fix(policies): vla jepa prepare model input to take index 0 and not index -1

* fix(datasets): stop frame errors being treated as shard exhaustion in StreamingLeRobotDataset

StreamingLeRobotDataset.__iter__ caught every RuntimeError and treated all as exhausted shard. Real errors like video decode failure made each shard get dropped on the first frame, so iteration ended while yeilding zero frames with no errors.

Shard exahustion is StopIteration raised from make_frame generator, which python converts to RuntimeError with StopIteration as __cause__. Added check to tell StopIteration from everything else, consuming real shard exhaustion while re-raising everything else.

Added a test that injects a decode failure and asserts iteration raises instead of returning on empty stream.

Fixes #4066

* refactor(datasets): exception streaming

---------

Co-authored-by: Mohit Yadav <mohitydv09@gmail.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

@@ -61,15 +61,19 @@ 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
-git lfs install && git lfs pull
-hf auth login                               # required to push datasets/policies
-```
+# 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
 
-Contributors can alternatively use `uv sync --locked --extra feetech` (see `AGENTS.md`).
+# 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
+```
 
 **4.2 Find USB ports** — run once per arm, unplug when prompted.
 

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/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/robocasa.mdx

@@ -82,6 +82,8 @@ By default the env samples objects only from the `lightwheel` registry (what `--
 
 All eval snippets below mirror the CI command (see `.github/workflows/benchmark_tests.yml`). The `--rename_map` argument maps RoboCasa's native camera keys (`robot0_agentview_left` / `robot0_eye_in_hand` / `robot0_agentview_right`) onto the three-camera (`camera1` / `camera2` / `camera3`) input layout the released `smolvla_robocasa` policy was trained on.
 
+By default, each task uses the rollout horizon registered by RoboCasa. Set `--env.episode_length=<steps>` to apply the same explicit horizon to every selected task.
+
 ### Single-task evaluation (recommended for quick iteration)
 
 ```bash

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

@@ -50,11 +50,11 @@ 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 \

examples/isaac_teleop_to_so101/common.py

@@ -44,6 +44,7 @@ from typing import Protocol
 
 import numpy as np
 
+from lerobot.lerobot_types import RobotAction, RobotObservation
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.processor import (
     RobotProcessorPipeline,
@@ -56,7 +57,6 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
     EEBoundsAndSafety,
     InverseKinematicsEEToJoints,
 )
-from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.constants import HF_LEROBOT_CALIBRATION, HF_LEROBOT_HOME, TELEOPERATORS
 from lerobot.utils.robot_utils import precise_sleep
 

examples/isaac_teleop_to_so101/isaac_teleop/teleop_so101_leader_arm.py

@@ -38,7 +38,7 @@ from typing import TYPE_CHECKING
 
 import numpy as np
 
-from lerobot.types import RobotAction
+from lerobot.lerobot_types import RobotAction
 
 from .base import _GRIPPER_MOTOR_SCALE, IsaacTeleopTeleoperator, _isaacteleop_available
 from .config_isaac_teleop import SO101LeaderArmConfig

examples/isaac_teleop_to_so101/isaac_teleop/teleop_xr_controller.py

@@ -32,7 +32,7 @@ from typing import TYPE_CHECKING, Any
 
 import numpy as np
 
-from lerobot.types import RobotAction
+from lerobot.lerobot_types import RobotAction
 
 from .base import IsaacTeleopTeleoperator, _isaacteleop_available
 from .config_isaac_teleop import XRControllerConfig

examples/isaac_teleop_to_so101/isaac_teleop/xr_controller_processor.py

@@ -26,8 +26,8 @@ from __future__ import annotations
 from dataclasses import dataclass
 
 from lerobot.configs.types import FeatureType, PipelineFeatureType, PolicyFeature
+from lerobot.lerobot_types import RobotAction
 from lerobot.processor import ProcessorStepRegistry, RobotActionProcessorStep
-from lerobot.types import RobotAction
 from lerobot.utils.rotation import Rotation
 
 from .base import _GRIPPER_MOTOR_SCALE

examples/phone_to_so100/evaluate.py

@@ -21,6 +21,7 @@ from lerobot.cameras.opencv import OpenCVCameraConfig
 from lerobot.common.control_utils import predict_action
 from lerobot.configs import FeatureType, PolicyFeature
 from lerobot.datasets import LeRobotDataset, aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.lerobot_types import RobotAction, RobotObservation
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.policies import make_pre_post_processors
 from lerobot.policies.act import ACTPolicy
@@ -38,7 +39,6 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
     ForwardKinematicsJointsToEE,
     InverseKinematicsEEToJoints,
 )
-from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.constants import ACTION, OBS_STR
 from lerobot.utils.feature_utils import build_dataset_frame, combine_feature_dicts
 from lerobot.utils.keyboard_input import init_keyboard_listener

examples/phone_to_so100/record.py

@@ -16,6 +16,7 @@
 
 from lerobot.cameras.opencv import OpenCVCameraConfig
 from lerobot.datasets import LeRobotDataset, aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.lerobot_types import RobotAction, RobotObservation
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.processor import (
     RobotProcessorPipeline,
@@ -36,7 +37,6 @@ from lerobot.scripts.lerobot_record import record_loop
 from lerobot.teleoperators.phone import Phone, PhoneConfig
 from lerobot.teleoperators.phone.config_phone import PhoneOS
 from lerobot.teleoperators.phone.phone_processor import MapPhoneActionToRobotAction
-from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.feature_utils import combine_feature_dicts
 from lerobot.utils.keyboard_input import init_keyboard_listener
 from lerobot.utils.utils import log_say

examples/phone_to_so100/replay.py

@@ -17,6 +17,7 @@
 import time
 
 from lerobot.datasets import LeRobotDataset
+from lerobot.lerobot_types import RobotAction, RobotObservation
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.processor import (
     RobotProcessorPipeline,
@@ -27,7 +28,6 @@ from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
 from lerobot.robots.so_follower.robot_kinematic_processor import (
     InverseKinematicsEEToJoints,
 )
-from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.constants import ACTION
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import log_say

examples/phone_to_so100/rollout.py

@@ -27,6 +27,7 @@ Highlight, or DAgger via ``lerobot-rollout --strategy.type=...``.
 
 from lerobot.cameras.opencv import OpenCVCameraConfig
 from lerobot.configs import PreTrainedConfig
+from lerobot.lerobot_types import RobotAction, RobotObservation
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.processor import (
     RobotProcessorPipeline,
@@ -43,7 +44,6 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
 from lerobot.rollout import BaseStrategyConfig, RolloutConfig, build_rollout_context
 from lerobot.rollout.inference import SyncInferenceConfig
 from lerobot.rollout.strategies import BaseStrategy
-from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.process import ProcessSignalHandler
 from lerobot.utils.utils import init_logging
 

examples/phone_to_so100/teleoperate.py

@@ -15,6 +15,7 @@
 
 import time
 
+from lerobot.lerobot_types import RobotAction, RobotObservation
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.processor import (
     RobotProcessorPipeline,
@@ -31,7 +32,6 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
 from lerobot.teleoperators.phone import Phone, PhoneConfig
 from lerobot.teleoperators.phone.config_phone import PhoneOS
 from lerobot.teleoperators.phone.phone_processor import MapPhoneActionToRobotAction
-from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
 

examples/so100_to_so100_EE/evaluate.py

@@ -21,6 +21,7 @@ from lerobot.cameras.opencv import OpenCVCameraConfig
 from lerobot.common.control_utils import predict_action
 from lerobot.configs import FeatureType, PolicyFeature
 from lerobot.datasets import LeRobotDataset, aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.lerobot_types import RobotAction, RobotObservation
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.policies import make_pre_post_processors
 from lerobot.policies.act import ACTPolicy
@@ -38,7 +39,6 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
     ForwardKinematicsJointsToEE,
     InverseKinematicsEEToJoints,
 )
-from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.constants import ACTION, OBS_STR
 from lerobot.utils.feature_utils import build_dataset_frame, combine_feature_dicts
 from lerobot.utils.keyboard_input import init_keyboard_listener

examples/so100_to_so100_EE/record.py

@@ -17,6 +17,7 @@
 
 from lerobot.cameras.opencv import OpenCVCameraConfig
 from lerobot.datasets import LeRobotDataset, aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.lerobot_types import RobotAction, RobotObservation
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.processor import (
     RobotProcessorPipeline,
@@ -33,7 +34,6 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
 )
 from lerobot.scripts.lerobot_record import record_loop
 from lerobot.teleoperators.so_leader import SO100Leader, SO100LeaderConfig
-from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.feature_utils import combine_feature_dicts
 from lerobot.utils.keyboard_input import init_keyboard_listener
 from lerobot.utils.utils import log_say

examples/so100_to_so100_EE/replay.py

@@ -18,6 +18,7 @@
 import time
 
 from lerobot.datasets import LeRobotDataset
+from lerobot.lerobot_types import RobotAction, RobotObservation
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.processor import (
     RobotProcessorPipeline,
@@ -28,7 +29,6 @@ from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
 from lerobot.robots.so_follower.robot_kinematic_processor import (
     InverseKinematicsEEToJoints,
 )
-from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.constants import ACTION
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import log_say

examples/so100_to_so100_EE/rollout.py

@@ -25,6 +25,7 @@ forward/inverse kinematics.
 
 from lerobot.cameras.opencv import OpenCVCameraConfig
 from lerobot.configs import PreTrainedConfig
+from lerobot.lerobot_types import RobotAction, RobotObservation
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.processor import (
     RobotProcessorPipeline,
@@ -41,7 +42,6 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
 from lerobot.rollout import BaseStrategyConfig, RolloutConfig, build_rollout_context
 from lerobot.rollout.inference import SyncInferenceConfig
 from lerobot.rollout.strategies import BaseStrategy
-from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.process import ProcessSignalHandler
 from lerobot.utils.utils import init_logging
 

examples/so100_to_so100_EE/teleoperate.py

@@ -16,6 +16,7 @@
 
 import time
 
+from lerobot.lerobot_types import RobotAction, RobotObservation
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.processor import (
     RobotProcessorPipeline,
@@ -30,7 +31,6 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
     InverseKinematicsEEToJoints,
 )
 from lerobot.teleoperators.so_leader import SO100Leader, SO100LeaderConfig
-from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
 

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/async_inference/policy_server.py

@@ -38,6 +38,7 @@ import draccus
 import grpc
 import torch
 
+from lerobot.lerobot_types import PolicyAction
 from lerobot.policies import get_policy_class, make_pre_post_processors
 from lerobot.processor import PolicyProcessorPipeline
 from lerobot.transport import (
@@ -45,7 +46,6 @@ from lerobot.transport import (
     services_pb2_grpc,  # type: ignore
 )
 from lerobot.transport.utils import receive_bytes_in_chunks
-from lerobot.types import PolicyAction
 
 from .configs import PolicyServerConfig
 from .constants import SUPPORTED_POLICIES

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."""
@@ -164,17 +172,25 @@ class OpenCVCamera(Camera):
                 f"Failed to open {self}.Run `lerobot-find-cameras opencv` to find available cameras."
             )
 
-        self._configure_capture_settings()
-        self._start_read_thread()
+        try:
+            self._configure_capture_settings()
+            self._start_read_thread()
 
-        if warmup and self.warmup_s > 0:
-            start_time = time.time()
-            while time.time() - start_time < self.warmup_s:
-                self.async_read(timeout_ms=self.warmup_s * 1000)
-                time.sleep(0.1)
-            with self.frame_lock:
-                if self.latest_frame is None:
-                    raise ConnectionError(f"{self} failed to capture frames during warmup.")
+            if warmup and self.warmup_s > 0:
+                start_time = time.time()
+                while time.time() - start_time < self.warmup_s:
+                    self.async_read(timeout_ms=self.warmup_s * 1000)
+                    time.sleep(0.1)
+                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.")
 
@@ -312,32 +328,36 @@ class OpenCVCamera(Camera):
 
         for target in targets_to_scan:
             camera = cv2.VideoCapture(target)
-            if camera.isOpened():
-                default_width = int(camera.get(cv2.CAP_PROP_FRAME_WIDTH))
-                default_height = int(camera.get(cv2.CAP_PROP_FRAME_HEIGHT))
-                default_fps = camera.get(cv2.CAP_PROP_FPS)
-                default_format = camera.get(cv2.CAP_PROP_FORMAT)
+            try:
+                if camera.isOpened():
+                    default_width = int(camera.get(cv2.CAP_PROP_FRAME_WIDTH))
+                    default_height = int(camera.get(cv2.CAP_PROP_FRAME_HEIGHT))
+                    default_fps = camera.get(cv2.CAP_PROP_FPS)
+                    default_format = camera.get(cv2.CAP_PROP_FORMAT)
 
-                # 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)])
+                    # 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)]
+                    )
 
-                camera_info = {
-                    "name": f"OpenCV Camera @ {target}",
-                    "type": "OpenCV",
-                    "id": target,
-                    "backend_api": camera.getBackendName(),
-                    "default_stream_profile": {
-                        "format": default_format,
-                        "fourcc": default_fourcc,
-                        "width": default_width,
-                        "height": default_height,
-                        "fps": default_fps,
-                    },
-                }
+                    camera_info = {
+                        "name": f"OpenCV Camera @ {target}",
+                        "type": "OpenCV",
+                        "id": target,
+                        "backend_api": camera.getBackendName(),
+                        "default_stream_profile": {
+                            "format": default_format,
+                            "fourcc": default_fourcc,
+                            "width": default_width,
+                            "height": default_height,
+                            "fps": default_fps,
+                        },
+                    }
 
-                found_cameras_info.append(camera_info)
+                    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.
         """
@@ -190,22 +206,31 @@ class RealSenseCamera(Camera):
                 f"Failed to open {self}.Run `lerobot-find-cameras realsense` to find available cameras."
             ) from e
 
-        self._configure_capture_settings()
-        self._start_read_thread()
+        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.
-        self.warmup_s = max(self.warmup_s, 1)
+            # 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.
+            self.warmup_s = max(self.warmup_s, 1)
 
-        warmup_read = self.async_read if self.use_rgb else self.async_read_depth
-        start_time = time.time()
-        while time.time() - start_time < self.warmup_s:
-            warmup_read(timeout_ms=self.warmup_s * 1000)
-            time.sleep(0.1)
-        with self.frame_lock:
-            if (self.use_rgb and self.latest_color_frame is None) or (
-                self.use_depth and self.latest_depth_frame is None
-            ):
-                raise ConnectionError(f"{self} failed to capture frames during warmup.")
+            warmup_read = self.async_read if self.use_rgb else self.async_read_depth
+            start_time = time.time()
+            while time.time() - start_time < self.warmup_s:
+                warmup_read(timeout_ms=self.warmup_s * 1000)
+                time.sleep(0.1)
+            with self.frame_lock:
+                if (self.use_rgb and self.latest_color_frame is None) or (
+                    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/common/control_utils.py

@@ -35,9 +35,9 @@ else:
 
 if TYPE_CHECKING:
     from lerobot.datasets import LeRobotDataset
+from lerobot.lerobot_types import PolicyAction
 from lerobot.processor import PolicyProcessorPipeline
 from lerobot.robots import Robot
-from lerobot.types import PolicyAction
 
 
 def predict_action(

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.
@@ -406,15 +438,16 @@ def aggregate_videos(
             videos_idx[key]["dst_file_durations"] = {}
 
     for key, video_idx in videos_idx.items():
-        unique_chunk_file_pairs = {
-            (chunk, file)
-            for chunk, file in zip(
-                src_meta.episodes[f"videos/{key}/chunk_index"],
-                src_meta.episodes[f"videos/{key}/file_index"],
-                strict=False,
-            )
-        }
-        unique_chunk_file_pairs = sorted(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"],
+                    src_meta.episodes[f"videos/{key}/file_index"],
+                    strict=False,
+                )
+            }
+        )
 
         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 = {
-        (c, f)
-        for c, f in zip(
-            src_meta.episodes["data/chunk_index"], src_meta.episodes["data/file_index"], strict=False
-        )
-    }
-
-    unique_chunk_file_ids = sorted(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,
+            )
+        }
+    )
     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,
@@ -580,16 +628,16 @@ 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 = {
-        (c, f)
-        for c, f in zip(
-            src_meta.episodes["meta/episodes/chunk_index"],
-            src_meta.episodes["meta/episodes/file_index"],
-            strict=False,
-        )
-    }
-
-    chunk_file_ids = sorted(chunk_file_ids)
+    chunk_file_ids: list[ChunkFile] = sorted(
+        {
+            (c, f)
+            for c, f in zip(
+                src_meta.episodes["meta/episodes/chunk_index"],
+                src_meta.episodes["meta/episodes/file_index"],
+                strict=False,
+            )
+        }
+    )
     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_writer.py

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

src/lerobot/datasets/pipeline_features.py

@@ -17,8 +17,8 @@ from collections.abc import Sequence
 from typing import Any
 
 from lerobot.configs import PipelineFeatureType
+from lerobot.lerobot_types import RobotAction, RobotObservation
 from lerobot.processor import DataProcessorPipeline
-from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.constants import ACTION, OBS_IMAGES, OBS_STATE, OBS_STR
 from lerobot.utils.feature_utils import hw_to_dataset_features
 

src/lerobot/datasets/streaming_dataset.py

@@ -58,6 +58,10 @@ class LookAheadError(Exception):
     pass
 
 
+class _ShardExhaustedError(Exception):
+    """Raised when a streaming dataset shard has no more items."""
+
+
 class Backtrackable[T]:
     """
     Wrap any iterator/iterable so you can step back up to `history` items
@@ -178,7 +182,7 @@ class Backtrackable[T]:
         """
         Check if we can go back `steps` items without raising an IndexError.
         """
-        return steps <= len(self._back_buf) + self._cursor
+        return steps < len(self._back_buf) + self._cursor
 
     def can_peek_ahead(self, steps: int = 1) -> bool:
         """
@@ -422,10 +426,7 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
                     else:
                         frames_buffer.append(frame)
                     break  # random shard sampled, switch shard
-            except (
-                RuntimeError,
-                StopIteration,
-            ):  # NOTE: StopIteration inside a generator throws a RuntimeError since python 3.7
+            except _ShardExhaustedError:
                 del idx_to_backtrack_dataset[shard_key]  # Remove exhausted shard, onto another shard
 
         # Once shards are all exhausted, shuffle the buffer and yield the remaining frames
@@ -503,7 +504,11 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
 
     def make_frame(self, dataset_iterator: Backtrackable) -> Generator:
         """Makes a frame starting from a dataset iterator"""
-        item = next(dataset_iterator)
+        try:
+            item = next(dataset_iterator)
+        except StopIteration as e:
+            # Translate exhaustion here, before PEP 479 turns it into an indistinguishable RuntimeError.
+            raise _ShardExhaustedError from e
         item = item_to_torch(item)
 
         updates = []  # list of "updates" to apply to the item retrieved from hf_dataset (w/o camera features)

src/lerobot/envs/configs.py

@@ -507,7 +507,7 @@ class MetaworldEnv(EnvConfig):
 class RoboCasaEnv(EnvConfig):
     task: str = "CloseFridge"
     fps: int = 20
-    episode_length: int = 1000
+    episode_length: int | None = None
     obs_type: str = "pixels_agent_pos"
     render_mode: str = "rgb_array"
     camera_name: str = "robot0_agentview_left,robot0_eye_in_hand,robot0_agentview_right"

src/lerobot/envs/libero.py

@@ -30,7 +30,7 @@ from gymnasium import spaces
 from libero.libero import benchmark, get_libero_path
 from libero.libero.envs import OffScreenRenderEnv
 
-from lerobot.types import RobotObservation
+from lerobot.lerobot_types import RobotObservation
 
 from .utils import _LazyAsyncVectorEnv, parse_camera_names
 
@@ -384,7 +384,12 @@ class LiberoEnv(gym.Env):
 
     def close(self):
         if self._env is not None:
-            self._env.close()
+            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

@@ -25,7 +25,7 @@ import metaworld.policies as policies
 import numpy as np
 from gymnasium import spaces
 
-from lerobot.types import RobotObservation
+from lerobot.lerobot_types import RobotObservation
 
 from .utils import _LazyAsyncVectorEnv
 
@@ -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/robocasa.py

@@ -25,7 +25,7 @@ import gymnasium as gym
 import numpy as np
 from gymnasium import spaces
 
-from lerobot.types import RobotObservation
+from lerobot.lerobot_types import RobotObservation
 
 from .utils import _LazyAsyncVectorEnv, parse_camera_names
 
@@ -98,6 +98,19 @@ def _resolve_tasks(task: str) -> tuple[list[str], str | None]:
     return names, None
 
 
+def _get_task_horizon(task: str) -> int:
+    """Return the rollout horizon registered by RoboCasa for a task."""
+    from robocasa.utils.dataset_registry_utils import get_task_horizon
+
+    try:
+        return int(get_task_horizon(task))
+    except ValueError as exc:
+        raise ValueError(
+            f"No RoboCasa horizon is registered for task '{task}'. "
+            "Set `--env.episode_length=<steps>` explicitly."
+        ) from exc
+
+
 def convert_action(flat_action: np.ndarray) -> dict[str, Any]:
     """Split a flat (12,) action vector into a RoboCasa action dict.
 
@@ -154,7 +167,7 @@ class RoboCasaEnv(gym.Env):
 
         self.camera_name = parse_camera_names(camera_name)
 
-        self._max_episode_steps = episode_length if episode_length is not None else 1000
+        self._max_episode_steps = episode_length if episode_length is not None else _get_task_horizon(task)
 
         # Deferred — created on first reset() inside the worker subprocess
         # to avoid inheriting stale GPU/EGL contexts across fork().

src/lerobot/envs/robotwin.py

@@ -28,7 +28,7 @@ import numpy as np
 import torch
 from gymnasium import spaces
 
-from lerobot.types import RobotObservation
+from lerobot.lerobot_types import RobotObservation
 from lerobot.utils.import_utils import _scipy_available
 
 from .utils import _LazyAsyncVectorEnv
@@ -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

@@ -37,7 +37,7 @@ import numpy as np
 from gymnasium import spaces
 from scipy.spatial.transform import Rotation
 
-from lerobot.types import RobotObservation
+from lerobot.lerobot_types import RobotObservation
 
 from .utils import _LazyAsyncVectorEnv
 
@@ -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/eo1/processor_eo1.py

@@ -22,6 +22,7 @@ from typing import TYPE_CHECKING, Any
 import torch
 
 from lerobot.configs.types import FeatureType, PipelineFeatureType, PolicyFeature
+from lerobot.lerobot_types import TransitionKey
 from lerobot.processor import (
     ComplementaryDataProcessorStep,
     PolicyAction,
@@ -31,7 +32,6 @@ from lerobot.processor import (
     make_default_policy_processor_steps,
     make_policy_processor_pipelines,
 )
-from lerobot.types import TransitionKey
 from lerobot.utils.constants import OBS_STATE
 from lerobot.utils.import_utils import _transformers_available, require_package
 

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

@@ -21,6 +21,7 @@ from typing import Any
 import torch
 
 from lerobot.configs import FeatureType, PipelineFeatureType, PolicyFeature
+from lerobot.lerobot_types import EnvTransition, TransitionKey
 from lerobot.processor import (
     AddBatchDimensionProcessorStep,
     DeviceProcessorStep,
@@ -40,7 +41,6 @@ from lerobot.processor.converters import (
     policy_action_to_transition,
     transition_to_policy_action,
 )
-from lerobot.types import EnvTransition, TransitionKey
 from lerobot.utils.constants import (
     ACTION,
     DONE,
@@ -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

@@ -28,6 +28,7 @@ if TYPE_CHECKING:
 
 from lerobot.configs import FeatureType, PreTrainedConfig
 from lerobot.envs import EnvConfig, env_to_policy_features
+from lerobot.lerobot_types import PolicyAction
 from lerobot.processor import (
     AbsoluteActionsProcessorStep,
     PolicyProcessorPipeline,
@@ -37,19 +38,25 @@ from lerobot.processor import (
     transition_to_batch,
     transition_to_policy_action,
 )
-from lerobot.types import PolicyAction
 from lerobot.utils.constants import (
     ACTION,
     POLICY_POSTPROCESSOR_DEFAULT_NAME,
     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

@@ -50,6 +50,7 @@ if TYPE_CHECKING or _datasets_available:
 else:
     LeRobotDataset = None
 
+from lerobot.lerobot_types import EnvTransition, TransitionKey
 from lerobot.processor import (
     AbsoluteActionsProcessorStep,
     AddBatchDimensionProcessorStep,
@@ -66,7 +67,6 @@ from lerobot.processor import (
     transition_to_batch,
     transition_to_policy_action,
 )
-from lerobot.types import EnvTransition, TransitionKey
 from lerobot.utils.constants import (
     ACTION,
     OBS_IMAGE,
@@ -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/molmoact2/processor_molmoact2.py

@@ -36,6 +36,7 @@ import torch
 from torch import Tensor
 
 from lerobot.configs import FeatureType, PipelineFeatureType, PolicyFeature
+from lerobot.lerobot_types import EnvTransition, TransitionKey
 from lerobot.processor import (
     AddBatchDimensionProcessorStep,
     DeviceProcessorStep,
@@ -49,7 +50,6 @@ from lerobot.processor import (
     policy_action_to_transition,
     transition_to_policy_action,
 )
-from lerobot.types import EnvTransition, TransitionKey
 from lerobot.utils.constants import (
     ACTION,
     OBS_IMAGES,

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/pi05/processor_pi05.py

@@ -22,6 +22,7 @@ import numpy as np
 import torch
 
 from lerobot.configs import PipelineFeatureType, PolicyFeature
+from lerobot.lerobot_types import EnvTransition, TransitionKey
 from lerobot.processor import (
     AbsoluteActionsProcessorStep,
     PolicyAction,
@@ -33,7 +34,6 @@ from lerobot.processor import (
     make_default_policy_processor_steps,
     make_policy_processor_pipelines,
 )
-from lerobot.types import EnvTransition, TransitionKey
 from lerobot.utils.constants import OBS_STATE
 
 from .configuration_pi05 import PI05Config

src/lerobot/policies/pi0_fast/processor_pi0_fast.py

@@ -22,6 +22,7 @@ import numpy as np
 import torch
 
 from lerobot.configs import PipelineFeatureType, PolicyFeature
+from lerobot.lerobot_types import EnvTransition, TransitionKey
 from lerobot.processor import (
     AbsoluteActionsProcessorStep,
     ActionTokenizerProcessorStep,
@@ -34,7 +35,6 @@ from lerobot.processor import (
     make_default_policy_processor_steps,
     make_policy_processor_pipelines,
 )
-from lerobot.types import EnvTransition, TransitionKey
 from lerobot.utils.constants import OBS_STATE
 
 from .configuration_pi0_fast import PI0FastConfig

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/policies/utils.py

@@ -22,7 +22,7 @@ import torch
 from torch import nn
 
 from lerobot.configs import FeatureType, PolicyFeature, PreTrainedConfig
-from lerobot.types import PolicyAction, RobotAction, RobotObservation
+from lerobot.lerobot_types import PolicyAction, RobotAction, RobotObservation
 from lerobot.utils.constants import ACTION, OBS_STR
 from lerobot.utils.feature_utils import build_dataset_frame
 

src/lerobot/policies/vla_jepa/configuration_vla_jepa.py

@@ -15,11 +15,13 @@
 from __future__ import annotations
 
 from dataclasses import dataclass, field
+from typing import Any
 
 from lerobot.configs.policies import PreTrainedConfig
-from lerobot.configs.types import NormalizationMode
+from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
 from lerobot.optim.optimizers import AdamWConfig
 from lerobot.optim.schedulers import CosineDecayWithWarmupSchedulerConfig
+from lerobot.utils.constants import OBS_STATE
 
 
 @PreTrainedConfig.register_subclass("vla_jepa")
@@ -122,6 +124,13 @@ class VLAJEPAConfig(PreTrainedConfig):
         if self.robot_state_feature is not None:
             self.state_dim = self.robot_state_feature.shape[0]
 
+    def set_dataset_feature_metadata(self, dataset_features: dict[str, Any]) -> None:
+        """Add `observation.state` to `input_features` if missing, so it gets normalized."""
+        if OBS_STATE in self.input_features or OBS_STATE not in dataset_features:
+            return
+        shape = tuple(dataset_features[OBS_STATE]["shape"])
+        self.input_features[OBS_STATE] = PolicyFeature(type=FeatureType.STATE, shape=shape)
+
     def get_optimizer_preset(self) -> AdamWConfig:
         return AdamWConfig(
             lr=self.optimizer_lr,

src/lerobot/policies/vla_jepa/modeling_vla_jepa.py

@@ -399,7 +399,8 @@ class VLAJEPAPolicy(PreTrainedPolicy):
         state = batch.get(OBS_STATE)
         if state is not None:
             if state.ndim > 2:
-                state = state[:, -1, :]
+                # deltas are forward-looking here, so index 0 is the current observation, not -1.
+                state = state[:, 0, :]
             inputs["state"] = (state.unsqueeze(1) if state.ndim == 2 else state).float()  # [B, 1, dim]
 
         return inputs

src/lerobot/policies/xvla/modeling_xvla.py

@@ -150,7 +150,7 @@ class XVLAModel(nn.Module):
         # Freeze or unfreeze policy transformer
         if not self.config.train_policy_transformer:
             for name, param in self.transformer.named_parameters():
-                if "soft_prompts" not in name:
+                if "soft_prompt" not in name:
                     param.requires_grad = False
 
         # Freeze or unfreeze soft prompts

src/lerobot/policies/xvla/processor_xvla.py

@@ -21,6 +21,7 @@ import numpy as np
 import torch
 
 from lerobot.configs import PipelineFeatureType, PolicyFeature
+from lerobot.lerobot_types import EnvTransition, TransitionKey
 from lerobot.processor import (
     ObservationProcessorStep,
     PolicyAction,
@@ -31,7 +32,6 @@ from lerobot.processor import (
     make_default_policy_processor_steps,
     make_policy_processor_pipelines,
 )
-from lerobot.types import EnvTransition, TransitionKey
 from lerobot.utils.constants import (
     IMAGENET_STATS,
     OBS_IMAGES,

src/lerobot/processor/__init__.py

@@ -14,7 +14,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from lerobot.types import (
+from lerobot.lerobot_types import (
     EnvAction,
     EnvTransition,
     PolicyAction,

src/lerobot/processor/batch_processor.py

@@ -25,7 +25,7 @@ from dataclasses import dataclass, field
 from torch import Tensor
 
 from lerobot.configs import PipelineFeatureType, PolicyFeature
-from lerobot.types import EnvTransition, PolicyAction
+from lerobot.lerobot_types import EnvTransition, PolicyAction
 from lerobot.utils.constants import OBS_ENV_STATE, OBS_IMAGE, OBS_IMAGES, OBS_STATE
 
 from .pipeline import (

src/lerobot/processor/converters.py

@@ -23,7 +23,7 @@ from typing import Any
 import numpy as np
 import torch
 
-from lerobot.types import EnvTransition, PolicyAction, RobotAction, RobotObservation, TransitionKey
+from lerobot.lerobot_types import EnvTransition, PolicyAction, RobotAction, RobotObservation, TransitionKey
 from lerobot.utils.constants import ACTION, DONE, INFO, OBS_PREFIX, REWARD, TRUNCATED
 
 

src/lerobot/processor/delta_action_processor.py

@@ -17,7 +17,7 @@
 from dataclasses import dataclass
 
 from lerobot.configs import FeatureType, PipelineFeatureType, PolicyFeature
-from lerobot.types import PolicyAction, RobotAction
+from lerobot.lerobot_types import PolicyAction, RobotAction
 
 from .pipeline import ActionProcessorStep, ProcessorStepRegistry, RobotActionProcessorStep
 
@@ -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/device_processor.py

@@ -25,7 +25,7 @@ from typing import Any
 import torch
 
 from lerobot.configs import PipelineFeatureType, PolicyFeature
-from lerobot.types import EnvTransition, PolicyAction, TransitionKey
+from lerobot.lerobot_types import EnvTransition, PolicyAction, TransitionKey
 from lerobot.utils.device_utils import get_safe_torch_device
 
 from .pipeline import ProcessorStep, ProcessorStepRegistry

src/lerobot/processor/factory.py

@@ -20,7 +20,7 @@ from typing import Any
 import torch
 
 from lerobot.configs.policies import PreTrainedConfig
-from lerobot.types import PolicyAction, RobotAction, RobotObservation
+from lerobot.lerobot_types import PolicyAction, RobotAction, RobotObservation
 from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
 
 from .batch_processor import AddBatchDimensionProcessorStep

src/lerobot/processor/gym_action_processor.py

@@ -17,7 +17,7 @@
 from dataclasses import dataclass
 
 from lerobot.configs import PipelineFeatureType, PolicyFeature
-from lerobot.types import EnvAction, EnvTransition, PolicyAction, TransitionKey
+from lerobot.lerobot_types import EnvAction, EnvTransition, PolicyAction, TransitionKey
 
 from .converters import to_tensor
 from .hil_processor import TELEOP_ACTION_KEY

src/lerobot/processor/hil_processor.py

@@ -29,7 +29,7 @@ from lerobot.teleoperators.utils import TeleopEvents
 if TYPE_CHECKING:
     from lerobot.teleoperators.teleoperator import Teleoperator
 
-from lerobot.types import EnvTransition, PolicyAction, TransitionKey
+from lerobot.lerobot_types import EnvTransition, PolicyAction, TransitionKey
 
 from .pipeline import (
     ComplementaryDataProcessorStep,

src/lerobot/processor/normalize_processor.py

@@ -25,7 +25,7 @@ import torch
 from torch import Tensor
 
 from lerobot.configs import FeatureType, NormalizationMode, PipelineFeatureType, PolicyFeature
-from lerobot.types import EnvTransition, PolicyAction, TransitionKey
+from lerobot.lerobot_types import EnvTransition, PolicyAction, TransitionKey
 
 if TYPE_CHECKING:
     from lerobot.datasets import LeRobotDataset

src/lerobot/processor/pipeline.py

@@ -45,7 +45,14 @@ from huggingface_hub import hf_hub_download
 from safetensors.torch import load_file, save_file
 
 from lerobot.configs import PipelineFeatureType, PolicyFeature
-from lerobot.types import EnvAction, EnvTransition, PolicyAction, RobotAction, RobotObservation, TransitionKey
+from lerobot.lerobot_types import (
+    EnvAction,
+    EnvTransition,
+    PolicyAction,
+    RobotAction,
+    RobotObservation,
+    TransitionKey,
+)
 from lerobot.utils.constants import HF_LEROBOT_HOME
 from lerobot.utils.hub import HubMixin
 
@@ -713,6 +720,8 @@ 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 +740,7 @@ 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 +930,7 @@ 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 +954,7 @@ 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 +972,7 @@ 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 +986,9 @@ 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 +1152,7 @@ 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 +1171,7 @@ 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 +1192,7 @@ 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 +1206,12 @@ 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/processor/policy_robot_bridge.py

@@ -20,7 +20,7 @@ from typing import Any
 import torch
 
 from lerobot.configs import FeatureType, PipelineFeatureType, PolicyFeature
-from lerobot.types import PolicyAction, RobotAction
+from lerobot.lerobot_types import PolicyAction, RobotAction
 from lerobot.utils.constants import ACTION
 
 from .pipeline import ActionProcessorStep, ProcessorStepRegistry

src/lerobot/processor/relative_action_processor.py

@@ -20,7 +20,7 @@ import torch
 from torch import Tensor
 
 from lerobot.configs import PipelineFeatureType, PolicyFeature
-from lerobot.types import EnvTransition, TransitionKey
+from lerobot.lerobot_types import EnvTransition, TransitionKey
 from lerobot.utils.constants import OBS_STATE
 
 from .delta_action_processor import MapDeltaActionToRobotActionStep, MapTensorToDeltaActionDictStep

src/lerobot/processor/render_messages_processor.py

@@ -23,7 +23,7 @@ from lerobot.configs import PipelineFeatureType, PolicyFeature
 from lerobot.configs.recipe import TrainingRecipe
 from lerobot.datasets.language import LANGUAGE_EVENTS, LANGUAGE_PERSISTENT
 from lerobot.datasets.language_render import render_sample
-from lerobot.types import EnvTransition, TransitionKey
+from lerobot.lerobot_types import EnvTransition, TransitionKey
 from lerobot.utils.utils import unwrap_scalar
 
 from .pipeline import ProcessorStep, ProcessorStepRegistry

src/lerobot/processor/tokenizer_processor.py

@@ -30,7 +30,7 @@ from typing import TYPE_CHECKING, Any
 import torch
 
 from lerobot.configs import FeatureType, PipelineFeatureType, PolicyFeature
-from lerobot.types import EnvTransition, RobotObservation, TransitionKey
+from lerobot.lerobot_types import EnvTransition, RobotObservation, TransitionKey
 from lerobot.utils.constants import (
     ACTION_TOKEN_MASK,
     ACTION_TOKENS,

src/lerobot/rewards/robometer/compute_rabc_weights.py

@@ -57,10 +57,10 @@ import torch
 from tqdm import tqdm
 
 from lerobot.datasets import LeRobotDataset
+from lerobot.lerobot_types import TransitionKey
 from lerobot.rewards.robometer.configuration_robometer import RobometerConfig
 from lerobot.rewards.robometer.modeling_robometer import RobometerRewardModel
 from lerobot.rewards.robometer.processor_robometer import RobometerEncoderProcessorStep
-from lerobot.types import TransitionKey
 
 DEFAULT_OUTPUT_FILENAME = "robometer_progress.parquet"
 

src/lerobot/rewards/robometer/processor_robometer.py

@@ -25,6 +25,7 @@ from PIL import Image
 from torch import Tensor
 
 from lerobot.configs import PipelineFeatureType, PolicyFeature
+from lerobot.lerobot_types import EnvTransition, TransitionKey
 from lerobot.processor import (
     AddBatchDimensionProcessorStep,
     DeviceProcessorStep,
@@ -39,7 +40,6 @@ from lerobot.rewards.robometer.configuration_robometer import (
     RobometerConfig,
 )
 from lerobot.rewards.robometer.modeling_robometer import ROBOMETER_FEATURE_PREFIX
-from lerobot.types import EnvTransition, TransitionKey
 from lerobot.utils.constants import (
     OBS_IMAGES,
     POLICY_POSTPROCESSOR_DEFAULT_NAME,

src/lerobot/rewards/sarm/processor_sarm.py

@@ -47,6 +47,7 @@ else:
     Faker = None  # type: ignore[assignment, misc]
 
 from lerobot.configs import FeatureType, PipelineFeatureType, PolicyFeature
+from lerobot.lerobot_types import EnvTransition, PolicyAction, TransitionKey
 from lerobot.processor import (
     AddBatchDimensionProcessorStep,
     DeviceProcessorStep,
@@ -58,7 +59,6 @@ from lerobot.processor import (
     policy_action_to_transition,
     transition_to_policy_action,
 )
-from lerobot.types import EnvTransition, PolicyAction, TransitionKey
 from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
 
 from .configuration_sarm import SARMConfig

src/lerobot/rewards/topreward/compute_rabc_weights.py

@@ -48,10 +48,10 @@ import torch
 from tqdm import tqdm
 
 from lerobot.datasets import LeRobotDataset
+from lerobot.lerobot_types import TransitionKey
 from lerobot.rewards.topreward.configuration_topreward import TOPRewardConfig
 from lerobot.rewards.topreward.modeling_topreward import TOPRewardModel
 from lerobot.rewards.topreward.processor_topreward import TOPRewardEncoderProcessorStep
-from lerobot.types import TransitionKey
 
 DEFAULT_OUTPUT_FILENAME = "topreward_progress.parquet"
 

src/lerobot/rewards/topreward/processor_topreward.py

@@ -23,6 +23,7 @@ import torch
 from torch import Tensor
 
 from lerobot.configs import PipelineFeatureType, PolicyFeature
+from lerobot.lerobot_types import EnvTransition, TransitionKey
 from lerobot.processor import (
     AddBatchDimensionProcessorStep,
     DeviceProcessorStep,
@@ -37,7 +38,6 @@ from lerobot.rewards.topreward.configuration_topreward import (
     DEFAULT_PROMPT_SUFFIX_TEMPLATE,
     TOPRewardConfig,
 )
-from lerobot.types import EnvTransition, TransitionKey
 from lerobot.utils.constants import (
     OBS_IMAGES,
     OBS_PREFIX,

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/base.py

@@ -28,7 +28,7 @@ from huggingface_hub.errors import HfHubHTTPError
 from safetensors.torch import load_file as load_safetensors, save_file as save_safetensors
 from torch.optim import Optimizer
 
-from lerobot.types import BatchType
+from lerobot.lerobot_types import BatchType
 from lerobot.utils.hub import HubMixin
 
 from .configs import RLAlgorithmConfig, TrainingStats

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/algorithms/sac/sac_algorithm.py

@@ -26,6 +26,7 @@ import torch.nn.functional as F  # noqa: N812
 from torch import Tensor
 from torch.optim import Optimizer
 
+from lerobot.lerobot_types import BatchType
 from lerobot.policies.gaussian_actor.modeling_gaussian_actor import (
     DISCRETE_DIMENSION_INDEX,
     MLP,
@@ -35,7 +36,6 @@ from lerobot.policies.gaussian_actor.modeling_gaussian_actor import (
     orthogonal_init,
 )
 from lerobot.policies.utils import get_device_from_parameters
-from lerobot.types import BatchType
 from lerobot.utils.constants import ACTION
 from lerobot.utils.transition import move_state_dict_to_device