"Unknown timestep strategy: {self.config.timestep_sampling_strategy}")
+
+    def compute_loss(self, model: nn.Module, batch: dict[str, Tensor], conditioning_vec: Tensor) -> Tensor:
+        data = batch[ACTION]
+        batch_size = data.shape[0]
+        device = data.device
+
+        noise = torch.randn_like(data)
+        t = self._sample_timesteps(batch_size, device)
+        t_expanded = t.view(-1, 1, 1)
+        x_t = t_expanded * data + (1 - (1 - self.config.sigma_min) * t_expanded) * noise
+
+        target_velocity = data - (1 - self.config.sigma_min) * noise
+        predicted_velocity = model(x_t, t, conditioning_vec=conditioning_vec)
+        loss = F.mse_loss(predicted_velocity, target_velocity, reduction="none")
+
+        if self.do_mask_loss_for_padding and "action_is_pad" in batch:
+            valid_mask = ~batch["action_is_pad"]
+            loss = loss * valid_mask.unsqueeze(-1)
+
+        return loss.mean()
+
+    def conditional_sample(self, model: nn.Module, batch_size: int, conditioning_vec: Tensor) -> Tensor:
+        device = next(model.parameters()).device
+        dtype = next(model.parameters()).dtype
+
+        x = torch.randn((batch_size, self.horizon, self.action_dim), dtype=dtype, device=device)
+
+        num_steps = self.config.num_integration_steps
+        time_grid = torch.linspace(0, 1, num_steps + 1, device=device)
+
+        if self.config.integration_method == "euler":
+            x = self._euler_integrate(model, x, time_grid, conditioning_vec)
+        elif self.config.integration_method == "rk4":
+            x = self._rk4_integrate(model, x, time_grid, conditioning_vec)
+        else:
+            raise ValueError(f"Unknown integration method: {self.config.integration_method}")
+
+        return x
+
+    def _euler_integrate(
+        self, model: nn.Module, x_init: Tensor, time_grid: Tensor, conditioning_vec: Tensor
+    ) -> Tensor:
+        x = x_init
+        for i in range(len(time_grid) - 1):
+            t_scalar = time_grid[i].item()
+            dt = (time_grid[i + 1] - time_grid[i]).item()
+            t_batch = torch.full((x.shape[0],), t_scalar, dtype=x.dtype, device=x.device)
+            with torch.no_grad():
+                velocity = model(x, t_batch, conditioning_vec=conditioning_vec)
+            x = x + dt * velocity
+        return x
+
+    def _rk4_integrate(
+        self, model: nn.Module, x_init: Tensor, time_grid: Tensor, conditioning_vec: Tensor
+    ) -> Tensor:
+        x = x_init
+
+        def dynamics(x_val: Tensor, t_scalar: float) -> Tensor:
+            t_batch = torch.full((x_val.shape[0],), t_scalar, dtype=x_val.dtype, device=x_val.device)
+            with torch.no_grad():
+                return model(x_val, t_batch, conditioning_vec=conditioning_vec)
+
+        for i in range(len(time_grid) - 1):
+            t = time_grid[i].item()
+            dt = (time_grid[i + 1] - time_grid[i]).item()
+
+            k1 = dynamics(x, t)
+            k2 = dynamics(x + dt * k1 / 2, t + dt / 2)
+            k3 = dynamics(x + dt * k2 / 2, t + dt / 2)
+            k4 = dynamics(x + dt * k3, t + dt)
+
+            x = x + dt / 6 * (k1 + 2 * k2 + 2 * k3 + k4)
+
+        return x

src/lerobot/policies/multi_task_dit/processor_multi_task_dit.py

@@ -0,0 +1,105 @@
+#!/usr/bin/env python
+
+# Copyright 2025 Bryson Jones 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 typing import Any
+
+import torch
+
+from lerobot.policies.multi_task_dit.configuration_multi_task_dit import MultiTaskDiTConfig
+from lerobot.processor import (
+    AddBatchDimensionProcessorStep,
+    DeviceProcessorStep,
+    NormalizerProcessorStep,
+    PolicyAction,
+    PolicyProcessorPipeline,
+    RenameObservationsProcessorStep,
+    TokenizerProcessorStep,
+    UnnormalizerProcessorStep,
+)
+from lerobot.processor.converters import policy_action_to_transition, transition_to_policy_action
+from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
+
+
+def make_multi_task_dit_pre_post_processors(
+    config: MultiTaskDiTConfig,
+    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
+) -> tuple[
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+    PolicyProcessorPipeline[PolicyAction, PolicyAction],
+]:
+    """
+    Constructs pre-processor and post-processor pipelines for a Multi-Task DiT policy.
+
+    The pre-processing pipeline prepares the input data for the model by:
+    1. Renaming features.
+    2. Adding a batch dimension.
+    3. Tokenizing the language task description (if present).
+    4. Moving the data to the specified device.
+    5. Normalizing the input and output features based on dataset statistics.
+
+    The post-processing pipeline handles the model's output by:
+    1. Unnormalizing the output features to their original scale.
+    2. Moving the data to the CPU.
+
+    Args:
+        config: The configuration object for the Multi-Task DiT policy,
+            containing feature definitions, normalization mappings, and device information.
+        dataset_stats: A dictionary of statistics used for normalization.
+            Defaults to None.
+
+    Returns:
+        A tuple containing the configured pre-processor and post-processor pipelines.
+    """
+
+    input_steps = [
+        RenameObservationsProcessorStep(rename_map={}),
+        AddBatchDimensionProcessorStep(),
+        TokenizerProcessorStep(
+            tokenizer_name=config.text_encoder_name,
+            padding=config.tokenizer_padding,
+            padding_side=config.tokenizer_padding_side,
+            max_length=config.tokenizer_max_length,
+            truncation=config.tokenizer_truncation,
+        ),
+        DeviceProcessorStep(device=config.device),
+        NormalizerProcessorStep(
+            features={**config.input_features, **config.output_features},
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+            device=config.device,
+        ),
+    ]
+    output_steps = [
+        UnnormalizerProcessorStep(
+            features=config.output_features,
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+        ),
+        DeviceProcessorStep(device="cpu"),
+    ]
+
+    return (
+        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
+            steps=input_steps,
+            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
+        ),
+        PolicyProcessorPipeline[PolicyAction, PolicyAction](
+            steps=output_steps,
+            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
+            to_transition=policy_action_to_transition,
+            to_output=transition_to_policy_action,
+        ),
+    )

src/lerobot/policies/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
+from typing import TYPE_CHECKING, Literal, TypedDict, Unpack
 
 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
 
@@ -32,13 +32,21 @@ 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 transformers.models.gemma.modeling_gemma import GemmaForCausalLM
-    from transformers.models.paligemma.modeling_paligemma import PaliGemmaForConditionalGeneration
+
+    from lerobot.policies.pi_gemma import (
+        PaliGemmaForConditionalGenerationWithPiGemma,
+        PiGemmaForCausalLM,
+        _gated_residual,
+        layernorm_forward,
+    )
 else:
     CONFIG_MAPPING = None
     modeling_gemma = None
-    GemmaForCausalLM = None
-    PaliGemmaForConditionalGeneration = None
+    PiGemmaForCausalLM = None
+    _gated_residual = None
+    layernorm_forward = None
+    PaliGemmaForConditionalGenerationWithPiGemma = None
+
 
 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.policies.pi0.configuration_pi0 import DEFAULT_IMAGE_SIZE, PI0Config
@@ -191,7 +199,7 @@ 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(-1.0, 1.0)
+        resized_images = resized_images.clamp(0.0, 1.0)
     else:
         raise ValueError(f"Unsupported image dtype: {images.dtype}")
 
@@ -202,7 +210,7 @@ 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 -1.0
+    constant_value = 0 if images.dtype == torch.uint8 else 0.0
     padded_images = F.pad(
         resized_images,
         (pad_w0, pad_w1, pad_h0, pad_h1),  # left, right, top, bottom
@@ -221,14 +229,14 @@ 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.language_model, gemma_expert.model]
+    models = [paligemma.model.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 = layer.input_layernorm(hidden_states, cond=adarms_cond[i])  # noqa: PLW2901
+        hidden_states, gate = layernorm_forward(layer.input_layernorm, hidden_states, adarms_cond[i])
         gates.append(gate)
         input_shape = hidden_states.shape[:-1]
         hidden_shape = (*input_shape, -1, layer.self_attn.head_dim)
@@ -254,10 +262,10 @@ def compute_layer_complete(
         query_states, key_states, cos, sin, unsqueeze_dim=1
     )
     batch_size = query_states.shape[0]
-    scaling = paligemma.language_model.layers[layer_idx].self_attn.scaling
+    scaling = paligemma.model.language_model.layers[layer_idx].self_attn.scaling
     # Attention computation
     att_output, _ = modeling_gemma.eager_attention_forward(
-        paligemma.language_model.layers[layer_idx].self_attn,
+        paligemma.model.language_model.layers[layer_idx].self_attn,
         query_states,
         key_states,
         value_states,
@@ -265,7 +273,7 @@ def compute_layer_complete(
         scaling,
     )
     # Get head_dim from the current layer, not from the model
-    head_dim = paligemma.language_model.layers[layer_idx].self_attn.head_dim
+    head_dim = paligemma.model.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 = []
@@ -277,15 +285,15 @@ 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 = modeling_gemma._gated_residual(hidden_states, out_emb, gates[i])  # noqa: SLF001
+        out_emb = _gated_residual(hidden_states, out_emb, gates[i])
         after_first_residual = out_emb.clone()
-        out_emb, gate = layer.post_attention_layernorm(out_emb, cond=adarms_cond[i])
+        out_emb, gate = layernorm_forward(layer.post_attention_layernorm, out_emb, 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 = modeling_gemma._gated_residual(after_first_residual, out_emb, gate)  # noqa: SLF001
+        out_emb = _gated_residual(after_first_residual, out_emb, gate)
         outputs_embeds.append(out_emb)
         start_pos = end_pos
     return outputs_embeds
@@ -358,7 +366,7 @@ 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.torch_dtype = "float32"
+        vlm_config_hf.text_config.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
@@ -366,7 +374,7 @@ 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.torch_dtype = "float32"
+        vlm_config_hf.vision_config.dtype = "float32"
 
         action_expert_config_hf = CONFIG_MAPPING["gemma"](
             head_dim=action_expert_config.head_dim,
@@ -377,13 +385,13 @@ class PaliGemmaWithExpertModel(
             num_key_value_heads=action_expert_config.num_kv_heads,
             vocab_size=257152,
             hidden_activation="gelu_pytorch_tanh",
-            torch_dtype="float32",
+            dtype="float32",
             use_adarms=use_adarms[1],
             adarms_cond_dim=action_expert_config.width if use_adarms[1] else None,
         )
 
-        self.paligemma = PaliGemmaForConditionalGeneration(config=vlm_config_hf)
-        self.gemma_expert = GemmaForCausalLM(config=action_expert_config_hf)
+        self.paligemma = PaliGemmaForConditionalGenerationWithPiGemma(config=vlm_config_hf)
+        self.gemma_expert = PiGemmaForCausalLM(config=action_expert_config_hf)
         self.gemma_expert.model.embed_tokens = None
 
         self.to_bfloat16_for_selected_params(precision)
@@ -398,10 +406,11 @@ 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.vision_model.embeddings.patch_embedding.weight",
-            "vision_tower.vision_model.embeddings.patch_embedding.bias",
-            "vision_tower.vision_model.embeddings.position_embedding.weight",
+            "vision_tower",
+            "multi_modal_projector",
             "input_layernorm",
             "post_attention_layernorm",
             "model.norm",
@@ -413,8 +422,8 @@ class PaliGemmaWithExpertModel(
 
     def _set_requires_grad(self):
         if self.freeze_vision_encoder:
-            self.paligemma.vision_tower.eval()
-            for param in self.paligemma.vision_tower.parameters():
+            self.paligemma.model.vision_tower.eval()
+            for param in self.paligemma.model.vision_tower.parameters():
                 param.requires_grad = False
         if self.train_expert_only:
             self.paligemma.eval()
@@ -424,15 +433,23 @@ class PaliGemmaWithExpertModel(
     def train(self, mode: bool = True):
         super().train(mode)
         if self.freeze_vision_encoder:
-            self.paligemma.vision_tower.eval()
+            self.paligemma.model.vision_tower.eval()
         if self.train_expert_only:
             self.paligemma.eval()
 
     def embed_image(self, image: torch.Tensor):
-        return self.paligemma.model.get_image_features(image)
+        # 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
 
     def embed_language_tokens(self, tokens: torch.Tensor):
-        return self.paligemma.language_model.embed_tokens(tokens)
+        return self.paligemma.model.language_model.embed_tokens(tokens)
 
     def forward(
         self,
@@ -446,7 +463,7 @@ class PaliGemmaWithExpertModel(
         if adarms_cond is None:
             adarms_cond = [None, None]
         if inputs_embeds[1] is None:
-            prefix_output = self.paligemma.language_model.forward(
+            prefix_output = self.paligemma.model.language_model.forward(
                 inputs_embeds=inputs_embeds[0],
                 attention_mask=attention_mask,
                 position_ids=position_ids,
@@ -470,7 +487,7 @@ class PaliGemmaWithExpertModel(
             prefix_output = None
             prefix_past_key_values = None
         else:
-            models = [self.paligemma.language_model, self.gemma_expert.model]
+            models = [self.paligemma.model.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
@@ -510,7 +527,7 @@ class PaliGemmaWithExpertModel(
             def compute_final_norms(inputs_embeds, adarms_cond):
                 outputs_embeds = []
                 for i, hidden_states in enumerate(inputs_embeds):
-                    out_emb, _ = models[i].norm(hidden_states, cond=adarms_cond[i])
+                    out_emb, _ = layernorm_forward(models[i].norm, hidden_states, adarms_cond[i])
                     outputs_embeds.append(out_emb)
                 return outputs_embeds
 
@@ -576,29 +593,19 @@ 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.language_model.gradient_checkpointing = True
-        self.paligemma_with_expert.paligemma.vision_tower.gradient_checkpointing = 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.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.language_model.gradient_checkpointing = False
-        self.paligemma_with_expert.paligemma.vision_tower.gradient_checkpointing = 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.gemma_expert.model.gradient_checkpointing = False
         logging.info("Disabled gradient checkpointing for PI0Pytorch model")
 
@@ -760,7 +767,7 @@ 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.language_model.layers[0].self_attn.q_proj.weight.dtype
+            self.paligemma_with_expert.paligemma.model.language_model.layers[0].self_attn.q_proj.weight.dtype
             == torch.bfloat16
         ):
             suffix_embs = suffix_embs.to(dtype=torch.bfloat16)
@@ -834,7 +841,7 @@ 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.language_model.config._attn_implementation = "eager"  # noqa: SLF001
+        self.paligemma_with_expert.paligemma.model.language_model.config._attn_implementation = "eager"  # noqa: SLF001
 
         _, past_key_values = self.paligemma_with_expert.forward(
             attention_mask=prefix_att_2d_masks_4d,
@@ -908,6 +915,7 @@ 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,
@@ -997,14 +1005,12 @@ class PI0Policy(PreTrainedPolicy):
         # Check if dataset_stats were provided in kwargs
         model = cls(config, **kwargs)
 
-        # Now manually load and remap the state dict
+        # Load state dict (expects keys with "model." prefix)
         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",
@@ -1012,7 +1018,7 @@ class PI0Policy(PreTrainedPolicy):
                     force_download=kwargs.get("force_download", False),
                     resume_download=kwargs.get("resume_download"),
                     proxies=kwargs.get("proxies"),
-                    use_auth_token=kwargs.get("use_auth_token"),
+                    token=kwargs.get("token"),
                     revision=kwargs.get("revision"),
                     local_files_only=kwargs.get("local_files_only", False),
                 )
@@ -1025,7 +1031,7 @@ 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
@@ -1070,7 +1076,7 @@ class PI0Policy(PreTrainedPolicy):
                 print("All keys loaded successfully!")
 
         except Exception as e:
-            print(f"Warning: Could not remap state dict keys: {e}")
+            print(f"Warning: Could not load state dict: {e}")
 
         return model
 
@@ -1120,6 +1126,14 @@ 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
+from typing import TYPE_CHECKING, Literal, TypedDict, Unpack
 
 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
 
@@ -32,14 +32,20 @@ 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 transformers.models.gemma.modeling_gemma import GemmaForCausalLM
-    from transformers.models.paligemma.modeling_paligemma import PaliGemmaForConditionalGeneration
+
+    from lerobot.policies.pi_gemma import (
+        PaliGemmaForConditionalGenerationWithPiGemma,
+        PiGemmaForCausalLM,
+        _gated_residual,
+        layernorm_forward,
+    )
 else:
     CONFIG_MAPPING = None
     modeling_gemma = None
-    GemmaForCausalLM = None
-    PaliGemmaForConditionalGeneration = None
-
+    PiGemmaForCausalLM = None
+    _gated_residual = None
+    layernorm_forward = None
+    PaliGemmaForConditionalGenerationWithPiGemma = 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
@@ -92,10 +98,11 @@ def create_sinusoidal_pos_embedding(  # see openpi `create_sinusoidal_pos_embedd
 
 
 def sample_beta(alpha, beta, bsize, device):  # see openpi `sample_beta` (exact copy)
-    alpha_t = torch.as_tensor(alpha, dtype=torch.float32, device=device)
-    beta_t = torch.as_tensor(beta, dtype=torch.float32, device=device)
+    # 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)
     dist = torch.distributions.Beta(alpha_t, beta_t)
-    return dist.sample((bsize,))
+    return dist.sample((bsize,)).to(device)
 
 
 def make_att_2d_masks(pad_masks, att_masks):  # see openpi `make_att_2d_masks` (exact copy)
@@ -189,7 +196,7 @@ 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(-1.0, 1.0)
+        resized_images = resized_images.clamp(0.0, 1.0)
     else:
         raise ValueError(f"Unsupported image dtype: {images.dtype}")
 
@@ -200,7 +207,7 @@ 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 -1.0
+    constant_value = 0 if images.dtype == torch.uint8 else 0.0
     padded_images = F.pad(
         resized_images,
         (pad_w0, pad_w1, pad_h0, pad_h1),  # left, right, top, bottom
@@ -219,14 +226,14 @@ 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.language_model, gemma_expert.model]
+    models = [paligemma.model.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 = layer.input_layernorm(hidden_states, cond=adarms_cond[i])  # noqa: PLW2901
+        hidden_states, gate = layernorm_forward(layer.input_layernorm, hidden_states, adarms_cond[i])
         gates.append(gate)
         input_shape = hidden_states.shape[:-1]
         hidden_shape = (*input_shape, -1, layer.self_attn.head_dim)
@@ -252,10 +259,10 @@ def compute_layer_complete(
         query_states, key_states, cos, sin, unsqueeze_dim=1
     )
     batch_size = query_states.shape[0]
-    scaling = paligemma.language_model.layers[layer_idx].self_attn.scaling
+    scaling = paligemma.model.language_model.layers[layer_idx].self_attn.scaling
     # Attention computation
     att_output, _ = modeling_gemma.eager_attention_forward(
-        paligemma.language_model.layers[layer_idx].self_attn,
+        paligemma.model.language_model.layers[layer_idx].self_attn,
         query_states,
         key_states,
         value_states,
@@ -263,7 +270,7 @@ def compute_layer_complete(
         scaling,
     )
     # Get head_dim from the current layer, not from the model
-    head_dim = paligemma.language_model.layers[layer_idx].self_attn.head_dim
+    head_dim = paligemma.model.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 = []
@@ -275,15 +282,15 @@ 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 = modeling_gemma._gated_residual(hidden_states, out_emb, gates[i])  # noqa: SLF001
+        out_emb = _gated_residual(hidden_states, out_emb, gates[i])
         after_first_residual = out_emb.clone()
-        out_emb, gate = layer.post_attention_layernorm(out_emb, cond=adarms_cond[i])
+        out_emb, gate = layernorm_forward(layer.post_attention_layernorm, out_emb, 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 = modeling_gemma._gated_residual(after_first_residual, out_emb, gate)  # noqa: SLF001
+        out_emb = _gated_residual(after_first_residual, out_emb, gate)
         outputs_embeds.append(out_emb)
         start_pos = end_pos
     return outputs_embeds
@@ -356,7 +363,7 @@ 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.torch_dtype = "float32"
+        vlm_config_hf.text_config.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
@@ -364,7 +371,7 @@ 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.torch_dtype = "float32"
+        vlm_config_hf.vision_config.dtype = "float32"
 
         action_expert_config_hf = CONFIG_MAPPING["gemma"](
             head_dim=action_expert_config.head_dim,
@@ -375,13 +382,13 @@ class PaliGemmaWithExpertModel(
             num_key_value_heads=action_expert_config.num_kv_heads,
             vocab_size=257152,
             hidden_activation="gelu_pytorch_tanh",
-            torch_dtype="float32",
+            dtype="float32",
             use_adarms=use_adarms[1],
             adarms_cond_dim=action_expert_config.width if use_adarms[1] else None,
         )
 
-        self.paligemma = PaliGemmaForConditionalGeneration(config=vlm_config_hf)
-        self.gemma_expert = GemmaForCausalLM(config=action_expert_config_hf)
+        self.paligemma = PaliGemmaForConditionalGenerationWithPiGemma(config=vlm_config_hf)
+        self.gemma_expert = PiGemmaForCausalLM(config=action_expert_config_hf)
         self.gemma_expert.model.embed_tokens = None
 
         self.to_bfloat16_for_selected_params(precision)
@@ -396,10 +403,11 @@ 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.vision_model.embeddings.patch_embedding.weight",
-            "vision_tower.vision_model.embeddings.patch_embedding.bias",
-            "vision_tower.vision_model.embeddings.position_embedding.weight",
+            "vision_tower",
+            "multi_modal_projector",
             "input_layernorm",
             "post_attention_layernorm",
             "model.norm",
@@ -411,8 +419,8 @@ class PaliGemmaWithExpertModel(
 
     def _set_requires_grad(self):
         if self.freeze_vision_encoder:
-            self.paligemma.vision_tower.eval()
-            for param in self.paligemma.vision_tower.parameters():
+            self.paligemma.model.vision_tower.eval()
+            for param in self.paligemma.model.vision_tower.parameters():
                 param.requires_grad = False
         if self.train_expert_only:
             self.paligemma.eval()
@@ -422,15 +430,23 @@ class PaliGemmaWithExpertModel(
     def train(self, mode: bool = True):
         super().train(mode)
         if self.freeze_vision_encoder:
-            self.paligemma.vision_tower.eval()
+            self.paligemma.model.vision_tower.eval()
         if self.train_expert_only:
             self.paligemma.eval()
 
     def embed_image(self, image: torch.Tensor):
-        return self.paligemma.model.get_image_features(image)
+        # 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
 
     def embed_language_tokens(self, tokens: torch.Tensor):
-        return self.paligemma.language_model.embed_tokens(tokens)
+        return self.paligemma.model.language_model.embed_tokens(tokens)
 
     def forward(
         self,
@@ -444,7 +460,7 @@ class PaliGemmaWithExpertModel(
         if adarms_cond is None:
             adarms_cond = [None, None]
         if inputs_embeds[1] is None:
-            prefix_output = self.paligemma.language_model.forward(
+            prefix_output = self.paligemma.model.language_model.forward(
                 inputs_embeds=inputs_embeds[0],
                 attention_mask=attention_mask,
                 position_ids=position_ids,
@@ -468,7 +484,7 @@ class PaliGemmaWithExpertModel(
             prefix_output = None
             prefix_past_key_values = None
         else:
-            models = [self.paligemma.language_model, self.gemma_expert.model]
+            models = [self.paligemma.model.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
@@ -508,7 +524,7 @@ class PaliGemmaWithExpertModel(
             def compute_final_norms(inputs_embeds, adarms_cond):
                 outputs_embeds = []
                 for i, hidden_states in enumerate(inputs_embeds):
-                    out_emb, _ = models[i].norm(hidden_states, cond=adarms_cond[i])
+                    out_emb, _ = layernorm_forward(models[i].norm, hidden_states, adarms_cond[i])
                     outputs_embeds.append(out_emb)
                 return outputs_embeds
 
@@ -573,29 +589,19 @@ 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.language_model.gradient_checkpointing = True
-        self.paligemma_with_expert.paligemma.vision_tower.gradient_checkpointing = 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.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.language_model.gradient_checkpointing = False
-        self.paligemma_with_expert.paligemma.vision_tower.gradient_checkpointing = 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.gemma_expert.model.gradient_checkpointing = False
         logging.info("Disabled gradient checkpointing for PI05Pytorch model")
 
@@ -737,7 +743,7 @@ 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.language_model.layers[0].self_attn.q_proj.weight.dtype
+            self.paligemma_with_expert.paligemma.model.language_model.layers[0].self_attn.q_proj.weight.dtype
             == torch.bfloat16
         ):
             suffix_embs = suffix_embs.to(dtype=torch.bfloat16)
@@ -808,7 +814,7 @@ 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.language_model.config._attn_implementation = "eager"  # noqa: SLF001
+        self.paligemma_with_expert.paligemma.model.language_model.config._attn_implementation = "eager"  # noqa: SLF001
 
         _, past_key_values = self.paligemma_with_expert.forward(
             attention_mask=prefix_att_2d_masks_4d,
@@ -880,6 +886,7 @@ 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,
@@ -969,14 +976,12 @@ class PI05Policy(PreTrainedPolicy):
         # Check if dataset_stats were provided in kwargs
         model = cls(config, **kwargs)
 
-        # Now manually load and remap the state dict
+        # Load state dict (expects keys with "model." prefix)
         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",
@@ -984,7 +989,7 @@ class PI05Policy(PreTrainedPolicy):
                     force_download=kwargs.get("force_download", False),
                     resume_download=kwargs.get("resume_download"),
                     proxies=kwargs.get("proxies"),
-                    use_auth_token=kwargs.get("use_auth_token"),
+                    token=kwargs.get("token"),
                     revision=kwargs.get("revision"),
                     local_files_only=kwargs.get("local_files_only", False),
                 )
@@ -997,7 +1002,7 @@ 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
@@ -1009,8 +1014,6 @@ 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
 
@@ -1044,7 +1047,7 @@ class PI05Policy(PreTrainedPolicy):
                 print("All keys loaded successfully!")
 
         except Exception as e:
-            print(f"Warning: Could not remap state dict keys: {e}")
+            print(f"Warning: Could not load state dict: {e}")
 
         return model
 
@@ -1098,6 +1101,14 @@ 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

@@ -23,7 +23,6 @@ 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,
@@ -68,9 +67,6 @@ 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 = "physical-intelligence/fast"
+    action_tokenizer_name: str = "lerobot/fast-action-tokenizer"
     temperature: float = 0.0
     max_decoding_steps: int = 256
     fast_skip_tokens: int = 128

src/lerobot/policies/pi0_fast/modeling_pi0_fast.py

@@ -19,13 +19,12 @@ import logging
 import math
 from collections import deque
 from pathlib import Path
-from typing import TYPE_CHECKING, Literal, TypedDict
+from typing import TYPE_CHECKING, Literal, TypedDict, Unpack
 
 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
 
@@ -38,11 +37,16 @@ else:
 if TYPE_CHECKING or _transformers_available:
     from transformers import AutoTokenizer
     from transformers.models.auto import CONFIG_MAPPING
-    from transformers.models.paligemma.modeling_paligemma import PaliGemmaForConditionalGeneration
+
+    from lerobot.policies.pi_gemma import (
+        PaliGemmaForConditionalGenerationWithPiGemma,
+        PiGemmaModel,
+    )
 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
@@ -121,7 +125,7 @@ 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(-1.0, 1.0)
+        resized_images = resized_images.clamp(0.0, 1.0)
     else:
         raise ValueError(f"Unsupported image dtype: {images.dtype}")
 
@@ -132,7 +136,7 @@ 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 -1.0
+    constant_value = 0 if images.dtype == torch.uint8 else 0.0
     padded_images = F.pad(
         resized_images,
         (pad_w0, pad_w1, pad_h0, pad_h1),  # left, right, top, bottom
@@ -206,16 +210,22 @@ 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.torch_dtype = "float32"
+        vlm_config_hf.text_config.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.torch_dtype = "float32"
+        vlm_config_hf.vision_config.dtype = "float32"
 
-        self.paligemma = PaliGemmaForConditionalGeneration(config=vlm_config_hf)
+        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.to_bfloat16_for_selected_params(precision)
 
@@ -228,10 +238,11 @@ 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.vision_model.embeddings.patch_embedding.weight",
-            "vision_tower.vision_model.embeddings.patch_embedding.bias",
-            "vision_tower.vision_model.embeddings.position_embedding.weight",
+            "vision_tower",
+            "multi_modal_projector",
             "input_layernorm",
             "post_attention_layernorm",
             "model.norm",
@@ -242,10 +253,18 @@ class PI0FastPaliGemma(nn.Module):
                 param.data = param.data.to(dtype=torch.float32)
 
     def embed_image(self, image: torch.Tensor):
-        return self.paligemma.model.get_image_features(image)
+        # 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
 
     def embed_language_tokens(self, tokens: torch.Tensor):
-        return self.paligemma.language_model.embed_tokens(tokens)
+        return self.paligemma.model.language_model.embed_tokens(tokens)
 
     def forward(
         self,
@@ -259,7 +278,7 @@ class PI0FastPaliGemma(nn.Module):
         if adarms_cond is None:
             adarms_cond = [None, None]
         if inputs_embeds[1] is None:
-            prefix_output = self.paligemma.language_model.forward(
+            prefix_output = self.paligemma.model.language_model.forward(
                 inputs_embeds=inputs_embeds[0],
                 attention_mask=attention_mask,
                 position_ids=position_ids,
@@ -306,24 +325,14 @@ 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.language_model.gradient_checkpointing_enable(
+        self.paligemma_with_expert.paligemma.model.language_model.gradient_checkpointing_enable(
             gradient_checkpointing_kwargs={"use_reentrant": False}
         )
-        self.paligemma_with_expert.paligemma.vision_tower.gradient_checkpointing_enable(
+        self.paligemma_with_expert.paligemma.model.vision_tower.gradient_checkpointing_enable(
             gradient_checkpointing_kwargs={"use_reentrant": False}
         )
         logging.info("Enabled gradient checkpointing for PI0FastPytorch model")
@@ -332,8 +341,8 @@ 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.language_model.gradient_checkpointing_disable()
-        self.paligemma_with_expert.paligemma.vision_tower.gradient_checkpointing_disable()
+        self.paligemma_with_expert.paligemma.model.language_model.gradient_checkpointing_disable()
+        self.paligemma_with_expert.paligemma.model.vision_tower.gradient_checkpointing_disable()
         logging.info("Disabled gradient checkpointing for PI0FastPytorch model")
 
     def _apply_checkpoint(self, func, *args, **kwargs):
@@ -523,7 +532,7 @@ class PI0FastPytorch(nn.Module):  # see openpi `PI0Pytorch`
 
         # Convert embeddings to bfloat16 if needed
         if (
-            self.paligemma_with_expert.paligemma.language_model.layers[0].self_attn.q_proj.weight.dtype
+            self.paligemma_with_expert.paligemma.model.language_model.layers[0].self_attn.q_proj.weight.dtype
             == torch.bfloat16
         ):
             prefix_embs = prefix_embs.to(dtype=torch.bfloat16)
@@ -616,7 +625,7 @@ class PI0FastPytorch(nn.Module):  # see openpi `PI0Pytorch`
         )
 
         if (
-            self.paligemma_with_expert.paligemma.language_model.layers[0].self_attn.q_proj.weight.dtype
+            self.paligemma_with_expert.paligemma.model.language_model.layers[0].self_attn.q_proj.weight.dtype
             == torch.bfloat16
         ):
             prefix_embs = prefix_embs.to(dtype=torch.bfloat16)
@@ -714,7 +723,7 @@ class PI0FastPytorch(nn.Module):  # see openpi `PI0Pytorch`
 
         # Ensure correct precision (bfloat16/float32)
         if (
-            self.paligemma_with_expert.paligemma.language_model.layers[0].self_attn.q_proj.weight.dtype
+            self.paligemma_with_expert.paligemma.model.language_model.layers[0].self_attn.q_proj.weight.dtype
             == torch.bfloat16
         ):
             prefix_embs = prefix_embs.to(dtype=torch.bfloat16)
@@ -897,14 +906,12 @@ class PI0FastPolicy(PreTrainedPolicy):
         # Check if dataset_stats were provided in kwargs
         model = cls(config, **kwargs)
 
-        # Now manually load and remap the state dict
+        # Load state dict (expects keys with "model." prefix)
         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",
@@ -912,7 +919,7 @@ class PI0FastPolicy(PreTrainedPolicy):
                     force_download=kwargs.get("force_download", False),
                     resume_download=kwargs.get("resume_download"),
                     proxies=kwargs.get("proxies"),
-                    use_auth_token=kwargs.get("use_auth_token"),
+                    token=kwargs.get("token"),
                     revision=kwargs.get("revision"),
                     local_files_only=kwargs.get("local_files_only", False),
                 )
@@ -925,8 +932,9 @@ 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
@@ -936,8 +944,6 @@ 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
 
@@ -971,7 +977,7 @@ class PI0FastPolicy(PreTrainedPolicy):
                 print("All keys loaded successfully!")
 
         except Exception as e:
-            print(f"Warning: Could not remap state dict keys: {e}")
+            print(f"Warning: Could not load state dict: {e}")
 
         return model
 

src/lerobot/policies/pi0_fast/processor_pi0_fast.py

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

@@ -0,0 +1,363 @@
+# 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
+from typing import TypedDict, TypeVar, Unpack
 
 import packaging
 import safetensors
@@ -28,7 +28,6 @@ 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"
+    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.
     device: str = "cpu"
     model_type: str = "cnn"  # "transformer" or "cnn"
     num_cameras: int = 2

src/lerobot/policies/smolvla/configuration_smolvla.py

@@ -106,6 +106,9 @@ 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

@@ -54,12 +54,11 @@ policy = SmolVLAPolicy.from_pretrained("lerobot/smolvla_base")
 
 import math
 from collections import deque
-from typing import TypedDict
+from typing import TypedDict, Unpack
 
 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
@@ -593,6 +592,12 @@ 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

@@ -77,7 +77,6 @@ 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 = "physical-intelligence/fast"
+    action_tokenizer_path: str | None = "lerobot/fast-action-tokenizer"
 
     # Action prediction mode: "diffusion" or "fast"
     prediction_mode: str = "diffusion"

src/lerobot/policies/wall_x/modeling_wall_x.py

@@ -261,10 +261,15 @@ class Qwen2_5_VLMoEForAction(Qwen2_5_VLForConditionalGeneration):
     and optional LoRA fine-tuning support.
     """
 
-    _tied_weights_keys = ["lm_head.weight"]
+    _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.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,
@@ -312,6 +317,11 @@ 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)
 
@@ -331,7 +341,7 @@ class Qwen2_5_VLMoEForAction(Qwen2_5_VLForConditionalGeneration):
                 force_download=kwargs.get("force_download", False),
                 resume_download=kwargs.get("resume_download"),
                 proxies=kwargs.get("proxies"),
-                use_auth_token=kwargs.get("use_auth_token"),
+                token=kwargs.get("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

@@ -21,6 +21,7 @@ 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)
@@ -38,6 +39,7 @@ 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

@@ -11,7 +11,6 @@ from transformers.activations import ACT2FN
 from transformers.cache_utils import (
     Cache,
     DynamicCache,
-    SlidingWindowCache,
     StaticCache,
 )
 from transformers.generation import GenerationMixin
@@ -31,6 +30,15 @@ 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
@@ -594,19 +602,40 @@ 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
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+
+        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 = {}
 
         inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
         self.register_buffer("inv_freq", inv_freq, persistent=False)
@@ -1567,7 +1596,7 @@ QWEN2_5_VL_INPUTS_DOCSTRING = r"""
 
 
 class Qwen2_5_VLForConditionalGeneration(Qwen2_5_VLPreTrainedModel, GenerationMixin):
-    _tied_weights_keys = ["lm_head.weight"]
+    _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.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, videos=None, return_tensors=return_tensors)
+        image_inputs = processor.image_processor(images=images, 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(images=None, videos=videos, return_tensors=return_tensors)
+        videos_inputs = processor.image_processor(videos=videos, return_tensors=return_tensors)
         video_grid_thw = videos_inputs["video_grid_thw"]
     else:
         videos_inputs = {}

src/lerobot/policies/xvla/configuration_florence2.py

@@ -276,6 +276,8 @@ 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

@@ -1951,7 +1951,10 @@ class Florence2Decoder(Florence2LanguagePreTrainedModel):
 
 
 class Florence2LanguageModel(Florence2LanguagePreTrainedModel):
-    _tied_weights_keys = ["encoder.embed_tokens.weight", "decoder.embed_tokens.weight"]
+    _tied_weights_keys = {
+        "encoder.embed_tokens.weight": "shared.weight",
+        "decoder.embed_tokens.weight": "shared.weight",
+    }
 
     def __init__(self, config: Florence2LanguageConfig):
         super().__init__(config)
@@ -2076,7 +2079,10 @@ class Florence2LanguageModel(Florence2LanguagePreTrainedModel):
 
 class Florence2LanguageForConditionalGeneration(Florence2LanguagePreTrainedModel, GenerationMixin):
     base_model_prefix = "model"
-    _tied_weights_keys = ["encoder.embed_tokens.weight", "decoder.embed_tokens.weight", "lm_head.weight"]
+    _tied_weights_keys = {
+        "model.encoder.embed_tokens.weight": "model.shared.weight",
+        "model.decoder.embed_tokens.weight": "model.shared.weight",
+    }
     _keys_to_ignore_on_load_missing = ["final_logits_bias"]
 
     def __init__(self, config: Florence2LanguageConfig):
@@ -2436,11 +2442,10 @@ FLORENCE2_INPUTS_DOCSTRING = r"""
     FLORENCE2_START_DOCSTRING,
 )
 class Florence2ForConditionalGeneration(Florence2PreTrainedModel):
-    _tied_weights_keys = [
-        "language_model.encoder.embed_tokens.weight",
-        "language_model.decoder.embed_tokens.weight",
-        "language_model.lm_head.weight",
-    ]
+    _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",
+    }
 
     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, TypeAlias, TypedDict
+from typing import Any, TypedDict
 
 import numpy as np
 import torch
@@ -36,10 +36,10 @@ class TransitionKey(str, Enum):
     COMPLEMENTARY_DATA = "complementary_data"
 
 
-PolicyAction: TypeAlias = torch.Tensor
-RobotAction: TypeAlias = dict[str, Any]
-EnvAction: TypeAlias = np.ndarray
-RobotObservation: TypeAlias = dict[str, Any]
+PolicyAction = torch.Tensor
+RobotAction = dict[str, Any]
+EnvAction = np.ndarray
+RobotObservation = 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, Generic, TypeAlias, TypedDict, TypeVar, cast
+from typing import Any, TypedDict, TypeVar, cast
 
 import torch
 from huggingface_hub import hf_hub_download
@@ -251,7 +251,7 @@ class ProcessorMigrationError(Exception):
 
 
 @dataclass
-class DataProcessorPipeline(HubMixin, Generic[TInput, TOutput]):
+class DataProcessorPipeline[TInput, TOutput](HubMixin):
     """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(HubMixin, Generic[TInput, TOutput]):
 
 
 # Type aliases for semantic clarity.
-RobotProcessorPipeline: TypeAlias = DataProcessorPipeline[TInput, TOutput]
-PolicyProcessorPipeline: TypeAlias = DataProcessorPipeline[TInput, TOutput]
+RobotProcessorPipeline = DataProcessorPipeline[TInput, TOutput]
+PolicyProcessorPipeline = 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., "physical-intelligence/fast").
+        tokenizer_name: The name of a pretrained processor from the Hugging Face Hub (e.g., "lerobot/fast-action-tokenizer").
         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

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

src/lerobot/robots/so_follower/so_follower.py

@@ -17,7 +17,6 @@
 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
@@ -230,5 +229,5 @@ class SOFollower(Robot):
         logger.info(f"{self} disconnected.")
 
 
-SO100Follower: TypeAlias = SOFollower
-SO101Follower: TypeAlias = SOFollower
+SO100Follower = SOFollower
+SO101Follower = SOFollower

src/lerobot/scripts/lerobot_calibrate.py

@@ -56,6 +56,7 @@ 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

@@ -132,10 +132,13 @@ 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["frame_index"][i].item())
+            rr.set_time("frame_index", sequence=batch["index"][i].item() - first_index)
             rr.set_time("timestamp", timestamp=batch["timestamp"][i].item())
 
             # display each camera image

src/lerobot/scripts/lerobot_edit_dataset.py

@@ -21,6 +21,9 @@ 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:
@@ -29,16 +32,31 @@ Delete episodes 0, 2, and 5 from a dataset:
         --operation.type delete_episodes \
         --operation.episode_indices "[0, 2, 5]"
 
-Delete episodes and save to a new dataset:
+Delete episodes from a local dataset at a specific path:
     lerobot-edit-dataset \
         --repo_id lerobot/pusht \
-        --new_repo_id lerobot/pusht_filtered \
+        --root /path/to/pusht \
         --operation.type delete_episodes \
         --operation.episode_indices "[0, 2, 5]"
 
-Split dataset by fractions:
+Delete episodes and save to a new dataset at a specific path and with a new repo_id:
     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):
+    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}'
 
@@ -56,15 +74,29 @@ Split into more than two splits:
 
 Merge multiple datasets:
     lerobot-edit-dataset \
-        --repo_id lerobot/pusht_merged \
+        --new_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.images.top']"
+        --operation.feature_names "['observation.image']"
 
 Modify tasks - set a single task for all episodes (WARNING: modifies in-place):
     lerobot-edit-dataset \
@@ -88,8 +120,8 @@ 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 \
-        --operation.type convert_image_to_video \
-        --operation.output_dir /path/to/output/pusht_video
+        --new_root /path/to/output/pusht_video \
+        --operation.type convert_image_to_video
 
 Convert image dataset to video format and save with new repo_id:
     lerobot-edit-dataset \
@@ -104,28 +136,6 @@ 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 \
@@ -189,6 +199,7 @@ class SplitConfig(OperationConfig):
 @dataclass
 class MergeConfig(OperationConfig):
     repo_ids: list[str] | None = None
+    roots: list[str] | None = None
 
 
 @OperationConfig.register_subclass("remove_feature")
@@ -222,59 +233,46 @@ 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:
-    repo_id: str
-    operation: (
-        DeleteEpisodesConfig | SplitConfig | MergeConfig | RemoveFeatureConfig | ConvertImageToVideoConfig | TrimEpisodeConfig
-    )
+    # Operation configuration.
     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 | 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")
+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
 
-        if dataset_path.exists():
-            if old_path.exists():
-                shutil.rmtree(old_path)
-            shutil.move(str(dataset_path), str(old_path))
+    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
 
-        output_dir = dataset_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")
 
-    return output_repo_id, output_dir
+        if input_path.exists():
+            if backup_path.exists():
+                shutil.rmtree(backup_path)
+            shutil.move(input_path, backup_path)
+
+    return output_repo_id, output_path
 
 
 def handle_delete_episodes(cfg: EditDatasetConfig) -> None:
@@ -286,11 +284,15 @@ 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, cfg.new_repo_id, Path(cfg.root) if cfg.root else None
+        cfg.repo_id,
+        new_repo_id=cfg.new_repo_id,
+        root=cfg.root,
+        new_root=cfg.new_root,
     )
 
-    if cfg.new_repo_id is None:
-        dataset.root = Path(str(dataset.root) + "_old")
+    # 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")
 
     logging.info(f"Deleting episodes {cfg.operation.episode_indices} from {cfg.repo_id}")
     new_dataset = delete_episodes(
@@ -317,19 +319,27 @@ 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)
+    split_datasets = split_dataset(
+        dataset,
+        splits=cfg.operation.splits,
+        output_dir=cfg.new_root,
+    )
 
     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_repo_id}")
+            logging.info(f"Pushing {split_name} split to hub as {split_ds.repo_id}")
             LeRobotDataset(split_ds.repo_id, root=split_ds.root).push_to_hub()
 
 
@@ -340,18 +350,29 @@ def handle_merge(cfg: EditDatasetConfig) -> None:
     if not cfg.operation.repo_ids:
         raise ValueError("repo_ids must be specified for merge operation")
 
-    if not cfg.repo_id:
-        raise ValueError("repo_id must be specified as the output repository for merged dataset")
+    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."
+        )
 
-    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]
+    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]
 
-    output_dir = Path(cfg.root) / cfg.repo_id if cfg.root else HF_LEROBOT_HOME / cfg.repo_id
+    output_dir = Path(cfg.new_root) if cfg.new_root else HF_LEROBOT_HOME / cfg.new_repo_id
 
-    logging.info(f"Merging datasets into {cfg.repo_id}")
+    logging.info(f"Merging datasets into {cfg.new_repo_id}")
     merged_dataset = merge_datasets(
         datasets,
-        output_repo_id=cfg.repo_id,
+        output_repo_id=cfg.new_repo_id,
         output_dir=output_dir,
     )
 
@@ -361,7 +382,7 @@ def handle_merge(cfg: EditDatasetConfig) -> None:
     )
 
     if cfg.push_to_hub:
-        logging.info(f"Pushing to hub as {cfg.repo_id}")
+        logging.info(f"Pushing to hub as {cfg.new_repo_id}")
         LeRobotDataset(merged_dataset.repo_id, root=output_dir).push_to_hub()
 
 
@@ -374,11 +395,15 @@ 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, cfg.new_repo_id, Path(cfg.root) if cfg.root else None
+        cfg.repo_id,
+        new_repo_id=cfg.new_repo_id,
+        root=cfg.root,
+        new_root=cfg.new_root,
     )
 
-    if cfg.new_repo_id is None:
-        dataset.root = Path(str(dataset.root) + "_old")
+    # 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")
 
     logging.info(f"Removing features {cfg.operation.feature_names} from {cfg.repo_id}")
     new_dataset = remove_feature(
@@ -396,92 +421,6 @@ 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")
@@ -492,9 +431,10 @@ 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")
 
-    # 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.")
+    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."
+        )
 
     dataset = LeRobotDataset(cfg.repo_id, root=cfg.root)
     logging.warning(f"Modifying dataset in-place at {dataset.root}. Original data will be overwritten.")
@@ -530,32 +470,30 @@ 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_repo_id, 2) operation.output_dir, 3) auto-generated name
+    # Priority: 1) new_root, 2) new_repo_id, 3) operation.output_dir, 4) 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_repo_id:
-        # Use new_repo_id for both local storage and hub push
+    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:
         output_repo_id = 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
+        output_dir = HF_LEROBOT_HOME / cfg.new_repo_id
         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}")
+        logging.info(f"Saving to local directory: {output_dir} as {output_repo_id}")
     else:
-        # Auto-generate name: append "_video" to original repo_id
         output_repo_id = f"{cfg.repo_id}_video"
-        # 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}")
+        output_dir = HF_LEROBOT_HOME / output_repo_id
+        logging.info(f"Saving to auto-generated location: {output_dir} as {output_repo_id}")
 
     logging.info(f"Converting dataset {cfg.repo_id} to video format")
 
@@ -630,8 +568,20 @@ 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":
@@ -646,8 +596,6 @@ 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

@@ -61,6 +61,7 @@ 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

@@ -125,6 +125,7 @@ from lerobot.teleoperators import (  # noqa: F401
     make_teleoperator_from_config,
     omx_leader,
     openarm_leader,
+    openarm_mini,
     reachy2_teleoperator,
     so_leader,
     unitree_g1,
@@ -154,7 +155,7 @@ 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').
+    # Root directory where the dataset will be stored (e.g. 'dataset/path'). If None, defaults to $HF_LEROBOT_HOME/repo_id.
     root: str | Path | None = None
     # Limit the frames per second.
     fps: int = 30
@@ -333,6 +334,7 @@ def record_loop(
         preprocessor.reset()
         postprocessor.reset()
 
+    no_action_count = 0
     timestamp = 0
     start_episode_t = time.perf_counter()
     while timestamp < control_time_s:
@@ -380,11 +382,13 @@ 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:
-            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."
-            )
+            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."
+                )
             continue
 
         # Applies a pipeline to the action, default is IdentityProcessor

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').
+    # Root directory where the dataset will be stored (e.g. 'dataset/path'). If None, defaults to $HF_LEROBOT_HOME/repo_id.
     root: str | Path | None = None
     # Limit the frames per second. By default, uses the policy fps.
     fps: int = 30

src/lerobot/scripts/lerobot_teleoperate.py

@@ -94,6 +94,7 @@ from lerobot.teleoperators import (  # noqa: F401
     make_teleoperator_from_config,
     omx_leader,
     openarm_leader,
+    openarm_mini,
     reachy2_teleoperator,
     so_leader,
     unitree_g1,

src/lerobot/scripts/lerobot_train.py

@@ -380,10 +380,10 @@ def train(cfg: TrainPipelineConfig, accelerator: Accelerator | None = None):
         "dataloading_s": AverageMeter("data_s", ":.3f"),
     }
 
-    # Use effective batch size for proper epoch calculation in distributed training
+    # Keep global batch size for logging; MetricsTracker handles world size internally.
     effective_batch_size = cfg.batch_size * accelerator.num_processes
     train_tracker = MetricsTracker(
-        effective_batch_size,
+        cfg.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("physical-intelligence/fast", trust_remote_code=True)
+    base_tokenizer = AutoProcessor.from_pretrained("lerobot/fast-action-tokenizer", 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

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

src/lerobot/teleoperators/so_leader/so_leader.py

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

src/lerobot/utils/io_utils.py

@@ -16,12 +16,10 @@
 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):
@@ -33,7 +31,7 @@ def write_video(video_path, stacked_frames, fps):
         imageio.mimsave(video_path, stacked_frames, fps=fps)
 
 
-def deserialize_json_into_object(fpath: Path, obj: T) -> T:
+def deserialize_json_into_object[T: JsonLike](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

@@ -104,9 +104,10 @@ 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
+        self.samples = self.steps * self._batch_size * world_size
         self.episodes = self.samples / self._avg_samples_per_ep
         self.epochs = self.samples / self._num_frames
         self.accelerator = accelerator
@@ -132,7 +133,8 @@ class MetricsTracker:
         Updates metrics that depend on 'step' for one step.
         """
         self.steps += 1
-        self.samples += self._batch_size * (self.accelerator.num_processes if self.accelerator else 1)
+        world_size = self.accelerator.num_processes if self.accelerator else 1
+        self.samples += self._batch_size * world_size
         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 sha256:19eaaa85f66ba4aa6388dbb83819ffad6ea4363247208f871a8dc385689f6fc8
+oid sha256:54aecbc1af72a4cd5e9261492f5e7601890517516257aacdf2a0ffb3ce281f1b
 size 992

tests/artifacts/policies/pusht_diffusion_/grad_stats.safetensors

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

tests/artifacts/policies/pusht_diffusion_/output_dict.safetensors

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

tests/artifacts/policies/pusht_diffusion_/param_stats.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:778fddbbaa64248cee35cb377c02cc2b6076f7ce5855146de677128900617ddf
+oid sha256:645bff922ac7bea63ad018ebf77c303c0e4cd2c1c0dc5ef3192865281bef3dc6
 size 47424

tests/async_inference/test_robot_client.py

@@ -231,3 +231,39 @@ def test_ready_to_send_observation_with_varying_threshold(robot_client, g_thresh
         robot_client.action_queue.put(act)
 
     assert robot_client._ready_to_send_observation() is expected
+
+
+# -----------------------------------------------------------------------------
+# Regression test: robot type registry populated by robot_client imports
+# -----------------------------------------------------------------------------
+
+
+def test_robot_client_registers_builtin_robot_types():
+    """Importing robot_client must populate RobotConfig's ChoiceRegistry.
+
+    This is a regression test for a bug introduced in #2425, where removing
+    robot module imports from robot_client.py caused RobotConfig's registry to
+    be empty, breaking CLI argument parsing with:
+      error: argument --robot.type: invalid choice: 'so101_follower' (choose from )
+
+    Robot types are registered via @RobotConfig.register_subclass() decorators
+    at import time, so all supported modules must be explicitly imported.
+    """
+    import lerobot.async_inference.robot_client  # noqa: F401
+    from lerobot.robots.config import RobotConfig
+
+    known_choices = RobotConfig.get_known_choices()
+
+    expected_robot_types = [
+        "so100_follower",
+        "so101_follower",
+        "koch_follower",
+        "omx_follower",
+        "bi_so_follower",
+    ]
+    for robot_type in expected_robot_types:
+        assert robot_type in known_choices, (
+            f"Robot type '{robot_type}' is not registered in RobotConfig's ChoiceRegistry. "
+            f"Ensure the corresponding module is imported in robot_client.py. "
+            f"Known choices: {sorted(known_choices)}"
+        )

tests/fixtures/dataset_factories.py

@@ -222,7 +222,7 @@ def tasks_factory():
     def _create_tasks(total_tasks: int = 3) -> pd.DataFrame:
         ids = list(range(total_tasks))
         tasks = [f"Perform action {i}." for i in ids]
-        df = pd.DataFrame({"task_index": ids}, index=tasks)
+        df = pd.DataFrame({"task_index": ids}, index=pd.Index(tasks, name="task"))
         return df
 
     return _create_tasks

tests/policies/hilserl/test_modeling_classifier.py

@@ -14,6 +14,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import pytest
 import torch
 
 from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
@@ -37,6 +38,9 @@ def test_classifier_output():
 
 
 @require_package("transformers")
+@pytest.mark.skip(
+    reason="helper2424/resnet10 needs to be updated to work with the latest version of transformers"
+)
 def test_binary_classifier_with_default_params():
     from lerobot.policies.sac.reward_model.modeling_classifier import Classifier
 
@@ -78,6 +82,9 @@ def test_binary_classifier_with_default_params():
 
 
 @require_package("transformers")
+@pytest.mark.skip(
+    reason="helper2424/resnet10 needs to be updated to work with the latest version of transformers"
+)
 def test_multiclass_classifier():
     from lerobot.policies.sac.reward_model.modeling_classifier import Classifier
 
@@ -117,6 +124,9 @@ def test_multiclass_classifier():
 
 
 @require_package("transformers")
+@pytest.mark.skip(
+    reason="helper2424/resnet10 needs to be updated to work with the latest version of transformers"
+)
 def test_default_device():
     from lerobot.policies.sac.reward_model.modeling_classifier import Classifier
 
@@ -129,6 +139,9 @@ def test_default_device():
 
 
 @require_package("transformers")
+@pytest.mark.skip(
+    reason="helper2424/resnet10 needs to be updated to work with the latest version of transformers"
+)
 def test_explicit_device_setup():
     from lerobot.policies.sac.reward_model.modeling_classifier import Classifier
 

tests/policies/multi_task_dit/test_multi_task_dit.py

@@ -0,0 +1,624 @@
+#!/usr/bin/env python
+
+# Copyright 2025 Bryson Jones 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.
+
+# ruff: noqa: E402
+
+"""Test script for Multi-Task DiT policy.
+
+To run tests locally:
+    python -m pytest tests/policies/multi_task_dit/test_multi_task_dit.py -v
+"""
+
+import os
+
+import pytest
+import torch
+from torch import Tensor
+
+pytest.importorskip("transformers")
+
+pytestmark = pytest.mark.skipif(
+    os.environ.get("CI") == "true" or os.environ.get("GITHUB_ACTIONS") == "true",
+    reason="This test requires local transformers installation and is not meant for CI",
+)
+
+from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
+from lerobot.policies.multi_task_dit.configuration_multi_task_dit import MultiTaskDiTConfig
+from lerobot.policies.multi_task_dit.modeling_multi_task_dit import MultiTaskDiTPolicy
+from lerobot.policies.multi_task_dit.processor_multi_task_dit import (
+    make_multi_task_dit_pre_post_processors,
+)
+from lerobot.utils.constants import (
+    ACTION,
+    OBS_IMAGES,
+    OBS_LANGUAGE_ATTENTION_MASK,
+    OBS_LANGUAGE_TOKENS,
+    OBS_STATE,
+)
+from lerobot.utils.random_utils import seeded_context, set_seed
+
+
+@pytest.fixture(autouse=True)
+def set_random_seed():
+    seed = 17
+    set_seed(seed)
+
+
+def create_train_batch(
+    batch_size: int = 2,
+    n_obs_steps: int = 2,
+    horizon: int = 16,
+    state_dim: int = 10,
+    action_dim: int = 10,
+    height: int = 224,
+    width: int = 224,
+) -> dict[str, Tensor]:
+    """Create a training batch with visual input and text."""
+    return {
+        "observation.state": torch.randn(batch_size, n_obs_steps, state_dim),
+        f"{OBS_IMAGES}.laptop": torch.rand(batch_size, n_obs_steps, 3, height, width),
+        ACTION: torch.randn(batch_size, horizon, action_dim),
+        "task": ["pick up the cube"] * batch_size,
+    }
+
+
+def create_observation_batch(
+    batch_size: int = 2, state_dim: int = 10, height: int = 224, width: int = 224
+) -> dict:
+    """Create observation batch for inference for a single timestep."""
+    return {
+        "observation.state": torch.randn(batch_size, state_dim),
+        f"{OBS_IMAGES}.laptop": torch.rand(batch_size, 3, height, width),
+        "task": ["pick up the red cube"] * batch_size,
+    }
+
+
+def create_config(
+    state_dim: int = 10,
+    action_dim: int = 10,
+    n_obs_steps: int = 2,
+    horizon: int = 16,
+    n_action_steps: int = 8,
+    with_visual: bool = True,
+    height: int = 224,
+    width: int = 224,
+) -> MultiTaskDiTConfig:
+    """Create a MultiTaskDiT config for testing.
+
+    Args:
+        state_dim: Dimension of state observations
+        action_dim: Dimension of actions
+        n_obs_steps: Number of observation steps
+        horizon: Action prediction horizon
+        n_action_steps: Number of action steps to execute
+        with_visual: Whether to include visual input (default: True)
+        height: Image height (only used if with_visual=True)
+        width: Image width (only used if with_visual=True)
+    """
+    input_features = {OBS_STATE: PolicyFeature(type=FeatureType.STATE, shape=(state_dim,))}
+
+    if with_visual:
+        input_features[f"{OBS_IMAGES}.laptop"] = PolicyFeature(
+            type=FeatureType.VISUAL, shape=(3, height, width)
+        )
+
+    config = MultiTaskDiTConfig(
+        input_features=input_features,
+        output_features={ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(action_dim,))},
+        n_obs_steps=n_obs_steps,
+        horizon=horizon,
+        n_action_steps=n_action_steps,
+        # Use smaller model for faster tests
+        hidden_dim=128,
+        num_layers=2,
+        num_heads=4,
+    )
+
+    config.validate_features()
+    return config
+
+
+@pytest.mark.parametrize("batch_size,state_dim,action_dim", [(2, 10, 10), (1, 6, 6)])
+def test_multi_task_dit_policy_forward(batch_size: int, state_dim: int, action_dim: int):
+    """Test forward pass (training mode)."""
+    n_obs_steps = 2
+    horizon = 16
+    n_action_steps = 8
+
+    config = create_config(
+        state_dim=state_dim,
+        action_dim=action_dim,
+        n_obs_steps=n_obs_steps,
+        horizon=horizon,
+        n_action_steps=n_action_steps,
+    )
+
+    policy = MultiTaskDiTPolicy(config=config)
+    policy.train()
+
+    # Use preprocessor to handle tokenization
+    config.normalization_mapping = {
+        "VISUAL": NormalizationMode.IDENTITY,
+        "STATE": NormalizationMode.IDENTITY,
+        "ACTION": NormalizationMode.IDENTITY,
+    }
+    preprocessor, _ = make_multi_task_dit_pre_post_processors(config=config, dataset_stats=None)
+
+    batch = create_train_batch(
+        batch_size=batch_size,
+        n_obs_steps=n_obs_steps,
+        horizon=horizon,
+        state_dim=state_dim,
+        action_dim=action_dim,
+    )
+
+    # Process batch through preprocessor to tokenize task text
+    processed_batch = preprocessor(batch)
+
+    # Test forward pass
+    loss, _ = policy.forward(processed_batch)
+    assert loss is not None
+    assert loss.item() is not None
+    assert loss.shape == ()
+
+    # Test backward pass
+    loss.backward()
+
+
+def test_multi_task_dit_pre_post_processors():
+    """Test pre and post processors for Multi-Task DiT policy."""
+    state_dim = 10
+    action_dim = 8
+    n_obs_steps = 2
+    horizon = 16
+
+    config = create_config(
+        state_dim=state_dim,
+        action_dim=action_dim,
+        n_obs_steps=n_obs_steps,
+        horizon=horizon,
+        n_action_steps=8,
+    )
+    config.device = "cpu"
+
+    # Set normalization mode to match the stats we're providing
+    config.normalization_mapping = {
+        "VISUAL": NormalizationMode.IDENTITY,
+        "STATE": NormalizationMode.MEAN_STD,  # Use MEAN_STD since we provide mean/std stats
+        "ACTION": NormalizationMode.MIN_MAX,
+    }
+
+    # Create dataset stats for normalization
+    dataset_stats = {
+        "observation.state": {
+            "mean": torch.zeros(state_dim),
+            "std": torch.ones(state_dim),
+        },
+        "action": {
+            "min": torch.full((action_dim,), -1.0),
+            "max": torch.ones(action_dim),
+        },
+    }
+
+    # Create processors
+    preprocessor, postprocessor = make_multi_task_dit_pre_post_processors(
+        config=config, dataset_stats=dataset_stats
+    )
+
+    # Test preprocessor with sample data
+    batch = {
+        "observation.state": torch.randn(state_dim),
+        f"{OBS_IMAGES}.laptop": torch.rand(3, 224, 224),
+        ACTION: torch.randn(action_dim),
+        "task": "pick up the cube",
+    }
+
+    processed_batch = preprocessor(batch)
+
+    # Check that data is batched
+    assert processed_batch["observation.state"].shape == (1, state_dim)
+    assert processed_batch[f"{OBS_IMAGES}.laptop"].shape == (1, 3, 224, 224)
+    assert processed_batch[ACTION].shape == (1, action_dim)
+    # Check that task text was tokenized
+    assert OBS_LANGUAGE_TOKENS in processed_batch
+    assert OBS_LANGUAGE_ATTENTION_MASK in processed_batch
+    assert processed_batch[OBS_LANGUAGE_TOKENS].shape[0] == 1  # batch dimension
+    assert processed_batch[OBS_LANGUAGE_ATTENTION_MASK].shape[0] == 1  # batch dimension
+
+    # Check that data is on correct device
+    assert processed_batch["observation.state"].device.type == "cpu"
+    assert processed_batch[f"{OBS_IMAGES}.laptop"].device.type == "cpu"
+    assert processed_batch[ACTION].device.type == "cpu"
+
+    # Test postprocessor with sample action (PolicyAction is just a torch.Tensor)
+    action = torch.randn(1, action_dim)
+    processed_action = postprocessor(action)
+
+    # Check that action is unnormalized and on CPU
+    assert processed_action.shape == (1, action_dim)
+    assert processed_action.device.type == "cpu"
+
+
+def test_multi_task_dit_pre_post_processors_normalization():
+    """Test that normalization and unnormalization work correctly with simple sanity check numbers."""
+    state_dim = 3
+    action_dim = 2
+
+    config = create_config(
+        state_dim=state_dim,
+        action_dim=action_dim,
+        n_obs_steps=2,
+        horizon=16,
+        n_action_steps=8,
+    )
+    config.device = "cpu"
+
+    # Set normalization mode to match the stats we're providing
+    config.normalization_mapping = {
+        "VISUAL": NormalizationMode.IDENTITY,
+        "STATE": NormalizationMode.MEAN_STD,  # Use MEAN_STD since we provide mean/std stats
+        "ACTION": NormalizationMode.MIN_MAX,
+    }
+
+    # Use simple stats that will actually transform the values
+    dataset_stats = {
+        "observation.state": {
+            "mean": torch.full((state_dim,), 5.0),
+            "std": torch.full((state_dim,), 2.0),
+        },
+        "action": {
+            "min": torch.zeros(action_dim),
+            "max": torch.full((action_dim,), 2.0),
+        },
+    }
+
+    # Create processors
+    preprocessor, postprocessor = make_multi_task_dit_pre_post_processors(
+        config=config, dataset_stats=dataset_stats
+    )
+
+    # Use simple input values
+    input_state = torch.tensor([7.0, 5.0, 3.0])  # Will normalize to [1.0, 0.0, -1.0]
+    input_action = torch.tensor([1.0, 2.0])  # Will normalize to [0.0, 1.0]
+
+    batch = {
+        "observation.state": input_state,
+        f"{OBS_IMAGES}.laptop": torch.rand(3, 224, 224),
+        ACTION: input_action,
+        "task": "test task",
+    }
+
+    # Process through preprocessor
+    processed_batch = preprocessor(batch)
+
+    # State normalization: (x - mean) / std
+    expected_normalized_state = torch.tensor([1.0, 0.0, -1.0])
+    assert torch.allclose(processed_batch["observation.state"][0], expected_normalized_state, atol=1e-5)
+
+    # Action normalization: (x - min) / (max - min) * 2 - 1
+    expected_normalized_action = torch.tensor([0.0, 1.0])
+    assert torch.allclose(processed_batch[ACTION][0], expected_normalized_action, atol=1e-5)
+
+    # Test unnormalization: should recover original values
+    normalized_action_tensor = processed_batch[ACTION][0:1]  # Keep batch dimension
+    unnormalized_action = postprocessor(normalized_action_tensor)
+
+    # Should recover original action values
+    assert torch.allclose(unnormalized_action[0], input_action, atol=1e-4)
+
+
+@pytest.mark.parametrize("batch_size,state_dim,action_dim", [(2, 10, 10), (1, 6, 6)])
+def test_multi_task_dit_policy_select_action(batch_size: int, state_dim: int, action_dim: int):
+    """Test select_action (inference mode)."""
+    n_obs_steps = 2
+    horizon = 16
+    n_action_steps = 8
+
+    config = create_config(
+        state_dim=state_dim,
+        action_dim=action_dim,
+        n_obs_steps=n_obs_steps,
+        horizon=horizon,
+        n_action_steps=n_action_steps,
+    )
+
+    policy = MultiTaskDiTPolicy(config=config)
+    policy.eval()
+    policy.reset()  # Reset queues before inference
+
+    # Create processors - use IDENTITY normalization when no stats provided
+    config.normalization_mapping = {
+        "VISUAL": NormalizationMode.IDENTITY,
+        "STATE": NormalizationMode.IDENTITY,
+        "ACTION": NormalizationMode.IDENTITY,
+    }
+    preprocessor, postprocessor = make_multi_task_dit_pre_post_processors(config=config, dataset_stats=None)
+
+    with torch.no_grad():
+        observation_batch = create_observation_batch(batch_size=batch_size, state_dim=state_dim)
+        # Process observation through preprocessor
+        processed_obs = preprocessor(observation_batch)
+        selected_action = policy.select_action(processed_obs)
+        # Process action through postprocessor (PolicyAction is just a torch.Tensor)
+        processed_action = postprocessor(selected_action)
+        assert processed_action.shape == (batch_size, action_dim)
+
+
+def test_multi_task_dit_policy_diffusion_objective():
+    """Test policy with diffusion objective."""
+    batch_size = 2
+    state_dim = 10
+    action_dim = 10
+    n_obs_steps = 2
+    horizon = 16
+    n_action_steps = 8
+
+    input_features = {
+        OBS_STATE: PolicyFeature(type=FeatureType.STATE, shape=(state_dim,)),
+        f"{OBS_IMAGES}.laptop": PolicyFeature(type=FeatureType.VISUAL, shape=(3, 224, 224)),
+    }
+
+    config = MultiTaskDiTConfig(
+        input_features=input_features,
+        output_features={ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(action_dim,))},
+        n_obs_steps=n_obs_steps,
+        horizon=horizon,
+        n_action_steps=n_action_steps,
+        # Use diffusion objective
+        objective="diffusion",
+        noise_scheduler_type="DDPM",
+        num_train_timesteps=100,
+        num_inference_steps=10,
+        # Smaller model for tests
+        hidden_dim=128,
+        num_layers=2,
+        num_heads=4,
+    )
+    config.validate_features()
+
+    policy = MultiTaskDiTPolicy(config=config)
+    policy.train()
+
+    # Use preprocessor to handle tokenization
+    config.normalization_mapping = {
+        "VISUAL": NormalizationMode.IDENTITY,
+        "STATE": NormalizationMode.IDENTITY,
+        "ACTION": NormalizationMode.IDENTITY,
+    }
+    preprocessor, _ = make_multi_task_dit_pre_post_processors(config=config, dataset_stats=None)
+
+    batch = create_train_batch(
+        batch_size=batch_size,
+        n_obs_steps=n_obs_steps,
+        horizon=horizon,
+        state_dim=state_dim,
+        action_dim=action_dim,
+    )
+
+    # Process batch through preprocessor to tokenize task text
+    processed_batch = preprocessor(batch)
+
+    # Test forward pass
+    loss, _ = policy.forward(processed_batch)
+    assert loss is not None
+    assert loss.item() is not None
+
+    # Test inference
+    policy.eval()
+    # Use IDENTITY normalization when no stats provided
+    config.normalization_mapping = {
+        "VISUAL": NormalizationMode.IDENTITY,
+        "STATE": NormalizationMode.IDENTITY,
+        "ACTION": NormalizationMode.IDENTITY,
+    }
+    preprocessor, postprocessor = make_multi_task_dit_pre_post_processors(config=config, dataset_stats=None)
+    with torch.no_grad():
+        observation_batch = create_observation_batch(batch_size=batch_size, state_dim=state_dim)
+        # Process observation through preprocessor
+        processed_obs = preprocessor(observation_batch)
+        selected_action = policy.select_action(processed_obs)
+        # Process action through postprocessor (PolicyAction is just a torch.Tensor)
+        processed_action = postprocessor(selected_action)
+        assert processed_action.shape == (batch_size, action_dim)
+
+
+def test_multi_task_dit_policy_flow_matching_objective():
+    """Test policy with flow matching objective."""
+    batch_size = 2
+    state_dim = 10
+    action_dim = 10
+    n_obs_steps = 2
+    horizon = 16
+    n_action_steps = 8
+
+    input_features = {
+        OBS_STATE: PolicyFeature(type=FeatureType.STATE, shape=(state_dim,)),
+        f"{OBS_IMAGES}.laptop": PolicyFeature(type=FeatureType.VISUAL, shape=(3, 224, 224)),
+    }
+
+    config = MultiTaskDiTConfig(
+        input_features=input_features,
+        output_features={ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(action_dim,))},
+        n_obs_steps=n_obs_steps,
+        horizon=horizon,
+        n_action_steps=n_action_steps,
+        # Use flow matching objective
+        objective="flow_matching",
+        sigma_min=0.0,
+        num_integration_steps=10,  # Fewer steps for faster tests
+        integration_method="euler",
+        # Smaller model for tests
+        hidden_dim=128,
+        num_layers=2,
+        num_heads=4,
+    )
+    config.validate_features()
+
+    policy = MultiTaskDiTPolicy(config=config)
+    policy.train()
+
+    # Use preprocessor to handle tokenization
+    config.normalization_mapping = {
+        "VISUAL": NormalizationMode.IDENTITY,
+        "STATE": NormalizationMode.IDENTITY,
+        "ACTION": NormalizationMode.IDENTITY,
+    }
+    preprocessor, _ = make_multi_task_dit_pre_post_processors(config=config, dataset_stats=None)
+
+    batch = create_train_batch(
+        batch_size=batch_size,
+        n_obs_steps=n_obs_steps,
+        horizon=horizon,
+        state_dim=state_dim,
+        action_dim=action_dim,
+    )
+
+    # Process batch through preprocessor to tokenize task text
+    processed_batch = preprocessor(batch)
+
+    # Test forward pass
+    loss, _ = policy.forward(processed_batch)
+    assert loss is not None
+    assert loss.item() is not None
+
+    # Test inference
+    policy.eval()
+    # Use IDENTITY normalization when no stats provided
+    config.normalization_mapping = {
+        "VISUAL": NormalizationMode.IDENTITY,
+        "STATE": NormalizationMode.IDENTITY,
+        "ACTION": NormalizationMode.IDENTITY,
+    }
+    preprocessor, postprocessor = make_multi_task_dit_pre_post_processors(config=config, dataset_stats=None)
+    with torch.no_grad():
+        observation_batch = create_observation_batch(batch_size=batch_size, state_dim=state_dim)
+        # Process observation through preprocessor
+        processed_obs = preprocessor(observation_batch)
+        selected_action = policy.select_action(processed_obs)
+        # Process action through postprocessor (PolicyAction is just a torch.Tensor)
+        processed_action = postprocessor(selected_action)
+        assert processed_action.shape == (batch_size, action_dim)
+
+
+def test_multi_task_dit_policy_save_and_load(tmp_path):
+    """Test that the policy can be saved and loaded correctly."""
+    root = tmp_path / "test_multi_task_dit_save_and_load"
+
+    state_dim = 10
+    action_dim = 10
+    batch_size = 2
+    n_obs_steps = 2
+    horizon = 16
+    n_action_steps = 8
+
+    config = create_config(
+        state_dim=state_dim,
+        action_dim=action_dim,
+        n_obs_steps=n_obs_steps,
+        horizon=horizon,
+        n_action_steps=n_action_steps,
+    )
+
+    policy = MultiTaskDiTPolicy(config=config)
+    policy.eval()
+
+    # Get device before saving
+    device = next(policy.parameters()).device
+
+    policy.save_pretrained(root)
+    loaded_policy = MultiTaskDiTPolicy.from_pretrained(root, config=config)
+
+    # Explicitly move loaded_policy to the same device
+    loaded_policy.to(device)
+    loaded_policy.eval()
+
+    batch = create_train_batch(
+        batch_size=batch_size,
+        n_obs_steps=n_obs_steps,
+        horizon=horizon,
+        state_dim=state_dim,
+        action_dim=action_dim,
+    )
+
+    # Use preprocessor to handle tokenization
+    config.normalization_mapping = {
+        "VISUAL": NormalizationMode.IDENTITY,
+        "STATE": NormalizationMode.IDENTITY,
+        "ACTION": NormalizationMode.IDENTITY,
+    }
+    preprocessor, postprocessor = make_multi_task_dit_pre_post_processors(config=config, dataset_stats=None)
+
+    with torch.no_grad():
+        with seeded_context(12):
+            # Process batch through preprocessor
+            processed_batch = preprocessor(batch)
+            # Move batch to the same device as the policy
+            for key in processed_batch:
+                if isinstance(processed_batch[key], torch.Tensor):
+                    processed_batch[key] = processed_batch[key].to(device)
+            # Collect policy values before saving
+            loss, _ = policy.forward(processed_batch)
+
+            observation_batch = create_observation_batch(batch_size=batch_size, state_dim=state_dim)
+            # Process observation through preprocessor
+            processed_obs = preprocessor(observation_batch)
+            actions = policy.select_action(processed_obs)
+
+        with seeded_context(12):
+            # Process batch through preprocessor
+            processed_batch = preprocessor(batch)
+            # Collect policy values after loading
+            loaded_loss, _ = loaded_policy.forward(processed_batch)
+
+            loaded_observation_batch = create_observation_batch(batch_size=batch_size, state_dim=state_dim)
+            processed_obs = preprocessor(loaded_observation_batch)
+            loaded_actions = loaded_policy.select_action(processed_obs)
+
+        # Compare state dicts
+        assert policy.state_dict().keys() == loaded_policy.state_dict().keys()
+        for k in policy.state_dict():
+            assert torch.allclose(policy.state_dict()[k], loaded_policy.state_dict()[k], atol=1e-6)
+
+        # Compare values before and after saving and loading
+        assert torch.allclose(loss, loaded_loss)
+        assert torch.allclose(actions, loaded_actions)
+
+
+def test_multi_task_dit_policy_get_optim_params():
+    """Test that the policy returns correct optimizer parameter groups."""
+    config = create_config(
+        state_dim=10,
+        action_dim=10,
+        n_obs_steps=2,
+        horizon=16,
+        n_action_steps=8,
+    )
+
+    policy = MultiTaskDiTPolicy(config=config)
+    param_groups = policy.get_optim_params()
+
+    # Should have 2 parameter groups: non-vision and vision encoder
+    assert len(param_groups) == 2
+
+    # First group is non-vision params (no lr specified, will use default)
+    assert "params" in param_groups[0]
+    assert len(param_groups[0]["params"]) > 0
+
+    # Second group is vision encoder params with different lr
+    assert "params" in param_groups[1]
+    assert "lr" in param_groups[1]
+    expected_lr = config.optimizer_lr * config.vision_encoder_lr_multiplier
+    assert param_groups[1]["lr"] == expected_lr

tests/policies/pi0_fast/test_pi0_fast_original_vs_lerobot.py

@@ -17,7 +17,6 @@
 """Test script to verify PI0Fast policy integration with LeRobot vs the original implementation"""
 # ruff: noqa: E402
 
-import os
 import random
 from copy import deepcopy
 from typing import Any
@@ -28,10 +27,6 @@ import torch
 
 pytest.importorskip("transformers")
 pytest.importorskip("scipy")
-pytestmark = pytest.mark.skipif(
-    os.environ.get("CI") == "true" or os.environ.get("GITHUB_ACTIONS") == "true",
-    reason="This test requires accepting the model license",
-)
 
 from lerobot.policies.pi0_fast.configuration_pi0_fast import PI0FastConfig
 from lerobot.policies.pi0_fast.modeling_pi0_fast import PI0FastPolicy
@@ -45,7 +40,7 @@ from lerobot.utils.constants import (
     OBS_LANGUAGE_TOKENS,
     OBS_STATE,
 )  # noqa: E402
-from tests.utils import require_cuda  # noqa: E402
+from tests.utils import require_cuda, require_hf_token  # noqa: E402
 
 # Constants
 DUMMY_ACTION_DIM = 7
@@ -53,22 +48,24 @@ DUMMY_STATE_DIM = 20
 IMAGE_HEIGHT = 224
 IMAGE_WIDTH = 224
 NUM_VIEWS = 2  # Number of camera views
-DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
+DEVICE = "cuda"
 MODEL_PATH_LEROBOT = "lerobot/pi0fast-base"
 
 # Expected action token shape: (batch_size, max_decoding_steps)
 EXPECTED_ACTION_TOKENS_SHAPE = (1, 2)
 
 # Expected first 5 action tokens (for reproducibility check)
-EXPECTED_ACTION_TOKENS_FIRST_5 = torch.tensor([255657, 255362])
+EXPECTED_ACTION_TOKENS_FIRST_5 = torch.tensor([255020, 255589])
 
 # Expected actions after detokenization
 EXPECTED_ACTIONS_SHAPE = (1, 2, 32)  # (batch_size, n_action_steps, action_dim)
-EXPECTED_ACTIONS_MEAN = 0.04419417306780815
-EXPECTED_ACTIONS_STD = 0.26231569051742554
-EXPECTED_ACTIONS_FIRST_5 = torch.tensor([0.0000, 1.4849, 0.0000, 0.0000, 0.0000])
+EXPECTED_ACTIONS_MEAN = 0.046403881162405014
+EXPECTED_ACTIONS_STD = 0.2607129216194153
+EXPECTED_ACTIONS_FIRST_5 = torch.tensor([0.0000, 0.3536, 0.0707, 0.0000, 0.0000])
 
 
+@require_cuda
+@require_hf_token
 def set_seed_all(seed: int):
     """Set random seed for all RNG sources to ensure reproducibility."""
     random.seed(seed)
@@ -85,6 +82,8 @@ def set_seed_all(seed: int):
     torch.use_deterministic_algorithms(True, warn_only=True)
 
 
+@require_cuda
+@require_hf_token
 def instantiate_lerobot_pi0_fast(
     from_pretrained: bool = False,
     model_path: str = MODEL_PATH_LEROBOT,
@@ -127,6 +126,8 @@ def instantiate_lerobot_pi0_fast(
     return policy, preprocessor, postprocessor
 
 
+@require_cuda
+@require_hf_token
 def create_dummy_data(device=DEVICE):
     """Create dummy data for testing both implementations."""
     batch_size = 1
@@ -158,28 +159,35 @@ def create_dummy_data(device=DEVICE):
 
 # Pytest fixtures
 @pytest.fixture(scope="module")
+@require_cuda
+@require_hf_token
 def pi0_fast_components():
     """Fixture to instantiate and provide all PI0Fast components for tests."""
     print(f"\nTesting with DEVICE='{DEVICE}'")
     print("\n[Setup] Instantiating LeRobot PI0Fast policy...")
     policy_obj, preprocessor_obj, postprocessor_obj = instantiate_lerobot_pi0_fast(from_pretrained=True)
     print("Model loaded successfully")
-    yield policy_obj, preprocessor_obj, postprocessor_obj
+    return policy_obj, preprocessor_obj, postprocessor_obj
 
 
 @pytest.fixture(scope="module")
+@require_cuda
+@require_hf_token
 def policy(pi0_fast_components):
     """Fixture to provide the PI0Fast policy for tests."""
     return pi0_fast_components[0]
 
 
 @pytest.fixture(scope="module")
+@require_cuda
+@require_hf_token
 def preprocessor(pi0_fast_components):
     """Fixture to provide the PI0Fast preprocessor for tests."""
     return pi0_fast_components[1]
 
 
 @require_cuda
+@require_hf_token
 def test_pi0_fast_preprocessor_alignment(policy, preprocessor):
     """Test that LeRobot PI0Fast preprocessor produces expected outputs."""
     print("\n" + "=" * 80)
@@ -227,6 +235,7 @@ def test_pi0_fast_preprocessor_alignment(policy, preprocessor):
 
 
 @require_cuda
+@require_hf_token
 def test_pi0_fast_action_generation(policy, preprocessor):
     """Test PI0Fast LeRobot implementation generates expected actions."""
     print("\n" + "=" * 80)
@@ -305,6 +314,7 @@ def test_pi0_fast_action_generation(policy, preprocessor):
 
 
 @require_cuda
+@require_hf_token
 def test_pi0_fast_inference_reproducibility(policy, preprocessor):
     """Test that PI0Fast inference is reproducible with the same seed."""
     print("\n" + "=" * 80)
@@ -346,6 +356,7 @@ def test_pi0_fast_inference_reproducibility(policy, preprocessor):
 
 
 @require_cuda
+@require_hf_token
 def test_pi0_fast_forward_pass_logits(policy, preprocessor):
     """Test PI0Fast forward pass and compare logits against expected values."""
     print("\n" + "=" * 80)
@@ -395,6 +406,7 @@ def test_pi0_fast_forward_pass_logits(policy, preprocessor):
 
 
 @require_cuda
+@require_hf_token
 def test_pi0_fast_action_token_sampling(policy, preprocessor):
     """Test PI0Fast action token sampling (autoregressive decoding)."""
     print("\n" + "=" * 80)
@@ -451,6 +463,7 @@ def test_pi0_fast_action_token_sampling(policy, preprocessor):
 
 
 @require_cuda
+@require_hf_token
 def test_pi0_fast_detokenization(policy, preprocessor):
     """Test PI0Fast action detokenization (FAST decoding)."""
     print("\n" + "=" * 80)

tests/policies/pi0_pi05/test_pi0.py

@@ -14,18 +14,12 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-"""Test script to verify PI0 policy integration with LeRobot, only meant to be run locally!"""
-
-import os
+"""Test script to verify PI0 policy integration with LeRobot"""
 
 import pytest
 import torch
 
-# Skip this entire module in CI
-pytestmark = pytest.mark.skipif(
-    os.environ.get("CI") == "true" or os.environ.get("GITHUB_ACTIONS") == "true",
-    reason="This test requires local OpenPI installation and is not meant for CI",
-)
+pytest.importorskip("transformers")
 
 from lerobot.policies.factory import make_policy_config  # noqa: E402
 from lerobot.policies.pi0 import (  # noqa: E402
@@ -34,10 +28,11 @@ from lerobot.policies.pi0 import (  # noqa: E402
     make_pi0_pre_post_processors,  # noqa: E402
 )
 from lerobot.utils.random_utils import set_seed  # noqa: E402
-from tests.utils import require_cuda  # noqa: E402
+from tests.utils import require_cuda, require_hf_token  # noqa: E402
 
 
 @require_cuda
+@require_hf_token
 def test_policy_instantiation():
     # Create config
     set_seed(42)
@@ -114,6 +109,7 @@ def test_policy_instantiation():
 
 
 @require_cuda
+@require_hf_token
 def test_config_creation():
     """Test policy config creation through factory."""
     try:

tests/policies/pi0_pi05/test_pi05.py

@@ -14,20 +14,12 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-"""Test script to verify PI0.5 (pi05) support in PI0 policy, only meant to be run locally!"""
-
-import os
+"""Test script to verify PI0.5 (pi05) support in PI0 policy"""
 
 import pytest
 import torch
 
-from lerobot.utils.random_utils import set_seed
-
-# Skip this entire module in CI
-pytestmark = pytest.mark.skipif(
-    os.environ.get("CI") == "true" or os.environ.get("GITHUB_ACTIONS") == "true",
-    reason="This test requires local OpenPI installation and is not meant for CI",
-)
+pytest.importorskip("transformers")
 
 from lerobot.policies.factory import make_policy_config  # noqa: E402
 from lerobot.policies.pi05 import (  # noqa: E402
@@ -35,10 +27,12 @@ from lerobot.policies.pi05 import (  # noqa: E402
     PI05Policy,
     make_pi05_pre_post_processors,  # noqa: E402
 )
-from tests.utils import require_cuda  # noqa: E402
+from lerobot.utils.random_utils import set_seed
+from tests.utils import require_cuda, require_hf_token  # noqa: E402
 
 
 @require_cuda
+@require_hf_token
 def test_policy_instantiation():
     # Create config
     set_seed(42)
@@ -151,6 +145,7 @@ def test_policy_instantiation():
 
 
 @require_cuda
+@require_hf_token
 def test_config_creation():
     """Test policy config creation through factory."""
     try:

tests/policies/pi0_pi05/test_pi05_original_vs_lerobot.py

@@ -14,7 +14,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-"""Test script to verify PI0OpenPI policy integration with LeRobot vs the original implementation, only meant to be run locally!"""
+"""Test script to verify PI0OpenPI policy integration with LeRobot vs the original implementation"""
 
 import os
 from copy import deepcopy

tests/policies/pi0_pi05/test_pi05_rtc.py

@@ -24,9 +24,10 @@ import torch
 # Skip this entire module in CI
 pytestmark = pytest.mark.skipif(
     os.environ.get("CI") == "true" or os.environ.get("GITHUB_ACTIONS") == "true",
-    reason="This test requires local OpenPI installation and is not meant for CI",
+    reason="TODO: This test seems to hang the CI",
 )
 
+
 from lerobot.configs.types import FeatureType, PolicyFeature, RTCAttentionSchedule  # noqa: E402
 from lerobot.policies.pi05 import PI05Config, PI05Policy, make_pi05_pre_post_processors  # noqa: E402
 from lerobot.policies.rtc.configuration_rtc import RTCConfig  # noqa: E402

tests/policies/pi0_pi05/test_pi0_original_vs_lerobot.py

@@ -14,7 +14,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-"""Test script to verify PI0 policy integration with LeRobot vs the original implementation, only meant to be run locally!"""
+"""Test script to verify PI0 policy integration with LeRobot vs the original implementation"""
 
 import os
 from copy import deepcopy

tests/policies/pi0_pi05/test_pi0_rtc.py

@@ -24,9 +24,10 @@ import torch
 # Skip this entire module in CI
 pytestmark = pytest.mark.skipif(
     os.environ.get("CI") == "true" or os.environ.get("GITHUB_ACTIONS") == "true",
-    reason="This test requires local OpenPI installation and is not meant for CI",
+    reason="TODO: This test seems to hang the CI",
 )
 
+
 from lerobot.configs.types import FeatureType, PolicyFeature, RTCAttentionSchedule  # noqa: E402
 from lerobot.policies.pi0 import PI0Config, PI0Policy, make_pi0_pre_post_processors  # noqa: E402
 from lerobot.policies.rtc.configuration_rtc import RTCConfig  # noqa: E402
@@ -88,6 +89,7 @@ def test_pi0_rtc_initialization_without_rtc_config():
     print("✓ PI0 RTC initialization without RTC config: Test passed")
 
 
+@require_cuda
 def test_pi0_rtc_inference_with_prev_chunk():
     """Test PI0 policy inference with RTC and previous chunk."""
     set_seed(42)

tests/policies/test_policies.py

@@ -143,12 +143,18 @@ def test_policy(ds_repo_id, env_name, env_kwargs, policy_name, policy_kwargs):
     Note: We test various combinations of policy and dataset. The combinations are by no means exhaustive,
           and for now we add tests as we see fit.
     """
+    if policy_name == "vqbet" and DEVICE == "mps":
+        pytest.skip("VQBet does not support MPS backend")
+    if policy_name == "act" and "aloha" in ds_repo_id and DEVICE == "mps":
+        pytest.skip("ACT with aloha has batch mutation issues on MPS")
+
     train_cfg = TrainPipelineConfig(
         # TODO(rcadene, aliberts): remove dataset download
         dataset=DatasetConfig(repo_id=ds_repo_id, episodes=[0]),
         policy=make_policy_config(policy_name, push_to_hub=False, **policy_kwargs),
         env=make_env_config(env_name, **env_kwargs),
     )
+    train_cfg.policy.device = DEVICE
     train_cfg.validate()
 
     # Check that we can make the policy object.
@@ -227,6 +233,7 @@ def test_act_backbone_lr():
         dataset=DatasetConfig(repo_id="lerobot/aloha_sim_insertion_scripted", episodes=[0]),
         policy=make_policy_config("act", optimizer_lr=0.01, optimizer_lr_backbone=0.001, push_to_hub=False),
     )
+    cfg.policy.device = DEVICE
     cfg.validate()  # Needed for auto-setting some parameters
 
     assert cfg.policy.optimizer_lr == 0.01