feat(lekiwi): support LeKiwi in lerobot-replay CLI (#3739)

Register the `lekiwi` robot module in `lerobot_replay.py` so episodes can be
replayed on a LeKiwi via `--robot.type=lekiwi_client`. The module is already
registered in `lerobot_calibrate.py` and `lerobot_setup_motors.py`; this fills
the gap so the replay CLI recognizes the same robot.

Replayed actions are loaded from the dataset as torch tensors, which
`json.dumps` cannot serialize when `LeKiwiClient.send_action` ships them over
ZMQ. Coerce each action value to a plain float before sending. This is scoped
to the LeKiwi network client and does not affect any other robot.

Co-authored-by: Steven Palma <imstevenpmwork@ieee.org>

.github/dependabot.yml

@@ -0,0 +1,11 @@
+version: 2
+updates:
+  - package-ecosystem: "github-actions"
+    directory: "/"
+    schedule:
+      interval: "weekly"
+    cooldown:
+      default-days: 7
+    groups:
+      actions:
+        patterns: ["*"]

AGENTS.md

@@ -51,6 +51,7 @@ pre-commit run --all-files                           # Lint + format (ruff, typo
 ## Notes
 
 - **Mypy is gradual**: strict only for `lerobot.envs`, `lerobot.configs`, `lerobot.optim`, `lerobot.model`, `lerobot.cameras`, `lerobot.motors`, `lerobot.transport`. Add type annotations when modifying these modules.
-- **Optional dependencies**: many policies, envs, and robots are behind extras (e.g., `lerobot[aloha]`). New imports for optional packages must be guarded or lazy. See `pyproject.toml [project.optional-dependencies]`.
+- **Imports**: prefer top-level imports; relative (`from .sibling import X`) across sibling files within a module, absolute (`from lerobot.module import X`) across modules.
+- **Optional dependencies**: many policies, envs, and robots are behind extras (e.g., `lerobot[aloha]`, see `pyproject.toml`). Guard optional imports with `TYPE_CHECKING or _foo_available` at module top + a `require_package(...)` check at use time. Reuse the `_foo_available` flags in `utils/import_utils.py`; don't call `is_package_available`.
 - **Video decoding**: datasets can store observations as video files. `LeRobotDataset` handles frame extraction, but tests need ffmpeg installed.
 - **Prioritize use of `uv run`** to execute Python commands (not raw `python` or `pip`).

docs/source/bring_your_own_policies.mdx

@@ -165,6 +165,8 @@ Batches are flat dictionaries keyed by the constants in [`lerobot.utils.constant
 
 LeRobot uses `PolicyProcessorPipeline`s to normalize inputs and de-normalize outputs around your policy. For a concrete reference, see [`processor_act.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/act/processor_act.py) or [`processor_diffusion.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/diffusion/processor_diffusion.py).
 
+Pay close attention here: processors are the most common reproducibility pain point. A mismatch in normalization mode (`IDENTITY` vs `MEAN_STD` vs `MIN_MAX` vs `QUANTILES`/`QUANTILE10`) or in which features get normalized will train and eval without erroring, yet silently wreck results. Make sure the modes match how the checkpoint was trained, that the required stats exist (e.g. `QUANTILES` needs `q01`/`q99`), and that the pre- and post-processors stay consistent.
+
 ```python
 # processor_my_policy.py
 from typing import Any
@@ -304,7 +306,9 @@ Mirror an existing policy that's structurally similar to yours; the diff is smal
 
 ### Heavy / optional dependencies
 
-Most policies need a heavy backbone (transformers, diffusers, a specific VLM SDK). The convention is **two-step gating**: a `TYPE_CHECKING`-guarded import at module top, and a `require_package` runtime check in the constructor. [`modeling_diffusion.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/diffusion/modeling_diffusion.py) is the canonical reference:
+Most policies need a heavy backbone (transformers, diffusers, a specific VLM SDK). Wherever one exists, prefer loading it e.g from `transformers` or `diffusers` rather than re-implementing the architecture in-tree.
+
+The convention is **two-step gating**: a `TYPE_CHECKING`-guarded import at module top, and a `require_package` runtime check in the constructor. [`modeling_diffusion.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/diffusion/modeling_diffusion.py) is the canonical reference:
 
 ```python
 from typing import TYPE_CHECKING
@@ -374,6 +378,7 @@ The general expectations are in [`CONTRIBUTING.md`](https://github.com/huggingfa
 - [ ] Optional deps live behind a `[project.optional-dependencies]` extra and the `TYPE_CHECKING + require_package` guard.
 - [ ] `tests/policies/` updated; backward-compat artifact committed & policy-specific tests.
 - [ ] `src/lerobot/policies/<name>/README.md` symlinked into `docs/source/policy_<name>_README.md`; user-facing `docs/source/<name>.mdx` written and added to `_toctree.yml`.
+- [ ] `lerobot-train --policy.type my_policy ...` runs end-to-end for at least a few steps + save a checkpoint that can be loaded and run by `lerobot-eval` or `lerobot-rollout`.
 - [ ] `templates/lerobot_modelcard_template.md` has a description entry and a `policy_docs` link for your policy.
 - [ ] The models table in the root `README.md` lists your policy in the right category, linking to your doc page.
 - [ ] At least one reproducible benchmark eval in the policy MDX with a published checkpoint (sim benchmark, or real-robot dataset + checkpoint).

docs/source/using_dataset_tools.mdx

@@ -252,6 +252,10 @@ lerobot-dataset-viz \
     --episode-index 0
 ```
 
+For a private or gated dataset, authenticate first with `hf auth login`, or set the
+`HF_TOKEN` environment variable. The Hub client then discovers the credential
+automatically; no token argument is needed.
+
 **From a local folder:**
 Add the `--root` option and set `--mode local`. For example, to search in `./my_local_data_dir/lerobot/pusht`:
 

pyproject.toml

@@ -155,7 +155,7 @@ accelerate-dep = ["accelerate>=1.14.0,<2.0.0"]
 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"]
-diffusers-dep = ["diffusers>=0.27.2,<0.36.0"]
+diffusers-dep = ["diffusers>=0.38.0,<0.40.0"]
 qwen-vl-utils-dep = ["qwen-vl-utils>=0.0.11,<0.1.0"]
 matplotlib-dep = ["matplotlib>=3.10.3,<4.0.0", "contourpy>=1.3.0,<2.0.0"] # NOTE: Explicitly listing contourpy helps the resolver converge faster.
 pyserial-dep = ["pyserial>=3.5,<4.0"]

src/lerobot/cameras/reachy2_camera/reachy2_camera.py

@@ -173,7 +173,8 @@ class Reachy2Camera(Camera):
             raise ValueError(
                 f"Invalid color mode '{self.color_mode}'. Expected {ColorMode.RGB} or {ColorMode.BGR}."
             )
-        if self.color_mode == ColorMode.RGB:
+        is_depth_frame = self.config.name == "depth" and self.config.image_type == "depth"
+        if not is_depth_frame and self.color_mode == ColorMode.RGB:
             frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
 
         self.latest_frame = frame

src/lerobot/cameras/realsense/camera_realsense.py

@@ -453,7 +453,7 @@ class RealSenseCamera(Camera):
             )
 
         processed_image = image
-        if self.color_mode == ColorMode.BGR:
+        if not depth_frame and self.color_mode == ColorMode.BGR:
             processed_image = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
 
         if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE, cv2.ROTATE_180]:

src/lerobot/configs/train.py

@@ -14,7 +14,6 @@
 import builtins
 import datetime as dt
 import json
-import multiprocessing
 import os
 import tempfile
 from dataclasses import dataclass, field
@@ -102,12 +101,6 @@ class TrainPipelineConfig(HubMixin):
     batch_size: int = 8
     prefetch_factor: int = 4
     persistent_workers: bool = True
-    # DataLoader worker start method. "spawn" is safer than "fork" with
-    # non-fork-safe libs (PyAV / torchcodec / ffmpeg), but adds some
-    # worker-startup time per run since workers re-import modules instead
-    # of inheriting parent state. Override with `--dataloader_multiprocessing_context=fork`
-    # when appropriate, or set it to `null` to use Python's platform default.
-    dataloader_multiprocessing_context: str | None = "spawn"
     steps: int = 100_000
     # Run policy in the simulation environment every N steps to measure reward/success (0 = disabled).
     env_eval_freq: int = 20_000
@@ -219,17 +212,6 @@ class TrainPipelineConfig(HubMixin):
             self.reward_model.pretrained_path = str(policy_dir)
 
     def validate(self) -> None:
-        available_contexts = multiprocessing.get_all_start_methods()
-        if (
-            self.dataloader_multiprocessing_context is not None
-            and self.dataloader_multiprocessing_context not in available_contexts
-        ):
-            raise ValueError(
-                "`dataloader_multiprocessing_context` must be None or one of "
-                f"{available_contexts} on this platform, got "
-                f"{self.dataloader_multiprocessing_context!r}."
-            )
-
         self._resolve_pretrained_from_cli()
 
         if self.policy is None and self.reward_model is None:

src/lerobot/datasets/dataset_metadata.py

@@ -73,6 +73,8 @@ class LeRobotDatasetMetadata:
         revision: str | None = None,
         force_cache_sync: bool = False,
         metadata_buffer_size: int = 10,
+        *,
+        token: str | bool | None = None,
     ):
         """Load or download metadata for an existing LeRobot dataset.
 
@@ -94,6 +96,10 @@ class LeRobotDatasetMetadata:
                 even when local files exist.
             metadata_buffer_size: Number of episode metadata records to buffer
                 in memory before flushing to parquet.
+            token: Authentication token used for Hub requests. Pass a string
+                token, ``True`` to require the locally stored token, ``False``
+                to disable authentication, or ``None`` to use the Hugging Face
+                Hub default.
         """
         self.repo_id = repo_id
         self.revision = revision if revision else CODEBASE_VERSION
@@ -113,9 +119,12 @@ class LeRobotDatasetMetadata:
             self._load_metadata()
         except (FileNotFoundError, NotADirectoryError):
             if is_valid_version(self.revision):
-                self.revision = get_safe_version(self.repo_id, self.revision)
+                if token is None:
+                    self.revision = get_safe_version(self.repo_id, self.revision)
+                else:
+                    self.revision = get_safe_version(self.repo_id, self.revision, token=token)
 
-            self._pull_from_repo(allow_patterns="meta/")
+            self._pull_from_repo(allow_patterns="meta/", token=token)
             self._load_metadata()
 
     def _flush_metadata_buffer(self) -> None:
@@ -220,7 +229,10 @@ class LeRobotDatasetMetadata:
         self,
         allow_patterns: list[str] | str | None = None,
         ignore_patterns: list[str] | str | None = None,
+        *,
+        token: str | bool | None = None,
     ) -> None:
+        token_kwargs = {} if token is None else {"token": token}
         if self._requested_root is None:
             self.root = Path(
                 snapshot_download(
@@ -230,6 +242,7 @@ class LeRobotDatasetMetadata:
                     cache_dir=HF_LEROBOT_HUB_CACHE,
                     allow_patterns=allow_patterns,
                     ignore_patterns=ignore_patterns,
+                    **token_kwargs,
                 )
             )
             return
@@ -242,6 +255,7 @@ class LeRobotDatasetMetadata:
             local_dir=self._requested_root,
             allow_patterns=allow_patterns,
             ignore_patterns=ignore_patterns,
+            **token_kwargs,
         )
         self.root = self._requested_root
 

src/lerobot/datasets/dataset_writer.py

@@ -172,6 +172,23 @@ class DatasetWriter:
     def _get_image_file_dir(self, episode_index: int, image_key: str) -> Path:
         return self._get_image_file_path(episode_index, image_key, frame_index=0).parent
 
+    def _get_episode_buffer_index(self) -> int:
+        episode_index = self.episode_buffer["episode_index"]
+        # episode_index is `int` when freshly created, but becomes `np.ndarray` after
+        # save_episode() mutates the buffer. Handle both types here.
+        if isinstance(episode_index, np.ndarray):
+            episode_index = episode_index.item() if episode_index.size == 1 else episode_index[0]
+        return int(episode_index)
+
+    def _delete_camera_frame_dirs(self, camera_keys: list[str]) -> None:
+        if self.image_writer is not None:
+            self._wait_image_writer()
+        episode_index = self._get_episode_buffer_index()
+        for camera_key in camera_keys:
+            img_dir = self._get_image_file_dir(episode_index, camera_key)
+            if img_dir.is_dir():
+                shutil.rmtree(img_dir)
+
     def _save_image(
         self, image: torch.Tensor | np.ndarray | PIL.Image.Image, fpath: Path, compress_level: int = 1
     ) -> None:
@@ -369,7 +386,9 @@ class DatasetWriter:
                 self._episodes_since_last_encoding = 0
 
         if episode_data is None:
-            self.clear_episode_buffer(delete_images=len(self._meta.image_keys) > 0)
+            if len(self._meta.image_keys) > 0:
+                self._delete_camera_frame_dirs(self._meta.image_keys)
+            self.episode_buffer = self._create_episode_buffer()
 
     def _batch_save_episode_video(self, start_episode: int, end_episode: int | None = None) -> None:
         """Batch save videos for multiple episodes."""
@@ -561,10 +580,10 @@ class DatasetWriter:
         return metadata
 
     def clear_episode_buffer(self, delete_images: bool = True) -> None:
-        """Discard the current episode buffer and optionally delete temp images.
+        """Discard the current episode buffer and optionally delete temp camera frames.
 
         Args:
-            delete_images: If ``True``, remove temporary image directories
+            delete_images: If ``True``, remove temporary camera frame directories
                 written for the current episode.
         """
         # Cancel streaming encoder if active
@@ -572,17 +591,7 @@ class DatasetWriter:
             self._streaming_encoder.cancel_episode()
 
         if delete_images:
-            if self.image_writer is not None:
-                self._wait_image_writer()
-            episode_index = self.episode_buffer["episode_index"]
-            # episode_index is `int` when freshly created, but becomes `np.ndarray` after
-            # save_episode() mutates the buffer. Handle both types here.
-            if isinstance(episode_index, np.ndarray):
-                episode_index = episode_index.item() if episode_index.size == 1 else episode_index[0]
-            for cam_key in self._meta.image_keys:
-                img_dir = self._get_image_file_dir(episode_index, cam_key)
-                if img_dir.is_dir():
-                    shutil.rmtree(img_dir)
+            self._delete_camera_frame_dirs(self._meta.camera_keys)
 
         self.episode_buffer = self._create_episode_buffer()
 

src/lerobot/datasets/lerobot_dataset.py

@@ -65,6 +65,8 @@ class LeRobotDataset(torch.utils.data.Dataset):
         encoder_threads: int | None = None,
         streaming_encoding: bool = False,
         encoder_queue_maxsize: int = 30,
+        *,
+        token: str | bool | None = None,
     ):
         """
         2 modes are available for instantiating this class, depending on 2 different use cases:
@@ -197,6 +199,11 @@ class LeRobotDataset(torch.utils.data.Dataset):
                 instead of writing PNG images first. This makes save_episode() near-instant. Defaults to False.
             encoder_queue_maxsize (int, optional): Maximum number of frames to buffer per camera when using
                 streaming encoding. Defaults to 30 (~1s at 30fps).
+            token: Authentication token used while downloading this dataset
+                from the Hub. Pass a string token, ``True`` to require the
+                locally stored token, ``False`` to disable authentication, or
+                ``None`` to use the Hugging Face Hub default. The token is not
+                retained on the dataset instance after initialization.
 
         Note:
             Write-mode parameters (``streaming_encoding``, ``batch_encoding_size``) passed to
@@ -220,7 +227,11 @@ class LeRobotDataset(torch.utils.data.Dataset):
 
         # Load metadata (sets self.root once from the resolved metadata root)
         self.meta = LeRobotDatasetMetadata(
-            self.repo_id, self._requested_root, self.revision, force_cache_sync=force_cache_sync
+            self.repo_id,
+            self._requested_root,
+            self.revision,
+            force_cache_sync=force_cache_sync,
+            token=token,
         )
         self.root = self.meta.root
         self.revision = self.meta.revision
@@ -260,8 +271,11 @@ class LeRobotDataset(torch.utils.data.Dataset):
         # Load actual data
         if force_cache_sync or not self.reader.try_load():
             if is_valid_version(self.revision):
-                self.revision = get_safe_version(self.repo_id, self.revision)
-            self._download(download_videos)
+                if token is None:
+                    self.revision = get_safe_version(self.repo_id, self.revision)
+                else:
+                    self.revision = get_safe_version(self.repo_id, self.revision, token=token)
+            self._download(download_videos, token=token)
             self.reader.load_and_activate()
 
         # Detect write-mode params for backward compatibility
@@ -626,10 +640,11 @@ class LeRobotDataset(torch.utils.data.Dataset):
                 hub_api.delete_tag(self.repo_id, tag=CODEBASE_VERSION, repo_type="dataset")
             hub_api.create_tag(self.repo_id, tag=CODEBASE_VERSION, revision=branch, repo_type="dataset")
 
-    def _download(self, download_videos: bool = True) -> None:
+    def _download(self, download_videos: bool = True, *, token: str | bool | None = None) -> None:
         """Downloads the dataset from the given 'repo_id' at the provided version."""
         ignore_patterns = None if download_videos else "videos/"
         files = None
+        token_kwargs = {} if token is None else {"token": token}
         if self.episodes is not None:
             # Reader is guaranteed to exist here (created in __init__ before _download)
             files = self.reader.get_episodes_file_paths()
@@ -643,6 +658,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
                     cache_dir=HF_LEROBOT_HUB_CACHE,
                     allow_patterns=files,
                     ignore_patterns=ignore_patterns,
+                    **token_kwargs,
                 )
             )
         else:
@@ -654,6 +670,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
                 local_dir=self._requested_root,
                 allow_patterns=files,
                 ignore_patterns=ignore_patterns,
+                **token_kwargs,
             )
             self.meta.root = self._requested_root
 
@@ -793,6 +810,8 @@ class LeRobotDataset(torch.utils.data.Dataset):
         image_writer_threads: int = 0,
         streaming_encoding: bool = False,
         encoder_queue_maxsize: int = 30,
+        *,
+        token: str | bool | None = None,
     ) -> "LeRobotDataset":
         """Resume recording on an existing dataset.
 
@@ -826,6 +845,8 @@ class LeRobotDataset(torch.utils.data.Dataset):
             streaming_encoding: If ``True``, encode video in real-time during
                 capture.
             encoder_queue_maxsize: Max buffered frames per camera for streaming.
+            token: Authentication token used if metadata must be downloaded
+                from the Hub. The token is not retained on the dataset instance.
 
         Returns:
             A :class:`LeRobotDataset` in write mode, ready to append episodes.
@@ -854,7 +875,11 @@ class LeRobotDataset(torch.utils.data.Dataset):
 
         # Load metadata (revision-safe when root is not provided)
         obj.meta = LeRobotDatasetMetadata(
-            obj.repo_id, obj._requested_root, obj.revision, force_cache_sync=force_cache_sync
+            obj.repo_id,
+            obj._requested_root,
+            obj.revision,
+            force_cache_sync=force_cache_sync,
+            token=token,
         )
 
         obj._encoder_threads = encoder_threads

src/lerobot/datasets/multi_dataset.py

@@ -48,6 +48,8 @@ class MultiLeRobotDataset(torch.utils.data.Dataset):
         tolerances_s: dict | None = None,
         download_videos: bool = True,
         video_backend: str | None = None,
+        *,
+        token: str | bool | None = None,
     ):
         super().__init__()
         self.repo_ids = repo_ids
@@ -65,6 +67,7 @@ class MultiLeRobotDataset(torch.utils.data.Dataset):
                 tolerance_s=self.tolerances_s[repo_id],
                 download_videos=download_videos,
                 video_backend=video_backend,
+                token=token,
             )
             for repo_id in repo_ids
         ]

src/lerobot/datasets/streaming_dataset.py

@@ -256,6 +256,8 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
         shuffle: bool = True,
         return_uint8: bool = False,
         depth_output_unit: str = DEFAULT_DEPTH_UNIT,
+        *,
+        token: str | bool | None = None,
     ):
         """Initialize a StreamingLeRobotDataset.
 
@@ -278,6 +280,11 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
             shuffle (bool, optional): Whether to shuffle the dataset across exhaustions. Defaults to True.
             depth_output_unit (str, optional): Physical unit depth maps are dequantized to ("m" or "mm").
                 Defaults to "mm".
+            token: Authentication token used while streaming this dataset from
+                the Hub. Pass a string token, ``True`` to require the locally
+                stored token, ``False`` to disable authentication, or ``None``
+                to use the Hugging Face Hub default. The token is not retained
+                on the dataset instance after initialization.
         """
         super().__init__()
         self.repo_id = repo_id
@@ -306,7 +313,11 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
 
         # Load metadata
         self.meta = LeRobotDatasetMetadata(
-            self.repo_id, self._requested_root, self.revision, force_cache_sync=force_cache_sync
+            self.repo_id,
+            self._requested_root,
+            self.revision,
+            force_cache_sync=force_cache_sync,
+            token=token,
         )
         self.root = self.meta.root
         self.revision = self.meta.revision
@@ -334,12 +345,14 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
             self.delta_timestamps = delta_timestamps
             self.delta_indices = get_delta_indices(self.delta_timestamps, self.fps)
 
+        token_kwargs = {} if token is None or self.streaming_from_local else {"token": token}
         self.hf_dataset: datasets.IterableDataset = load_dataset(
             self.repo_id if not self.streaming_from_local else str(self.root),
             split="train",
             streaming=self.streaming,
             data_files="data/*/*.parquet",
             revision=self.revision,
+            **token_kwargs,
         )
 
         self.num_shards = min(self.hf_dataset.num_shards, max_num_shards)

src/lerobot/datasets/utils.py

@@ -325,16 +325,19 @@ def check_version_compatibility(
         logging.warning(FUTURE_MESSAGE.format(repo_id=repo_id, version=v_check))
 
 
-def get_repo_versions(repo_id: str) -> list[packaging.version.Version]:
+def get_repo_versions(repo_id: str, *, token: str | bool | None = None) -> list[packaging.version.Version]:
     """Return available valid versions (branches and tags) on a given Hub repo.
 
     Args:
         repo_id (str): The repository ID on the Hugging Face Hub.
+        token: Authentication token used for Hub requests. Pass a string token,
+            ``True`` to require the locally stored token, ``False`` to disable
+            authentication, or ``None`` to use the Hugging Face Hub default.
 
     Returns:
         list[packaging.version.Version]: A list of valid versions found.
     """
-    api = HfApi()
+    api = HfApi() if token is None else HfApi(token=token)
     repo_refs = api.list_repo_refs(repo_id, repo_type="dataset")
     repo_refs = [b.name for b in repo_refs.branches + repo_refs.tags]
     repo_versions = []
@@ -345,7 +348,12 @@ def get_repo_versions(repo_id: str) -> list[packaging.version.Version]:
     return repo_versions
 
 
-def get_safe_version(repo_id: str, version: str | packaging.version.Version) -> str:
+def get_safe_version(
+    repo_id: str,
+    version: str | packaging.version.Version,
+    *,
+    token: str | bool | None = None,
+) -> str:
     """Return the specified version if available on repo, or the latest compatible one.
 
     If the exact version is not found, it looks for the latest version with the
@@ -354,6 +362,7 @@ def get_safe_version(repo_id: str, version: str | packaging.version.Version) ->
     Args:
         repo_id (str): The repository ID on the Hugging Face Hub.
         version (str | packaging.version.Version): The target version.
+        token: Authentication token forwarded to the Hub version lookup.
 
     Returns:
         str: The safe version string (e.g., "v1.2.3") to use as a revision.
@@ -366,7 +375,7 @@ def get_safe_version(repo_id: str, version: str | packaging.version.Version) ->
     target_version = (
         packaging.version.parse(version) if not isinstance(version, packaging.version.Version) else version
     )
-    hub_versions = get_repo_versions(repo_id)
+    hub_versions = get_repo_versions(repo_id) if token is None else get_repo_versions(repo_id, token=token)
 
     if not hub_versions:
         raise RevisionNotFoundError(

src/lerobot/envs/metaworld.py

@@ -155,6 +155,7 @@ class MetaworldEnv(gym.Env):
             env.model.cam_pos[2] = [0.75, 0.075, 0.7]
         env.reset()
         env._freeze_rand_vec = False  # otherwise no randomization
+        env.seeded_rand_vec = True  # use seeded RNG so reset(seed=X) controls object positions
         self._env = env
 
     def render(self) -> np.ndarray:
@@ -220,6 +221,8 @@ class MetaworldEnv(gym.Env):
         self._ensure_env()
         super().reset(seed=seed)
 
+        if seed is not None:
+            self._env.seed(seed)
         raw_obs, info = self._env.reset(seed=seed)
 
         observation = self._format_raw_obs(raw_obs)

src/lerobot/policies/factory.py

@@ -177,6 +177,7 @@ def make_pre_post_processors(
             return make_groot_pre_post_processors_from_pretrained(
                 config=policy_cfg,
                 pretrained_path=pretrained_path,
+                revision=pretrained_revision,
                 dataset_stats=kwargs.get("dataset_stats"),
                 dataset_meta=kwargs.get("dataset_meta"),
                 preprocessor_overrides=kwargs.get("preprocessor_overrides"),

src/lerobot/policies/groot/processor_groot.py

@@ -475,6 +475,7 @@ def make_groot_pre_post_processors_from_pretrained(
     config: GrootConfig,
     pretrained_path: str,
     *,
+    revision: str | None = None,
     dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
     dataset_meta: Any | None = None,
     preprocessor_overrides: dict[str, Any] | None = None,
@@ -511,6 +512,7 @@ def make_groot_pre_post_processors_from_pretrained(
 
     preprocessor, postprocessor = _load_groot_processor_pipelines(
         pretrained_path,
+        revision=revision,
         preprocessor_overrides=preprocessor_overrides,
         postprocessor_overrides=postprocessor_overrides,
         preprocessor_config_filename=preprocessor_config_filename,
@@ -526,6 +528,7 @@ def make_groot_pre_post_processors_from_pretrained(
 def _load_groot_processor_pipelines(
     pretrained_path: str,
     *,
+    revision: str | None,
     preprocessor_overrides: dict[str, Any],
     postprocessor_overrides: dict[str, Any],
     preprocessor_config_filename: str,
@@ -540,6 +543,7 @@ def _load_groot_processor_pipelines(
     preprocessor = PolicyProcessorPipeline.from_pretrained(
         pretrained_model_name_or_path=pretrained_path,
         config_filename=preprocessor_config_filename,
+        revision=revision,
         overrides=preprocessor_overrides,
         to_transition=batch_to_transition,
         to_output=transition_to_batch,
@@ -547,6 +551,7 @@ def _load_groot_processor_pipelines(
     postprocessor = PolicyProcessorPipeline.from_pretrained(
         pretrained_model_name_or_path=pretrained_path,
         config_filename=postprocessor_config_filename,
+        revision=revision,
         overrides=postprocessor_overrides,
         to_transition=policy_action_to_transition,
         to_output=transition_to_policy_action,

src/lerobot/processor/pipeline.py

@@ -713,6 +713,8 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
             ProcessorMigrationError: If the model requires migration to processor format.
         """
         model_id = str(pretrained_model_name_or_path)
+        model_path = Path(model_id)
+        is_local_source = model_path.is_dir() or model_path.is_file()
         hub_download_kwargs = {
             "force_download": force_download,
             "resume_download": resume_download,
@@ -731,7 +733,7 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
 
         # 3. Build steps with overrides
         steps, validated_overrides = cls._build_steps_with_overrides(
-            loaded_config, overrides or {}, model_id, base_path, hub_download_kwargs
+            loaded_config, overrides or {}, model_id, base_path, hub_download_kwargs, is_local_source
         )
 
         # 4. Validate that all overrides were used
@@ -921,6 +923,7 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
         model_id: str,
         base_path: Path | None,
         hub_download_kwargs: dict[str, Any],
+        is_local_source: bool = False,
     ) -> tuple[list[ProcessorStep], set[str]]:
         """Build all processor steps with overrides and state loading.
 
@@ -944,7 +947,7 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
         3. **State Loading** (via _load_step_state):
            - **If step has "state_file"**: Load tensor state from .safetensors
            - **Local first**: Check base_path/state_file.safetensors
-           - **Hub fallback**: Download state file if not found locally
+           - **Hub fallback**: Download state file if the pipeline was loaded from the Hub
            - **Optional**: Only load if step has load_state_dict method
 
         4. **Override Tracking**:
@@ -962,6 +965,7 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
             model_id: The model identifier (needed for Hub state file downloads)
             base_path: Local directory path for finding state files
             hub_download_kwargs: Parameters for hf_hub_download (tokens, cache, etc.)
+            is_local_source: Whether model_id resolved to a local directory or config file.
 
         Returns:
             Tuple of (instantiated_steps_list, unused_override_keys)
@@ -975,7 +979,9 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
         steps, remaining_override_keys = cls._build_steps_from_config(loaded_config, overrides)
 
         for step_instance, step_entry in zip(steps, loaded_config["steps"], strict=True):
-            cls._load_step_state(step_instance, step_entry, model_id, base_path, hub_download_kwargs)
+            cls._load_step_state(
+                step_instance, step_entry, model_id, base_path, hub_download_kwargs, is_local_source
+            )
 
         return steps, remaining_override_keys
 
@@ -1139,6 +1145,7 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
         model_id: str,
         base_path: Path | None,
         hub_download_kwargs: dict[str, Any],
+        is_local_source: bool = False,
     ) -> None:
         """Load state dictionary for a processor step if available.
 
@@ -1157,7 +1164,7 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
            - **Use case**: Loading from local saved model directory
 
         2. **Hub download fallback**: Download state file from repository
-           - **When triggered**: Local file not found or base_path is None
+           - **When triggered**: Local file not found and the pipeline source is a Hub repo
            - **Process**: Use hf_hub_download with same parameters as config
            - **Example**: Download "normalize_step_0.safetensors" from "user/repo"
            - **Result**: Downloaded to local cache, path returned
@@ -1178,6 +1185,7 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
             model_id: The model identifier (used for Hub downloads if needed)
             base_path: Local directory path for finding state files (None for Hub-only)
             hub_download_kwargs: Parameters for hf_hub_download (tokens, cache, etc.)
+            is_local_source: Whether model_id resolved to a local directory or config file.
 
         Note:
             This method modifies step_instance in-place and returns None.
@@ -1191,6 +1199,12 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
         # Try local file first
         if base_path and (base_path / state_filename).exists():
             state_path = str(base_path / state_filename)
+        elif is_local_source:
+            state_path = base_path / state_filename if base_path else Path(state_filename)
+            raise FileNotFoundError(
+                f"State file '{state_filename}' was not found for local processor pipeline "
+                f"'{model_id}' at '{state_path}'."
+            )
         else:
             # Download from Hub
             state_path = hf_hub_download(

src/lerobot/robots/bi_so_follower/bi_so_follower.py

@@ -58,6 +58,9 @@ class BiSOFollower(BimanualMixin, Robot):
             port=config.left_arm_config.port,
             disable_torque_on_disconnect=config.left_arm_config.disable_torque_on_disconnect,
             max_relative_target=config.left_arm_config.max_relative_target,
+            position_p_coefficient=config.left_arm_config.position_p_coefficient,
+            position_i_coefficient=config.left_arm_config.position_i_coefficient,
+            position_d_coefficient=config.left_arm_config.position_d_coefficient,
             use_degrees=config.left_arm_config.use_degrees,
             cameras=left_arm_cameras,
         )
@@ -68,6 +71,9 @@ class BiSOFollower(BimanualMixin, Robot):
             port=config.right_arm_config.port,
             disable_torque_on_disconnect=config.right_arm_config.disable_torque_on_disconnect,
             max_relative_target=config.right_arm_config.max_relative_target,
+            position_p_coefficient=config.right_arm_config.position_p_coefficient,
+            position_i_coefficient=config.right_arm_config.position_i_coefficient,
+            position_d_coefficient=config.right_arm_config.position_d_coefficient,
             use_degrees=config.right_arm_config.use_degrees,
             cameras=config.right_arm_config.cameras,
         )

src/lerobot/robots/lekiwi/lekiwi_client.py

@@ -323,6 +323,10 @@ class LeKiwiClient(Robot):
             np.ndarray: the action sent to the motors, potentially clipped.
         """
 
+        # Action values may be torch tensors (e.g. replayed from a dataset) or numpy
+        # scalars; json.dumps only serializes Python primitives, so coerce each value to a
+        # plain float before sending.
+        action = {key: float(value) for key, value in action.items()}
         self.zmq_cmd_socket.send_string(json.dumps(action))  # action is in motor space
 
         # TODO(Steven): Remove the np conversion when it is possible to record a non-numpy array value

src/lerobot/robots/so_follower/config_so_follower.py

@@ -41,6 +41,11 @@ class SOFollowerConfig:
     # Set to `True` for backward compatibility with previous policies/dataset
     use_degrees: bool = True
 
+    # Position-mode PID gains written to Feetech STS3215 motors at connect time.
+    position_p_coefficient: int = 16
+    position_i_coefficient: int = 0
+    position_d_coefficient: int = 32
+
 
 @RobotConfig.register_subclass("so101_follower")
 @RobotConfig.register_subclass("so100_follower")

src/lerobot/robots/so_follower/so_follower.py

@@ -161,11 +161,9 @@ class SOFollower(Robot):
             self.bus.configure_motors()
             for motor in self.bus.motors:
                 self.bus.write("Operating_Mode", motor, OperatingMode.POSITION.value)
-                # Set P_Coefficient to lower value to avoid shakiness (Default is 32)
-                self.bus.write("P_Coefficient", motor, 16)
-                # Set I_Coefficient and D_Coefficient to default value 0 and 32
-                self.bus.write("I_Coefficient", motor, 0)
-                self.bus.write("D_Coefficient", motor, 32)
+                self.bus.write("P_Coefficient", motor, self.config.position_p_coefficient)
+                self.bus.write("I_Coefficient", motor, self.config.position_i_coefficient)
+                self.bus.write("D_Coefficient", motor, self.config.position_d_coefficient)
 
                 if motor == "gripper":
                     self.bus.write("Max_Torque_Limit", motor, 500)  # 50% of max torque to avoid burnout

src/lerobot/rollout/configs.py

@@ -326,8 +326,17 @@ class RolloutConfig:
 
         policy_path = parser.get_path_arg("policy")
         if policy_path:
-            cli_overrides = parser.get_cli_overrides("policy")
-            self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=cli_overrides)
+            yaml_overrides = parser.get_yaml_overrides("policy")
+            cli_overrides = parser.get_cli_overrides("policy") or []
+            policy_overrides = yaml_overrides + cli_overrides
+            pretrained_revision = parser.parse_arg("pretrained_revision", cli_overrides)
+            if pretrained_revision is None:
+                pretrained_revision = parser.parse_arg("pretrained_revision", yaml_overrides)
+            self.policy = PreTrainedConfig.from_pretrained(
+                policy_path,
+                revision=pretrained_revision,
+                cli_overrides=policy_overrides,
+            )
             self.policy.pretrained_path = policy_path
         if self.policy is None:
             raise ValueError("--policy.path is required for rollout")

src/lerobot/rollout/context.py

@@ -27,7 +27,7 @@ from threading import Event
 
 import torch
 
-from lerobot.configs import FeatureType
+from lerobot.configs import FeatureType, PreTrainedConfig
 from lerobot.datasets import (
     LeRobotDataset,
     aggregate_pipeline_dataset_features,
@@ -159,6 +159,35 @@ class RolloutContext:
 # ---------------------------------------------------------------------------
 
 
+def _load_pretrained_policy(policy_config: PreTrainedConfig) -> PreTrainedPolicy:
+    """Load policy weights, keeping adapter and base-model revisions independent."""
+    pretrained_revision = policy_config.pretrained_revision
+    policy_class = get_policy_class(policy_config.type)
+
+    if not policy_config.use_peft:
+        return policy_class.from_pretrained(
+            policy_config.pretrained_path,
+            config=policy_config,
+            revision=pretrained_revision,
+        )
+
+    from peft import PeftConfig, PeftModel
+
+    peft_path = policy_config.pretrained_path
+    peft_config = PeftConfig.from_pretrained(peft_path, revision=pretrained_revision)
+    policy = policy_class.from_pretrained(
+        pretrained_name_or_path=peft_config.base_model_name_or_path,
+        config=policy_config,
+        revision=peft_config.revision,
+    )
+    return PeftModel.from_pretrained(
+        policy,
+        peft_path,
+        config=peft_config,
+        revision=pretrained_revision,
+    )
+
+
 def build_rollout_context(
     cfg: RolloutConfig,
     shutdown_event: Event,
@@ -176,7 +205,6 @@ def build_rollout_context(
     # --- 1. Policy (heavy I/O, but no hardware yet) -------------------
     logger.info("Loading policy from '%s'...", cfg.policy.pretrained_path)
     policy_config = cfg.policy
-    policy_class = get_policy_class(policy_config.type)
 
     if hasattr(policy_config, "compile_model"):
         policy_config.compile_model = cfg.use_torch_compile
@@ -187,17 +215,7 @@ def build_rollout_context(
             "Please use `cpu` or `cuda` backend."
         )
 
-    if policy_config.use_peft:
-        from peft import PeftConfig, PeftModel
-
-        peft_path = policy_config.pretrained_path
-        peft_config = PeftConfig.from_pretrained(peft_path)
-        policy = policy_class.from_pretrained(
-            pretrained_name_or_path=peft_config.base_model_name_or_path, config=policy_config
-        )
-        policy = PeftModel.from_pretrained(policy, peft_path, config=peft_config)
-    else:
-        policy = policy_class.from_pretrained(policy_config.pretrained_path, config=policy_config)
+    policy = _load_pretrained_policy(policy_config)
 
     if is_rtc:
         policy.config.rtc_config = cfg.inference.rtc
@@ -392,6 +410,7 @@ def build_rollout_context(
     preprocessor, postprocessor = make_pre_post_processors(
         policy_cfg=policy_config,
         pretrained_path=cfg.policy.pretrained_path,
+        pretrained_revision=policy_config.pretrained_revision,
         dataset_stats=dataset_stats,
         preprocessor_overrides={
             "device_processor": {"device": cfg.device},

src/lerobot/scripts/convert_dataset_v21_to_v30.py

@@ -61,6 +61,7 @@ import pyarrow as pa
 import tqdm
 from datasets import Dataset, Features, Image
 from huggingface_hub import HfApi, snapshot_download
+from huggingface_hub.errors import RevisionNotFoundError
 from requests import HTTPError
 
 from lerobot.datasets import CODEBASE_VERSION, LeRobotDataset, aggregate_stats
@@ -521,7 +522,7 @@ def convert_dataset(
         hub_api = HfApi()
         try:
             hub_api.delete_tag(repo_id, tag=CODEBASE_VERSION, repo_type="dataset")
-        except HTTPError as e:
+        except (HTTPError, RevisionNotFoundError) as e:
             print(f"tag={CODEBASE_VERSION} probably doesn't exist. Skipping exception ({e})")
             pass
         hub_api.delete_files(

src/lerobot/scripts/lerobot_eval.py

@@ -453,6 +453,9 @@ def eval_policy(
             raise exc from None
 
     start = time.time()
+    # Preserve the mode for direct callers. eval_policy_all scopes the mode
+    # around all tasks so parallel evaluations cannot race with each other.
+    was_training = policy.training
     policy.eval()
 
     # Determine how many batched rollouts we need to get n_episodes. Note that if n_episodes is not evenly
@@ -674,6 +677,8 @@ def eval_policy(
     if save_predicted_video:
         info["predicted_video_paths"] = predicted_video_paths
 
+    policy.train(was_training)
+
     return info
 
 
@@ -1010,40 +1015,48 @@ def eval_policy_all(
         recording_private=recording_private,
     )
 
-    if max_parallel_tasks <= 1:
-        prefetch_thread: threading.Thread | None = None
-        for i, (task_group, task_id, env) in enumerate(tasks):
-            if prefetch_thread is not None:
-                prefetch_thread.join()
-                prefetch_thread = None
+    # Set the shared policy's mode before launching any workers. Restoring it
+    # inside individual tasks would let one task enable training mode while
+    # another task is still evaluating.
+    was_training = policy.training
+    policy.eval()
+    try:
+        if max_parallel_tasks <= 1:
+            prefetch_thread: threading.Thread | None = None
+            for i, (task_group, task_id, env) in enumerate(tasks):
+                if prefetch_thread is not None:
+                    prefetch_thread.join()
+                    prefetch_thread = None
 
-            try:
-                tg, tid, metrics = task_runner(task_group, task_id, env)
-                _accumulate_to(tg, metrics)
-                per_task_infos.append({"task_group": tg, "task_id": tid, "metrics": metrics})
-            finally:
-                env.close()
-                # Prefetch next task's workers *after* closing current env to prevent
-                # GPU memory overlap between consecutive tasks.
-                if i + 1 < len(tasks):
-                    next_env = tasks[i + 1][2]
-                    if hasattr(next_env, "_ensure"):
-                        prefetch_thread = threading.Thread(target=next_env._ensure, daemon=True)
-                        prefetch_thread.start()
-    else:
-        with cf.ThreadPoolExecutor(max_workers=max_parallel_tasks) as executor:
-            fut2meta = {}
-            for task_group, task_id, env in tasks:
-                fut = executor.submit(task_runner, task_group, task_id, env)
-                fut2meta[fut] = (task_group, task_id, env)
-            for fut in cf.as_completed(fut2meta):
-                tg, tid, env = fut2meta[fut]
                 try:
-                    tg, tid, metrics = fut.result()
+                    tg, tid, metrics = task_runner(task_group, task_id, env)
                     _accumulate_to(tg, metrics)
                     per_task_infos.append({"task_group": tg, "task_id": tid, "metrics": metrics})
                 finally:
                     env.close()
+                    # Prefetch next task's workers *after* closing current env to prevent
+                    # GPU memory overlap between consecutive tasks.
+                    if i + 1 < len(tasks):
+                        next_env = tasks[i + 1][2]
+                        if hasattr(next_env, "_ensure"):
+                            prefetch_thread = threading.Thread(target=next_env._ensure, daemon=True)
+                            prefetch_thread.start()
+        else:
+            with cf.ThreadPoolExecutor(max_workers=max_parallel_tasks) as executor:
+                fut2meta = {}
+                for task_group, task_id, env in tasks:
+                    fut = executor.submit(task_runner, task_group, task_id, env)
+                    fut2meta[fut] = (task_group, task_id, env)
+                for fut in cf.as_completed(fut2meta):
+                    tg, tid, env = fut2meta[fut]
+                    try:
+                        tg, tid, metrics = fut.result()
+                        _accumulate_to(tg, metrics)
+                        per_task_infos.append({"task_group": tg, "task_id": tid, "metrics": metrics})
+                    finally:
+                        env.close()
+    finally:
+        policy.train(was_training)
 
     # compute aggregated metrics helper (robust to lists/scalars)
     def _agg_from_list(xs):

src/lerobot/scripts/lerobot_record.py

@@ -453,9 +453,11 @@ def record(
                 encoder_queue_maxsize=cfg.dataset.encoder_queue_maxsize,
             )
 
-        robot.connect()
+        # Connect the teleoperator before the robot so the robot isn't left idle (and possibly
+        # tripping a firmware watchdog) during teleop init. Matches lerobot_teleoperate.py.
         if teleop is not None:
             teleop.connect()
+        robot.connect()
 
         listener, events = init_keyboard_listener()
 

src/lerobot/scripts/lerobot_replay.py

@@ -61,6 +61,7 @@ from lerobot.robots import (  # noqa: F401
     earthrover_mini_plus,
     hope_jr,
     koch_follower,
+    lekiwi,
     make_robot_from_config,
     omx_follower,
     openarm_follower,

src/lerobot/scripts/lerobot_train.py

@@ -71,16 +71,6 @@ from lerobot.utils.utils import (
 from .lerobot_eval import eval_policy_all
 
 
-def _dataloader_worker_kwargs(cfg: TrainPipelineConfig) -> dict[str, Any]:
-    """Return worker-only DataLoader options, disabling them for single-process loading."""
-    workers_enabled = cfg.num_workers > 0
-    return {
-        "prefetch_factor": cfg.prefetch_factor if workers_enabled else None,
-        "persistent_workers": cfg.persistent_workers and workers_enabled,
-        "multiprocessing_context": cfg.dataloader_multiprocessing_context if workers_enabled else None,
-    }
-
-
 def update_policy(
     train_metrics: MetricsTracker,
     policy: PreTrainedPolicy,
@@ -483,7 +473,8 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
         pin_memory=device.type == "cuda",
         drop_last=False,
         collate_fn=collate_fn,
-        **_dataloader_worker_kwargs(cfg),
+        prefetch_factor=cfg.prefetch_factor if cfg.num_workers > 0 else None,
+        persistent_workers=cfg.persistent_workers and cfg.num_workers > 0,
     )
 
     # Build eval dataloader if a held-out split exists
@@ -509,7 +500,8 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
             pin_memory=device.type == "cuda",
             drop_last=False,
             collate_fn=eval_collate_fn,
-            **_dataloader_worker_kwargs(cfg),
+            prefetch_factor=cfg.prefetch_factor if cfg.num_workers > 0 else None,
+            persistent_workers=cfg.persistent_workers and cfg.num_workers > 0,
         )
 
     # Prepare everything with accelerator

tests/cameras/test_opencv.py

@@ -26,7 +26,7 @@ import cv2
 import numpy as np
 import pytest
 
-from lerobot.cameras.configs import Cv2Rotation
+from lerobot.cameras.configs import ColorMode, Cv2Rotation
 from lerobot.cameras.opencv import OpenCVCamera, OpenCVCameraConfig
 from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
 
@@ -132,6 +132,28 @@ def test_read(index_or_path):
         assert isinstance(img, np.ndarray)
 
 
+@pytest.mark.parametrize("index_or_path", TEST_IMAGE_PATHS, ids=TEST_IMAGE_SIZES)
+def test_color_mode_conversion(index_or_path):
+    """RGB and BGR reads of the same frame must differ only by a channel-axis reversal."""
+    rgb_config = OpenCVCameraConfig(index_or_path=index_or_path, color_mode=ColorMode.RGB, warmup_s=0)
+    bgr_config = OpenCVCameraConfig(index_or_path=index_or_path, color_mode=ColorMode.BGR, warmup_s=0)
+    with OpenCVCamera(rgb_config) as rgb_cam:
+        rgb = rgb_cam.read()
+    with OpenCVCamera(bgr_config) as bgr_cam:
+        bgr = bgr_cam.read()
+
+    assert rgb.shape == bgr.shape
+    np.testing.assert_array_equal(rgb, bgr[..., ::-1])
+
+
+def test_postprocess_invalid_color_mode():
+    config = OpenCVCameraConfig(index_or_path=DEFAULT_PNG_FILE_PATH)
+    camera = OpenCVCamera(config)
+    camera.color_mode = "invalid"
+    with pytest.raises(ValueError):
+        camera._postprocess_image(np.zeros((120, 160, 3), dtype=np.uint8))
+
+
 def test_read_before_connect():
     config = OpenCVCameraConfig(index_or_path=DEFAULT_PNG_FILE_PATH)
 

tests/cameras/test_reachy2_camera.py

@@ -22,6 +22,7 @@ import pytest
 
 pytest.importorskip("reachy2_sdk")
 
+from lerobot.cameras.configs import ColorMode
 from lerobot.cameras.reachy2_camera import Reachy2Camera, Reachy2CameraConfig
 from lerobot.utils.errors import DeviceNotConnectedError
 
@@ -33,28 +34,19 @@ PARAMS = [
 ]
 
 
-def _make_cam_manager_mock():
+def _make_cam_manager_mock(color_frame, depth_frame=None):
     c = MagicMock(name="CameraManagerMock")
 
     teleop = MagicMock(name="TeleopCam")
     teleop.width = 640
     teleop.height = 480
-    teleop.get_frame = MagicMock(
-        side_effect=lambda *_, **__: (
-            np.zeros((480, 640, 3), dtype=np.uint8),
-            time.time(),
-        )
-    )
+    teleop.get_frame = MagicMock(side_effect=lambda *_, **__: (color_frame, time.time()))
 
     depth = MagicMock(name="DepthCam")
     depth.width = 640
     depth.height = 480
-    depth.get_frame = MagicMock(
-        side_effect=lambda *_, **__: (
-            np.zeros((480, 640, 3), dtype=np.uint8),
-            time.time(),
-        )
-    )
+    depth.get_frame = MagicMock(side_effect=lambda *_, **__: (color_frame, time.time()))
+    depth.get_depth_frame = MagicMock(side_effect=lambda *_, **__: (depth_frame, time.time()))
 
     c.is_connected.return_value = True
     c.teleop = teleop
@@ -84,12 +76,14 @@ def _make_cam_manager_mock():
     # ids=["teleop-left", "teleop-right", "torso-rgb", "torso-depth"],
     ids=["teleop-left", "teleop-right", "torso-rgb"],
 )
-def camera(request):
+def camera(request, img_array_factory):
     name, image_type = request.param
+    color_frame = img_array_factory(height=480, width=640)
+    depth_frame = img_array_factory(height=480, width=640, channels=1, dtype=np.uint16)[..., 0]
     with (
         patch(
             "lerobot.cameras.reachy2_camera.reachy2_camera.CameraManager",
-            side_effect=lambda *a, **k: _make_cam_manager_mock(),
+            side_effect=lambda *a, **k: _make_cam_manager_mock(color_frame, depth_frame),
         ),
     ):
         config = Reachy2CameraConfig(name=name, image_type=image_type)
@@ -188,6 +182,41 @@ def test_read_latest_too_old(camera):
         _ = camera.read_latest(max_age_ms=0)  # immediately too old
 
 
+def test_color_mode_conversion(img_array_factory):
+    """teleop frames are native BGR: RGB reverses the channel axis, BGR is passed through."""
+    frame = img_array_factory(height=8, width=8)
+
+    outputs = {}
+    for color_mode in (ColorMode.RGB, ColorMode.BGR):
+        with patch(
+            "lerobot.cameras.reachy2_camera.reachy2_camera.CameraManager",
+            side_effect=lambda *a, **k: _make_cam_manager_mock(frame),
+        ):
+            cam = Reachy2Camera(Reachy2CameraConfig(name="teleop", image_type="left", color_mode=color_mode))
+            cam.connect()
+            outputs[color_mode] = cam.read()
+            cam.disconnect()
+
+    np.testing.assert_array_equal(outputs[ColorMode.BGR], frame)
+    np.testing.assert_array_equal(outputs[ColorMode.RGB], frame[..., ::-1])
+
+
+def test_depth_frame_not_color_converted(img_array_factory):
+    """A depth/depth frame must be returned as-is, without BGR<->RGB conversion."""
+    color_frame = img_array_factory(height=8, width=8)
+    depth = img_array_factory(height=8, width=8, channels=1, dtype=np.uint16)[..., 0]
+    with patch(
+        "lerobot.cameras.reachy2_camera.reachy2_camera.CameraManager",
+        side_effect=lambda *a, **k: _make_cam_manager_mock(color_frame, depth_frame=depth),
+    ):
+        cam = Reachy2Camera(Reachy2CameraConfig(name="depth", image_type="depth"))
+        cam.connect()
+        out = cam.read()
+        cam.disconnect()
+
+    np.testing.assert_array_equal(out, depth)
+
+
 def test_wrong_camera_name():
     with pytest.raises(ValueError):
         _ = Reachy2CameraConfig(name="wrong-name", image_type="left")

tests/cameras/test_realsense.py

@@ -25,7 +25,7 @@ from unittest.mock import patch
 import numpy as np
 import pytest
 
-from lerobot.cameras.configs import Cv2Rotation
+from lerobot.cameras.configs import ColorMode, Cv2Rotation
 from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
 
 pytest.importorskip("pyrealsense2")
@@ -109,6 +109,32 @@ def test_read_depth():
     assert isinstance(img, np.ndarray)
 
 
+# These exercise _postprocess_image directly rather than read(): the bag playback returns
+# non-deterministic frames we can't compare against, and the depth read() path is skipped
+# (see test_read_depth) with the current pyrealsense2 version.
+def test_color_mode_conversion(img_array_factory):
+    """RGB (native for RealSense) is passed through; BGR reverses the channel axis."""
+    color = img_array_factory(height=3, width=4)
+
+    outputs = {}
+    for color_mode in (ColorMode.RGB, ColorMode.BGR):
+        camera = RealSenseCamera(RealSenseCameraConfig(serial_number_or_name="042", color_mode=color_mode))
+        camera.capture_height, camera.capture_width = color.shape[:2]
+        outputs[color_mode] = camera._postprocess_image(color)
+
+    np.testing.assert_array_equal(outputs[ColorMode.RGB], color)
+    np.testing.assert_array_equal(outputs[ColorMode.BGR], color[..., ::-1])
+
+
+def test_depth_frame_not_color_converted(img_array_factory):
+    """Depth frames must bypass color conversion, even when a BGR color_mode is set."""
+    camera = RealSenseCamera(RealSenseCameraConfig(serial_number_or_name="042", color_mode=ColorMode.BGR))
+    depth = img_array_factory(height=3, width=4, channels=1, dtype=np.uint16)[..., 0]
+    camera.capture_height, camera.capture_width = depth.shape
+
+    np.testing.assert_array_equal(camera._postprocess_image(depth, depth_frame=True), depth)
+
+
 def test_read_before_connect():
     config = RealSenseCameraConfig(serial_number_or_name="042")
     camera = RealSenseCamera(config)

tests/datasets/test_dataset_utils.py

@@ -14,16 +14,21 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from types import SimpleNamespace
+from unittest.mock import Mock
+
 import pytest
 import torch
+from packaging.version import Version
 
 pytest.importorskip("datasets", reason="datasets is required (install lerobot[dataset])")
 
 from datasets import Dataset  # noqa: E402
 from huggingface_hub import DatasetCard
 
+import lerobot.datasets.utils as dataset_utils
 from lerobot.datasets.io_utils import hf_transform_to_torch
-from lerobot.datasets.utils import create_lerobot_dataset_card
+from lerobot.datasets.utils import create_lerobot_dataset_card, get_repo_versions, get_safe_version
 from lerobot.utils.constants import ACTION, OBS_IMAGES
 from lerobot.utils.feature_utils import combine_feature_dicts
 
@@ -57,6 +62,30 @@ def test_default_parameters():
     ]
 
 
+@pytest.mark.parametrize("token", ["hf_test_token", True, False])
+def test_get_repo_versions_forwards_token(monkeypatch, token):
+    api = Mock()
+    api.list_repo_refs.return_value = SimpleNamespace(
+        branches=[SimpleNamespace(name="v3.0")],
+        tags=[],
+    )
+    hf_api = Mock(return_value=api)
+    monkeypatch.setattr(dataset_utils, "HfApi", hf_api)
+
+    assert get_repo_versions("private/repo", token=token) == [Version("3.0")]
+    hf_api.assert_called_once_with(token=token)
+    api.list_repo_refs.assert_called_once_with("private/repo", repo_type="dataset")
+
+
+@pytest.mark.parametrize("token", ["hf_test_token", True, False])
+def test_get_safe_version_forwards_token(monkeypatch, token):
+    get_versions = Mock(return_value=[Version("3.0")])
+    monkeypatch.setattr(dataset_utils, "get_repo_versions", get_versions)
+
+    assert get_safe_version("private/repo", "v3.0", token=token) == "v3.0"
+    get_versions.assert_called_once_with("private/repo", token=token)
+
+
 def test_with_tags():
     tags = ["tag1", "tag2"]
     card = create_lerobot_dataset_card(tags=tags)

tests/datasets/test_dataset_writer.py

@@ -204,6 +204,38 @@ def test_clear_resets_buffer(tmp_path):
     assert dataset.writer.episode_buffer["size"] == 0
 
 
+def test_clear_removes_video_frame_staging_dir(tmp_path):
+    """clear_episode_buffer() removes PNG staging dirs for video features."""
+    video_key = "observation.images.cam"
+    features = {
+        video_key: {
+            "dtype": "video",
+            "shape": (64, 96, 3),
+            "names": ["height", "width", "channels"],
+        },
+        "action": {"dtype": "float32", "shape": (2,), "names": None},
+    }
+    dataset = LeRobotDataset.create(
+        repo_id=DUMMY_REPO_ID,
+        fps=DEFAULT_FPS,
+        features=features,
+        root=tmp_path / "ds",
+        use_videos=True,
+    )
+
+    dataset.add_frame(_make_frame(features))
+    video_staging_dir = (
+        dataset.root
+        / Path(DEFAULT_IMAGE_PATH.format(image_key=video_key, episode_index=0, frame_index=0)).parent
+    )
+    assert video_staging_dir.is_dir()
+
+    dataset.clear_episode_buffer()
+
+    assert dataset.writer.episode_buffer["size"] == 0
+    assert not video_staging_dir.exists()
+
+
 def test_finalize_is_idempotent(tmp_path):
     """Calling finalize() twice does not raise."""
     dataset = LeRobotDataset.create(

tests/datasets/test_lerobot_dataset.py

@@ -20,6 +20,7 @@ property delegation, and the full create-record-finalize-read lifecycle.
 """
 
 from pathlib import Path
+from types import SimpleNamespace
 from unittest.mock import Mock
 
 import pytest
@@ -191,6 +192,48 @@ def test_metadata_without_root_uses_hub_cache_snapshot_download(
     }
 
 
+@pytest.mark.parametrize("token", ["hf_test_token", True, False])
+def test_metadata_download_forwards_token(tmp_path, monkeypatch, token):
+    snapshot_root = tmp_path / "snapshot"
+    snapshot_download = Mock(return_value=str(snapshot_root))
+    get_safe_version = Mock(return_value="v3.0")
+    load_metadata = Mock(side_effect=[FileNotFoundError, None])
+    monkeypatch.setattr(dataset_metadata_module, "snapshot_download", snapshot_download)
+    monkeypatch.setattr(dataset_metadata_module, "get_safe_version", get_safe_version)
+    monkeypatch.setattr(LeRobotDatasetMetadata, "_load_metadata", load_metadata)
+
+    meta = LeRobotDatasetMetadata(
+        repo_id=DUMMY_REPO_ID,
+        revision="v3.0",
+        token=token,
+    )
+
+    assert meta.root == snapshot_root
+    assert not hasattr(meta, "_token")
+    get_safe_version.assert_called_once_with(DUMMY_REPO_ID, "v3.0", token=token)
+    assert snapshot_download.call_args.kwargs["token"] is token
+
+
+@pytest.mark.parametrize("token", ["hf_test_token", True, False])
+def test_data_download_forwards_token(tmp_path, monkeypatch, token):
+    snapshot_root = tmp_path / "snapshot"
+    snapshot_download = Mock(return_value=str(snapshot_root))
+    monkeypatch.setattr(lerobot_dataset_module, "snapshot_download", snapshot_download)
+
+    dataset = LeRobotDataset.__new__(LeRobotDataset)
+    dataset.repo_id = DUMMY_REPO_ID
+    dataset.revision = "main"
+    dataset.episodes = None
+    dataset._requested_root = None
+    dataset.meta = SimpleNamespace(root=None)
+    dataset.reader = SimpleNamespace(root=None)
+
+    dataset._download(token=token)
+
+    assert dataset.root == snapshot_root
+    assert snapshot_download.call_args.kwargs["token"] is token
+
+
 def test_without_root_reads_different_revisions_from_distinct_snapshot_roots(
     tmp_path,
     info_factory,

tests/datasets/test_streaming.py

@@ -13,12 +13,16 @@
 # 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 types import SimpleNamespace
+from unittest.mock import Mock
+
 import numpy as np
 import pytest
 import torch
 
 pytest.importorskip("datasets", reason="datasets is required (install lerobot[dataset])")
 
+import lerobot.datasets.streaming_dataset as streaming_dataset_module
 from lerobot.datasets.streaming_dataset import StreamingLeRobotDataset
 from lerobot.datasets.utils import safe_shard
 from lerobot.utils.constants import ACTION
@@ -71,6 +75,40 @@ def get_frames_expected_order(streaming_ds: StreamingLeRobotDataset) -> list[int
     return expected_indices
 
 
+@pytest.mark.parametrize("token", ["hf_test_token", True, False])
+@pytest.mark.parametrize("from_local", [False, True])
+def test_streaming_dataset_forwards_hub_token_only_for_remote_data(tmp_path, monkeypatch, token, from_local):
+    requested_root = tmp_path / "local" if from_local else None
+    metadata = SimpleNamespace(
+        root=requested_root or tmp_path / "snapshot",
+        revision=streaming_dataset_module.CODEBASE_VERSION,
+        _version=streaming_dataset_module.CODEBASE_VERSION,
+        features={},
+        depth_keys=[],
+        image_keys=[],
+        rescale_depth_stats=Mock(),
+    )
+    metadata_cls = Mock(return_value=metadata)
+    load_dataset = Mock(return_value=SimpleNamespace(num_shards=1))
+    monkeypatch.setattr(streaming_dataset_module, "LeRobotDatasetMetadata", metadata_cls)
+    monkeypatch.setattr(streaming_dataset_module, "load_dataset", load_dataset)
+
+    dataset = StreamingLeRobotDataset(DUMMY_REPO_ID, root=requested_root, token=token)
+
+    metadata_cls.assert_called_once_with(
+        DUMMY_REPO_ID,
+        requested_root,
+        streaming_dataset_module.CODEBASE_VERSION,
+        force_cache_sync=False,
+        token=token,
+    )
+    if from_local:
+        assert "token" not in load_dataset.call_args.kwargs
+    else:
+        assert load_dataset.call_args.kwargs["token"] is token
+    assert not hasattr(dataset, "_token")
+
+
 def test_single_frame_consistency(tmp_path, lerobot_dataset_factory):
     """Test if are correctly accessed"""
     ds_num_frames = 400

tests/processor/test_pipeline_from_pretrained_helpers.py

@@ -26,8 +26,17 @@ import tempfile
 from pathlib import Path
 
 import pytest
+import torch
+from safetensors.torch import save_file
 
-from lerobot.processor.pipeline import DataProcessorPipeline, ProcessorMigrationError
+from lerobot.configs import PipelineFeatureType, PolicyFeature
+from lerobot.processor.pipeline import (
+    DataProcessorPipeline,
+    ProcessorMigrationError,
+    ProcessorStep,
+    ProcessorStepRegistry,
+)
+from lerobot.types import EnvTransition
 
 # Simplified Config Loading Tests
 
@@ -98,6 +107,140 @@ def test_load_config_nonexistent_path_tries_hub():
         DataProcessorPipeline._load_config("nonexistent/path", "processor.json", {})
 
 
+def test_from_pretrained_local_directory_missing_state_does_not_call_hub(monkeypatch):
+    """Local processor dirs must fail locally when a state file is missing."""
+
+    @ProcessorStepRegistry.register("local_missing_state_step")
+    class LocalMissingStateStep(ProcessorStep):
+        def __call__(self, transition: EnvTransition) -> EnvTransition:
+            return transition
+
+        def transform_features(
+            self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+        ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+            return features
+
+        def load_state_dict(self, state: dict[str, torch.Tensor]) -> None:
+            pass
+
+    try:
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            tmp_path = Path(tmp_dir)
+            config = {
+                "name": "LocalMissingStatePipeline",
+                "steps": [{"registry_name": "local_missing_state_step", "state_file": "missing.safetensors"}],
+            }
+            (tmp_path / "processor.json").write_text(json.dumps(config))
+
+            def fail_hub_download(*args, **kwargs):
+                pytest.fail("local missing processor state should not call hf_hub_download")
+
+            monkeypatch.setattr("lerobot.processor.pipeline.hf_hub_download", fail_hub_download)
+
+            with pytest.raises(FileNotFoundError, match="missing.safetensors.*local processor pipeline"):
+                DataProcessorPipeline.from_pretrained(tmp_path, config_filename="processor.json")
+    finally:
+        ProcessorStepRegistry.unregister("local_missing_state_step")
+
+
+def test_from_pretrained_local_config_file_missing_state_does_not_call_hub(monkeypatch):
+    """Local single-file processor configs must also keep missing state resolution local."""
+
+    @ProcessorStepRegistry.register("local_file_missing_state_step")
+    class LocalFileMissingStateStep(ProcessorStep):
+        def __call__(self, transition: EnvTransition) -> EnvTransition:
+            return transition
+
+        def transform_features(
+            self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+        ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+            return features
+
+        def load_state_dict(self, state: dict[str, torch.Tensor]) -> None:
+            pass
+
+    try:
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            tmp_path = Path(tmp_dir)
+            config_path = tmp_path / "processor.json"
+            config = {
+                "name": "LocalFileMissingStatePipeline",
+                "steps": [
+                    {"registry_name": "local_file_missing_state_step", "state_file": "missing.safetensors"}
+                ],
+            }
+            config_path.write_text(json.dumps(config))
+
+            def fail_hub_download(*args, **kwargs):
+                pytest.fail("local missing processor state should not call hf_hub_download")
+
+            monkeypatch.setattr("lerobot.processor.pipeline.hf_hub_download", fail_hub_download)
+
+            with pytest.raises(FileNotFoundError, match="missing.safetensors.*local processor pipeline"):
+                DataProcessorPipeline.from_pretrained(config_path, config_filename="ignored.json")
+    finally:
+        ProcessorStepRegistry.unregister("local_file_missing_state_step")
+
+
+def test_from_pretrained_hub_source_missing_local_state_still_calls_hub(monkeypatch, tmp_path):
+    """Hub sources still fall back to hf_hub_download for state files."""
+
+    @ProcessorStepRegistry.register("hub_state_step")
+    class HubStateStep(ProcessorStep):
+        def __init__(self):
+            self.value = torch.tensor(0)
+
+        def __call__(self, transition: EnvTransition) -> EnvTransition:
+            return transition
+
+        def transform_features(
+            self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+        ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+            return features
+
+        def load_state_dict(self, state: dict[str, torch.Tensor]) -> None:
+            self.value = state["value"]
+
+    try:
+        state_path = tmp_path / "downloaded.safetensors"
+        save_file({"value": torch.tensor(7)}, state_path)
+        loaded_config = {
+            "name": "HubStatePipeline",
+            "steps": [{"registry_name": "hub_state_step", "state_file": "hub_state.safetensors"}],
+        }
+        calls = []
+
+        def fake_load_config(cls, model_id, config_filename, hub_download_kwargs):
+            return loaded_config, tmp_path / "hub_cache"
+
+        def fake_hub_download(**kwargs):
+            calls.append(kwargs)
+            return str(state_path)
+
+        monkeypatch.setattr(DataProcessorPipeline, "_load_config", classmethod(fake_load_config))
+        monkeypatch.setattr("lerobot.processor.pipeline.hf_hub_download", fake_hub_download)
+
+        pipeline = DataProcessorPipeline.from_pretrained("user/repo", config_filename="processor.json")
+
+        assert calls == [
+            {
+                "repo_id": "user/repo",
+                "filename": "hub_state.safetensors",
+                "repo_type": "model",
+                "force_download": False,
+                "resume_download": None,
+                "proxies": None,
+                "token": None,
+                "cache_dir": None,
+                "local_files_only": False,
+                "revision": None,
+            }
+        ]
+        assert pipeline.steps[0].value.item() == 7
+    finally:
+        ProcessorStepRegistry.unregister("hub_state_step")
+
+
 # Config Validation Tests
 
 

tests/robots/test_so100_follower.py

@@ -109,3 +109,22 @@ def test_send_action(follower):
 
     goal_pos = {m: (i + 1) * 10 for i, m in enumerate(follower.bus.motors)}
     follower.bus.sync_write.assert_called_once_with("Goal_Position", goal_pos)
+
+
+def test_configure_writes_position_pid_coefficients():
+    bus_mock = _make_bus_mock()
+    bus_mock.motors = ["shoulder_pan"]
+    robot = MagicMock()
+    robot.bus = bus_mock
+    robot.config = SO100FollowerConfig(
+        port="/dev/null",
+        position_p_coefficient=32,
+        position_i_coefficient=1,
+        position_d_coefficient=16,
+    )
+
+    SO100Follower.configure(robot)
+
+    bus_mock.write.assert_any_call("P_Coefficient", "shoulder_pan", 32)
+    bus_mock.write.assert_any_call("I_Coefficient", "shoulder_pan", 1)
+    bus_mock.write.assert_any_call("D_Coefficient", "shoulder_pan", 16)

tests/test_rollout.py

@@ -17,6 +17,8 @@
 from __future__ import annotations
 
 import dataclasses
+import sys
+from types import SimpleNamespace
 from unittest.mock import MagicMock
 
 import pytest
@@ -106,6 +108,109 @@ def test_sentry_config_defaults():
     assert cfg.target_video_file_size_mb is None
 
 
+def test_rollout_config_passes_policy_pretrained_revision(monkeypatch):
+    from lerobot.configs import PreTrainedConfig, parser
+    from lerobot.rollout import RolloutConfig
+    from tests.mocks.mock_robot import MockRobotConfig
+
+    captured = {}
+
+    def fake_from_pretrained(cls, pretrained_name_or_path, **kwargs):
+        captured["pretrained_name_or_path"] = pretrained_name_or_path
+        captured.update(kwargs)
+        return SimpleNamespace(device="cpu", pretrained_revision=kwargs["revision"])
+
+    monkeypatch.setattr(parser, "get_yaml_overrides", lambda _: ["--pretrained_revision=yaml-sha"])
+    monkeypatch.setattr(
+        sys,
+        "argv",
+        ["lerobot-rollout", "--policy.path=user/policy", "--policy.pretrained_revision=cli-sha"],
+    )
+    monkeypatch.setattr(PreTrainedConfig, "from_pretrained", classmethod(fake_from_pretrained))
+
+    cfg = RolloutConfig(robot=MockRobotConfig())
+
+    assert captured["pretrained_name_or_path"] == "user/policy"
+    assert captured["revision"] == "cli-sha"
+    assert captured["cli_overrides"] == [
+        "--pretrained_revision=yaml-sha",
+        "--pretrained_revision=cli-sha",
+    ]
+    assert cfg.policy.pretrained_path == "user/policy"
+    assert cfg.policy.pretrained_revision == "cli-sha"
+
+
+def test_load_pretrained_policy_passes_revision(monkeypatch):
+    import lerobot.rollout.context as rollout_context
+
+    policy_config = SimpleNamespace(
+        type="mock",
+        use_peft=False,
+        pretrained_path="user/policy",
+        pretrained_revision="policy-sha",
+    )
+    policy_class = MagicMock()
+    loaded_policy = MagicMock()
+    policy_class.from_pretrained.return_value = loaded_policy
+    monkeypatch.setattr(rollout_context, "get_policy_class", lambda _: policy_class)
+
+    policy = rollout_context._load_pretrained_policy(policy_config)
+
+    assert policy is loaded_policy
+    policy_class.from_pretrained.assert_called_once_with(
+        "user/policy",
+        config=policy_config,
+        revision="policy-sha",
+    )
+
+
+def test_load_pretrained_peft_policy_keeps_adapter_and_base_revisions_separate(monkeypatch):
+    import lerobot.rollout.context as rollout_context
+
+    policy_config = SimpleNamespace(
+        type="mock",
+        use_peft=True,
+        pretrained_path="user/adapter",
+        pretrained_revision="adapter-sha",
+    )
+    policy_class = MagicMock()
+    base_policy = MagicMock()
+    policy_class.from_pretrained.return_value = base_policy
+    monkeypatch.setattr(rollout_context, "get_policy_class", lambda _: policy_class)
+
+    peft_config = SimpleNamespace(
+        base_model_name_or_path="user/base-policy",
+        revision="base-sha",
+    )
+    peft_config_from_pretrained = MagicMock(return_value=peft_config)
+    adapted_policy = MagicMock()
+    peft_model_from_pretrained = MagicMock(return_value=adapted_policy)
+    monkeypatch.setitem(
+        sys.modules,
+        "peft",
+        SimpleNamespace(
+            PeftConfig=SimpleNamespace(from_pretrained=peft_config_from_pretrained),
+            PeftModel=SimpleNamespace(from_pretrained=peft_model_from_pretrained),
+        ),
+    )
+
+    policy = rollout_context._load_pretrained_policy(policy_config)
+
+    assert policy is adapted_policy
+    peft_config_from_pretrained.assert_called_once_with("user/adapter", revision="adapter-sha")
+    policy_class.from_pretrained.assert_called_once_with(
+        pretrained_name_or_path="user/base-policy",
+        config=policy_config,
+        revision="base-sha",
+    )
+    peft_model_from_pretrained.assert_called_once_with(
+        base_policy,
+        "user/adapter",
+        config=peft_config,
+        revision="adapter-sha",
+    )
+
+
 # ---------------------------------------------------------------------------
 # RolloutRingBuffer
 # ---------------------------------------------------------------------------