chore(format)

* fix(utils): validate precise_sleep spin/margin args

Negative spin_threshold/sleep_margin make remaining arithmetic wrong
and can overshoot. Reject them early; cover the no-op path.

* test: drop flaky wall-clock assertion in no-op test

Per review: the 50ms wall-clock check can exceed its bound on a preempted
CI worker even when precise_sleep returns immediately. The direct calls
already exercise the non-positive no-op path, so the assertion is redundant.

* chore(tests): remove precise_sleep test negative values

---------

Co-authored-by: Bartok9 <danielrpike9@gmail.com>

AGENT_GUIDE.md

@@ -61,16 +61,20 @@ Full details in [`docs/source/so101.mdx`](./docs/source/so101.mdx) and [`docs/so
 **4.1 Install**
 
 ```bash
-pip install 'lerobot[feetech]'              # SO-100/SO-101 motor stack
-# pip install 'lerobot[all]'                # everything
-# pip install 'lerobot[aloha,pusht]'        # specific features
-# pip install 'lerobot[smolvla]'            # add SmolVLA deps
+# uv (recommended — see AGENTS.md and CLAUDE.md)
+uv sync --locked --extra feetech          # SO-100/SO-101 motor stack
+# uv sync --locked --extra all            # everything
+# uv sync --locked --extra smolvla        # add SmolVLA deps
+
+# pip (alternative, e.g. when not working from source)
+# pip install 'lerobot[feetech]'
+# pip install 'lerobot[all]'
+# pip install 'lerobot[smolvla]'
+
 git lfs install && git lfs pull
 hf auth login                             # required to push datasets/policies
 ```
 
-Contributors can alternatively use `uv sync --locked --extra feetech` (see `AGENTS.md`).
-
 **4.2 Find USB ports** — run once per arm, unplug when prompted.
 
 ```bash
@@ -321,11 +325,10 @@ SmolVLA ships with `freeze_vision_encoder=True`. Unfreezing usually **improves p
 
 ```bash
 lerobot-train ... --policy.type=smolvla \
-  --policy.fine_tune_vision_encoder=true
+  --policy.freeze_vision_encoder=false \
+  --policy.train_expert_only=false
 ```
 
-This selectively trains the vision encoder and connector while leaving the language model frozen. Their learning rate defaults to `0.1 × optimizer_lr`; adjust it with `--policy.vision_encoder_lr_multiplier` if needed.
-
 ### 7.7 Signals to stop / keep going
 
 - Train loss plateaus → stop, save a Hub checkpoint.

docs/source/cameras.mdx

@@ -164,8 +164,8 @@ includes the range reported by the sensor. Requesting an unsupported control als
 Omitted controls leave the sensor's existing automatic or manual setting unchanged. These options
 require `use_rgb=True`.
 
-On the RealSense D405, the color stream is provided by the Stereo Module, so changing manual
-exposure or gain also affects the depth stream.
+Manual color controls require a dedicated RGB module. Cameras without one, such as the RealSense
+D405, do not support them and raise an error at connection time.
 
 </hfoption>
 </hfoptions>

docs/source/smolvla.mdx

@@ -70,19 +70,6 @@ cd lerobot && lerobot-train \
   GPU allows it, as long as loading times remain short.
 </Tip>
 
-For tasks that require adapting visual features, such as distinguishing new colors or shapes, selectively
-fine-tune the vision encoder and its connector:
-
-```bash
-lerobot-train ... \
-  --policy.path=lerobot/smolvla_base \
-  --policy.fine_tune_vision_encoder=true
-```
-
-This keeps the language model frozen with the default `train_expert_only=true` setting and trains the vision
-path at `0.1` times the main learning rate by default. Fine-tuning the vision encoder increases memory use and
-can reduce the model's general visual knowledge, so enable it only when the frozen encoder is insufficient.
-
 Fine-tuning is an art. For a complete overview of the options for finetuning, run
 
 ```bash

src/lerobot/cameras/realsense/camera_realsense.py

@@ -365,11 +365,12 @@ class RealSenseCamera(Camera):
         return self._async_read(timeout_ms=10000, read_depth=read_depth)
 
     def _get_color_sensor(self) -> "rs.sensor":
-        """Returns the sensor that controls the color stream.
+        """Returns the dedicated "RGB Camera" sensor that controls the color stream.
 
-        Most RealSense cameras expose "RGB Camera" for color. The D405 has no
-        separate RGB module — its color stream comes from "Stereo Module".
-        We try RGB Camera first, then fall back to Stereo Module.
+        Manual color controls are only applied to a dedicated RGB module. Cameras
+        without one (e.g. the D405, whose color stream comes from the shared
+        "Stereo Module") are unsupported, so we never fall back to another sensor
+        to avoid altering the depth stream.
         """
         if self.rs_profile is None:
             raise RuntimeError(f"{self}: rs_profile must be initialized before use.")
@@ -377,12 +378,14 @@ class RealSenseCamera(Camera):
         device = self.rs_profile.get_device()
         sensors = {s.get_info(rs.camera_info.name): s for s in device.query_sensors()}
 
-        for name in ("RGB Camera", "Stereo Module"):
-            if name in sensors:
-                return sensors[name]
+        if "RGB Camera" in sensors:
+            return sensors["RGB Camera"]
 
         available = list(sensors.keys())
-        raise RuntimeError(f"{self}: no color sensor found. Available sensors: {available}")
+        raise RuntimeError(
+            f"{self}: manual color controls require a dedicated 'RGB Camera' module, which this camera does not have. ",
+            f"Available sensors: {available}.",
+        )
 
     def _set_sensor_option(self, sensor: "rs.sensor", option: "rs.option", value: float, label: str) -> None:
         """Sets a sensor option, re-raising range errors with actionable diagnostics."""

src/lerobot/policies/smolvla/configuration_smolvla.py

@@ -67,8 +67,6 @@ class SmolVLAConfig(PreTrainedConfig):
 
     # Finetuning settings
     freeze_vision_encoder: bool = True
-    fine_tune_vision_encoder: bool = False  # Fine-tune vision + connector; takes priority over freezing.
-    vision_encoder_lr_multiplier: float = 0.1
     train_expert_only: bool = True
     train_state_proj: bool = True
 
@@ -112,12 +110,6 @@ class SmolVLAConfig(PreTrainedConfig):
         super().__post_init__()
 
         """Input validation (not exhaustive)."""
-        if self.fine_tune_vision_encoder:
-            self.freeze_vision_encoder = False
-        if self.vision_encoder_lr_multiplier <= 0:
-            raise ValueError(
-                f"`vision_encoder_lr_multiplier` must be positive, got {self.vision_encoder_lr_multiplier}."
-            )
         if self.n_action_steps > self.chunk_size:
             raise ValueError(
                 f"The chunk size is the upper bound for the number of action steps per model invocation. Got "

src/lerobot/policies/smolvla/modeling_smolvla.py

@@ -186,28 +186,9 @@ class SmolVLAPolicy(PreTrainedPolicy):
             if model_value is not None:
                 model_value.rtc_processor = self.rtc_processor
 
-    def get_optim_params(self):
-        if not self.config.fine_tune_vision_encoder:
+    def get_optim_params(self) -> dict:
         return self.parameters()
 
-        vision_params = []
-        other_params = []
-        for name, param in self.named_parameters():
-            if not param.requires_grad:
-                continue
-            if ".vision_model." in name or ".connector." in name:
-                vision_params.append(param)
-            else:
-                other_params.append(param)
-
-        return [
-            {"params": other_params},
-            {
-                "params": vision_params,
-                "lr": self.config.optimizer_lr * self.config.vision_encoder_lr_multiplier,
-            },
-        ]
-
     def _get_action_chunk(
         self, batch: dict[str, Tensor], noise: Tensor | None = None, **kwargs: Unpack[ActionSelectKwargs]
     ) -> Tensor:
@@ -512,7 +493,6 @@ class VLAFlowMatching(nn.Module):
         self.vlm_with_expert = SmolVLMWithExpertModel(
             model_id=self.config.vlm_model_name,
             freeze_vision_encoder=self.config.freeze_vision_encoder,
-            fine_tune_vision_encoder=self.config.fine_tune_vision_encoder,
             train_expert_only=self.config.train_expert_only,
             load_vlm_weights=self.config.load_vlm_weights,
             attention_mode=self.config.attention_mode,

src/lerobot/policies/smolvla/smolvlm_with_expert.py

@@ -78,7 +78,6 @@ class SmolVLMWithExpertModel(nn.Module):
         load_vlm_weights: bool = True,
         train_expert_only: bool = True,
         freeze_vision_encoder: bool = False,
-        fine_tune_vision_encoder: bool = False,
         attention_mode: str = "self_attn",
         num_expert_layers: int = -1,
         num_vlm_layers: int = -1,
@@ -142,7 +141,6 @@ class SmolVLMWithExpertModel(nn.Module):
         self.num_key_value_heads = self.config.text_config.num_key_value_heads
 
         self.freeze_vision_encoder = freeze_vision_encoder
-        self.fine_tune_vision_encoder = fine_tune_vision_encoder
         self.train_expert_only = train_expert_only
         self.attention_mode = attention_mode
         self.expert_hidden_size = lm_expert_config.hidden_size
@@ -152,6 +150,10 @@ class SmolVLMWithExpertModel(nn.Module):
         return self.vlm.model
 
     def set_requires_grad(self):
+        if self.freeze_vision_encoder:
+            self.get_vlm_model().vision_model.eval()
+            for params in self.get_vlm_model().vision_model.parameters():
+                params.requires_grad = False
         if self.train_expert_only:
             self.vlm.eval()
             for params in self.vlm.parameters():
@@ -174,18 +176,6 @@ class SmolVLMWithExpertModel(nn.Module):
             for name, params in self.vlm.named_parameters():
                 if any(k in name for k in frozen_layers):
                     params.requires_grad = False
-
-        if self.freeze_vision_encoder:
-            self.get_vlm_model().vision_model.eval()
-            for params in self.get_vlm_model().vision_model.parameters():
-                params.requires_grad = False
-
-        if self.fine_tune_vision_encoder:
-            for params in self.get_vlm_model().vision_model.parameters():
-                params.requires_grad = True
-            for params in self.get_vlm_model().connector.parameters():
-                params.requires_grad = True
-
         # To avoid unused params issue with distributed training
         for name, params in self.lm_expert.named_parameters():
             if "lm_head" in name:
@@ -194,15 +184,11 @@ class SmolVLMWithExpertModel(nn.Module):
     def train(self, mode: bool = True):
         super().train(mode)
 
-        if self.train_expert_only:
-            self.vlm.eval()
-
         if self.freeze_vision_encoder:
             self.get_vlm_model().vision_model.eval()
 
-        if self.fine_tune_vision_encoder:
-            self.get_vlm_model().vision_model.train(mode)
-            self.get_vlm_model().connector.train(mode)
+        if self.train_expert_only:
+            self.vlm.eval()
 
     def embed_image(self, image: torch.Tensor):
         patch_attention_mask = None

src/lerobot/rl/buffer.py

@@ -18,7 +18,7 @@ import functools
 import threading
 from collections.abc import Callable, Sequence
 from contextlib import suppress
-from typing import TypedDict
+from typing import NotRequired, TypedDict
 
 import torch
 import torch.nn.functional as F  # noqa: N812
@@ -36,7 +36,7 @@ class BatchTransition(TypedDict):
     next_state: dict[str, torch.Tensor]
     done: torch.Tensor
     truncated: torch.Tensor
-    complementary_info: dict[str, torch.Tensor | float | int] | None = None
+    complementary_info: NotRequired[dict[str, torch.Tensor | float | int] | None]
 
 
 def random_crop_vectorized(images: torch.Tensor, output_size: tuple) -> torch.Tensor:

src/lerobot/scripts/lerobot_find_cameras.py

@@ -28,7 +28,6 @@ lerobot-find-cameras
 # NOTE(Steven): macOS cameras sometimes report different FPS at init time, not an issue here as we don't specify FPS when opening the cameras, but the information displayed might not be truthful.
 
 import argparse
-import concurrent.futures
 import logging
 import time
 from pathlib import Path
@@ -133,7 +132,7 @@ def save_image(
     camera_identifier: str | int,
     images_dir: Path,
     camera_type: str,
-):
+) -> None:
     """
     Saves a single image to disk using Pillow. Handles color conversion if necessary.
     """
@@ -152,7 +151,7 @@ def save_image(
         logger.error(f"Failed to save image for camera {camera_identifier} (type {camera_type}): {e}")
 
 
-def create_camera_instance(cam_meta: dict[str, Any]) -> dict[str, Any] | None:
+def create_camera_instance(cam_meta: dict[str, Any], *, warmup_s: int = 1) -> dict[str, Any] | None:
     """Create and connect to a camera instance based on metadata."""
     cam_type = cam_meta.get("type")
     cam_id = cam_meta.get("id")
@@ -165,12 +164,14 @@ def create_camera_instance(cam_meta: dict[str, Any]) -> dict[str, Any] | None:
             cv_config = OpenCVCameraConfig(
                 index_or_path=cam_id,
                 color_mode=ColorMode.RGB,
+                warmup_s=warmup_s,
             )
             instance = OpenCVCamera(cv_config)
         elif cam_type == "RealSense":
             rs_config = RealSenseCameraConfig(
                 serial_number_or_name=cam_id,
                 color_mode=ColorMode.RGB,
+                warmup_s=warmup_s,
             )
             instance = RealSenseCamera(rs_config)
         else:
@@ -188,9 +189,7 @@ def create_camera_instance(cam_meta: dict[str, Any]) -> dict[str, Any] | None:
         return None
 
 
-def process_camera_image(
-    cam_dict: dict[str, Any], output_dir: Path, current_time: float
-) -> concurrent.futures.Future | None:
+def process_camera_image(cam_dict: dict[str, Any], output_dir: Path, current_time: float) -> None:
     """Capture and process an image from a single camera."""
     cam = cam_dict["instance"]
     meta = cam_dict["meta"]
@@ -200,7 +199,7 @@ def process_camera_image(
     try:
         image_data = cam.read()
 
-        return save_image(
+        save_image(
             image_data,
             cam_id_str,
             output_dir,
@@ -215,10 +214,9 @@ def process_camera_image(
     return None
 
 
-def cleanup_cameras(cameras_to_use: list[dict[str, Any]]):
+def cleanup_camera(cam_dict: dict[str, Any]) -> None:
     """Disconnect all cameras."""
-    logger.info(f"Disconnecting {len(cameras_to_use)} cameras...")
-    for cam_dict in cameras_to_use:
+    logger.info(f"Disconnecting camera with ID {cam_dict['meta'].get('id')}...")
     try:
         if cam_dict["instance"] and cam_dict["instance"].is_connected:
             cam_dict["instance"].disconnect()
@@ -230,6 +228,7 @@ def save_images_from_all_cameras(
     output_dir: Path,
     record_time_s: float = 2.0,
     camera_type: str | None = None,
+    warmup_s: int = 1,
 ):
     """
     Connects to detected cameras (optionally filtered by type) and saves images from each.
@@ -240,6 +239,7 @@ def save_images_from_all_cameras(
         record_time_s: Duration in seconds to record images.
         camera_type: Optional string to filter cameras ("realsense" or "opencv").
                             If None, uses all detected cameras.
+        warmup_s: Duration in seconds to warmup camera before recording images.
     """
     output_dir.mkdir(parents=True, exist_ok=True)
     logger.info(f"Saving images to {output_dir}")
@@ -249,39 +249,23 @@ def save_images_from_all_cameras(
         logger.warning("No cameras detected matching the criteria. Cannot save images.")
         return
 
-    cameras_to_use = []
-    for cam_meta in all_camera_metadata:
-        camera_instance = create_camera_instance(cam_meta)
-        if camera_instance:
-            cameras_to_use.append(camera_instance)
+    logger.info(
+        f"Starting image capture for {record_time_s} seconds from {len(all_camera_metadata)} cameras."
+    )
 
-    if not cameras_to_use:
-        logger.warning("No cameras could be connected. Aborting image save.")
-        return
-
-    logger.info(f"Starting image capture for {record_time_s} seconds from {len(cameras_to_use)} cameras.")
-    start_time = time.perf_counter()
-
-    with concurrent.futures.ThreadPoolExecutor(max_workers=len(cameras_to_use) * 2) as executor:
     try:
+        for cam_meta in all_camera_metadata:
+            cam_dict = create_camera_instance(cam_meta, warmup_s=warmup_s)
+            if cam_dict is None:
+                continue
+            start_time = time.perf_counter()
             while time.perf_counter() - start_time < record_time_s:
-                futures = []
                 current_capture_time = time.perf_counter()
-
-                for cam_dict in cameras_to_use:
-                    future = process_camera_image(cam_dict, output_dir, current_capture_time)
-                    if future:
-                        futures.append(future)
-
-                if futures:
-                    concurrent.futures.wait(futures)
-
+                process_camera_image(cam_dict, output_dir, current_capture_time)
+            cleanup_camera(cam_dict)
     except KeyboardInterrupt:
         logger.info("Capture interrupted by user.")
     finally:
-            print("\nFinalizing image saving...")
-            executor.shutdown(wait=True)
-            cleanup_cameras(cameras_to_use)
         print(f"Image capture finished. Images saved to {output_dir}")
 
 
@@ -291,7 +275,6 @@ def main():
     parser = argparse.ArgumentParser(
         description="Unified camera utility script for listing cameras and capturing images."
     )
-
     parser.add_argument(
         "camera_type",
         type=str,
@@ -309,8 +292,14 @@ def main():
     parser.add_argument(
         "--record-time-s",
         type=float,
-        default=6.0,
-        help="Time duration to attempt capturing frames. Default: 6 seconds.",
+        default=2.0,
+        help="Time duration to attempt capturing frames. Default: 2 seconds.",
+    )
+    parser.add_argument(
+        "--warmup-s",
+        type=int,
+        default=1,
+        help="Time duration to warmup camera before attempting to capture frames. Default: 1 second.",
     )
     args = parser.parse_args()
     save_images_from_all_cameras(**vars(args))

src/lerobot/teleoperators/phone/teleop_phone.py

@@ -171,7 +171,13 @@ class IOSPhone(BasePhone, Teleoperator):
         # HEBI provides orientation in w, x, y, z format.
         # Scipy's Rotation expects x, y, z, w.
         quat_xyzw = np.concatenate((ar_quat[1:], [ar_quat[0]]))  # wxyz to xyzw
+        # ARKit can emit zero/NaN quaternions before tracking is ready or on a
+        # dropped packet. Rotation.from_quat now rejects those; degrade the same
+        # way as a missing pose so teleop stays alive mid-session.
+        try:
             rot = Rotation.from_quat(quat_xyzw)
+        except ValueError:
+            return False, None, None, None
         pos = ar_pos - rot.apply(self.config.camera_offset)
         return True, pos, rot, pose
 

src/lerobot/utils/device_utils.py

@@ -37,16 +37,25 @@ def auto_select_torch_device() -> torch.device:
 
 # TODO(Steven): Remove log. log shouldn't be an argument, this should be handled by the logger level
 def get_safe_torch_device(try_device: str, log: bool = False) -> torch.device:
-    """Given a string, return a torch.device with checks on whether the device is available."""
+    """Given a string, return a torch.device with checks on whether the device is available.
+
+    Raises:
+        ValueError: If the requested device family is known but not available on
+            this machine (``AssertionError`` was previously used and is easy to
+            mistake for a programmer bug under ``python -O`` where asserts vanish).
+    """
     try_device = str(try_device)
     if try_device.startswith("cuda"):
-        assert torch.cuda.is_available()
+        if not torch.cuda.is_available():
+            raise ValueError(f"Requested device {try_device!r} but CUDA is not available.")
         device = torch.device(try_device)
     elif try_device == "mps":
-        assert torch.backends.mps.is_available()
+        if not torch.backends.mps.is_available():
+            raise ValueError("Requested device 'mps' but MPS is not available.")
         device = torch.device("mps")
     elif try_device == "xpu":
-        assert torch.xpu.is_available()
+        if not torch.xpu.is_available():
+            raise ValueError("Requested device 'xpu' but XPU is not available.")
         device = torch.device("xpu")
     elif try_device == "cpu":
         device = torch.device("cpu")

src/lerobot/utils/io_utils.py

@@ -32,21 +32,21 @@ def load_json(fpath: Path) -> Any:
     Returns:
         Any: The data loaded from the JSON file.
     """
-    with open(fpath) as f:
+    with open(fpath, encoding="utf-8") as f:
         return json.load(f)
 
 
-def write_json(data: dict, fpath: Path) -> None:
-    """Write data to a JSON file.
+def write_json(data: JsonLike, fpath: Path) -> None:
+    """Write JSON-serializable data to a file.
 
     Creates parent directories if they don't exist.
 
     Args:
-        data (dict): The dictionary to write.
+        data: JSON-serializable data to write.
         fpath (Path): The path to the output JSON file.
     """
     fpath.parent.mkdir(exist_ok=True, parents=True)
-    with open(fpath, "w") as f:
+    with open(fpath, "w", encoding="utf-8") as f:
         json.dump(data, f, indent=4, ensure_ascii=False)
 
 

src/lerobot/utils/robot_utils.py

@@ -30,6 +30,10 @@ def precise_sleep(seconds: float, spin_threshold: float = 0.010, sleep_margin: f
     """
     if seconds <= 0:
         return
+    if spin_threshold < 0:
+        raise ValueError(f"spin_threshold must be >= 0, got {spin_threshold}")
+    if sleep_margin < 0:
+        raise ValueError(f"sleep_margin must be >= 0, got {sleep_margin}")
 
     system = platform.system()
     # On macOS and Windows the scheduler / sleep granularity can make

src/lerobot/utils/rotation.py

@@ -29,9 +29,12 @@ class Rotation:
     def __init__(self, quat: np.ndarray) -> None:
         """Initialize rotation from quaternion [x, y, z, w]."""
         self._quat = np.asarray(quat, dtype=float)
-        # Normalize quaternion
+        if self._quat.shape != (4,):
+            raise ValueError(f"Quaternion must have shape (4,), got {self._quat.shape}")
+        # Normalize quaternion. Reject the zero vector — it has no orientation.
         norm = np.linalg.norm(self._quat)
-        if norm > 0:
+        if norm <= 0.0 or not np.isfinite(norm):
+            raise ValueError(f"Quaternion must be a non-zero finite vector; got {self._quat} (norm={norm})")
         self._quat = self._quat / norm
 
     @classmethod

src/lerobot/utils/transition.py

@@ -14,7 +14,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import TypedDict
+from typing import NotRequired, TypedDict
 
 import torch
 
@@ -28,7 +28,7 @@ class Transition(TypedDict):
     next_state: dict[str, torch.Tensor]
     done: bool
     truncated: bool
-    complementary_info: dict[str, torch.Tensor | float | int] | None = None
+    complementary_info: NotRequired[dict[str, torch.Tensor | float | int] | None]
 
 
 def move_transition_to_device(transition: Transition, device: str = "cpu") -> Transition:

src/lerobot/utils/utils.py

@@ -24,7 +24,6 @@ import sys
 import time
 from collections.abc import Iterator
 from copy import copy, deepcopy
-from datetime import datetime
 from pathlib import Path
 from statistics import mean
 from typing import TYPE_CHECKING, Any
@@ -61,14 +60,16 @@ def init_logging(
         accelerator: Optional Accelerator instance (for multi-GPU detection)
     """
 
-    def custom_format(record: logging.LogRecord) -> str:
-        dt = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
-        fnameline = f"{record.pathname}:{record.lineno}"
-        pid_str = f"[PID: {os.getpid()}] " if display_pid else ""
-        return f"{record.levelname} {pid_str}{dt} {fnameline[-15:]:>15} {record.getMessage()}"
+    class LeRobotFormatter(logging.Formatter):
+        def format(self, record: logging.LogRecord) -> str:
+            record.lerobot_location = f"{record.pathname}:{record.lineno}"[-15:]
+            record.lerobot_pid = f"[PID: {os.getpid()}] " if display_pid else ""
+            return super().format(record)
 
-    formatter = logging.Formatter()
-    formatter.format = custom_format
+    formatter = LeRobotFormatter(
+        "%(levelname)s %(lerobot_pid)s%(asctime)s %(lerobot_location)15s %(message)s",
+        datefmt="%Y-%m-%d %H:%M:%S",
+    )
 
     logger = logging.getLogger()
     logger.setLevel(logging.NOTSET)

tests/cameras/test_realsense.py

@@ -322,15 +322,16 @@ def test_get_color_sensor_prefers_rgb_camera():
     assert camera._get_color_sensor() is rgb
 
 
-def test_get_color_sensor_falls_back_to_stereo_module():
-    """D405 has no separate RGB module; color comes from Stereo Module."""
+def test_get_color_sensor_raises_without_dedicated_rgb_module():
+    """D405 has no separate RGB module; we refuse to touch the shared Stereo Module."""
     config = RealSenseCameraConfig(serial_number_or_name="042")
     camera = RealSenseCamera(config)
 
     stereo = _make_mock_sensor("Stereo Module")
     _attach_mock_color_sensor(camera, stereo)
 
-    assert camera._get_color_sensor() is stereo
+    with pytest.raises(RuntimeError, match="dedicated 'RGB Camera' module"):
+        camera._get_color_sensor()
 
 
 def test_get_color_sensor_raises_with_available_sensors():

tests/utils/test_device_utils.py

@@ -0,0 +1,36 @@
+#!/usr/bin/env python
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from unittest.mock import patch
+
+import pytest
+import torch
+
+from lerobot.utils.device_utils import get_safe_torch_device, is_torch_device_available
+
+
+def test_cpu_always_available():
+    assert get_safe_torch_device("cpu") == torch.device("cpu")
+    assert is_torch_device_available("cpu")
+
+
+def test_missing_cuda_raises_valueerror():
+    with patch("torch.cuda.is_available", return_value=False), pytest.raises(ValueError, match="CUDA"):
+        get_safe_torch_device("cuda")
+
+
+def test_missing_mps_raises_valueerror():
+    with patch("torch.backends.mps.is_available", return_value=False), pytest.raises(ValueError, match="MPS"):
+        get_safe_torch_device("mps")

tests/utils/test_rotation.py

@@ -0,0 +1,46 @@
+#!/usr/bin/env python
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import numpy as np
+import pytest
+
+from lerobot.utils.rotation import Rotation
+
+
+def test_zero_quaternion_rejected():
+    with pytest.raises(ValueError, match="non-zero"):
+        Rotation(np.zeros(4))
+
+
+def test_non_finite_quaternion_rejected():
+    with pytest.raises(ValueError, match="non-zero|finite"):
+        Rotation(np.array([np.nan, 0.0, 0.0, 1.0]))
+
+
+def test_wrong_shape_rejected():
+    with pytest.raises(ValueError, match="shape"):
+        Rotation(np.array([1.0, 0.0, 0.0]))
+
+
+def test_identity_roundtrip():
+    r = Rotation.from_rotvec(np.zeros(3))
+    assert np.allclose(r.as_rotvec(), 0.0)
+    assert np.allclose(r.as_matrix(), np.eye(3))
+
+
+def test_rotvec_roundtrip():
+    rotvec = np.array([0.1, -0.2, 0.3])
+    r = Rotation.from_rotvec(rotvec)
+    assert np.allclose(r.as_rotvec(), rotvec, atol=1e-6)