fix(num_frames): fixing redundant frames count in conversion script (#3091)

* fix(ci): log in HF for gated repo in nightly workflows

* fix(ci): add env var

* fix(ci): remove 10 min limit for multi-gpu nightly

.github/workflows/fast_tests.yml

@@ -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 +90,10 @@ jobs:
       - name: Install lerobot with test extras
         run: uv sync --extra "test"
 
+      - name: Login to Hugging Face
+        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

@@ -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 +88,11 @@ jobs:
       - name: Install lerobot with all extras
         run: uv sync --extra all # TODO(Steven): Make flash-attn optional
 
+      - name: Login to Hugging Face
+        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 +168,7 @@ 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: ${{ needs.build-and-push-docker.outputs.image_tag }} # zizmor: ignore[unpinned-images]
       options: --gpus all --shm-size "16gb"
@@ -173,6 +180,10 @@ jobs:
         shell: bash
         working-directory: /lerobot
     steps:
+      - name: Login to Hugging Face
+        run: |
+          hf auth login --token "$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
       - name: Run pytest on GPU

.github/workflows/nightly.yml

@@ -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: ${{ needs.build-docker-cpu-nightly.outputs.image_tag }} # zizmor: ignore[unpinned-images]
       options: --shm-size "16gb"
@@ -130,6 +131,10 @@ jobs:
         shell: bash
         working-directory: /lerobot
     steps:
+      - name: Login to Hugging Face
+        run: |
+          hf auth login --token "$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 +151,7 @@ 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: ${{ needs.build-docker-gpu-nightly.outputs.image_tag }} # zizmor: ignore[unpinned-images]
       options: --gpus all --shm-size "16gb"
@@ -157,6 +163,10 @@ jobs:
         shell: bash
         working-directory: /lerobot
     steps:
+      - name: Login to Hugging Face
+        run: |
+          hf auth login --token "$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 +184,7 @@ 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: ${{ needs.build-docker-gpu-nightly.outputs.image_tag }} # zizmor: ignore[unpinned-images]
       options: --gpus all --shm-size "16gb"
@@ -185,6 +196,10 @@ jobs:
         shell: bash
         working-directory: /lerobot
     steps:
+      - name: Login to Hugging Face
+        run: |
+          hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
+          hf auth whoami
       - name: Verify GPU availability
         run: |
           nvidia-smi
@@ -193,4 +208,3 @@ jobs:
       - 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

.github/workflows/unbound_deps_tests.yml

@@ -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 +80,10 @@ jobs:
 
       - name: Install lerobot with all extras
         run: uv sync --extra all # TODO(Steven): Make flash-attn optional
-
+      - name: Login to Hugging Face
+        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 +141,7 @@ 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: ${{ needs.build-and-push-docker.outputs.image_tag }} # zizmor: ignore[unpinned-images]
       options: --gpus all --shm-size "16gb"
@@ -148,6 +153,10 @@ jobs:
         shell: bash
         working-directory: /lerobot
     steps:
+      - name: Login to Hugging Face
+        run: |
+          hf auth login --token "$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

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
 

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/earthrover_mini_plus.mdx

@@ -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/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/pi0fast.mdx

@@ -52,7 +52,7 @@ This approach can transform **any existing VLM** into a VLA by training it to pr
 
 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 +114,15 @@ lerobot-train \
 
 ### Key Training Parameters
 
-| Parameter                              | Description                                        | Default                      |
-| -------------------------------------- | -------------------------------------------------- | ---------------------------- |
-| `--policy.gradient_checkpointing=true` | Reduces memory usage significantly during training | `false`                      |
-| `--policy.dtype=bfloat16`              | Use mixed precision training for efficiency        | `float32`                    |
-| `--policy.chunk_size`                  | Number of action steps to predict (action horizon) | `50`                         |
-| `--policy.n_action_steps`              | Number of action steps to execute                  | `50`                         |
-| `--policy.max_action_tokens`           | Maximum number of FAST tokens per action chunk     | `256`                        |
-| `--policy.action_tokenizer_name`       | FAST tokenizer to use                              | `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
 

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

pyproject.toml

@@ -25,7 +25,7 @@ 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" }
@@ -61,7 +61,7 @@ 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[cli]>=1.0.0,<2.0.0",
     "accelerate>=1.10.0,<2.0.0",
 
     # Core dependencies
@@ -96,9 +96,12 @@ 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"]
+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"]
 
 # Motors
 feetech = ["feetech-servo-sdk>=1.0.0,<2.0.0"]
@@ -129,17 +132,17 @@ phone = ["hebi-py>=2.8.0,<2.12.0", "teleop>=0.1.0,<0.2.0", "fastapi<1.0"]
 
 # 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"]
 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 +151,13 @@ 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", "matplotlib>=3.10.3,<4.0.0", "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"]
+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"]
@@ -176,8 +179,8 @@ 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]",
@@ -397,85 +400,3 @@ 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/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

@@ -89,8 +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")
@@ -152,7 +152,7 @@ def split_dataset(
         dataset: The source LeRobotDataset to split.
         splits: Either a dict mapping split names to episode indices, or a dict mapping
                 split names to fractions (must sum to <= 1.0).
-        output_dir: 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
@@ -243,8 +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")
@@ -288,8 +288,8 @@ def modify_features(
         dataset: The source LeRobotDataset.
         add_features: Optional dict mapping feature names to (feature_values, feature_info) tuples.
         remove_features: Optional feature name(s) to remove. Can be a single string or list.
-        output_dir: 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.
@@ -390,8 +390,8 @@ def add_features(
     Args:
         dataset: The source LeRobotDataset.
         features: Dictionary mapping feature names to (feature_values, feature_info) tuples.
-        output_dir: 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.
@@ -427,8 +427,8 @@ def remove_feature(
     Args:
         dataset: The source LeRobotDataset.
         feature_names: Name(s) of features to remove. Can be a single string or list.
-        output_dir: 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.
@@ -1475,7 +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}")
@@ -1529,7 +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",
@@ -1548,8 +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)
@@ -1600,6 +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,

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

src/lerobot/datasets/utils.py

@@ -341,6 +341,7 @@ 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
 
 

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/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/groot/eagle2_hg_model/image_processing_eagle2_5_vl_fast.py

@@ -14,7 +14,7 @@ from transformers.image_processing_utils import (
 )
 from transformers.image_processing_utils_fast import (
     BaseImageProcessorFast,
-    DefaultFastImageProcessorKwargs,
+    ImagesKwargs,
     group_images_by_shape,
     reorder_images,
 )
@@ -77,7 +77,7 @@ def crop(img: torch.Tensor, left: int, top: int, right: int, bottom: int) -> tor
     return img[:, top:bottom, left:right]
 
 
-class Eagle25VLFastImageProcessorKwargs(DefaultFastImageProcessorKwargs):
+class Eagle25VLFastImageProcessorKwargs(ImagesKwargs):
     max_dynamic_tiles: int | None
     min_dynamic_tiles: int | None
     use_thumbnail: bool | None

src/lerobot/policies/pi0/modeling_pi0.py

@@ -15,6 +15,7 @@
 # limitations under the License.
 
 import builtins
+import copy
 import logging
 import math
 from collections import deque
@@ -32,13 +33,21 @@ 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 +200,7 @@ 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 +211,7 @@ 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 +230,14 @@ 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 +263,10 @@ 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 +274,7 @@ 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 +286,15 @@ 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 +367,7 @@ 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 +375,7 @@ 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 +386,13 @@ 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 +407,11 @@ 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 +423,8 @@ 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 +434,23 @@ 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 +464,7 @@ 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 +488,7 @@ 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 +528,7 @@ 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 +594,19 @@ 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 +768,7 @@ 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 +842,7 @@ 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 +916,7 @@ 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,
@@ -995,23 +1004,14 @@ class PI0Policy(PreTrainedPolicy):
 
         # Initialize model without loading weights
         # Check if dataset_stats were provided in kwargs
-        if _transformers_available:
-            from transformers.modeling_utils import no_init_weights
+        model = cls(config, **kwargs)
 
-            with no_init_weights():
-                model = cls(config, **kwargs)
-            model.model.paligemma_with_expert.paligemma.tie_weights()
-        else:
-            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",
@@ -1019,7 +1019,7 @@ 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),
                 )
@@ -1032,7 +1032,7 @@ 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
@@ -1077,7 +1077,7 @@ 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
 
@@ -1127,6 +1127,14 @@ 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

@@ -15,6 +15,7 @@
 # limitations under the License.
 
 import builtins
+import copy
 import logging
 import math
 from collections import deque
@@ -32,14 +33,20 @@ 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 +99,11 @@ 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 +197,7 @@ 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 +208,7 @@ 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 +227,14 @@ 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 +260,10 @@ 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 +271,7 @@ 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 +283,15 @@ 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 +364,7 @@ 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 +372,7 @@ 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 +383,13 @@ 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 +404,11 @@ 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 +420,8 @@ 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 +431,23 @@ 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 +461,7 @@ 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 +485,7 @@ 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 +525,7 @@ 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 +590,19 @@ 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 +744,7 @@ 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 +815,7 @@ 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 +887,7 @@ 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,
@@ -967,23 +975,14 @@ class PI05Policy(PreTrainedPolicy):
 
         # Initialize model without loading weights
         # Check if dataset_stats were provided in kwargs
-        if _transformers_available:
-            from transformers.modeling_utils import no_init_weights
+        model = cls(config, **kwargs)
 
-            with no_init_weights():
-                model = cls(config, **kwargs)
-            model.model.paligemma_with_expert.paligemma.tie_weights()
-        else:
-            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",
@@ -991,7 +990,7 @@ 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),
                 )
@@ -1004,7 +1003,7 @@ 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
@@ -1016,8 +1015,6 @@ 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
 
@@ -1051,7 +1048,7 @@ 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
 
@@ -1105,6 +1102,14 @@ 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

@@ -38,11 +38,16 @@ 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 +126,7 @@ 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 +137,7 @@ 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 +211,22 @@ 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 +239,11 @@ 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 +254,18 @@ 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 +279,7 @@ 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 +326,14 @@ 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 +342,8 @@ 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 +533,7 @@ 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 +626,7 @@ 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 +724,7 @@ 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)
@@ -895,23 +905,14 @@ class PI0FastPolicy(PreTrainedPolicy):
 
         # Initialize model without loading weights
         # Check if dataset_stats were provided in kwargs
-        if _transformers_available:
-            from transformers.modeling_utils import no_init_weights
+        model = cls(config, **kwargs)
 
-            with no_init_weights():
-                model = cls(config, **kwargs)
-            model.model.paligemma_with_expert.paligemma.tie_weights()
-        else:
-            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",
@@ -919,7 +920,7 @@ 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),
                 )
@@ -932,8 +933,9 @@ 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
@@ -943,8 +945,6 @@ 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
 
@@ -978,7 +978,7 @@ 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/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

@@ -593,6 +593,12 @@ 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/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/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/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 \
@@ -167,6 +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")
@@ -200,36 +233,46 @@ class ConvertImageToVideoConfig(OperationConfig):
 @OperationConfig.register_subclass("info")
 @dataclass
 class InfoConfig(OperationConfig):
-    type: str = "info"
     show_features: bool = False
 
 
 @dataclass
 class EditDatasetConfig:
-    repo_id: str
+    # 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:
@@ -241,11 +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(
@@ -272,19 +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()
 
 
@@ -295,18 +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,
     )
 
@@ -316,7 +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()
 
 
@@ -329,11 +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(
@@ -361,9 +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.")
@@ -399,32 +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")
 
@@ -499,8 +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":

src/lerobot/scripts/lerobot_record.py

@@ -155,7 +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

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_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/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/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/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
@@ -53,22 +48,23 @@ 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
 def set_seed_all(seed: int):
     """Set random seed for all RNG sources to ensure reproducibility."""
     random.seed(seed)
@@ -85,6 +81,7 @@ def set_seed_all(seed: int):
     torch.use_deterministic_algorithms(True, warn_only=True)
 
 
+@require_cuda
 def instantiate_lerobot_pi0_fast(
     from_pretrained: bool = False,
     model_path: str = MODEL_PATH_LEROBOT,
@@ -127,6 +124,7 @@ def instantiate_lerobot_pi0_fast(
     return policy, preprocessor, postprocessor
 
 
+@require_cuda
 def create_dummy_data(device=DEVICE):
     """Create dummy data for testing both implementations."""
     batch_size = 1
@@ -158,22 +156,25 @@ def create_dummy_data(device=DEVICE):
 
 # Pytest fixtures
 @pytest.fixture(scope="module")
+@require_cuda
 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
 def policy(pi0_fast_components):
     """Fixture to provide the PI0Fast policy for tests."""
     return pi0_fast_components[0]
 
 
 @pytest.fixture(scope="module")
+@require_cuda
 def preprocessor(pi0_fast_components):
     """Fixture to provide the PI0Fast preprocessor for tests."""
     return pi0_fast_components[1]

tests/policies/pi0_pi05/test_pi0.py

@@ -16,17 +16,8 @@
 
 """Test script to verify PI0 policy integration with LeRobot, only meant to be run locally!"""
 
-import os
-
-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",
-)
-
 from lerobot.policies.factory import make_policy_config  # noqa: E402
 from lerobot.policies.pi0 import (  # noqa: E402
     PI0Config,

tests/policies/pi0_pi05/test_pi05.py

@@ -16,25 +16,15 @@
 
 """Test script to verify PI0.5 (pi05) support in PI0 policy, only meant to be run locally!"""
 
-import os
-
-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",
-)
-
 from lerobot.policies.factory import make_policy_config  # noqa: E402
 from lerobot.policies.pi05 import (  # noqa: E402
     PI05Config,
     PI05Policy,
     make_pi05_pre_post_processors,  # noqa: E402
 )
+from lerobot.utils.random_utils import set_seed
 from tests.utils import require_cuda  # noqa: E402
 
 

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_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_sac_policy.py

@@ -305,6 +305,9 @@ def test_sac_policy_with_visual_input(batch_size: int, state_dim: int, action_di
     [(1, 6, 6, "helper2424/resnet10"), (1, 6, 6, "facebook/convnext-base-224")],
 )
 @pytest.mark.skipif(not TRANSFORMERS_AVAILABLE, reason="Transformers are not installed")
+@pytest.mark.skip(
+    reason="helper2424/resnet10 needs to be updated to work with the latest version of transformers"
+)
 def test_sac_policy_with_pretrained_encoder(
     batch_size: int, state_dim: int, action_dim: int, vision_encoder_name: str
 ):

tests/policies/wall_x/test_wallx.py

@@ -16,8 +16,6 @@
 
 """Test script to verify Wall-X policy integration with LeRobot, only meant to be run locally!"""
 
-import os
-
 import pytest
 import torch
 
@@ -26,19 +24,15 @@ pytest.importorskip("peft")
 pytest.importorskip("transformers")
 pytest.importorskip("torchdiffeq")
 
-# 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 Wall-X installation and is not meant for CI",
-)
-
 from lerobot.policies.factory import make_policy_config  # noqa: E402
 from lerobot.policies.wall_x import WallXConfig  # noqa: E402
 from lerobot.policies.wall_x.modeling_wall_x import WallXPolicy  # noqa: E402
 from lerobot.policies.wall_x.processor_wall_x import make_wall_x_pre_post_processors  # noqa: E402
 from lerobot.utils.random_utils import set_seed  # noqa: E402
+from tests.utils import require_cuda  # noqa: E402
 
 
+@require_cuda
 def test_policy_instantiation():
     # Create config
     set_seed(42)
@@ -123,6 +117,7 @@ def test_policy_instantiation():
         raise
 
 
+@require_cuda
 def test_config_creation():
     """Test policy config creation through factory."""
     try:
@@ -134,8 +129,3 @@ def test_config_creation():
     except Exception as e:
         print(f"Config creation failed: {e}")
         raise
-
-
-if __name__ == "__main__":
-    test_policy_instantiation()
-    test_config_creation()

tests/scripts/test_edit_dataset_parsing.py

@@ -27,6 +27,7 @@ from lerobot.scripts.lerobot_edit_dataset import (
     OperationConfig,
     RemoveFeatureConfig,
     SplitConfig,
+    _validate_config,
 )
 
 
@@ -51,11 +52,23 @@ class TestOperationTypeParsing:
         ],
     )
     def test_operation_type_resolves_correct_class(self, type_name, expected_cls):
-        cfg = parse_cfg(["--repo_id", "test/repo", "--operation.type", type_name])
+        cfg = parse_cfg(
+            ["--repo_id", "test/repo", "--new_repo_id", "test/merged", "--operation.type", type_name]
+        )
         assert isinstance(cfg.operation, expected_cls), (
             f"Expected {expected_cls.__name__}, got {type(cfg.operation).__name__}"
         )
 
+    def test_merge_requires_new_repo_id(self):
+        cfg = parse_cfg(["--operation.type", "merge"])
+        with pytest.raises(ValueError, match="--new_repo_id is required for merge"):
+            _validate_config(cfg)
+
+    def test_non_merge_requires_repo_id(self):
+        cfg = parse_cfg(["--operation.type", "delete_episodes"])
+        with pytest.raises(ValueError, match="--repo_id is required for delete_episodes"):
+            _validate_config(cfg)
+
     @pytest.mark.parametrize(
         "type_name, expected_cls",
         [
@@ -69,6 +82,8 @@ class TestOperationTypeParsing:
         ],
     )
     def test_get_choice_name_roundtrips(self, type_name, expected_cls):
-        cfg = parse_cfg(["--repo_id", "test/repo", "--operation.type", type_name])
+        cfg = parse_cfg(
+            ["--repo_id", "test/repo", "--new_repo_id", "test/merged", "--operation.type", type_name]
+        )
         resolved_name = OperationConfig.get_choice_name(type(cfg.operation))
         assert resolved_name == type_name

tests/utils/test_logging_utils.py

@@ -24,6 +24,11 @@ def mock_metrics():
     return {"loss": AverageMeter("loss", ":.3f"), "accuracy": AverageMeter("accuracy", ":.2f")}
 
 
+class MockAccelerator:
+    def __init__(self, num_processes: int):
+        self.num_processes = num_processes
+
+
 def test_average_meter_initialization():
     meter = AverageMeter("loss", ":.2f")
     assert meter.name == "loss"
@@ -82,6 +87,37 @@ def test_metrics_tracker_step(mock_metrics):
     assert tracker.epochs == tracker.samples / 1000
 
 
+def test_metrics_tracker_initialization_with_accelerator(mock_metrics):
+    tracker = MetricsTracker(
+        batch_size=32,
+        num_frames=1000,
+        num_episodes=50,
+        metrics=mock_metrics,
+        initial_step=10,
+        accelerator=MockAccelerator(num_processes=2),
+    )
+    assert tracker.steps == 10
+    assert tracker.samples == 10 * 32 * 2
+    assert tracker.episodes == tracker.samples / (1000 / 50)
+    assert tracker.epochs == tracker.samples / 1000
+
+
+def test_metrics_tracker_step_with_accelerator(mock_metrics):
+    tracker = MetricsTracker(
+        batch_size=32,
+        num_frames=1000,
+        num_episodes=50,
+        metrics=mock_metrics,
+        initial_step=5,
+        accelerator=MockAccelerator(num_processes=2),
+    )
+    tracker.step()
+    assert tracker.steps == 6
+    assert tracker.samples == (5 * 32 * 2) + (32 * 2)
+    assert tracker.episodes == tracker.samples / (1000 / 50)
+    assert tracker.epochs == tracker.samples / 1000
+
+
 def test_metrics_tracker_getattr(mock_metrics):
     tracker = MetricsTracker(batch_size=32, num_frames=1000, num_episodes=50, metrics=mock_metrics)
     assert tracker.loss == mock_metrics["loss"]