docs(update): updating docs with the task modification features

Part of #2326.

Signed-off-by: 陈伟 <woshei0a0a0a@qq.com>

.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: ["*"]

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

docker/Dockerfile.internal

@@ -68,17 +68,16 @@ ENV HOME=/home/user_lerobot \
 # issues with MuJoCo and OpenGL drivers.
 RUN uv venv --python python${PYTHON_VERSION}
 
-# Install Python dependencies for caching
+# Install third-party dependencies separately for layer caching
 COPY --chown=user_lerobot:user_lerobot setup.py pyproject.toml uv.lock README.md MANIFEST.in ./
-COPY --chown=user_lerobot:user_lerobot src/ src/
-
-RUN uv sync --locked --extra all --no-cache
+RUN uv sync --locked --extra all --no-install-project --no-cache
 
 RUN chmod +x /lerobot/.venv/lib/python${PYTHON_VERSION}/site-packages/triton/backends/nvidia/bin/ptxas
 
-# Copy the rest of the application source code
+# Copy the application source code and install the local project
 # Make sure to have the git-LFS files for testing
 COPY --chown=user_lerobot:user_lerobot . .
+RUN uv sync --locked --extra all --no-cache
 
 # Set the default command
 CMD ["/bin/bash"]

docker/Dockerfile.user

@@ -60,15 +60,14 @@ ENV HOME=/home/user_lerobot \
 # run other Python projects in the same container without dependency conflicts.
 RUN uv venv
 
-# Install Python dependencies for caching
+# Install third-party dependencies separately for layer caching
 COPY --chown=user_lerobot:user_lerobot setup.py pyproject.toml uv.lock README.md MANIFEST.in ./
-COPY --chown=user_lerobot:user_lerobot src/ src/
+RUN uv sync --locked --extra all --no-install-project --no-cache
 
-RUN uv sync --locked --extra all --no-cache
-
-# Copy the rest of the application code
+# Copy the application code and install the local project
 # Make sure to have the git-LFS files for testing
 COPY --chown=user_lerobot:user_lerobot . .
+RUN uv sync --locked --extra all --no-cache
 
 # Set the default command
 CMD ["/bin/bash"]

docs/source/backwardcomp.mdx

@@ -58,7 +58,7 @@ final_action = postprocessor(action)
 
 ## Hardware API redesign
 
-PR [#777](https://github.com/huggingface/lerobot/pull/777) improves the LeRobot calibration but is **not backward-compatible**. Below is a overview of what changed and how you can continue to work with datasets created before this pull request.
+PR [#777](https://github.com/huggingface/lerobot/pull/777) improves the LeRobot calibration but is **not backward-compatible**. Below is an overview of what changed and how you can continue to work with datasets created before this pull request.
 
 ### What changed?
 
@@ -129,8 +129,8 @@ python examples/backward_compatibility/replay.py \
 
 Policies output actions in the same format as the datasets (`torch.Tensors`). Therefore, the same transformations should be applied.
 
-To find these transformations, we recommend to first try and and replay an episode of the dataset your policy was trained on using the section above.
-Then, add these same transformations on your inference script (shown here in the `record.py` script):
+To find these transformations, we recommend first replaying an episode of the dataset your policy was trained on using the section above.
+Then, add these same transformations to your inference script (shown here in the `record.py` script):
 
 ```diff
 action_values = predict_action(

docs/source/cameras.mdx

@@ -136,6 +136,10 @@ config = RealSenseCameraConfig(
     height=480,
     color_mode=ColorMode.RGB,
     use_depth=True,
+    # Optional fixed color controls. Omit them to leave the current sensor settings unchanged.
+    exposure=120,
+    gain=64,
+    white_balance=4600,
     rotation=Cv2Rotation.NO_ROTATION
 )
 
@@ -154,6 +158,15 @@ finally:
 ```
 <!-- prettier-ignore-end -->
 
+Manual color controls disable the corresponding automatic exposure or white-balance mode. Their
+supported ranges vary by camera model; an invalid value raises an error at connection time that
+includes the range reported by the sensor. Requesting an unsupported control also raises an error.
+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.
+
 </hfoption>
 </hfoptions>
 

docs/source/env_processor.mdx

@@ -88,20 +88,6 @@ policy_preprocessor = NormalizerProcessorStep(stats=dataset_stats)
 
 The same policy can work with different environment processors, and the same environment processor can work with different policies:
 
-````python
-# Use SmolVLA policy with LIBERO environment
-# Use SmolVLA policy with LIBERO environment
-libero_preprocessor, libero_postprocessor = make_env_pre_post_processors(
-    env_cfg=libero_cfg,
-    policy_cfg=smolvla_cfg,
-)
-smolvla_preprocessor, smolvla_postprocessor = make_pre_post_processors(smolvla_cfg)
-# Or use ACT policy with the same LIBERO environment
-libero_preprocessor, libero_postprocessor = make_env_pre_post_processors(
-    env_cfg=libero_cfg,
-    policy_cfg=act_cfg,
-)
-act_preprocessor, act_postprocessor = make_pre_post_processors(act_cfg)
 ```python
 # Use SmolVLA policy with LIBERO environment
 libero_preprocessor, libero_postprocessor = make_env_pre_post_processors(
@@ -116,6 +102,7 @@ libero_preprocessor, libero_postprocessor = make_env_pre_post_processors(
     policy_cfg=act_cfg,
 )
 act_preprocessor, act_postprocessor = make_pre_post_processors(act_cfg)
+```
 
 ### 3. **Easier Experimentation**
 
@@ -145,7 +132,7 @@ class LiberoVelocityProcessorStep(ObservationProcessorStep):
         state = torch.cat([eef_pos, eef_axisangle, eef_vel,
                           gripper_pos, gripper_vel], dim=-1)  # 14D
         return state
-````
+```
 
 ### 4. **Cleaner Environment Code**
 

docs/source/feetech.mdx

@@ -40,10 +40,10 @@ This tutorial guides you through updating the firmware of Feetech motors using t
 For each motor you want to update:
 
 1. **Select the motor** from the list by clicking on it
-2. **Click on Upgrade tab**:
-3. **Click on Online button**:
-   - If an potential firmware update is found, it will be displayed in the box
-4. **Click on Upgrade button**:
+2. **Click the Upgrade tab**:
+3. **Click the Online button**:
+   - If a potential firmware update is found, it will be displayed in the box
+4. **Click the Upgrade button**:
    - The update progress will be displayed
 
 ## Step 6: Verify Update

docs/source/hil_data_collection.mdx

@@ -59,6 +59,7 @@ The `lerobot-rollout --strategy.type=dagger` mode requires **teleoperators with
 
 - `bi_openarm_mini` - Bimanual OpenArm Mini
 - `so_leader` - SO100 / SO101 leader arm
+- `bi_so_leader` - Bimanual SO100 / SO101 leader arms
 
 > [!IMPORTANT]
 > The provided commands default to `bi_openarm_follower` + `bi_openarm_mini`.

docs/source/hope_jr.mdx

@@ -211,7 +211,7 @@ Record, Replay and Train with Hope-JR is still experimental.
 
 ### Record
 
-This step records the dataset, which can be seen as an example [here](https://huggingface.co/datasets/nepyope/hand_record_test_with_video_data/settings).
+This step records the dataset, which can be seen as an example [here](https://huggingface.co/datasets/nepyope/hand_record_test_with_video_data).
 
 ```bash
 lerobot-record \

docs/source/inference.mdx

@@ -187,14 +187,13 @@ Teleop is optional — if omitted the robot holds its position during the reset
 | `ESC`       | Stop the recording session       |
 
 | Flag                                            | Description                                                                |
-| ----------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| ----------------------------------------------- | -------------------------------------------------------------------------- |
 | `--dataset.num_episodes`                        | Number of episodes to record                                               |
 | `--dataset.episode_time_s`                      | Duration of each recording episode in seconds                              |
 | `--dataset.reset_time_s`                        | Duration of the reset phase between episodes in seconds                    |
 | `--teleop.type`                                 | Optional. Teleoperator to drive the robot during resets                    |
 | `--strategy.reset_to_initial_position`          | Whether to reset the robot to its initial position between episodes        |
 | `--strategy.smooth_leader_to_follower_handover` | Whether to turn on or off the leader -> follower smooth handover behavior. |
-| `--strategy.smooth_handover`                    | Smoothly hand control to the teleop at reset start (default: true). Disable for clutch-style teleops that re-reference at the current robot pose on engage |
 
 ---
 

docs/source/integrate_hardware.mdx

@@ -18,7 +18,7 @@ If you're using Feetech or Dynamixel motors, LeRobot provides built-in bus inter
 - [`DynamixelMotorsBus`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/motors/dynamixel/dynamixel.py) – for controlling Dynamixel servos
 
 Please refer to the [`MotorsBus`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/motors/motors_bus.py) abstract class to learn about its API.
-For a good example of how it can be used, you can have a look at our own [SO101 follower implementation](https://github.com/huggingface/lerobot/blob/main/src/lerobot/robots/so_follower/so101_follower/so101_follower.py)
+For a good example of how it can be used, you can have a look at our own [SO101 follower implementation](https://github.com/huggingface/lerobot/blob/main/src/lerobot/robots/so_follower/so_follower.py)
 
 Use these if compatible. Otherwise, you'll need to find or write a Python interface (not covered in this tutorial):
 

docs/source/lekiwi.mdx

@@ -51,7 +51,7 @@ In addition to these instructions, you need to install the Feetech SDK & ZeroMQ
 pip install -e ".[lekiwi]"
 ```
 
-Great :hugs:! You are now done installing LeRobot, and we can begin assembling the SO100/SO101 arms and the mobile base :robot:.
+Great 🤗! You are now done installing LeRobot, and we can begin assembling the SO100/SO101 arms and the mobile base 🤖.
 Every time you now want to use LeRobot, you can go to the `~/lerobot` folder where we installed LeRobot and run one of the commands.
 
 # Step-by-Step Assembly Instructions

docs/source/pi0fast.mdx

@@ -174,7 +174,7 @@ The model takes images, text instructions, and robot state as input, and outputs
 
 ## Reproducing π₀Fast results
 
-We reproduce the results of π₀Fast on the LIBERO benchmark using the LeRobot implementation. We take the LeRobot PiFast base model [lerobot/pi0fast-base](https://huggingface.co/lerobot/pi0fast-base) and finetune for an additional 40kk steps in bfloat16, with batch size of 256 on 8 H100 GPUs using the [HuggingFace LIBERO dataset](https://huggingface.co/datasets/HuggingFaceVLA/libero).
+We reproduce the results of π₀Fast on the LIBERO benchmark using the LeRobot implementation. We take the LeRobot PiFast base model [lerobot/pi0fast-base](https://huggingface.co/lerobot/pi0fast-base) and finetune for an additional 40k steps in bfloat16, with batch size of 256 on 8 H100 GPUs using the [HuggingFace LIBERO dataset](https://huggingface.co/datasets/HuggingFaceVLA/libero).
 
 The finetuned model can be found here:
 

docs/source/processors_robots_teleop.mdx

@@ -22,7 +22,7 @@ With processors, you choose the learning features you want to use for your polic
 ## Three pipelines
 
 We often compose three pipelines. Depending on your setup, some can be empty if action and observation spaces already match.
-Each of these pipelines handle different conversions between different action and observation spaces. Below is a quick explanation of each pipeline.
+Each of these pipelines handles different conversions between different action and observation spaces. Below is a quick explanation of each pipeline.
 
 1. Pipeline 1: Teleop action space → dataset action space (phone pose → EE targets)
 2. Pipeline 2: Dataset action space → robot command space (EE targets → joints)
@@ -74,15 +74,15 @@ In the phone to SO-100 follower examples we use the following adapters:
 - `robot_action_to_transition`: transforms the teleop action dict to a pipeline transition.
 - `transition_to_robot_action`: transforms the pipeline transition to a robot action dict.
 - `observation_to_transition`: transforms the robot observation dict to a pipeline transition.
-- `transition_to_observation`: transforms the pipeline transition to a observation dict.
+- `transition_to_observation`: transforms the pipeline transition to an observation dict.
 
-Checkout [src/lerobot/processor/converters.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/processor/converters.py) for more details.
+Check out [src/lerobot/processor/converters.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/processor/converters.py) for more details.
 
 ## Dataset feature contracts
 
 Dataset features are determined by the keys saved in the dataset. Each step can declare what features it modifies in a contract called `transform_features(...)`. Once you build a processor, the processor can then aggregate all of these features with `aggregate_pipeline_dataset_features()` and merge multiple feature dicts with `combine_feature_dicts(...)`.
 
-Below is and example of how we declare features with the `transform_features` method in the phone to SO-100 follower examples:
+Below is an example of how we declare features with the `transform_features` method in the phone to SO-100 follower examples:
 
 ```python
     def transform_features(

docs/source/rtc.mdx

@@ -57,7 +57,7 @@ policy_cfg.rtc_config = RTCConfig(
 policy = PI0Policy.from_pretrained("lerobot/pi0_base", policy_cfg=policy_cfg, device="cuda")
 
 # Now use predict_action_chunk with RTC parameters
-inference_delay = 4  # How many steps of inference latency, this values should be calculated based on the inference latency of the policy
+inference_delay = 4  # How many steps of inference latency, this value should be calculated based on the inference latency of the policy
 
 # Initialize the action queue
 action_queue = ActionQueue(policy_cfg.rtc_config)
@@ -100,7 +100,7 @@ Typical values: 8-12 steps
 RTCConfig(execution_horizon=10)
 ```
 
-**`max_guidance_weight`**: How strongly to enforce consistency with the previous chunk. This is a hyperparameter that can be tuned to balance the smoothness of the transitions and the reactivity of the policy. For 10 steps flow matching (SmolVLA, Pi0, Pi0.5), a value of 10.0 is a optimal value.
+**`max_guidance_weight`**: How strongly to enforce consistency with the previous chunk. This is a hyperparameter that can be tuned to balance the smoothness of the transitions and the reactivity of the policy. For 10 steps flow matching (SmolVLA, Pi0, Pi0.5), a value of 10.0 is an optimal value.
 
 **`prefix_attention_schedule`**: How to weight consistency across the overlap region.
 

docs/source/smolvla.mdx

@@ -93,7 +93,7 @@ lerobot-train --help
 
 ## Evaluate the finetuned model and run it in real-time
 
-Similarly for when recording an episode, it is recommended that you are logged in to the HuggingFace Hub. You can follow the corresponding steps: [Record a dataset](./il_robots).
+Similarly for when recording an episode, it is recommended that you are logged in to the HuggingFace Hub. You can follow the corresponding steps: [Record a dataset](./il_robots#record-a-dataset).
 Once you are logged in, you can run inference in your setup by doing:
 
 ```bash

docs/source/so101.mdx

@@ -338,7 +338,7 @@ It is advisable to install one 3-pin cable in the motor after placing them befor
 <hfoption id="Leader">
 
 - Mount the leader holder onto the wrist and secure it with 4 M3x6mm screws.
-- Attach the handle to motor 5 using 1 M2x6mm screw.
+- Attach the handle to the leader holder using 1 M2x6mm screw.
 - Insert the gripper motor, secure it with 2 M2x6mm screws on each side, attach a motor horn using a M3x6mm horn screw.
 - Attach the follower trigger with 4 M3x6mm screws.
 

docs/source/using_dataset_tools.mdx

@@ -11,9 +11,10 @@ LeRobot provides several utilities for manipulating datasets:
 3. **Merge Datasets** - Combine multiple datasets into one. The datasets must have identical features, and episodes are concatenated in the order specified in `repo_ids`
 4. **Add Features** - Add new features to a dataset
 5. **Remove Features** - Remove features from a dataset
-6. **Convert to Video** - Convert image-based datasets to video format for efficient storage (RGB and depth cameras are encoded with separate encoders)
-7. **Re-encode Videos** - Re-encode an existing video dataset's RGB and/or depth streams with new encoder settings
-8. **Show the Info of Datasets** - Show the summary of datasets information such as number of episode etc.
+6. **Modify Tasks** - Change the natural-language task descriptions associated with episodes
+7. **Convert to Video** - Convert image-based datasets to video format for efficient storage (RGB and depth cameras are encoded with separate encoders)
+8. **Re-encode Videos** - Re-encode an existing video dataset's RGB and/or depth streams with new encoder settings
+9. **Show the Info of Datasets** - Show the summary of datasets information such as number of episode etc.
 
 The core implementation is in `lerobot.datasets.dataset_tools`.
 An example script detailing how to use the tools API is available in `examples/dataset/use_dataset_tools.py`.
@@ -50,11 +51,11 @@ lerobot-edit-dataset \
 Divide a dataset into multiple subsets.
 
 ```bash
-# Split by fractions (e.g. 80% train, 20% test, 20% val)
+# Split by fractions (e.g. 60% train, 20% val, 20% test)
 lerobot-edit-dataset \
     --repo_id lerobot/pusht \
     --operation.type split \
-    --operation.splits '{"train": 0.8, "test": 0.2, "val": 0.2}'
+    --operation.splits '{"train": 0.6, "val": 0.2, "test": 0.2}'
 
 # Split by specific episode indices
 lerobot-edit-dataset \
@@ -89,6 +90,53 @@ lerobot-edit-dataset \
     --operation.feature_names "['observation.images.top']"
 ```
 
+#### Modify Tasks
+
+Change the natural-language task descriptions attached to episodes. This is useful for fixing typos, standardizing wording, or re-labeling episodes.
+
+> [!WARNING]
+> `modify_tasks` modifies the dataset **in-place** (updating `meta/tasks.parquet`, the `task_index` column in the data files, the `tasks` column in the episode metadata, and `total_tasks` in `meta/info.json`). The `--new_repo_id` and `--new_root` parameters are ignored for this operation.
+
+```bash
+# Set a single task for all episodes
+lerobot-edit-dataset \
+    --repo_id lerobot/pusht \
+    --operation.type modify_tasks \
+    --operation.new_task "Pick up the cube and place it"
+
+# Set different tasks for specific episodes
+lerobot-edit-dataset \
+    --repo_id lerobot/pusht \
+    --operation.type modify_tasks \
+    --operation.episode_tasks '{"0": "Task A", "1": "Task B", "2": "Task A"}'
+
+# Replace existing task strings wherever they appear
+lerobot-edit-dataset \
+    --repo_id lerobot/pusht \
+    --operation.type modify_tasks \
+    --operation.task_replacements '{"Pick up the red cube": "Lift the red cube"}'
+
+# Combine modes in a single run
+lerobot-edit-dataset \
+    --repo_id lerobot/pusht \
+    --operation.type modify_tasks \
+    --operation.new_task "Default task" \
+    --operation.task_replacements '{"Pick up the red cube": "Lift the red cube"}' \
+    --operation.episode_tasks '{"5": "Special task for episode 5"}'
+```
+
+**Parameters:**
+
+- `new_task`: A single task string used as the default for episodes not otherwise covered.
+- `episode_tasks`: Mapping from episode index to task string.
+- `task_replacements`: Mapping from existing task strings to their replacements, applied to episodes whose current task matches a key. Every key must be an existing task in the dataset.
+
+The modes can be combined in a single run. Per episode, the task is resolved with the following precedence:
+
+`episode_tasks` > `task_replacements` > `new_task` > original task
+
+At least one of `new_task`, `episode_tasks`, or `task_replacements` must be specified. An episode that ends up with no task raises an error.
+
 #### Convert to Video
 
 Convert an image-based dataset to video format, creating a new LeRobotDataset where images are stored as videos. This is useful for reducing storage requirements and improving data loading performance. The new dataset will have the exact same structure as the original, but with images encoded as MP4 videos in the proper LeRobot format.

pyproject.toml

@@ -67,7 +67,7 @@ dependencies = [
     "einops>=0.8.0,<0.9.0",
 
     # Config & Hub
-    "draccus==0.10.0", # TODO: Relax version constraint
+    "draccus>=0.11.6,<0.12.0",
     "huggingface-hub>=1.0.0,<2.0.0",
     "requests>=2.32.0,<3.0.0",
 
@@ -494,6 +494,19 @@ ignore_errors = true
 module = "lerobot.envs.*"
 ignore_errors = false
 
+[[tool.mypy.overrides]]
+module = "lerobot.annotations.*"
+ignore_errors = false
+disallow_untyped_defs = true
+disallow_incomplete_defs = true
+check_untyped_defs = true
+
+[[tool.mypy.overrides]]
+module = "lerobot.transforms.*"
+ignore_errors = false
+disallow_untyped_defs = true
+disallow_incomplete_defs = true
+check_untyped_defs = true
 
 # [[tool.mypy.overrides]]
 # module = "lerobot.utils.*"

src/lerobot/cameras/opencv/camera_opencv.py

@@ -120,14 +120,22 @@ class OpenCVCamera(Camera):
         self.rotation: int | None = get_cv2_rotation(config.rotation)
         self.backend: int = config.backend
 
-        if self.height and self.width:
-            self.capture_width, self.capture_height = self.width, self.height
-            if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE]:
-                self.capture_width, self.capture_height = self.height, self.width
+        self.capture_width: int | None = None
+        self.capture_height: int | None = None
+        self._reset_connection_settings()
 
     def __str__(self) -> str:
         return f"{self.__class__.__name__}({self.index_or_path})"
 
+    def _reset_connection_settings(self) -> None:
+        """Restore settings that may have been auto-detected during a failed connection."""
+        self.fps = self.config.fps
+        self.width = self.config.width
+        self.height = self.config.height
+        self.capture_width, self.capture_height = self.width, self.height
+        if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE]:
+            self.capture_width, self.capture_height = self.height, self.width
+
     @property
     def is_connected(self) -> bool:
         """Checks if the camera is currently connected and opened."""
@@ -164,6 +172,7 @@ class OpenCVCamera(Camera):
                 f"Failed to open {self}.Run `lerobot-find-cameras opencv` to find available cameras."
             )
 
+        try:
             self._configure_capture_settings()
             self._start_read_thread()
 
@@ -175,6 +184,13 @@ class OpenCVCamera(Camera):
                 with self.frame_lock:
                     if self.latest_frame is None:
                         raise ConnectionError(f"{self} failed to capture frames during warmup.")
+        except BaseException:
+            try:
+                self._cleanup_resources()
+            except Exception:
+                logger.exception(f"Failed to fully clean up {self} after connect() failed.")
+            self._reset_connection_settings()
+            raise
 
         logger.info(f"{self} connected.")
 
@@ -312,6 +328,7 @@ class OpenCVCamera(Camera):
 
         for target in targets_to_scan:
             camera = cv2.VideoCapture(target)
+            try:
                 if camera.isOpened():
                     default_width = int(camera.get(cv2.CAP_PROP_FRAME_WIDTH))
                     default_height = int(camera.get(cv2.CAP_PROP_FRAME_HEIGHT))
@@ -321,7 +338,9 @@ class OpenCVCamera(Camera):
                     # Get FOURCC code and convert to string
                     default_fourcc_code = camera.get(cv2.CAP_PROP_FOURCC)
                     default_fourcc_code_int = int(default_fourcc_code)
-                default_fourcc = "".join([chr((default_fourcc_code_int >> 8 * i) & 0xFF) for i in range(4)])
+                    default_fourcc = "".join(
+                        [chr((default_fourcc_code_int >> 8 * i) & 0xFF) for i in range(4)]
+                    )
 
                     camera_info = {
                         "name": f"OpenCV Camera @ {target}",
@@ -338,6 +357,7 @@ class OpenCVCamera(Camera):
                     }
 
                     found_cameras_info.append(camera_info)
+            finally:
                 camera.release()
 
         return found_cameras_info
@@ -496,6 +516,26 @@ class OpenCVCamera(Camera):
             self.latest_timestamp = None
             self.new_frame_event.clear()
 
+    def _cleanup_resources(self) -> None:
+        """Stop background reads and release the capture, including after partial setup."""
+        read_thread = self.thread
+        videocapture = self.videocapture
+
+        try:
+            self._stop_read_thread()
+        finally:
+            self.videocapture = None
+            try:
+                if videocapture is not None:
+                    videocapture.release()
+            finally:
+                # Releasing the device may unblock a hardware read that outlived
+                # the first bounded join in _stop_read_thread().
+                if read_thread is not None and read_thread.is_alive():
+                    read_thread.join(timeout=2.0)
+                    if read_thread.is_alive():  # pragma: no cover
+                        logger.warning(f"{self} read thread remained alive after releasing the capture.")
+
     @check_if_not_connected
     def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
         """
@@ -586,16 +626,6 @@ class OpenCVCamera(Camera):
         if not self.is_connected and self.thread is None:
             raise DeviceNotConnectedError(f"{self} not connected.")
 
-        if self.thread is not None:
-            self._stop_read_thread()
-
-        if self.videocapture is not None:
-            self.videocapture.release()
-            self.videocapture = None
-
-        with self.frame_lock:
-            self.latest_frame = None
-            self.latest_timestamp = None
-            self.new_frame_event.clear()
+        self._cleanup_resources()
 
         logger.info(f"{self} disconnected.")

src/lerobot/cameras/realsense/camera_realsense.py

@@ -121,6 +121,9 @@ class RealSenseCamera(Camera):
 
         self.config = config
 
+        self.width: int | None = config.width
+        self.height: int | None = config.height
+
         if config.serial_number_or_name.isdigit():
             self.serial_number = config.serial_number_or_name
         else:
@@ -131,6 +134,9 @@ class RealSenseCamera(Camera):
         self.use_rgb = config.use_rgb
         self.use_depth = config.use_depth
         self.warmup_s = config.warmup_s
+        self.exposure: int | None = config.exposure
+        self.gain: int | None = config.gain
+        self.white_balance: int | None = config.white_balance
 
         self.rs_pipeline: rs.pipeline | None = None
         self.rs_profile: rs.pipeline_profile | None = None
@@ -145,14 +151,23 @@ class RealSenseCamera(Camera):
 
         self.rotation: int | None = get_cv2_rotation(config.rotation)
 
-        if self.height and self.width:
-            self.capture_width, self.capture_height = self.width, self.height
-            if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE]:
-                self.capture_width, self.capture_height = self.height, self.width
+        self.capture_width: int | None = None
+        self.capture_height: int | None = None
+        self._reset_connection_settings()
 
     def __str__(self) -> str:
         return f"{self.__class__.__name__}({self.serial_number})"
 
+    def _reset_connection_settings(self) -> None:
+        """Restore settings that may have been auto-detected during a failed connection."""
+        self.fps = self.config.fps
+        self.width = self.config.width
+        self.height = self.config.height
+        self.warmup_s = self.config.warmup_s
+        self.capture_width, self.capture_height = self.width, self.height
+        if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE]:
+            self.capture_width, self.capture_height = self.height, self.width
+
     @property
     def is_connected(self) -> bool:
         """Checks if the camera pipeline is started and streams are active."""
@@ -172,7 +187,8 @@ class RealSenseCamera(Camera):
 
         Raises:
             DeviceAlreadyConnectedError: If the camera is already connected.
-            ValueError: If the configuration is invalid (e.g., missing serial/name, name not unique).
+            ValueError: If the configuration is invalid, a requested sensor option is unsupported,
+                or a requested sensor value is invalid.
             ConnectionError: If the camera is found but fails to start the pipeline or no RealSense devices are detected at all.
             RuntimeError: If the pipeline starts but fails to apply requested settings.
         """
@@ -190,7 +206,9 @@ class RealSenseCamera(Camera):
                 f"Failed to open {self}.Run `lerobot-find-cameras realsense` to find available cameras."
             ) from e
 
+        try:
             self._configure_capture_settings()
+            self._configure_sensor_options()
             self._start_read_thread()
 
             # NOTE(Steven/Caroline): Enforcing at least one second of warmup as RS cameras need a bit of time before the first read. If we don't wait, the first read from the warmup will raise.
@@ -206,6 +224,13 @@ class RealSenseCamera(Camera):
                     self.use_depth and self.latest_depth_frame is None
                 ):
                     raise ConnectionError(f"{self} failed to capture frames during warmup.")
+        except BaseException:
+            try:
+                self._cleanup_resources()
+            except Exception:
+                logger.exception(f"Failed to fully clean up {self} after connect() failed.")
+            self._reset_connection_settings()
+            raise
 
         logger.info(f"{self} connected.")
 
@@ -339,6 +364,111 @@ class RealSenseCamera(Camera):
         self.new_frame_event.clear()
         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.
+
+        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.
+        """
+        if self.rs_profile is None:
+            raise RuntimeError(f"{self}: rs_profile must be initialized before use.")
+
+        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]
+
+        available = list(sensors.keys())
+        raise RuntimeError(f"{self}: no color sensor found. 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."""
+        try:
+            sensor.set_option(option, value)
+        except Exception as e:
+            range_info = ""
+            try:
+                option_range = sensor.get_option_range(option)
+                range_info = (
+                    f" (supported range: min={option_range.min}, max={option_range.max}, "
+                    f"step={option_range.step}, default={option_range.default})"
+                )
+            except Exception:
+                range_info = " (option range unavailable)"
+            raise ValueError(
+                f"{self}: failed to set {label} to {value}{range_info}. Original error: {e}"
+            ) from e
+
+    def _configure_sensor_options(self) -> None:
+        """Applies manual sensor options (exposure, gain, white balance) to the color sensor.
+
+        When exposure or gain is set, auto-exposure is disabled first. When white_balance
+        is set, auto white balance is disabled first. An omitted option is left unchanged,
+        and configuration is skipped entirely if all options are omitted.
+
+        Raises:
+            ValueError: If the sensor does not support a requested option or a requested
+                value is invalid. Invalid-value errors include the option name, requested
+                value, and supported range when available.
+        """
+        if self.exposure is None and self.gain is None and self.white_balance is None:
+            return
+
+        color_sensor = self._get_color_sensor()
+
+        requested_options = (
+            (rs.option.exposure, self.exposure, "exposure"),
+            (rs.option.gain, self.gain, "gain"),
+            (rs.option.white_balance, self.white_balance, "white balance"),
+        )
+        unsupported_options = [
+            label
+            for option, value, label in requested_options
+            if value is not None and not color_sensor.supports(option)
+        ]
+        if unsupported_options:
+            raise ValueError(
+                f"{self}: color sensor does not support requested manual options: {unsupported_options}."
+            )
+
+        manual_exposure_requested = self.exposure is not None or self.gain is not None
+        if manual_exposure_requested:
+            if color_sensor.supports(rs.option.enable_auto_exposure):
+                self._set_sensor_option(color_sensor, rs.option.enable_auto_exposure, 0, "auto-exposure")
+                logger.info(f"{self} auto-exposure disabled.")
+            else:
+                logger.warning(
+                    f"{self} sensor does not support disabling auto-exposure; "
+                    "applying manual exposure/gain directly."
+                )
+
+        if self.exposure is not None:
+            self._set_sensor_option(color_sensor, rs.option.exposure, self.exposure, "exposure")
+            logger.info(f"{self} exposure set to {self.exposure}.")
+
+        if self.gain is not None:
+            self._set_sensor_option(color_sensor, rs.option.gain, self.gain, "gain")
+            logger.info(f"{self} gain set to {self.gain}.")
+
+        if self.white_balance is not None:
+            if color_sensor.supports(rs.option.enable_auto_white_balance):
+                self._set_sensor_option(
+                    color_sensor, rs.option.enable_auto_white_balance, 0, "auto white balance"
+                )
+                logger.info(f"{self} auto white balance disabled.")
+            else:
+                logger.warning(
+                    f"{self} sensor does not support disabling auto white balance; "
+                    "applying manual white balance directly."
+                )
+            self._set_sensor_option(
+                color_sensor, rs.option.white_balance, self.white_balance, "white balance"
+            )
+            logger.info(f"{self} white balance set to {self.white_balance}.")
+
     @check_if_not_connected
     def read_depth(self, timeout_ms: int = 200) -> NDArray[Any]:
         """
@@ -541,6 +671,27 @@ class RealSenseCamera(Camera):
             self.latest_timestamp = None
             self.new_frame_event.clear()
 
+    def _cleanup_resources(self) -> None:
+        """Stop background reads and stop the pipeline, including after partial setup."""
+        read_thread = self.thread
+        rs_pipeline = self.rs_pipeline
+
+        try:
+            self._stop_read_thread()
+        finally:
+            self.rs_pipeline = None
+            self.rs_profile = None
+            try:
+                if rs_pipeline is not None:
+                    rs_pipeline.stop()
+            finally:
+                # Stopping the pipeline may unblock a hardware read that outlived
+                # the first bounded join in _stop_read_thread().
+                if read_thread is not None and read_thread.is_alive():
+                    read_thread.join(timeout=2.0)
+                    if read_thread.is_alive():  # pragma: no cover
+                        logger.warning(f"{self} read thread remained alive after stopping the pipeline.")
+
     def _async_read(self, timeout_ms: float, read_depth: bool = False) -> NDArray[Any]:
         """Shared helper for :meth:`async_read`/:meth:`async_read_depth`: return the latest buffered frame."""
         if self.thread is None or not self.thread.is_alive():
@@ -684,18 +835,5 @@ class RealSenseCamera(Camera):
                 f"Attempted to disconnect {self}, but it appears already disconnected."
             )
 
-        if self.thread is not None:
-            self._stop_read_thread()
-
-        if self.rs_pipeline is not None:
-            self.rs_pipeline.stop()
-            self.rs_pipeline = None
-            self.rs_profile = None
-
-        with self.frame_lock:
-            self.latest_color_frame = None
-            self.latest_depth_frame = None
-            self.latest_timestamp = None
-            self.new_frame_event.clear()
-
+        self._cleanup_resources()
         logger.info(f"{self} disconnected.")

src/lerobot/cameras/realsense/configuration_realsense.py

@@ -46,6 +46,17 @@ class RealSenseCameraConfig(CameraConfig):
         use_depth: Whether to enable depth stream. Defaults to False.
         rotation: Image rotation setting (0°, 90°, 180°, or 270°). Defaults to no rotation.
         warmup_s: Time reading frames before returning from connect (in seconds)
+        exposure: Manual exposure value for the color sensor. When set, auto-exposure is
+            disabled and this fixed value is used. Valid ranges are camera-model specific
+            and reported if the value is rejected. Defaults to None (leave unchanged).
+        gain: Manual gain value for the color sensor. When set, auto-exposure is disabled
+            and this fixed gain is used, which also freezes exposure at its current value
+            when no exposure is configured. Valid ranges are camera-model specific and
+            reported if the value is rejected. Defaults to None (leave unchanged).
+        white_balance: Manual white balance value for the color sensor. When set, auto
+            white balance is disabled and this fixed value is used. Valid ranges are
+            camera-model specific and reported if the value is rejected. Defaults to None
+            (leave unchanged).
 
     Note:
         - Either name or serial_number must be specified.
@@ -61,6 +72,9 @@ class RealSenseCameraConfig(CameraConfig):
     use_depth: bool = False
     rotation: Cv2Rotation = Cv2Rotation.NO_ROTATION
     warmup_s: int = 1
+    exposure: int | None = None
+    gain: int | None = None
+    white_balance: int | None = None
 
     def __post_init__(self) -> None:
         self.color_mode = ColorMode(self.color_mode)
@@ -69,6 +83,18 @@ class RealSenseCameraConfig(CameraConfig):
         if not self.use_rgb and not self.use_depth:
             raise ValueError("At least one of `use_rgb` or `use_depth` must be enabled.")
 
+        manual_color_options = {
+            "exposure": self.exposure,
+            "gain": self.gain,
+            "white_balance": self.white_balance,
+        }
+        configured_color_options = [name for name, value in manual_color_options.items() if value is not None]
+        if configured_color_options and not self.use_rgb:
+            raise ValueError(
+                "Manual color sensor options require `use_rgb=True`. "
+                f"Configured options: {configured_color_options}."
+            )
+
         values = (self.fps, self.width, self.height)
         if any(v is not None for v in values) and any(v is None for v in values):
             raise ValueError(

src/lerobot/configs/dataset.py

@@ -71,13 +71,19 @@ class DatasetRecordConfig:
     # Number of threads per encoder instance. None = auto (codec default).
     # Lower values reduce CPU usage, maps to 'lp' (via svtav1-params) for libsvtav1 and 'threads' for h264/hevc..
     encoder_threads: int | None = None
+    # Skip appending the date-time tag to repo_id, keeping the user-provided name as-is
+    # (e.g. self-managed versioned names intended for a later `lerobot-edit-dataset merge`).
+    no_stamp: bool = False
 
     def stamp_repo_id(self) -> None:
         """Append a date-time tag to ``repo_id`` so each recording session gets a unique name.
 
         Must be called explicitly at dataset *creation* time — not on resume,
         where the existing ``repo_id`` (already stamped) must be preserved.
+        No-op when ``no_stamp`` is set, preserving a user-managed ``repo_id``.
         """
+        if self.no_stamp:
+            return
         if self.repo_id:
             timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
             self.repo_id = f"{self.repo_id}_{timestamp}"

src/lerobot/configs/policies.py

@@ -163,8 +163,10 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):  # type: igno
         return None
 
     def _save_pretrained(self, save_directory: Path) -> None:
-        with open(save_directory / CONFIG_NAME, "w") as f, draccus.config_type("json"):
-            draccus.dump(self, f, indent=4)
+        # Encode against the base class so draccus includes the choice "type" key,
+        # which `from_pretrained` needs to resolve the concrete subclass.
+        with open(save_directory / CONFIG_NAME, "w") as f:
+            json.dump(draccus.encode(self, PreTrainedConfig), f, indent=4)
 
     @classmethod
     def from_pretrained(

src/lerobot/configs/rewards.py

@@ -103,8 +103,10 @@ class RewardModelConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):
         pass
 
     def _save_pretrained(self, save_directory: Path) -> None:
-        with open(save_directory / CONFIG_NAME, "w") as f, draccus.config_type("json"):
-            draccus.dump(self, f, indent=4)
+        # Encode against the base class so draccus includes the choice "type" key,
+        # which `from_pretrained` needs to resolve the concrete subclass.
+        with open(save_directory / CONFIG_NAME, "w") as f:
+            json.dump(draccus.encode(self, RewardModelConfig), f, indent=4)
 
     @classmethod
     def from_pretrained(

src/lerobot/configs/train.py

@@ -14,6 +14,7 @@
 import builtins
 import datetime as dt
 import json
+import multiprocessing
 import os
 import tempfile
 from dataclasses import dataclass, field
@@ -101,6 +102,12 @@ 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
@@ -187,7 +194,11 @@ class TrainPipelineConfig(HubMixin):
             )
 
         if Path(config_path).resolve().exists():
-            policy_dir = Path(config_path).parent
+            # `config_path` may point at the checkpoint's train_config.json or at its
+            # pretrained_model/ directory (both documented above) — resolve either to
+            # the pretrained_model/ directory.
+            config_path_obj = Path(config_path)
+            policy_dir = config_path_obj.parent if config_path_obj.is_file() else config_path_obj
             self.checkpoint_path = policy_dir.parent
         elif self.job.is_remote:
             return
@@ -212,6 +223,17 @@ 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/aggregate.py

@@ -19,6 +19,7 @@ import copy
 import logging
 import shutil
 from pathlib import Path
+from typing import Any, NotRequired, TypedDict
 
 import datasets
 import pandas as pd
@@ -49,8 +50,32 @@ from .utils import (
 )
 from .video_utils import concatenate_video_files, get_video_duration_in_s
 
+logger = logging.getLogger(__name__)
 
-def merge_video_feature_info_for_aggregate(all_metadata: list[LeRobotDatasetMetadata]) -> dict[str, dict]:
+type FeatureDict = dict[str, dict[str, Any]]
+type ChunkFile = tuple[int, int]
+
+
+class IndexState(TypedDict):
+    chunk: int
+    file: int
+    src_to_dst: NotRequired[dict[ChunkFile, ChunkFile]]
+
+
+class VideoIndex(TypedDict):
+    chunk: int
+    file: int
+    latest_duration: float
+    episode_duration: float
+    src_to_offset: NotRequired[dict[ChunkFile, float]]
+    src_to_dst: NotRequired[dict[ChunkFile, ChunkFile]]
+    dst_file_durations: NotRequired[dict[ChunkFile, float]]
+
+
+type VideoIndexState = dict[str, VideoIndex]
+
+
+def merge_video_feature_info_for_aggregate(all_metadata: list[LeRobotDatasetMetadata]) -> FeatureDict:
     """Create a merged video feature info dictionary for aggregation. The video encoder info is merged field-by-field: each key is kept only when every source agrees; otherwise that key is set to ``null`` (or ``{}`` for ``video.extra_options``) and a warning is logged.
 
     Args:
@@ -59,14 +84,14 @@ def merge_video_feature_info_for_aggregate(all_metadata: list[LeRobotDatasetMeta
     Returns:
         dict: A dictionary of merged video feature info.
     """
-    merged_info = copy.deepcopy(all_metadata[0].features)
+    merged_info: FeatureDict = copy.deepcopy(all_metadata[0].features)
     video_keys = [k for k in merged_info if merged_info[k].get("dtype") == "video"]
 
     for vk in video_keys:
         video_infos = [m.features.get(vk, {}).get("info") or {} for m in all_metadata]
         base_video_info = video_infos[0]
 
-        merged_encoder_info: dict = {}
+        merged_encoder_info: dict[str, Any] = {}
         fallback_keys: list[str] = []
         for info_key in VIDEO_ENCODER_INFO_KEYS:
             values = [info.get(info_key, None) for info in video_infos]
@@ -80,7 +105,7 @@ def merge_video_feature_info_for_aggregate(all_metadata: list[LeRobotDatasetMeta
                 merged_encoder_info[info_key] = {} if info_key == "video.extra_options" else None
 
         if fallback_keys:
-            logging.warning(
+            logger.warning(
                 f"Merging heterogeneous or incomplete video encoder metadata for feature {vk}. "
                 f"Setting these keys to null: {fallback_keys}.",
             )
@@ -92,7 +117,7 @@ def merge_video_feature_info_for_aggregate(all_metadata: list[LeRobotDatasetMeta
     return merged_info
 
 
-def validate_all_metadata(all_metadata: list[LeRobotDatasetMetadata]):
+def validate_all_metadata(all_metadata: list[LeRobotDatasetMetadata]) -> tuple[int, str | None, FeatureDict]:
     """Validates that all dataset metadata have consistent properties.
 
     Ensures all datasets have the same fps, robot_type, and features to guarantee
@@ -129,7 +154,9 @@ def validate_all_metadata(all_metadata: list[LeRobotDatasetMetadata]):
     return fps, robot_type, features
 
 
-def update_data_df(df, src_meta, dst_meta):
+def update_data_df(
+    df: pd.DataFrame, src_meta: LeRobotDatasetMetadata, dst_meta: LeRobotDatasetMetadata
+) -> pd.DataFrame:
     """Updates a data DataFrame with new indices and task mappings for aggregation.
 
     Adjusts episode indices, frame indices, and task indices to account for
@@ -154,12 +181,12 @@ def update_data_df(df, src_meta, dst_meta):
 
 
 def update_meta_data(
-    df,
-    dst_meta,
-    meta_idx,
-    data_idx,
-    videos_idx,
-):
+    df: pd.DataFrame,
+    dst_meta: LeRobotDatasetMetadata,
+    meta_idx: IndexState,
+    data_idx: IndexState,
+    videos_idx: VideoIndexState,
+) -> pd.DataFrame:
     """Updates metadata DataFrame with new chunk, file, and timestamp indices.
 
     Adjusts all indices and timestamps to account for previously aggregated
@@ -289,7 +316,7 @@ def aggregate_datasets(
     chunk_size: int | None = None,
     concatenate_videos: bool = True,
     concatenate_data: bool = True,
-):
+) -> None:
     """Aggregates multiple LeRobot datasets into a single unified dataset.
 
     This is the main function that orchestrates the aggregation process by:
@@ -309,7 +336,7 @@ def aggregate_datasets(
         concatenate_videos: When False, keep one mp4 per source file instead of packing into shards.
         concatenate_data: When False, keep one parquet per source file instead of packing into shards.
     """
-    logging.info("Start aggregate_datasets")
+    logger.info("Start aggregate_datasets")
 
     if data_files_size_in_mb is None:
         data_files_size_in_mb = DEFAULT_DATA_FILE_SIZE_IN_MB
@@ -341,15 +368,15 @@ def aggregate_datasets(
         video_files_size_in_mb=video_files_size_in_mb,
     )
 
-    logging.info("Find all tasks")
+    logger.info("Find all tasks")
     unique_tasks = pd.concat([m.tasks for m in all_metadata]).index.unique()
     dst_meta.tasks = pd.DataFrame(
         {"task_index": range(len(unique_tasks))}, index=pd.Index(unique_tasks, name="task")
     )
 
-    meta_idx = {"chunk": 0, "file": 0}
-    data_idx = {"chunk": 0, "file": 0}
-    videos_idx = {
+    meta_idx: IndexState = {"chunk": 0, "file": 0}
+    data_idx: IndexState = {"chunk": 0, "file": 0}
+    videos_idx: VideoIndexState = {
         key: {"chunk": 0, "file": 0, "latest_duration": 0, "episode_duration": 0} for key in video_keys
     }
 
@@ -373,12 +400,17 @@ def aggregate_datasets(
         dst_meta.info.total_frames += src_meta.total_frames
 
     finalize_aggregation(dst_meta, all_metadata)
-    logging.info("Aggregation complete.")
+    logger.info("Aggregation complete.")
 
 
 def aggregate_videos(
-    src_meta, dst_meta, videos_idx, video_files_size_in_mb, chunk_size, concatenate_videos=True
-):
+    src_meta: LeRobotDatasetMetadata,
+    dst_meta: LeRobotDatasetMetadata,
+    videos_idx: VideoIndexState,
+    video_files_size_in_mb: float,
+    chunk_size: int,
+    concatenate_videos: bool = True,
+) -> VideoIndexState:
     """Aggregates video chunks from a source dataset into the destination dataset.
 
     Handles video file concatenation and rotation based on file size limits.
@@ -406,7 +438,8 @@ def aggregate_videos(
             videos_idx[key]["dst_file_durations"] = {}
 
     for key, video_idx in videos_idx.items():
-        unique_chunk_file_pairs = {
+        unique_chunk_file_pairs: list[ChunkFile] = sorted(
+            {
                 (chunk, file)
                 for chunk, file in zip(
                     src_meta.episodes[f"videos/{key}/chunk_index"],
@@ -414,7 +447,7 @@ def aggregate_videos(
                     strict=False,
                 )
             }
-        unique_chunk_file_pairs = sorted(unique_chunk_file_pairs)
+        )
 
         chunk_idx = video_idx["chunk"]
         file_idx = video_idx["file"]
@@ -489,7 +522,14 @@ def aggregate_videos(
     return videos_idx
 
 
-def aggregate_data(src_meta, dst_meta, data_idx, data_files_size_in_mb, chunk_size, concatenate_data=True):
+def aggregate_data(
+    src_meta: LeRobotDatasetMetadata,
+    dst_meta: LeRobotDatasetMetadata,
+    data_idx: IndexState,
+    data_files_size_in_mb: float,
+    chunk_size: int,
+    concatenate_data: bool = True,
+) -> IndexState:
     """Aggregates data chunks from a source dataset into the destination dataset.
 
     Reads source data files, updates indices to match the aggregated dataset,
@@ -510,14 +550,16 @@ def aggregate_data(src_meta, dst_meta, data_idx, data_files_size_in_mb, chunk_si
     Returns:
         dict: Updated data_idx with current chunk and file indices.
     """
-    unique_chunk_file_ids = {
+    unique_chunk_file_ids: list[ChunkFile] = sorted(
+        {
             (c, f)
             for c, f in zip(
-            src_meta.episodes["data/chunk_index"], src_meta.episodes["data/file_index"], strict=False
+                src_meta.episodes["data/chunk_index"],
+                src_meta.episodes["data/file_index"],
+                strict=False,
             )
         }
-
-    unique_chunk_file_ids = sorted(unique_chunk_file_ids)
+    )
     contains_images = len(dst_meta.image_keys) > 0
 
     # retrieve features schema for proper image typing in parquet
@@ -525,7 +567,7 @@ def aggregate_data(src_meta, dst_meta, data_idx, data_files_size_in_mb, chunk_si
 
     # Track source to destination file mapping for metadata update
     # This is critical for handling datasets that are already results of a merge
-    src_to_dst: dict[tuple[int, int], tuple[int, int]] = {}
+    src_to_dst: dict[ChunkFile, ChunkFile] = {}
 
     for src_chunk_idx, src_file_idx in unique_chunk_file_ids:
         src_path = src_meta.root / DEFAULT_DATA_PATH.format(
@@ -564,7 +606,13 @@ def aggregate_data(src_meta, dst_meta, data_idx, data_files_size_in_mb, chunk_si
     return data_idx
 
 
-def aggregate_metadata(src_meta, dst_meta, meta_idx, data_idx, videos_idx):
+def aggregate_metadata(
+    src_meta: LeRobotDatasetMetadata,
+    dst_meta: LeRobotDatasetMetadata,
+    meta_idx: IndexState,
+    data_idx: IndexState,
+    videos_idx: VideoIndexState,
+) -> IndexState:
     """Aggregates metadata from a source dataset into the destination dataset.
 
     Reads source metadata files, updates all indices and timestamps,
@@ -580,7 +628,8 @@ def aggregate_metadata(src_meta, dst_meta, meta_idx, data_idx, videos_idx):
     Returns:
         dict: Updated meta_idx with current chunk and file indices.
     """
-    chunk_file_ids = {
+    chunk_file_ids: list[ChunkFile] = sorted(
+        {
             (c, f)
             for c, f in zip(
                 src_meta.episodes["meta/episodes/chunk_index"],
@@ -588,8 +637,7 @@ def aggregate_metadata(src_meta, dst_meta, meta_idx, data_idx, videos_idx):
                 strict=False,
             )
         }
-
-    chunk_file_ids = sorted(chunk_file_ids)
+    )
     for chunk_idx, file_idx in chunk_file_ids:
         src_path = src_meta.root / DEFAULT_EPISODES_PATH.format(chunk_index=chunk_idx, file_index=file_idx)
         df = pd.read_parquet(src_path)
@@ -622,16 +670,16 @@ def aggregate_metadata(src_meta, dst_meta, meta_idx, data_idx, videos_idx):
 def append_or_create_parquet_file(
     df: pd.DataFrame,
     src_path: Path,
-    idx: dict[str, int],
+    idx: IndexState,
     max_mb: float,
     chunk_size: int,
     default_path: str,
     contains_images: bool = False,
-    aggr_root: Path = None,
+    aggr_root: Path | None = None,
     hf_features: datasets.Features | None = None,
     concatenate: bool = True,
     one_row_group_per_episode: bool = False,
-) -> tuple[dict[str, int], tuple[int, int]]:
+) -> tuple[IndexState, ChunkFile]:
     """Appends data to an existing parquet file or creates a new one based on size constraints.
 
     Manages file rotation when size limits are exceeded to prevent individual files
@@ -654,7 +702,13 @@ def append_or_create_parquet_file(
     Returns:
         tuple: (updated_idx, (dst_chunk, dst_file)) where updated_idx is the index dict
                and (dst_chunk, dst_file) is the actual destination file the data was written to.
+
+    Raises:
+        ValueError: If aggr_root is not provided.
     """
+    if aggr_root is None:
+        raise ValueError("aggr_root must be provided.")
+
     dst_chunk, dst_file = idx["chunk"], idx["file"]
     dst_path = aggr_root / default_path.format(chunk_index=dst_chunk, file_index=dst_file)
 
@@ -698,7 +752,9 @@ def append_or_create_parquet_file(
     return idx, (dst_chunk, dst_file)
 
 
-def finalize_aggregation(aggr_meta, all_metadata):
+def finalize_aggregation(
+    aggr_meta: LeRobotDatasetMetadata, all_metadata: list[LeRobotDatasetMetadata]
+) -> None:
     """Finalizes the dataset aggregation by writing summary files and statistics.
 
     Writes the tasks file, info file with total counts and splits, and
@@ -708,16 +764,16 @@ def finalize_aggregation(aggr_meta, all_metadata):
         aggr_meta: Aggregated dataset metadata.
         all_metadata: List of all source dataset metadata objects.
     """
-    logging.info("write tasks")
+    logger.info("write tasks")
     write_tasks(aggr_meta.tasks, aggr_meta.root)
 
-    logging.info("write info")
+    logger.info("write info")
     aggr_meta.info.total_tasks = len(aggr_meta.tasks)
     aggr_meta.info.total_episodes = sum(m.total_episodes for m in all_metadata)
     aggr_meta.info.total_frames = sum(m.total_frames for m in all_metadata)
     aggr_meta.info.splits = {"train": f"0:{sum(m.total_episodes for m in all_metadata)}"}
     write_info(aggr_meta.info, aggr_meta.root)
 
-    logging.info("write stats")
+    logger.info("write stats")
     aggr_meta.stats = aggregate_stats([m.stats for m in all_metadata])
     write_stats(aggr_meta.stats, aggr_meta.root)

src/lerobot/datasets/dataset_metadata.py

@@ -188,8 +188,8 @@ class LeRobotDatasetMetadata:
     def _load_metadata(self):
         self.info = load_info(self.root)
         check_version_compatibility(self.repo_id, self._version, CODEBASE_VERSION)
-        self.tasks = load_tasks(self.root)
-        self.episodes = load_episodes(self.root)
+        self.tasks = load_tasks(self.root) if self.total_tasks > 0 else None
+        self.episodes = load_episodes(self.root) if self.total_episodes > 0 else None
         self.stats = load_stats(self.root)
 
     def ensure_readable(self) -> None:

src/lerobot/datasets/dataset_tools.py

@@ -1435,15 +1435,18 @@ def modify_tasks(
     dataset: LeRobotDataset,
     new_task: str | None = None,
     episode_tasks: dict[int, str] | None = None,
+    task_replacements: dict[str, str] | None = None,
 ) -> LeRobotDataset:
     """Modify tasks in a LeRobotDataset.
 
     This function allows you to either:
     1. Set a single task for the entire dataset (using `new_task`)
     2. Set specific tasks for specific episodes (using `episode_tasks`)
+    3. Replace existing task strings wherever they appear (using `task_replacements`)
 
-    You can combine both: `new_task` sets the default, and `episode_tasks` overrides
-    specific episodes.
+    Per episode, the task is resolved with precedence:
+    `episode_tasks` > `task_replacements` > `new_task` > original task. An episode that ends
+    up with no task (none of the above apply and it had no original task) raises an error.
 
     The dataset is modified in-place, updating only the task-related files:
     - meta/tasks.parquet
@@ -1453,11 +1456,14 @@ def modify_tasks(
 
     Args:
         dataset: The source LeRobotDataset to modify.
-        new_task: A single task string to apply to all episodes. If None and episode_tasks
-            is also None, raises an error.
-        episode_tasks: Optional dict mapping episode indices to their task strings.
-            Overrides `new_task` for specific episodes.
+        new_task: Default task applied to any episode not covered by `episode_tasks` or a
+            matching `task_replacements` entry.
+        episode_tasks: Optional dict mapping episode indices to task strings. Takes precedence
+            over both `task_replacements` and `new_task`.
+        task_replacements: Optional dict mapping existing task strings to new ones. Applied to
+            episodes whose current task matches a key. Every key must be an existing task.
 
+    At least one of `new_task`, `episode_tasks`, or `task_replacements` must be provided.
 
     Examples:
         Set a single task for all episodes:
@@ -1475,11 +1481,17 @@ def modify_tasks(
                 new_task="Default task",
                 episode_tasks={5: "Special task for episode 5"}
             )
-    """
-    if new_task is None and episode_tasks is None:
-        raise ValueError("Must specify at least one of new_task or episode_tasks")
 
-    if episode_tasks is not None:
+        Replace existing task strings in-place:
+            dataset = modify_tasks(
+                dataset,
+                task_replacements={"Pick up the cube": "Lift the cube"}
+            )
+    """
+    if not new_task and not episode_tasks and not task_replacements:
+        raise ValueError("Must specify at least one of new_task, episode_tasks, or task_replacements")
+
+    if episode_tasks:
         valid_indices = set(range(dataset.meta.total_episodes))
         invalid = set(episode_tasks.keys()) - valid_indices
         if invalid:
@@ -1489,19 +1501,29 @@ def modify_tasks(
     if dataset.meta.episodes is None:
         dataset.meta.episodes = load_episodes(dataset.root)
 
+    if task_replacements:
+        current_tasks = set(dataset.meta.tasks.index)
+        invalid_tasks = set(task_replacements) - current_tasks
+        if invalid_tasks:
+            raise ValueError(f"Task replacements reference unknown tasks: {sorted(invalid_tasks)}")
+
     # Build the mapping from episode index to task string
     episode_to_task: dict[int, str] = {}
     for ep_idx in range(dataset.meta.total_episodes):
+        original_tasks = dataset.meta.episodes[ep_idx]["tasks"]
+        original_task = original_tasks[0] if original_tasks else None
+
         if episode_tasks and ep_idx in episode_tasks:
             episode_to_task[ep_idx] = episode_tasks[ep_idx]
-        elif new_task is not None:
+        elif task_replacements and original_task in task_replacements:
+            episode_to_task[ep_idx] = task_replacements[original_task]
+        elif new_task:
             episode_to_task[ep_idx] = new_task
-        else:
+        elif original_task:
             # Keep original task if not overridden and no default provided
-            original_tasks = dataset.meta.episodes[ep_idx]["tasks"]
-            if not original_tasks:
-                raise ValueError(f"Episode {ep_idx} has no tasks and no default task was provided")
-            episode_to_task[ep_idx] = original_tasks[0]
+            episode_to_task[ep_idx] = original_task
+        else:
+            raise ValueError(f"Episode {ep_idx} has no task; provide new_task or episode_tasks")
 
     # Collect all unique tasks and create new task mapping
     unique_tasks = sorted(set(episode_to_task.values()))

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/envs/libero.py

@@ -384,7 +384,12 @@ class LiberoEnv(gym.Env):
 
     def close(self):
         if self._env is not None:
+            try:
                 self._env.close()
+            finally:
+                # LIBERO deletes its inner env on close, so this wrapper must
+                # be recreated before the next reset.
+                self._env = None
 
 
 def _make_env_fns(

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/envs/robotwin.py

@@ -384,7 +384,9 @@ class RoboTwinEnv(gym.Env):
 
         self._env: Any | None = None  # deferred — created on first reset() inside worker
         self._step_count: int = 0
-        self._black_frame = np.zeros((self.observation_height, self.observation_width, 3), dtype=np.uint8)
+        self._black_frame: np.ndarray = np.zeros(
+            (self.observation_height, self.observation_width, 3), dtype=np.uint8
+        )
 
         image_spaces = {
             cam: spaces.Box(

src/lerobot/envs/vlabench.py

@@ -373,7 +373,7 @@ class VLABenchEnv(gym.Env):
 
         if action.shape[0] != 7:
             # Unknown layout — fall back to zero-pad so the sim doesn't crash.
-            padded = np.zeros(ctrl_dim, dtype=np.float64)
+            padded: np.ndarray = np.zeros(ctrl_dim, dtype=np.float64)
             padded[: min(action.shape[0], ctrl_dim)] = action[:ctrl_dim]
             return padded
 

src/lerobot/motors/dynamixel/tables.py

@@ -122,6 +122,9 @@ MODEL_ENCODING_TABLE = {
     "xm430-w350": X_SERIES_ENCODINGS_TABLE,
     "xm540-w270": X_SERIES_ENCODINGS_TABLE,
     "xc430-w150": X_SERIES_ENCODINGS_TABLE,
+    "xh540-w150": X_SERIES_ENCODINGS_TABLE,
+    "xc330-t288": X_SERIES_ENCODINGS_TABLE,
+    "xc330-t181": X_SERIES_ENCODINGS_TABLE,
 }
 
 # {model: model_resolution}
@@ -134,6 +137,9 @@ MODEL_RESOLUTION = {
     "xm430-w350": 4096,
     "xm540-w270": 4096,
     "xc430-w150": 4096,
+    "xh540-w150": 4096,
+    "xc330-t288": 4096,
+    "xc330-t181": 4096,
 }
 
 # {model: model_number}
@@ -145,6 +151,9 @@ MODEL_NUMBER_TABLE = {
     "xm430-w350": 1020,
     "xm540-w270": 1120,
     "xc430-w150": 1070,
+    "xh540-w150": 1110,
+    "xc330-t288": 1220,
+    "xc330-t181": 1210,
 }
 
 # {model: available_operating_modes}
@@ -156,6 +165,9 @@ MODEL_OPERATING_MODES = {
     "xm430-w350": [0, 1, 3, 4, 5, 16],
     "xm540-w270": [0, 1, 3, 4, 5, 16],
     "xc430-w150": [1, 3, 4, 16],
+    "xh540-w150": [0, 1, 3, 4, 5, 16],
+    "xc330-t288": [0, 1, 3, 4, 5, 16],
+    "xc330-t181": [0, 1, 3, 4, 5, 16],
 }
 
 MODEL_CONTROL_TABLE = {
@@ -166,6 +178,9 @@ MODEL_CONTROL_TABLE = {
     "xm430-w350": X_SERIES_CONTROL_TABLE,
     "xm540-w270": X_SERIES_CONTROL_TABLE,
     "xc430-w150": X_SERIES_CONTROL_TABLE,
+    "xh540-w150": X_SERIES_CONTROL_TABLE,
+    "xc330-t288": X_SERIES_CONTROL_TABLE,
+    "xc330-t181": X_SERIES_CONTROL_TABLE,
 }
 
 MODEL_BAUDRATE_TABLE = {
@@ -176,6 +191,9 @@ MODEL_BAUDRATE_TABLE = {
     "xm430-w350": X_SERIES_BAUDRATE_TABLE,
     "xm540-w270": X_SERIES_BAUDRATE_TABLE,
     "xc430-w150": X_SERIES_BAUDRATE_TABLE,
+    "xh540-w150": X_SERIES_BAUDRATE_TABLE,
+    "xc330-t288": X_SERIES_BAUDRATE_TABLE,
+    "xc330-t181": X_SERIES_BAUDRATE_TABLE,
 }
 
 AVAILABLE_BAUDRATES = [

src/lerobot/policies/evo1/modeling_evo1.py

@@ -42,6 +42,9 @@ class Evo1Policy(PreTrainedPolicy):
     config_class = Evo1Config
     name = "evo1"
 
+    def supports_rtc(self) -> bool:
+        return True
+
     def __init__(self, config: Evo1Config, *, vlm_hub_kwargs: dict | None = None, **kwargs):
         super().__init__(config)
         config.validate_features()

src/lerobot/policies/evo1/processor_evo1.py

@@ -302,6 +302,33 @@ def _pad_evo1_stats(
     return padded_stats
 
 
+def _refresh_evo1_normalization_steps(
+    config: Evo1Config,
+    preprocessor: PolicyProcessorPipeline,
+    postprocessor: PolicyProcessorPipeline,
+) -> None:
+    """Re-pad checkpoint-loaded (un)normalizer stats/features to EVO1's fixed widths.
+
+    Loading a checkpoint injects the raw dataset stats (unpadded to max_state_dim/max_action_dim)
+    into the (un)normalizer via the generic override path in make_pre_post_processors. Those stats
+    and their declared features must be re-padded/reshaped to EVO1's fixed widths, otherwise
+    normalization fails against the padded state/action tensors (e.g. state padded to 24 vs. 8-dim
+    LIBERO stats). Padding is a no-op when stats are already at the target width.
+    """
+    normalization_features = _evo1_normalization_features(config)
+    action_features = _evo1_action_features(config)
+    for step in preprocessor.steps:
+        if isinstance(step, NormalizerProcessorStep):
+            step.features = normalization_features
+            step.stats = _pad_evo1_stats(config, step.stats)
+            step.to(device=step.device, dtype=step.dtype)
+    for step in postprocessor.steps:
+        if isinstance(step, UnnormalizerProcessorStep):
+            step.features = action_features
+            step.stats = _pad_evo1_stats(config, step.stats)
+            step.to(device=step.device, dtype=step.dtype)
+
+
 def reconcile_evo1_processors(
     config: Evo1Config,
     preprocessor: PolicyProcessorPipeline,
@@ -309,16 +336,19 @@ def reconcile_evo1_processors(
 ) -> tuple[PolicyProcessorPipeline, PolicyProcessorPipeline]:
     """Reconcile checkpoint-loaded pipelines with the current EVO1 config.
 
-    Two things cannot be restored from a serialized pipeline alone: the EVO1 batch converter
-    (converters are plain functions and are never serialized), and eval-time CLI overrides of the
-    action postprocessing flags (`postprocess_action_dim`, `binarize_gripper`, `gripper_*`). This
-    restores the converter and rebuilds the action step from the current config so those overrides
-    take effect.
+    Three things cannot be restored from a serialized pipeline alone: the EVO1 batch converter
+    (converters are plain functions and are never serialized), eval-time CLI overrides of the
+    action postprocessing flags (`postprocess_action_dim`, `binarize_gripper`, `gripper_*`), and the
+    (un)normalizer stats/features when the generic override path injects raw, unpadded dataset
+    stats. This restores the converter, re-pads the normalization stats to EVO1's fixed widths, and
+    rebuilds the action step from the current config so those overrides take effect.
     """
     # Pipelines reloaded from a checkpoint come back with the default batch converter, which drops
     # non-observation extras (embodiment_id, state_mask, custom task fields) needed by EVO1.
     preprocessor.to_transition = evo1_batch_to_transition
 
+    _refresh_evo1_normalization_steps(config, preprocessor, postprocessor)
+
     action_step = Evo1ActionProcessorStep(
         action_dim=_evo1_action_dim(config),
         binarize_gripper=config.binarize_gripper,

src/lerobot/policies/factory.py

@@ -44,12 +44,19 @@ from lerobot.utils.constants import (
     POLICY_PREPROCESSOR_DEFAULT_NAME,
 )
 from lerobot.utils.feature_utils import dataset_to_policy_features
+from lerobot.utils.import_utils import _peft_available, require_package
 
 from .evo1.configuration_evo1 import Evo1Config
 from .groot.configuration_groot import GrootConfig
 from .pretrained import PreTrainedPolicy
 from .utils import validate_visual_features_consistency
 
+if TYPE_CHECKING or _peft_available:
+    from peft import PeftConfig, PeftModel
+else:
+    PeftConfig = None
+    PeftModel = None
+
 
 def _reconnect_relative_absolute_steps(
     preprocessor: PolicyProcessorPipeline, postprocessor: PolicyProcessorPipeline
@@ -177,6 +184,7 @@ 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"),
@@ -333,12 +341,15 @@ def make_policy(
         # Load a pretrained PEFT model on top of the policy. The pretrained path points to the folder/repo
         # of the adapter and the adapter's config contains the path to the base policy. So we need the
         # adapter config first, then load the correct policy and then apply PEFT.
-        from peft import PeftConfig, PeftModel
+        require_package("peft", extra="peft")
 
         logging.info("Loading policy's PEFT adapter.")
 
         peft_pretrained_path = str(cfg.pretrained_path)
-        peft_config = PeftConfig.from_pretrained(peft_pretrained_path)
+        peft_config = PeftConfig.from_pretrained(
+            peft_pretrained_path,
+            revision=cfg.pretrained_revision,
+        )
 
         kwargs["pretrained_name_or_path"] = peft_config.base_model_name_or_path
         if not kwargs["pretrained_name_or_path"]:
@@ -349,9 +360,14 @@ def make_policy(
                 "the adapter was trained."
             )
 
+        kwargs["revision"] = peft_config.revision
         policy = policy_cls.from_pretrained(**kwargs)
         policy = PeftModel.from_pretrained(
-            policy, peft_pretrained_path, config=peft_config, is_trainable=True
+            policy,
+            peft_pretrained_path,
+            config=peft_config,
+            revision=cfg.pretrained_revision,
+            is_trainable=True,
         )
 
     else:

src/lerobot/policies/gaussian_actor/configuration_gaussian_actor.py

@@ -37,13 +37,19 @@ def is_image_feature(key: str) -> bool:
 @dataclass
 class ConcurrencyConfig:
     """Configuration for the concurrency of the actor and learner.
+
     Possible values are:
     - "threads": Use threads for the actor and learner.
     - "processes": Use processes for the actor and learner.
+
+    ``multiprocessing_context`` selects the process-wide start method when
+    processes are used. Set it to ``None`` to preserve Python's default or a
+    method already selected by the embedding application.
     """
 
     actor: str = "threads"
     learner: str = "threads"
+    multiprocessing_context: str | None = "spawn"
 
 
 @dataclass

src/lerobot/policies/groot/modeling_groot.py

@@ -68,6 +68,9 @@ class GrootPolicy(PreTrainedPolicy):
     name = "groot"
     config_class = GrootConfig
 
+    def supports_rtc(self) -> bool:
+        return True
+
     def __init__(self, config: GrootConfig, **kwargs):
         """Initialize Groot policy wrapper."""
         require_package("transformers", extra="groot")

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/policies/molmoact2/modeling_molmoact2.py

@@ -43,11 +43,22 @@ from torch.distributions import Beta
 
 from lerobot.policies.pretrained import PreTrainedPolicy
 from lerobot.utils.constants import ACTION
-from lerobot.utils.import_utils import _scipy_available, _transformers_available, require_package
+from lerobot.utils.import_utils import (
+    _peft_available,
+    _scipy_available,
+    _transformers_available,
+    require_package,
+)
 
 from ..rtc.modeling_rtc import RTCProcessor
 from .configuration_molmoact2 import MolmoAct2Config
 
+if TYPE_CHECKING or _peft_available:
+    from peft import LoraConfig, get_peft_model
+else:
+    LoraConfig = None
+    get_peft_model = None
+
 logger = logging.getLogger(__name__)
 
 
@@ -509,6 +520,9 @@ class MolmoAct2Policy(PreTrainedPolicy):
     config_class = MolmoAct2Config
     name = "molmoact2"
 
+    def supports_rtc(self) -> bool:
+        return self.config.inference_action_mode == "continuous"
+
     def __init__(
         self,
         config: MolmoAct2Config,
@@ -1731,13 +1745,11 @@ class MolmoAct2Policy(PreTrainedPolicy):
 
     def _build_inner_lora_config(self):
         require_package("peft", extra="molmoact2")
-        from peft import LoraConfig
 
         return LoraConfig(**self._get_inner_peft_targets())
 
     def _apply_lora_adapters(self) -> None:
         require_package("peft", extra="molmoact2")
-        from peft import get_peft_model
 
         peft_config = self._build_inner_lora_config()
         self._validate_peft_config(peft_config)

src/lerobot/policies/pi0/modeling_pi0.py

@@ -749,6 +749,9 @@ class PI0Policy(PreTrainedPolicy):
     config_class = PI0Config
     name = "pi0"
 
+    def supports_rtc(self) -> bool:
+        return True
+
     def __init__(
         self,
         config: PI0Config,

src/lerobot/policies/pi05/modeling_pi05.py

@@ -714,6 +714,9 @@ class PI05Policy(PreTrainedPolicy):
     config_class = PI05Config
     name = "pi05"
 
+    def supports_rtc(self) -> bool:
+        return True
+
     def __init__(
         self,
         config: PI05Config,

src/lerobot/policies/pretrained.py

@@ -34,14 +34,22 @@ from lerobot.configs import PreTrainedConfig
 from lerobot.configs.train import TrainPipelineConfig
 from lerobot.utils.device_utils import resolve_safetensors_device
 from lerobot.utils.hub import HubMixin
+from lerobot.utils.import_utils import _peft_available, require_package
 
 from .utils import log_model_loading_keys
 
-T = TypeVar("T", bound="PreTrainedPolicy")
+if TYPE_CHECKING or _peft_available:
+    from peft import PEFT_TYPE_TO_CONFIG_MAPPING, PeftType, get_peft_model
+else:
+    PEFT_TYPE_TO_CONFIG_MAPPING = None
+    PeftType = None
+    get_peft_model = None
 
 if TYPE_CHECKING:
     from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
 
+T = TypeVar("T", bound="PreTrainedPolicy")
+
 
 def _build_card_context(
     cfg: TrainPipelineConfig | None,
@@ -241,6 +249,10 @@ class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
         """
         raise NotImplementedError
 
+    def supports_rtc(self) -> bool:
+        """Whether this policy implements Real-Time Chunking inference semantics."""
+        return False
+
     # TODO(aliberts, rcadene): split into 'forward' and 'compute_loss'?
     @abc.abstractmethod
     def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict | None]:
@@ -384,7 +396,7 @@ class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
             peft_cli_overrides: Optional dict of CLI overrides (method_type, target_modules, r, etc.)
                 These are merged with policy defaults to build the final config.
         """
-        from peft import get_peft_model
+        require_package("peft", extra="peft")
 
         # If user provided a complete config, use it directly (with overrides)
         if peft_config is not None:
@@ -455,7 +467,7 @@ class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
         Returns:
             Preprocessed dict with renamed keys and init_type mapped to method-specific key.
         """
-        from peft import PeftType
+        require_package("peft", extra="peft")
 
         cli_overrides = cli_overrides.copy()
 
@@ -480,7 +492,7 @@ class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
 
     def _build_peft_config(self, cli_overrides: dict):
         """Build a PEFT config from policy defaults and CLI overrides."""
-        from peft import PEFT_TYPE_TO_CONFIG_MAPPING, PeftType
+        require_package("peft", extra="peft")
 
         # Determine PEFT method type (default to LORA)
         method_type_str = cli_overrides.get("method_type") or "lora"
@@ -507,7 +519,7 @@ class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
 
     def _apply_peft_cli_overrides(self, peft_config, cli_overrides: dict):
         """Apply CLI overrides to an existing PEFT config."""
-        from peft import PEFT_TYPE_TO_CONFIG_MAPPING, PeftType
+        require_package("peft", extra="peft")
 
         # Get method type from existing config or CLI override
         method_type_str = cli_overrides.get("method_type")

src/lerobot/policies/smolvla/modeling_smolvla.py

@@ -145,6 +145,9 @@ class SmolVLAPolicy(PreTrainedPolicy):
     config_class = SmolVLAConfig
     name = "smolvla"
 
+    def supports_rtc(self) -> bool:
+        return True
+
     def __init__(
         self,
         config: SmolVLAConfig,

src/lerobot/policies/smolvla/smolvlm_with_expert.py

@@ -168,14 +168,23 @@ class SmolVLMWithExpertModel(nn.Module):
                 last_layers.append(self.num_vlm_layers - 2)
             frozen_layers = [
                 "lm_head",
-                "text_model.model.norm.weight",
+                "text_model.norm.weight",
             ]
             for layer in last_layers:
-                frozen_layers.append(f"text_model.model.layers.{layer}.")
+                frozen_layers.append(f"text_model.layers.{layer}.")
 
+            unmatched_patterns = set(frozen_layers)
             for name, params in self.vlm.named_parameters():
-                if any(k in name for k in frozen_layers):
+                matched_patterns = [k for k in frozen_layers if k in name]
+                if matched_patterns:
                     params.requires_grad = False
+                    unmatched_patterns.difference_update(matched_patterns)
+            if unmatched_patterns:
+                raise RuntimeError(
+                    "Some frozen layer patterns matched no VLM parameters, so the corresponding layers "
+                    "would silently remain trainable (parameter naming may have changed in transformers): "
+                    f"{sorted(unmatched_patterns)}"
+                )
         # To avoid unused params issue with distributed training
         for name, params in self.lm_expert.named_parameters():
             if "lm_head" in name:

src/lerobot/processor/delta_action_processor.py

@@ -132,10 +132,20 @@ class MapDeltaActionToRobotActionStep(RobotActionProcessorStep):
     def transform_features(
         self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
     ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
-        for axis in ["x", "y", "z", "gripper"]:
+        for axis in ["x", "y", "z"]:
             features[PipelineFeatureType.ACTION].pop(f"delta_{axis}", None)
+        features[PipelineFeatureType.ACTION].pop("gripper", None)
 
-        for feat in ["enabled", "target_x", "target_y", "target_z", "target_wx", "target_wy", "target_wz"]:
+        for feat in [
+            "enabled",
+            "target_x",
+            "target_y",
+            "target_z",
+            "target_wx",
+            "target_wy",
+            "target_wz",
+            "gripper_vel",
+        ]:
             features[PipelineFeatureType.ACTION][f"{feat}"] = PolicyFeature(
                 type=FeatureType.ACTION, shape=(1,)
             )

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/rl/actor.py

@@ -91,7 +91,7 @@ from lerobot.robots import so_follower  # noqa: F401
 from lerobot.teleoperators import gamepad, so_leader  # noqa: F401
 from lerobot.teleoperators.utils import TeleopEvents
 from lerobot.utils.device_utils import get_safe_torch_device
-from lerobot.utils.process import ProcessSignalHandler
+from lerobot.utils.process import ProcessSignalHandler, ensure_multiprocessing_start_method
 from lerobot.utils.random_utils import set_seed
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.transition import (
@@ -124,9 +124,7 @@ def actor_cli(cfg: TrainRLServerPipelineConfig):
     cfg.validate()
     display_pid = False
     if not use_threads(cfg):
-        import torch.multiprocessing as mp
-
-        mp.set_start_method("spawn")
+        ensure_multiprocessing_start_method(cfg.policy.concurrency.multiprocessing_context)
         display_pid = True
 
     # Create logs directory to ensure it exists

src/lerobot/rl/algorithms/configs.py

@@ -16,6 +16,7 @@ from __future__ import annotations
 
 import abc
 import builtins
+import json
 import logging
 import os
 from dataclasses import dataclass, field
@@ -78,8 +79,10 @@ class RLAlgorithmConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):
 
     def _save_pretrained(self, save_directory: Path) -> None:
         """Serialize this config as ``config.json`` inside ``save_directory``."""
-        with open(save_directory / CONFIG_NAME, "w") as f, draccus.config_type("json"):
-            draccus.dump(self, f, indent=4)
+        # Encode against the base class so draccus includes the choice "type" key,
+        # which `from_pretrained` needs to resolve the concrete subclass.
+        with open(save_directory / CONFIG_NAME, "w") as f:
+            json.dump(draccus.encode(self, RLAlgorithmConfig), f, indent=4)
 
     @classmethod
     def from_pretrained(

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/rl/learner.py

@@ -102,7 +102,7 @@ from lerobot.utils.constants import (
 )
 from lerobot.utils.device_utils import get_safe_torch_device
 from lerobot.utils.io_utils import load_json, write_json
-from lerobot.utils.process import ProcessSignalHandler
+from lerobot.utils.process import ProcessSignalHandler, ensure_multiprocessing_start_method
 from lerobot.utils.random_utils import set_seed
 from lerobot.utils.utils import (
     format_big_number,
@@ -123,9 +123,7 @@ def train_cli(cfg: TrainRLServerPipelineConfig):
     # Fail fast with a friendly error if the optional ``hilserl`` extra is missing.
     require_package("grpcio", extra="hilserl", import_name="grpc")
     if not use_threads(cfg):
-        import torch.multiprocessing as mp
-
-        mp.set_start_method("spawn")
+        ensure_multiprocessing_start_method(cfg.policy.concurrency.multiprocessing_context)
 
     # Use the job_name from the config
     train(

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

@@ -46,6 +46,12 @@ class SOFollowerConfig:
     position_i_coefficient: int = 0
     position_d_coefficient: int = 32
 
+    # Number of extra attempts when a `sync_read` of the motors fails. Feetech buses can occasionally
+    # return a corrupted status packet ("Incorrect status packet!"), especially when several joints move
+    # at once, which otherwise aborts the control loop. Retries are immediate (no sleep) and only happen on
+    # failure, so the steady-state read cost is unchanged.
+    num_read_retries: int = 2
+
 
 @RobotConfig.register_subclass("so101_follower")
 @RobotConfig.register_subclass("so100_follower")

src/lerobot/robots/so_follower/robot_kinematic_processor.py

@@ -510,10 +510,10 @@ class ForwardKinematicsJointsToEEAction(RobotActionProcessorStep):
         # We only use the ee pose in the dataset, so we don't need the joint positions
         for n in self.motor_names:
             features[PipelineFeatureType.ACTION].pop(f"{n}.pos", None)
-        # We specify the dataset features of this step that we want to be stored in the dataset
+        # Store end-effector features as actions in the dataset schema
         for k in ["x", "y", "z", "wx", "wy", "wz", "gripper_pos"]:
             features[PipelineFeatureType.ACTION][f"ee.{k}"] = PolicyFeature(
-                type=FeatureType.STATE, shape=(1,)
+                type=FeatureType.ACTION, shape=(1,)
             )
         return features
 

src/lerobot/robots/so_follower/so_follower.py

@@ -180,7 +180,7 @@ class SOFollower(Robot):
     def get_observation(self) -> RobotObservation:
         # Read arm position
         start = time.perf_counter()
-        obs_dict = self.bus.sync_read("Present_Position")
+        obs_dict = self.bus.sync_read("Present_Position", num_retry=self.config.num_read_retries)
         obs_dict = {f"{motor}.pos": val for motor, val in obs_dict.items()}
         dt_ms = (time.perf_counter() - start) * 1e3
         logger.debug(f"{self} read state: {dt_ms:.1f}ms")
@@ -221,7 +221,7 @@ class SOFollower(Robot):
         # Cap goal position when too far away from present position.
         # /!\ Slower fps expected due to reading from the follower.
         if self.config.max_relative_target is not None:
-            present_pos = self.bus.sync_read("Present_Position")
+            present_pos = self.bus.sync_read("Present_Position", num_retry=self.config.num_read_retries)
             goal_present_pos = {key: (g_pos, present_pos[key]) for key, g_pos in goal_pos.items()}
             goal_pos = ensure_safe_goal_position(goal_present_pos, self.config.max_relative_target)
 

src/lerobot/rollout/configs.py

@@ -149,15 +149,6 @@ class EpisodicStrategyConfig(RolloutStrategyConfig):
     # Note that leader -> follower handover is only supported when the leader has `send_feedback` capability.
     smooth_leader_to_follower_handover: bool = True
 
-    # Whether to turn on or off the smooth handover behavior at the start of the
-    # reset phase: the leader is driven to the follower position (actuated
-    # teleops, see `smooth_leader_to_follower_handover`), or the follower is
-    # slid to the teleop pose (non-actuated teleops). Disable for clutch-style
-    # teleoperators (e.g. VR controllers) that re-reference at the current robot
-    # pose on engage: the handover is already continuous there, and the blocking
-    # interpolation only delays the start of the reset phase.
-    smooth_handover: bool = True
-
 
 @RolloutStrategyConfig.register_subclass("dagger")
 @dataclass
@@ -335,8 +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

@@ -24,10 +24,11 @@ from __future__ import annotations
 import logging
 from dataclasses import dataclass, field
 from threading import Event
+from typing import TYPE_CHECKING
 
 import torch
 
-from lerobot.configs import FeatureType
+from lerobot.configs import FeatureType, PreTrainedConfig
 from lerobot.datasets import (
     LeRobotDataset,
     aggregate_pipeline_dataset_features,
@@ -47,6 +48,7 @@ from lerobot.processor.relative_action_processor import RelativeActionsProcessor
 from lerobot.robots import make_robot_from_config
 from lerobot.teleoperators import Teleoperator, make_teleoperator_from_config
 from lerobot.utils.feature_utils import combine_feature_dicts, hw_to_dataset_features
+from lerobot.utils.import_utils import _peft_available, require_package
 
 from .configs import BaseStrategyConfig, DAggerStrategyConfig, RolloutConfig
 from .inference import (
@@ -55,8 +57,15 @@ from .inference import (
     SyncInferenceConfig,
     create_inference_engine,
 )
+from .inference.rtc import supports_rtc_inference
 from .robot_wrapper import ThreadSafeRobot
 
+if TYPE_CHECKING or _peft_available:
+    from peft import PeftConfig, PeftModel
+else:
+    PeftConfig = None
+    PeftModel = None
+
 logger = logging.getLogger(__name__)
 
 
@@ -159,6 +168,35 @@ 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,
+        )
+
+    require_package("peft", extra="peft")
+
+    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 +214,6 @@ 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
@@ -187,19 +224,15 @@ 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:
+        if not supports_rtc_inference(policy):
+            raise ValueError(
+                f"RTC inference is not supported by policy type '{policy_config.type}': "
+                "the policy must implement RTC semantics and predict_action_chunk must accept "
+                "inference_delay and prev_chunk_left_over. Use '--inference.type=sync' instead."
+            )
         policy.config.rtc_config = cfg.inference.rtc
         if hasattr(policy, "init_rtc_processor"):
             policy.init_rtc_processor()
@@ -276,12 +309,22 @@ def build_rollout_context(
     # ``observation_features`` values are either a tuple (camera shape) or the
     # ``float`` type itself used as a sentinel for scalar motor features —
     # see ``dict[str, type | tuple]`` annotation on ``Robot.observation_features``.
+    # Keep cameras (tuple) plus both joint-position (.pos) and base-velocity (.vel)
+    # scalar state features. LeKiwi's observation.state is 9-dim (6 arm .pos +
+    # x/y/theta.vel) and the policy was trained/normalized on all 9; the old .pos-only
+    # filter fed a 6-dim state into a 9-dim normalizer → RuntimeError (size 6 vs 9).
+    # Pure-arm robots have no .vel state keys, so this is a no-op for them.
     observation_features_hw = {
         k: v
         for k, v in all_obs_features.items()
-        if isinstance(v, tuple) or (v is float and k.endswith(".pos"))
+        if isinstance(v, tuple) or (v is float and k.endswith((".pos", ".vel")))
     }
-    action_features_hw = {k: v for k, v in robot.action_features.items() if k.endswith(".pos")}
+    # Keep both joint-position (.pos) and base-velocity (.vel) action features so
+    # mobile manipulators command the base too (e.g. LeKiwi: 6 arm .pos +
+    # x/y/theta.vel = 9-dim action). Pure-arm robots have no .vel keys, so this is
+    # a no-op for them. Without the .vel keys the base velocities are silently
+    # dropped from dataset_features[ACTION]/ordered_action_keys and the base never moves.
+    action_features_hw = {k: v for k, v in robot.action_features.items() if k.endswith((".pos", ".vel"))}
 
     # The action side is always needed: sync inference reads action names from
     # ``dataset_features[ACTION]`` to map policy tensors back to robot actions.
@@ -392,6 +435,7 @@ 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/rollout/inference/rtc.py

@@ -22,6 +22,7 @@ way via ``notify_observation``.
 
 from __future__ import annotations
 
+import inspect
 import logging
 import math
 import time
@@ -62,6 +63,23 @@ _RTC_JOIN_TIMEOUT_S: float = 3.0
 # ---------------------------------------------------------------------------
 
 
+def supports_rtc_inference(policy: PreTrainedPolicy) -> bool:
+    """Whether a policy declares RTC support and accepts the RTC call shape."""
+    supports_rtc = getattr(policy, "supports_rtc", None)
+    if not callable(supports_rtc) or not supports_rtc():
+        return False
+
+    try:
+        inspect.signature(policy.predict_action_chunk).bind(
+            object(),
+            inference_delay=0,
+            prev_chunk_left_over=None,
+        )
+    except (TypeError, ValueError):
+        return False
+    return True
+
+
 def _normalize_prev_actions_length(prev_actions: torch.Tensor, target_steps: int) -> torch.Tensor:
     """Pad or truncate RTC prefix actions to a fixed length for stable compiled inference."""
     if prev_actions.ndim != 2:

src/lerobot/rollout/strategies/episodic.py

@@ -143,10 +143,6 @@ class EpisodicStrategy(RolloutStrategy):
                             # position so the operator takes over without fighting the arm.
                             # For non-actuated teleops: slide the follower to the teleop's current
                             # pose instead, since the leader cannot be driven.
-                            # Disabled entirely with --strategy.smooth_handover=false (useful for
-                            # clutch-style teleops that re-reference at the current robot pose on
-                            # engage).
-                            if self.config.smooth_handover:
                             obs = robot.get_observation()
                             current_pos = {k: v for k, v in obs.items() if k.endswith(".pos")}
                             if (

src/lerobot/scripts/augment_dataset_quantile_stats.py

@@ -36,6 +36,7 @@ python src/lerobot/scripts/augment_dataset_quantile_stats.py \
 import argparse
 import concurrent.futures
 import logging
+import os
 from pathlib import Path
 
 import numpy as np
@@ -52,6 +53,7 @@ from lerobot.datasets import (
     get_feature_stats,
     write_stats,
 )
+from lerobot.datasets.compute_stats import sample_indices
 from lerobot.utils.utils import init_logging
 
 
@@ -77,12 +79,14 @@ def has_quantile_stats(stats: dict[str, dict] | None, quantile_list_keys: list[s
     return False
 
 
-def process_single_episode(dataset: LeRobotDataset, episode_idx: int) -> dict:
+def process_single_episode(dataset: LeRobotDataset, episode_idx: int, use_sampling: bool = True) -> dict:
     """Process a single episode and return its statistics.
 
     Args:
         dataset: The LeRobot dataset
         episode_idx: Index of the episode to process
+        use_sampling: If True, sub-sample image/video frames per episode to bound
+            memory. If False, use every frame (exact, higher memory).
 
     Returns:
         Dictionary containing episode statistics
@@ -92,16 +96,31 @@ def process_single_episode(dataset: LeRobotDataset, episode_idx: int) -> dict:
     start_idx = dataset.meta.episodes[episode_idx]["dataset_from_index"]
     end_idx = dataset.meta.episodes[episode_idx]["dataset_to_index"]
 
-    collected_data: dict[str, list] = {}
-    for idx in range(start_idx, end_idx):
-        item = dataset[idx]
-        for key, value in item.items():
-            if key not in dataset.features:
-                continue
+    episode_len = end_idx - start_idx
 
-            if key not in collected_data:
-                collected_data[key] = []
-            collected_data[key].append(value)
+    # Images/video are the memory hog, so sub-sample those frames per episode;
+    # numeric columns are cheap, so read them in full (exact).
+    image_keys = [k for k in dataset.features if dataset.features[k]["dtype"] in ("image", "video")]
+    numeric_keys = [
+        k for k in dataset.features if dataset.features[k]["dtype"] not in ("image", "video", "string")
+    ]
+
+    collected_data: dict[str, list] = {}
+
+    # Numeric features: every frame, read directly from the underlying table.
+    if numeric_keys:
+        numeric_cols = dataset.hf_dataset.select_columns(numeric_keys)[start_idx:end_idx]
+        for key in numeric_keys:
+            collected_data[key] = [torch.as_tensor(v) for v in numeric_cols[key]]
+
+    # Image/video features: decode only a sampled subset of frames.
+    if image_keys:
+        sampled_offsets = sample_indices(episode_len) if use_sampling else list(range(episode_len))
+        for offset in sampled_offsets:
+            item = dataset[start_idx + offset]
+            for key in image_keys:
+                if key in item:
+                    collected_data.setdefault(key, []).append(item[key])
 
     ep_stats = {}
     for key, data_list in collected_data.items():
@@ -131,11 +150,13 @@ def process_single_episode(dataset: LeRobotDataset, episode_idx: int) -> dict:
     return ep_stats
 
 
-def compute_quantile_stats_for_dataset(dataset: LeRobotDataset) -> dict[str, dict]:
+def compute_quantile_stats_for_dataset(dataset: LeRobotDataset, use_sampling: bool = True) -> dict[str, dict]:
     """Compute quantile statistics for all episodes in the dataset.
 
     Args:
         dataset: The LeRobot dataset to compute statistics for
+        use_sampling: If True, sub-sample image/video frames per episode to bound
+            memory. If False, use every frame (exact, higher memory).
 
     Returns:
         Dictionary containing aggregated statistics with quantiles
@@ -153,15 +174,15 @@ def compute_quantile_stats_for_dataset(dataset: LeRobotDataset) -> dict[str, dic
     if has_videos:
         logging.info("Dataset contains video keys - using sequential processing for thread safety")
         for episode_idx in tqdm(range(dataset.num_episodes), desc="Processing episodes"):
-            ep_stats = process_single_episode(dataset, episode_idx)
+            ep_stats = process_single_episode(dataset, episode_idx, use_sampling)
             episode_stats_list.append(ep_stats)
     else:
         logging.info("Dataset has no video keys - using parallel processing for better performance")
-        max_workers = min(dataset.num_episodes, 16)
+        max_workers = min(dataset.num_episodes, int(os.environ.get("LEROBOT_STATS_MAX_WORKERS", 16)))
 
         with concurrent.futures.ThreadPoolExecutor(max_workers=max_workers) as executor:
             future_to_episode = {
-                executor.submit(process_single_episode, dataset, episode_idx): episode_idx
+                executor.submit(process_single_episode, dataset, episode_idx, use_sampling): episode_idx
                 for episode_idx in range(dataset.num_episodes)
             }
 
@@ -188,6 +209,7 @@ def augment_dataset_with_quantile_stats(
     repo_id: str,
     root: str | Path | None = None,
     overwrite: bool = False,
+    use_sampling: bool = True,
 ) -> None:
     """Augment a dataset with quantile statistics if they are missing.
 
@@ -195,6 +217,8 @@ def augment_dataset_with_quantile_stats(
         repo_id: Repository ID of the dataset
         root: Local root directory for the dataset
         overwrite: Overwrite existing quantile statistics if they already exist
+        use_sampling: If True, sub-sample image/video frames per episode to bound
+            memory. If False, use every frame (exact, higher memory).
     """
     logging.info(f"Loading dataset: {repo_id}")
     dataset = LeRobotDataset(
@@ -208,7 +232,7 @@ def augment_dataset_with_quantile_stats(
 
     logging.info("Dataset does not contain quantile statistics. Computing them now...")
 
-    new_stats = compute_quantile_stats_for_dataset(dataset)
+    new_stats = compute_quantile_stats_for_dataset(dataset, use_sampling=use_sampling)
 
     logging.info("Updating dataset metadata with new quantile statistics")
     dataset.meta.stats = new_stats
@@ -248,6 +272,14 @@ def main():
         action="store_true",
         help="Overwrite existing quantile statistics if they already exist",
     )
+    parser.add_argument(
+        "--no-sampling",
+        action="store_true",
+        help=(
+            "Compute stats over every frame (exact, higher memory). By default, "
+            "image/video frames are sub-sampled per episode to bound memory."
+        ),
+    )
 
     args = parser.parse_args()
     root = Path(args.root) if args.root else None
@@ -258,6 +290,7 @@ def main():
         repo_id=args.repo_id,
         root=root,
         overwrite=args.overwrite,
+        use_sampling=not args.no_sampling,
     )
 
 

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
@@ -93,6 +94,8 @@ from lerobot.datasets.video_utils import concatenate_video_files, get_video_dura
 from lerobot.utils.constants import HF_LEROBOT_HOME
 from lerobot.utils.utils import flatten_dict, init_logging
 
+logger = logging.getLogger(__name__)
+
 V21 = "v2.1"
 V30 = "v3.0"
 
@@ -475,11 +478,11 @@ def convert_dataset(
     # First check if the dataset already has a v3.0 version
     if root is None and not force_conversion:
         try:
-            print("Trying to download v3.0 version of the dataset from the hub...")
+            logger.info("Trying to download v3.0 version of the dataset from the hub...")
             snapshot_download(repo_id, repo_type="dataset", revision=V30, local_dir=HF_LEROBOT_HOME / repo_id)
             return
         except Exception:
-            print("Dataset does not have an uploaded v3.0 version. Continuing with conversion.")
+            logger.info("Dataset does not have an uploaded v3.0 version. Continuing with conversion.")
 
     # Set root based on whether local dataset path is provided
     use_local_dataset = False
@@ -487,7 +490,7 @@ def convert_dataset(
     if root.exists():
         validate_local_dataset_version(root)
         use_local_dataset = True
-        print(f"Using local dataset at {root}")
+        logger.info(f"Using local dataset at {root}")
 
     old_root = root.parent / f"{root.name}_old"
     new_root = root.parent / f"{root.name}_v30"
@@ -521,8 +524,8 @@ def convert_dataset(
         hub_api = HfApi()
         try:
             hub_api.delete_tag(repo_id, tag=CODEBASE_VERSION, repo_type="dataset")
-        except HTTPError as e:
-            print(f"tag={CODEBASE_VERSION} probably doesn't exist. Skipping exception ({e})")
+        except (HTTPError, RevisionNotFoundError) as e:
+            logger.warning(f"tag={CODEBASE_VERSION} probably doesn't exist. Skipping exception ({e})")
             pass
         hub_api.delete_files(
             delete_patterns=["data/chunk*/episode_*", "meta/*.jsonl", "videos/chunk*"],

src/lerobot/scripts/lerobot_annotate.py

@@ -154,14 +154,14 @@ def _push_to_hub(root: Path, cfg: AnnotationPipelineConfig) -> None:
     repo_id = cfg.new_repo_id or cfg.repo_id
     commit_message = cfg.push_commit_message or "Add steerable annotations (lerobot-annotate)"
     api = HfApi()
-    print(f"[lerobot-annotate] creating/locating dataset repo {repo_id}...", flush=True)
+    logger.info(f"[lerobot-annotate] creating/locating dataset repo {repo_id}...")
     api.create_repo(
         repo_id=repo_id,
         repo_type="dataset",
         private=cfg.push_private,
         exist_ok=True,
     )
-    print(f"[lerobot-annotate] uploading {root} -> {repo_id}...", flush=True)
+    logger.info(f"[lerobot-annotate] uploading {root} -> {repo_id}...")
     commit_info = api.upload_folder(
         folder_path=str(root),
         repo_id=repo_id,
@@ -172,7 +172,7 @@ def _push_to_hub(root: Path, cfg: AnnotationPipelineConfig) -> None:
         # at the source dataset; a fresh card is generated below instead.
         ignore_patterns=[".annotate_staging/**", "**/.DS_Store", "README.md"],
     )
-    print(f"[lerobot-annotate] uploaded to https://huggingface.co/datasets/{repo_id}", flush=True)
+    logger.info(f"[lerobot-annotate] uploaded to https://huggingface.co/datasets/{repo_id}")
 
     dataset_info = load_info(root)
     card = create_lerobot_dataset_card(dataset_info=dataset_info, license="apache-2.0", repo_id=repo_id)
@@ -200,14 +200,13 @@ def _push_to_hub(root: Path, cfg: AnnotationPipelineConfig) -> None:
         with suppress(RevisionNotFoundError):
             api.delete_tag(repo_id, tag=version_tag, repo_type="dataset")
         api.create_tag(**tag_kwargs)
-        print(f"[lerobot-annotate] tagged {repo_id} as {version_tag}", flush=True)
+        logger.info(f"[lerobot-annotate] tagged {repo_id} as {version_tag}")
     except Exception as exc:  # noqa: BLE001
-        print(
+        logger.warning(
             f"[lerobot-annotate] WARNING: could not create tag {version_tag!r} on {repo_id}: {exc}. "
             "Dataset is uploaded but ``LeRobotDataset`` won't be able to load it until it's tagged. "
             "Run: from huggingface_hub import HfApi; "
-            f"HfApi().create_tag({repo_id!r}, tag={version_tag!r}, repo_type='dataset', exist_ok=True)",
-            flush=True,
+            f"HfApi().create_tag({repo_id!r}, tag={version_tag!r}, repo_type='dataset', exist_ok=True)"
         )
 
 

src/lerobot/scripts/lerobot_dataset_viz.py

@@ -89,6 +89,8 @@ from lerobot.datasets import LeRobotDataset
 from lerobot.utils.constants import ACTION, DONE, OBS_STATE, REWARD, SUCCESS
 from lerobot.utils.utils import init_logging
 
+logger = logging.getLogger(__name__)
+
 DEFAULT_FOXGLOVE_PORT = 8765
 DEFAULT_RERUN_PORT = 9090
 
@@ -299,7 +301,7 @@ def visualize_dataset(
             while True:
                 time.sleep(1)
         except KeyboardInterrupt:
-            print("Ctrl-C received. Exiting.")
+            logger.info("Ctrl-C received. Exiting.")
 
 
 def main():

src/lerobot/scripts/lerobot_edit_dataset.py

@@ -127,6 +127,12 @@ Modify tasks - set default task with overrides for specific episodes (WARNING: m
         --operation.new_task "Default task" \
         --operation.episode_tasks '{"5": "Special task for episode 5"}'
 
+Modify tasks - replace existing task strings in-place (WARNING: modifies in-place):
+    lerobot-edit-dataset \
+        --repo_id lerobot/pusht \
+        --operation.type modify_tasks \
+        --operation.task_replacements '{"Pick up the red cube": "Lift the red cube"}'
+
 Convert image dataset to video format and save locally:
     lerobot-edit-dataset \
         --repo_id lerobot/pusht_image \
@@ -303,6 +309,7 @@ class RemoveFeatureConfig(OperationConfig):
 class ModifyTasksConfig(OperationConfig):
     new_task: str | None = None
     episode_tasks: dict[str, str] | None = None
+    task_replacements: dict[str, str] | None = None
 
 
 @OperationConfig.register_subclass("convert_image_to_video")
@@ -551,9 +558,12 @@ def handle_modify_tasks(cfg: EditDatasetConfig) -> None:
 
     new_task = cfg.operation.new_task
     episode_tasks_raw = cfg.operation.episode_tasks
+    task_replacements = cfg.operation.task_replacements
 
-    if new_task is None and episode_tasks_raw is None:
-        raise ValueError("Must specify at least one of new_task or episode_tasks for modify_tasks operation")
+    if new_task is None and episode_tasks_raw is None and task_replacements is None:
+        raise ValueError(
+            "Must specify at least one of new_task, episode_tasks, or task_replacements for modify_tasks operation"
+        )
 
     if cfg.new_repo_id is not None or cfg.new_root is not None:
         logging.warning(
@@ -573,11 +583,14 @@ def handle_modify_tasks(cfg: EditDatasetConfig) -> None:
         logging.info(f"  Default task: '{new_task}'")
     if episode_tasks:
         logging.info(f"  Episode-specific tasks: {episode_tasks}")
+    if task_replacements:
+        logging.info(f"  Task replacements: {task_replacements}")
 
     modified_dataset = modify_tasks(
         dataset,
         new_task=new_task,
         episode_tasks=episode_tasks,
+        task_replacements=task_replacements,
     )
 
     logging.info(f"Dataset modified at {dataset.root}")

src/lerobot/scripts/lerobot_eval.py

@@ -62,7 +62,7 @@ from dataclasses import asdict
 from functools import partial
 from pathlib import Path
 from pprint import pformat
-from typing import Any, TypedDict
+from typing import TYPE_CHECKING, Any, TypedDict
 
 import einops
 import gymnasium as gym
@@ -87,7 +87,7 @@ from lerobot.processor import PolicyProcessorPipeline
 from lerobot.types import PolicyAction
 from lerobot.utils.constants import ACTION, DONE, OBS_IMAGE, OBS_IMAGES, OBS_STR, REWARD
 from lerobot.utils.device_utils import get_safe_torch_device
-from lerobot.utils.import_utils import register_third_party_plugins
+from lerobot.utils.import_utils import _peft_available, register_third_party_plugins, require_package
 from lerobot.utils.io_utils import write_video
 from lerobot.utils.random_utils import set_seed
 from lerobot.utils.utils import (
@@ -95,6 +95,14 @@ from lerobot.utils.utils import (
     inside_slurm,
 )
 
+if TYPE_CHECKING or _peft_available:
+    from peft import PeftModel
+else:
+    PeftModel = None
+
+
+logger = logging.getLogger(__name__)
+
 
 def _env_features_to_dataset_features(env_features: dict) -> dict:
     """Convert EnvConfig.features to the dict format expected by LeRobotDataset.create()."""
@@ -444,15 +452,16 @@ def eval_policy(
         exc = ValueError(
             f"Policy of type 'PreTrainedPolicy' is expected, but type '{type(policy)}' was provided."
         )
-        try:
-            from peft import PeftModel
-
+        if not _peft_available:
+            raise exc
+        require_package("peft", extra="peft")
         if not isinstance(policy, PeftModel):
             raise exc
-        except ImportError:
-            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
@@ -555,7 +564,7 @@ def eval_policy(
         if seeds:
             all_seeds.extend(seeds)
         else:
-            all_seeds.append(None)
+            all_seeds.extend([None] * env.num_envs)
 
         # FIXME: episode_data is either None or it doesn't exist
         if return_episode_data:
@@ -674,6 +683,8 @@ def eval_policy(
     if save_predicted_video:
         info["predicted_video_paths"] = predicted_video_paths
 
+    policy.train(was_training)
+
     return info
 
 
@@ -791,13 +802,13 @@ def eval_main(cfg: EvalPipelineConfig):
             recording_repo_id=cfg.eval.recording_repo_id,
             recording_private=cfg.eval.recording_private,
         )
-        print("Overall Aggregated Metrics:")
-        print(info["overall"])
+        logger.info("Overall Aggregated Metrics:")
+        logger.info(info["overall"])
 
         # Print per-suite stats
         for task_group, task_group_info in info.items():
-            print(f"\nAggregated Metrics for {task_group}:")
-            print(task_group_info)
+            logger.info(f"\nAggregated Metrics for {task_group}:")
+            logger.info(task_group_info)
     # Close all vec envs
     close_envs(envs)
 
@@ -1010,6 +1021,12 @@ def eval_policy_all(
         recording_private=recording_private,
     )
 
+    # 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):
@@ -1044,6 +1061,8 @@ def eval_policy_all(
                         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_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
@@ -40,6 +39,7 @@ from PIL import Image
 from lerobot.cameras import ColorMode
 from lerobot.cameras.opencv import OpenCVCamera, OpenCVCameraConfig
 from lerobot.cameras.realsense import RealSenseCamera, RealSenseCameraConfig
+from lerobot.utils.utils import init_logging
 
 logger = logging.getLogger(__name__)
 
@@ -132,7 +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.
     """
@@ -151,7 +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")
@@ -164,12 +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:
@@ -187,9 +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"]
@@ -199,7 +199,7 @@ def process_camera_image(
     try:
         image_data = cam.read()
 
-        return save_image(
+        save_image(
             image_data,
             cam_id_str,
             output_dir,
@@ -214,10 +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()
@@ -229,6 +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.
@@ -239,6 +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}")
@@ -248,47 +249,32 @@ 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}")
 
 
 def main():
+    init_logging()
+
     parser = argparse.ArgumentParser(
         description="Unified camera utility script for listing cameras and capturing images."
     )
-
     parser.add_argument(
         "camera_type",
         type=str,
@@ -306,8 +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/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_rollout.py

@@ -165,6 +165,7 @@ from lerobot.robots import (  # noqa: F401
     earthrover_mini_plus,
     hope_jr,
     koch_follower,
+    lekiwi,
     omx_follower,
     openarm_follower,
     reachy2,

src/lerobot/scripts/lerobot_train.py

@@ -22,7 +22,8 @@ import dataclasses
 import logging
 import sys
 import time
-from contextlib import nullcontext
+from collections.abc import Iterator
+from contextlib import contextmanager, nullcontext
 from pprint import pformat
 from typing import TYPE_CHECKING, Any
 
@@ -57,7 +58,7 @@ from lerobot.optim.factory import make_optimizer_and_scheduler
 from lerobot.policies import PreTrainedPolicy, make_policy, make_pre_post_processors
 from lerobot.rewards import make_reward_pre_post_processors
 from lerobot.utils.collate import lerobot_collate_fn
-from lerobot.utils.import_utils import register_third_party_plugins
+from lerobot.utils.import_utils import _peft_available, register_third_party_plugins, require_package
 from lerobot.utils.logging_utils import AverageMeter, MetricsTracker
 from lerobot.utils.random_utils import set_seed
 from lerobot.utils.utils import (
@@ -68,9 +69,38 @@ from lerobot.utils.utils import (
     inside_slurm,
 )
 
+if TYPE_CHECKING or _peft_available:
+    from peft import PeftModel
+else:
+    PeftModel = None
+
 from .lerobot_eval import eval_policy_all
 
 
+@contextmanager
+def _make_eval_envs(cfg: TrainPipelineConfig) -> Iterator[dict[str, dict[int, Any]]]:
+    """Create evaluation environments for one run and always dispose of them."""
+    envs = make_env(
+        cfg.env,
+        n_envs=cfg.eval.batch_size,
+        use_async_envs=cfg.eval.use_async_envs,
+    )
+    try:
+        yield envs
+    finally:
+        close_envs(envs)
+
+
+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,
@@ -197,8 +227,6 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
     if cfg.job.is_remote:
         return submit_to_hf(cfg)
 
-    from lerobot.utils.import_utils import require_package
-
     require_package("accelerate", extra="training")
     from accelerate import Accelerator
     from accelerate.utils import DistributedDataParallelKwargs, DistributedType
@@ -267,14 +295,6 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
     if not is_main_process:
         dataset, eval_dataset = make_train_eval_datasets(cfg)
 
-    # Create environment used for evaluating checkpoints during training on simulation data.
-    # On real-world data, no need to create an environment as evaluations are done outside train.py,
-    # using the eval.py instead, with gym_dora environment and dora-rs.
-    eval_env = None
-    if cfg.env_eval_freq > 0 and cfg.env is not None and is_main_process:
-        logging.info("Creating env")
-        eval_env = make_env(cfg.env, n_envs=cfg.eval.batch_size, use_async_envs=cfg.eval.use_async_envs)
-
     if cfg.is_reward_model_training:
         if is_main_process:
             logging.info("Creating reward model")
@@ -302,7 +322,7 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
     if cfg.peft is not None:
         if cfg.is_reward_model_training:
             raise ValueError("PEFT is only supported for policy training. ")
-        from peft import PeftModel
+        require_package("peft", extra="peft")
 
         if isinstance(policy, PeftModel):
             logging.info("PEFT adapter already loaded from checkpoint, skipping wrap_with_peft.")
@@ -328,7 +348,6 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
         preprocessor_overrides = {
             "device_processor": {"device": device.type},
             "normalizer_processor": {
-                "stats": dataset.meta.stats,
                 "features": {**policy.config.input_features, **policy.config.output_features},
                 "norm_map": policy.config.normalization_mapping,
             },
@@ -336,11 +355,17 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
         }
         postprocessor_overrides = {
             "unnormalizer_processor": {
-                "stats": dataset.meta.stats,
                 "features": policy.config.output_features,
                 "norm_map": policy.config.normalization_mapping,
             },
         }
+        # On resume, the checkpoint's saved processor stats are authoritative: they may have
+        # been adapted by the policy (e.g. EVO1 pads state/action stats to max_state_dim),
+        # and force-feeding raw dataset stats over them crashes normalization (#4006).
+        # This mirrors the `dataset_stats` kwarg above, which is also skipped on resume.
+        if not cfg.resume:
+            preprocessor_overrides["normalizer_processor"]["stats"] = dataset.meta.stats
+            postprocessor_overrides["unnormalizer_processor"]["stats"] = dataset.meta.stats
         if getattr(active_cfg, "use_relative_actions", False):
             preprocessor_overrides["relative_actions_processor"] = {
                 "enabled": True,
@@ -473,8 +498,7 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
         pin_memory=device.type == "cuda",
         drop_last=False,
         collate_fn=collate_fn,
-        prefetch_factor=cfg.prefetch_factor if cfg.num_workers > 0 else None,
-        persistent_workers=cfg.persistent_workers and cfg.num_workers > 0,
+        **_dataloader_worker_kwargs(cfg),
     )
 
     # Build eval dataloader if a held-out split exists
@@ -500,8 +524,7 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
             pin_memory=device.type == "cuda",
             drop_last=False,
             collate_fn=eval_collate_fn,
-            prefetch_factor=cfg.prefetch_factor if cfg.num_workers > 0 else None,
-            persistent_workers=cfg.persistent_workers and cfg.num_workers > 0,
+            **_dataloader_worker_kwargs(cfg),
         )
 
     # Prepare everything with accelerator
@@ -684,7 +707,7 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
             if is_main_process:
                 step_id = get_step_identifier(step, cfg.steps)
                 logging.info(f"Eval policy at step {step}")
-                with torch.no_grad(), accelerator.autocast():
+                with _make_eval_envs(cfg) as eval_env, torch.no_grad(), accelerator.autocast():
                     eval_info = eval_policy_all(
                         envs=eval_env,  # dict[suite][task_id] -> vec_env
                         policy=accelerator.unwrap_model(policy),
@@ -732,9 +755,6 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
     if is_main_process:
         progbar.close()
 
-    if eval_env:
-        close_envs(eval_env)
-
     is_fsdp = accelerator.distributed_type == DistributedType.FSDP
     model_state_dict = accelerator.get_state_dict(policy) if is_fsdp else None
     if is_main_process:

src/lerobot/scripts/lerobot_train_tokenizer.py

@@ -45,6 +45,7 @@ lerobot-train-tokenizer \
 """
 
 import json
+import logging
 from dataclasses import dataclass
 from pathlib import Path
 from typing import TYPE_CHECKING
@@ -63,6 +64,9 @@ else:
 from lerobot.configs import NormalizationMode, parser
 from lerobot.datasets import LeRobotDataset
 from lerobot.utils.constants import ACTION, OBS_STATE
+from lerobot.utils.utils import init_logging
+
+logger = logging.getLogger(__name__)
 
 
 @dataclass
@@ -274,11 +278,8 @@ def process_episode(args):
 
         return action_chunks
 
-    except Exception as e:
-        print(f"Error processing episode {ep_idx}: {e}")
-        import traceback
-
-        traceback.print_exc()
+    except Exception:
+        logger.exception("Error processing episode %s", ep_idx)
         return None
 
 
@@ -300,10 +301,10 @@ def train_fast_tokenizer(
     Returns:
         Trained FAST tokenizer
     """
-    print(f"Training FAST tokenizer on {len(action_chunks)} action chunks...")
-    print(f"Action chunk shape: {action_chunks.shape}")
-    print(f"Vocab size: {vocab_size}")
-    print(f"DCT scale: {scale}")
+    logger.info(f"Training FAST tokenizer on {len(action_chunks)} action chunks...")
+    logger.info(f"Action chunk shape: {action_chunks.shape}")
+    logger.info(f"Vocab size: {vocab_size}")
+    logger.info(f"DCT scale: {scale}")
 
     # download the tokenizer source code (not pretrained weights)
     # we'll train a new tokenizer on our own data
@@ -314,7 +315,7 @@ def train_fast_tokenizer(
 
     # train the new tokenizer on our action data using .fit()
     # this trains the BPE tokenizer on DCT coefficients
-    print("Training new tokenizer (this may take a few minutes)...")
+    logger.info("Training new tokenizer (this may take a few minutes)...")
     tokenizer = base_tokenizer.fit(
         action_data_list,
         scale=scale,
@@ -322,21 +323,21 @@ def train_fast_tokenizer(
         time_horizon=action_chunks.shape[1],  # action_horizon
         action_dim=action_chunks.shape[2],  # encoded dimensions
     )
-    print("✓ Tokenizer training complete!")
+    logger.info("✓ Tokenizer training complete!")
 
     # validate it works
     sample_chunk = action_chunks[0]
     encoded = tokenizer(sample_chunk[None])[0]
     if isinstance(encoded, list):
         encoded = np.array(encoded)
-    print(f"Sample encoding: {len(encoded)} tokens for chunk shape {sample_chunk.shape}")
+    logger.info(f"Sample encoding: {len(encoded)} tokens for chunk shape {sample_chunk.shape}")
 
     return tokenizer
 
 
 def compute_compression_stats(tokenizer, action_chunks: np.ndarray):
     """Compute compression statistics."""
-    print("\nComputing compression statistics...")
+    logger.info("\nComputing compression statistics...")
 
     # sample for stats (use max 1000 chunks for speed)
     sample_size = min(1000, len(action_chunks))
@@ -366,12 +367,12 @@ def compute_compression_stats(tokenizer, action_chunks: np.ndarray):
         "max_token_length": float(np.max(token_lengths)),
     }
 
-    print("Compression Statistics:")
-    print(f"  Average compression ratio: {stats['compression_ratio']:.2f}x")
-    print(f"  Mean token length: {stats['mean_token_length']:.1f}")
-    print(f"  P99 token length: {stats['p99_token_length']:.0f}")
-    print(f"  Min token length: {stats['min_token_length']:.0f}")
-    print(f"  Max token length: {stats['max_token_length']:.0f}")
+    logger.info("Compression Statistics:")
+    logger.info(f"  Average compression ratio: {stats['compression_ratio']:.2f}x")
+    logger.info(f"  Mean token length: {stats['mean_token_length']:.1f}")
+    logger.info(f"  P99 token length: {stats['p99_token_length']:.0f}")
+    logger.info(f"  Min token length: {stats['min_token_length']:.0f}")
+    logger.info(f"  Max token length: {stats['max_token_length']:.0f}")
 
     return stats
 
@@ -385,9 +386,9 @@ def train_tokenizer(cfg: TokenizerTrainingConfig):
         cfg: TokenizerTrainingConfig dataclass with all configuration parameters
     """
     # load dataset
-    print(f"Loading dataset: {cfg.repo_id}")
+    logger.info(f"Loading dataset: {cfg.repo_id}")
     dataset = LeRobotDataset(repo_id=cfg.repo_id, root=cfg.root)
-    print(f"Dataset loaded: {dataset.num_episodes} episodes, {dataset.num_frames} frames")
+    logger.info(f"Dataset loaded: {dataset.num_episodes} episodes, {dataset.num_frames} frames")
 
     # parse normalization mode
     try:
@@ -397,7 +398,7 @@ def train_tokenizer(cfg: TokenizerTrainingConfig):
             f"Invalid normalization_mode: {cfg.normalization_mode}. "
             f"Must be one of: {', '.join([m.value for m in NormalizationMode])}"
         ) from err
-    print(f"Normalization mode: {norm_mode.value}")
+    logger.info(f"Normalization mode: {norm_mode.value}")
 
     # parse encoded dimensions
     encoded_dim_ranges = []
@@ -406,38 +407,38 @@ def train_tokenizer(cfg: TokenizerTrainingConfig):
         encoded_dim_ranges.append((start, end))
 
     total_encoded_dims = sum(end - start for start, end in encoded_dim_ranges)
-    print(f"Encoding {total_encoded_dims} dimensions: {cfg.encoded_dims}")
+    logger.info(f"Encoding {total_encoded_dims} dimensions: {cfg.encoded_dims}")
 
     # parse relative dimensions
     relative_dim_list = None
     if cfg.relative_dims is not None and cfg.relative_dims.strip():
         relative_dim_list = [int(d.strip()) for d in cfg.relative_dims.split(",")]
-        print(f"Relative dimensions: {relative_dim_list}")
+        logger.info(f"Relative dimensions: {relative_dim_list}")
     else:
-        print("No relative dimensions specified")
+        logger.info("No relative dimensions specified")
 
-    print(f"Use relative transform: {cfg.use_relative_transform}")
+    logger.info(f"Use relative transform: {cfg.use_relative_transform}")
     if cfg.use_relative_transform and (relative_dim_list is None or len(relative_dim_list) == 0):
-        print(
+        logger.warning(
             "Warning: use_relative_transform=True but no relative_dims specified. "
             "No relative transform will be applied."
         )
 
-    print(f"Action horizon: {cfg.action_horizon}")
-    print(f"State key: {cfg.state_key}")
+    logger.info(f"Action horizon: {cfg.action_horizon}")
+    logger.info(f"State key: {cfg.state_key}")
 
     # determine episodes to process
     num_episodes = dataset.num_episodes
     if cfg.max_episodes is not None:
         num_episodes = min(cfg.max_episodes, num_episodes)
 
-    print(f"Processing {num_episodes} episodes...")
+    logger.info(f"Processing {num_episodes} episodes...")
 
     # process episodes sequentially (to avoid pickling issues with dataset)
     all_chunks = []
     for ep_idx in range(num_episodes):
         if ep_idx % 10 == 0:
-            print(f"  Processing episode {ep_idx}/{num_episodes}...")
+            logger.info(f"  Processing episode {ep_idx}/{num_episodes}...")
 
         chunks = process_episode(
             (
@@ -455,19 +456,19 @@ def train_tokenizer(cfg: TokenizerTrainingConfig):
 
     # concatenate all chunks
     all_chunks = np.concatenate(all_chunks, axis=0)
-    print(f"Collected {len(all_chunks)} action chunks")
+    logger.info(f"Collected {len(all_chunks)} action chunks")
 
     # extract only encoded dimensions FIRST (before normalization)
     encoded_chunks = []
     for start, end in encoded_dim_ranges:
         encoded_chunks.append(all_chunks[:, :, start:end])
     encoded_chunks = np.concatenate(encoded_chunks, axis=-1)  # [N, H, D_encoded]
-    print(f"Extracted {encoded_chunks.shape[-1]} encoded dimensions")
+    logger.info(f"Extracted {encoded_chunks.shape[-1]} encoded dimensions")
 
     # apply normalization to encoded dimensions
-    print("\nBefore normalization - overall stats:")
-    print(f"  Min: {np.min(encoded_chunks):.4f}, Max: {np.max(encoded_chunks):.4f}")
-    print(f"  Mean: {np.mean(encoded_chunks):.4f}, Std: {np.std(encoded_chunks):.4f}")
+    logger.info("\nBefore normalization - overall stats:")
+    logger.info(f"  Min: {np.min(encoded_chunks):.4f}, Max: {np.max(encoded_chunks):.4f}")
+    logger.info(f"  Mean: {np.mean(encoded_chunks):.4f}, Std: {np.std(encoded_chunks):.4f}")
 
     # get normalization stats from dataset
     norm_stats = dataset.meta.stats
@@ -489,9 +490,9 @@ def train_tokenizer(cfg: TokenizerTrainingConfig):
                     encoded_stats[stat_name] = stat_array[encoded_dim_indices]
 
         if encoded_stats:
-            print(f"\nNormalization stats for encoded dimensions (mode: {norm_mode.value}):")
+            logger.info(f"\nNormalization stats for encoded dimensions (mode: {norm_mode.value}):")
             for stat_name, stat_values in encoded_stats.items():
-                print(
+                logger.info(
                     f"  {stat_name}: shape={stat_values.shape}, "
                     f"range=[{np.min(stat_values):.4f}, {np.max(stat_values):.4f}]"
                 )
@@ -499,27 +500,27 @@ def train_tokenizer(cfg: TokenizerTrainingConfig):
             # apply normalization based on mode
             try:
                 encoded_chunks = apply_normalization(encoded_chunks, encoded_stats, norm_mode, eps=1e-8)
-                print(f"\nApplied {norm_mode.value} normalization")
+                logger.info(f"\nApplied {norm_mode.value} normalization")
             except ValueError as e:
-                print(f"Warning: {e}. Using raw actions without normalization.")
+                logger.warning(f"Warning: {e}. Using raw actions without normalization.")
 
-            print("\nAfter normalization - overall stats:")
-            print(f"  Min: {np.min(encoded_chunks):.4f}, Max: {np.max(encoded_chunks):.4f}")
-            print(f"  Mean: {np.mean(encoded_chunks):.4f}, Std: {np.std(encoded_chunks):.4f}")
+            logger.info("\nAfter normalization - overall stats:")
+            logger.info(f"  Min: {np.min(encoded_chunks):.4f}, Max: {np.max(encoded_chunks):.4f}")
+            logger.info(f"  Mean: {np.mean(encoded_chunks):.4f}, Std: {np.std(encoded_chunks):.4f}")
 
-            print("\nPer-dimension stats (after normalization):")
+            logger.info("\nPer-dimension stats (after normalization):")
             for d in range(encoded_chunks.shape[-1]):
                 dim_data = encoded_chunks[:, :, d]
-                print(
+                logger.info(
                     f"  Dim {d}: min={np.min(dim_data):7.4f}, max={np.max(dim_data):7.4f}, "
                     f"mean={np.mean(dim_data):7.4f}, std={np.std(dim_data):7.4f}"
                 )
         else:
-            print("Warning: Could not extract stats for encoded dimensions, using raw actions")
+            logger.warning("Warning: Could not extract stats for encoded dimensions, using raw actions")
     else:
-        print("Warning: No normalization stats found in dataset, using raw actions")
+        logger.warning("Warning: No normalization stats found in dataset, using raw actions")
 
-    print(f"Encoded chunks shape: {encoded_chunks.shape}")
+    logger.info(f"Encoded chunks shape: {encoded_chunks.shape}")
 
     # train FAST tokenizer
     tokenizer = train_fast_tokenizer(
@@ -561,8 +562,8 @@ def train_tokenizer(cfg: TokenizerTrainingConfig):
     with open(output_path / "metadata.json", "w") as f:
         json.dump(metadata, f, indent=2)
 
-    print(f"\nSaved FAST tokenizer to {output_path}")
-    print(f"Metadata: {json.dumps(metadata, indent=2)}")
+    logger.info(f"\nSaved FAST tokenizer to {output_path}")
+    logger.info(f"Metadata: {json.dumps(metadata, indent=2)}")
 
     # push to Hugging Face Hub if requested
     if cfg.push_to_hub:
@@ -570,10 +571,10 @@ def train_tokenizer(cfg: TokenizerTrainingConfig):
         hub_repo_id = cfg.hub_repo_id
         if hub_repo_id is None:
             hub_repo_id = output_path.name
-            print(f"\nNo hub_repo_id provided, using: {hub_repo_id}")
+            logger.info(f"\nNo hub_repo_id provided, using: {hub_repo_id}")
 
-        print(f"\nPushing tokenizer to Hugging Face Hub: {hub_repo_id}")
-        print(f"   Private: {cfg.hub_private}")
+        logger.info(f"\nPushing tokenizer to Hugging Face Hub: {hub_repo_id}")
+        logger.info(f"   Private: {cfg.hub_private}")
 
         try:
             # use the tokenizer's push_to_hub method
@@ -593,14 +594,15 @@ def train_tokenizer(cfg: TokenizerTrainingConfig):
                 commit_message="Upload tokenizer metadata",
             )
 
-            print(f"Successfully pushed tokenizer to: https://huggingface.co/{hub_repo_id}")
+            logger.info(f"Successfully pushed tokenizer to: https://huggingface.co/{hub_repo_id}")
         except Exception as e:
-            print(f"Error pushing to hub: {e}")
-            print("   Make sure you're logged in with `huggingface-cli login`")
+            logger.error(f"Error pushing to hub: {e}")
+            logger.error("   Make sure you're logged in with `huggingface-cli login`")
 
 
 def main():
     """CLI entry point that parses arguments and runs the tokenizer training."""
+    init_logging()
     train_tokenizer()
 
 

src/lerobot/teleoperators/bi_so_leader/bi_so_leader.py

@@ -67,7 +67,15 @@ class BiSOLeader(BimanualMixin, Teleoperator):
 
     @cached_property
     def feedback_features(self) -> dict[str, type]:
-        return {}
+        # Bimanual teleop has feedback (can be actuated for handover).
+        # Return the same structure as action_features for consistency with left/right arms.
+        left_arm_features = self.left_arm.feedback_features
+        right_arm_features = self.right_arm.feedback_features
+
+        return {
+            **{f"left_{k}": v for k, v in left_arm_features.items()},
+            **{f"right_{k}": v for k, v in right_arm_features.items()},
+        }
 
     def setup_motors(self) -> None:
         self.left_arm.setup_motors()
@@ -87,6 +95,43 @@ class BiSOLeader(BimanualMixin, Teleoperator):
 
         return action_dict
 
+    def enable_torque(self) -> None:
+        """Enable torque on both leader arms for smooth handover."""
+        self.left_arm.enable_torque()
+        self.right_arm.enable_torque()
+
+    def disable_torque(self) -> None:
+        """Disable torque on both leader arms to allow human control."""
+        self.left_arm.disable_torque()
+        self.right_arm.disable_torque()
+
+    @check_if_not_connected
     def send_feedback(self, feedback: dict[str, float]) -> None:
-        # TODO: Implement force feedback
-        raise NotImplementedError
+        """Route bimanual feedback to left and right arms with proper prefix stripping.
+
+        Receives feedback dict with keys like: left_shoulder_pan.pos, right_shoulder_pan.pos, ...
+        Splits and routes to each arm by removing the prefix.
+
+        This enables DAgger smooth handover: when transitioning from policy control to human
+        intervention, both leader arms are commanded to the follower's current pose to avoid
+        discontinuities.
+        """
+        # Split feedback by arm prefix
+        left_feedback = {}
+        right_feedback = {}
+
+        for key, value in feedback.items():
+            if key.startswith("left_"):
+                # Strip "left_" prefix and pass to left arm
+                stripped_key = key[5:]  # len("left_") == 5
+                left_feedback[stripped_key] = value
+            elif key.startswith("right_"):
+                # Strip "right_" prefix and pass to right arm
+                stripped_key = key[6:]  # len("right_") == 6
+                right_feedback[stripped_key] = value
+
+        # Send to each arm
+        if left_feedback:
+            self.left_arm.send_feedback(left_feedback)
+        if right_feedback:
+            self.right_arm.send_feedback(right_feedback)

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/teleoperators/so_leader/config_so_leader.py

@@ -29,6 +29,12 @@ class SOLeaderConfig:
     # Whether to use degrees for angles
     use_degrees: bool = True
 
+    # Number of extra attempts when a `sync_read` of the motors fails. Feetech buses can occasionally
+    # return a corrupted status packet ("Incorrect status packet!"), especially when several joints move
+    # at once, which otherwise aborts the teleoperation loop. Retries are immediate (no sleep) and only
+    # happen on failure, so the steady-state read cost is unchanged.
+    num_read_retries: int = 2
+
 
 @TeleoperatorConfig.register_subclass("so101_leader")
 @TeleoperatorConfig.register_subclass("so100_leader")

src/lerobot/teleoperators/so_leader/so_leader.py

@@ -145,7 +145,7 @@ class SOLeader(Teleoperator):
     @check_if_not_connected
     def get_action(self) -> dict[str, float]:
         start = time.perf_counter()
-        action = self.bus.sync_read("Present_Position")
+        action = self.bus.sync_read("Present_Position", num_retry=self.config.num_read_retries)
         action = {f"{motor}.pos": val for motor, val in action.items()}
         dt_ms = (time.perf_counter() - start) * 1e3
         logger.debug(f"{self} read action: {dt_ms:.1f}ms")

src/lerobot/transforms/transforms.py

@@ -41,7 +41,7 @@ class RandomSubsetApply(Transform):
 
     def __init__(
         self,
-        transforms: Sequence[Callable],
+        transforms: Sequence[Callable[..., Any]],
         p: list[float] | None = None,
         n_subset: int | None = None,
         random_order: bool = False,
@@ -50,7 +50,7 @@ class RandomSubsetApply(Transform):
         if not isinstance(transforms, Sequence):
             raise TypeError("Argument transforms should be a sequence of callables")
         if p is None:
-            p = [1] * len(transforms)
+            p = [1.0] * len(transforms)
         elif len(p) != len(transforms):
             raise ValueError(
                 f"Length of p doesn't match the number of transforms: {len(p)} != {len(transforms)}"
@@ -69,7 +69,7 @@ class RandomSubsetApply(Transform):
         self.n_subset = n_subset
         self.random_order = random_order
 
-        self.selected_transforms = None
+        self.selected_transforms: list[Callable[..., Any]] = []
 
     def forward(self, *inputs: Any) -> Any:
         needs_unpacking = len(inputs) > 1
@@ -119,7 +119,7 @@ class SharpnessJitter(Transform):
         super().__init__()
         self.sharpness = self._check_input(sharpness)
 
-    def _check_input(self, sharpness):
+    def _check_input(self, sharpness: float | Sequence[float]) -> tuple[float, float]:
         if isinstance(sharpness, (int | float)):
             if sharpness < 0:
                 raise ValueError("If sharpness is a single number, it must be non negative.")
@@ -215,7 +215,7 @@ class ImageTransformsConfig:
     )
 
 
-def make_transform_from_config(cfg: ImageTransformConfig):
+def make_transform_from_config(cfg: ImageTransformConfig) -> Transform:
     if cfg.type == "SharpnessJitter":
         return SharpnessJitter(**cfg.kwargs)
 
@@ -236,8 +236,8 @@ class ImageTransforms(Transform):
         super().__init__()
         self._cfg = cfg
 
-        self.weights = []
-        self.transforms = {}
+        self.weights: list[float] = []
+        self.transforms: dict[str, Transform] = {}
         for tf_name, tf_cfg in cfg.tfs.items():
             if tf_cfg.weight <= 0.0:
                 continue

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/process.py

@@ -16,11 +16,39 @@
 # limitations under the License.
 
 import logging
+import multiprocessing
 import os
 import signal
 import sys
 
 
+def ensure_multiprocessing_start_method(start_method: str | None) -> None:
+    """Set a multiprocessing start method once, or verify the existing method matches.
+
+    Passing ``None`` leaves Python's process-wide default untouched. This is useful
+    when LeRobot is embedded in an application that owns multiprocessing setup.
+    """
+    if start_method is None:
+        return
+
+    available_methods = multiprocessing.get_all_start_methods()
+    if start_method not in available_methods:
+        raise ValueError(
+            f"Multiprocessing start method must be one of {available_methods} on this platform, "
+            f"got {start_method!r}."
+        )
+
+    current_method = multiprocessing.get_start_method(allow_none=True)
+    if current_method is None:
+        multiprocessing.set_start_method(start_method)
+    elif current_method != start_method:
+        raise RuntimeError(
+            f"Multiprocessing start method is already {current_method!r}; cannot change it to "
+            f"{start_method!r}. Set the configured multiprocessing context to null to keep the "
+            "application's existing method, or launch LeRobot in a fresh process."
+        )
+
+
 class ProcessSignalHandler:
     """Utility class to attach graceful shutdown signal handlers.
 

src/lerobot/utils/random_utils.py

@@ -85,6 +85,8 @@ def serialize_torch_rng_state() -> dict[str, torch.Tensor]:
     torch_rng_state_dict = {"torch_rng_state": torch.get_rng_state()}
     if torch.cuda.is_available():
         torch_rng_state_dict["torch_cuda_rng_state"] = torch.cuda.get_rng_state()
+    if torch.backends.mps.is_available():
+        torch_rng_state_dict["torch_mps_rng_state"] = torch.mps.get_rng_state()
     return torch_rng_state_dict
 
 
@@ -95,6 +97,8 @@ def deserialize_torch_rng_state(rng_state_dict: dict[str, torch.Tensor]) -> None
     torch.set_rng_state(rng_state_dict["torch_rng_state"])
     if torch.cuda.is_available() and "torch_cuda_rng_state" in rng_state_dict:
         torch.cuda.set_rng_state(rng_state_dict["torch_cuda_rng_state"])
+    if torch.backends.mps.is_available() and "torch_mps_rng_state" in rng_state_dict:
+        torch.mps.set_rng_state(rng_state_dict["torch_mps_rng_state"])
 
 
 def serialize_rng_state() -> dict[str, torch.Tensor]:

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)
@@ -133,10 +134,13 @@ def say(text: str, blocking: bool = False):
     else:
         raise RuntimeError("Unsupported operating system for text-to-speech.")
 
+    try:
         if blocking:
-        subprocess.run(cmd, check=True)
+            subprocess.run(cmd, check=True, timeout=5)
         else:
             subprocess.Popen(cmd, creationflags=subprocess.CREATE_NO_WINDOW if system == "Windows" else 0)
+    except (FileNotFoundError, subprocess.TimeoutExpired) as e:
+        logging.warning("Text-to-speech command failed: %s | Error: %s", cmd, e)
 
 
 def log_say(text: str, play_sounds: bool = True, blocking: bool = False):

tests/cameras/test_opencv.py

@@ -20,7 +20,7 @@
 # ```
 
 from pathlib import Path
-from unittest.mock import patch
+from unittest.mock import MagicMock, patch
 
 import cv2
 import numpy as np
@@ -123,6 +123,73 @@ def test_invalid_width_connect():
         camera.connect(warmup=False)
 
 
+def test_connect_cleans_up_after_settings_failure_and_allows_retry():
+    config = OpenCVCameraConfig(index_or_path=DEFAULT_PNG_FILE_PATH, warmup_s=0)
+    camera = OpenCVCamera(config)
+    opened_captures = []
+
+    def fail_settings():
+        opened_captures.append(camera.videocapture)
+        raise RuntimeError("settings failed")
+
+    with (
+        patch.object(camera, "_configure_capture_settings", side_effect=fail_settings),
+        pytest.raises(RuntimeError, match="settings failed"),
+    ):
+        camera.connect(warmup=False)
+
+    assert camera.videocapture is None
+    assert camera.thread is None
+    assert not camera.is_connected
+    assert opened_captures[0] is not None
+    assert not opened_captures[0].isOpened()
+
+    camera.connect(warmup=False)
+    assert camera.is_connected
+    camera.disconnect()
+
+
+def test_connect_cleans_up_after_warmup_failure_and_allows_retry():
+    config = OpenCVCameraConfig(index_or_path=DEFAULT_PNG_FILE_PATH, warmup_s=1)
+    camera = OpenCVCamera(config)
+    read_threads = []
+
+    def fail_warmup(*_args, **_kwargs):
+        read_threads.append(camera.thread)
+        raise TimeoutError("no frame")
+
+    with (
+        patch.object(camera, "async_read", side_effect=fail_warmup),
+        pytest.raises(TimeoutError, match="no frame"),
+    ):
+        camera.connect()
+
+    assert camera.videocapture is None
+    assert camera.thread is None
+    assert not camera.is_connected
+    assert read_threads[0] is not None
+    assert not read_threads[0].is_alive()
+
+    camera.connect(warmup=False)
+    assert camera.is_connected
+    camera.disconnect()
+
+
+def test_find_cameras_releases_unopened_handles():
+    module_path = OpenCVCamera.__module__
+    unopened_capture = MagicMock()
+    unopened_capture.isOpened.return_value = False
+
+    with (
+        patch(f"{module_path}.platform.system", return_value="Darwin"),
+        patch(f"{module_path}.MAX_OPENCV_INDEX", 1),
+        patch(f"{module_path}.cv2.VideoCapture", return_value=unopened_capture),
+    ):
+        assert OpenCVCamera.find_cameras() == []
+
+    unopened_capture.release.assert_called_once_with()
+
+
 @pytest.mark.parametrize("index_or_path", TEST_IMAGE_PATHS, ids=TEST_IMAGE_SIZES)
 def test_read(index_or_path):
     config = OpenCVCameraConfig(index_or_path=index_or_path, warmup_s=0)

tests/cameras/test_realsense.py

@@ -20,7 +20,7 @@
 # ```
 
 from pathlib import Path
-from unittest.mock import patch
+from unittest.mock import MagicMock, patch
 
 import numpy as np
 import pytest
@@ -30,6 +30,8 @@ from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnected
 
 pytest.importorskip("pyrealsense2")
 
+import pyrealsense2 as rs
+
 from lerobot.cameras.realsense import RealSenseCamera, RealSenseCameraConfig
 
 TEST_ARTIFACTS_DIR = Path(__file__).parent.parent / "artifacts" / "cameras"
@@ -61,6 +63,17 @@ def test_abc_implementation():
     _ = RealSenseCamera(config)
 
 
+@pytest.mark.parametrize("option", ["exposure", "gain", "white_balance"])
+def test_manual_color_option_requires_rgb(option):
+    with pytest.raises(ValueError, match="use_rgb=True"):
+        RealSenseCameraConfig(
+            serial_number_or_name="042",
+            use_rgb=False,
+            use_depth=True,
+            **{option: 100},
+        )
+
+
 def test_connect():
     config = RealSenseCameraConfig(serial_number_or_name="042", warmup_s=0)
 
@@ -83,6 +96,27 @@ def test_connect_invalid_camera_path(patch_realsense):
         camera.connect(warmup=False)
 
 
+def test_connect_cleans_up_when_sensor_configuration_fails():
+    config = RealSenseCameraConfig(serial_number_or_name="042", exposure=120)
+    camera = RealSenseCamera(config)
+    pipeline = MagicMock()
+    pipeline.start.return_value = MagicMock()
+
+    with (
+        patch("lerobot.cameras.realsense.camera_realsense.rs.pipeline", return_value=pipeline),
+        patch.object(camera, "_configure_rs_pipeline_config"),
+        patch.object(camera, "_configure_capture_settings"),
+        patch.object(camera, "_configure_sensor_options", side_effect=ValueError("invalid exposure")),
+        pytest.raises(ValueError, match="invalid exposure"),
+    ):
+        camera.connect(warmup=False)
+
+    pipeline.stop.assert_called_once_with()
+    assert camera.rs_pipeline is None
+    assert camera.rs_profile is None
+    assert not camera.is_connected
+
+
 def test_invalid_width_connect():
     config = RealSenseCameraConfig(serial_number_or_name="042", width=99999, height=480, fps=30)
     camera = RealSenseCamera(config)
@@ -91,6 +125,33 @@ def test_invalid_width_connect():
         camera.connect(warmup=False)
 
 
+def test_connect_cleans_up_after_warmup_failure_and_allows_retry():
+    config = RealSenseCameraConfig(serial_number_or_name="042", width=640, height=480, fps=30)
+    camera = RealSenseCamera(config)
+    read_threads = []
+
+    def fail_warmup(*_args, **_kwargs):
+        read_threads.append(camera.thread)
+        raise TimeoutError("no frame")
+
+    with (
+        patch.object(camera, "async_read", side_effect=fail_warmup),
+        pytest.raises(TimeoutError, match="no frame"),
+    ):
+        camera.connect()
+
+    assert camera.rs_pipeline is None
+    assert camera.rs_profile is None
+    assert camera.thread is None
+    assert not camera.is_connected
+    assert read_threads[0] is not None
+    assert not read_threads[0].is_alive()
+
+    camera.connect(warmup=False)
+    assert camera.is_connected
+    camera.disconnect()
+
+
 def test_read():
     config = RealSenseCameraConfig(serial_number_or_name="042", width=640, height=480, fps=30, warmup_s=0)
     with RealSenseCamera(config) as camera:
@@ -228,6 +289,203 @@ def test_read_latest_too_old():
             _ = camera.read_latest(max_age_ms=0)  # immediately too old
 
 
+def _make_mock_sensor(name: str, supported_options: set | None = None) -> MagicMock:
+    """Build a fake rs.sensor that reports a name and a configurable supported-options set."""
+    supported = supported_options if supported_options is not None else set()
+    sensor = MagicMock()
+    sensor.get_info.return_value = name
+    sensor.supports.side_effect = lambda opt: opt in supported
+    return sensor
+
+
+def _attach_mock_color_sensor(camera: RealSenseCamera, sensor: MagicMock) -> None:
+    """Wire camera.rs_profile so _get_color_sensor finds the given sensor."""
+    profile = MagicMock()
+    device = MagicMock()
+    device.query_sensors.return_value = [sensor]
+    profile.get_device.return_value = device
+    camera.rs_profile = profile
+
+
+def test_get_color_sensor_prefers_rgb_camera():
+    config = RealSenseCameraConfig(serial_number_or_name="042")
+    camera = RealSenseCamera(config)
+
+    rgb = _make_mock_sensor("RGB Camera")
+    stereo = _make_mock_sensor("Stereo Module")
+    profile = MagicMock()
+    device = MagicMock()
+    device.query_sensors.return_value = [stereo, rgb]
+    profile.get_device.return_value = device
+    camera.rs_profile = profile
+
+    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."""
+    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
+
+
+def test_get_color_sensor_raises_with_available_sensors():
+    config = RealSenseCameraConfig(serial_number_or_name="042")
+    camera = RealSenseCamera(config)
+
+    other = _make_mock_sensor("Motion Module")
+    _attach_mock_color_sensor(camera, other)
+
+    with pytest.raises(RuntimeError, match="Motion Module"):
+        camera._get_color_sensor()
+
+
+def test_configure_sensor_options_skipped_when_none():
+    config = RealSenseCameraConfig(serial_number_or_name="042")
+    camera = RealSenseCamera(config)
+
+    with patch.object(RealSenseCamera, "_get_color_sensor") as mock_get:
+        camera._configure_sensor_options()
+        mock_get.assert_not_called()
+
+
+def test_configure_sensor_options_applies_all_values():
+    config = RealSenseCameraConfig(serial_number_or_name="042", exposure=120, gain=64, white_balance=4600)
+    camera = RealSenseCamera(config)
+
+    sensor = _make_mock_sensor(
+        "RGB Camera",
+        supported_options={
+            rs.option.enable_auto_exposure,
+            rs.option.exposure,
+            rs.option.gain,
+            rs.option.enable_auto_white_balance,
+            rs.option.white_balance,
+        },
+    )
+    _attach_mock_color_sensor(camera, sensor)
+
+    camera._configure_sensor_options()
+
+    sensor.set_option.assert_any_call(rs.option.enable_auto_exposure, 0)
+    sensor.set_option.assert_any_call(rs.option.exposure, 120)
+    sensor.set_option.assert_any_call(rs.option.gain, 64)
+    sensor.set_option.assert_any_call(rs.option.enable_auto_white_balance, 0)
+    sensor.set_option.assert_any_call(rs.option.white_balance, 4600)
+
+
+@pytest.mark.parametrize(
+    ("config_field", "option", "label"),
+    [
+        ("exposure", rs.option.exposure, "exposure"),
+        ("gain", rs.option.gain, "gain"),
+        ("white_balance", rs.option.white_balance, "white balance"),
+    ],
+)
+def test_configure_sensor_options_raises_when_requested_option_is_unsupported(config_field, option, label):
+    config = RealSenseCameraConfig(serial_number_or_name="042", **{config_field: 100})
+    camera = RealSenseCamera(config)
+
+    sensor = _make_mock_sensor("RGB Camera", supported_options=set())
+    _attach_mock_color_sensor(camera, sensor)
+
+    with pytest.raises(ValueError, match=label):
+        camera._configure_sensor_options()
+
+    sensor.supports.assert_any_call(option)
+    sensor.set_option.assert_not_called()
+
+
+@pytest.mark.parametrize(
+    ("config_field", "option", "value"),
+    [
+        ("exposure", rs.option.exposure, 120),
+        ("gain", rs.option.gain, 64),
+    ],
+)
+def test_configure_sensor_options_exposure_or_gain_disables_auto_exposure(config_field, option, value):
+    """white_balance=None should not touch auto white balance."""
+    config = RealSenseCameraConfig(serial_number_or_name="042", **{config_field: value})
+    camera = RealSenseCamera(config)
+
+    sensor = _make_mock_sensor(
+        "RGB Camera",
+        supported_options={rs.option.enable_auto_exposure, option},
+    )
+    _attach_mock_color_sensor(camera, sensor)
+
+    camera._configure_sensor_options()
+
+    calls = [call.args for call in sensor.set_option.call_args_list]
+    assert (rs.option.enable_auto_exposure, 0) in calls
+    assert (option, value) in calls
+    for opt, _ in calls:
+        assert opt != rs.option.enable_auto_white_balance
+        assert opt != rs.option.white_balance
+
+
+def test_configure_sensor_options_warns_when_auto_exposure_control_is_unsupported(caplog):
+    config = RealSenseCameraConfig(serial_number_or_name="042", exposure=120)
+    camera = RealSenseCamera(config)
+
+    sensor = _make_mock_sensor("RGB Camera", supported_options={rs.option.exposure})
+    _attach_mock_color_sensor(camera, sensor)
+
+    with caplog.at_level("WARNING"):
+        camera._configure_sensor_options()
+
+    sensor.set_option.assert_called_once_with(rs.option.exposure, 120)
+    assert "does not support disabling auto-exposure" in caplog.text
+
+
+def test_configure_sensor_options_warns_when_auto_white_balance_control_is_unsupported(caplog):
+    config = RealSenseCameraConfig(serial_number_or_name="042", white_balance=4600)
+    camera = RealSenseCamera(config)
+
+    sensor = _make_mock_sensor("RGB Camera", supported_options={rs.option.white_balance})
+    _attach_mock_color_sensor(camera, sensor)
+
+    with caplog.at_level("WARNING"):
+        camera._configure_sensor_options()
+
+    sensor.set_option.assert_called_once_with(rs.option.white_balance, 4600)
+    assert "does not support disabling auto white balance" in caplog.text
+
+
+def test_configure_sensor_options_out_of_range_raises_value_error():
+    """set_option errors should be re-raised as ValueError with range diagnostics."""
+    config = RealSenseCameraConfig(serial_number_or_name="042", exposure=999999)
+    camera = RealSenseCamera(config)
+
+    sensor = _make_mock_sensor(
+        "RGB Camera",
+        supported_options={rs.option.enable_auto_exposure, rs.option.exposure},
+    )
+
+    def fake_set_option(option, value):
+        if option == rs.option.exposure:
+            raise RuntimeError("value out of range")
+
+    sensor.set_option.side_effect = fake_set_option
+
+    option_range = MagicMock(min=1, max=10000, step=1, default=156)
+    sensor.get_option_range.return_value = option_range
+
+    _attach_mock_color_sensor(camera, sensor)
+
+    with pytest.raises(ValueError, match="exposure") as exc_info:
+        camera._configure_sensor_options()
+
+    msg = str(exc_info.value)
+    assert "999999" in msg
+    assert "min=1" in msg
+    assert "max=10000" in msg
+
+
 @pytest.mark.parametrize(
     "rotation",
     [

tests/configs/test_resume_from_hub.py

@@ -66,3 +66,27 @@ def test_from_pretrained_raises_when_no_root_config_and_no_checkpoints(monkeypat
 
     with pytest.raises(FileNotFoundError, match="train_config.json not found"):
         TrainPipelineConfig.from_pretrained("user/empty-repo")
+
+
+@pytest.mark.parametrize("pass_dir", [False, True])
+def test_resolve_resume_checkpoint_accepts_file_or_pretrained_model_dir(tmp_path, monkeypatch, pass_dir):
+    """`--config_path` may point at the checkpoint's train_config.json or at its
+    pretrained_model/ directory; both must resolve `policy.pretrained_path` to the
+    pretrained_model/ directory (regression test for the directory case, which
+    previously resolved one level too high and failed on model.safetensors)."""
+    pretrained_dir = tmp_path / "checkpoints" / "000002" / "pretrained_model"
+    pretrained_dir.mkdir(parents=True)
+    (pretrained_dir / "train_config.json").touch()
+    target = pretrained_dir if pass_dir else pretrained_dir / "train_config.json"
+
+    from lerobot.policies.act.configuration_act import ACTConfig
+
+    cfg = tc.draccus.parse(TrainPipelineConfig, args=["--dataset.repo_id", "u/d"])
+    cfg.policy = ACTConfig()
+    cfg.resume = True
+    monkeypatch.setattr(tc.parser, "parse_arg", lambda name: str(target) if name == "config_path" else None)
+
+    cfg._resolve_resume_checkpoint()
+
+    assert cfg.policy.pretrained_path == pretrained_dir
+    assert cfg.checkpoint_path == pretrained_dir.parent

tests/datasets/test_augment_quantile_stats.py

@@ -0,0 +1,104 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import numpy as np
+import pytest
+
+pytest.importorskip("datasets", reason="datasets is required (install lerobot[dataset])")
+
+from lerobot.scripts.augment_dataset_quantile_stats import (
+    compute_quantile_stats_for_dataset,
+    has_quantile_stats,
+)
+
+
+def _numeric_keys(dataset):
+    return [k for k, v in dataset.features.items() if v["dtype"] not in ("image", "video", "string")]
+
+
+def _image_keys(dataset):
+    return [k for k, v in dataset.features.items() if v["dtype"] in ("image", "video")]
+
+
+def test_numeric_stats_are_unaffected_by_sampling(tmp_path, lerobot_dataset_factory):
+    """Sampling only touches image/video frames; numeric features are read in
+    full either way, so their stats must be identical with and without sampling."""
+    dataset = lerobot_dataset_factory(
+        root=tmp_path / "ds", total_episodes=2, total_frames=400, use_videos=False
+    )
+
+    exact = compute_quantile_stats_for_dataset(dataset, use_sampling=False)
+    sampled = compute_quantile_stats_for_dataset(dataset, use_sampling=True)
+
+    numeric_keys = _numeric_keys(dataset)
+    assert numeric_keys, "fixture should expose numeric features"
+    for key in numeric_keys:
+        if key not in exact:
+            continue
+        for stat in ("mean", "std", "q01", "q50", "q99"):
+            if stat in exact[key]:
+                np.testing.assert_allclose(
+                    sampled[key][stat],
+                    exact[key][stat],
+                    rtol=1e-6,
+                    atol=1e-6,
+                    err_msg=f"numeric feature '{key}' stat '{stat}' changed under sampling",
+                )
+
+
+def test_image_sampling_reduces_data_but_keeps_stats_close(tmp_path, lerobot_dataset_factory):
+    """For images, sampling should reduce the number of samples considered while
+    keeping the resulting statistics close to the exact ones."""
+    dataset = lerobot_dataset_factory(
+        root=tmp_path / "ds", total_episodes=2, total_frames=400, use_videos=False
+    )
+
+    exact = compute_quantile_stats_for_dataset(dataset, use_sampling=False)
+    sampled = compute_quantile_stats_for_dataset(dataset, use_sampling=True)
+
+    image_keys = _image_keys(dataset)
+    assert image_keys, "fixture should expose at least one image feature"
+    for key in image_keys:
+        # sampling actually looked at fewer pixels
+        assert sampled[key]["count"][0] < exact[key]["count"][0]
+        # but per-channel mean stays close
+        np.testing.assert_allclose(
+            sampled[key]["mean"],
+            exact[key]["mean"],
+            rtol=0.15,
+            err_msg=f"image feature '{key}' mean drifted too far under sampling",
+        )
+
+
+def test_short_episodes_use_all_frames(tmp_path, lerobot_dataset_factory):
+    """With episodes shorter than the sampling floor, sampling is a no-op and
+    must produce exactly the same stats as the exact path."""
+    dataset = lerobot_dataset_factory(
+        root=tmp_path / "ds", total_episodes=2, total_frames=40, use_videos=False
+    )
+
+    exact = compute_quantile_stats_for_dataset(dataset, use_sampling=False)
+    sampled = compute_quantile_stats_for_dataset(dataset, use_sampling=True)
+
+    for key in _image_keys(dataset):
+        assert sampled[key]["count"][0] == exact[key]["count"][0]
+
+
+def test_quantile_stats_present_after_compute(tmp_path, lerobot_dataset_factory):
+    """The computed stats should contain quantile keys for the dataset."""
+    dataset = lerobot_dataset_factory(
+        root=tmp_path / "ds", total_episodes=2, total_frames=200, use_videos=False
+    )
+    stats = compute_quantile_stats_for_dataset(dataset, use_sampling=True)
+    assert has_quantile_stats(stats)

tests/datasets/test_dataset_tools.py

@@ -1125,9 +1125,61 @@ def test_modify_tasks_default_with_overrides(sample_dataset):
             assert ep_data["tasks"][0] == default_task
 
 
+def test_modify_tasks_replacements(sample_dataset):
+    """Test replacing task strings based on their current values."""
+    modified_dataset = modify_tasks(
+        sample_dataset,
+        task_replacements={
+            "task_0": "Pick the cube",
+            "task_1": "Place the cube",
+        },
+    )
+
+    assert len(modified_dataset.meta.tasks) == 2
+    assert "Pick the cube" in modified_dataset.meta.tasks.index
+    assert "Place the cube" in modified_dataset.meta.tasks.index
+
+    for ep_idx in range(5):
+        expected_task = "Pick the cube" if ep_idx % 2 == 0 else "Place the cube"
+        assert modified_dataset.meta.episodes[ep_idx]["tasks"][0] == expected_task
+
+
+def test_modify_tasks_replacements_with_episode_overrides(sample_dataset):
+    """Test that explicit episode overrides take precedence over replacements."""
+    modified_dataset = modify_tasks(
+        sample_dataset,
+        task_replacements={
+            "task_0": "Pick the cube",
+            "task_1": "Place the cube",
+        },
+        episode_tasks={1: "Inspect the cube"},
+    )
+
+    assert modified_dataset.meta.episodes[0]["tasks"][0] == "Pick the cube"
+    assert modified_dataset.meta.episodes[1]["tasks"][0] == "Inspect the cube"
+    assert modified_dataset.meta.episodes[3]["tasks"][0] == "Place the cube"
+    assert len(modified_dataset.meta.tasks) == 3
+
+
+def test_modify_tasks_default_task_and_replacements(sample_dataset):
+    """Test that new_task acts as the default for episodes not matched by task_replacements."""
+    modified_dataset = modify_tasks(
+        sample_dataset,
+        new_task="Default task",
+        task_replacements={"task_0": "Pick the cube"},
+    )
+
+    for ep_idx in range(5):
+        expected_task = "Pick the cube" if ep_idx % 2 == 0 else "Default task"
+        assert modified_dataset.meta.episodes[ep_idx]["tasks"][0] == expected_task
+    assert len(modified_dataset.meta.tasks) == 2
+
+
 def test_modify_tasks_no_task_specified(sample_dataset):
     """Test error when no task is specified."""
-    with pytest.raises(ValueError, match="Must specify at least one of new_task or episode_tasks"):
+    with pytest.raises(
+        ValueError, match="Must specify at least one of new_task, episode_tasks, or task_replacements"
+    ):
         modify_tasks(sample_dataset)
 
 
@@ -1137,6 +1189,12 @@ def test_modify_tasks_invalid_episode_indices(sample_dataset):
         modify_tasks(sample_dataset, episode_tasks={10: "Task", 20: "Task"})
 
 
+def test_modify_tasks_invalid_task_replacements(sample_dataset):
+    """Test error when task replacements refer to unknown task strings."""
+    with pytest.raises(ValueError, match="Task replacements reference unknown tasks"):
+        modify_tasks(sample_dataset, task_replacements={"missing_task": "New task"})
+
+
 def test_modify_tasks_updates_info_json(sample_dataset):
     """Test that total_tasks is updated in info.json."""
     episode_tasks = {0: "Task A", 1: "Task B", 2: "Task C", 3: "Task A", 4: "Task B"}

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

@@ -482,6 +482,20 @@ def test_add_frame_works_in_write_mode(tmp_path):
 # ── Resume mode ──────────────────────────────────────────────────────
 
 
+def test_resume_freshly_created_empty_dataset(tmp_path):
+    """resume() accepts a local dataset created before any episode was recorded."""
+    root = tmp_path / "resume_empty_ds"
+    LeRobotDataset.create(repo_id=DUMMY_REPO_ID, fps=DEFAULT_FPS, features=SIMPLE_FEATURES, root=root)
+
+    resumed = LeRobotDataset.resume(repo_id=DUMMY_REPO_ID, root=root)
+
+    assert isinstance(resumed.writer, DatasetWriter)
+    assert resumed.meta.total_episodes == 0
+    assert resumed.meta.total_frames == 0
+    assert resumed.meta.tasks is None
+    assert resumed.meta.episodes is None
+
+
 def test_resume_creates_writer(tmp_path):
     """After resume(), writer is a DatasetWriter."""
     root = tmp_path / "resume_ds"

tests/motors/test_feetech.py

@@ -294,6 +294,19 @@ def test__sync_read(addr, length, ids_values, mock_motors, dummy_motors):
     assert read_values == ids_values
 
 
+def test__sync_read_retries_after_transient_failure(mock_motors, dummy_motors):
+    addr, length, ids_values = (10, 4, {1: 1337})
+    stub = mock_motors.build_sync_read_stub(addr, length, ids_values, num_invalid_try=1)
+    bus = FeetechMotorsBus(port=mock_motors.port, motors=dummy_motors)
+    bus.connect(handshake=False)
+
+    read_values, read_comm = bus._sync_read(addr, length, list(ids_values), num_retry=1)
+
+    assert read_comm == scs.COMM_SUCCESS
+    assert read_values == ids_values
+    assert mock_motors.stubs[stub].calls == 2
+
+
 @pytest.mark.parametrize("raise_on_error", (True, False))
 def test__sync_read_comm(raise_on_error, mock_motors, dummy_motors):
     addr, length, ids_values = (10, 4, {1: 1337})

tests/policies/evo1/test_evo1.py

@@ -496,6 +496,60 @@ def test_evo1_processor_save_load_round_trip_applies_config_overrides(tmp_path):
     assert "embodiment_id" in processed
 
 
+def test_reconcile_evo1_processors_repads_overridden_stats(tmp_path):
+    """Loading a checkpoint and injecting raw (unpadded) dataset stats must be re-padded.
+
+    Regression test: lerobot-train passes the raw dataset stats as normalizer/unnormalizer
+    overrides when resuming from a checkpoint (e.g. stage2 from a stage1 checkpoint). Those stats
+    are at the dataset dims (e.g. LIBERO state=8/action=7), but EVO1 pads state/action to
+    max_state_dim/max_action_dim before normalization, so reconcile_evo1_processors must re-pad the
+    stats or normalization crashes with a shape mismatch.
+    """
+    config = make_config()
+    preprocessor, postprocessor = make_evo1_pre_post_processors(config, dataset_stats=make_stats())
+    preprocessor.save_pretrained(tmp_path)
+    postprocessor.save_pretrained(tmp_path)
+
+    # Reload with the generic override path injecting raw, unpadded dataset stats.
+    raw_stats = make_stats()
+    loaded_pre = PolicyProcessorPipeline.from_pretrained(
+        tmp_path,
+        config_filename=f"{POLICY_PREPROCESSOR_DEFAULT_NAME}.json",
+        overrides={"normalizer_processor": {"stats": raw_stats}},
+        to_transition=batch_to_transition,
+        to_output=transition_to_batch,
+    )
+    loaded_post = PolicyProcessorPipeline.from_pretrained(
+        tmp_path,
+        config_filename=f"{POLICY_POSTPROCESSOR_DEFAULT_NAME}.json",
+        overrides={"unnormalizer_processor": {"stats": raw_stats}},
+        to_transition=policy_action_to_transition,
+        to_output=transition_to_policy_action,
+    )
+
+    # Sanity: the override really injected unpadded stats before reconciliation.
+    normalizer = next(step for step in loaded_pre.steps if isinstance(step, NormalizerProcessorStep))
+    assert normalizer._tensor_stats[OBS_STATE]["min"].shape == (STATE_DIM,)
+
+    loaded_pre, loaded_post = reconcile_evo1_processors(config, loaded_pre, loaded_post)
+
+    normalizer = next(step for step in loaded_pre.steps if isinstance(step, NormalizerProcessorStep))
+    unnormalizer = next(step for step in loaded_post.steps if isinstance(step, UnnormalizerProcessorStep))
+    assert normalizer._tensor_stats[OBS_STATE]["min"].shape == (MAX_STATE_DIM,)
+    assert normalizer._tensor_stats[ACTION]["min"].shape == (MAX_ACTION_DIM,)
+    assert unnormalizer._tensor_stats[ACTION]["min"].shape == (MAX_ACTION_DIM,)
+
+    # Normalizing a padded state must not raise (this is the exact runtime path that crashed).
+    processed = loaded_pre(
+        {
+            "task": "pick the block",
+            OBS_STATE: torch.zeros(STATE_DIM),
+            f"{OBS_IMAGES}.front": torch.rand(3, 16, 16),
+        }
+    )
+    assert processed[OBS_STATE].shape == (1, MAX_STATE_DIM)
+
+
 def test_evo1_policy_forward_and_inference_use_batched_embedding(monkeypatch):
     monkeypatch.setattr(modeling_evo1, "Evo1Model", DummyEvo1Model)
     policy = modeling_evo1.Evo1Policy(make_config())

tests/policies/test_factory_peft_revision.py

@@ -0,0 +1,83 @@
+# 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 types import SimpleNamespace
+from unittest.mock import MagicMock
+
+import torch
+
+import lerobot.policies.factory as policy_factory
+
+
+def test_make_policy_keeps_peft_adapter_and_base_revisions_separate(monkeypatch):
+    cfg = SimpleNamespace(
+        type="mock",
+        device="cpu",
+        pretrained_path="user/adapter",
+        pretrained_revision="adapter-sha",
+        use_peft=True,
+        input_features={},
+        output_features={},
+    )
+    dataset_meta = SimpleNamespace(features={}, stats={})
+
+    base_policy = torch.nn.Linear(1, 1)
+    policy_from_pretrained = MagicMock(return_value=base_policy)
+    policy_class = SimpleNamespace(from_pretrained=policy_from_pretrained)
+    monkeypatch.setattr(policy_factory, "get_policy_class", lambda _: policy_class)
+    monkeypatch.setattr(policy_factory, "dataset_to_policy_features", lambda _: {})
+    monkeypatch.setattr(policy_factory, "validate_visual_features_consistency", lambda *args: None)
+
+    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 = torch.nn.Linear(1, 1)
+    peft_model_from_pretrained = MagicMock(return_value=adapted_policy)
+    require_package = MagicMock()
+    monkeypatch.setattr(policy_factory, "require_package", require_package)
+    monkeypatch.setattr(
+        policy_factory,
+        "PeftConfig",
+        SimpleNamespace(from_pretrained=peft_config_from_pretrained),
+    )
+    monkeypatch.setattr(
+        policy_factory,
+        "PeftModel",
+        SimpleNamespace(from_pretrained=peft_model_from_pretrained),
+    )
+
+    policy = policy_factory.make_policy(cfg, ds_meta=dataset_meta)
+
+    assert policy is adapted_policy
+    require_package.assert_called_once_with("peft", extra="peft")
+    peft_config_from_pretrained.assert_called_once_with(
+        "user/adapter",
+        revision="adapter-sha",
+    )
+    policy_from_pretrained.assert_called_once_with(
+        config=cfg,
+        dataset_stats=dataset_meta.stats,
+        dataset_meta=dataset_meta,
+        pretrained_name_or_path="user/base-policy",
+        revision="base-sha",
+    )
+    peft_model_from_pretrained.assert_called_once_with(
+        base_policy,
+        "user/adapter",
+        config=peft_config,
+        revision="adapter-sha",
+        is_trainable=True,
+    )

tests/policies/test_gaussian_actor_config.py

@@ -113,6 +113,7 @@ def test_gaussian_actor_config_default_initialization():
     # Concurrency configuration
     assert config.concurrency.actor == "threads"
     assert config.concurrency.learner == "threads"
+    assert config.concurrency.multiprocessing_context == "spawn"
 
     assert isinstance(config.actor_network_kwargs, ActorNetworkConfig)
     assert isinstance(config.policy_kwargs, PolicyConfig)
@@ -152,6 +153,7 @@ def test_concurrency_config():
     config = ConcurrencyConfig()
     assert config.actor == "threads"
     assert config.learner == "threads"
+    assert config.multiprocessing_context == "spawn"
 
 
 def test_gaussian_actor_config_custom_initialization():

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_robot_kinematic_processor.py

@@ -0,0 +1,45 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import pytest
+
+from lerobot.configs import FeatureType, PipelineFeatureType, PolicyFeature
+from lerobot.robots.so_follower.robot_kinematic_processor import (
+    ForwardKinematicsJointsToEEAction,
+    ForwardKinematicsJointsToEEObservation,
+)
+
+MOTOR_NAMES = ["shoulder_pan", "shoulder_lift", "elbow_flex", "wrist_flex", "wrist_roll", "gripper"]
+EE_KEYS = {f"ee.{k}" for k in ["x", "y", "z", "wx", "wy", "wz", "gripper_pos"]}
+
+
+def _joint_bucket(feature_type: FeatureType) -> dict[str, PolicyFeature]:
+    return {f"{n}.pos": PolicyFeature(type=feature_type, shape=(1,)) for n in MOTOR_NAMES}
+
+
+@pytest.mark.parametrize(
+    ("step_cls", "bucket", "feature_type"),
+    [
+        (ForwardKinematicsJointsToEEAction, PipelineFeatureType.ACTION, FeatureType.ACTION),
+        (ForwardKinematicsJointsToEEObservation, PipelineFeatureType.OBSERVATION, FeatureType.STATE),
+    ],
+)
+def test_fk_feature_schema(step_cls, bucket, feature_type):
+    features = {PipelineFeatureType.ACTION: {}, PipelineFeatureType.OBSERVATION: {}}
+    features[bucket] = _joint_bucket(feature_type)
+    out = step_cls(kinematics=None, motor_names=MOTOR_NAMES).transform_features(features)[bucket]
+    assert set(out) == EE_KEYS
+    assert {feature.type for feature in out.values()} == {feature_type}