refactor(utils): enhance task handling in add_envs_task function

- Improved the `add_envs_task` function to validate the output of `task_description` and `task` calls, ensuring they return lists of strings.
- Removed the use of `else` statement for environments without language instructions, simplifying the logic and enhancing readability.
- Streamlined the observation dictionary handling by ensuring consistent data types for task attributes.

.github/PULL_REQUEST_TEMPLATE.md

@@ -30,7 +30,7 @@ pytest -sx tests/test_stuff.py::test_something
 ```
 
 ```bash
-python -m lerobot.scripts.train --some.option=true
+lerobot-train --some.option=true
 ```
 
 ## SECTION TO REMOVE BEFORE SUBMITTING YOUR PR

.github/workflows/nightly.yml

@@ -29,8 +29,8 @@ on:
 env:
   UV_VERSION: "0.8.0"
   PYTHON_VERSION: "3.10"
-  DOCKER_IMAGE_NAME_CPU: huggingface/lerobot-gpu:latest
-  DOCKER_IMAGE_NAME_GPU: huggingface/lerobot-cpu:latest
+  DOCKER_IMAGE_NAME_CPU: huggingface/lerobot-cpu:latest
+  DOCKER_IMAGE_NAME_GPU: huggingface/lerobot-gpu:latest
 
 # Ensures that only the latest commit is built, canceling older runs.
 concurrency:

Makefile

@@ -44,7 +44,7 @@ test-end-to-end:
 	${MAKE} DEVICE=$(DEVICE) test-smolvla-ete-eval
 
 test-act-ete-train:
-	python -m lerobot.scripts.train \
+	lerobot-train \
 		--policy.type=act \
 		--policy.dim_model=64 \
 		--policy.n_action_steps=20 \
@@ -68,12 +68,12 @@ test-act-ete-train:
 		--output_dir=tests/outputs/act/
 
 test-act-ete-train-resume:
-	python -m lerobot.scripts.train \
+	lerobot-train \
 		--config_path=tests/outputs/act/checkpoints/000002/pretrained_model/train_config.json \
 		--resume=true
 
 test-act-ete-eval:
-	python -m lerobot.scripts.eval \
+	lerobot-eval \
 		--policy.path=tests/outputs/act/checkpoints/000004/pretrained_model \
 		--policy.device=$(DEVICE) \
 		--env.type=aloha \
@@ -82,7 +82,7 @@ test-act-ete-eval:
 		--eval.batch_size=1
 
 test-diffusion-ete-train:
-	python -m lerobot.scripts.train \
+	lerobot-train \
 		--policy.type=diffusion \
 		--policy.down_dims='[64,128,256]' \
 		--policy.diffusion_step_embed_dim=32 \
@@ -106,7 +106,7 @@ test-diffusion-ete-train:
 		--output_dir=tests/outputs/diffusion/
 
 test-diffusion-ete-eval:
-	python -m lerobot.scripts.eval \
+	lerobot-eval \
 		--policy.path=tests/outputs/diffusion/checkpoints/000002/pretrained_model \
 		--policy.device=$(DEVICE) \
 		--env.type=pusht \
@@ -115,7 +115,7 @@ test-diffusion-ete-eval:
 		--eval.batch_size=1
 
 test-tdmpc-ete-train:
-	python -m lerobot.scripts.train \
+	lerobot-train \
 		--policy.type=tdmpc \
 		--policy.device=$(DEVICE) \
 		--policy.push_to_hub=false \
@@ -137,7 +137,7 @@ test-tdmpc-ete-train:
 		--output_dir=tests/outputs/tdmpc/
 
 test-tdmpc-ete-eval:
-	python -m lerobot.scripts.eval \
+	lerobot-eval \
 		--policy.path=tests/outputs/tdmpc/checkpoints/000002/pretrained_model \
 		--policy.device=$(DEVICE) \
 		--env.type=xarm \
@@ -148,7 +148,7 @@ test-tdmpc-ete-eval:
 
 
 test-smolvla-ete-train:
-	python -m lerobot.scripts.train \
+	lerobot-train \
 		--policy.type=smolvla \
 		--policy.n_action_steps=20 \
 		--policy.chunk_size=20 \
@@ -171,7 +171,7 @@ test-smolvla-ete-train:
 		--output_dir=tests/outputs/smolvla/
 
 test-smolvla-ete-eval:
-	python -m lerobot.scripts.eval \
+	lerobot-eval \
 		--policy.path=tests/outputs/smolvla/checkpoints/000004/pretrained_model \
 		--policy.device=$(DEVICE) \
 		--env.type=aloha \

README.md

@@ -6,7 +6,7 @@
 
 <div align="center">
 
-[![Tests](https://github.com/huggingface/lerobot/actions/workflows/nightly.yml/badge.svg?branch=main)](https://github.com/huggingface/lerobot/actions/workflows/nighty.yml?query=branch%3Amain)
+[![Tests](https://github.com/huggingface/lerobot/actions/workflows/nightly.yml/badge.svg?branch=main)](https://github.com/huggingface/lerobot/actions/workflows/nightly.yml?query=branch%3Amain)
 [![Python versions](https://img.shields.io/pypi/pyversions/lerobot)](https://www.python.org/downloads/)
 [![License](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](https://github.com/huggingface/lerobot/blob/main/LICENSE)
 [![Status](https://img.shields.io/pypi/status/lerobot)](https://pypi.org/project/lerobot/)
@@ -276,7 +276,7 @@ Check out [example 2](https://github.com/huggingface/lerobot/blob/main/examples/
 We also provide a more capable script to parallelize the evaluation over multiple environments during the same rollout. Here is an example with a pretrained model hosted on [lerobot/diffusion_pusht](https://huggingface.co/lerobot/diffusion_pusht):
 
 ```bash
-python -m lerobot.scripts.eval \
+lerobot-eval \
     --policy.path=lerobot/diffusion_pusht \
     --env.type=pusht \
     --eval.batch_size=10 \
@@ -288,10 +288,10 @@ python -m lerobot.scripts.eval \
 Note: After training your own policy, you can re-evaluate the checkpoints with:
 
 ```bash
-python -m lerobot.scripts.eval --policy.path={OUTPUT_DIR}/checkpoints/last/pretrained_model
+lerobot-eval --policy.path={OUTPUT_DIR}/checkpoints/last/pretrained_model
 ```
 
-See `python -m lerobot.scripts.eval --help` for more instructions.
+See `lerobot-eval --help` for more instructions.
 
 ### Train your own policy
 
@@ -303,7 +303,7 @@ A link to the wandb logs for the run will also show up in yellow in your termina
 
 \<img src="https://raw.githubusercontent.com/huggingface/lerobot/main/media/wandb.png" alt="WandB logs example"\>
 
-Note: For efficiency, during training every checkpoint is evaluated on a low number of episodes. You may use `--eval.n_episodes=500` to evaluate on more episodes than the default. Or, after training, you may want to re-evaluate your best checkpoints on more episodes or change the evaluation settings. See `python -m lerobot.scripts.eval --help` for more instructions.
+Note: For efficiency, during training every checkpoint is evaluated on a low number of episodes. You may use `--eval.n_episodes=500` to evaluate on more episodes than the default. Or, after training, you may want to re-evaluate your best checkpoints on more episodes or change the evaluation settings. See `lerobot-eval --help` for more instructions.
 
 #### Reproduce state-of-the-art (SOTA)
 
@@ -311,7 +311,7 @@ We provide some pretrained policies on our [hub page](https://huggingface.co/ler
 You can reproduce their training by loading the config from their run. Simply running:
 
 ```bash
-python -m lerobot.scripts.train --config_path=lerobot/diffusion_pusht
+lerobot-train --config_path=lerobot/diffusion_pusht
 ```
 
 reproduces SOTA results for Diffusion Policy on the PushT task.

docker/Dockerfile.user

@@ -29,7 +29,7 @@ ENV DEBIAN_FRONTEND=noninteractive \
 
 # Install system dependencies and uv (as root)
 RUN apt-get update && apt-get install -y --no-install-recommends \
-    build-essential git curl libglib2.0-0 libegl1-mesa ffmpeg \
+    build-essential git curl libglib2.0-0 libegl1-mesa-dev ffmpeg \
     libusb-1.0-0-dev speech-dispatcher libgeos-dev portaudio19-dev \
     && curl -LsSf https://astral.sh/uv/install.sh | sh \
     && mv /root/.local/bin/uv /usr/local/bin/uv \

docs/source/_toctree.yml

@@ -35,10 +35,14 @@
     title: Koch v1.1
   - local: lekiwi
     title: LeKiwi
+  - local: reachy2
+    title: Reachy 2
   title: "Robots"
 - sections:
   - local: notebooks
     title: Notebooks
+  - local: feetech
+    title: Updating Feetech Firmware
   title: "Resources"
 - sections:
   - local: contributing

docs/source/cameras.mdx

@@ -9,7 +9,7 @@ To instantiate a camera, you need a camera identifier. This identifier might cha
 To find the camera indices of the cameras plugged into your system, run the following script:
 
 ```bash
-python -m lerobot.find_cameras opencv # or realsense for Intel Realsense cameras
+lerobot-find-cameras opencv # or realsense for Intel Realsense cameras
 ```
 
 The output will look something like this if you have two cameras connected:

docs/source/feetech.mdx

@@ -0,0 +1,71 @@
+# Feetech Motor Firmware Update
+
+This tutorial guides you through updating the firmware of Feetech motors using the official Feetech software.
+
+## Prerequisites
+
+- Windows computer (Feetech software is only available for Windows)
+- Feetech motor control board
+- USB cable to connect the control board to your computer
+- Feetech motors connected to the control board
+
+## Step 1: Download Feetech Software
+
+1. Visit the official Feetech software download page: [https://www.feetechrc.com/software.html](https://www.feetechrc.com/software.html)
+2. Download the latest version of the Feetech debugging software (FD)
+3. Install the software on your Windows computer
+
+## Step 2: Hardware Setup
+
+1. Connect your Feetech motors to the motor control board
+2. Connect the motor control board to your Windows computer via USB cable
+3. Ensure power is supplied to the motors
+
+## Step 3: Configure Connection
+
+1. Launch the Feetech debugging software
+2. Select the correct COM port from the port dropdown menu
+   - If unsure which port to use, check Windows Device Manager under "Ports (COM & LPT)"
+3. Set the appropriate baud rate (typically 1000000 for most Feetech motors)
+4. Click "Open" to establish communication with the control board
+
+## Step 4: Scan for Motors
+
+1. Once connected, click the "Search" button to detect all connected motors
+2. The software will automatically discover and list all motors on the bus
+3. Each motor will appear with its ID number
+
+## Step 5: Update Firmware
+
+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**:
+   - The update progress will be displayed
+
+## Step 6: Verify Update
+
+1. After the update completes, the software should automatically refresh the motor information
+2. Verify that the firmware version has been updated to the expected version
+
+## Important Notes
+
+⚠️ **Warning**: Do not disconnect power or USB during firmware updates, it will potentially brick the motor.
+
+## Bonus: Motor Debugging on Linux/macOS
+
+For debugging purposes only, you can use the open-source Feetech Debug Tool:
+
+- **Repository**: [FT_SCServo_Debug_Qt](https://github.com/CarolinePascal/FT_SCServo_Debug_Qt/tree/fix/port-search-timer)
+
+### Installation Instructions
+
+Follow the instructions in the repository to install the tool, for Ubuntu you can directly install it, for MacOS you need to build it from source.
+
+**Limitations:**
+
+- This tool is for debugging and parameter adjustment only
+- Firmware updates must still be done on Windows with official Feetech software

docs/source/hilserl.mdx

@@ -4,7 +4,13 @@ In this tutorial you will go through the full Human-in-the-Loop Sample-Efficient
 
 HIL-SERL is a sample-efficient reinforcement learning algorithm that combines human demonstrations with online learning and human interventions. The approach starts from a small set of human demonstrations, uses them to train a reward classifier, and then employs an actor-learner architecture where humans can intervene during policy execution to guide exploration and correct unsafe behaviors. In this tutorial, you'll use a gamepad to provide interventions and control the robot during the learning process.
 
-It combines three key ingredients: 1. **Offline demonstrations & reward classifier:** a handful of human-teleop episodes plus a vision-based success detector give the policy a shaped starting point. 2. **On-robot actor / learner loop with human interventions:** a distributed Soft Actor Critic (SAC) learner updates the policy while an actor explores on the physical robot; the human can jump in at any time to correct dangerous or unproductive behaviour. 3. **Safety & efficiency tools:** joint/end-effector (EE) bounds, crop region of interest (ROI) preprocessing and WandB monitoring keep the data useful and the hardware safe.
+It combines three key ingredients:
+
+1. **Offline demonstrations & reward classifier:** a handful of human-teleop episodes plus a vision-based success detector give the policy a shaped starting point.
+
+2. **On-robot actor / learner loop with human interventions:** a distributed Soft Actor Critic (SAC) learner updates the policy while an actor explores on the physical robot; the human can jump in at any time to correct dangerous or unproductive behaviour.
+
+3. **Safety & efficiency tools:** joint/end-effector (EE) bounds, crop region of interest (ROI) preprocessing and WandB monitoring keep the data useful and the hardware safe.
 
 Together these elements let HIL-SERL reach near-perfect task success and faster cycle times than imitation-only baselines.
 
@@ -56,30 +62,243 @@ pip install -e ".[hilserl]"
 
 ### Understanding Configuration
 
-The training process begins with proper configuration for the HILSerl environment. The configuration class of interest is `HILSerlRobotEnvConfig` in `lerobot/envs/configs.py`. Which is defined as:
+The training process begins with proper configuration for the HILSerl environment. The main configuration class is `GymManipulatorConfig` in `lerobot/scripts/rl/gym_manipulator.py`, which contains nested `HILSerlRobotEnvConfig` and `DatasetConfig`. The configuration is organized into focused, nested sub-configs:
 
 <!-- prettier-ignore-start -->
 ```python
+class GymManipulatorConfig:
+    env: HILSerlRobotEnvConfig    # Environment configuration (nested)
+    dataset: DatasetConfig    # Dataset recording/replay configuration (nested)
+    mode: str | None = None    # "record", "replay", or None (for training)
+    device: str = "cpu"    # Compute device
+
 class HILSerlRobotEnvConfig(EnvConfig):
     robot: RobotConfig | None = None    # Main robot agent (defined in `lerobot/robots`)
-    teleop: TeleoperatorConfig | None = None    # Teleoperator agent, e.g., gamepad or leader arm, (defined in `lerobot/teleoperators`)
-    wrapper: EnvTransformConfig | None = None    # Environment wrapper settings; check `lerobot/scripts/server/gym_manipulator.py`
-    fps: int = 10    # Control frequency
+    teleop: TeleoperatorConfig | None = None    # Teleoperator agent, e.g., gamepad or leader arm
+    processor: HILSerlProcessorConfig    # Processing pipeline configuration (nested)
     name: str = "real_robot"    # Environment name
-    mode: str = None    # "record", "replay", or None (for training)
-    repo_id: str | None = None    # LeRobot dataset repository ID
-    dataset_root: str | None = None    # Local dataset root (optional)
-    task: str = ""    # Task identifier
-    num_episodes: int = 10    # Number of episodes for recording
-    episode: int = 0    # episode index for replay
-    device: str = "cuda"    # Compute device
-    push_to_hub: bool = True    # Whether to push the recorded datasets to Hub
-    pretrained_policy_name_or_path: str | None = None    # For policy loading
-    reward_classifier_pretrained_path: str | None = None    # For reward model
-    number_of_steps_after_success: int = 0    # For reward classifier, collect more positive examples after a success to train a classifier
+    task: str | None = None    # Task identifier
+    fps: int = 10    # Control frequency
+
+# Nested processor configuration
+class HILSerlProcessorConfig:
+    control_mode: str = "gamepad"    # Control mode
+    observation: ObservationConfig | None = None    # Observation processing settings
+    image_preprocessing: ImagePreprocessingConfig | None = None    # Image crop/resize settings
+    gripper: GripperConfig | None = None    # Gripper control and penalty settings
+    reset: ResetConfig | None = None    # Environment reset and timing settings
+    inverse_kinematics: InverseKinematicsConfig | None = None    # IK processing settings
+    reward_classifier: RewardClassifierConfig | None = None    # Reward classifier settings
+    max_gripper_pos: float | None = 100.0    # Maximum gripper position
+
+# Sub-configuration classes
+class ObservationConfig:
+    add_joint_velocity_to_observation: bool = False    # Add joint velocities to state
+    add_current_to_observation: bool = False    # Add motor currents to state
+    add_ee_pose_to_observation: bool = False    # Add end-effector pose to state
+    display_cameras: bool = False    # Display camera feeds during execution
+
+class ImagePreprocessingConfig:
+    crop_params_dict: dict[str, tuple[int, int, int, int]] | None = None    # Image cropping parameters
+    resize_size: tuple[int, int] | None = None    # Target image size
+
+class GripperConfig:
+    use_gripper: bool = True    # Enable gripper control
+    gripper_penalty: float = 0.0    # Penalty for inappropriate gripper usage
+    gripper_penalty_in_reward: bool = False    # Include gripper penalty in reward
+
+class ResetConfig:
+    fixed_reset_joint_positions: Any | None = None    # Joint positions for reset
+    reset_time_s: float = 5.0    # Time to wait during reset
+    control_time_s: float = 20.0    # Maximum episode duration
+    terminate_on_success: bool = True    # Whether to terminate episodes on success detection
+
+class InverseKinematicsConfig:
+    urdf_path: str | None = None    # Path to robot URDF file
+    target_frame_name: str | None = None    # End-effector frame name
+    end_effector_bounds: dict[str, list[float]] | None = None    # EE workspace bounds
+    end_effector_step_sizes: dict[str, float] | None = None    # EE step sizes per axis
+
+class RewardClassifierConfig:
+    pretrained_path: str | None = None    # Path to pretrained reward classifier
+    success_threshold: float = 0.5    # Success detection threshold
+    success_reward: float = 1.0    # Reward value for successful episodes
+
+# Dataset configuration
+class DatasetConfig:
+    repo_id: str    # LeRobot dataset repository ID
+    task: str    # Task identifier
+    root: str | None = None    # Local dataset root directory
+    num_episodes_to_record: int = 5    # Number of episodes for recording
+    replay_episode: int | None = None    # Episode index for replay
+    push_to_hub: bool = False    # Whether to push datasets to Hub
 ```
 <!-- prettier-ignore-end -->
 
+### Processor Pipeline Architecture
+
+HIL-SERL uses a modular processor pipeline architecture that processes robot observations and actions through a series of composable steps. The pipeline is divided into two main components:
+
+#### Environment Processor Pipeline
+
+The environment processor (`env_processor`) handles incoming observations and environment state:
+
+1. **VanillaObservationProcessorStep**: Converts raw robot observations into standardized format
+2. **JointVelocityProcessorStep** (optional): Adds joint velocity information to observations
+3. **MotorCurrentProcessorStep** (optional): Adds motor current readings to observations
+4. **ForwardKinematicsJointsToEE** (optional): Computes end-effector pose from joint positions
+5. **ImageCropResizeProcessorStep** (optional): Crops and resizes camera images
+6. **TimeLimitProcessorStep** (optional): Enforces episode time limits
+7. **GripperPenaltyProcessorStep** (optional): Applies penalties for inappropriate gripper usage
+8. **RewardClassifierProcessorStep** (optional): Automated reward detection using vision models
+9. **AddBatchDimensionProcessorStep**: Converts data to batch format for neural network processing
+10. **DeviceProcessorStep**: Moves data to the specified compute device (CPU/GPU)
+
+#### Action Processor Pipeline
+
+The action processor (`action_processor`) handles outgoing actions and human interventions:
+
+1. **AddTeleopActionAsComplimentaryDataStep**: Captures teleoperator actions for logging
+2. **AddTeleopEventsAsInfoStep**: Records intervention events and episode control signals
+3. **AddRobotObservationAsComplimentaryData**: Stores raw robot state for processing
+4. **InterventionActionProcessorStep**: Handles human interventions and episode termination
+5. **Inverse Kinematics Pipeline** (when enabled):
+   - **MapDeltaActionToRobotActionStep**: Converts delta actions to robot action format
+   - **EEReferenceAndDelta**: Computes end-effector reference and delta movements
+   - **EEBoundsAndSafety**: Enforces workspace safety bounds
+   - **InverseKinematicsEEToJoints**: Converts end-effector actions to joint targets
+   - **GripperVelocityToJoint**: Handles gripper control commands
+
+#### Configuration Examples
+
+**Basic Observation Processing**:
+
+```json
+{
+  "env": {
+    "processor": {
+      "observation": {
+        "add_joint_velocity_to_observation": true,
+        "add_current_to_observation": false,
+        "display_cameras": false
+      }
+    }
+  }
+}
+```
+
+**Image Processing**:
+
+```json
+{
+  "env": {
+    "processor": {
+      "image_preprocessing": {
+        "crop_params_dict": {
+          "observation.images.front": [180, 250, 120, 150],
+          "observation.images.side": [180, 207, 180, 200]
+        },
+        "resize_size": [128, 128]
+      }
+    }
+  }
+}
+```
+
+**Inverse Kinematics Setup**:
+
+```json
+{
+  "env": {
+    "processor": {
+      "inverse_kinematics": {
+        "urdf_path": "path/to/robot.urdf",
+        "target_frame_name": "end_effector",
+        "end_effector_bounds": {
+          "min": [0.16, -0.08, 0.03],
+          "max": [0.24, 0.2, 0.1]
+        },
+        "end_effector_step_sizes": {
+          "x": 0.02,
+          "y": 0.02,
+          "z": 0.02
+        }
+      }
+    }
+  }
+}
+```
+
+### Advanced Observation Processing
+
+The HIL-SERL framework supports additional observation processing features that can improve policy learning:
+
+#### Joint Velocity Processing
+
+Enable joint velocity estimation to provide the policy with motion information:
+
+```json
+{
+  "env": {
+    "processor": {
+      "observation": {
+        "add_joint_velocity_to_observation": true
+      }
+    }
+  }
+}
+```
+
+This processor:
+
+- Estimates joint velocities using finite differences between consecutive joint position readings
+- Adds velocity information to the observation state vector
+- Useful for policies that need motion awareness for dynamic tasks
+
+#### Motor Current Processing
+
+Monitor motor currents to detect contact forces and load conditions:
+
+```json
+{
+  "env": {
+    "processor": {
+      "observation": {
+        "add_current_to_observation": true
+      }
+    }
+  }
+}
+```
+
+This processor:
+
+- Reads motor current values from the robot's control system
+- Adds current measurements to the observation state vector
+- Helps detect contact events, object weights, and mechanical resistance
+- Useful for contact-rich manipulation tasks
+
+#### Combined Observation Processing
+
+You can enable multiple observation processing features simultaneously:
+
+```json
+{
+  "env": {
+    "processor": {
+      "observation": {
+        "add_joint_velocity_to_observation": true,
+        "add_current_to_observation": true,
+        "add_ee_pose_to_observation": false,
+        "display_cameras": false
+      }
+    }
+  }
+}
+```
+
+**Note**: Enabling additional observation features increases the state space dimensionality, which may require adjusting your policy network architecture and potentially collecting more training data.
+
 ### Finding Robot Workspace Bounds
 
 Before collecting demonstrations, you need to determine the appropriate operational bounds for your robot.
@@ -130,22 +349,56 @@ With the bounds defined, you can safely collect demonstrations for training. Tra
 
 Create a configuration file for recording demonstrations (or edit an existing one like [env_config_so100.json](https://huggingface.co/datasets/aractingi/lerobot-example-config-files/blob/main/env_config_so100.json)):
 
-1. Set `mode` to `"record"`
-2. Specify a unique `repo_id` for your dataset (e.g., "username/task_name")
-3. Set `num_episodes` to the number of demonstrations you want to collect
-4. Set `crop_params_dict` to `null` initially (we'll determine crops later)
-5. Configure `robot`, `cameras`, and other hardware settings
+1. Set `mode` to `"record"` at the root level
+2. Specify a unique `repo_id` for your dataset in the `dataset` section (e.g., "username/task_name")
+3. Set `num_episodes_to_record` in the `dataset` section to the number of demonstrations you want to collect
+4. Set `env.processor.image_preprocessing.crop_params_dict` to `{}` initially (we'll determine crops later)
+5. Configure `env.robot`, `env.teleop`, and other hardware settings in the `env` section
 
 Example configuration section:
 
 ```json
-"mode": "record",
-"repo_id": "username/pick_lift_cube",
-"dataset_root": null,
-"task": "pick_and_lift",
-"num_episodes": 15,
-"episode": 0,
-"push_to_hub": true
+{
+  "env": {
+    "type": "gym_manipulator",
+    "name": "real_robot",
+    "fps": 10,
+    "processor": {
+      "control_mode": "gamepad",
+      "observation": {
+        "display_cameras": false
+      },
+      "image_preprocessing": {
+        "crop_params_dict": {},
+        "resize_size": [128, 128]
+      },
+      "gripper": {
+        "use_gripper": true,
+        "gripper_penalty": 0.0
+      },
+      "reset": {
+        "reset_time_s": 5.0,
+        "control_time_s": 20.0
+      }
+    },
+    "robot": {
+      // ... robot configuration ...
+    },
+    "teleop": {
+      // ... teleoperator configuration ...
+    }
+  },
+  "dataset": {
+    "repo_id": "username/pick_lift_cube",
+    "root": null,
+    "task": "pick_and_lift",
+    "num_episodes_to_record": 15,
+    "replay_episode": 0,
+    "push_to_hub": true
+  },
+  "mode": "record",
+  "device": "cpu"
+}
 ```
 
 ### Using a Teleoperation Device
@@ -191,10 +444,20 @@ The gamepad provides a very convenient way to control the robot and the episode
 To setup the gamepad, you need to set the `control_mode` to `"gamepad"` and define the `teleop` section in the configuration file.
 
 ```json
+{
+  "env": {
     "teleop": {
-        "type": "gamepad",
-        "use_gripper": true
+      "type": "gamepad",
+      "use_gripper": true
     },
+    "processor": {
+      "control_mode": "gamepad",
+      "gripper": {
+        "use_gripper": true
+      }
+    }
+  }
+}
 ```
 
 <p align="center">
@@ -216,11 +479,21 @@ The SO101 leader arm has reduced gears that allows it to move and track the foll
 To setup the SO101 leader, you need to set the `control_mode` to `"leader"` and define the `teleop` section in the configuration file.
 
 ```json
+{
+  "env": {
     "teleop": {
-        "type": "so101_leader",
-        "port": "/dev/tty.usbmodem585A0077921", # check your port number
-        "use_degrees": true
+      "type": "so101_leader",
+      "port": "/dev/tty.usbmodem585A0077921",
+      "use_degrees": true
     },
+    "processor": {
+      "control_mode": "leader",
+      "gripper": {
+        "use_gripper": true
+      }
+    }
+  }
+}
 ```
 
 In order to annotate the success/failure of the episode, **you will need** to use a keyboard to press `s` for success, `esc` for failure.
@@ -251,7 +524,7 @@ python -m lerobot.scripts.rl.gym_manipulator --config_path src/lerobot/configs/e
 
 During recording:
 
-1. The robot will reset to the initial position defined in the configuration file `fixed_reset_joint_positions`
+1. The robot will reset to the initial position defined in the configuration file `env.processor.reset.fixed_reset_joint_positions`
 2. Complete the task successfully
 3. The episode ends with a reward of 1 when you press the "success" button
 4. If the time limit is reached, or the fail button is pressed, the episode ends with a reward of 0
@@ -310,11 +583,19 @@ observation.images.front: [180, 250, 120, 150]
 Add these crop parameters to your training configuration:
 
 ```json
-"crop_params_dict": {
-    "observation.images.side": [180, 207, 180, 200],
-    "observation.images.front": [180, 250, 120, 150]
-},
-"resize_size": [128, 128]
+{
+  "env": {
+    "processor": {
+      "image_preprocessing": {
+        "crop_params_dict": {
+          "observation.images.side": [180, 207, 180, 200],
+          "observation.images.front": [180, 250, 120, 150]
+        },
+        "resize_size": [128, 128]
+      }
+    }
+  }
+}
 ```
 
 **Recommended image resolution**
@@ -343,26 +624,52 @@ python -m lerobot.scripts.rl.gym_manipulator --config_path src/lerobot/configs/r
 
 **Key Parameters for Data Collection**
 
-- **mode**: set it to `"record"` to collect a dataset
-- **repo_id**: `"hf_username/dataset_name"`, name of the dataset and repo on the hub
-- **num_episodes**: Number of episodes to record
-- **number_of_steps_after_success**: Number of additional frames to record after a success (reward=1) is detected
-- **fps**: Number of frames per second to record
-- **push_to_hub**: Whether to push the dataset to the hub
+- **mode**: set it to `"record"` to collect a dataset (at root level)
+- **dataset.repo_id**: `"hf_username/dataset_name"`, name of the dataset and repo on the hub
+- **dataset.num_episodes_to_record**: Number of episodes to record
+- **env.processor.reset.terminate_on_success**: Whether to automatically terminate episodes when success is detected (default: `true`)
+- **env.fps**: Number of frames per second to record
+- **dataset.push_to_hub**: Whether to push the dataset to the hub
 
-The `number_of_steps_after_success` parameter is crucial as it allows you to collect more positive examples. When a success is detected, the system will continue recording for the specified number of steps while maintaining the reward=1 label. Otherwise, there won't be enough states in the dataset labeled to 1 to train a good classifier.
+The `env.processor.reset.terminate_on_success` parameter allows you to control episode termination behavior. When set to `false`, episodes will continue even after success is detected, allowing you to collect more positive examples with the reward=1 label. This is crucial for training reward classifiers as it provides more success state examples in your dataset. When set to `true` (default), episodes terminate immediately upon success detection.
+
+**Important**: For reward classifier training, set `terminate_on_success: false` to collect sufficient positive examples. For regular HIL-SERL training, keep it as `true` to enable automatic episode termination when the task is completed successfully.
 
 Example configuration section for data collection:
 
 ```json
 {
+  "env": {
+    "type": "gym_manipulator",
+    "name": "real_robot",
+    "fps": 10,
+    "processor": {
+      "reset": {
+        "reset_time_s": 5.0,
+        "control_time_s": 20.0,
+        "terminate_on_success": false
+      },
+      "gripper": {
+        "use_gripper": true
+      }
+    },
+    "robot": {
+      // ... robot configuration ...
+    },
+    "teleop": {
+      // ... teleoperator configuration ...
+    }
+  },
+  "dataset": {
+    "repo_id": "hf_username/dataset_name",
+    "dataset_root": "data/your_dataset",
+    "task": "reward_classifier_task",
+    "num_episodes_to_record": 20,
+    "replay_episode": null,
+    "push_to_hub": true
+  },
   "mode": "record",
-  "repo_id": "hf_username/dataset_name",
-  "dataset_root": "data/your_dataset",
-  "num_episodes": 20,
-  "push_to_hub": true,
-  "fps": 10,
-  "number_of_steps_after_success": 15
+  "device": "cpu"
 }
 ```
 
@@ -412,7 +719,7 @@ Example configuration for training the [reward classifier](https://huggingface.c
 To train the classifier, use the `train.py` script with your configuration:
 
 ```bash
-python -m lerobot.scripts.train --config_path path/to/reward_classifier_train_config.json
+lerobot-train --config_path path/to/reward_classifier_train_config.json
 ```
 
 **Deploying and Testing the Model**
@@ -421,9 +728,17 @@ To use your trained reward classifier, configure the `HILSerlRobotEnvConfig` to
 
 <!-- prettier-ignore-start -->
 ```python
-env_config = HILSerlRobotEnvConfig(
-    reward_classifier_pretrained_path="path_to_your_pretrained_trained_model",
-    # Other environment parameters
+config = GymManipulatorConfig(
+    env=HILSerlRobotEnvConfig(
+        processor=HILSerlProcessorConfig(
+            reward_classifier=RewardClassifierConfig(
+                pretrained_path="path_to_your_pretrained_trained_model"
+            )
+        ),
+        # Other environment parameters
+    ),
+    dataset=DatasetConfig(...),
+    mode=None  # For training
 )
 ```
 <!-- prettier-ignore-end -->
@@ -432,7 +747,18 @@ or set the argument in the json config file.
 
 ```json
 {
-  "reward_classifier_pretrained_path": "path_to_your_pretrained_model"
+  "env": {
+    "processor": {
+      "reward_classifier": {
+        "pretrained_path": "path_to_your_pretrained_model",
+        "success_threshold": 0.7,
+        "success_reward": 1.0
+      },
+      "reset": {
+        "terminate_on_success": true
+      }
+    }
+  }
 }
 ```
 
@@ -458,7 +784,7 @@ The reward classifier will automatically provide rewards based on the visual inp
 3. **Train the classifier**:
 
    ```bash
-   python -m lerobot.scripts.train --config_path src/lerobot/configs/reward_classifier_train_config.json
+   lerobot-train --config_path src/lerobot/configs/reward_classifier_train_config.json
    ```
 
 4. **Test the classifier**:

docs/source/hilserl_sim.mdx

@@ -32,9 +32,12 @@ To use `gym_hil` with LeRobot, you need to create a configuration file. An examp
 
 ```json
 {
-  "type": "hil",
-  "name": "franka_sim",
-  "task": "PandaPickCubeGamepad-v0",
+  "env": {
+    "type": "gym_manipulator",
+    "name": "gym_hil",
+    "task": "PandaPickCubeGamepad-v0",
+    "fps": 10
+  },
   "device": "cuda"
 }
 ```
@@ -45,28 +48,40 @@ Available tasks:
 - `PandaPickCubeGamepad-v0`: With gamepad control
 - `PandaPickCubeKeyboard-v0`: With keyboard control
 
-### Gym Wrappers Configuration
+### Processor Configuration
 
 ```json
-"wrapper": {
-    "gripper_penalty": -0.02,
-    "control_time_s": 15.0,
-    "use_gripper": true,
-    "fixed_reset_joint_positions": [0.0, 0.195, 0.0, -2.43, 0.0, 2.62, 0.785],
-    "end_effector_step_sizes": {
-        "x": 0.025,
-        "y": 0.025,
-        "z": 0.025
-    },
-    "control_mode": "gamepad"
+{
+  "env": {
+    "processor": {
+      "control_mode": "gamepad",
+      "gripper": {
+        "use_gripper": true,
+        "gripper_penalty": -0.02
+      },
+      "reset": {
+        "control_time_s": 15.0,
+        "fixed_reset_joint_positions": [
+          0.0, 0.195, 0.0, -2.43, 0.0, 2.62, 0.785
+        ]
+      },
+      "inverse_kinematics": {
+        "end_effector_step_sizes": {
+          "x": 0.025,
+          "y": 0.025,
+          "z": 0.025
+        }
+      }
     }
+  }
+}
 ```
 
 Important parameters:
 
-- `gripper_penalty`: Penalty for excessive gripper movement
-- `use_gripper`: Whether to enable gripper control
-- `end_effector_step_sizes`: Size of the steps in the x,y,z axes of the end-effector
+- `gripper.gripper_penalty`: Penalty for excessive gripper movement
+- `gripper.use_gripper`: Whether to enable gripper control
+- `inverse_kinematics.end_effector_step_sizes`: Size of the steps in the x,y,z axes of the end-effector
 - `control_mode`: Set to `"gamepad"` to use a gamepad controller
 
 ## Running with HIL RL of LeRobot
@@ -75,39 +90,50 @@ Important parameters:
 
 To run the environment, set mode to null:
 
-<!-- prettier-ignore-start -->
-```python
+```bash
 python -m lerobot.scripts.rl.gym_manipulator --config_path path/to/gym_hil_env.json
 ```
-<!-- prettier-ignore-end -->
 
 ### Recording a Dataset
 
 To collect a dataset, set the mode to `record` whilst defining the repo_id and number of episodes to record:
 
-<!-- prettier-ignore-start -->
-```python
+```json
+{
+  "env": {
+    "type": "gym_manipulator",
+    "name": "gym_hil",
+    "task": "PandaPickCubeGamepad-v0"
+  },
+  "dataset": {
+    "repo_id": "username/sim_dataset",
+    "root": null,
+    "task": "pick_cube",
+    "num_episodes_to_record": 10,
+    "replay_episode": null,
+    "push_to_hub": true
+  },
+  "mode": "record"
+}
+```
+
+```bash
 python -m lerobot.scripts.rl.gym_manipulator --config_path path/to/gym_hil_env.json
 ```
-<!-- prettier-ignore-end -->
 
 ### Training a Policy
 
 To train a policy, checkout the configuration example available [here](https://huggingface.co/datasets/aractingi/lerobot-example-config-files/blob/main/train_gym_hil_env.json) and run the actor and learner servers:
 
-<!-- prettier-ignore-start -->
-```python
+```bash
 python -m lerobot.scripts.rl.actor --config_path path/to/train_gym_hil_env.json
 ```
-<!-- prettier-ignore-end -->
 
 In a different terminal, run the learner server:
 
-<!-- prettier-ignore-start -->
-```python
+```bash
 python -m lerobot.scripts.rl.learner --config_path path/to/train_gym_hil_env.json
 ```
-<!-- prettier-ignore-end -->
 
 The simulation environment provides a safe and repeatable way to develop and test your Human-In-the-Loop reinforcement learning components before deploying to real robots.
 

docs/source/hope_jr.mdx

@@ -19,7 +19,7 @@ pip install -e ".[hopejr]"
 Before starting calibration and operation, you need to identify the USB ports for each HopeJR component. Run this script to find the USB ports for the arm, hand, glove, and exoskeleton:
 
 ```bash
-python -m lerobot.find_port
+lerobot-find-port
 ```
 
 This will display the available USB ports and their associated devices. Make note of the port paths (e.g., `/dev/tty.usbmodem58760433331`, `/dev/tty.usbmodem11301`) as you'll need to specify them in the `--robot.port` and `--teleop.port` parameters when recording data, replaying episodes, or running teleoperation scripts.
@@ -31,7 +31,7 @@ Before performing teleoperation, HopeJR's limbs need to be calibrated. Calibrati
 ### 1.1 Calibrate Robot Hand
 
 ```bash
-python -m lerobot.calibrate \
+lerobot-calibrate \
     --robot.type=hope_jr_hand \
     --robot.port=/dev/tty.usbmodem58760432281 \
     --robot.id=blue \
@@ -81,7 +81,7 @@ Once you have set the appropriate boundaries for all joints, click "Save" to sav
 ### 1.2 Calibrate Teleoperator Glove
 
 ```bash
-python -m lerobot.calibrate \
+lerobot-calibrate \
     --teleop.type=homunculus_glove \
     --teleop.port=/dev/tty.usbmodem11201 \
     --teleop.id=red \
@@ -120,7 +120,7 @@ Once calibration is complete, the system will save the calibration to `/Users/yo
 ### 1.3 Calibrate Robot Arm
 
 ```bash
-python -m lerobot.calibrate \
+lerobot-calibrate \
     --robot.type=hope_jr_arm \
     --robot.port=/dev/tty.usbserial-1110 \
     --robot.id=white
@@ -146,7 +146,7 @@ Use the calibration interface to set the range boundaries for each joint. Move e
 ### 1.4 Calibrate Teleoperator Exoskeleton
 
 ```bash
-python -m lerobot.calibrate \
+lerobot-calibrate \
     --teleop.type=homunculus_arm \
     --teleop.port=/dev/tty.usbmodem11201 \
     --teleop.id=black
@@ -178,7 +178,7 @@ Due to global variable conflicts in the Feetech middleware, teleoperation for ar
 ### Hand
 
 ```bash
-python -m lerobot.teleoperate \
+lerobot-teleoperate \
     --robot.type=hope_jr_hand \
     --robot.port=/dev/tty.usbmodem58760432281 \
     --robot.id=blue \
@@ -194,7 +194,7 @@ python -m lerobot.teleoperate \
 ### Arm
 
 ```bash
-python -m lerobot.teleoperate \
+lerobot-teleoperate \
     --robot.type=hope_jr_arm \
     --robot.port=/dev/tty.usbserial-1110 \
     --robot.id=white \
@@ -214,7 +214,7 @@ Record, Replay and Train with Hope-JR is still experimental.
 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).
 
 ```bash
-python -m lerobot.record \
+lerobot-record \
     --robot.type=hope_jr_hand \
     --robot.port=/dev/tty.usbmodem58760432281 \
     --robot.id=right \
@@ -236,7 +236,7 @@ python -m lerobot.record \
 ### Replay
 
 ```bash
-python -m lerobot.replay \
+lerobot-replay \
     --robot.type=hope_jr_hand \
     --robot.port=/dev/tty.usbmodem58760432281 \
     --robot.id=right \
@@ -248,7 +248,7 @@ python -m lerobot.replay \
 ### Train
 
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
   --dataset.repo_id=nepyope/hand_record_test_with_video_data \
   --policy.type=act \
   --output_dir=outputs/train/hopejr_hand \
@@ -263,7 +263,7 @@ python -m lerobot.scripts.train \
 This training run can be viewed as an example [here](https://wandb.ai/tino/lerobot/runs/rp0k8zvw?nw=nwusertino).
 
 ```bash
-python -m lerobot.record \
+lerobot-record \
   --robot.type=hope_jr_hand \
   --robot.port=/dev/tty.usbmodem58760432281 \
   --robot.id=right \

docs/source/il_robots.mdx

@@ -45,7 +45,7 @@ Note that the `id` associated with a robot is used to store the calibration file
 <hfoptions id="teleoperate_so101">
 <hfoption id="Command">
 ```bash
-python -m lerobot.teleoperate \
+lerobot-teleoperate \
     --robot.type=so101_follower \
     --robot.port=/dev/tty.usbmodem58760431541 \
     --robot.id=my_awesome_follower_arm \
@@ -101,7 +101,7 @@ With `rerun`, you can teleoperate again while simultaneously visualizing the cam
 <hfoptions id="teleoperate_koch_camera">
 <hfoption id="Command">
 ```bash
-python -m lerobot.teleoperate \
+lerobot-teleoperate \
     --robot.type=koch_follower \
     --robot.port=/dev/tty.usbmodem58760431541 \
     --robot.id=my_awesome_follower_arm \
@@ -174,7 +174,7 @@ Now you can record a dataset. To record 5 episodes and upload your dataset to th
 <hfoptions id="record">
 <hfoption id="Command">
 ```bash
-python -m lerobot.record \
+lerobot-record \
     --robot.type=so101_follower \
     --robot.port=/dev/tty.usbmodem585A0076841 \
     --robot.id=my_awesome_follower_arm \
@@ -376,7 +376,7 @@ You can replay the first episode on your robot with either the command below or
 <hfoptions id="replay">
 <hfoption id="Command">
 ```bash
-python -m lerobot.replay \
+lerobot-replay \
     --robot.type=so101_follower \
     --robot.port=/dev/tty.usbmodem58760431541 \
     --robot.id=my_awesome_follower_arm \
@@ -428,10 +428,10 @@ Your robot should replicate movements similar to those you recorded. For example
 
 ## Train a policy
 
-To train a policy to control your robot, use the [`python -m lerobot.scripts.train`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
+To train a policy to control your robot, use the [`lerobot-train`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
 
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
   --dataset.repo_id=${HF_USER}/so101_test \
   --policy.type=act \
   --output_dir=outputs/train/act_so101_test \
@@ -453,7 +453,7 @@ Training should take several hours. You will find checkpoints in `outputs/train/
 To resume training from a checkpoint, below is an example command to resume from `last` checkpoint of the `act_so101_test` policy:
 
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
   --config_path=outputs/train/act_so101_test/checkpoints/last/pretrained_model/train_config.json \
   --resume=true
 ```
@@ -490,7 +490,7 @@ You can use the `record` script from [`lerobot/record.py`](https://github.com/hu
 <hfoptions id="eval">
 <hfoption id="Command">
 ```bash
-python -m lerobot.record  \
+lerobot-record  \
   --robot.type=so100_follower \
   --robot.port=/dev/ttyACM1 \
   --robot.cameras="{ up: {type: opencv, index_or_path: /dev/video10, width: 640, height: 480, fps: 30}, side: {type: intelrealsense, serial_number_or_name: 233522074606, width: 640, height: 480, fps: 30}}" \
@@ -519,11 +519,14 @@ from lerobot.utils.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import _init_rerun
 from lerobot.record import record_loop
+from lerobot.policies.factory import make_processor
 
 NUM_EPISODES = 5
 FPS = 30
 EPISODE_TIME_SEC = 60
 TASK_DESCRIPTION = "My task description"
+HF_MODEL_ID = "<hf_username>/<model_repo_id>"
+HF_DATASET_ID = "<hf_username>/<eval_dataset_repo_id>"
 
 # Create the robot configuration
 camera_config = {"front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS)}
@@ -535,7 +538,7 @@ robot_config = SO100FollowerConfig(
 robot = SO100Follower(robot_config)
 
 # Initialize the policy
-policy = ACTPolicy.from_pretrained("<hf_username>/<my_policy_repo_id>")
+policy = ACTPolicy.from_pretrained(HF_MODEL_ID)
 
 # Configure the dataset features
 action_features = hw_to_dataset_features(robot.action_features, "action")
@@ -544,7 +547,7 @@ dataset_features = {**action_features, **obs_features}
 
 # Create the dataset
 dataset = LeRobotDataset.create(
-    repo_id="<hf_username>/eval_<dataset_repo_id>",
+    repo_id=HF_DATASET_ID,
     fps=FPS,
     features=dataset_features,
     robot_type=robot.name,
@@ -559,6 +562,12 @@ _init_rerun(session_name="recording")
 # Connect the robot
 robot.connect()
 
+preprocessor, postprocessor = make_processor(
+    policy_cfg=policy,
+    pretrained_path=HF_MODEL_ID,
+    dataset_stats=dataset.meta.stats,
+)
+
 for episode_idx in range(NUM_EPISODES):
     log_say(f"Running inference, recording eval episode {episode_idx + 1} of {NUM_EPISODES}")
 
@@ -568,6 +577,8 @@ for episode_idx in range(NUM_EPISODES):
         events=events,
         fps=FPS,
         policy=policy,
+        preprocessor=preprocessor,
+        postprocessor=postprocessor,
         dataset=dataset,
         control_time_s=EPISODE_TIME_SEC,
         single_task=TASK_DESCRIPTION,

docs/source/il_sim.mdx

@@ -24,11 +24,36 @@ pip install -e ".[hilserl]"
 
 To use `gym_hil` with LeRobot, you need to use a configuration file. An example config file can be found [here](https://huggingface.co/datasets/aractingi/lerobot-example-config-files/blob/main/env_config_gym_hil_il.json).
 
-To teleoperate and collect a dataset, we need to modify this config file and you should add your `repo_id` here: `"repo_id": "il_gym",` and `"num_episodes": 30,` and make sure you set `mode` to `record`, "mode": "record".
+To teleoperate and collect a dataset, we need to modify this config file. Here's an example configuration for imitation learning data collection:
 
-If you do not have a Nvidia GPU also change `"device": "cuda"` parameter in the config file (for example to `mps` for MacOS).
+```json
+{
+  "env": {
+    "type": "gym_manipulator",
+    "name": "gym_hil",
+    "task": "PandaPickCubeGamepad-v0",
+    "fps": 10
+  },
+  "dataset": {
+    "repo_id": "your_username/il_gym",
+    "root": null,
+    "task": "pick_cube",
+    "num_episodes_to_record": 30,
+    "replay_episode": null,
+    "push_to_hub": true
+  },
+  "mode": "record",
+  "device": "cuda"
+}
+```
 
-By default the config file assumes you use a controller. To use your keyboard please change the envoirment specified at `"task"` in the config file and set it to `"PandaPickCubeKeyboard-v0"`.
+Key configuration points:
+
+- Set your `repo_id` in the `dataset` section: `"repo_id": "your_username/il_gym"`
+- Set `num_episodes_to_record: 30` to collect 30 demonstration episodes
+- Ensure `mode` is set to `"record"`
+- If you don't have an NVIDIA GPU, change `"device": "cuda"` to `"mps"` for macOS or `"cpu"`
+- To use keyboard instead of gamepad, change `"task"` to `"PandaPickCubeKeyboard-v0"`
 
 Then we can run this command to start:
 
@@ -96,10 +121,10 @@ If you uploaded your dataset to the hub you can [visualize your dataset online](
 
 ## Train a policy
 
-To train a policy to control your robot, use the [`python -m lerobot.scripts.train`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
+To train a policy to control your robot, use the [`lerobot-train`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
 
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
   --dataset.repo_id=${HF_USER}/il_gym \
   --policy.type=act \
   --output_dir=outputs/train/il_sim_test \
@@ -140,9 +165,32 @@ huggingface-cli upload ${HF_USER}/il_sim_test${CKPT} \
 
 ## Evaluate your policy in Sim
 
-To evaluate your policy we have to use the config file that can be found [here](https://huggingface.co/datasets/aractingi/lerobot-example-config-files/blob/main/eval_config_gym_hil.json).
+To evaluate your policy we have to use a configuration file. An example can be found [here](https://huggingface.co/datasets/aractingi/lerobot-example-config-files/blob/main/eval_config_gym_hil.json).
 
-Make sure to replace the `repo_id` with the dataset you trained on, for example `pepijn223/il_sim_dataset` and replace the `pretrained_policy_name_or_path` with your model id, for example `pepijn223/il_sim_model`
+Here's an example evaluation configuration:
+
+```json
+{
+  "env": {
+    "type": "gym_manipulator",
+    "name": "gym_hil",
+    "task": "PandaPickCubeGamepad-v0",
+    "fps": 10
+  },
+  "dataset": {
+    "repo_id": "your_username/il_sim_dataset",
+    "dataset_root": null,
+    "task": "pick_cube"
+  },
+  "pretrained_policy_name_or_path": "your_username/il_sim_model",
+  "device": "cuda"
+}
+```
+
+Make sure to replace:
+
+- `repo_id` with the dataset you trained on (e.g., `your_username/il_sim_dataset`)
+- `pretrained_policy_name_or_path` with your model ID (e.g., `your_username/il_sim_model`)
 
 Then you can run this command to visualize your trained policy
 

docs/source/koch.mdx

@@ -31,7 +31,7 @@ pip install -e ".[dynamixel]"
 To find the port for each bus servo adapter, run this script:
 
 ```bash
-python -m lerobot.find_port
+lerobot-find-port
 ```
 
 <hfoptions id="example">
@@ -98,7 +98,7 @@ For a visual reference on how to set the motor ids please refer to [this video](
 <hfoption id="Command">
 
 ```bash
-python -m lerobot.setup_motors \
+lerobot-setup-motors \
     --robot.type=koch_follower \
     --robot.port=/dev/tty.usbmodem575E0031751 # <- paste here the port found at previous step
 ```
@@ -174,7 +174,7 @@ Do the same steps for the leader arm but modify the command or script accordingl
 <hfoption id="Command">
 
 ```bash
-python -m lerobot.setup_motors \
+lerobot-setup-motors \
     --teleop.type=koch_leader \
     --teleop.port=/dev/tty.usbmodem575E0031751 \  # <- paste here the port found at previous step
 ```
@@ -211,7 +211,7 @@ Run the following command or API example to calibrate the follower arm:
 <hfoption id="Command">
 
 ```bash
-python -m lerobot.calibrate \
+lerobot-calibrate \
     --robot.type=koch_follower \
     --robot.port=/dev/tty.usbmodem58760431551 \ # <- The port of your robot
     --robot.id=my_awesome_follower_arm  # <- Give the robot a unique name
@@ -249,7 +249,7 @@ Do the same steps to calibrate the leader arm, run the following command or API
 <hfoption id="Command">
 
 ```bash
-python -m lerobot.calibrate \
+lerobot-calibrate \
     --teleop.type=koch_leader \
     --teleop.port=/dev/tty.usbmodem58760431551 \ # <- The port of your robot
     --teleop.id=my_awesome_leader_arm  # <- Give the robot a unique name

docs/source/lekiwi.mdx

@@ -60,7 +60,7 @@ First, we will assemble the two SO100/SO101 arms. One to attach to the mobile ba
 To find the port for each bus servo adapter, run this script:
 
 ```bash
-python -m lerobot.find_port
+lerobot-find-port
 ```
 
 <hfoptions id="example">
@@ -116,7 +116,7 @@ The instructions for configuring the motors can be found in the SO101 [docs](./s
 You can run this command to setup motors for LeKiwi. It will first setup the motors for arm (id 6..1) and then setup motors for wheels (9,8,7)
 
 ```bash
-python -m lerobot.setup_motors \
+lerobot-setup-motors \
     --robot.type=lekiwi \
     --robot.port=/dev/tty.usbmodem58760431551 # <- paste here the port found at previous step
 ```
@@ -174,7 +174,7 @@ The calibration process is very important because it allows a neural network tra
 Make sure the arm is connected to the Raspberry Pi and run this script or API example (on the Raspberry Pi via SSH) to launch calibration of the follower arm:
 
 ```bash
-python -m lerobot.calibrate \
+lerobot-calibrate \
     --robot.type=lekiwi \
     --robot.id=my_awesome_kiwi # <- Give the robot a unique name
 ```
@@ -193,7 +193,7 @@ Then, to calibrate the leader arm (which is attached to the laptop/pc). Run the
 <hfoption id="Command">
 
 ```bash
-python -m lerobot.calibrate \
+lerobot-calibrate \
     --teleop.type=so100_leader \
     --teleop.port=/dev/tty.usbmodem58760431551 \ # <- The port of your robot
     --teleop.id=my_awesome_leader_arm # <- Give the robot a unique name

docs/source/reachy2.mdx

@@ -0,0 +1,288 @@
+# Reachy 2
+
+Reachy 2 is an open-source humanoid robot made by Pollen Robotics, specifically designed for the development of embodied AI and real-world applications.
+Check out [Pollen Robotics website](https://www.pollen-robotics.com/reachy/), or access [Reachy 2 documentation](https://docs.pollen-robotics.com/) for more information on the platform!
+
+## Teleoperate Reachy 2
+
+Currently, there are two ways to teleoperate Reachy 2:
+
+- Pollen Robotics’ VR teleoperation (not included in LeRobot).
+- Robot-to-robot teleoperation (use one Reachy 2 to control another).
+
+## Reachy 2 Simulation
+
+**(Linux only)** You can run Reachy 2 in simulation (Gazebo or MuJoCo) using the provided [Docker image](https://hub.docker.com/r/pollenrobotics/reachy2_core).
+
+1. Install [Docker Engine](https://docs.docker.com/engine/).
+2. Run (for MuJoCo):
+
+```
+docker run --rm -it \
+  --name reachy \
+  --privileged \
+  --network host \
+  --ipc host \
+  --device-cgroup-rule='c 189:* rwm' \
+  --group-add audio \
+  -e ROS_DOMAIN_ID="$ROS_DOMAIN_ID" \
+  -e DISPLAY="$DISPLAY" \
+  -e RCUTILS_CONSOLE_OUTPUT_FORMAT="[{severity}]: {message}" \
+  -e REACHY2_CORE_SERVICE_FAKE="${REACHY2_CORE_SERVICE_FAKE:-true}" \
+  -v /dev:/dev \
+  -v "$HOME/.reachy_config":/home/reachy/.reachy_config_override \
+  -v "$HOME/.reachy.log":/home/reachy/.ros/log \
+  -v /usr/lib/x86_64-linux-gnu:/opt/host-libs \
+  --entrypoint /package/launch.sh \
+  pollenrobotics/reachy2_core:1.7.5.9_deploy \
+  start_rviz:=true start_sdk_server:=true mujoco:=true
+```
+
+> If MuJoCo runs slowly (low simulation frequency), append `-e LD_LIBRARY_PATH="/opt/host-libs:$LD_LIBRARY_PATH" \` to the previous command to improve performance:
+>
+> ```
+> docker run --rm -it \
+>   --name reachy \
+>   --privileged \
+>   --network host \
+>   --ipc host \
+>   --device-cgroup-rule='c 189:* rwm' \
+>   --group-add audio \
+>   -e ROS_DOMAIN_ID="$ROS_DOMAIN_ID" \
+>   -e DISPLAY="$DISPLAY" \
+>   -e RCUTILS_CONSOLE_OUTPUT_FORMAT="[{severity}]: {message}" \
+>   -e REACHY2_CORE_SERVICE_FAKE="${REACHY2_CORE_SERVICE_FAKE:-true}" \
+>   -e LD_LIBRARY_PATH="/opt/host-libs:$LD_LIBRARY_PATH" \
+>   -v /dev:/dev \
+>   -v "$HOME/.reachy_config":/home/reachy/.reachy_config_override \
+>   -v "$HOME/.reachy.log":/home/reachy/.ros/log \
+>   -v /usr/lib/x86_64-linux-gnu:/opt/host-libs \
+>   --entrypoint /package/launch.sh \
+>   pollenrobotics/reachy2_core:1.7.5.9_deploy \
+>   start_rviz:=true start_sdk_server:=true mujoco:=true
+> ```
+
+## Setup
+
+### Prerequisites
+
+- On your robot, check the **service images** meet the minimum versions:
+  - **reachy2-core >= 1.7.5.2**
+  - **webrtc >= 2.0.1.1**
+
+Then, if you want to use VR teleoperation:
+
+- Install the [Reachy 2 teleoperation application](https://docs.pollen-robotics.com/teleoperation/teleoperation-introduction/discover-teleoperation/).
+  Use version **>=v1.2.0**
+
+We recommend using two computers: one for teleoperation (Windows required) and another for recording with LeRobot.
+
+### Install LeRobot
+
+Follow the [installation instructions](https://github.com/huggingface/lerobot#installation) to install LeRobot.
+
+Install LeRobot with Reachy 2 dependencies:
+
+```bash
+pip install -e ".[reachy2]"
+```
+
+### (Optional but recommended) Install pollen_data_acquisition_server
+
+How you manage Reachy 2 recording sessions is up to you, but the **easiest** way is to use this server so you can control sessions directly from the VR teleoperation app.
+
+> **Note:** Currently, only the VR teleoperation application works as a client for this server, so this step primarily targets teleoperation. You’re free to develop custom clients to manage sessions to your needs.
+
+In your LeRobot environment, install the server from source:
+
+```bash
+git clone https://github.com/pollen-robotics/pollen_data_acquisition_server.git
+cd pollen_data_acquisition_server
+pip install -e .
+```
+
+Find the [pollen_data_acquisition_server documentation here](https://github.com/pollen-robotics/pollen_data_acquisition_server).
+
+## Step 1: Recording
+
+### Get Reachy 2 IP address
+
+Before starting teleoperation and data recording, find the [robot's IP address](https://docs.pollen-robotics.com/getting-started/setup-reachy2/connect-reachy2/).
+We strongly recommend connecting all devices (PC and robot) via **Ethernet**.
+
+### Launch recording
+
+There are two ways to manage recording sessions when using the Reachy 2 VR teleoperation application:
+
+- **Using the data acquisition server (recommended for VR teleop)**: The VR app orchestrates sessions (via the server it tells LeRobot when to create datasets, start/stop episodes) while also controlling the robot’s motions.
+- **Using LeRobot’s record script**: LeRobot owns session control and decides when to start/stop episodes. If you also use the VR teleop app, it’s only for motion control.
+
+### Option 1: Using Pollen data acquisition server (recommended for VR teleop)
+
+Make sure you have installed pollen_data_acquisition_server, as explained in the Setup section.
+
+Launch the data acquisition server to be able to manage your session directly from the teleoperation application:
+
+```bash
+python -m pollen_data_acquisition_server.server
+```
+
+Then get into the teleoperation application and choose "Data acquisition session".
+You can finally setup your session by following the screens displayed.
+
+> Even without the VR app, you can use the `pollen_data_acquisition_server` with your own client implementation.
+
+### Option 2: Using lerobot.record
+
+Reachy 2 is fully supported by LeRobot’s recording features.
+If you choose this option but still want to use the VR teleoperation application, select "Standard session" in the app.
+
+**Example: start a recording without the mobile base:**
+First add reachy2 and reachy2_teleoperator to the imports of the record script. Then you can use the following command:
+
+```bash
+python -m lerobot.record \
+    --robot.type=reachy2 \
+    --robot.ip_address=192.168.0.200 \
+    --robot.id=r2-0000 \
+    --robot.use_external_commands=true \
+    --robot.with_mobile_base=false \
+    --teleop.type=reachy2_teleoperator \
+    --teleop.ip_address=192.168.0.200 \
+    --teleop.with_mobile_base=false \
+    --dataset.repo_id=pollen_robotics/record_test \
+    --dataset.single_task="Reachy 2 recording test" \
+    --dataset.num_episodes=1 \
+    --dataset.episode_time_s=5 \
+    --dataset.fps=15 \
+    --dataset.push_to_hub=true \
+    --dataset.private=true \
+    --display_data=true
+```
+
+#### Specific Options
+
+**Extended setup overview (all options included):**
+
+```bash
+python -m lerobot.record \
+    --robot.type=reachy2 \
+    --robot.ip_address=192.168.0.200 \
+    --robot.use_external_commands=true \
+    --robot.with_mobile_base=true \
+    --robot.with_l_arm=true \
+    --robot.with_r_arm=true \
+    --robot.with_neck=true \
+    --robot.with_antennas=true \
+    --robot.with_left_teleop_camera=true \
+    --robot.with_right_teleop_camera=true \
+    --robot.with_torso_camera=false \
+    --robot.disable_torque_on_disconnect=false \
+    --robot.max_relative_target=5.0 \
+    --teleop.type=reachy2_teleoperator \
+    --teleop.ip_address=192.168.0.200 \
+    --teleop.use_present_position=false \
+    --teleop.with_mobile_base=false \
+    --teleop.with_l_arm=true \
+    --teleop.with_r_arm=true \
+    --teleop.with_neck=true \
+    --teleop.with_antennas=true \
+    --dataset.repo_id=pollen_robotics/record_test \
+    --dataset.single_task="Reachy 2 recording test" \
+    --dataset.num_episodes=1 \
+    --dataset.episode_time_s=5 \
+    --dataset.fps=15 \
+    --dataset.push_to_hub=true \
+    --dataset.private=true \
+    --display_data=true
+```
+
+##### `--robot.use_external_commands`
+
+Determine whether LeRobot robot.send_action() sends commands to the robot.
+**Must** be set to false while using the VR teleoperation application, as the app already sends commands.
+
+##### `--teleop.use_present_position`
+
+Determine whether the teleoperator reads the goal or present position of the robot.
+Must be set to true if a compliant Reachy 2 is used to control another one.
+
+##### Use the relevant parts
+
+From our initial tests, recording **all** joints when only some are moving can reduce model quality with certain policies.
+To avoid this, you can exclude specific parts from recording and replay using:
+
+````
+--robot.with_<part>=false
+```,
+with `<part>` being one of : `mobile_base`, `l_arm`, `r_arm", `neck`, `antennas`.
+It determine whether the corresponding part is recorded in the observations. True if not set.
+
+By default, **all parts are recorded**.
+
+The same per-part mechanism is available in `reachy2_teleoperator` as well.
+
+````
+
+--teleop.with\_<part>
+
+```
+with `<part>` being one of : `mobile_base`, `l_arm`, `r_arm", `neck`, `antennas`.
+Determine whether the corresponding part is recorded in the actions. True if not set.
+
+> **Important:** In a given session, the **enabled parts must match** on both the robot and the teleoperator.
+For example, if the robot runs with `--robot.with_mobile_base=false`, the teleoperator must disable the same part `--teleoperator.with_mobile_base=false`.
+
+##### Use the relevant cameras
+
+You can do the same for **cameras**. By default, only the **teleoperation cameras** are recorded (both `left_teleop_camera` and `right_teleop_camera`). Enable or disable each camera with:
+
+```
+
+--robot.with_left_teleop_camera=<true|false>
+--robot.with_right_teleop_camera=<true|false>
+--robot.with_torso_camera=<true|false>
+
+````
+
+
+## Step 2: Replay
+
+Make sure the robot is configured with the same parts as the dataset:
+
+```bash
+python -m lerobot.replay \
+    --robot.type=reachy2 \
+    --robot.ip_address=192.168.0.200 \
+    --robot.use_external_commands=false \
+    --robot.with_mobile_base=false \
+    --dataset.repo_id=pollen_robotics/record_test \
+    --dataset.episode=0
+    --display_data=true
+````
+
+## Step 3: Train
+
+```bash
+python -m lerobot.scripts.train \
+  --dataset.repo_id=pollen_robotics/record_test \
+  --policy.type=act \
+  --output_dir=outputs/train/reachy2_test \
+  --job_name=reachy2 \
+  --policy.device=mps \
+  --wandb.enable=true \
+  --policy.repo_id=pollen_robotics/record_test_policy
+```
+
+## Step 4: Evaluate
+
+```bash
+python -m lerobot.record \
+  --robot.type=reachy2 \
+  --robot.ip_address=192.168.0.200 \
+  --display_data=false \
+  --dataset.repo_id=pollen_robotics/eval_record_test \
+  --dataset.single_task="Evaluate reachy2 policy" \
+  --dataset.num_episodes=10 \
+  --policy.path=outputs/train/reachy2_test/checkpoints/last/pretrained_model
+```

docs/source/smolvla.mdx

@@ -54,7 +54,7 @@ If you don't have a gpu device, you can train using our notebook on [![Google Co
 Pass your dataset to the training script using `--dataset.repo_id`. If you want to test your installation, run the following command where we use one of the datasets we collected for the [SmolVLA Paper](https://huggingface.co/papers/2506.01844).
 
 ```bash
-cd lerobot && python -m lerobot.scripts.train \
+cd lerobot && lerobot-train \
   --policy.path=lerobot/smolvla_base \
   --dataset.repo_id=${HF_USER}/mydataset \
   --batch_size=64 \
@@ -73,7 +73,7 @@ cd lerobot && python -m lerobot.scripts.train \
 Fine-tuning is an art. For a complete overview of the options for finetuning, run
 
 ```bash
-python -m lerobot.scripts.train --help
+lerobot-train --help
 ```
 
 <p align="center">
@@ -97,7 +97,7 @@ Similarly for when recording an episode, it is recommended that you are logged i
 Once you are logged in, you can run inference in your setup by doing:
 
 ```bash
-python -m lerobot.record \
+lerobot-record \
   --robot.type=so101_follower \
   --robot.port=/dev/ttyACM0 \ # <- Use your port
   --robot.id=my_blue_follower_arm \ # <- Use your robot id

docs/source/so100.mdx

@@ -26,7 +26,7 @@ Unlike the SO-101, the motor connectors are not easily accessible once the arm i
 To find the port for each bus servo adapter, run this script:
 
 ```bash
-python -m lerobot.find_port
+lerobot-find-port
 ```
 
 <hfoptions id="example">
@@ -93,7 +93,7 @@ For a visual reference on how to set the motor ids please refer to [this video](
 <hfoption id="Command">
 
 ```bash
-python -m lerobot.setup_motors \
+lerobot-setup-motors \
     --robot.type=so100_follower \
     --robot.port=/dev/tty.usbmodem585A0076841  # <- paste here the port found at previous step
 ```
@@ -168,7 +168,7 @@ Do the same steps for the leader arm.
 <hfoptions id="setup_motors">
 <hfoption id="Command">
 ```bash
-python -m lerobot.setup_motors \
+lerobot-setup-motors \
     --teleop.type=so100_leader \
     --teleop.port=/dev/tty.usbmodem575E0031751  # <- paste here the port found at previous step
 ```
@@ -568,7 +568,7 @@ Run the following command or API example to calibrate the follower arm:
 <hfoption id="Command">
 
 ```bash
-python -m lerobot.calibrate \
+lerobot-calibrate \
     --robot.type=so100_follower \
     --robot.port=/dev/tty.usbmodem58760431551 \ # <- The port of your robot
     --robot.id=my_awesome_follower_arm # <- Give the robot a unique name
@@ -606,7 +606,7 @@ Do the same steps to calibrate the leader arm, run the following command or API
 <hfoption id="Command">
 
 ```bash
-python -m lerobot.calibrate \
+lerobot-calibrate \
     --teleop.type=so100_leader \
     --teleop.port=/dev/tty.usbmodem58760431551 \ # <- The port of your robot
     --teleop.id=my_awesome_leader_arm # <- Give the robot a unique name

docs/source/so101.mdx

@@ -162,7 +162,7 @@ It is advisable to install one 3-pin cable in the motor after placing them befor
 To find the port for each bus servo adapter, connect MotorBus to your computer via USB and power. Run the following script and disconnect the MotorBus when prompted:
 
 ```bash
-python -m lerobot.find_port
+lerobot-find-port
 ```
 
 <hfoptions id="example">
@@ -240,7 +240,7 @@ Connect the usb cable from your computer and the power supply to the follower ar
 <hfoption id="Command">
 
 ```bash
-python -m lerobot.setup_motors \
+lerobot-setup-motors \
     --robot.type=so101_follower \
     --robot.port=/dev/tty.usbmodem585A0076841  # <- paste here the port found at previous step
 ```
@@ -316,7 +316,7 @@ Do the same steps for the leader arm.
 <hfoption id="Command">
 
 ```bash
-python -m lerobot.setup_motors \
+lerobot-setup-motors \
     --teleop.type=so101_leader \
     --teleop.port=/dev/tty.usbmodem575E0031751  # <- paste here the port found at previous step
 ```
@@ -353,7 +353,7 @@ Run the following command or API example to calibrate the follower arm:
 <hfoption id="Command">
 
 ```bash
-python -m lerobot.calibrate \
+lerobot-calibrate \
     --robot.type=so101_follower \
     --robot.port=/dev/tty.usbmodem58760431551 \ # <- The port of your robot
     --robot.id=my_awesome_follower_arm # <- Give the robot a unique name
@@ -402,7 +402,7 @@ Do the same steps to calibrate the leader arm, run the following command or API
 <hfoption id="Command">
 
 ```bash
-python -m lerobot.calibrate \
+lerobot-calibrate \
     --teleop.type=so101_leader \
     --teleop.port=/dev/tty.usbmodem58760431551 \ # <- The port of your robot
     --teleop.id=my_awesome_leader_arm # <- Give the robot a unique name

examples/4_train_policy_with_script.md

@@ -62,7 +62,7 @@ By default, every field takes its default value specified in the dataclass. If a
 Let's say that we want to train [Diffusion Policy](../src/lerobot/policies/diffusion) on the [pusht](https://huggingface.co/datasets/lerobot/pusht) dataset, using the [gym_pusht](https://github.com/huggingface/gym-pusht) environment for evaluation. The command to do so would look like this:
 
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
     --dataset.repo_id=lerobot/pusht \
     --policy.type=diffusion \
     --env.type=pusht
@@ -77,7 +77,7 @@ Let's break this down:
 Let's see another example. Let's say you've been training [ACT](../src/lerobot/policies/act) on [lerobot/aloha_sim_insertion_human](https://huggingface.co/datasets/lerobot/aloha_sim_insertion_human) using the [gym-aloha](https://github.com/huggingface/gym-aloha) environment for evaluation with:
 
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
     --policy.type=act \
     --dataset.repo_id=lerobot/aloha_sim_insertion_human \
     --env.type=aloha \
@@ -90,7 +90,7 @@ We now want to train a different policy for aloha on another task. We'll change
 Looking at the [`AlohaEnv`](../src/lerobot/envs/configs.py) config, the task is `"AlohaInsertion-v0"` by default, which corresponds to the task we trained on in the command above. The [gym-aloha](https://github.com/huggingface/gym-aloha?tab=readme-ov-file#description) environment also has the `AlohaTransferCube-v0` task which corresponds to this other task we want to train on. Putting this together, we can train this new policy on this different task using:
 
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
     --policy.type=act \
     --dataset.repo_id=lerobot/aloha_sim_transfer_cube_human \
     --env.type=aloha \
@@ -127,7 +127,7 @@ Now, let's assume that we want to reproduce the run just above. That run has pro
 We can then simply load the config values from this file using:
 
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
     --config_path=outputs/train/act_aloha_transfer/checkpoints/last/pretrained_model/ \
     --output_dir=outputs/train/act_aloha_transfer_2
 ```
@@ -137,7 +137,7 @@ python -m lerobot.scripts.train \
 Similarly to Hydra, we can still override some parameters in the CLI if we want to, e.g.:
 
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
     --config_path=outputs/train/act_aloha_transfer/checkpoints/last/pretrained_model/ \
     --output_dir=outputs/train/act_aloha_transfer_2
     --policy.n_action_steps=80
@@ -148,7 +148,7 @@ python -m lerobot.scripts.train \
 `--config_path` can also accept the repo_id of a repo on the hub that contains a `train_config.json` file, e.g. running:
 
 ```bash
-python -m lerobot.scripts.train --config_path=lerobot/diffusion_pusht
+lerobot-train --config_path=lerobot/diffusion_pusht
 ```
 
 will start a training run with the same configuration used for training [lerobot/diffusion_pusht](https://huggingface.co/lerobot/diffusion_pusht)
@@ -160,7 +160,7 @@ Being able to resume a training run is important in case it crashed or aborted f
 Let's reuse the command from the previous run and add a few more options:
 
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
     --policy.type=act \
     --dataset.repo_id=lerobot/aloha_sim_transfer_cube_human \
     --env.type=aloha \
@@ -179,7 +179,7 @@ INFO 2025-01-24 16:10:56 ts/train.py:263 Checkpoint policy after step 100
 Now let's simulate a crash by killing the process (hit `ctrl`+`c`). We can then simply resume this run from the last checkpoint available with:
 
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
     --config_path=outputs/train/run_resumption/checkpoints/last/pretrained_model/ \
     --resume=true
 ```
@@ -190,7 +190,7 @@ Another reason for which you might want to resume a run is simply to extend trai
 You could double the number of steps of the previous run with:
 
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
     --config_path=outputs/train/run_resumption/checkpoints/last/pretrained_model/ \
     --resume=true \
     --steps=200000
@@ -224,7 +224,7 @@ In addition to the features currently in Draccus, we've added a special `.path`
 For example, we could fine-tune a [policy pre-trained on the aloha transfer task](https://huggingface.co/lerobot/act_aloha_sim_transfer_cube_human) on the aloha insertion task. We can achieve this with:
 
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
     --policy.path=lerobot/act_aloha_sim_transfer_cube_human \
     --dataset.repo_id=lerobot/aloha_sim_insertion_human \
     --env.type=aloha \
@@ -270,7 +270,7 @@ We'll summarize here the main use cases to remember from this tutorial.
 #### Train a policy from scratch – CLI
 
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
     --policy.type=act \  # <- select 'act' policy
     --env.type=pusht \  # <- select 'pusht' environment
     --dataset.repo_id=lerobot/pusht  # <- train on this dataset
@@ -279,7 +279,7 @@ python -m lerobot.scripts.train \
 #### Train a policy from scratch - config file + CLI
 
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
     --config_path=path/to/pretrained_model \  # <- can also be a repo_id
     --policy.n_action_steps=80  # <- you may still override values
 ```
@@ -287,7 +287,7 @@ python -m lerobot.scripts.train \
 #### Resume/continue a training run
 
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
     --config_path=checkpoint/pretrained_model/ \
     --resume=true \
     --steps=200000  # <- you can change some training parameters
@@ -296,7 +296,7 @@ python -m lerobot.scripts.train \
 #### Fine-tuning
 
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
     --policy.path=lerobot/act_aloha_sim_transfer_cube_human \  # <- can also be a local path to a checkpoint
     --dataset.repo_id=lerobot/aloha_sim_insertion_human \
     --env.type=aloha \

examples/backward_compatibility/replay.py

@@ -18,7 +18,7 @@ Replays the actions of an episode from a dataset on a robot.
 Example:
 
 ```shell
-python -m lerobot.replay \
+lerobot-replay \
     --robot.type=so100_follower \
     --robot.port=/dev/tty.usbmodem58760431541 \
     --robot.id=black \

examples/lekiwi/evaluate.py

@@ -1,6 +1,7 @@
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.datasets.utils import hw_to_dataset_features
 from lerobot.policies.act.modeling_act import ACTPolicy
+from lerobot.policies.factory import make_pre_post_processors
 from lerobot.record import record_loop
 from lerobot.robots.lekiwi import LeKiwiClient, LeKiwiClientConfig
 from lerobot.utils.control_utils import init_keyboard_listener
@@ -11,12 +12,14 @@ NUM_EPISODES = 2
 FPS = 30
 EPISODE_TIME_SEC = 60
 TASK_DESCRIPTION = "My task description"
+HF_MODEL_ID = "<hf_username>/<model_repo_id>"
+HF_DATASET_ID = "<hf_username>/<eval_dataset_repo_id>"
 
 # Create the robot and teleoperator configurations
 robot_config = LeKiwiClientConfig(remote_ip="172.18.134.136", id="lekiwi")
 robot = LeKiwiClient(robot_config)
 
-policy = ACTPolicy.from_pretrained("<hf_username>/<policy_repo_id>")
+policy = ACTPolicy.from_pretrained(HF_MODEL_ID)
 
 # Configure the dataset features
 action_features = hw_to_dataset_features(robot.action_features, "action")
@@ -25,7 +28,7 @@ dataset_features = {**action_features, **obs_features}
 
 # Create the dataset
 dataset = LeRobotDataset.create(
-    repo_id="<hf_username>/<eval_dataset_repo_id>",
+    repo_id=HF_DATASET_ID,
     fps=FPS,
     features=dataset_features,
     robot_type=robot.name,
@@ -43,6 +46,12 @@ listener, events = init_keyboard_listener()
 if not robot.is_connected:
     raise ValueError("Robot is not connected!")
 
+preprocessor, postprocessor = make_pre_post_processors(
+    policy_cfg=policy,
+    pretrained_path=HF_MODEL_ID,
+    dataset_stats=dataset.meta.stats,
+)
+
 recorded_episodes = 0
 while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
     log_say(f"Running inference, recording eval episode {recorded_episodes} of {NUM_EPISODES}")
@@ -53,6 +62,8 @@ while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
         events=events,
         fps=FPS,
         policy=policy,
+        preprocessor=preprocessor,
+        postprocessor=postprocessor,
         dataset=dataset,
         control_time_s=EPISODE_TIME_SEC,
         single_task=TASK_DESCRIPTION,

examples/lekiwi/teleoperate.py

@@ -38,7 +38,7 @@ while True:
     keyboard_keys = keyboard.get_action()
     base_action = robot._from_keyboard_to_base_action(keyboard_keys)
 
-    log_rerun_data(observation, {**arm_action, **base_action})
+    log_rerun_data(observation=observation, action={**arm_action, **base_action})
 
     action = {**arm_action, **base_action} if len(base_action) > 0 else arm_action
 

examples/phone_to_so100/evaluate.py

@@ -0,0 +1,159 @@
+# !/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.datasets.utils import combine_feature_dicts
+from lerobot.model.kinematics import RobotKinematics
+from lerobot.policies.act.modeling_act import ACTPolicy
+from lerobot.policies.factory import make_pre_post_processors
+from lerobot.processor import RobotProcessorPipeline
+from lerobot.processor.converters import (
+    identity_transition,
+    observation_to_transition,
+    transition_to_action,
+)
+from lerobot.record import record_loop
+from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
+from lerobot.robots.so100_follower.robot_kinematic_processor import (
+    AddRobotObservationAsComplimentaryData,
+    ForwardKinematicsJointsToEE,
+    InverseKinematicsEEToJoints,
+)
+from lerobot.robots.so100_follower.so100_follower import SO100Follower
+from lerobot.utils.control_utils import init_keyboard_listener
+from lerobot.utils.utils import log_say
+from lerobot.utils.visualization_utils import _init_rerun
+
+NUM_EPISODES = 5
+FPS = 30
+EPISODE_TIME_SEC = 60
+TASK_DESCRIPTION = "My task description"
+HF_MODEL_ID = "<hf_username>/<model_repo_id>"
+HF_DATASET_ID = "<hf_username>/<dataset_repo_id>"
+
+# Initialize the robot with degrees
+camera_config = {"front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS)}
+robot_config = SO100FollowerConfig(
+    port="/dev/tty.usbmodem58760434471",
+    id="my_awesome_follower_arm",
+    cameras=camera_config,
+    use_degrees=True,
+)
+
+# Initialize the robot
+robot = SO100Follower(robot_config)
+
+# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+kinematics_solver = RobotKinematics(
+    urdf_path="./src/lerobot/teleoperators/sim/so101_new_calib.urdf",
+    target_frame_name="gripper_frame_link",
+    joint_names=list(robot.bus.motors.keys()),
+)
+
+# Build pipeline to convert ee pose action to joint action
+robot_ee_to_joints_processor = RobotProcessorPipeline(
+    steps=[
+        AddRobotObservationAsComplimentaryData(robot=robot),
+        InverseKinematicsEEToJoints(
+            kinematics=kinematics_solver,
+            motor_names=list(robot.bus.motors.keys()),
+            initial_guess_current_joints=True,
+        ),
+    ],
+    to_transition=identity_transition,
+    to_output=transition_to_action,
+)
+
+# Build pipeline to convert joint observation to ee pose observation
+robot_joints_to_ee_pose_processor = RobotProcessorPipeline(
+    steps=[
+        ForwardKinematicsJointsToEE(kinematics=kinematics_solver, motor_names=list(robot.bus.motors.keys()))
+    ],
+    to_transition=observation_to_transition,
+    to_output=identity_transition,
+)
+
+# Build dataset action and gripper features
+action_ee_and_gripper = aggregate_pipeline_dataset_features(
+    pipeline=robot_ee_to_joints_processor,
+    initial_features=create_initial_features(),
+    use_videos=True,
+    patterns=["action.ee", "action.gripper.pos", "observation.state.gripper.pos"],
+)  # Get all ee action features + gripper pos action features
+
+# Build dataset observation features
+obs_ee = aggregate_pipeline_dataset_features(
+    pipeline=robot_joints_to_ee_pose_processor,
+    initial_features=create_initial_features(observation=robot.observation_features),
+    use_videos=True,
+    patterns=["observation.state.ee"],
+)  # Get all ee observation features
+
+dataset_features = combine_feature_dicts(obs_ee, action_ee_and_gripper)
+
+print("All dataset features: ", dataset_features)
+
+# Create the dataset
+dataset = LeRobotDataset.create(
+    repo_id=HF_DATASET_ID,
+    fps=FPS,
+    features=dataset_features,
+    robot_type=robot.name,
+    use_videos=True,
+    image_writer_threads=4,
+)
+
+# Initialize the keyboard listener and rerun visualization
+_, events = init_keyboard_listener()
+_init_rerun(session_name="recording_phone")
+
+# Connect the robot and teleoperator
+robot.connect()
+
+episode_idx = 0
+
+policy = ACTPolicy.from_pretrained(HF_MODEL_ID)
+preprocessor, postprocessor = make_pre_post_processors(
+    policy_cfg=policy,
+    pretrained_path=HF_MODEL_ID,
+    dataset_stats=dataset.meta.stats,
+)
+
+for episode_idx in range(NUM_EPISODES):
+    log_say(f"Running inference, recording eval episode {episode_idx + 1} of {NUM_EPISODES}")
+
+    record_loop(
+        robot=robot,
+        events=events,
+        fps=FPS,
+        policy=policy,
+        preprocessor=preprocessor,
+        postprocessor=postprocessor,
+        dataset=dataset,
+        control_time_s=EPISODE_TIME_SEC,
+        single_task=TASK_DESCRIPTION,
+        display_data=True,
+        robot_action_processor=robot_ee_to_joints_processor,
+        robot_observation_processor=robot_joints_to_ee_pose_processor,
+    )
+    dataset.save_episode()
+
+# Clean up
+log_say("Stop recording")
+robot.disconnect()
+dataset.push_to_hub()

examples/phone_to_so100/record.py

@@ -0,0 +1,216 @@
+# !/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.datasets.utils import combine_feature_dicts
+from lerobot.model.kinematics import RobotKinematics
+from lerobot.processor import RobotProcessorPipeline
+from lerobot.processor.converters import (
+    action_to_transition,
+    identity_transition,
+    observation_to_transition,
+    transition_to_action,
+)
+from lerobot.record import record_loop
+from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
+from lerobot.robots.so100_follower.robot_kinematic_processor import (
+    AddRobotObservationAsComplimentaryData,
+    EEBoundsAndSafety,
+    EEReferenceAndDelta,
+    ForwardKinematicsJointsToEE,
+    GripperVelocityToJoint,
+    InverseKinematicsEEToJoints,
+)
+from lerobot.robots.so100_follower.so100_follower import SO100Follower
+from lerobot.teleoperators.phone.config_phone import PhoneConfig, PhoneOS
+from lerobot.teleoperators.phone.phone_processor import MapPhoneActionToRobotAction
+from lerobot.teleoperators.phone.teleop_phone import Phone
+from lerobot.utils.control_utils import init_keyboard_listener
+from lerobot.utils.utils import log_say
+from lerobot.utils.visualization_utils import _init_rerun
+
+NUM_EPISODES = 10
+FPS = 30
+EPISODE_TIME_SEC = 60
+RESET_TIME_SEC = 30
+TASK_DESCRIPTION = "My task description"
+HF_REPO_ID = "<hf_username>/<dataset_repo_id>"
+
+# Initialize the robot and teleoperator
+camera_config = {"front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS)}
+robot_config = SO100FollowerConfig(
+    port="/dev/tty.usbmodem58760434471",
+    id="my_awesome_follower_arm",
+    cameras=camera_config,
+    use_degrees=True,
+)
+teleop_config = PhoneConfig(phone_os=PhoneOS.IOS)  # or PhoneOS.ANDROID
+
+# Initialize the robot and teleoperator
+robot = SO100Follower(robot_config)
+phone = Phone(teleop_config)
+
+# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+kinematics_solver = RobotKinematics(
+    urdf_path="./src/lerobot/teleoperators/sim/so101_new_calib.urdf",
+    target_frame_name="gripper_frame_link",
+    joint_names=list(robot.bus.motors.keys()),
+)
+
+# Build pipeline to convert phone action to ee pose action
+phone_to_robot_ee_pose_processor = RobotProcessorPipeline(
+    steps=[
+        MapPhoneActionToRobotAction(platform=teleop_config.phone_os),
+        AddRobotObservationAsComplimentaryData(robot=robot),
+        EEReferenceAndDelta(
+            kinematics=kinematics_solver,
+            end_effector_step_sizes={"x": 0.5, "y": 0.5, "z": 0.5},
+            motor_names=list(robot.bus.motors.keys()),
+        ),
+        EEBoundsAndSafety(
+            end_effector_bounds={"min": [-1.0, -1.0, -1.0], "max": [1.0, 1.0, 1.0]},
+            max_ee_step_m=0.20,
+            max_ee_twist_step_rad=0.50,
+        ),
+    ],
+    to_transition=action_to_transition,
+    to_output=identity_transition,
+)
+
+# Build pipeline to convert ee pose action to joint action
+robot_ee_to_joints_processor = RobotProcessorPipeline(
+    steps=[
+        InverseKinematicsEEToJoints(
+            kinematics=kinematics_solver,
+            motor_names=list(robot.bus.motors.keys()),
+            initial_guess_current_joints=True,
+        ),
+        GripperVelocityToJoint(
+            motor_names=list(robot.bus.motors.keys()),
+            speed_factor=20.0,
+        ),
+    ],
+    to_transition=identity_transition,
+    to_output=transition_to_action,
+)
+
+# Build pipeline to convert joint observation to ee pose observation
+robot_joints_to_ee_pose = RobotProcessorPipeline(
+    steps=[
+        ForwardKinematicsJointsToEE(kinematics=kinematics_solver, motor_names=list(robot.bus.motors.keys()))
+    ],
+    to_transition=observation_to_transition,
+    to_output=identity_transition,
+)
+
+# Build dataset ee action features
+action_ee = aggregate_pipeline_dataset_features(
+    pipeline=phone_to_robot_ee_pose_processor,
+    initial_features=create_initial_features(action=phone.action_features),
+    use_videos=True,
+    patterns=["action.ee"],
+)
+
+# Get gripper pos action features
+gripper = aggregate_pipeline_dataset_features(
+    pipeline=robot_ee_to_joints_processor,
+    initial_features=create_initial_features(),
+    use_videos=True,
+    patterns=["action.gripper.pos", "observation.state.gripper.pos"],
+)
+
+# Build dataset ee observation features
+observation_ee = aggregate_pipeline_dataset_features(
+    pipeline=robot_joints_to_ee_pose,
+    initial_features=create_initial_features(observation=robot.observation_features),
+    use_videos=True,
+    patterns=["observation.state.ee"],
+)
+
+dataset_features = combine_feature_dicts(action_ee, gripper, observation_ee)
+
+print("All dataset features: ", dataset_features)
+
+# Create the dataset
+dataset = LeRobotDataset.create(
+    repo_id=HF_REPO_ID,
+    fps=FPS,
+    features=dataset_features,
+    robot_type=robot.name,
+    use_videos=True,
+    image_writer_threads=4,
+)
+
+# Initialize the keyboard listener and rerun visualization
+_, events = init_keyboard_listener()
+_init_rerun(session_name="recording_phone")
+
+# Connect the robot and teleoperator
+robot.connect()
+phone.connect()
+
+episode_idx = 0
+while episode_idx < NUM_EPISODES and not events["stop_recording"]:
+    log_say(f"Recording episode {episode_idx + 1} of {NUM_EPISODES}")
+
+    record_loop(
+        robot=robot,
+        events=events,
+        fps=FPS,
+        teleop=phone,
+        dataset=dataset,
+        control_time_s=EPISODE_TIME_SEC,
+        single_task=TASK_DESCRIPTION,
+        display_data=True,
+        teleop_action_processor=phone_to_robot_ee_pose_processor,
+        robot_action_processor=robot_ee_to_joints_processor,
+        robot_observation_processor=robot_joints_to_ee_pose,
+    )
+
+    # Reset the environment if not stopping or re-recording
+    if not events["stop_recording"] and (episode_idx < NUM_EPISODES - 1 or events["rerecord_episode"]):
+        log_say("Reset the environment")
+        record_loop(
+            robot=robot,
+            events=events,
+            fps=FPS,
+            teleop=phone,
+            control_time_s=RESET_TIME_SEC,
+            single_task=TASK_DESCRIPTION,
+            display_data=True,
+            teleop_action_processor=phone_to_robot_ee_pose_processor,
+            robot_action_processor=robot_ee_to_joints_processor,
+            robot_observation_processor=robot_joints_to_ee_pose,
+        )
+
+    if events["rerecord_episode"]:
+        log_say("Re-recording episode")
+        events["rerecord_episode"] = False
+        events["exit_early"] = False
+        dataset.clear_episode_buffer()
+        continue
+
+    dataset.save_episode()
+    episode_idx += 1
+
+# Clean up
+log_say("Stop recording")
+robot.disconnect()
+phone.disconnect()
+dataset.push_to_hub()

examples/phone_to_so100/replay.py

@@ -0,0 +1,81 @@
+# !/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import time
+
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.model.kinematics import RobotKinematics
+from lerobot.processor import RobotProcessorPipeline
+from lerobot.processor.converters import action_to_transition, transition_to_action
+from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
+from lerobot.robots.so100_follower.robot_kinematic_processor import (
+    AddRobotObservationAsComplimentaryData,
+    InverseKinematicsEEToJoints,
+)
+from lerobot.robots.so100_follower.so100_follower import SO100Follower
+from lerobot.utils.robot_utils import busy_wait
+from lerobot.utils.utils import log_say
+
+EPISODE_IDX = 0
+HF_REPO_ID = "<hf_username>/<dataset_repo_id>"
+
+robot_config = SO100FollowerConfig(
+    port="/dev/tty.usbmodem58760434471", id="my_awesome_follower_arm", use_degrees=True
+)
+robot = SO100Follower(robot_config)
+robot.connect()
+
+dataset = LeRobotDataset(HF_REPO_ID, episodes=[EPISODE_IDX])
+actions = dataset.hf_dataset.select_columns("action")
+
+# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+kinematics_solver = RobotKinematics(
+    urdf_path="./src/lerobot/teleoperators/sim/so101_new_calib.urdf",
+    target_frame_name="gripper_frame_link",
+    joint_names=list(robot.bus.motors.keys()),
+)
+
+# Build pipeline to convert ee pose action to joint action
+robot_ee_to_joints_processor = RobotProcessorPipeline(
+    steps=[
+        AddRobotObservationAsComplimentaryData(robot=robot),
+        InverseKinematicsEEToJoints(
+            kinematics=kinematics_solver,
+            motor_names=list(robot.bus.motors.keys()),
+            initial_guess_current_joints=False,  # Because replay is open loop
+        ),
+    ],
+    to_transition=action_to_transition,
+    to_output=transition_to_action,
+)
+
+robot_ee_to_joints_processor.reset()
+
+log_say(f"Replaying episode {EPISODE_IDX}")
+for idx in range(dataset.num_frames):
+    t0 = time.perf_counter()
+
+    ee_action = {
+        name: float(actions[idx]["action"][i]) for i, name in enumerate(dataset.features["action"]["names"])
+    }
+
+    joint_action = robot_ee_to_joints_processor(ee_action)
+    action_sent = robot.send_action(joint_action)
+
+    busy_wait(1.0 / dataset.fps - (time.perf_counter() - t0))
+
+robot.disconnect()

examples/phone_to_so100/teleoperate.py

@@ -0,0 +1,93 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specif
+
+import time
+
+from lerobot.model.kinematics import RobotKinematics
+from lerobot.processor import RobotProcessorPipeline
+from lerobot.processor.converters import action_to_transition, transition_to_action
+from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
+from lerobot.robots.so100_follower.robot_kinematic_processor import (
+    AddRobotObservationAsComplimentaryData,
+    EEBoundsAndSafety,
+    EEReferenceAndDelta,
+    GripperVelocityToJoint,
+    InverseKinematicsEEToJoints,
+)
+from lerobot.robots.so100_follower.so100_follower import SO100Follower
+from lerobot.teleoperators.phone.config_phone import PhoneConfig, PhoneOS
+from lerobot.teleoperators.phone.phone_processor import MapPhoneActionToRobotAction
+from lerobot.teleoperators.phone.teleop_phone import Phone
+
+# Initialize the robot and teleoperator
+robot_config = SO100FollowerConfig(
+    port="/dev/tty.usbmodem58760434471", id="my_awesome_follower_arm", use_degrees=True
+)
+teleop_config = PhoneConfig(phone_os=PhoneOS.IOS)  # or PhoneOS.ANDROID
+
+# Initialize the robot and teleoperator
+robot = SO100Follower(robot_config)
+teleop_device = Phone(teleop_config)
+
+# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+kinematics_solver = RobotKinematics(
+    urdf_path="./src/lerobot/teleoperators/sim/so101_new_calib.urdf",
+    target_frame_name="gripper_frame_link",
+    joint_names=list(robot.bus.motors.keys()),
+)
+
+# Build pipeline to convert phone action to ee pose action to joint action
+phone_to_robot_joints_processor = RobotProcessorPipeline(
+    steps=[
+        MapPhoneActionToRobotAction(platform=teleop_config.phone_os),
+        AddRobotObservationAsComplimentaryData(robot=robot),
+        EEReferenceAndDelta(
+            kinematics=kinematics_solver,
+            end_effector_step_sizes={"x": 0.5, "y": 0.5, "z": 0.5},
+            motor_names=list(robot.bus.motors.keys()),
+        ),
+        EEBoundsAndSafety(
+            end_effector_bounds={"min": [-1.0, -1.0, -1.0], "max": [1.0, 1.0, 1.0]},
+            max_ee_step_m=0.10,
+            max_ee_twist_step_rad=0.50,
+        ),
+        InverseKinematicsEEToJoints(
+            kinematics=kinematics_solver,
+            motor_names=list(robot.bus.motors.keys()),
+        ),
+        GripperVelocityToJoint(
+            motor_names=list(robot.bus.motors.keys()),
+            speed_factor=20.0,
+        ),
+    ],
+    to_transition=action_to_transition,
+    to_output=transition_to_action,
+)
+
+robot.connect()
+teleop_device.connect()
+
+print("Starting teleop loop. Move your phone to teleoperate the robot.")
+while True:
+    # Get teleop observation
+    phone_obs = teleop_device.get_action()
+
+    # Phone -> EE pose -> Joints transition
+    joint_action = phone_to_robot_joints_processor(phone_obs)
+
+    if joint_action:
+        robot.send_action(joint_action)
+
+    time.sleep(0.01)

pyproject.toml

@@ -25,7 +25,7 @@ discord = "https://discord.gg/s3KuuzsPFb"
 
 [project]
 name = "lerobot"
-version = "0.3.3"
+version = "0.3.4"
 description = "🤗 LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch"
 readme = "README.md"
 license = { text = "Apache-2.0" }
@@ -95,7 +95,7 @@ dependencies = [
 # Common
 pygame-dep = ["pygame>=2.5.1"]
 placo-dep = ["placo>=0.9.6"]
-transformers-dep = ["transformers>=4.50.3,<4.52.0"] # TODO: Bumb dependency
+transformers-dep = ["transformers<=4.52.0"]
 grpcio-dep = ["grpcio==1.73.1", "protobuf==6.31.0"]
 
 # Motors
@@ -106,11 +106,13 @@ dynamixel = ["dynamixel-sdk>=3.7.31"]
 gamepad = ["lerobot[pygame-dep]", "hidapi>=0.14.0"]
 hopejr = ["lerobot[feetech]", "lerobot[pygame-dep]"]
 lekiwi = ["lerobot[feetech]", "pyzmq>=26.2.1"]
+reachy2 = ["reachy2_sdk>=1.0.14"]
 kinematics = ["lerobot[placo-dep]"]
 intelrealsense = [
     "pyrealsense2>=2.55.1.6486 ; sys_platform != 'darwin'",
     "pyrealsense2-macosx>=2.54 ; sys_platform == 'darwin'",
 ]
+phone = ["hebi-py>=2.8.0", "teleop>=0.1.0"]
 # stretch = [
 #     "hello-robot-stretch-body>=0.7.27 ; sys_platform == 'linux'",
 #     "pyrender @ git+https://github.com/mmatl/pyrender.git ; sys_platform == 'linux'",
@@ -141,6 +143,7 @@ all = [
     "lerobot[gamepad]",
     "lerobot[hopejr]",
     "lerobot[lekiwi]",
+    "lerobot[reachy2]",
     "lerobot[kinematics]",
     "lerobot[intelrealsense]",
     "lerobot[pi0]",
@@ -152,7 +155,8 @@ all = [
     "lerobot[video_benchmark]",
     "lerobot[aloha]",
     "lerobot[pusht]",
-    "lerobot[xarm]"
+    "lerobot[xarm]",
+    "lerobot[phone]",
 ]
 
 [project.scripts]

src/lerobot/calibrate.py

@@ -18,7 +18,7 @@ Helper to recalibrate your device (robot or teleoperator).
 Example:
 
 ```shell
-python -m lerobot.calibrate \
+lerobot-calibrate \
     --teleop.type=so100_leader \
     --teleop.port=/dev/tty.usbmodem58760431551 \
     --teleop.id=blue

src/lerobot/cameras/opencv/camera_opencv.py

@@ -60,7 +60,7 @@ class OpenCVCamera(Camera):
     or port changes, especially on Linux. Use the provided utility script to find
     available camera indices or paths:
     ```bash
-    python -m lerobot.find_cameras opencv
+    lerobot-find-cameras opencv
     ```
 
     The camera's default settings (FPS, resolution, color mode) are used unless
@@ -165,8 +165,7 @@ class OpenCVCamera(Camera):
             self.videocapture.release()
             self.videocapture = None
             raise ConnectionError(
-                f"Failed to open {self}."
-                f"Run `python -m lerobot.find_cameras opencv` to find available cameras."
+                f"Failed to open {self}.Run `lerobot-find-cameras opencv` to find available cameras."
             )
 
         self._configure_capture_settings()

src/lerobot/cameras/reachy2_camera/__init__.py

@@ -0,0 +1,16 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .configuration_reachy2_camera import Reachy2CameraConfig
+from .reachy2_camera import Reachy2Camera

src/lerobot/cameras/reachy2_camera/configuration_reachy2_camera.py

@@ -0,0 +1,78 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+
+from ..configs import CameraConfig, ColorMode
+
+
+@CameraConfig.register_subclass("reachy2_camera")
+@dataclass
+class Reachy2CameraConfig(CameraConfig):
+    """Configuration class for Reachy 2 camera devices.
+
+    This class provides configuration options for Reachy 2 cameras,
+    supporting both the teleop and depth cameras. It includes settings
+    for resolution, frame rate, color mode, and the selection of the cameras.
+
+    Example configurations:
+    ```python
+    # Basic configurations
+    Reachy2CameraConfig(
+        name="teleop",
+        image_type="left",
+        ip_address="192.168.0.200",  # IP address of the robot
+        fps=15,
+        width=640,
+        height=480,
+        color_mode=ColorMode.RGB,
+    )  # Left teleop camera, 640x480 @ 15FPS
+    ```
+
+    Attributes:
+        name: Name of the camera device. Can be "teleop" or "depth".
+        image_type: Type of image stream. For "teleop" camera, can be "left" or "right".
+                    For "depth" camera, can be "rgb" or "depth". (depth is not supported yet)
+        fps: Requested frames per second for the color stream.
+        width: Requested frame width in pixels for the color stream.
+        height: Requested frame height in pixels for the color stream.
+        color_mode: Color mode for image output (RGB or BGR). Defaults to RGB.
+        ip_address: IP address of the robot. Defaults to "localhost".
+        port: Port number for the camera server. Defaults to 50065.
+
+    Note:
+        - Only 3-channel color output (RGB/BGR) is currently supported.
+    """
+
+    name: str
+    image_type: str
+    color_mode: ColorMode = ColorMode.RGB
+    ip_address: str | None = "localhost"
+    port: int = 50065
+    # use_depth: bool = False
+
+    def __post_init__(self):
+        if self.name not in ["teleop", "depth"]:
+            raise ValueError(f"`name` is expected to be 'teleop' or 'depth', but {self.name} is provided.")
+        if (self.name == "teleop" and self.image_type not in ["left", "right"]) or (
+            self.name == "depth" and self.image_type not in ["rgb", "depth"]
+        ):
+            raise ValueError(
+                f"`image_type` is expected to be 'left' or 'right' for teleop camera, and 'rgb' or 'depth' for depth camera, but {self.image_type} is provided."
+            )
+
+        if self.color_mode not in ["rgb", "bgr"]:
+            raise ValueError(
+                f"`color_mode` is expected to be 'rgb' or 'bgr', but {self.color_mode} is provided."
+            )

src/lerobot/cameras/reachy2_camera/reachy2_camera.py

@@ -0,0 +1,288 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Provides the Reachy2Camera class for capturing frames from Reachy 2 cameras using Reachy 2's CameraManager.
+"""
+
+import logging
+import os
+import platform
+import time
+from threading import Event, Lock, Thread
+from typing import Any
+
+# Fix MSMF hardware transform compatibility for Windows before importing cv2
+if platform.system() == "Windows" and "OPENCV_VIDEOIO_MSMF_ENABLE_HW_TRANSFORMS" not in os.environ:
+    os.environ["OPENCV_VIDEOIO_MSMF_ENABLE_HW_TRANSFORMS"] = "0"
+import cv2
+import numpy as np
+from reachy2_sdk.media.camera import CameraView
+from reachy2_sdk.media.camera_manager import CameraManager
+
+from lerobot.errors import DeviceNotConnectedError
+
+from ..camera import Camera
+from .configuration_reachy2_camera import ColorMode, Reachy2CameraConfig
+
+logger = logging.getLogger(__name__)
+
+
+class Reachy2Camera(Camera):
+    """
+    Manages Reachy 2 camera using Reachy 2 CameraManager.
+
+    This class provides a high-level interface to connect to, configure, and read
+    frames from Reachy 2 cameras. It supports both synchronous and asynchronous
+    frame reading.
+
+    An Reachy2Camera instance requires a camera name (e.g., "teleop") and an image
+    type (e.g., "left") to be specified in the configuration.
+
+    The camera's default settings (FPS, resolution, color mode) are used unless
+    overridden in the configuration.
+    """
+
+    def __init__(self, config: Reachy2CameraConfig):
+        """
+        Initializes the Reachy2Camera instance.
+
+        Args:
+            config: The configuration settings for the camera.
+        """
+        super().__init__(config)
+
+        self.config = config
+
+        self.fps = config.fps
+        self.color_mode = config.color_mode
+
+        self.cam_manager: CameraManager | None = None
+
+        self.thread: Thread | None = None
+        self.stop_event: Event | None = None
+        self.frame_lock: Lock = Lock()
+        self.latest_frame: np.ndarray | None = None
+        self.new_frame_event: Event = Event()
+
+    def __str__(self) -> str:
+        return f"{self.__class__.__name__}({self.config.name}, {self.config.image_type})"
+
+    @property
+    def is_connected(self) -> bool:
+        """Checks if the camera is currently connected and opened."""
+        if self.config.name == "teleop":
+            return self.cam_manager._grpc_connected and self.cam_manager.teleop if self.cam_manager else False
+        elif self.config.name == "depth":
+            return self.cam_manager._grpc_connected and self.cam_manager.depth if self.cam_manager else False
+        else:
+            raise ValueError(f"Invalid camera name '{self.config.name}'. Expected 'teleop' or 'depth'.")
+
+    def connect(self, warmup: bool = True):
+        """
+        Connects to the Reachy2 CameraManager as specified in the configuration.
+        """
+        self.cam_manager = CameraManager(host=self.config.ip_address, port=self.config.port)
+        self.cam_manager.initialize_cameras()
+
+        logger.info(f"{self} connected.")
+
+    @staticmethod
+    def find_cameras(ip_address: str = "localhost", port: int = 50065) -> list[dict[str, Any]]:
+        """
+        Detects available Reachy 2 cameras.
+
+        Returns:
+            List[Dict[str, Any]]: A list of dictionaries,
+            where each dictionary contains 'name', 'stereo',
+            and the default profile properties (width, height, fps).
+        """
+        initialized_cameras = []
+        camera_manager = CameraManager(host=ip_address, port=port)
+
+        for camera in [camera_manager.teleop, camera_manager.depth]:
+            if camera is None:
+                continue
+
+            height, width, _, _, _, _, _ = camera.get_parameters()
+
+            camera_info = {
+                "name": camera._cam_info.name,
+                "stereo": camera._cam_info.stereo,
+                "default_profile": {
+                    "width": width,
+                    "height": height,
+                    "fps": 30,
+                },
+            }
+            initialized_cameras.append(camera_info)
+
+        camera_manager.disconnect()
+        return initialized_cameras
+
+    def read(self, color_mode: ColorMode | None = None) -> np.ndarray:
+        """
+        Reads a single frame synchronously from the camera.
+
+        This is a blocking call.
+
+        Args:
+            color_mode (Optional[ColorMode]): If specified, overrides the default
+                color mode (`self.color_mode`) for this read operation (e.g.,
+                request RGB even if default is BGR).
+
+        Returns:
+            np.ndarray: The captured frame as a NumPy array in the format
+                       (height, width, channels), using the specified or default
+                       color mode and applying any configured rotation.
+        """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+
+        start_time = time.perf_counter()
+
+        frame = None
+
+        if self.cam_manager is None:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+        else:
+            if self.config.name == "teleop" and hasattr(self.cam_manager, "teleop"):
+                if self.config.image_type == "left":
+                    frame = self.cam_manager.teleop.get_frame(CameraView.LEFT, size=(640, 480))[0]
+                elif self.config.image_type == "right":
+                    frame = self.cam_manager.teleop.get_frame(CameraView.RIGHT, size=(640, 480))[0]
+            elif self.config.name == "depth" and hasattr(self.cam_manager, "depth"):
+                if self.config.image_type == "depth":
+                    frame = self.cam_manager.depth.get_depth_frame()[0]
+                elif self.config.image_type == "rgb":
+                    frame = self.cam_manager.depth.get_frame(size=(640, 480))[0]
+
+            if frame is None:
+                return np.empty((0, 0, 3), dtype=np.uint8)
+
+            if self.config.color_mode == "rgb":
+                frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+
+        read_duration_ms = (time.perf_counter() - start_time) * 1e3
+        logger.debug(f"{self} read took: {read_duration_ms:.1f}ms")
+
+        return frame
+
+    def _read_loop(self):
+        """
+        Internal loop run by the background thread for asynchronous reading.
+
+        On each iteration:
+        1. Reads a color frame
+        2. Stores result in latest_frame (thread-safe)
+        3. Sets new_frame_event to notify listeners
+
+        Stops on DeviceNotConnectedError, logs other errors and continues.
+        """
+        while not self.stop_event.is_set():
+            try:
+                color_image = self.read()
+
+                with self.frame_lock:
+                    self.latest_frame = color_image
+                self.new_frame_event.set()
+
+            except DeviceNotConnectedError:
+                break
+            except Exception as e:
+                logger.warning(f"Error reading frame in background thread for {self}: {e}")
+
+    def _start_read_thread(self) -> None:
+        """Starts or restarts the background read thread if it's not running."""
+        if self.thread is not None and self.thread.is_alive():
+            self.thread.join(timeout=0.1)
+        if self.stop_event is not None:
+            self.stop_event.set()
+
+        self.stop_event = Event()
+        self.thread = Thread(target=self._read_loop, args=(), name=f"{self}_read_loop")
+        self.thread.daemon = True
+        self.thread.start()
+
+    def _stop_read_thread(self) -> None:
+        """Signals the background read thread to stop and waits for it to join."""
+        if self.stop_event is not None:
+            self.stop_event.set()
+
+        if self.thread is not None and self.thread.is_alive():
+            self.thread.join(timeout=2.0)
+
+        self.thread = None
+        self.stop_event = None
+
+    def async_read(self, timeout_ms: float = 200) -> np.ndarray:
+        """
+        Reads the latest available frame asynchronously.
+
+        This method retrieves the most recent frame captured by the background
+        read thread. It does not block waiting for the camera hardware directly,
+        but may wait up to timeout_ms for the background thread to provide a frame.
+
+        Args:
+            timeout_ms (float): Maximum time in milliseconds to wait for a frame
+                to become available. Defaults to 200ms (0.2 seconds).
+
+        Returns:
+            np.ndarray: The latest captured frame as a NumPy array in the format
+                       (height, width, channels), processed according to configuration.
+
+        Raises:
+            DeviceNotConnectedError: If the camera is not connected.
+            TimeoutError: If no frame becomes available within the specified timeout.
+            RuntimeError: If an unexpected error occurs.
+        """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+
+        if self.thread is None or not self.thread.is_alive():
+            self._start_read_thread()
+
+        if not self.new_frame_event.wait(timeout=timeout_ms / 1000.0):
+            thread_alive = self.thread is not None and self.thread.is_alive()
+            raise TimeoutError(
+                f"Timed out waiting for frame from camera {self} after {timeout_ms} ms. "
+                f"Read thread alive: {thread_alive}."
+            )
+
+        with self.frame_lock:
+            frame = self.latest_frame
+            self.new_frame_event.clear()
+
+        if frame is None:
+            raise RuntimeError(f"Internal error: Event set but no frame available for {self}.")
+
+        return frame
+
+    def disconnect(self):
+        """
+        Stops the background read thread (if running).
+
+        Raises:
+            DeviceNotConnectedError: If the camera is already disconnected.
+        """
+        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.cam_manager is not None:
+            self.cam_manager.disconnect()
+
+        logger.info(f"{self} disconnected.")

src/lerobot/cameras/realsense/camera_realsense.py

@@ -51,7 +51,7 @@ class RealSenseCamera(Camera):
 
     Use the provided utility script to find available camera indices and default profiles:
     ```bash
-    python -m lerobot.find_cameras realsense
+    lerobot-find-cameras realsense
     ```
 
     A `RealSenseCamera` instance requires a configuration object specifying the
@@ -176,8 +176,7 @@ class RealSenseCamera(Camera):
             self.rs_profile = None
             self.rs_pipeline = None
             raise ConnectionError(
-                f"Failed to open {self}."
-                "Run `python -m lerobot.find_cameras realsense` to find available cameras."
+                f"Failed to open {self}.Run `lerobot-find-cameras realsense` to find available cameras."
             ) from e
 
         self._configure_capture_settings()

src/lerobot/cameras/utils.py

@@ -37,8 +37,14 @@ def make_cameras_from_configs(camera_configs: dict[str, CameraConfig]) -> dict[s
             from .realsense.camera_realsense import RealSenseCamera
 
             cameras[key] = RealSenseCamera(cfg)
+
+        elif cfg.type == "reachy2_camera":
+            from .reachy2_camera.reachy2_camera import Reachy2Camera
+
+            cameras[key] = Reachy2Camera(cfg)
+
         else:
-            raise ValueError(f"The motor type '{cfg.type}' is not valid.")
+            raise ValueError(f"The camera type '{cfg.type}' is not valid.")
 
     return cameras
 

src/lerobot/configs/policies.py

@@ -26,7 +26,7 @@ from huggingface_hub import hf_hub_download
 from huggingface_hub.constants import CONFIG_NAME
 from huggingface_hub.errors import HfHubHTTPError
 
-from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
+from lerobot.configs.types import FeatureType, PolicyFeature
 from lerobot.constants import ACTION, OBS_STATE
 from lerobot.optim.optimizers import OptimizerConfig
 from lerobot.optim.schedulers import LRSchedulerConfig
@@ -53,7 +53,6 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):
     """
 
     n_obs_steps: int = 1
-    normalization_mapping: dict[str, NormalizationMode] = field(default_factory=dict)
 
     input_features: dict[str, PolicyFeature] = field(default_factory=dict)
     output_features: dict[str, PolicyFeature] = field(default_factory=dict)

src/lerobot/configs/types.py

@@ -24,6 +24,12 @@ class FeatureType(str, Enum):
     ENV = "ENV"
     ACTION = "ACTION"
     REWARD = "REWARD"
+    LANGUAGE = "LANGUAGE"
+
+
+class PipelineFeatureType(str, Enum):
+    ACTION = "ACTION"
+    OBSERVATION = "OBSERVATION"
 
 
 class NormalizationMode(str, Enum):

src/lerobot/constants.py

@@ -21,8 +21,14 @@ OBS_ENV_STATE = "observation.environment_state"
 OBS_STATE = "observation.state"
 OBS_IMAGE = "observation.image"
 OBS_IMAGES = "observation.images"
+OBS_LANGUAGE = "observation.language"
 ACTION = "action"
 REWARD = "next.reward"
+TRUNCATED = "next.truncated"
+DONE = "next.done"
+
+OBS_LANGUAGE_TOKENS = "observation.language.tokens"
+OBS_LANGUAGE_ATTENTION_MASK = "observation.language.attention_mask"
 
 ROBOTS = "robots"
 ROBOT_TYPE = "robot_type"
@@ -39,6 +45,9 @@ OPTIMIZER_STATE = "optimizer_state.safetensors"
 OPTIMIZER_PARAM_GROUPS = "optimizer_param_groups.json"
 SCHEDULER_STATE = "scheduler_state.json"
 
+POLICY_PREPROCESSOR_DEFAULT_NAME = "policy_preprocessor"
+POLICY_POSTPROCESSOR_DEFAULT_NAME = "policy_postprocessor"
+
 if "LEROBOT_HOME" in os.environ:
     raise ValueError(
         f"You have a 'LEROBOT_HOME' environment variable set to '{os.getenv('LEROBOT_HOME')}'.\n"

src/lerobot/datasets/lerobot_dataset.py

@@ -825,6 +825,8 @@ class LeRobotDataset(torch.utils.data.Dataset):
         """
         if not episode_data:
             episode_buffer = self.episode_buffer
+        else:
+            episode_buffer = episode_data
 
         validate_episode_buffer(episode_buffer, self.meta.total_episodes, self.features)
 

src/lerobot/datasets/pipeline_features.py

@@ -0,0 +1,141 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import re
+from collections.abc import Sequence
+from typing import Any
+
+from lerobot.configs.types import PipelineFeatureType
+from lerobot.constants import ACTION, OBS_IMAGES, OBS_STATE
+from lerobot.datasets.utils import hw_to_dataset_features
+from lerobot.processor import DataProcessorPipeline
+
+
+def create_initial_features(
+    action: dict[str, Any] | None, observation: dict[str, Any] | None
+) -> dict[PipelineFeatureType, dict[str, Any]]:
+    """
+    Creates the initial features dict for the dataset from action and observation specs.
+
+    Args:
+        action: A dictionary of action feature names to their types/shapes.
+        observation: A dictionary of observation feature names to their types/shapes.
+
+    Returns:
+        The initial features dictionary structured by PipelineFeatureType.
+    """
+    features = {PipelineFeatureType.ACTION: {}, PipelineFeatureType.OBSERVATION: {}}
+    if action:
+        features[PipelineFeatureType.ACTION] = action
+    if observation:
+        features[PipelineFeatureType.OBSERVATION] = observation
+    return features
+
+
+# Helper to filter state/action keys based on regex patterns.
+def should_keep(key: str, patterns: tuple[str]) -> bool:
+    if patterns is None:
+        return True
+    return any(re.search(pat, key) for pat in patterns)
+
+
+def strip_prefix(key: str, prefixes_to_strip: tuple[str]) -> str:
+    for prefix in prefixes_to_strip:
+        if key.startswith(prefix):
+            return key[len(prefix) :]
+    return key
+
+
+# Define prefixes to strip from feature keys for clean names.
+# Handles both fully qualified (e.g., "action.state") and short (e.g., "state") forms.
+PREFIXES_TO_STRIP = tuple(
+    f"{token}." for const in (ACTION, OBS_STATE, OBS_IMAGES) for token in (const, const.split(".")[-1])
+)
+
+
+def aggregate_pipeline_dataset_features(
+    pipeline: DataProcessorPipeline,
+    initial_features: dict[PipelineFeatureType, dict[str, Any]],
+    *,
+    use_videos: bool = True,
+    patterns: Sequence[str] | None = None,
+) -> dict[str, dict]:
+    """
+    Aggregates and filters pipeline features to create a dataset-ready features dictionary.
+
+    This function transforms initial features using the pipeline, categorizes them as action or observations
+    (image or state), filters them based on `use_videos` and `patterns`, and finally
+    formats them for use with a Hugging Face LeRobot Dataset.
+
+    Args:
+        pipeline: The DataProcessorPipeline to apply.
+        initial_features: A dictionary of raw feature specs for actions and observations.
+        use_videos: If False, image features are excluded.
+        patterns: A sequence of regex patterns to filter action and state features.
+                  Image features are not affected by this filter.
+
+    Returns:
+        A dictionary of features formatted for a Hugging Face LeRobot Dataset.
+    """
+    all_features = pipeline.transform_features(initial_features)
+
+    # Intermediate storage for categorized and filtered features.
+    processed_features: dict[str, dict[str, Any]] = {
+        "action": {},
+        "observation": {},
+    }
+    images_token = OBS_IMAGES.split(".")[-1]
+
+    # Iterate through all features transformed by the pipeline.
+    for ptype, feats in all_features.items():
+        if ptype not in [PipelineFeatureType.ACTION, PipelineFeatureType.OBSERVATION]:
+            continue
+
+        for key, value in feats.items():
+            # 1. Categorize the feature.
+            is_action = ptype == PipelineFeatureType.ACTION
+            # Observations are classified as images if their key matches image-related tokens or if the shape of the feature is 3.
+            # All other observations are treated as state.
+            is_image = not is_action and (
+                (isinstance(value, tuple) and len(value) == 3)
+                or (
+                    key.startswith(f"{OBS_IMAGES}.")
+                    or key.startswith(f"{images_token}.")
+                    or f".{images_token}." in key
+                )
+            )
+
+            # 2. Apply filtering rules.
+            if is_image and not use_videos:
+                continue
+            if not is_image and not should_keep(key, patterns):
+                continue
+
+            # 3. Add the feature to the appropriate group with a clean name.
+            name = strip_prefix(key, PREFIXES_TO_STRIP)
+            if is_action:
+                processed_features["action"][name] = value
+            else:
+                processed_features["observation"][name] = value
+
+    # Convert the processed features into the final dataset format.
+    dataset_features = {}
+    if processed_features["action"]:
+        dataset_features.update(hw_to_dataset_features(processed_features["action"], ACTION, use_videos))
+    if processed_features["observation"]:
+        dataset_features.update(
+            hw_to_dataset_features(processed_features["observation"], "observation", use_videos)
+        )
+
+    return dataset_features

src/lerobot/datasets/utils.py

@@ -75,13 +75,20 @@ DEFAULT_FEATURES = {
 
 
 def flatten_dict(d: dict, parent_key: str = "", sep: str = "/") -> dict:
-    """Flatten a nested dictionary structure by collapsing nested keys into one key with a separator.
+    """Flatten a nested dictionary by joining keys with a separator.
 
-    For example:
-    ```
-    >>> dct = {"a": {"b": 1, "c": {"d": 2}}, "e": 3}`
-    >>> print(flatten_dict(dct))
-    {"a/b": 1, "a/c/d": 2, "e": 3}
+    Example:
+        >>> dct = {"a": {"b": 1, "c": {"d": 2}}, "e": 3}
+        >>> print(flatten_dict(dct))
+        {'a/b': 1, 'a/c/d': 2, 'e': 3}
+
+    Args:
+        d (dict): The dictionary to flatten.
+        parent_key (str): The base key to prepend to the keys in this level.
+        sep (str): The separator to use between keys.
+
+    Returns:
+        dict: A flattened dictionary.
     """
     items = []
     for k, v in d.items():
@@ -94,6 +101,20 @@ def flatten_dict(d: dict, parent_key: str = "", sep: str = "/") -> dict:
 
 
 def unflatten_dict(d: dict, sep: str = "/") -> dict:
+    """Unflatten a dictionary with delimited keys into a nested dictionary.
+
+    Example:
+        >>> flat_dct = {"a/b": 1, "a/c/d": 2, "e": 3}
+        >>> print(unflatten_dict(flat_dct))
+        {'a': {'b': 1, 'c': {'d': 2}}, 'e': 3}
+
+    Args:
+        d (dict): A dictionary with flattened keys.
+        sep (str): The separator used in the keys.
+
+    Returns:
+        dict: A nested dictionary.
+    """
     outdict = {}
     for key, value in d.items():
         parts = key.split(sep)
@@ -107,6 +128,16 @@ def unflatten_dict(d: dict, sep: str = "/") -> dict:
 
 
 def get_nested_item(obj: DictLike, flattened_key: str, sep: str = "/") -> Any:
+    """Access an item in a nested dictionary using a flattened key.
+
+    Args:
+        obj (DictLike): The nested dictionary-like object.
+        flattened_key (str): A key with parts separated by `sep`.
+        sep (str): The separator used in the flattened key.
+
+    Returns:
+        Any: The value from the nested dictionary.
+    """
     split_keys = flattened_key.split(sep)
     getter = obj[split_keys[0]]
     if len(split_keys) == 1:
@@ -119,6 +150,19 @@ def get_nested_item(obj: DictLike, flattened_key: str, sep: str = "/") -> Any:
 
 
 def serialize_dict(stats: dict[str, torch.Tensor | np.ndarray | dict]) -> dict:
+    """Serialize a dictionary containing tensors or numpy arrays to be JSON-compatible.
+
+    Converts torch.Tensor, np.ndarray, and np.generic types to lists or native Python types.
+
+    Args:
+        stats (dict): A dictionary that may contain non-serializable numeric types.
+
+    Returns:
+        dict: A dictionary with all values converted to JSON-serializable types.
+
+    Raises:
+        NotImplementedError: If a value has an unsupported type.
+    """
     serialized_dict = {}
     for key, value in flatten_dict(stats).items():
         if isinstance(value, (torch.Tensor, np.ndarray)):
@@ -133,6 +177,17 @@ def serialize_dict(stats: dict[str, torch.Tensor | np.ndarray | dict]) -> dict:
 
 
 def embed_images(dataset: datasets.Dataset) -> datasets.Dataset:
+    """Embed image bytes into the dataset table before saving to Parquet.
+
+    This function prepares a Hugging Face dataset for serialization by converting
+    image objects into an embedded format that can be stored in Arrow/Parquet.
+
+    Args:
+        dataset (datasets.Dataset): The input dataset, possibly containing image features.
+
+    Returns:
+        datasets.Dataset: The dataset with images embedded in the table storage.
+    """
     # Embed image bytes into the table before saving to parquet
     format = dataset.format
     dataset = dataset.with_format("arrow")
@@ -142,38 +197,94 @@ def embed_images(dataset: datasets.Dataset) -> datasets.Dataset:
 
 
 def load_json(fpath: Path) -> Any:
+    """Load data from a JSON file.
+
+    Args:
+        fpath (Path): Path to the JSON file.
+
+    Returns:
+        Any: The data loaded from the JSON file.
+    """
     with open(fpath) as f:
         return json.load(f)
 
 
 def write_json(data: dict, fpath: Path) -> None:
+    """Write data to a JSON file.
+
+    Creates parent directories if they don't exist.
+
+    Args:
+        data (dict): The dictionary to write.
+        fpath (Path): The path to the output JSON file.
+    """
     fpath.parent.mkdir(exist_ok=True, parents=True)
     with open(fpath, "w") as f:
         json.dump(data, f, indent=4, ensure_ascii=False)
 
 
 def load_jsonlines(fpath: Path) -> list[Any]:
+    """Load data from a JSON Lines file.
+
+    Args:
+        fpath (Path): Path to the JSON Lines file.
+
+    Returns:
+        list[Any]: A list of objects loaded from the file.
+    """
     with jsonlines.open(fpath, "r") as reader:
         return list(reader)
 
 
 def write_jsonlines(data: dict, fpath: Path) -> None:
+    """Write a list of dictionaries to a JSON Lines file.
+
+    Creates parent directories if they don't exist.
+
+    Args:
+        data (dict): The list of dictionaries to write.
+        fpath (Path): The path to the output JSON Lines file.
+    """
     fpath.parent.mkdir(exist_ok=True, parents=True)
     with jsonlines.open(fpath, "w") as writer:
         writer.write_all(data)
 
 
 def append_jsonlines(data: dict, fpath: Path) -> None:
+    """Append a dictionary to a JSON Lines file.
+
+    Creates parent directories if they don't exist.
+
+    Args:
+        data (dict): The dictionary to append.
+        fpath (Path): The path to the JSON Lines file.
+    """
     fpath.parent.mkdir(exist_ok=True, parents=True)
     with jsonlines.open(fpath, "a") as writer:
         writer.write(data)
 
 
 def write_info(info: dict, local_dir: Path):
+    """Write dataset info metadata to its standard file path.
+
+    Args:
+        info (dict): The dataset information dictionary.
+        local_dir (Path): The root directory of the dataset.
+    """
     write_json(info, local_dir / INFO_PATH)
 
 
 def load_info(local_dir: Path) -> dict:
+    """Load dataset info metadata from its standard file path.
+
+    Also converts shape lists to tuples for consistency.
+
+    Args:
+        local_dir (Path): The root directory of the dataset.
+
+    Returns:
+        dict: The dataset information dictionary.
+    """
     info = load_json(local_dir / INFO_PATH)
     for ft in info["features"].values():
         ft["shape"] = tuple(ft["shape"])
@@ -181,16 +292,40 @@ def load_info(local_dir: Path) -> dict:
 
 
 def write_stats(stats: dict, local_dir: Path):
+    """Serialize and write dataset statistics to their standard file path.
+
+    Args:
+        stats (dict): The statistics dictionary (can contain tensors/numpy arrays).
+        local_dir (Path): The root directory of the dataset.
+    """
     serialized_stats = serialize_dict(stats)
     write_json(serialized_stats, local_dir / STATS_PATH)
 
 
 def cast_stats_to_numpy(stats) -> dict[str, dict[str, np.ndarray]]:
+    """Recursively cast numerical values in a stats dictionary to numpy arrays.
+
+    Args:
+        stats (dict): The statistics dictionary.
+
+    Returns:
+        dict: The statistics dictionary with values cast to numpy arrays.
+    """
     stats = {key: np.array(value) for key, value in flatten_dict(stats).items()}
     return unflatten_dict(stats)
 
 
 def load_stats(local_dir: Path) -> dict[str, dict[str, np.ndarray]]:
+    """Load dataset statistics and cast numerical values to numpy arrays.
+
+    Returns None if the stats file doesn't exist.
+
+    Args:
+        local_dir (Path): The root directory of the dataset.
+
+    Returns:
+        A dictionary of statistics or None if the file is not found.
+    """
     if not (local_dir / STATS_PATH).exists():
         return None
     stats = load_json(local_dir / STATS_PATH)
@@ -198,6 +333,13 @@ def load_stats(local_dir: Path) -> dict[str, dict[str, np.ndarray]]:
 
 
 def write_task(task_index: int, task: dict, local_dir: Path):
+    """Write a single task to the tasks metadata file.
+
+    Args:
+        task_index (int): The index of the task.
+        task (dict): The task description dictionary.
+        local_dir (Path): The root directory of the dataset.
+    """
     task_dict = {
         "task_index": task_index,
         "task": task,
@@ -206,6 +348,16 @@ def write_task(task_index: int, task: dict, local_dir: Path):
 
 
 def load_tasks(local_dir: Path) -> tuple[dict, dict]:
+    """Load tasks from the tasks metadata file.
+
+    Args:
+        local_dir (Path): The root directory of the dataset.
+
+    Returns:
+        A tuple containing:
+        - A dictionary mapping task index to task description.
+        - A dictionary mapping task description to task index.
+    """
     tasks = load_jsonlines(local_dir / TASKS_PATH)
     tasks = {item["task_index"]: item["task"] for item in sorted(tasks, key=lambda x: x["task_index"])}
     task_to_task_index = {task: task_index for task_index, task in tasks.items()}
@@ -213,15 +365,36 @@ def load_tasks(local_dir: Path) -> tuple[dict, dict]:
 
 
 def write_episode(episode: dict, local_dir: Path):
+    """Write a single episode's metadata to the episodes metadata file.
+
+    Args:
+        episode (dict): The episode metadata dictionary.
+        local_dir (Path): The root directory of the dataset.
+    """
     append_jsonlines(episode, local_dir / EPISODES_PATH)
 
 
 def load_episodes(local_dir: Path) -> dict:
+    """Load episode metadata from the episodes metadata file.
+
+    Args:
+        local_dir (Path): The root directory of the dataset.
+
+    Returns:
+        dict: A dictionary mapping episode index to episode metadata.
+    """
     episodes = load_jsonlines(local_dir / EPISODES_PATH)
     return {item["episode_index"]: item for item in sorted(episodes, key=lambda x: x["episode_index"])}
 
 
 def write_episode_stats(episode_index: int, episode_stats: dict, local_dir: Path):
+    """Write statistics for a single episode to the episode stats file.
+
+    Args:
+        episode_index (int): The index of the episode.
+        episode_stats (dict): The statistics for the episode.
+        local_dir (Path): The root directory of the dataset.
+    """
     # We wrap episode_stats in a dictionary since `episode_stats["episode_index"]`
     # is a dictionary of stats and not an integer.
     episode_stats = {"episode_index": episode_index, "stats": serialize_dict(episode_stats)}
@@ -229,6 +402,14 @@ def write_episode_stats(episode_index: int, episode_stats: dict, local_dir: Path
 
 
 def load_episodes_stats(local_dir: Path) -> dict:
+    """Load per-episode statistics from the episode stats file.
+
+    Args:
+        local_dir (Path): The root directory of the dataset.
+
+    Returns:
+        dict: A dictionary mapping episode index to its statistics dictionary.
+    """
     episodes_stats = load_jsonlines(local_dir / EPISODES_STATS_PATH)
     return {
         item["episode_index"]: cast_stats_to_numpy(item["stats"])
@@ -239,12 +420,35 @@ def load_episodes_stats(local_dir: Path) -> dict:
 def backward_compatible_episodes_stats(
     stats: dict[str, dict[str, np.ndarray]], episodes: list[int]
 ) -> dict[str, dict[str, np.ndarray]]:
+    """Create a per-episode stats dictionary from a global stats dictionary.
+
+    This is used for backward compatibility with older datasets that only had global stats.
+
+    Args:
+        stats (dict): The global dataset statistics.
+        episodes (list[int]): A list of episode indices.
+
+    Returns:
+        dict: A dictionary mapping each episode index to the global stats.
+    """
     return dict.fromkeys(episodes, stats)
 
 
 def load_image_as_numpy(
     fpath: str | Path, dtype: np.dtype = np.float32, channel_first: bool = True
 ) -> np.ndarray:
+    """Load an image from a file into a numpy array.
+
+    Args:
+        fpath (str | Path): Path to the image file.
+        dtype (np.dtype): The desired data type of the output array. If floating,
+            pixels are scaled to [0, 1].
+        channel_first (bool): If True, converts the image to (C, H, W) format.
+            Otherwise, it remains in (H, W, C) format.
+
+    Returns:
+        np.ndarray: The image as a numpy array.
+    """
     img = PILImage.open(fpath).convert("RGB")
     img_array = np.array(img, dtype=dtype)
     if channel_first:  # (H, W, C) -> (C, H, W)
@@ -255,10 +459,19 @@ def load_image_as_numpy(
 
 
 def hf_transform_to_torch(items_dict: dict[torch.Tensor | None]):
-    """Get a transform function that convert items from Hugging Face dataset (pyarrow)
-    to torch tensors. Importantly, images are converted from PIL, which corresponds to
-    a channel last representation (h w c) of uint8 type, to a torch image representation
-    with channel first (c h w) of float32 type in range [0,1].
+    """Convert a batch from a Hugging Face dataset to torch tensors.
+
+    This transform function converts items from Hugging Face dataset format (pyarrow)
+    to torch tensors. Importantly, images are converted from PIL objects (H, W, C, uint8)
+    to a torch image representation (C, H, W, float32) in the range [0, 1]. Other
+    types are converted to torch.tensor.
+
+    Args:
+        items_dict (dict): A dictionary representing a batch of data from a
+            Hugging Face dataset.
+
+    Returns:
+        dict: The batch with items converted to torch tensors.
     """
     for key in items_dict:
         first_item = items_dict[key][0]
@@ -273,6 +486,14 @@ def hf_transform_to_torch(items_dict: dict[torch.Tensor | None]):
 
 
 def is_valid_version(version: str) -> bool:
+    """Check if a string is a valid PEP 440 version.
+
+    Args:
+        version (str): The version string to check.
+
+    Returns:
+        bool: True if the version string is valid, False otherwise.
+    """
     try:
         packaging.version.parse(version)
         return True
@@ -286,6 +507,18 @@ def check_version_compatibility(
     current_version: str | packaging.version.Version,
     enforce_breaking_major: bool = True,
 ) -> None:
+    """Check for version compatibility between a dataset and the current codebase.
+
+    Args:
+        repo_id (str): The repository ID for logging purposes.
+        version_to_check (str | packaging.version.Version): The version of the dataset.
+        current_version (str | packaging.version.Version): The current version of the codebase.
+        enforce_breaking_major (bool): If True, raise an error on major version mismatch.
+
+    Raises:
+        BackwardCompatibilityError: If the dataset version is from a newer, incompatible
+            major version of the codebase.
+    """
     v_check = (
         packaging.version.parse(version_to_check)
         if not isinstance(version_to_check, packaging.version.Version)
@@ -303,7 +536,14 @@ def check_version_compatibility(
 
 
 def get_repo_versions(repo_id: str) -> list[packaging.version.Version]:
-    """Returns available valid versions (branches and tags) on given repo."""
+    """Return available valid versions (branches and tags) on a given Hub repo.
+
+    Args:
+        repo_id (str): The repository ID on the Hugging Face Hub.
+
+    Returns:
+        list[packaging.version.Version]: A list of valid versions found.
+    """
     api = HfApi()
     repo_refs = api.list_repo_refs(repo_id, repo_type="dataset")
     repo_refs = [b.name for b in repo_refs.branches + repo_refs.tags]
@@ -316,9 +556,22 @@ def get_repo_versions(repo_id: str) -> list[packaging.version.Version]:
 
 
 def get_safe_version(repo_id: str, version: str | packaging.version.Version) -> str:
-    """
-    Returns the version if available on repo or the latest compatible one.
-    Otherwise, will throw a `CompatibilityError`.
+    """Return the specified version if available on repo, or the latest compatible one.
+
+    If the exact version is not found, it looks for the latest version with the
+    same major version number that is less than or equal to the target minor version.
+
+    Args:
+        repo_id (str): The repository ID on the Hugging Face Hub.
+        version (str | packaging.version.Version): The target version.
+
+    Returns:
+        str: The safe version string (e.g., "v1.2.3") to use as a revision.
+
+    Raises:
+        RevisionNotFoundError: If the repo has no version tags.
+        BackwardCompatibilityError: If only older major versions are available.
+        ForwardCompatibilityError: If only newer major versions are available.
     """
     target_version = (
         packaging.version.parse(version) if not isinstance(version, packaging.version.Version) else version
@@ -360,6 +613,17 @@ def get_safe_version(repo_id: str, version: str | packaging.version.Version) ->
 
 
 def get_hf_features_from_features(features: dict) -> datasets.Features:
+    """Convert a LeRobot features dictionary to a `datasets.Features` object.
+
+    Args:
+        features (dict): A LeRobot-style feature dictionary.
+
+    Returns:
+        datasets.Features: The corresponding Hugging Face `datasets.Features` object.
+
+    Raises:
+        ValueError: If a feature has an unsupported shape.
+    """
     hf_features = {}
     for key, ft in features.items():
         if ft["dtype"] == "video":
@@ -387,6 +651,14 @@ def get_hf_features_from_features(features: dict) -> datasets.Features:
 
 
 def _validate_feature_names(features: dict[str, dict]) -> None:
+    """Validate that feature names do not contain invalid characters.
+
+    Args:
+        features (dict): The LeRobot features dictionary.
+
+    Raises:
+        ValueError: If any feature name contains '/'.
+    """
     invalid_features = {name: ft for name, ft in features.items() if "/" in name}
     if invalid_features:
         raise ValueError(f"Feature names should not contain '/'. Found '/' in '{invalid_features}'.")
@@ -395,6 +667,22 @@ def _validate_feature_names(features: dict[str, dict]) -> None:
 def hw_to_dataset_features(
     hw_features: dict[str, type | tuple], prefix: str, use_video: bool = True
 ) -> dict[str, dict]:
+    """Convert hardware-specific features to a LeRobot dataset feature dictionary.
+
+    This function takes a dictionary describing hardware outputs (like joint states
+    or camera image shapes) and formats it into the standard LeRobot feature
+    specification.
+
+    Args:
+        hw_features (dict): Dictionary mapping feature names to their type (float for
+            joints) or shape (tuple for images).
+        prefix (str): The prefix to add to the feature keys (e.g., "observation"
+            or "action").
+        use_video (bool): If True, image features are marked as "video", otherwise "image".
+
+    Returns:
+        dict: A LeRobot features dictionary.
+    """
     features = {}
     joint_fts = {key: ftype for key, ftype in hw_features.items() if ftype is float}
     cam_fts = {key: shape for key, shape in hw_features.items() if isinstance(shape, tuple)}
@@ -427,6 +715,20 @@ def hw_to_dataset_features(
 def build_dataset_frame(
     ds_features: dict[str, dict], values: dict[str, Any], prefix: str
 ) -> dict[str, np.ndarray]:
+    """Construct a single data frame from raw values based on dataset features.
+
+    A "frame" is a dictionary containing all the data for a single timestep,
+    formatted as numpy arrays according to the feature specification.
+
+    Args:
+        ds_features (dict): The LeRobot dataset features dictionary.
+        values (dict): A dictionary of raw values from the hardware/environment.
+        prefix (str): The prefix to filter features by (e.g., "observation"
+            or "action").
+
+    Returns:
+        dict: A dictionary representing a single frame of data.
+    """
     frame = {}
     for key, ft in ds_features.items():
         if key in DEFAULT_FEATURES or not key.startswith(prefix):
@@ -440,6 +742,21 @@ def build_dataset_frame(
 
 
 def dataset_to_policy_features(features: dict[str, dict]) -> dict[str, PolicyFeature]:
+    """Convert dataset features to policy features.
+
+    This function transforms the dataset's feature specification into a format
+    that a policy can use, classifying features by type (e.g., visual, state,
+    action) and ensuring correct shapes (e.g., channel-first for images).
+
+    Args:
+        features (dict): The LeRobot dataset features dictionary.
+
+    Returns:
+        dict: A dictionary mapping feature keys to `PolicyFeature` objects.
+
+    Raises:
+        ValueError: If an image feature does not have a 3D shape.
+    """
     # TODO(aliberts): Implement "type" in dataset features and simplify this
     policy_features = {}
     for key, ft in features.items():
@@ -470,6 +787,58 @@ def dataset_to_policy_features(features: dict[str, dict]) -> dict[str, PolicyFea
     return policy_features
 
 
+def combine_feature_dicts(*dicts: dict) -> dict:
+    """Merge LeRobot grouped feature dicts.
+
+    - For 1D numeric specs (dtype not image/video/string) with "names": we merge the names and recompute the shape.
+    - For others (e.g. `observation.images.*`), the last one wins (if they are identical).
+
+    Args:
+        *dicts: A variable number of LeRobot feature dictionaries to merge.
+
+    Returns:
+        dict: A single merged feature dictionary.
+
+    Raises:
+        ValueError: If there's a dtype mismatch for a feature being merged.
+    """
+    out: dict = {}
+    for d in dicts:
+        for key, value in d.items():
+            if not isinstance(value, dict):
+                out[key] = value
+                continue
+
+            dtype = value.get("dtype")
+            shape = value.get("shape")
+            is_vector = (
+                dtype not in ("image", "video", "string")
+                and isinstance(shape, tuple)
+                and len(shape) == 1
+                and "names" in value
+            )
+
+            if is_vector:
+                # Initialize or retrieve the accumulating dict for this feature key
+                target = out.setdefault(key, {"dtype": dtype, "names": [], "shape": (0,)})
+                # Ensure consistent data types across merged entries
+                if "dtype" in target and dtype != target["dtype"]:
+                    raise ValueError(f"dtype mismatch for '{key}': {target['dtype']} vs {dtype}")
+
+                # Merge feature names: append only new ones to preserve order without duplicates
+                seen = set(target["names"])
+                for n in value["names"]:
+                    if n not in seen:
+                        target["names"].append(n)
+                        seen.add(n)
+                # Recompute the shape to reflect the updated number of features
+                target["shape"] = (len(target["names"]),)
+            else:
+                # For images/videos and non-1D entries: override with the latest definition
+                out[key] = value
+    return out
+
+
 def create_empty_dataset_info(
     codebase_version: str,
     fps: int,
@@ -477,6 +846,18 @@ def create_empty_dataset_info(
     use_videos: bool,
     robot_type: str | None = None,
 ) -> dict:
+    """Create a template dictionary for a new dataset's `info.json`.
+
+    Args:
+        codebase_version (str): The version of the LeRobot codebase.
+        fps (int): The frames per second of the data.
+        features (dict): The LeRobot features dictionary for the dataset.
+        use_videos (bool): Whether the dataset will store videos.
+        robot_type (str | None): The type of robot used, if any.
+
+    Returns:
+        dict: A dictionary with the initial dataset metadata.
+    """
     return {
         "codebase_version": codebase_version,
         "robot_type": robot_type,
@@ -497,6 +878,18 @@ def create_empty_dataset_info(
 def get_episode_data_index(
     episode_dicts: dict[dict], episodes: list[int] | None = None
 ) -> dict[str, torch.Tensor]:
+    """Calculate the start and end indices for each episode in a flattened dataset.
+
+    Args:
+        episode_dicts (dict): A dictionary mapping episode index to episode metadata,
+            which must contain a "length" key.
+        episodes (list[int] | None): An optional list of episode indices to consider.
+            If None, all episodes are used.
+
+    Returns:
+        dict: A dictionary with "from" and "to" keys, containing torch tensors
+            with the start and end indices for each episode.
+    """
     episode_lengths = {ep_idx: ep_dict["length"] for ep_idx, ep_dict in episode_dicts.items()}
     if episodes is not None:
         episode_lengths = {ep_idx: episode_lengths[ep_idx] for ep_idx in episodes}
@@ -516,16 +909,19 @@ def check_timestamps_sync(
     tolerance_s: float,
     raise_value_error: bool = True,
 ) -> bool:
-    """
-    This check is to make sure that each timestamp is separated from the next by (1/fps) +/- tolerance
-    to account for possible numerical error.
+    """Check if timestamps are separated by (1/fps) +/- tolerance.
+
+    This check ensures that consecutive timestamps within an episode are spaced
+    correctly, accounting for possible numerical errors. It ignores the boundaries
+    between episodes.
 
     Args:
         timestamps (np.ndarray): Array of timestamps in seconds.
         episode_indices (np.ndarray): Array indicating the episode index for each timestamp.
-        episode_data_index (dict[str, np.ndarray]): A dictionary that includes 'to',
+        episode_data_index (dict): A dictionary that includes 'to',
             which identifies indices for the end of each episode.
-        fps (int): Frames per second. Used to check the expected difference between consecutive timestamps.
+        fps (int): Frames per second. Used to check the expected difference between
+            consecutive timestamps.
         tolerance_s (float): Allowed deviation from the expected (1/fps) difference.
         raise_value_error (bool): Whether to raise a ValueError if the check fails.
 
@@ -533,7 +929,8 @@ def check_timestamps_sync(
         bool: True if all checked timestamp differences lie within tolerance, False otherwise.
 
     Raises:
-        ValueError: If the check fails and `raise_value_error` is True.
+        ValueError: If `timestamps` and `episode_indices` shapes do not match, or if
+            the check fails and `raise_value_error` is True.
     """
     if timestamps.shape != episode_indices.shape:
         raise ValueError(
@@ -584,9 +981,23 @@ def check_timestamps_sync(
 def check_delta_timestamps(
     delta_timestamps: dict[str, list[float]], fps: int, tolerance_s: float, raise_value_error: bool = True
 ) -> bool:
-    """This will check if all the values in delta_timestamps are multiples of 1/fps +/- tolerance.
-    This is to ensure that these delta_timestamps added to any timestamp from a dataset will themselves be
-    actual timestamps from the dataset.
+    """Check if delta timestamps are multiples of 1/fps +/- tolerance.
+
+    This ensures that adding these delta timestamps to any existing timestamp in
+    the dataset will result in a value that aligns with the dataset's frame rate.
+
+    Args:
+        delta_timestamps (dict): A dictionary where values are lists of time
+            deltas in seconds.
+        fps (int): The frames per second of the dataset.
+        tolerance_s (float): The allowed tolerance in seconds.
+        raise_value_error (bool): If True, raises an error on failure.
+
+    Returns:
+        bool: True if all deltas are valid, False otherwise.
+
+    Raises:
+        ValueError: If any delta is outside the tolerance and `raise_value_error` is True.
     """
     outside_tolerance = {}
     for key, delta_ts in delta_timestamps.items():
@@ -612,6 +1023,15 @@ def check_delta_timestamps(
 
 
 def get_delta_indices(delta_timestamps: dict[str, list[float]], fps: int) -> dict[str, list[int]]:
+    """Convert delta timestamps in seconds to delta indices in frames.
+
+    Args:
+        delta_timestamps (dict): A dictionary of time deltas in seconds.
+        fps (int): The frames per second of the dataset.
+
+    Returns:
+        dict: A dictionary of frame delta indices.
+    """
     delta_indices = {}
     for key, delta_ts in delta_timestamps.items():
         delta_indices[key] = [round(d * fps) for d in delta_ts]
@@ -620,9 +1040,17 @@ def get_delta_indices(delta_timestamps: dict[str, list[float]], fps: int) -> dic
 
 
 def cycle(iterable):
-    """The equivalent of itertools.cycle, but safe for Pytorch dataloaders.
+    """Create a dataloader-safe cyclical iterator.
 
-    See https://github.com/pytorch/pytorch/issues/23900 for information on why itertools.cycle is not safe.
+    This is an equivalent of `itertools.cycle` but is safe for use with
+    PyTorch DataLoaders with multiple workers.
+    See https://github.com/pytorch/pytorch/issues/23900 for details.
+
+    Args:
+        iterable: The iterable to cycle over.
+
+    Yields:
+        Items from the iterable, restarting from the beginning when exhausted.
     """
     iterator = iter(iterable)
     while True:
@@ -633,8 +1061,14 @@ def cycle(iterable):
 
 
 def create_branch(repo_id, *, branch: str, repo_type: str | None = None) -> None:
-    """Create a branch on a existing Hugging Face repo. Delete the branch if it already
-    exists before creating it.
+    """Create a branch on an existing Hugging Face repo.
+
+    Deletes the branch if it already exists before creating it.
+
+    Args:
+        repo_id (str): The ID of the repository.
+        branch (str): The name of the branch to create.
+        repo_type (str | None): The type of the repository (e.g., "dataset").
     """
     api = HfApi()
 
@@ -652,9 +1086,20 @@ def create_lerobot_dataset_card(
     dataset_info: dict | None = None,
     **kwargs,
 ) -> DatasetCard:
-    """
-    Keyword arguments will be used to replace values in src/lerobot/datasets/card_template.md.
-    Note: If specified, license must be one of https://huggingface.co/docs/hub/repositories-licenses.
+    """Create a `DatasetCard` for a LeRobot dataset.
+
+    Keyword arguments are used to replace values in the card template.
+    Note: If specified, `license` must be a valid license identifier from
+    https://huggingface.co/docs/hub/repositories-licenses.
+
+    Args:
+        tags (list | None): A list of tags to add to the dataset card.
+        dataset_info (dict | None): The dataset's info dictionary, which will
+            be displayed on the card.
+        **kwargs: Additional keyword arguments to populate the card template.
+
+    Returns:
+        DatasetCard: The generated dataset card object.
     """
     card_tags = ["LeRobot"]
 
@@ -686,19 +1131,16 @@ def create_lerobot_dataset_card(
 
 
 class IterableNamespace(SimpleNamespace):
-    """
-    A namespace object that supports both dictionary-like iteration and dot notation access.
-    Automatically converts nested dictionaries into IterableNamespaces.
+    """A namespace object that supports both dictionary-like iteration and dot notation.
 
-    This class extends SimpleNamespace to provide:
-    - Dictionary-style iteration over keys
-    - Access to items via both dot notation (obj.key) and brackets (obj["key"])
-    - Dictionary-like methods: items(), keys(), values()
-    - Recursive conversion of nested dictionaries
+    This class extends `SimpleNamespace` to provide dictionary-style iteration,
+    access to items via brackets (`obj["key"]`), and dictionary-like methods
+    (`items()`, `keys()`, `values()`). Nested dictionaries are recursively
+    converted to `IterableNamespace` objects.
 
     Args:
-        dictionary: Optional dictionary to initialize the namespace
-        **kwargs: Additional keyword arguments passed to SimpleNamespace
+        dictionary (dict, optional): A dictionary to initialize the namespace with.
+        **kwargs: Additional keyword arguments to initialize the namespace.
 
     Examples:
         >>> data = {"name": "Alice", "details": {"age": 25}}
@@ -712,10 +1154,16 @@ class IterableNamespace(SimpleNamespace):
         >>> for key, value in ns.items():
         ...     print(f"{key}: {value}")
         name: Alice
-        details: IterableNamespace(age=25)
+        details: <__main__.IterableNamespace object at ...>
     """
 
     def __init__(self, dictionary: dict[str, Any] = None, **kwargs):
+        """Initialize the IterableNamespace.
+
+        Args:
+            dictionary (dict, optional): Dictionary to populate the namespace.
+            **kwargs: Keyword arguments to populate the namespace.
+        """
         super().__init__(**kwargs)
         if dictionary is not None:
             for key, value in dictionary.items():
@@ -725,22 +1173,46 @@ class IterableNamespace(SimpleNamespace):
                     setattr(self, key, value)
 
     def __iter__(self) -> Iterator[str]:
+        """Return an iterator over the keys of the namespace."""
         return iter(vars(self))
 
     def __getitem__(self, key: str) -> Any:
+        """Allow bracket-style access to attributes.
+
+        Args:
+            key (str): The name of the attribute.
+
+        Returns:
+            Any: The value of the attribute.
+        """
         return vars(self)[key]
 
     def items(self):
+        """Return a view of the namespace's (key, value) pairs."""
         return vars(self).items()
 
     def values(self):
+        """Return a view of the namespace's values."""
         return vars(self).values()
 
     def keys(self):
+        """Return a view of the namespace's keys."""
         return vars(self).keys()
 
 
 def validate_frame(frame: dict, features: dict):
+    """Validate a single data frame against the dataset's feature specification.
+
+    Checks for missing/extra features, and validates the dtype and shape of each
+    provided feature.
+
+    Args:
+        frame (dict): The data frame to validate.
+        features (dict): The LeRobot features dictionary for the dataset.
+
+    Raises:
+        ValueError: If the frame does not match the feature specification.
+    """
     expected_features = set(features) - set(DEFAULT_FEATURES)
     actual_features = set(frame)
 
@@ -755,6 +1227,15 @@ def validate_frame(frame: dict, features: dict):
 
 
 def validate_features_presence(actual_features: set[str], expected_features: set[str]):
+    """Check for missing or extra features in a frame.
+
+    Args:
+        actual_features (set[str]): The set of feature names present in the frame.
+        expected_features (set[str]): The set of feature names expected in the frame.
+
+    Returns:
+        str: An error message string if there's a mismatch, otherwise an empty string.
+    """
     error_message = ""
     missing_features = expected_features - actual_features
     extra_features = actual_features - expected_features
@@ -770,6 +1251,19 @@ def validate_features_presence(actual_features: set[str], expected_features: set
 
 
 def validate_feature_dtype_and_shape(name: str, feature: dict, value: np.ndarray | PILImage.Image | str):
+    """Validate the dtype and shape of a single feature's value.
+
+    Args:
+        name (str): The name of the feature.
+        feature (dict): The feature specification from the LeRobot features dictionary.
+        value: The value of the feature to validate.
+
+    Returns:
+        str: An error message if validation fails, otherwise an empty string.
+
+    Raises:
+        NotImplementedError: If the feature dtype is not supported for validation.
+    """
     expected_dtype = feature["dtype"]
     expected_shape = feature["shape"]
     if is_valid_numpy_dtype_string(expected_dtype):
@@ -785,6 +1279,17 @@ def validate_feature_dtype_and_shape(name: str, feature: dict, value: np.ndarray
 def validate_feature_numpy_array(
     name: str, expected_dtype: str, expected_shape: list[int], value: np.ndarray
 ):
+    """Validate a feature that is expected to be a numpy array.
+
+    Args:
+        name (str): The name of the feature.
+        expected_dtype (str): The expected numpy dtype as a string.
+        expected_shape (list[int]): The expected shape.
+        value (np.ndarray): The numpy array to validate.
+
+    Returns:
+        str: An error message if validation fails, otherwise an empty string.
+    """
     error_message = ""
     if isinstance(value, np.ndarray):
         actual_dtype = value.dtype
@@ -802,6 +1307,18 @@ def validate_feature_numpy_array(
 
 
 def validate_feature_image_or_video(name: str, expected_shape: list[str], value: np.ndarray | PILImage.Image):
+    """Validate a feature that is expected to be an image or video frame.
+
+    Accepts `np.ndarray` (channel-first or channel-last) or `PIL.Image.Image`.
+
+    Args:
+        name (str): The name of the feature.
+        expected_shape (list[str]): The expected shape (C, H, W).
+        value: The image data to validate.
+
+    Returns:
+        str: An error message if validation fails, otherwise an empty string.
+    """
     # Note: The check of pixels range ([0,1] for float and [0,255] for uint8) is done by the image writer threads.
     error_message = ""
     if isinstance(value, np.ndarray):
@@ -818,12 +1335,35 @@ def validate_feature_image_or_video(name: str, expected_shape: list[str], value:
 
 
 def validate_feature_string(name: str, value: str):
+    """Validate a feature that is expected to be a string.
+
+    Args:
+        name (str): The name of the feature.
+        value (str): The value to validate.
+
+    Returns:
+        str: An error message if validation fails, otherwise an empty string.
+    """
     if not isinstance(value, str):
         return f"The feature '{name}' is expected to be of type 'str', but type '{type(value)}' provided instead.\n"
     return ""
 
 
 def validate_episode_buffer(episode_buffer: dict, total_episodes: int, features: dict):
+    """Validate the episode buffer before it's written to disk.
+
+    Ensures the buffer has the required keys, contains at least one frame, and
+    has features consistent with the dataset's specification.
+
+    Args:
+        episode_buffer (dict): The buffer containing data for a single episode.
+        total_episodes (int): The current total number of episodes in the dataset.
+        features (dict): The LeRobot features dictionary for the dataset.
+
+    Raises:
+        ValueError: If the buffer is invalid.
+        NotImplementedError: If the episode index is manually set and doesn't match.
+    """
     if "size" not in episode_buffer:
         raise ValueError("size key not found in episode_buffer")
 

src/lerobot/envs/configs.py

@@ -161,35 +161,73 @@ class XarmEnv(EnvConfig):
 
 
 @dataclass
-class VideoRecordConfig:
-    """Configuration for video recording in ManiSkill environments."""
-
-    enabled: bool = False
-    record_dir: str = "videos"
-    trajectory_name: str = "trajectory"
+class ImagePreprocessingConfig:
+    crop_params_dict: dict[str, tuple[int, int, int, int]] | None = None
+    resize_size: tuple[int, int] | None = None
 
 
 @dataclass
-class EnvTransformConfig:
-    """Configuration for environment wrappers."""
+class RewardClassifierConfig:
+    """Configuration for reward classification."""
+
+    pretrained_path: str | None = None
+    success_threshold: float = 0.5
+    success_reward: float = 1.0
+
+
+@dataclass
+class InverseKinematicsConfig:
+    """Configuration for inverse kinematics processing."""
+
+    urdf_path: str | None = None
+    target_frame_name: str | None = None
+    end_effector_bounds: dict[str, list[float]] | None = None
+    end_effector_step_sizes: dict[str, float] | None = None
+
+
+@dataclass
+class ObservationConfig:
+    """Configuration for observation processing."""
 
-    # ee_action_space_params: EEActionSpaceConfig = field(default_factory=EEActionSpaceConfig)
-    control_mode: str = "gamepad"
-    display_cameras: bool = False
     add_joint_velocity_to_observation: bool = False
     add_current_to_observation: bool = False
     add_ee_pose_to_observation: bool = False
-    crop_params_dict: dict[str, tuple[int, int, int, int]] | None = None
-    resize_size: tuple[int, int] | None = None
-    control_time_s: float = 20.0
-    fixed_reset_joint_positions: Any | None = None
-    reset_time_s: float = 5.0
+    display_cameras: bool = False
+
+
+@dataclass
+class GripperConfig:
+    """Configuration for gripper control and penalties."""
+
     use_gripper: bool = True
-    gripper_quantization_threshold: float | None = 0.8
     gripper_penalty: float = 0.0
     gripper_penalty_in_reward: bool = False
 
 
+@dataclass
+class ResetConfig:
+    """Configuration for environment reset behavior."""
+
+    fixed_reset_joint_positions: Any | None = None
+    reset_time_s: float = 5.0
+    control_time_s: float = 20.0
+    terminate_on_success: bool = True
+
+
+@dataclass
+class HILSerlProcessorConfig:
+    """Configuration for environment processing pipeline."""
+
+    control_mode: str = "gamepad"
+    observation: ObservationConfig | None = None
+    image_preprocessing: ImagePreprocessingConfig | None = None
+    gripper: GripperConfig | None = None
+    reset: ResetConfig | None = None
+    inverse_kinematics: InverseKinematicsConfig | None = None
+    reward_classifier: RewardClassifierConfig | None = None
+    max_gripper_pos: float | None = 100.0
+
+
 @EnvConfig.register_subclass(name="gym_manipulator")
 @dataclass
 class HILSerlRobotEnvConfig(EnvConfig):
@@ -197,77 +235,10 @@ class HILSerlRobotEnvConfig(EnvConfig):
 
     robot: RobotConfig | None = None
     teleop: TeleoperatorConfig | None = None
-    wrapper: EnvTransformConfig | None = None
-    fps: int = 10
+    processor: HILSerlProcessorConfig = field(default_factory=HILSerlProcessorConfig)
+
     name: str = "real_robot"
-    mode: str | None = None  # Either "record", "replay", None
-    repo_id: str | None = None
-    dataset_root: str | None = None
-    task: str | None = ""
-    num_episodes: int = 10  # only for record mode
-    episode: int = 0
-    device: str = "cuda"
-    push_to_hub: bool = True
-    pretrained_policy_name_or_path: str | None = None
-    reward_classifier_pretrained_path: str | None = None
-    # For the reward classifier, to record more positive examples after a success
-    number_of_steps_after_success: int = 0
 
     @property
     def gym_kwargs(self) -> dict:
         return {}
-
-
-@EnvConfig.register_subclass("hil")
-@dataclass
-class HILEnvConfig(EnvConfig):
-    """Configuration for the HIL environment."""
-
-    name: str = "PandaPickCube"
-    task: str | None = "PandaPickCubeKeyboard-v0"
-    use_viewer: bool = True
-    gripper_penalty: float = 0.0
-    use_gamepad: bool = True
-    state_dim: int = 18
-    action_dim: int = 4
-    fps: int = 100
-    episode_length: int = 100
-    video_record: VideoRecordConfig = field(default_factory=VideoRecordConfig)
-    features: dict[str, PolicyFeature] = field(
-        default_factory=lambda: {
-            "action": PolicyFeature(type=FeatureType.ACTION, shape=(4,)),
-            "observation.image": PolicyFeature(type=FeatureType.VISUAL, shape=(3, 128, 128)),
-            "observation.state": PolicyFeature(type=FeatureType.STATE, shape=(18,)),
-        }
-    )
-    features_map: dict[str, str] = field(
-        default_factory=lambda: {
-            "action": ACTION,
-            "observation.image": OBS_IMAGE,
-            "observation.state": OBS_STATE,
-        }
-    )
-    ################# args from hilserlrobotenv
-    reward_classifier_pretrained_path: str | None = None
-    robot_config: RobotConfig | None = None
-    teleop_config: TeleoperatorConfig | None = None
-    wrapper: EnvTransformConfig | None = None
-    mode: str | None = None  # Either "record", "replay", None
-    repo_id: str | None = None
-    dataset_root: str | None = None
-    num_episodes: int = 10  # only for record mode
-    episode: int = 0
-    device: str = "cuda"
-    push_to_hub: bool = True
-    pretrained_policy_name_or_path: str | None = None
-    # For the reward classifier, to record more positive examples after a success
-    number_of_steps_after_success: int = 0
-    ############################
-
-    @property
-    def gym_kwargs(self) -> dict:
-        return {
-            "use_viewer": self.use_viewer,
-            "use_gamepad": self.use_gamepad,
-            "gripper_penalty": self.gripper_penalty,
-        }

src/lerobot/envs/factory.py

@@ -17,7 +17,7 @@ import importlib
 
 import gymnasium as gym
 
-from lerobot.envs.configs import AlohaEnv, EnvConfig, HILEnvConfig, PushtEnv, XarmEnv
+from lerobot.envs.configs import AlohaEnv, EnvConfig, PushtEnv, XarmEnv
 
 
 def make_env_config(env_type: str, **kwargs) -> EnvConfig:
@@ -27,8 +27,6 @@ def make_env_config(env_type: str, **kwargs) -> EnvConfig:
         return PushtEnv(**kwargs)
     elif env_type == "xarm":
         return XarmEnv(**kwargs)
-    elif env_type == "hil":
-        return HILEnvConfig(**kwargs)
     else:
         raise ValueError(f"Policy type '{env_type}' is not available.")
 

src/lerobot/envs/utils.py

@@ -127,9 +127,29 @@ def check_env_attributes_and_types(env: gym.vector.VectorEnv) -> None:
 def add_envs_task(env: gym.vector.VectorEnv, observation: dict[str, Any]) -> dict[str, Any]:
     """Adds task feature to the observation dict with respect to the first environment attribute."""
     if hasattr(env.envs[0], "task_description"):
-        observation["task"] = env.call("task_description")
+        task_result = env.call("task_description")
+
+        if isinstance(task_result, tuple):
+            task_result = list(task_result)
+
+        if not isinstance(task_result, list):
+            raise TypeError(f"Expected task_description to return a list, got {type(task_result)}")
+        if not all(isinstance(item, str) for item in task_result):
+            raise TypeError("All items in task_description result must be strings")
+
+        observation["task"] = task_result
     elif hasattr(env.envs[0], "task"):
-        observation["task"] = env.call("task")
+        task_result = env.call("task")
+
+        if isinstance(task_result, tuple):
+            task_result = list(task_result)
+
+        if not isinstance(task_result, list):
+            raise TypeError(f"Expected task to return a list, got {type(task_result)}")
+        if not all(isinstance(item, str) for item in task_result):
+            raise TypeError("All items in task result must be strings")
+
+        observation["task"] = task_result
     else:  #  For envs without language instructions, e.g. aloha transfer cube and etc.
         num_envs = observation[list(observation.keys())[0]].shape[0]
         observation["task"] = ["" for _ in range(num_envs)]

src/lerobot/find_cameras.py

@@ -20,7 +20,7 @@ Helper to find the camera devices available in your system.
 Example:
 
 ```shell
-python -m lerobot.find_cameras
+lerobot-find-cameras
 ```
 """
 

src/lerobot/find_port.py

@@ -18,7 +18,7 @@ Helper to find the USB port associated with your MotorsBus.
 Example:
 
 ```shell
-python -m lerobot.find_port
+lerobot-find-port
 ```
 """
 

src/lerobot/motors/dynamixel/tables.py

@@ -107,6 +107,8 @@ X_SERIES_ENCODINGS_TABLE = {
     "Goal_PWM": X_SERIES_CONTROL_TABLE["Goal_PWM"][1],
     "Goal_Current": X_SERIES_CONTROL_TABLE["Goal_Current"][1],
     "Goal_Velocity": X_SERIES_CONTROL_TABLE["Goal_Velocity"][1],
+    "Goal_Position": X_SERIES_CONTROL_TABLE["Goal_Position"][1],
+    "Present_Position": X_SERIES_CONTROL_TABLE["Present_Position"][1],
     "Present_PWM": X_SERIES_CONTROL_TABLE["Present_PWM"][1],
     "Present_Current": X_SERIES_CONTROL_TABLE["Present_Current"][1],
     "Present_Velocity": X_SERIES_CONTROL_TABLE["Present_Velocity"][1],

src/lerobot/motors/motors_bus.py

@@ -222,7 +222,7 @@ class MotorsBus(abc.ABC):
     A MotorsBus subclass instance requires a port (e.g. `FeetechMotorsBus(port="/dev/tty.usbmodem575E0031751"`)).
     To find the port, you can run our utility script:
     ```bash
-    python -m lerobot.find_port.py
+    lerobot-find-port.py
     >>> Finding all available ports for the MotorsBus.
     >>> ["/dev/tty.usbmodem575E0032081", "/dev/tty.usbmodem575E0031751"]
     >>> Remove the usb cable from your MotorsBus and press Enter when done.
@@ -446,7 +446,7 @@ class MotorsBus(abc.ABC):
         except (FileNotFoundError, OSError, serial.SerialException) as e:
             raise ConnectionError(
                 f"\nCould not connect on port '{self.port}'. Make sure you are using the correct port."
-                "\nTry running `python -m lerobot.find_port`\n"
+                "\nTry running `lerobot-find-port`\n"
             ) from e
 
     @abc.abstractmethod

src/lerobot/policies/__init__.py

@@ -15,6 +15,17 @@
 from .act.configuration_act import ACTConfig as ACTConfig
 from .diffusion.configuration_diffusion import DiffusionConfig as DiffusionConfig
 from .pi0.configuration_pi0 import PI0Config as PI0Config
+from .pi0.processor_pi0 import Pi0NewLineProcessor
 from .smolvla.configuration_smolvla import SmolVLAConfig as SmolVLAConfig
+from .smolvla.processor_smolvla import SmolVLANewLineProcessor
 from .tdmpc.configuration_tdmpc import TDMPCConfig as TDMPCConfig
 from .vqbet.configuration_vqbet import VQBeTConfig as VQBeTConfig
+
+__all__ = [
+    "ACTConfig",
+    "DiffusionConfig",
+    "PI0Config",
+    "SmolVLAConfig",
+    "TDMPCConfig",
+    "VQBeTConfig",
+]

src/lerobot/policies/act/modeling_act.py

@@ -35,7 +35,6 @@ from torchvision.ops.misc import FrozenBatchNorm2d
 
 from lerobot.constants import ACTION, OBS_IMAGES
 from lerobot.policies.act.configuration_act import ACTConfig
-from lerobot.policies.normalize import Normalize, Unnormalize
 from lerobot.policies.pretrained import PreTrainedPolicy
 
 
@@ -51,27 +50,16 @@ class ACTPolicy(PreTrainedPolicy):
     def __init__(
         self,
         config: ACTConfig,
-        dataset_stats: dict[str, dict[str, Tensor]] | None = None,
     ):
         """
         Args:
             config: Policy configuration class instance or None, in which case the default instantiation of
                     the configuration class is used.
-            dataset_stats: Dataset statistics to be used for normalization. If not passed here, it is expected
-                that they will be passed with a call to `load_state_dict` before the policy is used.
         """
         super().__init__(config)
         config.validate_features()
         self.config = config
 
-        self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
-        self.normalize_targets = Normalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-        self.unnormalize_outputs = Unnormalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-
         self.model = ACT(config)
 
         if config.temporal_ensemble_coeff is not None:
@@ -137,23 +125,19 @@ class ACTPolicy(PreTrainedPolicy):
         """Predict a chunk of actions given environment observations."""
         self.eval()
 
-        batch = self.normalize_inputs(batch)
         if self.config.image_features:
             batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
             batch[OBS_IMAGES] = [batch[key] for key in self.config.image_features]
 
         actions = self.model(batch)[0]
-        actions = self.unnormalize_outputs({ACTION: actions})[ACTION]
         return actions
 
     def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict]:
         """Run the batch through the model and compute the loss for training or validation."""
-        batch = self.normalize_inputs(batch)
         if self.config.image_features:
             batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
             batch[OBS_IMAGES] = [batch[key] for key in self.config.image_features]
 
-        batch = self.normalize_targets(batch)
         actions_hat, (mu_hat, log_sigma_x2_hat) = self.model(batch)
 
         l1_loss = (

src/lerobot/policies/act/processor_act.py

@@ -0,0 +1,89 @@
+#!/usr/bin/env python
+
+# Copyright 2024 Tony Z. Zhao and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import torch
+
+from lerobot.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
+from lerobot.policies.act.configuration_act import ACTConfig
+from lerobot.processor import (
+    AddBatchDimensionProcessorStep,
+    DeviceProcessorStep,
+    NormalizerProcessorStep,
+    PolicyProcessorPipeline,
+    ProcessorKwargs,
+    RenameObservationsProcessorStep,
+    UnnormalizerProcessorStep,
+)
+
+
+def make_act_pre_post_processors(
+    config: ACTConfig,
+    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
+    preprocessor_kwargs: ProcessorKwargs | None = None,
+    postprocessor_kwargs: ProcessorKwargs | None = None,
+) -> tuple[PolicyProcessorPipeline, PolicyProcessorPipeline]:
+    """Creates the pre- and post-processing pipelines for the ACT policy.
+
+    The pre-processing pipeline handles normalization, batching, and device placement for the model inputs.
+    The post-processing pipeline handles unnormalization and moves the model outputs back to the CPU.
+
+    Args:
+        config (ACTConfig): The ACT policy configuration object.
+        dataset_stats (dict[str, dict[str, torch.Tensor]] | None): A dictionary containing dataset
+            statistics (e.g., mean and std) used for normalization. Defaults to None.
+        preprocessor_kwargs (ProcessorKwargs | None): Extra keyword arguments to pass to the
+            preprocessor pipeline's constructor. Defaults to None.
+        postprocessor_kwargs (ProcessorKwargs | None): Extra keyword arguments to pass to the
+            postprocessor pipeline's constructor. Defaults to None.
+
+    Returns:
+        tuple[PolicyProcessorPipeline, PolicyProcessorPipeline]: A tuple containing the
+        pre-processor pipeline and the post-processor pipeline.
+    """
+    if preprocessor_kwargs is None:
+        preprocessor_kwargs = {}
+    if postprocessor_kwargs is None:
+        postprocessor_kwargs = {}
+
+    input_steps = [
+        RenameObservationsProcessorStep(rename_map={}),
+        AddBatchDimensionProcessorStep(),
+        DeviceProcessorStep(device=config.device),
+        NormalizerProcessorStep(
+            features={**config.input_features, **config.output_features},
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+            device=config.device,
+        ),
+    ]
+    output_steps = [
+        DeviceProcessorStep(device="cpu"),
+        UnnormalizerProcessorStep(
+            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
+        ),
+    ]
+
+    return (
+        PolicyProcessorPipeline(
+            steps=input_steps,
+            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
+            **preprocessor_kwargs,
+        ),
+        PolicyProcessorPipeline(
+            steps=output_steps,
+            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
+            **postprocessor_kwargs,
+        ),
+    )

src/lerobot/policies/diffusion/modeling_diffusion.py

@@ -35,7 +35,6 @@ from torch import Tensor, nn
 
 from lerobot.constants import ACTION, OBS_ENV_STATE, OBS_IMAGES, OBS_STATE
 from lerobot.policies.diffusion.configuration_diffusion import DiffusionConfig
-from lerobot.policies.normalize import Normalize, Unnormalize
 from lerobot.policies.pretrained import PreTrainedPolicy
 from lerobot.policies.utils import (
     get_device_from_parameters,
@@ -57,7 +56,6 @@ class DiffusionPolicy(PreTrainedPolicy):
     def __init__(
         self,
         config: DiffusionConfig,
-        dataset_stats: dict[str, dict[str, Tensor]] | None = None,
     ):
         """
         Args:
@@ -70,14 +68,6 @@ class DiffusionPolicy(PreTrainedPolicy):
         config.validate_features()
         self.config = config
 
-        self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
-        self.normalize_targets = Normalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-        self.unnormalize_outputs = Unnormalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-
         # queues are populated during rollout of the policy, they contain the n latest observations and actions
         self._queues = None
 
@@ -106,9 +96,6 @@ class DiffusionPolicy(PreTrainedPolicy):
         batch = {k: torch.stack(list(self._queues[k]), dim=1) for k in batch if k in self._queues}
         actions = self.diffusion.generate_actions(batch)
 
-        # TODO(rcadene): make above methods return output dictionary?
-        actions = self.unnormalize_outputs({ACTION: actions})[ACTION]
-
         return actions
 
     @torch.no_grad()
@@ -137,7 +124,6 @@ class DiffusionPolicy(PreTrainedPolicy):
         if ACTION in batch:
             batch.pop(ACTION)
 
-        batch = self.normalize_inputs(batch)
         if self.config.image_features:
             batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
             batch[OBS_IMAGES] = torch.stack([batch[key] for key in self.config.image_features], dim=-4)
@@ -153,11 +139,9 @@ class DiffusionPolicy(PreTrainedPolicy):
 
     def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, None]:
         """Run the batch through the model and compute the loss for training or validation."""
-        batch = self.normalize_inputs(batch)
         if self.config.image_features:
             batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
             batch[OBS_IMAGES] = torch.stack([batch[key] for key in self.config.image_features], dim=-4)
-        batch = self.normalize_targets(batch)
         loss = self.diffusion.compute_loss(batch)
         # no output_dict so returning None
         return loss, None

src/lerobot/policies/diffusion/processor_diffusion.py

@@ -0,0 +1,96 @@
+#!/usr/bin/env python
+
+# Copyright 2024 Columbia Artificial Intelligence, Robotics Lab,
+# and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import torch
+
+from lerobot.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
+from lerobot.policies.diffusion.configuration_diffusion import DiffusionConfig
+from lerobot.processor import (
+    AddBatchDimensionProcessorStep,
+    DeviceProcessorStep,
+    NormalizerProcessorStep,
+    PolicyProcessorPipeline,
+    ProcessorKwargs,
+    RenameObservationsProcessorStep,
+    UnnormalizerProcessorStep,
+)
+
+
+def make_diffusion_pre_post_processors(
+    config: DiffusionConfig,
+    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
+    preprocessor_kwargs: ProcessorKwargs | None = None,
+    postprocessor_kwargs: ProcessorKwargs | None = None,
+) -> tuple[PolicyProcessorPipeline, PolicyProcessorPipeline]:
+    """
+    Constructs pre-processor and post-processor pipelines for a diffusion policy.
+
+    The pre-processing pipeline prepares the input data for the model by:
+    1. Renaming features (if a `rename_map` is provided in `preprocessor_kwargs`).
+    2. Normalizing the input and output features based on dataset statistics.
+    3. Adding a batch dimension.
+    4. Moving the data to the specified device.
+
+    The post-processing pipeline handles the model's output by:
+    1. Moving the data to the CPU.
+    2. Unnormalizing the output features to their original scale.
+
+    Args:
+        config: The configuration object for the diffusion policy,
+            containing feature definitions, normalization mappings, and device information.
+        dataset_stats: A dictionary of statistics used for normalization.
+            Defaults to None.
+        preprocessor_kwargs: Additional keyword arguments
+            for the pre-processor pipeline. Defaults to an empty dictionary.
+        postprocessor_kwargs: Additional keyword arguments
+            for the post-processor pipeline. Defaults to an empty dictionary.
+
+    Returns:
+        A tuple containing the configured pre-processor and post-processor pipelines.
+    """
+    if preprocessor_kwargs is None:
+        preprocessor_kwargs = {}
+    if postprocessor_kwargs is None:
+        postprocessor_kwargs = {}
+
+    input_steps = [
+        RenameObservationsProcessorStep(rename_map={}),
+        AddBatchDimensionProcessorStep(),
+        DeviceProcessorStep(device=config.device),
+        NormalizerProcessorStep(
+            features={**config.input_features, **config.output_features},
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+        ),
+    ]
+    output_steps = [
+        DeviceProcessorStep(device="cpu"),
+        UnnormalizerProcessorStep(
+            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
+        ),
+    ]
+    return (
+        PolicyProcessorPipeline(
+            steps=input_steps,
+            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
+            **preprocessor_kwargs,
+        ),
+        PolicyProcessorPipeline(
+            steps=output_steps,
+            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
+            **postprocessor_kwargs,
+        ),
+    )

src/lerobot/policies/factory.py

@@ -14,12 +14,17 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import logging
+from __future__ import annotations
 
-from torch import nn
+import logging
+from typing import Any, TypedDict
+
+import torch
+from typing_extensions import Unpack
 
 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.configs.types import FeatureType
+from lerobot.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
 from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
 from lerobot.datasets.utils import dataset_to_policy_features
 from lerobot.envs.configs import EnvConfig
@@ -34,10 +39,26 @@ from lerobot.policies.sac.reward_model.configuration_classifier import RewardCla
 from lerobot.policies.smolvla.configuration_smolvla import SmolVLAConfig
 from lerobot.policies.tdmpc.configuration_tdmpc import TDMPCConfig
 from lerobot.policies.vqbet.configuration_vqbet import VQBeTConfig
+from lerobot.processor import PolicyProcessorPipeline, ProcessorKwargs
 
 
-def get_policy_class(name: str) -> PreTrainedPolicy:
-    """Get the policy's class and config class given a name (matching the policy class' `name` attribute)."""
+def get_policy_class(name: str) -> type[PreTrainedPolicy]:
+    """
+    Retrieves a policy class by its registered name.
+
+    This function uses dynamic imports to avoid loading all policy classes into memory
+    at once, improving startup time and reducing dependencies.
+
+    Args:
+        name: The name of the policy. Supported names are "tdmpc", "diffusion", "act",
+              "vqbet", "pi0", "pi0fast", "sac", "reward_classifier", "smolvla".
+
+    Returns:
+        The policy class corresponding to the given name.
+
+    Raises:
+        NotImplementedError: If the policy name is not recognized.
+    """
     if name == "tdmpc":
         from lerobot.policies.tdmpc.modeling_tdmpc import TDMPCPolicy
 
@@ -79,6 +100,24 @@ def get_policy_class(name: str) -> PreTrainedPolicy:
 
 
 def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
+    """
+    Instantiates a policy configuration object based on the policy type.
+
+    This factory function simplifies the creation of policy configuration objects by
+    mapping a string identifier to the corresponding config class.
+
+    Args:
+        policy_type: The type of the policy. Supported types include "tdmpc",
+                     "diffusion", "act", "vqbet", "pi0", "pi0fast", "sac", "smolvla",
+                     "reward_classifier".
+        **kwargs: Keyword arguments to be passed to the configuration class constructor.
+
+    Returns:
+        An instance of a `PreTrainedConfig` subclass.
+
+    Raises:
+        ValueError: If the `policy_type` is not recognized.
+    """
     if policy_type == "tdmpc":
         return TDMPCConfig(**kwargs)
     elif policy_type == "diffusion":
@@ -101,30 +140,210 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
         raise ValueError(f"Policy type '{policy_type}' is not available.")
 
 
+class ProcessorConfigKwargs(TypedDict, total=False):
+    """
+    A TypedDict defining the keyword arguments for processor configuration.
+
+    This provides type hints for the optional arguments passed to `make_pre_post_processors`,
+    improving code clarity and enabling static analysis.
+
+    Attributes:
+        preprocessor_config_filename: The filename for the preprocessor configuration.
+        postprocessor_config_filename: The filename for the postprocessor configuration.
+        preprocessor_overrides: A dictionary of overrides for the preprocessor configuration.
+        postprocessor_overrides: A dictionary of overrides for the postprocessor configuration.
+        dataset_stats: Dataset statistics for normalization.
+        preprocessor_kwargs: Additional arguments for the `PolicyProcessorPipeline`.
+        postprocessor_kwargs: Additional arguments for the `PolicyProcessorPipeline`.
+    """
+
+    preprocessor_config_filename: str | None
+    postprocessor_config_filename: str | None
+    preprocessor_overrides: dict[str, Any] | None
+    postprocessor_overrides: dict[str, Any] | None
+    dataset_stats: dict[str, dict[str, torch.Tensor]] | None
+    preprocessor_kwargs: ProcessorKwargs | None
+    postprocessor_kwargs: ProcessorKwargs | None
+
+
+def make_pre_post_processors(
+    policy_cfg: PreTrainedConfig,
+    pretrained_path: str | None = None,
+    **kwargs: Unpack[ProcessorConfigKwargs],
+) -> tuple[PolicyProcessorPipeline, PolicyProcessorPipeline]:
+    """
+    Create or load pre- and post-processor pipelines for a given policy.
+
+    This function acts as a factory. It can either load existing processor pipelines
+    from a pretrained path or create new ones from scratch based on the policy
+    configuration. Each policy type has a dedicated factory function for its
+    processors (e.g., `make_tdmpc_pre_post_processors`).
+
+    Args:
+        policy_cfg: The configuration of the policy for which to create processors.
+        pretrained_path: An optional path to load pretrained processor pipelines from.
+            If provided, pipelines are loaded from this path.
+        **kwargs: Keyword arguments for processor configuration, as defined in
+            `ProcessorConfigKwargs`.
+
+    Returns:
+        A tuple containing the input (pre-processor) and output (post-processor) pipelines.
+
+    Raises:
+        NotImplementedError: If a processor factory is not implemented for the given
+            policy configuration type.
+    """
+    if pretrained_path:
+        # Extract preprocessor and postprocessor kwargs
+        preprocessor_kwargs = kwargs.get("preprocessor_kwargs", {})
+        postprocessor_kwargs = kwargs.get("postprocessor_kwargs", {})
+
+        return (
+            PolicyProcessorPipeline.from_pretrained(
+                pretrained_model_name_or_path=pretrained_path,
+                config_filename=kwargs.get(
+                    "preprocessor_config_filename", f"{POLICY_PREPROCESSOR_DEFAULT_NAME}.json"
+                ),
+                overrides=kwargs.get("preprocessor_overrides", {}),
+                to_transition=preprocessor_kwargs.get("to_transition"),
+                to_output=preprocessor_kwargs.get("to_output"),
+            ),
+            PolicyProcessorPipeline.from_pretrained(
+                pretrained_model_name_or_path=pretrained_path,
+                config_filename=kwargs.get(
+                    "postprocessor_config_filename", f"{POLICY_POSTPROCESSOR_DEFAULT_NAME}.json"
+                ),
+                overrides=kwargs.get("postprocessor_overrides", {}),
+                to_transition=postprocessor_kwargs.get("to_transition"),
+                to_output=postprocessor_kwargs.get("to_output"),
+            ),
+        )
+
+    # Create a new processor based on policy type
+    if isinstance(policy_cfg, TDMPCConfig):
+        from lerobot.policies.tdmpc.processor_tdmpc import make_tdmpc_pre_post_processors
+
+        processors = make_tdmpc_pre_post_processors(
+            config=policy_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+            preprocessor_kwargs=kwargs.get("preprocessor_kwargs"),
+            postprocessor_kwargs=kwargs.get("postprocessor_kwargs"),
+        )
+
+    elif isinstance(policy_cfg, DiffusionConfig):
+        from lerobot.policies.diffusion.processor_diffusion import make_diffusion_pre_post_processors
+
+        processors = make_diffusion_pre_post_processors(
+            config=policy_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+            preprocessor_kwargs=kwargs.get("preprocessor_kwargs"),
+            postprocessor_kwargs=kwargs.get("postprocessor_kwargs"),
+        )
+
+    elif isinstance(policy_cfg, ACTConfig):
+        from lerobot.policies.act.processor_act import make_act_pre_post_processors
+
+        processors = make_act_pre_post_processors(
+            config=policy_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+            preprocessor_kwargs=kwargs.get("preprocessor_kwargs"),
+            postprocessor_kwargs=kwargs.get("postprocessor_kwargs"),
+        )
+
+    elif isinstance(policy_cfg, VQBeTConfig):
+        from lerobot.policies.vqbet.processor_vqbet import make_vqbet_pre_post_processors
+
+        processors = make_vqbet_pre_post_processors(
+            config=policy_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+            preprocessor_kwargs=kwargs.get("preprocessor_kwargs"),
+            postprocessor_kwargs=kwargs.get("postprocessor_kwargs"),
+        )
+
+    elif isinstance(policy_cfg, PI0Config):
+        from lerobot.policies.pi0.processor_pi0 import make_pi0_pre_post_processors
+
+        processors = make_pi0_pre_post_processors(
+            config=policy_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+            preprocessor_kwargs=kwargs.get("preprocessor_kwargs"),
+            postprocessor_kwargs=kwargs.get("postprocessor_kwargs"),
+        )
+
+    elif isinstance(policy_cfg, PI0FASTConfig):
+        from lerobot.policies.pi0fast.processor_pi0fast import make_pi0fast_pre_post_processors
+
+        processors = make_pi0fast_pre_post_processors(
+            config=policy_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+            preprocessor_kwargs=kwargs.get("preprocessor_kwargs"),
+            postprocessor_kwargs=kwargs.get("postprocessor_kwargs"),
+        )
+
+    elif isinstance(policy_cfg, SACConfig):
+        from lerobot.policies.sac.processor_sac import make_sac_pre_post_processors
+
+        processors = make_sac_pre_post_processors(
+            config=policy_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+            preprocessor_kwargs=kwargs.get("preprocessor_kwargs"),
+            postprocessor_kwargs=kwargs.get("postprocessor_kwargs"),
+        )
+
+    elif isinstance(policy_cfg, RewardClassifierConfig):
+        from lerobot.policies.sac.reward_model.processor_classifier import make_classifier_processor
+
+        processors = make_classifier_processor(
+            config=policy_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+            preprocessor_kwargs=kwargs.get("preprocessor_kwargs"),
+            postprocessor_kwargs=kwargs.get("postprocessor_kwargs"),
+        )
+
+    elif isinstance(policy_cfg, SmolVLAConfig):
+        from lerobot.policies.smolvla.processor_smolvla import make_smolvla_pre_post_processors
+
+        processors = make_smolvla_pre_post_processors(
+            config=policy_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+            preprocessor_kwargs=kwargs.get("preprocessor_kwargs"),
+            postprocessor_kwargs=kwargs.get("postprocessor_kwargs"),
+        )
+
+    else:
+        raise NotImplementedError(f"Processor for policy type '{policy_cfg.type}' is not implemented.")
+
+    return processors
+
+
 def make_policy(
     cfg: PreTrainedConfig,
     ds_meta: LeRobotDatasetMetadata | None = None,
     env_cfg: EnvConfig | None = None,
 ) -> PreTrainedPolicy:
-    """Make an instance of a policy class.
+    """
+    Instantiate a policy model.
 
-    This function exists because (for now) we need to parse features from either a dataset or an environment
-    in order to properly dimension and instantiate a policy for that dataset or environment.
+    This factory function handles the logic of creating a policy, which requires
+    determining the input and output feature shapes. These shapes can be derived
+    either from a `LeRobotDatasetMetadata` object or an `EnvConfig` object. The function
+    can either initialize a new policy from scratch or load a pretrained one.
 
     Args:
-        cfg (PreTrainedConfig): The config of the policy to make. If `pretrained_path` is set, the policy will
-            be loaded with the weights from that path.
-        ds_meta (LeRobotDatasetMetadata | None, optional): Dataset metadata to take input/output shapes and
-            statistics to use for (un)normalization of inputs/outputs in the policy. Defaults to None.
-        env_cfg (EnvConfig | None, optional): The config of a gym environment to parse features from. Must be
-            provided if ds_meta is not. Defaults to None.
-
-    Raises:
-        ValueError: Either ds_meta or env and env_cfg must be provided.
-        NotImplementedError: if the policy.type is 'vqbet' and the policy device 'mps' (due to an incompatibility)
+        cfg: The configuration for the policy to be created. If `cfg.pretrained_path` is
+             set, the policy will be loaded with weights from that path.
+        ds_meta: Dataset metadata used to infer feature shapes and types. Also provides
+                 statistics for normalization layers.
+        env_cfg: Environment configuration used to infer feature shapes and types.
+                 One of `ds_meta` or `env_cfg` must be provided.
 
     Returns:
-        PreTrainedPolicy: _description_
+        An instantiated and device-placed policy model.
+
+    Raises:
+        ValueError: If both or neither of `ds_meta` and `env_cfg` are provided.
+        NotImplementedError: If attempting to use an unsupported policy-backend
+                             combination (e.g., VQBeT with 'mps').
     """
     if bool(ds_meta) == bool(env_cfg):
         raise ValueError("Either one of a dataset metadata or a sim env must be provided.")
@@ -147,7 +366,6 @@ def make_policy(
     kwargs = {}
     if ds_meta is not None:
         features = dataset_to_policy_features(ds_meta.features)
-        kwargs["dataset_stats"] = ds_meta.stats
     else:
         if not cfg.pretrained_path:
             logging.warning(
@@ -155,6 +373,8 @@ def make_policy(
                 "rather than a dataset. Normalization modules inside the policy will have infinite values "
                 "by default without stats from a dataset."
             )
+        if env_cfg is None:
+            raise ValueError("env_cfg cannot be None when ds_meta is not provided")
         features = env_to_policy_features(env_cfg)
 
     cfg.output_features = {key: ft for key, ft in features.items() if ft.type is FeatureType.ACTION}
@@ -171,7 +391,7 @@ def make_policy(
         policy = policy_cls(**kwargs)
 
     policy.to(cfg.device)
-    assert isinstance(policy, nn.Module)
+    assert isinstance(policy, torch.nn.Module)
 
     # policy = torch.compile(policy, mode="reduce-overhead")
 

src/lerobot/policies/normalize.py

@@ -1,420 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-import numpy as np
-import torch
-from torch import Tensor, nn
-
-from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
-
-
-def create_stats_buffers(
-    features: dict[str, PolicyFeature],
-    norm_map: dict[str, NormalizationMode],
-    stats: dict[str, dict[str, Tensor]] | None = None,
-) -> dict[str, dict[str, nn.ParameterDict]]:
-    """
-    Create buffers per modality (e.g. "observation.image", "action") containing their mean, std, min, max
-    statistics.
-
-    Args: (see Normalize and Unnormalize)
-
-    Returns:
-        dict: A dictionary where keys are modalities and values are `nn.ParameterDict` containing
-            `nn.Parameters` set to `requires_grad=False`, suitable to not be updated during backpropagation.
-    """
-    stats_buffers = {}
-
-    for key, ft in features.items():
-        norm_mode = norm_map.get(ft.type, NormalizationMode.IDENTITY)
-        if norm_mode is NormalizationMode.IDENTITY:
-            continue
-
-        assert isinstance(norm_mode, NormalizationMode)
-
-        shape = tuple(ft.shape)
-
-        if ft.type is FeatureType.VISUAL:
-            # sanity checks
-            assert len(shape) == 3, f"number of dimensions of {key} != 3 ({shape=}"
-            c, h, w = shape
-            assert c < h and c < w, f"{key} is not channel first ({shape=})"
-            # override image shape to be invariant to height and width
-            shape = (c, 1, 1)
-
-        # Note: we initialize mean, std, min, max to infinity. They should be overwritten
-        # downstream by `stats` or `policy.load_state_dict`, as expected. During forward,
-        # we assert they are not infinity anymore.
-
-        buffer = {}
-        if norm_mode is NormalizationMode.MEAN_STD:
-            mean = torch.ones(shape, dtype=torch.float32) * torch.inf
-            std = torch.ones(shape, dtype=torch.float32) * torch.inf
-            buffer = nn.ParameterDict(
-                {
-                    "mean": nn.Parameter(mean, requires_grad=False),
-                    "std": nn.Parameter(std, requires_grad=False),
-                }
-            )
-        elif norm_mode is NormalizationMode.MIN_MAX:
-            min = torch.ones(shape, dtype=torch.float32) * torch.inf
-            max = torch.ones(shape, dtype=torch.float32) * torch.inf
-            buffer = nn.ParameterDict(
-                {
-                    "min": nn.Parameter(min, requires_grad=False),
-                    "max": nn.Parameter(max, requires_grad=False),
-                }
-            )
-
-        # TODO(aliberts, rcadene): harmonize this to only use one framework (np or torch)
-        if stats:
-            if isinstance(stats[key]["mean"], np.ndarray):
-                if norm_mode is NormalizationMode.MEAN_STD:
-                    buffer["mean"].data = torch.from_numpy(stats[key]["mean"]).to(dtype=torch.float32)
-                    buffer["std"].data = torch.from_numpy(stats[key]["std"]).to(dtype=torch.float32)
-                elif norm_mode is NormalizationMode.MIN_MAX:
-                    buffer["min"].data = torch.from_numpy(stats[key]["min"]).to(dtype=torch.float32)
-                    buffer["max"].data = torch.from_numpy(stats[key]["max"]).to(dtype=torch.float32)
-            elif isinstance(stats[key]["mean"], torch.Tensor):
-                # Note: The clone is needed to make sure that the logic in save_pretrained doesn't see duplicated
-                # tensors anywhere (for example, when we use the same stats for normalization and
-                # unnormalization). See the logic here
-                # https://github.com/huggingface/safetensors/blob/079781fd0dc455ba0fe851e2b4507c33d0c0d407/bindings/python/py_src/safetensors/torch.py#L97.
-                if norm_mode is NormalizationMode.MEAN_STD:
-                    buffer["mean"].data = stats[key]["mean"].clone().to(dtype=torch.float32)
-                    buffer["std"].data = stats[key]["std"].clone().to(dtype=torch.float32)
-                elif norm_mode is NormalizationMode.MIN_MAX:
-                    buffer["min"].data = stats[key]["min"].clone().to(dtype=torch.float32)
-                    buffer["max"].data = stats[key]["max"].clone().to(dtype=torch.float32)
-            else:
-                type_ = type(stats[key]["mean"])
-                raise ValueError(f"np.ndarray or torch.Tensor expected, but type is '{type_}' instead.")
-
-        stats_buffers[key] = buffer
-    return stats_buffers
-
-
-def _no_stats_error_str(name: str) -> str:
-    return (
-        f"`{name}` is infinity. You should either initialize with `stats` as an argument, or use a "
-        "pretrained model."
-    )
-
-
-class Normalize(nn.Module):
-    """Normalizes data (e.g. "observation.image") for more stable and faster convergence during training."""
-
-    def __init__(
-        self,
-        features: dict[str, PolicyFeature],
-        norm_map: dict[str, NormalizationMode],
-        stats: dict[str, dict[str, Tensor]] | None = None,
-    ):
-        """
-        Args:
-            shapes (dict): A dictionary where keys are input modalities (e.g. "observation.image") and values
-            are their shapes (e.g. `[3,96,96]`]). These shapes are used to create the tensor buffer containing
-            mean, std, min, max statistics. If the provided `shapes` contain keys related to images, the shape
-            is adjusted to be invariant to height and width, assuming a channel-first (c, h, w) format.
-            modes (dict): A dictionary where keys are output modalities (e.g. "observation.image") and values
-                are their normalization modes among:
-                    - "mean_std": subtract the mean and divide by standard deviation.
-                    - "min_max": map to [-1, 1] range.
-            stats (dict, optional): A dictionary where keys are output modalities (e.g. "observation.image")
-                and values are dictionaries of statistic types and their values (e.g.
-                `{"mean": torch.randn(3,1,1)}, "std": torch.randn(3,1,1)}`). If provided, as expected for
-                training the model for the first time, these statistics will overwrite the default buffers. If
-                not provided, as expected for finetuning or evaluation, the default buffers should to be
-                overwritten by a call to `policy.load_state_dict(state_dict)`. That way, initializing the
-                dataset is not needed to get the stats, since they are already in the policy state_dict.
-        """
-        super().__init__()
-        self.features = features
-        self.norm_map = norm_map
-        self.stats = stats
-        stats_buffers = create_stats_buffers(features, norm_map, stats)
-        for key, buffer in stats_buffers.items():
-            setattr(self, "buffer_" + key.replace(".", "_"), buffer)
-
-    # TODO(rcadene): should we remove torch.no_grad?
-    @torch.no_grad()
-    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
-        # TODO: Remove this shallow copy
-        batch = dict(batch)  # shallow copy avoids mutating the input batch
-        for key, ft in self.features.items():
-            if key not in batch:
-                # FIXME(aliberts, rcadene): This might lead to silent fail!
-                continue
-
-            norm_mode = self.norm_map.get(ft.type, NormalizationMode.IDENTITY)
-            if norm_mode is NormalizationMode.IDENTITY:
-                continue
-
-            buffer = getattr(self, "buffer_" + key.replace(".", "_"))
-
-            if norm_mode is NormalizationMode.MEAN_STD:
-                mean = buffer["mean"]
-                std = buffer["std"]
-                assert not torch.isinf(mean).any(), _no_stats_error_str("mean")
-                assert not torch.isinf(std).any(), _no_stats_error_str("std")
-                batch[key] = (batch[key] - mean) / (std + 1e-8)
-            elif norm_mode is NormalizationMode.MIN_MAX:
-                min = buffer["min"]
-                max = buffer["max"]
-                assert not torch.isinf(min).any(), _no_stats_error_str("min")
-                assert not torch.isinf(max).any(), _no_stats_error_str("max")
-                # normalize to [0,1]
-                batch[key] = (batch[key] - min) / (max - min + 1e-8)
-                # normalize to [-1, 1]
-                batch[key] = batch[key] * 2 - 1
-            else:
-                raise ValueError(norm_mode)
-        return batch
-
-
-class Unnormalize(nn.Module):
-    """
-    Similar to `Normalize` but unnormalizes output data (e.g. `{"action": torch.randn(b,c)}`) in their
-    original range used by the environment.
-    """
-
-    def __init__(
-        self,
-        features: dict[str, PolicyFeature],
-        norm_map: dict[str, NormalizationMode],
-        stats: dict[str, dict[str, Tensor]] | None = None,
-    ):
-        """
-        Args:
-            shapes (dict): A dictionary where keys are input modalities (e.g. "observation.image") and values
-            are their shapes (e.g. `[3,96,96]`]). These shapes are used to create the tensor buffer containing
-            mean, std, min, max statistics. If the provided `shapes` contain keys related to images, the shape
-            is adjusted to be invariant to height and width, assuming a channel-first (c, h, w) format.
-            modes (dict): A dictionary where keys are output modalities (e.g. "observation.image") and values
-                are their normalization modes among:
-                    - "mean_std": subtract the mean and divide by standard deviation.
-                    - "min_max": map to [-1, 1] range.
-            stats (dict, optional): A dictionary where keys are output modalities (e.g. "observation.image")
-                and values are dictionaries of statistic types and their values (e.g.
-                `{"mean": torch.randn(3,1,1)}, "std": torch.randn(3,1,1)}`). If provided, as expected for
-                training the model for the first time, these statistics will overwrite the default buffers. If
-                not provided, as expected for finetuning or evaluation, the default buffers should to be
-                overwritten by a call to `policy.load_state_dict(state_dict)`. That way, initializing the
-                dataset is not needed to get the stats, since they are already in the policy state_dict.
-        """
-        super().__init__()
-        self.features = features
-        self.norm_map = norm_map
-        self.stats = stats
-        # `self.buffer_observation_state["mean"]` contains `torch.tensor(state_dim)`
-        stats_buffers = create_stats_buffers(features, norm_map, stats)
-        for key, buffer in stats_buffers.items():
-            setattr(self, "buffer_" + key.replace(".", "_"), buffer)
-
-    # TODO(rcadene): should we remove torch.no_grad?
-    @torch.no_grad()
-    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
-        batch = dict(batch)  # shallow copy avoids mutating the input batch
-        for key, ft in self.features.items():
-            if key not in batch:
-                continue
-
-            norm_mode = self.norm_map.get(ft.type, NormalizationMode.IDENTITY)
-            if norm_mode is NormalizationMode.IDENTITY:
-                continue
-
-            buffer = getattr(self, "buffer_" + key.replace(".", "_"))
-
-            if norm_mode is NormalizationMode.MEAN_STD:
-                mean = buffer["mean"]
-                std = buffer["std"]
-                assert not torch.isinf(mean).any(), _no_stats_error_str("mean")
-                assert not torch.isinf(std).any(), _no_stats_error_str("std")
-                batch[key] = batch[key] * std + mean
-            elif norm_mode is NormalizationMode.MIN_MAX:
-                min = buffer["min"]
-                max = buffer["max"]
-                assert not torch.isinf(min).any(), _no_stats_error_str("min")
-                assert not torch.isinf(max).any(), _no_stats_error_str("max")
-                batch[key] = (batch[key] + 1) / 2
-                batch[key] = batch[key] * (max - min) + min
-            else:
-                raise ValueError(norm_mode)
-        return batch
-
-
-# TODO (azouitine): We should replace all normalization on the policies with register_buffer normalization
-#       and remove the `Normalize` and `Unnormalize` classes.
-def _initialize_stats_buffers(
-    module: nn.Module,
-    features: dict[str, PolicyFeature],
-    norm_map: dict[str, NormalizationMode],
-    stats: dict[str, dict[str, Tensor]] | None = None,
-) -> None:
-    """Register statistics buffers (mean/std or min/max) on the given *module*.
-
-    The logic matches the previous constructors of `NormalizeBuffer` and `UnnormalizeBuffer`,
-    but is factored out so it can be reused by both classes and stay in sync.
-    """
-    for key, ft in features.items():
-        norm_mode = norm_map.get(ft.type, NormalizationMode.IDENTITY)
-        if norm_mode is NormalizationMode.IDENTITY:
-            continue
-
-        shape: tuple[int, ...] = tuple(ft.shape)
-        if ft.type is FeatureType.VISUAL:
-            # reduce spatial dimensions, keep channel dimension only
-            c, *_ = shape
-            shape = (c, 1, 1)
-
-        prefix = key.replace(".", "_")
-
-        if norm_mode is NormalizationMode.MEAN_STD:
-            mean = torch.full(shape, torch.inf, dtype=torch.float32)
-            std = torch.full(shape, torch.inf, dtype=torch.float32)
-
-            if stats and key in stats and "mean" in stats[key] and "std" in stats[key]:
-                mean_data = stats[key]["mean"]
-                std_data = stats[key]["std"]
-                if isinstance(mean_data, torch.Tensor):
-                    # Note: The clone is needed to make sure that the logic in save_pretrained doesn't see duplicated
-                    # tensors anywhere (for example, when we use the same stats for normalization and
-                    # unnormalization). See the logic here
-                    # https://github.com/huggingface/safetensors/blob/079781fd0dc455ba0fe851e2b4507c33d0c0d407/bindings/python/py_src/safetensors/torch.py#L97.
-                    mean = mean_data.clone().to(dtype=torch.float32)
-                    std = std_data.clone().to(dtype=torch.float32)
-                else:
-                    raise ValueError(f"Unsupported stats type for key '{key}' (expected ndarray or Tensor).")
-
-            module.register_buffer(f"{prefix}_mean", mean)
-            module.register_buffer(f"{prefix}_std", std)
-            continue
-
-        if norm_mode is NormalizationMode.MIN_MAX:
-            min_val = torch.full(shape, torch.inf, dtype=torch.float32)
-            max_val = torch.full(shape, torch.inf, dtype=torch.float32)
-
-            if stats and key in stats and "min" in stats[key] and "max" in stats[key]:
-                min_data = stats[key]["min"]
-                max_data = stats[key]["max"]
-                if isinstance(min_data, torch.Tensor):
-                    min_val = min_data.clone().to(dtype=torch.float32)
-                    max_val = max_data.clone().to(dtype=torch.float32)
-                else:
-                    raise ValueError(f"Unsupported stats type for key '{key}' (expected ndarray or Tensor).")
-
-            module.register_buffer(f"{prefix}_min", min_val)
-            module.register_buffer(f"{prefix}_max", max_val)
-            continue
-
-        raise ValueError(norm_mode)
-
-
-class NormalizeBuffer(nn.Module):
-    """Same as `Normalize` but statistics are stored as registered buffers rather than parameters."""
-
-    def __init__(
-        self,
-        features: dict[str, PolicyFeature],
-        norm_map: dict[str, NormalizationMode],
-        stats: dict[str, dict[str, Tensor]] | None = None,
-    ):
-        super().__init__()
-        self.features = features
-        self.norm_map = norm_map
-
-        _initialize_stats_buffers(self, features, norm_map, stats)
-
-    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
-        batch = dict(batch)
-        for key, ft in self.features.items():
-            if key not in batch:
-                continue
-
-            norm_mode = self.norm_map.get(ft.type, NormalizationMode.IDENTITY)
-            if norm_mode is NormalizationMode.IDENTITY:
-                continue
-
-            prefix = key.replace(".", "_")
-
-            if norm_mode is NormalizationMode.MEAN_STD:
-                mean = getattr(self, f"{prefix}_mean")
-                std = getattr(self, f"{prefix}_std")
-                assert not torch.isinf(mean).any(), _no_stats_error_str("mean")
-                assert not torch.isinf(std).any(), _no_stats_error_str("std")
-                batch[key] = (batch[key] - mean) / (std + 1e-8)
-                continue
-
-            if norm_mode is NormalizationMode.MIN_MAX:
-                min_val = getattr(self, f"{prefix}_min")
-                max_val = getattr(self, f"{prefix}_max")
-                assert not torch.isinf(min_val).any(), _no_stats_error_str("min")
-                assert not torch.isinf(max_val).any(), _no_stats_error_str("max")
-                batch[key] = (batch[key] - min_val) / (max_val - min_val + 1e-8)
-                batch[key] = batch[key] * 2 - 1
-                continue
-
-            raise ValueError(norm_mode)
-
-        return batch
-
-
-class UnnormalizeBuffer(nn.Module):
-    """Inverse operation of `NormalizeBuffer`. Uses registered buffers for statistics."""
-
-    def __init__(
-        self,
-        features: dict[str, PolicyFeature],
-        norm_map: dict[str, NormalizationMode],
-        stats: dict[str, dict[str, Tensor]] | None = None,
-    ):
-        super().__init__()
-        self.features = features
-        self.norm_map = norm_map
-
-        _initialize_stats_buffers(self, features, norm_map, stats)
-
-    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
-        # batch = dict(batch)
-        for key, ft in self.features.items():
-            if key not in batch:
-                continue
-
-            norm_mode = self.norm_map.get(ft.type, NormalizationMode.IDENTITY)
-            if norm_mode is NormalizationMode.IDENTITY:
-                continue
-
-            prefix = key.replace(".", "_")
-
-            if norm_mode is NormalizationMode.MEAN_STD:
-                mean = getattr(self, f"{prefix}_mean")
-                std = getattr(self, f"{prefix}_std")
-                assert not torch.isinf(mean).any(), _no_stats_error_str("mean")
-                assert not torch.isinf(std).any(), _no_stats_error_str("std")
-                batch[key] = batch[key] * std + mean
-                continue
-
-            if norm_mode is NormalizationMode.MIN_MAX:
-                min_val = getattr(self, f"{prefix}_min")
-                max_val = getattr(self, f"{prefix}_max")
-                assert not torch.isinf(min_val).any(), _no_stats_error_str("min")
-                assert not torch.isinf(max_val).any(), _no_stats_error_str("max")
-                batch[key] = (batch[key] + 1) / 2
-                batch[key] = batch[key] * (max_val - min_val) + min_val
-                continue
-
-            raise ValueError(norm_mode)
-
-        return batch

src/lerobot/policies/pi0/modeling_pi0.py

@@ -30,7 +30,7 @@ pip install -e ".[pi0]"
 
 Example of finetuning the pi0 pretrained model (`pi0_base` in `openpi`):
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
 --policy.path=lerobot/pi0 \
 --dataset.repo_id=danaaubakirova/koch_test
 ```
@@ -38,7 +38,7 @@ python -m lerobot.scripts.train \
 Example of finetuning the pi0 neural network with PaliGemma and expert Gemma
 pretrained with VLM default parameters before pi0 finetuning:
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
 --policy.type=pi0 \
 --dataset.repo_id=danaaubakirova/koch_test
 ```
@@ -56,18 +56,15 @@ from collections import deque
 import torch
 import torch.nn.functional as F  # noqa: N812
 from torch import Tensor, nn
-from transformers import AutoTokenizer
 
-from lerobot.constants import ACTION, OBS_STATE
-from lerobot.policies.normalize import Normalize, Unnormalize
+from lerobot.constants import ACTION, OBS_LANGUAGE, OBS_STATE
 from lerobot.policies.pi0.configuration_pi0 import PI0Config
 from lerobot.policies.pi0.paligemma_with_expert import (
     PaliGemmaWithExpertConfig,
     PaliGemmaWithExpertModel,
 )
 from lerobot.policies.pretrained import PreTrainedPolicy
-from lerobot.policies.utils import log_model_loading_keys
-from lerobot.utils.utils import get_safe_dtype, init_logging
+from lerobot.utils.utils import get_safe_dtype
 
 
 def create_sinusoidal_pos_embedding(
@@ -223,28 +220,17 @@ class PI0Policy(PreTrainedPolicy):
     def __init__(
         self,
         config: PI0Config,
-        dataset_stats: dict[str, dict[str, Tensor]] | None = None,
     ):
         """
         Args:
             config: Policy configuration class instance or None, in which case the default instantiation of
                     the configuration class is used.
-            dataset_stats: Dataset statistics to be used for normalization. If not passed here, it is expected
-                that they will be passed with a call to `load_state_dict` before the policy is used.
         """
 
         super().__init__(config)
         config.validate_features()
         self.config = config
-        self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
-        self.normalize_targets = Normalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-        self.unnormalize_outputs = Unnormalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
 
-        self.language_tokenizer = AutoTokenizer.from_pretrained("google/paligemma-3b-pt-224")
         self.model = PI0FlowMatching(config)
 
         self.reset()
@@ -253,99 +239,6 @@ class PI0Policy(PreTrainedPolicy):
         """This should be called whenever the environment is reset."""
         self._action_queue = deque([], maxlen=self.config.n_action_steps)
 
-    @classmethod
-    def _transform_state_dict_keys(cls, state_dict: dict) -> dict:
-        """
-        Transform state dict keys to match expected model structure.
-
-        Transformations:
-        - model.paligemma_with_expert.paligemma.language_model.lm_head ->
-          model.paligemma_with_expert.paligemma.lm_head
-        - model.paligemma_with_expert.paligemma.language_model.model ->
-          model.paligemma_with_expert.paligemma.model.language_model
-        - model.paligemma_with_expert.paligemma.vision_tower ->
-          model.paligemma_with_expert.paligemma.model.vision_tower
-        - model.paligemma_with_expert.paligemma.multi_modal_projector ->
-          model.paligemma_with_expert.paligemma.model.multi_modal_projector
-
-        Also handles tied weights between lm_head.weight and
-        embed_tokens.weight.
-        """
-        import re
-
-        transformed_dict = {}
-
-        transformations = [
-            (
-                re.compile(r"\.paligemma_with_expert\.paligemma\.language_model\.lm_head"),
-                ".paligemma_with_expert.paligemma.lm_head",
-            ),
-            (
-                re.compile(r"\.paligemma_with_expert\.paligemma\.language_model\.model"),
-                ".paligemma_with_expert.paligemma.model.language_model",
-            ),
-            (
-                re.compile(r"\.paligemma_with_expert\.paligemma\.vision_tower"),
-                ".paligemma_with_expert.paligemma.model.vision_tower",
-            ),
-            (
-                re.compile(r"\.paligemma_with_expert\.paligemma\.multi_modal_projector"),
-                ".paligemma_with_expert.paligemma.model.multi_modal_projector",
-            ),
-        ]
-
-        for key, value in state_dict.items():
-            new_key = key
-            for pattern, replacement in transformations:
-                new_key = pattern.sub(replacement, new_key)
-            transformed_dict[new_key] = value
-
-        # Handle tied weights: lm_head.weight and embed_tokens.weight share memory
-        lm_head_key = None
-        embed_tokens_key = None
-
-        for key in transformed_dict:
-            if key.endswith(".paligemma_with_expert.paligemma.lm_head.weight"):
-                lm_head_key = key
-            elif key.endswith(".paligemma_with_expert.paligemma.model.language_model.embed_tokens.weight"):
-                embed_tokens_key = key
-            if lm_head_key and embed_tokens_key:
-                break
-
-        if lm_head_key and not embed_tokens_key:
-            embed_tokens_key = lm_head_key.replace(
-                ".lm_head.weight", ".model.language_model.embed_tokens.weight"
-            )
-            transformed_dict[embed_tokens_key] = transformed_dict[lm_head_key]
-        elif embed_tokens_key and not lm_head_key:
-            lm_head_key = embed_tokens_key.replace(
-                ".model.language_model.embed_tokens.weight", ".lm_head.weight"
-            )
-            transformed_dict[lm_head_key] = transformed_dict[embed_tokens_key]
-
-        return transformed_dict
-
-    @classmethod
-    def _load_as_safetensor(
-        cls, model: "PI0Policy", model_file: str, map_location: str, strict: bool
-    ) -> "PI0Policy":
-        """Override to apply key transformations before loading."""
-        from safetensors.torch import load_file
-
-        init_logging()
-        # Load the state dict from file safely
-        state_dict = load_file(model_file, device=map_location)
-
-        # Apply key transformations
-        transformed_state_dict = cls._transform_state_dict_keys(state_dict)
-
-        # Load the transformed state dict
-        msg = model.load_state_dict(transformed_state_dict, strict=strict)
-
-        # Log message
-        log_model_loading_keys(msg.missing_keys, msg.unexpected_keys)
-        return model
-
     def get_optim_params(self) -> dict:
         return self.parameters()
 
@@ -377,14 +270,13 @@ class PI0Policy(PreTrainedPolicy):
         if self.config.adapt_to_pi_aloha:
             batch[OBS_STATE] = self._pi_aloha_decode_state(batch[OBS_STATE])
 
-        batch = self.normalize_inputs(batch)
-
         # Action queue logic for n_action_steps > 1. When the action_queue is depleted, populate it by
         # querying the policy.
         if len(self._action_queue) == 0:
             images, img_masks = self.prepare_images(batch)
             state = self.prepare_state(batch)
-            lang_tokens, lang_masks = self.prepare_language(batch)
+            lang_tokens = batch[f"{OBS_LANGUAGE}.tokens"]
+            lang_masks = batch[f"{OBS_LANGUAGE}.attention_mask"]
 
             actions = self.model.sample_actions(
                 images, img_masks, lang_tokens, lang_masks, state, noise=noise
@@ -394,8 +286,6 @@ class PI0Policy(PreTrainedPolicy):
             original_action_dim = self.config.action_feature.shape[0]
             actions = actions[:, :, :original_action_dim]
 
-            actions = self.unnormalize_outputs({"action": actions})["action"]
-
             if self.config.adapt_to_pi_aloha:
                 actions = self._pi_aloha_encode_actions(actions)
 
@@ -410,12 +300,10 @@ class PI0Policy(PreTrainedPolicy):
             batch[OBS_STATE] = self._pi_aloha_decode_state(batch[OBS_STATE])
             batch[ACTION] = self._pi_aloha_encode_actions_inv(batch[ACTION])
 
-        batch = self.normalize_inputs(batch)
-        batch = self.normalize_targets(batch)
-
         images, img_masks = self.prepare_images(batch)
         state = self.prepare_state(batch)
-        lang_tokens, lang_masks = self.prepare_language(batch)
+        lang_tokens = batch[f"{OBS_LANGUAGE}.tokens"]
+        lang_masks = batch[f"{OBS_LANGUAGE}.attention_mask"]
         actions = self.prepare_action(batch)
         actions_is_pad = batch.get("action_is_pad")
 
@@ -482,26 +370,6 @@ class PI0Policy(PreTrainedPolicy):
 
         return images, img_masks
 
-    def prepare_language(self, batch) -> tuple[Tensor, Tensor]:
-        """Tokenize the text input"""
-        device = batch[OBS_STATE].device
-        tasks = batch["task"]
-
-        # PaliGemma prompt has to end with a new line
-        tasks = [task if task.endswith("\n") else f"{task}\n" for task in tasks]
-
-        tokenized_prompt = self.language_tokenizer.__call__(
-            tasks,
-            padding="max_length",
-            padding_side="right",
-            max_length=self.config.tokenizer_max_length,
-            return_tensors="pt",
-        )
-        lang_tokens = tokenized_prompt["input_ids"].to(device=device)
-        lang_masks = tokenized_prompt["attention_mask"].to(device=device, dtype=torch.bool)
-
-        return lang_tokens, lang_masks
-
     def _pi_aloha_decode_state(self, state):
         # Flip the joints.
         for motor_idx in [1, 2, 8, 9]:
@@ -567,7 +435,7 @@ class PI0FlowMatching(nn.Module):
     └──────────────────────────────┘
     """
 
-    def __init__(self, config):
+    def __init__(self, config: PI0Config):
         super().__init__()
         self.config = config
 

src/lerobot/policies/pi0/processor_pi0.py

@@ -0,0 +1,167 @@
+#!/usr/bin/env python
+
+# Copyright 2025 Physical Intelligence and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import torch
+
+from lerobot.configs.types import PipelineFeatureType, PolicyFeature
+from lerobot.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
+from lerobot.policies.pi0.configuration_pi0 import PI0Config
+from lerobot.processor import (
+    AddBatchDimensionProcessorStep,
+    ComplementaryDataProcessorStep,
+    DeviceProcessorStep,
+    NormalizerProcessorStep,
+    PolicyProcessorPipeline,
+    ProcessorKwargs,
+    ProcessorStep,
+    ProcessorStepRegistry,
+    RenameObservationsProcessorStep,
+    TokenizerProcessorStep,
+    UnnormalizerProcessorStep,
+)
+
+
+@ProcessorStepRegistry.register(name="pi0_new_line_processor")
+class Pi0NewLineProcessor(ComplementaryDataProcessorStep):
+    """
+    Ensures that the task description string ends with a newline character.
+
+    This processing step is required for compatibility with the PaliGemma tokenizer,
+    which expects a newline at the end of the text prompt. It handles both single
+    strings and lists of strings for the 'task' key in complementary data.
+    """
+
+    def complementary_data(self, complementary_data):
+        """
+        Adds a newline to the 'task' field if it doesn't already have one.
+
+        Args:
+            complementary_data: A dictionary that may contain a 'task' key with a
+                                string or list of strings.
+
+        Returns:
+            A new dictionary with the modified 'task' field.
+        """
+        if "task" not in complementary_data:
+            return complementary_data
+
+        task = complementary_data["task"]
+        if task is None:
+            return complementary_data
+
+        new_complementary_data = dict(complementary_data)
+
+        # Handle both string and list of strings
+        if isinstance(task, str):
+            # Single string: add newline if not present
+            if not task.endswith("\n"):
+                new_complementary_data["task"] = f"{task}\n"
+        elif isinstance(task, list) and all(isinstance(t, str) for t in task):
+            # List of strings: add newline to each if not present
+            new_complementary_data["task"] = [t if t.endswith("\n") else f"{t}\n" for t in task]
+        # If task is neither string nor list of strings, leave unchanged
+
+        return new_complementary_data
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        """
+        This step does not alter the feature definitions.
+
+        Args:
+            features: The input feature dictionary.
+
+        Returns:
+            The unchanged feature dictionary.
+        """
+        return features
+
+
+def make_pi0_pre_post_processors(
+    config: PI0Config,
+    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
+    preprocessor_kwargs: ProcessorKwargs | None = None,
+    postprocessor_kwargs: ProcessorKwargs | None = None,
+) -> tuple[PolicyProcessorPipeline, PolicyProcessorPipeline]:
+    """
+    Constructs pre-processor and post-processor pipelines for the PI0 policy.
+
+    The pre-processing pipeline prepares input data for the model by:
+    1. Renaming features to match pretrained configurations.
+    2. Normalizing input and output features based on dataset statistics.
+    3. Adding a batch dimension.
+    4. Appending a newline character to the task description for tokenizer compatibility.
+    5. Tokenizing the text prompt using the PaliGemma tokenizer.
+    6. Moving all data to the specified device.
+
+    The post-processing pipeline handles the model's output by:
+    1. Moving data to the CPU.
+    2. Unnormalizing the output features to their original scale.
+
+    Args:
+        config: The configuration object for the PI0 policy.
+        dataset_stats: A dictionary of statistics for normalization.
+        preprocessor_kwargs: Additional arguments for the pre-processor pipeline.
+        postprocessor_kwargs: Additional arguments for the post-processor pipeline.
+
+    Returns:
+        A tuple containing the configured pre-processor and post-processor pipelines.
+    """
+    if preprocessor_kwargs is None:
+        preprocessor_kwargs = {}
+    if postprocessor_kwargs is None:
+        postprocessor_kwargs = {}
+
+    # Add remaining processors
+    input_steps: list[ProcessorStep] = [
+        RenameObservationsProcessorStep(rename_map={}),  # To mimic the same processor as pretrained one
+        AddBatchDimensionProcessorStep(),
+        Pi0NewLineProcessor(),  # Add newlines before tokenization for PaliGemma
+        TokenizerProcessorStep(
+            tokenizer_name="google/paligemma-3b-pt-224",
+            max_length=config.tokenizer_max_length,
+            padding_side="right",
+            padding="max_length",
+        ),
+        DeviceProcessorStep(device=config.device),
+        NormalizerProcessorStep(
+            features={**config.input_features, **config.output_features},
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+        ),
+    ]
+
+    output_steps: list[ProcessorStep] = [
+        DeviceProcessorStep(device="cpu"),
+        UnnormalizerProcessorStep(
+            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
+        ),
+    ]
+
+    return (
+        PolicyProcessorPipeline(
+            steps=input_steps,
+            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
+            **preprocessor_kwargs,
+        ),
+        PolicyProcessorPipeline(
+            steps=output_steps,
+            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
+            **postprocessor_kwargs,
+        ),
+    )

src/lerobot/policies/pi0fast/modeling_pi0fast.py

@@ -25,14 +25,14 @@ Disclaimer: It is not expected to perform as well as the original implementation
 
 Example of finetuning the pi0+FAST pretrained model (`pi0_fast_base` in `openpi`):
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
 --policy.path=lerobot/pi0fast_base \
 --dataset.repo_id=danaaubakirova/koch_test
 ```
 
 Example of training the pi0+FAST neural network with from scratch:
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
 --policy.type=pi0fast \
 --dataset.repo_id=danaaubakirova/koch_test
 ```
@@ -58,7 +58,6 @@ from transformers.cache_utils import HybridCache, StaticCache
 from transformers.models.auto import CONFIG_MAPPING
 
 from lerobot.constants import ACTION, OBS_STATE
-from lerobot.policies.normalize import Normalize, Unnormalize
 from lerobot.policies.pi0fast.configuration_pi0fast import PI0FASTConfig
 from lerobot.policies.pretrained import PreTrainedPolicy
 
@@ -146,14 +145,6 @@ class PI0FASTPolicy(PreTrainedPolicy):
         config.validate_features()
         self.config = config
 
-        self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
-        self.normalize_targets = Normalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-        self.unnormalize_outputs = Unnormalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-
         self.language_tokenizer = AutoProcessor.from_pretrained("google/paligemma-3b-pt-224")
         self.model = PI0FAST(config)
 
@@ -221,8 +212,6 @@ class PI0FASTPolicy(PreTrainedPolicy):
         if self.config.adapt_to_pi_aloha:
             batch[OBS_STATE] = self._pi_aloha_decode_state(batch[OBS_STATE])
 
-        batch = self.normalize_inputs(batch)
-
         # Action queue logic for n_action_steps > 1. When the action_queue is depleted, populate it by
         # querying the policy.
         if len(self._action_queue) == 0:
@@ -235,8 +224,6 @@ class PI0FASTPolicy(PreTrainedPolicy):
             ]  # self.config.max_action_dim  # self.config.action_feature.shape[0]
             actions = actions[:, :, :original_action_dim]
 
-            actions = self.unnormalize_outputs({"action": actions})["action"]
-
             if self.config.adapt_to_pi_aloha:
                 actions = self._pi_aloha_encode_actions(actions)
 
@@ -249,8 +236,6 @@ class PI0FASTPolicy(PreTrainedPolicy):
         if self.config.adapt_to_pi_aloha:
             batch[OBS_STATE] = self._pi_aloha_decode_state(batch[OBS_STATE])
             batch[ACTION] = self._pi_aloha_encode_actions_inv(batch[ACTION])
-        batch = self.normalize_inputs(batch)
-        batch = self.normalize_targets(batch)
         loss_dict = self.model.forward(batch)
         return loss_dict["loss"], loss_dict
 

src/lerobot/policies/pi0fast/processor_pi0fast.py

@@ -0,0 +1,92 @@
+#!/usr/bin/env python
+
+# Copyright 2025 Physical Intelligence and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+
+from lerobot.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
+from lerobot.policies.pi0fast.configuration_pi0fast import PI0FASTConfig
+from lerobot.processor import (
+    AddBatchDimensionProcessorStep,
+    DeviceProcessorStep,
+    NormalizerProcessorStep,
+    PolicyProcessorPipeline,
+    ProcessorKwargs,
+    RenameObservationsProcessorStep,
+    UnnormalizerProcessorStep,
+)
+
+
+def make_pi0fast_pre_post_processors(
+    config: PI0FASTConfig,
+    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
+    preprocessor_kwargs: ProcessorKwargs | None = None,
+    postprocessor_kwargs: ProcessorKwargs | None = None,
+) -> tuple[PolicyProcessorPipeline, PolicyProcessorPipeline]:
+    """
+    Constructs pre-processor and post-processor pipelines for the PI0Fast policy.
+
+    The pre-processing pipeline prepares input data for the model by:
+    1. Renaming features to match pretrained configurations.
+    2. Normalizing input and output features based on dataset statistics.
+    3. Adding a batch dimension.
+    4. Moving all data to the specified device.
+
+    The post-processing pipeline handles the model's output by:
+    1. Moving data to the CPU.
+    2. Unnormalizing the output features to their original scale.
+
+    Args:
+        config: The configuration object for the PI0Fast policy.
+        dataset_stats: A dictionary of statistics for normalization.
+        preprocessor_kwargs: Additional arguments for the pre-processor pipeline.
+        postprocessor_kwargs: Additional arguments for the post-processor pipeline.
+
+    Returns:
+        A tuple containing the configured pre-processor and post-processor pipelines.
+    """
+    if preprocessor_kwargs is None:
+        preprocessor_kwargs = {}
+    if postprocessor_kwargs is None:
+        postprocessor_kwargs = {}
+
+    input_steps = [
+        RenameObservationsProcessorStep(rename_map={}),  # To mimic the same processor as pretrained one
+        AddBatchDimensionProcessorStep(),
+        DeviceProcessorStep(device=config.device),
+        NormalizerProcessorStep(
+            features={**config.input_features, **config.output_features},
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+        ),
+    ]
+    output_steps = [
+        DeviceProcessorStep(device="cpu"),
+        UnnormalizerProcessorStep(
+            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
+        ),
+    ]
+    return (
+        PolicyProcessorPipeline(
+            steps=input_steps,
+            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
+            **preprocessor_kwargs,
+        ),
+        PolicyProcessorPipeline(
+            steps=output_steps,
+            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
+            **postprocessor_kwargs,
+        ),
+    )

src/lerobot/policies/sac/modeling_sac.py

@@ -28,7 +28,6 @@ import torch.nn.functional as F  # noqa: N812
 from torch import Tensor
 from torch.distributions import MultivariateNormal, TanhTransform, Transform, TransformedDistribution
 
-from lerobot.policies.normalize import NormalizeBuffer
 from lerobot.policies.pretrained import PreTrainedPolicy
 from lerobot.policies.sac.configuration_sac import SACConfig, is_image_feature
 from lerobot.policies.utils import get_device_from_parameters
@@ -45,7 +44,6 @@ class SACPolicy(
     def __init__(
         self,
         config: SACConfig | None = None,
-        dataset_stats: dict[str, dict[str, Tensor]] | None = None,
     ):
         super().__init__(config)
         config.validate_features()
@@ -53,7 +51,6 @@ class SACPolicy(
 
         # Determine action dimension and initialize all components
         continuous_action_dim = config.output_features["action"].shape[0]
-        self._init_normalization(dataset_stats)
         self._init_encoders()
         self._init_critics(continuous_action_dim)
         self._init_actor(continuous_action_dim)
@@ -88,8 +85,7 @@ class SACPolicy(
 
         observations_features = None
         if self.shared_encoder and self.actor.encoder.has_images:
-            # Cache and normalize image features
-            observations_features = self.actor.encoder.get_cached_image_features(batch, normalize=True)
+            observations_features = self.actor.encoder.get_cached_image_features(batch)
 
         actions, _, _ = self.actor(batch, observations_features)
 
@@ -391,28 +387,12 @@ class SACPolicy(
         actor_loss = ((self.temperature * log_probs) - min_q_preds).mean()
         return actor_loss
 
-    def _init_normalization(self, dataset_stats):
-        """Initialize input/output normalization modules."""
-        self.normalize_inputs = nn.Identity()
-        self.normalize_targets = nn.Identity()
-        if self.config.dataset_stats is not None:
-            params = _convert_normalization_params_to_tensor(self.config.dataset_stats)
-            self.normalize_inputs = NormalizeBuffer(
-                self.config.input_features, self.config.normalization_mapping, params
-            )
-            stats = dataset_stats or params
-            self.normalize_targets = NormalizeBuffer(
-                self.config.output_features, self.config.normalization_mapping, stats
-            )
-
     def _init_encoders(self):
         """Initialize shared or separate encoders for actor and critic."""
         self.shared_encoder = self.config.shared_encoder
-        self.encoder_critic = SACObservationEncoder(self.config, self.normalize_inputs)
+        self.encoder_critic = SACObservationEncoder(self.config)
         self.encoder_actor = (
-            self.encoder_critic
-            if self.shared_encoder
-            else SACObservationEncoder(self.config, self.normalize_inputs)
+            self.encoder_critic if self.shared_encoder else SACObservationEncoder(self.config)
         )
 
     def _init_critics(self, continuous_action_dim):
@@ -424,9 +404,7 @@ class SACPolicy(
             )
             for _ in range(self.config.num_critics)
         ]
-        self.critic_ensemble = CriticEnsemble(
-            encoder=self.encoder_critic, ensemble=heads, output_normalization=self.normalize_targets
-        )
+        self.critic_ensemble = CriticEnsemble(encoder=self.encoder_critic, ensemble=heads)
         target_heads = [
             CriticHead(
                 input_dim=self.encoder_critic.output_dim + continuous_action_dim,
@@ -434,9 +412,7 @@ class SACPolicy(
             )
             for _ in range(self.config.num_critics)
         ]
-        self.critic_target = CriticEnsemble(
-            encoder=self.encoder_critic, ensemble=target_heads, output_normalization=self.normalize_targets
-        )
+        self.critic_target = CriticEnsemble(encoder=self.encoder_critic, ensemble=target_heads)
         self.critic_target.load_state_dict(self.critic_ensemble.state_dict())
 
         if self.config.use_torch_compile:
@@ -490,10 +466,9 @@ class SACPolicy(
 class SACObservationEncoder(nn.Module):
     """Encode image and/or state vector observations."""
 
-    def __init__(self, config: SACConfig, input_normalizer: nn.Module) -> None:
+    def __init__(self, config: SACConfig) -> None:
         super().__init__()
         self.config = config
-        self.input_normalization = input_normalizer
         self._init_image_layers()
         self._init_state_layers()
         self._compute_output_dim()
@@ -568,11 +543,10 @@ class SACObservationEncoder(nn.Module):
     def forward(
         self, obs: dict[str, Tensor], cache: dict[str, Tensor] | None = None, detach: bool = False
     ) -> Tensor:
-        obs = self.input_normalization(obs)
         parts = []
         if self.has_images:
             if cache is None:
-                cache = self.get_cached_image_features(obs, normalize=False)
+                cache = self.get_cached_image_features(obs)
             parts.append(self._encode_images(cache, detach))
         if self.has_env:
             parts.append(self.env_encoder(obs["observation.environment_state"]))
@@ -585,7 +559,7 @@ class SACObservationEncoder(nn.Module):
             "No parts to concatenate, you should have at least one image or environment state or state"
         )
 
-    def get_cached_image_features(self, obs: dict[str, Tensor], normalize: bool = False) -> dict[str, Tensor]:
+    def get_cached_image_features(self, obs: dict[str, Tensor]) -> dict[str, Tensor]:
         """Extract and optionally cache image features from observations.
 
         This function processes image observations through the vision encoder once and returns
@@ -597,26 +571,17 @@ class SACObservationEncoder(nn.Module):
         - The vision encoder forward pass is typically the main computational bottleneck during training and inference
         - Caching these features can provide 2-4x speedup in training and inference
 
-        Normalization behavior:
-        - When called from inside forward(): set normalize=False since inputs are already normalized
-        - When called from outside forward(): set normalize=True to ensure proper input normalization
-
         Usage patterns:
-        - Called in select_action() with normalize=True
+        - Called in select_action()
         - Called in learner.py's get_observation_features() to pre-compute features for all policy components
-        - Called internally by forward() with normalize=False
+        - Called internally by forward()
 
         Args:
             obs: Dictionary of observation tensors containing image keys
-            normalize: Whether to normalize observations before encoding
-                      Set to True when calling directly from outside the encoder's forward method
-                      Set to False when calling from within forward() where inputs are already normalized
 
         Returns:
             Dictionary mapping image keys to their corresponding encoded features
         """
-        if normalize:
-            obs = self.input_normalization(obs)
         batched = torch.cat([obs[k] for k in self.image_keys], dim=0)
         out = self.image_encoder(batched)
         chunks = torch.chunk(out, len(self.image_keys), dim=0)
@@ -747,7 +712,6 @@ class CriticEnsemble(nn.Module):
     Args:
         encoder (SACObservationEncoder): encoder for observations.
         ensemble (List[CriticHead]): list of critic heads.
-        output_normalization (nn.Module): normalization layer for actions.
         init_final (float | None): optional initializer scale for final layers.
 
     Forward returns a tensor of shape (num_critics, batch_size) containing Q-values.
@@ -757,13 +721,11 @@ class CriticEnsemble(nn.Module):
         self,
         encoder: SACObservationEncoder,
         ensemble: list[CriticHead],
-        output_normalization: nn.Module,
         init_final: float | None = None,
     ):
         super().__init__()
         self.encoder = encoder
         self.init_final = init_final
-        self.output_normalization = output_normalization
         self.critics = nn.ModuleList(ensemble)
 
     def forward(
@@ -775,11 +737,6 @@ class CriticEnsemble(nn.Module):
         device = get_device_from_parameters(self)
         # Move each tensor in observations to device
         observations = {k: v.to(device) for k, v in observations.items()}
-        # NOTE: We normalize actions it helps for sample efficiency
-        actions: dict[str, torch.tensor] = {"action": actions}
-        # NOTE: Normalization layer took dict in input and outputs a dict that why
-        actions = self.output_normalization(actions)["action"]
-        actions = actions.to(device)
 
         obs_enc = self.encoder(observations, cache=observation_features)
 

src/lerobot/policies/sac/processor_sac.py

@@ -0,0 +1,93 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team.
+# All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+
+from lerobot.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
+from lerobot.policies.sac.configuration_sac import SACConfig
+from lerobot.processor import (
+    AddBatchDimensionProcessorStep,
+    DeviceProcessorStep,
+    NormalizerProcessorStep,
+    PolicyProcessorPipeline,
+    ProcessorKwargs,
+    RenameObservationsProcessorStep,
+    UnnormalizerProcessorStep,
+)
+
+
+def make_sac_pre_post_processors(
+    config: SACConfig,
+    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
+    preprocessor_kwargs: ProcessorKwargs | None = None,
+    postprocessor_kwargs: ProcessorKwargs | None = None,
+) -> tuple[PolicyProcessorPipeline, PolicyProcessorPipeline]:
+    """
+    Constructs pre-processor and post-processor pipelines for the SAC policy.
+
+    The pre-processing pipeline prepares input data for the model by:
+    1. Renaming features to match pretrained configurations.
+    2. Normalizing input and output features based on dataset statistics.
+    3. Adding a batch dimension.
+    4. Moving all data to the specified device.
+
+    The post-processing pipeline handles the model's output by:
+    1. Moving data to the CPU.
+    2. Unnormalizing the output features to their original scale.
+
+    Args:
+        config: The configuration object for the SAC policy.
+        dataset_stats: A dictionary of statistics for normalization.
+        preprocessor_kwargs: Additional arguments for the pre-processor pipeline.
+        postprocessor_kwargs: Additional arguments for the post-processor pipeline.
+
+    Returns:
+        A tuple containing the configured pre-processor and post-processor pipelines.
+    """
+    if preprocessor_kwargs is None:
+        preprocessor_kwargs = {}
+    if postprocessor_kwargs is None:
+        postprocessor_kwargs = {}
+
+    input_steps = [
+        RenameObservationsProcessorStep(rename_map={}),
+        AddBatchDimensionProcessorStep(),
+        DeviceProcessorStep(device=config.device),
+        NormalizerProcessorStep(
+            features={**config.input_features, **config.output_features},
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+        ),
+    ]
+    output_steps = [
+        DeviceProcessorStep(device="cpu"),
+        UnnormalizerProcessorStep(
+            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
+        ),
+    ]
+    return (
+        PolicyProcessorPipeline(
+            steps=input_steps,
+            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
+            **preprocessor_kwargs,
+        ),
+        PolicyProcessorPipeline(
+            steps=output_steps,
+            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
+            **postprocessor_kwargs,
+        ),
+    )

src/lerobot/policies/sac/reward_model/modeling_classifier.py

@@ -20,7 +20,6 @@ import torch
 from torch import Tensor, nn
 
 from lerobot.constants import OBS_IMAGE, REWARD
-from lerobot.policies.normalize import Normalize, Unnormalize
 from lerobot.policies.pretrained import PreTrainedPolicy
 from lerobot.policies.sac.reward_model.configuration_classifier import RewardClassifierConfig
 
@@ -108,22 +107,12 @@ class Classifier(PreTrainedPolicy):
     def __init__(
         self,
         config: RewardClassifierConfig,
-        dataset_stats: dict[str, dict[str, Tensor]] | None = None,
     ):
         from transformers import AutoModel
 
         super().__init__(config)
         self.config = config
 
-        # Initialize normalization (standardized with the policy framework)
-        self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
-        self.normalize_targets = Normalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-        self.unnormalize_outputs = Unnormalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-
         # Set up encoder
         encoder = AutoModel.from_pretrained(self.config.model_name, trust_remote_code=True)
         # Extract vision model if we're given a multimodal model
@@ -247,10 +236,6 @@ class Classifier(PreTrainedPolicy):
 
     def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict[str, Tensor]]:
         """Standard forward pass for training compatible with train.py."""
-        # Normalize inputs if needed
-        batch = self.normalize_inputs(batch)
-        batch = self.normalize_targets(batch)
-
         # Extract images and labels
         images, labels = self.extract_images_and_labels(batch)
 

src/lerobot/policies/sac/reward_model/processor_classifier.py

@@ -0,0 +1,81 @@
+# !/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import torch
+
+from lerobot.policies.sac.reward_model.configuration_classifier import RewardClassifierConfig
+from lerobot.processor import (
+    DeviceProcessorStep,
+    IdentityProcessorStep,
+    NormalizerProcessorStep,
+    PolicyProcessorPipeline,
+    ProcessorKwargs,
+)
+
+
+def make_classifier_processor(
+    config: RewardClassifierConfig,
+    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
+    preprocessor_kwargs: ProcessorKwargs | None = None,
+    postprocessor_kwargs: ProcessorKwargs | None = None,
+) -> tuple[PolicyProcessorPipeline, PolicyProcessorPipeline]:
+    """
+    Constructs pre-processor and post-processor pipelines for the reward classifier.
+
+    The pre-processing pipeline prepares input data for the classifier by:
+    1. Normalizing both input and output features based on dataset statistics.
+    2. Moving the data to the specified device.
+
+    The post-processing pipeline handles the classifier's output by:
+    1. Moving the data to the CPU.
+    2. Applying an identity step, as no unnormalization is needed for the output logits.
+
+    Args:
+        config: The configuration object for the RewardClassifier.
+        dataset_stats: A dictionary of statistics for normalization.
+        preprocessor_kwargs: Additional arguments for the pre-processor pipeline.
+        postprocessor_kwargs: Additional arguments for the post-processor pipeline.
+
+    Returns:
+        A tuple containing the configured pre-processor and post-processor pipelines.
+    """
+    if preprocessor_kwargs is None:
+        preprocessor_kwargs = {}
+    if postprocessor_kwargs is None:
+        postprocessor_kwargs = {}
+
+    input_steps = [
+        NormalizerProcessorStep(
+            features=config.input_features, norm_map=config.normalization_mapping, stats=dataset_stats
+        ),
+        NormalizerProcessorStep(
+            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
+        ),
+        DeviceProcessorStep(device=config.device),
+    ]
+    output_steps = [DeviceProcessorStep(device="cpu"), IdentityProcessorStep()]
+
+    return (
+        PolicyProcessorPipeline(
+            steps=input_steps,
+            name="classifier_preprocessor",
+            **preprocessor_kwargs,
+        ),
+        PolicyProcessorPipeline(
+            steps=output_steps,
+            name="classifier_postprocessor",
+            **postprocessor_kwargs,
+        ),
+    )

src/lerobot/policies/smolvla/modeling_smolvla.py

@@ -28,7 +28,7 @@ pip install -e ".[smolvla]"
 
 Example of finetuning the smolvla pretrained model (`smolvla_base`):
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
 --policy.path=lerobot/smolvla_base \
 --dataset.repo_id=danaaubakirova/svla_so100_task1_v3 \
 --batch_size=64 \
@@ -38,7 +38,7 @@ python -m lerobot.scripts.train \
 Example of finetuning a smolVLA. SmolVLA is composed of a pretrained VLM,
 and an action expert.
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
 --policy.type=smolvla \
 --dataset.repo_id=danaaubakirova/svla_so100_task1_v3 \
 --batch_size=64 \
@@ -53,21 +53,13 @@ policy = SmolVLAPolicy.from_pretrained("lerobot/smolvla_base")
 """
 
 import math
-import os
-import re
 from collections import deque
 
-import safetensors
 import torch
 import torch.nn.functional as F  # noqa: N812
 from torch import Tensor, nn
-from transformers import AutoProcessor
 
-from lerobot.constants import ACTION, OBS_STATE
-from lerobot.policies.normalize import (
-    Normalize,
-    Unnormalize,
-)
+from lerobot.constants import ACTION, OBS_LANGUAGE, OBS_STATE
 from lerobot.policies.pretrained import PreTrainedPolicy
 from lerobot.policies.smolvla.configuration_smolvla import SmolVLAConfig
 from lerobot.policies.smolvla.smolvlm_with_expert import SmolVLMWithExpertModel
@@ -76,102 +68,6 @@ from lerobot.policies.utils import (
 )
 from lerobot.utils.utils import get_safe_dtype
 
-# Matches ".soNNN", optionally followed by "-something", up to the "_buffer_" marker
-_VARIANT_RE = re.compile(r"\.so\d+(?:-[\w]+)?_buffer_")
-
-
-def canonicalise(k: str) -> str:
-    """
-    Remove dataset-variant markers like '.so100-blue_' or '.so100_' from a
-    normalisation-buffer key.
-    """
-    return _VARIANT_RE.sub(".buffer_", k)
-
-
-def standardise_state_dict(
-    checkpoint: dict[str, torch.Tensor], ref_keys: set[str], *, verbose: bool = True
-) -> tuple[dict[str, torch.Tensor], list[str]]:
-    """
-    • Re-keys `checkpoint ` so that every entry matches the *reference* key set.
-    • If several variant keys collapse to the same canonical name we keep the
-      first one and log the collision.
-    • Returns the new dict + a list of entries that could not be matched.
-    """
-    out, collisions, unmatched = {}, {}, []
-
-    for k, v in checkpoint.items():
-        canon = canonicalise(k)
-        if canon in ref_keys:
-            if canon in out:  # duplicate after collapsing
-                collisions.setdefault(canon, []).append(k)
-            else:
-                out[canon] = v
-        else:
-            unmatched.append(k)
-
-    if verbose:
-        for canon, variants in collisions.items():
-            print(f"[standardise_state_dict] '{canon}'  ←  {variants}")
-        if unmatched:
-            print(f"[standardise_state_dict] kept {len(unmatched)} unmatched keys")
-
-    out.update({k: checkpoint[k] for k in unmatched})
-    return out, unmatched
-
-
-def rename_checkpoint_keys(checkpoint: dict, rename_str: str):
-    """
-    Renames keys in a checkpoint dictionary based on the given rename string.
-
-    Args:
-        checkpoint (dict): The checkpoint dictionary.
-        rename_str (str): A string specifying key mappings in the format "old1//new1,old2//new2".
-
-    Returns:
-        dict: The modified checkpoint with renamed keys.
-    """
-
-    rename_dict = dict(pair.split("//") for pair in rename_str.split(","))
-
-    new_checkpoint = {}
-    for k, v in checkpoint.items():
-        for old_key, new_key in rename_dict.items():
-            if old_key in k:
-                k = k.replace(old_key, new_key)
-        new_checkpoint[k] = v
-    return new_checkpoint
-
-
-def load_smolvla(
-    model: torch.nn.Module,
-    filename: str | os.PathLike,
-    *,
-    device: str = "cpu",
-    checkpoint_keys_mapping: str = "",
-) -> torch.nn.Module:
-    state_dict = safetensors.torch.load_file(filename, device=device)
-
-    # Optional user-supplied renames (e.g. "model._orig_mod.//model.")
-    if checkpoint_keys_mapping and "//" in checkpoint_keys_mapping:
-        state_dict = rename_checkpoint_keys(state_dict, checkpoint_keys_mapping)
-
-    state_dict, _ = standardise_state_dict(state_dict, set(model.state_dict().keys()))
-
-    # HACK(aliberts): to not overwrite normalization parameters as they should come from the dataset
-    norm_keys = ("normalize_inputs", "normalize_targets", "unnormalize_outputs")
-    state_dict = {k: v for k, v in state_dict.items() if not k.startswith(norm_keys)}
-
-    missing, unexpected = model.load_state_dict(state_dict, strict=False)
-
-    if not all(key.startswith(norm_keys) for key in missing) or unexpected:
-        raise RuntimeError(
-            "SmolVLA %d missing / %d unexpected keys",
-            len(missing),
-            len(unexpected),
-        )
-
-    return model
-
 
 def create_sinusoidal_pos_embedding(
     time: torch.tensor, dimension: int, min_period: float, max_period: float, device="cpu"
@@ -326,28 +222,17 @@ class SmolVLAPolicy(PreTrainedPolicy):
     def __init__(
         self,
         config: SmolVLAConfig,
-        dataset_stats: dict[str, dict[str, Tensor]] | None = None,
     ):
         """
         Args:
             config: Policy configuration class instance or None, in which case the default instantiation of
                     the configuration class is used.
-            dataset_stats: Dataset statistics to be used for normalization. If not passed here, it is expected
-                that they will be passed with a call to `load_state_dict` before the policy is used.
         """
 
         super().__init__(config)
         config.validate_features()
         self.config = config
-        self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
-        self.normalize_targets = Normalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-        self.unnormalize_outputs = Unnormalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
 
-        self.language_tokenizer = AutoProcessor.from_pretrained(self.config.vlm_model_name).tokenizer
         self.model = VLAFlowMatching(config)
         self.reset()
 
@@ -357,23 +242,6 @@ class SmolVLAPolicy(PreTrainedPolicy):
             ACTION: deque(maxlen=self.config.n_action_steps),
         }
 
-    # HACK(aliberts, danaaubakirova): we overwrite this classmethod here to fix smolVLA-specific issues
-    @classmethod
-    def _load_as_safetensor(
-        cls,
-        model: "SmolVLAPolicy",
-        model_file: str,
-        map_location: str,
-        strict: bool,
-    ):
-        safetensors.torch.load_model(model, model_file, strict=strict, device=map_location)
-        return load_smolvla(
-            model,
-            model_file,
-            device=map_location,
-            checkpoint_keys_mapping="model._orig_mod.//model.",
-        )
-
     def get_optim_params(self) -> dict:
         return self.parameters()
 
@@ -389,7 +257,8 @@ class SmolVLAPolicy(PreTrainedPolicy):
 
         images, img_masks = self.prepare_images(batch)
         state = self.prepare_state(batch)
-        lang_tokens, lang_masks = self.prepare_language(batch)
+        lang_tokens = batch[f"{OBS_LANGUAGE}.tokens"]
+        lang_masks = batch[f"{OBS_LANGUAGE}.attention_mask"]
 
         actions = self.model.sample_actions(images, img_masks, lang_tokens, lang_masks, state, noise=noise)
 
@@ -397,8 +266,6 @@ class SmolVLAPolicy(PreTrainedPolicy):
         original_action_dim = self.config.action_feature.shape[0]
         actions = actions[:, :, :original_action_dim]
 
-        actions = self.unnormalize_outputs({ACTION: actions})[ACTION]
-
         if self.config.adapt_to_pi_aloha:
             actions = self._pi_aloha_encode_actions(actions)
 
@@ -408,8 +275,6 @@ class SmolVLAPolicy(PreTrainedPolicy):
         if self.config.adapt_to_pi_aloha:
             batch[OBS_STATE] = self._pi_aloha_decode_state(batch[OBS_STATE])
 
-        batch = self.normalize_inputs(batch)
-
         return batch
 
     @torch.no_grad()
@@ -450,11 +315,11 @@ class SmolVLAPolicy(PreTrainedPolicy):
         if self.config.adapt_to_pi_aloha:
             batch[OBS_STATE] = self._pi_aloha_decode_state(batch[OBS_STATE])
             batch[ACTION] = self._pi_aloha_encode_actions_inv(batch[ACTION])
-        batch = self.normalize_inputs(batch)
-        batch = self.normalize_targets(batch)
+
         images, img_masks = self.prepare_images(batch)
         state = self.prepare_state(batch)
-        lang_tokens, lang_masks = self.prepare_language(batch)
+        lang_tokens = batch[f"{OBS_LANGUAGE}.tokens"]
+        lang_masks = batch[f"{OBS_LANGUAGE}.attention_mask"]
         actions = self.prepare_action(batch)
         actions_is_pad = batch.get("actions_id_pad")
         loss_dict = {}
@@ -518,30 +383,6 @@ class SmolVLAPolicy(PreTrainedPolicy):
             img_masks.append(mask)
         return images, img_masks
 
-    def prepare_language(self, batch) -> tuple[Tensor, Tensor]:
-        """Tokenize the text input"""
-        device = batch[OBS_STATE].device
-        tasks = batch["task"]
-        if isinstance(tasks, str):
-            tasks = [tasks]
-
-        if len(tasks) == 1:
-            tasks = [tasks[0] for _ in range(batch[OBS_STATE].shape[0])]
-
-        tasks = [task if task.endswith("\n") else f"{task}\n" for task in tasks]
-
-        tokenized_prompt = self.language_tokenizer.__call__(
-            tasks,
-            padding=self.config.pad_language_to,
-            padding_side="right",
-            max_length=self.config.tokenizer_max_length,
-            return_tensors="pt",
-        )
-        lang_tokens = tokenized_prompt["input_ids"].to(device=device)
-        lang_masks = tokenized_prompt["attention_mask"].to(device=device, dtype=torch.bool)
-
-        return lang_tokens, lang_masks
-
     def _pi_aloha_decode_state(self, state):
         # Flip the joints.
         for motor_idx in [1, 2, 8, 9]:

src/lerobot/policies/smolvla/processor_smolvla.py

@@ -0,0 +1,143 @@
+#!/usr/bin/env python
+
+# Copyright 2025 HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+
+from lerobot.configs.types import PipelineFeatureType, PolicyFeature
+from lerobot.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
+from lerobot.policies.smolvla.configuration_smolvla import SmolVLAConfig
+from lerobot.processor import (
+    AddBatchDimensionProcessorStep,
+    ComplementaryDataProcessorStep,
+    DeviceProcessorStep,
+    NormalizerProcessorStep,
+    PolicyProcessorPipeline,
+    ProcessorKwargs,
+    ProcessorStepRegistry,
+    RenameObservationsProcessorStep,
+    TokenizerProcessorStep,
+    UnnormalizerProcessorStep,
+)
+
+
+def make_smolvla_pre_post_processors(
+    config: SmolVLAConfig,
+    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
+    preprocessor_kwargs: ProcessorKwargs | None = None,
+    postprocessor_kwargs: ProcessorKwargs | None = None,
+) -> tuple[PolicyProcessorPipeline, PolicyProcessorPipeline]:
+    """
+    Constructs pre-processor and post-processor pipelines for the SmolVLA policy.
+
+    The pre-processing pipeline prepares input data for the model by:
+    1.  Renaming features to match pretrained configurations.
+    2.  Normalizing input and output features based on dataset statistics.
+    3.  Adding a batch dimension.
+    4.  Ensuring the language task description ends with a newline character.
+    5.  Tokenizing the language task description.
+    6.  Moving all data to the specified device.
+
+    The post-processing pipeline handles the model's output by:
+    1.  Moving data to the CPU.
+    2.  Unnormalizing the output actions to their original scale.
+
+    Args:
+        config: The configuration object for the SmolVLA policy.
+        dataset_stats: A dictionary of statistics for normalization.
+        preprocessor_kwargs: Additional arguments for the pre-processor pipeline.
+        postprocessor_kwargs: Additional arguments for the post-processor pipeline.
+
+    Returns:
+        A tuple containing the configured pre-processor and post-processor pipelines.
+    """
+    if preprocessor_kwargs is None:
+        preprocessor_kwargs = {}
+    if postprocessor_kwargs is None:
+        postprocessor_kwargs = {}
+
+    input_steps = [
+        RenameObservationsProcessorStep(rename_map={}),  # To mimic the same processor as pretrained one
+        AddBatchDimensionProcessorStep(),
+        SmolVLANewLineProcessor(),
+        TokenizerProcessorStep(
+            tokenizer_name=config.vlm_model_name,
+            padding=config.pad_language_to,
+            padding_side="right",
+            max_length=config.tokenizer_max_length,
+        ),
+        DeviceProcessorStep(device=config.device),
+        NormalizerProcessorStep(
+            features={**config.input_features, **config.output_features},
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+        ),
+    ]
+    output_steps = [
+        DeviceProcessorStep(device="cpu"),
+        UnnormalizerProcessorStep(
+            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
+        ),
+    ]
+    return (
+        PolicyProcessorPipeline(
+            steps=input_steps,
+            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
+            **preprocessor_kwargs,
+        ),
+        PolicyProcessorPipeline(
+            steps=output_steps,
+            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
+            **postprocessor_kwargs,
+        ),
+    )
+
+
+@ProcessorStepRegistry.register(name="smolvla_new_line_processor")
+class SmolVLANewLineProcessor(ComplementaryDataProcessorStep):
+    """
+    A processor step that ensures the 'task' description ends with a newline character.
+
+    This step is necessary for certain tokenizers (e.g., PaliGemma) that expect a
+    newline at the end of the prompt. It handles both single string tasks and lists
+    of string tasks.
+    """
+
+    def complementary_data(self, complementary_data):
+        if "task" not in complementary_data:
+            return complementary_data
+
+        task = complementary_data["task"]
+        if task is None:
+            return complementary_data
+
+        new_complementary_data = dict(complementary_data)
+
+        # Handle both string and list of strings
+        if isinstance(task, str):
+            # Single string: add newline if not present
+            if not task.endswith("\n"):
+                new_complementary_data["task"] = f"{task}\n"
+        elif isinstance(task, list) and all(isinstance(t, str) for t in task):
+            # List of strings: add newline to each if not present
+            new_complementary_data["task"] = [t if t.endswith("\n") else f"{t}\n" for t in task]
+        # If task is neither string nor list of strings, leave unchanged
+
+        return new_complementary_data
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        return features

src/lerobot/policies/tdmpc/modeling_tdmpc.py

@@ -36,7 +36,6 @@ import torch.nn.functional as F  # noqa: N812
 from torch import Tensor
 
 from lerobot.constants import ACTION, OBS_ENV_STATE, OBS_IMAGE, OBS_STATE, REWARD
-from lerobot.policies.normalize import Normalize, Unnormalize
 from lerobot.policies.pretrained import PreTrainedPolicy
 from lerobot.policies.tdmpc.configuration_tdmpc import TDMPCConfig
 from lerobot.policies.utils import get_device_from_parameters, get_output_shape, populate_queues
@@ -63,26 +62,19 @@ class TDMPCPolicy(PreTrainedPolicy):
     config_class = TDMPCConfig
     name = "tdmpc"
 
-    def __init__(self, config: TDMPCConfig, dataset_stats: dict[str, dict[str, Tensor]] | None = None):
+    def __init__(
+        self,
+        config: TDMPCConfig,
+    ):
         """
         Args:
             config: Policy configuration class instance or None, in which case the default instantiation of
                 the configuration class is used.
-            dataset_stats: Dataset statistics to be used for normalization. If not passed here, it is expected
-                that they will be passed with a call to `load_state_dict` before the policy is used.
         """
         super().__init__(config)
         config.validate_features()
         self.config = config
 
-        self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
-        self.normalize_targets = Normalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-        self.unnormalize_outputs = Unnormalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-
         self.model = TDMPCTOLD(config)
         self.model_target = deepcopy(self.model)
         for param in self.model_target.parameters():
@@ -137,7 +129,6 @@ class TDMPCPolicy(PreTrainedPolicy):
 
         actions = torch.clamp(actions, -1, +1)
 
-        actions = self.unnormalize_outputs({ACTION: actions})[ACTION]
         return actions
 
     @torch.no_grad()
@@ -147,11 +138,12 @@ class TDMPCPolicy(PreTrainedPolicy):
         if ACTION in batch:
             batch.pop(ACTION)
 
-        batch = self.normalize_inputs(batch)
-
         if self.config.image_features:
             batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
             batch[OBS_IMAGE] = batch[next(iter(self.config.image_features))]
+        # NOTE: for offline evaluation, we have action in the batch, so we need to pop it out
+        if ACTION in batch:
+            batch.pop(ACTION)
 
         self._queues = populate_queues(self._queues, batch)
 
@@ -320,11 +312,9 @@ class TDMPCPolicy(PreTrainedPolicy):
         """
         device = get_device_from_parameters(self)
 
-        batch = self.normalize_inputs(batch)
         if self.config.image_features:
             batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
             batch[OBS_IMAGE] = batch[next(iter(self.config.image_features))]
-        batch = self.normalize_targets(batch)
 
         info = {}
 

src/lerobot/policies/tdmpc/processor_tdmpc.py

@@ -0,0 +1,92 @@
+#!/usr/bin/env python
+
+# Copyright 2024 Nicklas Hansen, Xiaolong Wang, Hao Su,
+# and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import torch
+
+from lerobot.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
+from lerobot.policies.tdmpc.configuration_tdmpc import TDMPCConfig
+from lerobot.processor import (
+    AddBatchDimensionProcessorStep,
+    DeviceProcessorStep,
+    NormalizerProcessorStep,
+    PolicyProcessorPipeline,
+    ProcessorKwargs,
+    RenameObservationsProcessorStep,
+    UnnormalizerProcessorStep,
+)
+
+
+def make_tdmpc_pre_post_processors(
+    config: TDMPCConfig,
+    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
+    preprocessor_kwargs: ProcessorKwargs | None = None,
+    postprocessor_kwargs: ProcessorKwargs | None = None,
+) -> tuple[PolicyProcessorPipeline, PolicyProcessorPipeline]:
+    """
+    Constructs pre-processor and post-processor pipelines for the TDMPC policy.
+
+    The pre-processing pipeline prepares input data for the model by:
+    1. Renaming features to match pretrained configurations.
+    2. Normalizing input and output features based on dataset statistics.
+    3. Adding a batch dimension.
+    4. Moving all data to the specified device.
+
+    The post-processing pipeline handles the model's output by:
+    1. Moving data to the CPU.
+    2. Unnormalizing the output features to their original scale.
+
+    Args:
+        config: The configuration object for the TDMPC policy.
+        dataset_stats: A dictionary of statistics for normalization.
+        preprocessor_kwargs: Additional arguments for the pre-processor pipeline.
+        postprocessor_kwargs: Additional arguments for the post-processor pipeline.
+
+    Returns:
+        A tuple containing the configured pre-processor and post-processor pipelines.
+    """
+    if preprocessor_kwargs is None:
+        preprocessor_kwargs = {}
+    if postprocessor_kwargs is None:
+        postprocessor_kwargs = {}
+
+    input_steps = [
+        RenameObservationsProcessorStep(rename_map={}),
+        AddBatchDimensionProcessorStep(),
+        DeviceProcessorStep(device=config.device),
+        NormalizerProcessorStep(
+            features={**config.input_features, **config.output_features},
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+        ),
+    ]
+    output_steps = [
+        DeviceProcessorStep(device="cpu"),
+        UnnormalizerProcessorStep(
+            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
+        ),
+    ]
+    return (
+        PolicyProcessorPipeline(
+            steps=input_steps,
+            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
+            **preprocessor_kwargs,
+        ),
+        PolicyProcessorPipeline(
+            steps=output_steps,
+            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
+            **postprocessor_kwargs,
+        ),
+    )

src/lerobot/policies/vqbet/modeling_vqbet.py

@@ -28,7 +28,6 @@ import torchvision
 from torch import Tensor, nn
 
 from lerobot.constants import ACTION, OBS_IMAGES, OBS_STATE
-from lerobot.policies.normalize import Normalize, Unnormalize
 from lerobot.policies.pretrained import PreTrainedPolicy
 from lerobot.policies.utils import get_device_from_parameters, get_output_shape, populate_queues
 from lerobot.policies.vqbet.configuration_vqbet import VQBeTConfig
@@ -48,7 +47,6 @@ class VQBeTPolicy(PreTrainedPolicy):
     def __init__(
         self,
         config: VQBeTConfig | None = None,
-        dataset_stats: dict[str, dict[str, Tensor]] | None = None,
     ):
         """
         Args:
@@ -61,14 +59,6 @@ class VQBeTPolicy(PreTrainedPolicy):
         config.validate_features()
         self.config = config
 
-        self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
-        self.normalize_targets = Normalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-        self.unnormalize_outputs = Unnormalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-
         self.vqbet = VQBeTModel(config)
 
         self.reset()
@@ -128,7 +118,6 @@ class VQBeTPolicy(PreTrainedPolicy):
     def predict_action_chunk(self, batch: dict[str, Tensor]) -> Tensor:
         batch = {k: torch.stack(list(self._queues[k]), dim=1) for k in batch if k in self._queues}
         actions = self.vqbet(batch, rollout=True)[:, : self.config.action_chunk_size]
-        actions = self.unnormalize_outputs({ACTION: actions})[ACTION]
         return actions
 
     @torch.no_grad()
@@ -142,10 +131,12 @@ class VQBeTPolicy(PreTrainedPolicy):
         # NOTE: for offline evaluation, we have action in the batch, so we need to pop it out
         if ACTION in batch:
             batch.pop(ACTION)
-        batch = self.normalize_inputs(batch)
         batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
         # NOTE: It's important that this happens after stacking the images into a single key.
         batch["observation.images"] = torch.stack([batch[key] for key in self.config.image_features], dim=-4)
+        # NOTE: for offline evaluation, we have action in the batch, so we need to pop it out
+        if ACTION in batch:
+            batch.pop(ACTION)
 
         self._queues = populate_queues(self._queues, batch)
 
@@ -165,10 +156,8 @@ class VQBeTPolicy(PreTrainedPolicy):
 
     def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict]:
         """Run the batch through the model and compute the loss for training or validation."""
-        batch = self.normalize_inputs(batch)
         batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
         batch[OBS_IMAGES] = torch.stack([batch[key] for key in self.config.image_features], dim=-4)
-        batch = self.normalize_targets(batch)
         # VQ-BeT discretizes action using VQ-VAE before training BeT (please refer to section 3.2 in the VQ-BeT paper https://huggingface.co/papers/2403.03181)
         if not self.vqbet.action_head.vqvae_model.discretized.item():
             # loss: total loss of training RVQ

src/lerobot/policies/vqbet/processor_vqbet.py

@@ -0,0 +1,93 @@
+#!/usr/bin/env python
+
+# Copyright 2024 Seungjae Lee and Yibin Wang and Haritheja Etukuru
+# and H. Jin Kim and Nur Muhammad Mahi Shafiullah and Lerrel Pinto
+# and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import torch
+
+from lerobot.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
+from lerobot.policies.vqbet.configuration_vqbet import VQBeTConfig
+from lerobot.processor import (
+    AddBatchDimensionProcessorStep,
+    DeviceProcessorStep,
+    NormalizerProcessorStep,
+    PolicyProcessorPipeline,
+    ProcessorKwargs,
+    RenameObservationsProcessorStep,
+    UnnormalizerProcessorStep,
+)
+
+
+def make_vqbet_pre_post_processors(
+    config: VQBeTConfig,
+    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
+    preprocessor_kwargs: ProcessorKwargs | None = None,
+    postprocessor_kwargs: ProcessorKwargs | None = None,
+) -> tuple[PolicyProcessorPipeline, PolicyProcessorPipeline]:
+    """
+    Constructs pre-processor and post-processor pipelines for the VQ-BeT policy.
+
+    The pre-processing pipeline prepares input data for the model by:
+    1. Renaming features, allowing customization to match pretrained configurations.
+    2. Normalizing input and output features based on dataset statistics.
+    3. Adding a batch dimension.
+    4. Moving all data to the specified device.
+
+    The post-processing pipeline handles the model's output by:
+    1. Moving data to the CPU.
+    2. Unnormalizing the output features to their original scale.
+
+    Args:
+        config: The configuration object for the VQ-BeT policy.
+        dataset_stats: A dictionary of statistics for normalization.
+        preprocessor_kwargs: Additional arguments for the pre-processor pipeline.
+        postprocessor_kwargs: Additional arguments for the post-processor pipeline.
+
+    Returns:
+        A tuple containing the configured pre-processor and post-processor pipelines.
+    """
+    if preprocessor_kwargs is None:
+        preprocessor_kwargs = {}
+    if postprocessor_kwargs is None:
+        postprocessor_kwargs = {}
+
+    input_steps = [
+        RenameObservationsProcessorStep(rename_map={}),  # Let the possibility to the user to rename the keys
+        AddBatchDimensionProcessorStep(),
+        DeviceProcessorStep(device=config.device),
+        NormalizerProcessorStep(
+            features={**config.input_features, **config.output_features},
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+        ),
+    ]
+    output_steps = [
+        DeviceProcessorStep(device="cpu"),
+        UnnormalizerProcessorStep(
+            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
+        ),
+    ]
+    return (
+        PolicyProcessorPipeline(
+            steps=input_steps,
+            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
+            **preprocessor_kwargs,
+        ),
+        PolicyProcessorPipeline(
+            steps=output_steps,
+            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
+            **postprocessor_kwargs,
+        ),
+    )

src/lerobot/processor/__init__.py

@@ -14,41 +14,90 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from .device_processor import DeviceProcessor
-from .normalize_processor import NormalizerProcessor, UnnormalizerProcessor
-from .observation_processor import VanillaObservationProcessor
+from .batch_processor import AddBatchDimensionProcessorStep
+from .converters import (
+    batch_to_transition,
+    create_transition,
+    merge_transitions,
+    transition_to_batch,
+    transition_to_dataset_frame,
+)
+from .core import EnvTransition, TransitionKey
+from .delta_action_processor import MapDeltaActionToRobotActionStep, MapTensorToDeltaActionDictStep
+from .device_processor import DeviceProcessorStep
+from .gym_action_processor import Numpy2TorchActionProcessorStep, Torch2NumpyActionProcessorStep
+from .hil_processor import (
+    AddTeleopActionAsComplimentaryDataStep,
+    AddTeleopEventsAsInfoStep,
+    GripperPenaltyProcessorStep,
+    ImageCropResizeProcessorStep,
+    InterventionActionProcessorStep,
+    RewardClassifierProcessorStep,
+    TimeLimitProcessorStep,
+)
+from .joint_observations_processor import JointVelocityProcessorStep, MotorCurrentProcessorStep
+from .normalize_processor import NormalizerProcessorStep, UnnormalizerProcessorStep, hotswap_stats
+from .observation_processor import VanillaObservationProcessorStep
 from .pipeline import (
-    ActionProcessor,
-    DoneProcessor,
-    EnvTransition,
-    IdentityProcessor,
-    InfoProcessor,
-    ObservationProcessor,
+    ActionProcessorStep,
+    ComplementaryDataProcessorStep,
+    DataProcessorPipeline,
+    DoneProcessorStep,
+    IdentityProcessorStep,
+    InfoProcessorStep,
+    ObservationProcessorStep,
+    PolicyProcessorPipeline,
+    ProcessorKwargs,
     ProcessorStep,
     ProcessorStepRegistry,
-    RewardProcessor,
-    RobotProcessor,
-    TransitionKey,
-    TruncatedProcessor,
+    RewardProcessorStep,
+    RobotProcessorPipeline,
+    TruncatedProcessorStep,
 )
-from .rename_processor import RenameProcessor
+from .rename_processor import RenameObservationsProcessorStep
+from .tokenizer_processor import TokenizerProcessorStep
 
 __all__ = [
-    "ActionProcessor",
-    "DeviceProcessor",
-    "DoneProcessor",
+    "ActionProcessorStep",
+    "AddTeleopActionAsComplimentaryDataStep",
+    "AddTeleopEventsAsInfoStep",
+    "ComplementaryDataProcessorStep",
+    "batch_to_transition",
+    "create_transition",
+    "DeviceProcessorStep",
+    "DoneProcessorStep",
     "EnvTransition",
-    "IdentityProcessor",
-    "InfoProcessor",
-    "NormalizerProcessor",
-    "UnnormalizerProcessor",
-    "ObservationProcessor",
+    "GripperPenaltyProcessorStep",
+    "hotswap_stats",
+    "IdentityProcessorStep",
+    "ImageCropResizeProcessorStep",
+    "InfoProcessorStep",
+    "InterventionActionProcessorStep",
+    "JointVelocityProcessorStep",
+    "MapDeltaActionToRobotActionStep",
+    "MapTensorToDeltaActionDictStep",
+    "merge_transitions",
+    "MotorCurrentProcessorStep",
+    "NormalizerProcessorStep",
+    "Numpy2TorchActionProcessorStep",
+    "ObservationProcessorStep",
+    "PolicyProcessorPipeline",
+    "ProcessorKwargs",
     "ProcessorStep",
     "ProcessorStepRegistry",
-    "RenameProcessor",
-    "RewardProcessor",
-    "RobotProcessor",
+    "RenameObservationsProcessorStep",
+    "RewardClassifierProcessorStep",
+    "RewardProcessorStep",
+    "DataProcessorPipeline",
+    "TimeLimitProcessorStep",
+    "AddBatchDimensionProcessorStep",
+    "RobotProcessorPipeline",
+    "TokenizerProcessorStep",
+    "Torch2NumpyActionProcessorStep",
+    "transition_to_batch",
+    "transition_to_dataset_frame",
     "TransitionKey",
-    "TruncatedProcessor",
-    "VanillaObservationProcessor",
+    "TruncatedProcessorStep",
+    "UnnormalizerProcessorStep",
+    "VanillaObservationProcessorStep",
 ]

src/lerobot/processor/batch_processor.py

@@ -0,0 +1,250 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This script defines processor steps for adding a batch dimension to various components of an environment transition.
+
+These steps are designed to process actions, observations, and complementary data, making them suitable for batch processing by adding a leading dimension. This is a common requirement before feeding data into a neural network model.
+"""
+
+from dataclasses import dataclass, field
+
+from torch import Tensor
+
+from lerobot.configs.types import PipelineFeatureType, PolicyFeature
+from lerobot.constants import OBS_ENV_STATE, OBS_IMAGE, OBS_IMAGES, OBS_STATE
+
+from .core import EnvTransition
+from .pipeline import (
+    ActionProcessorStep,
+    ComplementaryDataProcessorStep,
+    ObservationProcessorStep,
+    ProcessorStep,
+    ProcessorStepRegistry,
+)
+
+
+@dataclass
+@ProcessorStepRegistry.register(name="to_batch_processor_action")
+class AddBatchDimensionActionStep(ActionProcessorStep):
+    """
+    Processor step to add a batch dimension to a 1D tensor action.
+
+    This is useful for creating a batch of size 1 from a single action sample.
+    """
+
+    def action(self, action: Tensor) -> Tensor:
+        """
+        Adds a batch dimension to the action if it's a 1D tensor.
+
+        Args:
+            action: The action tensor.
+
+        Returns:
+            The action tensor with an added batch dimension.
+        """
+        if not isinstance(action, Tensor) or action.dim() != 1:
+            return action
+        return action.unsqueeze(0)
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        """
+        Returns the input features unchanged.
+
+        Adding a batch dimension does not alter the feature definition.
+
+        Args:
+            features: A dictionary of policy features.
+
+        Returns:
+            The original dictionary of policy features.
+        """
+        return features
+
+
+@dataclass
+@ProcessorStepRegistry.register(name="to_batch_processor_observation")
+class AddBatchDimensionObservationStep(ObservationProcessorStep):
+    """
+    Processor step to add a batch dimension to observations.
+
+    It handles different types of observations:
+    - State vectors (1D tensors).
+    - Single images (3D tensors).
+    - Dictionaries of multiple images (3D tensors).
+    """
+
+    def observation(self, observation: dict[str, Tensor]) -> dict[str, Tensor]:
+        """
+        Adds a batch dimension to tensor-based observations in the observation dictionary.
+
+        Args:
+            observation: The observation dictionary.
+
+        Returns:
+            The observation dictionary with batch dimensions added to tensors.
+        """
+        # Process state observations - add batch dim if 1D
+        for state_key in [OBS_STATE, OBS_ENV_STATE]:
+            if state_key in observation:
+                state_value = observation[state_key]
+                if isinstance(state_value, Tensor) and state_value.dim() == 1:
+                    observation[state_key] = state_value.unsqueeze(0)
+
+        # Process single image observation - add batch dim if 3D
+        if OBS_IMAGE in observation:
+            image_value = observation[OBS_IMAGE]
+            if isinstance(image_value, Tensor) and image_value.dim() == 3:
+                observation[OBS_IMAGE] = image_value.unsqueeze(0)
+
+        # Process multiple image observations - add batch dim if 3D
+        for key, value in observation.items():
+            if key.startswith(f"{OBS_IMAGES}.") and isinstance(value, Tensor) and value.dim() == 3:
+                observation[key] = value.unsqueeze(0)
+        return observation
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        """
+        Returns the input features unchanged.
+
+        Adding a batch dimension does not alter the feature definition.
+
+        Args:
+            features: A dictionary of policy features.
+
+        Returns:
+            The original dictionary of policy features.
+        """
+        return features
+
+
+@dataclass
+@ProcessorStepRegistry.register(name="to_batch_processor_complementary_data")
+class AddBatchDimensionComplementaryDataStep(ComplementaryDataProcessorStep):
+    """
+    Processor step to add a batch dimension to complementary data fields.
+
+    Handles specific keys like 'task', 'index', and 'task_index' to make them batched.
+    - 'task' (str) is wrapped in a list.
+    - 'index' and 'task_index' (0D tensors) get a batch dimension.
+    """
+
+    def complementary_data(self, complementary_data: dict) -> dict:
+        """
+        Adds a batch dimension to specific fields in the complementary data dictionary.
+
+        Args:
+            complementary_data: The complementary data dictionary.
+
+        Returns:
+            The complementary data dictionary with batch dimensions added.
+        """
+        # Process task field - wrap string in list to add batch dimension
+        if "task" in complementary_data:
+            task_value = complementary_data["task"]
+            if isinstance(task_value, str):
+                complementary_data["task"] = [task_value]
+
+        # Process index field - add batch dim if 0D
+        if "index" in complementary_data:
+            index_value = complementary_data["index"]
+            if isinstance(index_value, Tensor) and index_value.dim() == 0:
+                complementary_data["index"] = index_value.unsqueeze(0)
+
+        # Process task_index field - add batch dim if 0D
+        if "task_index" in complementary_data:
+            task_index_value = complementary_data["task_index"]
+            if isinstance(task_index_value, Tensor) and task_index_value.dim() == 0:
+                complementary_data["task_index"] = task_index_value.unsqueeze(0)
+        return complementary_data
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        """
+        Returns the input features unchanged.
+
+        Adding a batch dimension does not alter the feature definition.
+
+        Args:
+            features: A dictionary of policy features.
+
+        Returns:
+            The original dictionary of policy features.
+        """
+        return features
+
+
+@dataclass
+@ProcessorStepRegistry.register(name="to_batch_processor")
+class AddBatchDimensionProcessorStep(ProcessorStep):
+    """
+    A composite processor step that adds a batch dimension to the entire environment transition.
+
+    This step combines individual processors for actions, observations, and complementary data
+    to create a batched transition (batch size 1) from a single-instance transition.
+
+    Attributes:
+        to_batch_action_processor: Processor for the action component.
+        to_batch_observation_processor: Processor for the observation component.
+        to_batch_complementary_data_processor: Processor for the complementary data component.
+    """
+
+    to_batch_action_processor: AddBatchDimensionActionStep = field(
+        default_factory=AddBatchDimensionActionStep
+    )
+    to_batch_observation_processor: AddBatchDimensionObservationStep = field(
+        default_factory=AddBatchDimensionObservationStep
+    )
+    to_batch_complementary_data_processor: AddBatchDimensionComplementaryDataStep = field(
+        default_factory=AddBatchDimensionComplementaryDataStep
+    )
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        """
+        Applies the batching process to all relevant parts of an environment transition.
+
+        Args:
+            transition: The environment transition to process.
+
+        Returns:
+            The environment transition with a batch dimension added.
+        """
+        transition = self.to_batch_action_processor(transition)
+        transition = self.to_batch_observation_processor(transition)
+        transition = self.to_batch_complementary_data_processor(transition)
+        return transition
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        """
+        Returns the input features unchanged.
+
+        Adding a batch dimension does not alter the feature definition.
+
+        Args:
+            features: A dictionary of policy features.
+
+        Returns:
+            The original dictionary of policy features.
+        """
+        # NOTE: We ignore the batch dimension when transforming features
+        return features

src/lerobot/processor/converters.py

@@ -0,0 +1,523 @@
+# !/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+from collections.abc import Sequence
+from copy import deepcopy
+from functools import singledispatch
+from typing import Any
+
+import numpy as np
+import torch
+
+from lerobot.constants import ACTION, DONE, OBS_IMAGES, OBS_STATE, REWARD, TRUNCATED
+
+from .core import EnvTransition, TransitionKey
+
+
+@singledispatch
+def to_tensor(
+    value: Any,
+    *,
+    dtype: torch.dtype | None = torch.float32,
+    device: torch.device | str | None = None,
+) -> torch.Tensor:
+    """
+    Convert various data types to PyTorch tensors with configurable options.
+
+    This is a unified tensor conversion function using single dispatch to handle
+    different input types appropriately.
+
+    Args:
+        value: Input value to convert (tensor, array, scalar, sequence, etc.).
+        dtype: Target tensor dtype. If None, preserves original dtype.
+        device: Target device for the tensor.
+
+    Returns:
+        A PyTorch tensor.
+
+    Raises:
+        TypeError: If the input type is not supported.
+    """
+    raise TypeError(f"Unsupported type for tensor conversion: {type(value)}")
+
+
+@to_tensor.register(torch.Tensor)
+def _(value: torch.Tensor, *, dtype=torch.float32, device=None, **kwargs) -> torch.Tensor:
+    """Handle conversion for existing PyTorch tensors."""
+    if dtype is not None:
+        value = value.to(dtype=dtype)
+    if device is not None:
+        value = value.to(device=device)
+    return value
+
+
+@to_tensor.register(np.ndarray)
+def _(
+    value: np.ndarray,
+    *,
+    dtype=torch.float32,
+    device=None,
+    **kwargs,
+) -> torch.Tensor:
+    """Handle conversion for numpy arrays."""
+    # Check for numpy scalars (0-dimensional arrays) and treat them as scalars.
+    if value.ndim == 0:
+        # Numpy scalars should be converted to 0-dimensional tensors.
+        scalar_value = value.item()
+        return torch.tensor(scalar_value, dtype=dtype, device=device)
+
+    # Create tensor from numpy array.
+    tensor = torch.from_numpy(value)
+
+    # Apply dtype and device conversion if specified.
+    if dtype is not None:
+        tensor = tensor.to(dtype=dtype)
+    if device is not None:
+        tensor = tensor.to(device=device)
+
+    return tensor
+
+
+@to_tensor.register(int)
+@to_tensor.register(float)
+@to_tensor.register(np.integer)
+@to_tensor.register(np.floating)
+def _(value, *, dtype=torch.float32, device=None, **kwargs) -> torch.Tensor:
+    """Handle conversion for scalar values including numpy scalars."""
+    return torch.tensor(value, dtype=dtype, device=device)
+
+
+@to_tensor.register(list)
+@to_tensor.register(tuple)
+def _(value: Sequence, *, dtype=torch.float32, device=None, **kwargs) -> torch.Tensor:
+    """Handle conversion for sequences (lists, tuples)."""
+    return torch.tensor(value, dtype=dtype, device=device)
+
+
+@to_tensor.register(dict)
+def _(value: dict, *, device=None, **kwargs) -> dict:
+    """Handle conversion for dictionaries by recursively converting their values to tensors."""
+    if not value:
+        return {}
+
+    result = {}
+    for key, sub_value in value.items():
+        if sub_value is None:
+            continue
+
+        if isinstance(sub_value, dict):
+            # Recursively process nested dictionaries.
+            result[key] = to_tensor(
+                sub_value,
+                device=device,
+                **kwargs,
+            )
+            continue
+
+        # Convert individual values to tensors.
+        result[key] = to_tensor(
+            sub_value,
+            device=device,
+            **kwargs,
+        )
+    return result
+
+
+def from_tensor_to_numpy(x: torch.Tensor | Any) -> np.ndarray | float | int | Any:
+    """
+    Convert a PyTorch tensor to a numpy array or scalar if applicable.
+
+    If the input is not a tensor, it is returned unchanged.
+
+    Args:
+        x: The input, which can be a tensor or any other type.
+
+    Returns:
+        A numpy array, a scalar, or the original input.
+    """
+    if isinstance(x, torch.Tensor):
+        return x.item() if x.numel() == 1 else x.detach().cpu().numpy()
+    return x
+
+
+def _is_image(arr: Any) -> bool:
+    """
+    Check if a given array is likely an image (uint8, 3D).
+
+    Args:
+        arr: The array to check.
+
+    Returns:
+        True if the array matches the image criteria, False otherwise.
+    """
+    return isinstance(arr, np.ndarray) and arr.dtype == np.uint8 and arr.ndim == 3
+
+
+def _split_obs_to_state_and_images(obs: dict[str, Any]) -> tuple[dict[str, Any], dict[str, Any]]:
+    """
+    Separate an observation dictionary into state and image components.
+
+    Args:
+        obs: The observation dictionary.
+
+    Returns:
+        A tuple containing two dictionaries: one for state and one for images.
+    """
+    state, images = {}, {}
+    for k, v in obs.items():
+        if "image" in k.lower() or _is_image(v):
+            images[k] = v
+        else:
+            state[k] = v
+    return state, images
+
+
+# Private Helper Functions (Common Logic)
+
+
+def _extract_complementary_data(batch: dict[str, Any]) -> dict[str, Any]:
+    """
+    Extract complementary data from a batch dictionary.
+
+    This includes padding flags, task description, and indices.
+
+    Args:
+        batch: The batch dictionary.
+
+    Returns:
+        A dictionary with the extracted complementary data.
+    """
+    pad_keys = {k: v for k, v in batch.items() if "_is_pad" in k}
+    task_key = {"task": batch["task"]} if "task" in batch else {}
+    index_key = {"index": batch["index"]} if "index" in batch else {}
+    task_index_key = {"task_index": batch["task_index"]} if "task_index" in batch else {}
+
+    return {**pad_keys, **task_key, **index_key, **task_index_key}
+
+
+def _merge_transitions(base: EnvTransition, other: EnvTransition) -> EnvTransition:
+    """
+    Merge two transitions, with the second one taking precedence in case of conflicts.
+
+    Args:
+        base: The base transition.
+        other: The transition to merge, which will overwrite base values.
+
+    Returns:
+        The merged transition dictionary.
+    """
+    out = deepcopy(base)
+
+    for key in (
+        TransitionKey.OBSERVATION,
+        TransitionKey.ACTION,
+        TransitionKey.INFO,
+        TransitionKey.COMPLEMENTARY_DATA,
+    ):
+        if other.get(key):
+            out.setdefault(key, {}).update(deepcopy(other[key]))
+
+    for k in (TransitionKey.REWARD, TransitionKey.DONE, TransitionKey.TRUNCATED):
+        if k in other:
+            out[k] = other[k]
+    return out
+
+
+# Core Conversion Functions
+
+
+def create_transition(
+    observation: dict[str, Any] | None = None,
+    action: dict[str, Any] | None = None,
+    reward: float = 0.0,
+    done: bool = False,
+    truncated: bool = False,
+    info: dict[str, Any] | None = None,
+    complementary_data: dict[str, Any] | None = None,
+) -> EnvTransition:
+    """
+    Create an `EnvTransition` dictionary with sensible defaults.
+
+    Args:
+        observation: Observation dictionary.
+        action: Action dictionary.
+        reward: Scalar reward value.
+        done: Episode termination flag.
+        truncated: Episode truncation flag.
+        info: Additional info dictionary.
+        complementary_data: Complementary data dictionary.
+
+    Returns:
+        A complete `EnvTransition` dictionary.
+    """
+    return {
+        TransitionKey.OBSERVATION: observation,
+        TransitionKey.ACTION: action,
+        TransitionKey.REWARD: reward,
+        TransitionKey.DONE: done,
+        TransitionKey.TRUNCATED: truncated,
+        TransitionKey.INFO: info if info is not None else {},
+        TransitionKey.COMPLEMENTARY_DATA: complementary_data if complementary_data is not None else {},
+    }
+
+
+def action_to_transition(action: dict[str, Any]) -> EnvTransition:
+    """
+    Convert a raw action dictionary into a standardized `EnvTransition`.
+
+    The keys in the action dictionary are prefixed with "action." and stored under
+    the `ACTION` key in the transition. Values are converted to tensors, except for
+    special types like `Rotation`.
+
+    Args:
+        action: The raw action dictionary from a teleoperation device or controller.
+
+    Returns:
+        An `EnvTransition` containing the formatted action.
+    """
+
+    return create_transition(observation={}, action=action)
+
+
+def observation_to_transition(observation: dict[str, Any]) -> EnvTransition:
+    """
+    Convert a raw robot observation dictionary into a standardized `EnvTransition`.
+
+    The observation is split into state and image components. State keys are prefixed
+    with "observation.state." and image keys with "observation.images.". The result is
+    stored under the `OBSERVATION` key in the transition.
+
+    Args:
+        observation: The raw observation dictionary from the environment.
+
+    Returns:
+        An `EnvTransition` containing the formatted observation.
+    """
+    state, images = _split_obs_to_state_and_images(observation)
+
+    image_observations = {f"{OBS_IMAGES}.{cam}": img for cam, img in images.items()}
+
+    return create_transition(observation={**state, **image_observations}, action={})
+
+
+def transition_to_action(transition: EnvTransition) -> dict[str, Any]:
+    """
+    Extract a raw action dictionary for a robot from an `EnvTransition`.
+
+    This function searches for keys in the format "action.*.pos" or "action.*.vel"
+    and converts them into a flat dictionary suitable for sending to a robot controller.
+
+    Args:
+        transition: The `EnvTransition` containing the action.
+
+    Returns:
+        A dictionary representing the raw robot action.
+    """
+    return transition.get(TransitionKey.ACTION)
+
+
+def merge_transitions(transitions: Sequence[EnvTransition] | EnvTransition) -> EnvTransition:
+    """
+    Merge a sequence of transitions into a single one.
+
+    If a single transition is provided, it is returned as is. For a sequence,
+    transitions are merged sequentially, with later transitions in the sequence
+    overwriting earlier ones.
+
+    Args:
+        transitions: A single transition or a sequence of them.
+
+    Returns:
+        A single merged `EnvTransition`.
+
+    Raises:
+        ValueError: If an empty sequence of transitions is provided.
+    """
+
+    if not isinstance(transitions, Sequence):  # Single transition
+        return transitions
+
+    items = list(transitions)
+    if not items:
+        raise ValueError("merge_transitions() requires a non-empty sequence of transitions")
+
+    result = items[0]
+    for t in items[1:]:
+        result = _merge_transitions(result, t)
+    return result
+
+
+def transition_to_dataset_frame(
+    transitions_or_transition: EnvTransition | Sequence[EnvTransition], features: dict[str, dict]
+) -> dict[str, Any]:
+    """
+    Convert one or more transitions into a flat dictionary suitable for a dataset frame.
+
+    This function processes `EnvTransition` objects according to a feature
+    specification, producing a format ready for training or evaluation.
+
+    Args:
+        transitions_or_transition: A single `EnvTransition` or a sequence to be merged.
+        features: A feature specification dictionary.
+
+    Returns:
+        A flat dictionary representing a single frame of data for a dataset.
+    """
+    action_names = features.get(ACTION, {}).get("names", [])
+    obs_state_names = features.get(OBS_STATE, {}).get("names", [])
+    image_keys = [k for k in features if k.startswith(OBS_IMAGES)]
+
+    tr = merge_transitions(transitions_or_transition)
+    obs = tr.get(TransitionKey.OBSERVATION, {}) or {}
+    act = tr.get(TransitionKey.ACTION, {}) or {}
+    batch: dict[str, Any] = {}
+
+    # Passthrough for images.
+    for k in image_keys:
+        if k in obs:
+            batch[k] = obs[k]
+
+    # Create observation.state vector.
+    if obs_state_names:
+        vals = [from_tensor_to_numpy(obs.get(f"{OBS_STATE}.{n}", 0.0)) for n in obs_state_names]
+        batch[OBS_STATE] = np.asarray(vals, dtype=np.float32)
+
+    # Create action vector.
+    if action_names:
+        vals = [from_tensor_to_numpy(act.get(f"{ACTION}.{n}", 0.0)) for n in action_names]
+        batch[ACTION] = np.asarray(vals, dtype=np.float32)
+
+    # Add transition metadata.
+    if tr.get(TransitionKey.REWARD) is not None:
+        reward_val = from_tensor_to_numpy(tr[TransitionKey.REWARD])
+        # Check if features expect array format, otherwise keep as scalar.
+        if REWARD in features and features[REWARD].get("shape") == (1,):
+            batch[REWARD] = np.array([reward_val], dtype=np.float32)
+        else:
+            batch[REWARD] = reward_val
+
+    if tr.get(TransitionKey.DONE) is not None:
+        done_val = from_tensor_to_numpy(tr[TransitionKey.DONE])
+        if DONE in features and features[DONE].get("shape") == (1,):
+            batch[DONE] = np.array([done_val], dtype=bool)
+        else:
+            batch[DONE] = done_val
+
+    if tr.get(TransitionKey.TRUNCATED) is not None:
+        truncated_val = from_tensor_to_numpy(tr[TransitionKey.TRUNCATED])
+        if TRUNCATED in features and features[TRUNCATED].get("shape") == (1,):
+            batch[TRUNCATED] = np.array([truncated_val], dtype=bool)
+        else:
+            batch[TRUNCATED] = truncated_val
+
+    # Add complementary data flags and task.
+    comp = tr.get(TransitionKey.COMPLEMENTARY_DATA) or {}
+    if comp:
+        # Padding flags.
+        for k, v in comp.items():
+            if k.endswith("_is_pad"):
+                batch[k] = v
+        # Task label.
+        if comp.get("task") is not None:
+            batch["task"] = comp["task"]
+
+    return batch
+
+
+def batch_to_transition(batch: dict[str, Any]) -> EnvTransition:
+    """
+    Convert a batch dictionary from a dataset/dataloader into an `EnvTransition`.
+
+    This function maps recognized keys from a batch to the `EnvTransition` structure,
+    filling in missing keys with sensible defaults.
+
+    Args:
+        batch: A batch dictionary.
+
+    Returns:
+        An `EnvTransition` dictionary.
+
+    Raises:
+        ValueError: If the input is not a dictionary.
+    """
+
+    # Validate input type.
+    if not isinstance(batch, dict):
+        raise ValueError(f"EnvTransition must be a dictionary. Got {type(batch).__name__}")
+
+    # Extract observation and complementary data keys.
+    observation_keys = {k: v for k, v in batch.items() if k.startswith("observation.")}
+    complementary_data = _extract_complementary_data(batch)
+
+    return create_transition(
+        observation=observation_keys if observation_keys else None,
+        action=batch.get("action"),
+        reward=batch.get("next.reward", 0.0),
+        done=batch.get("next.done", False),
+        truncated=batch.get("next.truncated", False),
+        info=batch.get("info", {}),
+        complementary_data=complementary_data if complementary_data else None,
+    )
+
+
+def transition_to_batch(transition: EnvTransition) -> dict[str, Any]:
+    """
+    Convert an `EnvTransition` back to the canonical batch format used in LeRobot.
+
+    This is the inverse of `batch_to_transition`.
+
+    Args:
+        transition: The `EnvTransition` to convert.
+
+    Returns:
+        A batch dictionary with canonical LeRobot field names.
+    """
+    batch = {
+        "action": transition.get(TransitionKey.ACTION),
+        "next.reward": transition.get(TransitionKey.REWARD, 0.0),
+        "next.done": transition.get(TransitionKey.DONE, False),
+        "next.truncated": transition.get(TransitionKey.TRUNCATED, False),
+        "info": transition.get(TransitionKey.INFO, {}),
+    }
+
+    # Add complementary data.
+    comp_data = transition.get(TransitionKey.COMPLEMENTARY_DATA, {})
+    if comp_data:
+        batch.update(comp_data)
+
+    # Flatten observation dictionary.
+    observation = transition.get(TransitionKey.OBSERVATION)
+    if isinstance(observation, dict):
+        batch.update(observation)
+
+    return batch
+
+
+def identity_transition(tr: EnvTransition) -> EnvTransition:
+    """
+    An identity function for transitions, returning the input unchanged.
+
+    Useful as a default or placeholder in processing pipelines.
+
+    Args:
+        tr: An `EnvTransition`.
+
+    Returns:
+        The same `EnvTransition`.
+    """
+    return tr

src/lerobot/processor/core.py

@@ -0,0 +1,49 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+from enum import Enum
+from typing import Any, TypedDict
+
+import torch
+
+
+class TransitionKey(str, Enum):
+    """Keys for accessing EnvTransition dictionary components."""
+
+    # TODO(Steven): Use consts
+    OBSERVATION = "observation"
+    ACTION = "action"
+    REWARD = "reward"
+    DONE = "done"
+    TRUNCATED = "truncated"
+    INFO = "info"
+    COMPLEMENTARY_DATA = "complementary_data"
+
+
+EnvTransition = TypedDict(
+    "EnvTransition",
+    {
+        TransitionKey.OBSERVATION.value: dict[str, Any] | None,
+        TransitionKey.ACTION.value: Any | torch.Tensor | None,
+        TransitionKey.REWARD.value: float | torch.Tensor | None,
+        TransitionKey.DONE.value: bool | torch.Tensor | None,
+        TransitionKey.TRUNCATED.value: bool | torch.Tensor | None,
+        TransitionKey.INFO.value: dict[str, Any] | None,
+        TransitionKey.COMPLEMENTARY_DATA.value: dict[str, Any] | None,
+    },
+)

src/lerobot/processor/delta_action_processor.py

@@ -0,0 +1,147 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+
+from torch import Tensor
+
+from lerobot.configs.types import FeatureType, PipelineFeatureType, PolicyFeature
+
+from .pipeline import ActionProcessorStep, ProcessorStepRegistry
+
+
+@ProcessorStepRegistry.register("map_tensor_to_delta_action_dict")
+@dataclass
+class MapTensorToDeltaActionDictStep(ActionProcessorStep):
+    """
+    Maps a flat action tensor from a policy to a structured delta action dictionary.
+
+    This step is typically used after a policy outputs a continuous action vector.
+    It decomposes the vector into named components for delta movements of the
+    end-effector (x, y, z) and optionally the gripper.
+
+    Attributes:
+        use_gripper: If True, assumes the 4th element of the tensor is the
+                     gripper action.
+    """
+
+    use_gripper: bool = True
+
+    def action(self, action: Tensor) -> dict:
+        if action.dim() > 1:
+            action = action.squeeze(0)
+
+        # TODO (maractingi): add rotation
+        delta_action = {
+            "delta_x": action[0],
+            "delta_y": action[1],
+            "delta_z": action[2],
+        }
+        if self.use_gripper:
+            delta_action["gripper"] = action[3]
+        return delta_action
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        features[PipelineFeatureType.ACTION]["delta_x"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        features[PipelineFeatureType.ACTION]["delta_y"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        features[PipelineFeatureType.ACTION]["delta_z"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        if self.use_gripper:
+            features[PipelineFeatureType.ACTION]["gripper"] = PolicyFeature(
+                type=FeatureType.ACTION, shape=(1,)
+            )
+        return features
+
+
+@ProcessorStepRegistry.register("map_delta_action_to_robot_action")
+@dataclass
+class MapDeltaActionToRobotActionStep(ActionProcessorStep):
+    """
+    Maps delta actions from teleoperators to robot target actions for inverse kinematics.
+
+    This step converts a dictionary of delta movements (e.g., from a gamepad)
+    into a target action format that includes an "enabled" flag and target
+    end-effector positions. It also handles scaling and noise filtering.
+
+    Attributes:
+        position_scale: A factor to scale the delta position inputs.
+        rotation_scale: A factor to scale the delta rotation inputs (currently unused).
+        noise_threshold: The magnitude below which delta inputs are considered noise
+                         and do not trigger an "enabled" state.
+    """
+
+    # Scale factors for delta movements
+    position_scale: float = 1.0
+    rotation_scale: float = 0.0  # No rotation deltas for gamepad/keyboard
+    noise_threshold: float = 1e-3  # 1 mm threshold to filter out noise
+
+    def action(self, action: dict) -> dict:
+        # NOTE (maractingi): Action can be a dict from the teleop_devices or a tensor from the policy
+        # TODO (maractingi): changing this target_xyz naming convention from the teleop_devices
+        delta_x = action.pop("delta_x", 0.0)
+        delta_y = action.pop("delta_y", 0.0)
+        delta_z = action.pop("delta_z", 0.0)
+        gripper = action.pop("gripper", 1.0)  # Default to "stay" (1.0)
+
+        # Determine if the teleoperator is actively providing input
+        # Consider enabled if any significant movement delta is detected
+        position_magnitude = (delta_x**2 + delta_y**2 + delta_z**2) ** 0.5  # Use Euclidean norm for position
+        enabled = position_magnitude > self.noise_threshold  # Small threshold to avoid noise
+
+        # Scale the deltas appropriately
+        scaled_delta_x = delta_x * self.position_scale
+        scaled_delta_y = delta_y * self.position_scale
+        scaled_delta_z = delta_z * self.position_scale
+
+        # For gamepad/keyboard, we don't have rotation input, so set to 0
+        # These could be extended in the future for more sophisticated teleoperators
+        target_wx = 0.0
+        target_wy = 0.0
+        target_wz = 0.0
+
+        # Update action with robot target format
+        action = {
+            "enabled": enabled,
+            "target_x": scaled_delta_x,
+            "target_y": scaled_delta_y,
+            "target_z": scaled_delta_z,
+            "target_wx": target_wx,
+            "target_wy": target_wy,
+            "target_wz": target_wz,
+            "gripper": float(gripper),
+        }
+
+        return action
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        """Transform features to match output format."""
+        features[PipelineFeatureType.ACTION].pop("delta_x", None)
+        features[PipelineFeatureType.ACTION].pop("delta_y", None)
+        features[PipelineFeatureType.ACTION].pop("delta_z", None)
+        features[PipelineFeatureType.ACTION].pop("gripper", None)
+
+        features[PipelineFeatureType.ACTION]["enabled"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        features[PipelineFeatureType.ACTION]["target_x"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        features[PipelineFeatureType.ACTION]["target_y"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        features[PipelineFeatureType.ACTION]["target_z"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        features[PipelineFeatureType.ACTION]["target_wx"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        features[PipelineFeatureType.ACTION]["target_wy"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        features[PipelineFeatureType.ACTION]["target_wz"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        features[PipelineFeatureType.ACTION]["gripper"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        return features

src/lerobot/processor/device_processor.py

@@ -13,70 +13,174 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+
+"""
+This script defines a processor step for moving environment transition data to a specific torch device and casting
+its floating-point precision.
+"""
+
 from dataclasses import dataclass
 from typing import Any
 
 import torch
 
-from lerobot.configs.types import PolicyFeature
-from lerobot.processor.pipeline import EnvTransition, TransitionKey
+from lerobot.configs.types import PipelineFeatureType, PolicyFeature
 from lerobot.utils.utils import get_safe_torch_device
 
+from .core import EnvTransition, TransitionKey
+from .pipeline import ProcessorStep, ProcessorStepRegistry
 
+
+@ProcessorStepRegistry.register("device_processor")
 @dataclass
-class DeviceProcessor:
-    """Processes transitions by moving tensors to the specified device.
+class DeviceProcessorStep(ProcessorStep):
+    """
+    Processor step to move all tensors within an `EnvTransition` to a specified device and optionally cast their
+    floating-point data type.
 
-    This processor ensures that all tensors in the transition are moved to the
-    specified device (CPU or GPU) before they are returned.
+    This is crucial for preparing data for model training or inference on hardware like GPUs.
+
+    Attributes:
+        device: The target device for tensors (e.g., "cpu", "cuda", "cuda:0").
+        float_dtype: The target floating-point dtype as a string (e.g., "float32", "float16", "bfloat16").
+                     If None, the dtype is not changed.
     """
 
-    device: torch.device = "cpu"
+    device: str = "cpu"
+    float_dtype: str | None = None
+
+    DTYPE_MAPPING = {
+        "float16": torch.float16,
+        "float32": torch.float32,
+        "float64": torch.float64,
+        "bfloat16": torch.bfloat16,
+        "half": torch.float16,
+        "float": torch.float32,
+        "double": torch.float64,
+    }
 
     def __post_init__(self):
-        self.device = get_safe_torch_device(self.device)
+        """
+        Initializes the processor by converting string configurations to torch objects.
+
+        This method sets up the `torch.device`, determines if transfers can be non-blocking, and validates the
+        `float_dtype` string, converting it to a `torch.dtype` object.
+        """
+        self.tensor_device: torch.device = get_safe_torch_device(self.device)
+        # Update device string in case a specific GPU was selected (e.g., "cuda" -> "cuda:0")
+        self.device = self.tensor_device.type
         self.non_blocking = "cuda" in str(self.device)
 
+        # Validate and convert float_dtype string to torch dtype
+        if self.float_dtype is not None:
+            if self.float_dtype not in self.DTYPE_MAPPING:
+                raise ValueError(
+                    f"Invalid float_dtype '{self.float_dtype}'. Available options: {list(self.DTYPE_MAPPING.keys())}"
+                )
+            self._target_float_dtype = self.DTYPE_MAPPING[self.float_dtype]
+        else:
+            self._target_float_dtype = None
+
+    def _process_tensor(self, tensor: torch.Tensor) -> torch.Tensor:
+        """
+        Moves a single tensor to the target device and casts its dtype.
+
+        Handles multi-GPU scenarios by not moving a tensor if it's already on a different CUDA device than
+        the target, which is useful when using frameworks like Accelerate.
+
+        Args:
+            tensor: The input torch.Tensor.
+
+        Returns:
+            The processed tensor on the correct device and with the correct dtype.
+        """
+        # Determine target device
+        if tensor.is_cuda and self.tensor_device.type == "cuda":
+            # Both tensor and target are on GPU - preserve tensor's GPU placement.
+            # This handles multi-GPU scenarios where Accelerate has already placed
+            # tensors on the correct GPU for each process.
+            target_device = tensor.device
+        else:
+            # Either tensor is on CPU, or we're configured for CPU.
+            # In both cases, use the configured device.
+            target_device = self.tensor_device
+
+        # Only move if necessary
+        if tensor.device != target_device:
+            tensor = tensor.to(target_device, non_blocking=self.non_blocking)
+
+        # Convert float dtype if specified and tensor is floating point
+        if self._target_float_dtype is not None and tensor.is_floating_point():
+            tensor = tensor.to(dtype=self._target_float_dtype)
+
+        return tensor
+
     def __call__(self, transition: EnvTransition) -> EnvTransition:
-        # Create a copy of the transition
+        """
+        Applies device and dtype conversion to all tensors in an environment transition.
+
+        It iterates through the transition, finds all `torch.Tensor` objects (including those nested in
+        dictionaries like `observation`), and processes them.
+
+        Args:
+            transition: The input `EnvTransition` object.
+
+        Returns:
+            A new `EnvTransition` object with all tensors moved to the target device and dtype.
+        """
         new_transition = transition.copy()
 
-        # Process observation tensors
-        observation = transition.get(TransitionKey.OBSERVATION)
-        if observation is not None:
-            new_observation = {
-                k: v.to(self.device, non_blocking=self.non_blocking) if isinstance(v, torch.Tensor) else v
-                for k, v in observation.items()
-            }
-            new_transition[TransitionKey.OBSERVATION] = new_observation
+        simple_tensor_keys = [
+            TransitionKey.ACTION,
+            TransitionKey.REWARD,
+            TransitionKey.DONE,
+            TransitionKey.TRUNCATED,
+        ]
 
-        # Process action tensor
-        action = transition.get(TransitionKey.ACTION)
-        if action is not None and isinstance(action, torch.Tensor):
-            new_transition[TransitionKey.ACTION] = action.to(self.device, non_blocking=self.non_blocking)
+        dict_tensor_keys = [
+            TransitionKey.OBSERVATION,
+            TransitionKey.COMPLEMENTARY_DATA,
+        ]
 
-        # Process reward tensor
-        reward = transition.get(TransitionKey.REWARD)
-        if reward is not None and isinstance(reward, torch.Tensor):
-            new_transition[TransitionKey.REWARD] = reward.to(self.device, non_blocking=self.non_blocking)
+        # Process simple, top-level tensors
+        for key in simple_tensor_keys:
+            value = transition.get(key)
+            if isinstance(value, torch.Tensor):
+                new_transition[key] = self._process_tensor(value)
 
-        # Process done tensor
-        done = transition.get(TransitionKey.DONE)
-        if done is not None and isinstance(done, torch.Tensor):
-            new_transition[TransitionKey.DONE] = done.to(self.device, non_blocking=self.non_blocking)
-
-        # Process truncated tensor
-        truncated = transition.get(TransitionKey.TRUNCATED)
-        if truncated is not None and isinstance(truncated, torch.Tensor):
-            new_transition[TransitionKey.TRUNCATED] = truncated.to(
-                self.device, non_blocking=self.non_blocking
-            )
+        # Process tensors nested within dictionaries
+        for key in dict_tensor_keys:
+            data_dict = transition.get(key)
+            if data_dict is not None:
+                new_data_dict = {
+                    k: self._process_tensor(v) if isinstance(v, torch.Tensor) else v
+                    for k, v in data_dict.items()
+                }
+                new_transition[key] = new_data_dict
 
         return new_transition
 
     def get_config(self) -> dict[str, Any]:
-        """Return configuration for serialization."""
-        return {"device": self.device}
+        """
+        Returns the serializable configuration of the processor.
 
-    def feature_contract(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
+        Returns:
+            A dictionary containing the device and float_dtype settings.
+        """
+        return {"device": self.device, "float_dtype": self.float_dtype}
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        """
+        Returns the input features unchanged.
+
+        Device and dtype transformations do not alter the fundamental definition of the features (e.g., shape).
+
+        Args:
+            features: A dictionary of policy features.
+
+        Returns:
+            The original dictionary of policy features.
+        """
         return features

src/lerobot/processor/gym_action_processor.py

@@ -0,0 +1,95 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+
+import numpy as np
+import torch
+
+from lerobot.configs.types import PipelineFeatureType, PolicyFeature
+
+from .converters import to_tensor
+from .pipeline import ActionProcessorStep, ProcessorStepRegistry
+
+
+@ProcessorStepRegistry.register("torch2numpy_action_processor")
+@dataclass
+class Torch2NumpyActionProcessorStep(ActionProcessorStep):
+    """
+    Converts a PyTorch tensor action to a NumPy array.
+
+    This step is useful when the output of a policy (typically a torch.Tensor)
+    needs to be passed to an environment or component that expects a NumPy array.
+
+    Attributes:
+        squeeze_batch_dim: If True, removes the first dimension of the array
+                           if it is of size 1. This is useful for converting a
+                           batched action of size (1, D) to a single action of size (D,).
+    """
+
+    squeeze_batch_dim: bool = True
+
+    def action(self, action: torch.Tensor) -> np.ndarray:
+        if not isinstance(action, torch.Tensor):
+            raise TypeError(
+                f"Expected torch.Tensor or None, got {type(action).__name__}. "
+                "Use appropriate processor for non-tensor actions."
+            )
+
+        numpy_action = action.detach().cpu().numpy()
+
+        # Remove batch dimensions but preserve action dimensions.
+        # Only squeeze if there's a batch dimension (first dim == 1).
+        if (
+            self.squeeze_batch_dim
+            and numpy_action.shape
+            and len(numpy_action.shape) > 1
+            and numpy_action.shape[0] == 1
+        ):
+            numpy_action = numpy_action.squeeze(0)
+
+        return numpy_action
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        return features
+
+
+@ProcessorStepRegistry.register("numpy2torch_action_processor")
+@dataclass
+class Numpy2TorchActionProcessorStep(ActionProcessorStep):
+    """
+    Converts a NumPy array action to a PyTorch tensor.
+
+    This step is useful for converting actions from environments or hardware,
+    which are often NumPy arrays, into PyTorch tensors that can be processed
+    by a policy or model.
+    """
+
+    def action(self, action: np.ndarray) -> torch.Tensor:
+        if not isinstance(action, np.ndarray):
+            raise TypeError(
+                f"Expected np.ndarray or None, got {type(action).__name__}. "
+                "Use appropriate processor for non-tensor actions."
+            )
+        torch_action = to_tensor(action, dtype=None)  # Preserve original dtype
+        return torch_action
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        return features

src/lerobot/processor/hil_processor.py

@@ -0,0 +1,592 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+import time
+from dataclasses import dataclass
+from typing import Any, Protocol, TypeVar, runtime_checkable
+
+import numpy as np
+import torch
+import torchvision.transforms.functional as F  # noqa: N812
+
+from lerobot.configs.types import PipelineFeatureType, PolicyFeature
+from lerobot.constants import ACTION
+from lerobot.teleoperators.teleoperator import Teleoperator
+from lerobot.teleoperators.utils import TeleopEvents
+
+from .core import EnvTransition, TransitionKey
+from .pipeline import (
+    ComplementaryDataProcessorStep,
+    InfoProcessorStep,
+    ObservationProcessorStep,
+    ProcessorStep,
+    ProcessorStepRegistry,
+    TruncatedProcessorStep,
+)
+
+GRIPPER_KEY = "gripper"
+DISCRETE_PENALTY_KEY = "discrete_penalty"
+TELEOP_ACTION_KEY = "teleop_action"
+
+
+@runtime_checkable
+class HasTeleopEvents(Protocol):
+    """
+    Minimal protocol for objects that provide teleoperation events.
+
+    This protocol defines the `get_teleop_events()` method, allowing processor
+    steps to interact with teleoperators that support event-based controls
+    (like episode termination or success flagging) without needing to know the
+    teleoperator's specific class.
+    """
+
+    def get_teleop_events(self) -> dict[str, Any]:
+        """
+        Get extra control events from the teleoperator.
+
+        Returns:
+            A dictionary containing control events such as:
+            - `is_intervention`: bool - Whether the human is currently intervening.
+            - `terminate_episode`: bool - Whether to terminate the current episode.
+            - `success`: bool - Whether the episode was successful.
+            - `rerecord_episode`: bool - Whether to rerecord the episode.
+        """
+        ...
+
+
+# Type variable constrained to Teleoperator subclasses that also implement events
+TeleopWithEvents = TypeVar("TeleopWithEvents", bound=Teleoperator)
+
+
+def _check_teleop_with_events(teleop: Teleoperator) -> None:
+    """
+    Runtime check that a teleoperator implements the `HasTeleopEvents` protocol.
+
+    Args:
+        teleop: The teleoperator instance to check.
+
+    Raises:
+        TypeError: If the teleoperator does not have a `get_teleop_events` method.
+    """
+    if not isinstance(teleop, HasTeleopEvents):
+        raise TypeError(
+            f"Teleoperator {type(teleop).__name__} must implement get_teleop_events() method. "
+            f"Compatible teleoperators: GamepadTeleop, KeyboardEndEffectorTeleop"
+        )
+
+
+@ProcessorStepRegistry.register("add_teleop_action_as_complementary_data")
+@dataclass
+class AddTeleopActionAsComplimentaryDataStep(ComplementaryDataProcessorStep):
+    """
+    Adds the raw action from a teleoperator to the transition's complementary data.
+
+    This is useful for human-in-the-loop scenarios where the human's input needs to
+    be available to downstream processors, for example, to override a policy's action
+    during an intervention.
+
+    Attributes:
+        teleop_device: The teleoperator instance to get the action from.
+    """
+
+    teleop_device: Teleoperator
+
+    def complementary_data(self, complementary_data: dict) -> dict:
+        """
+        Retrieves the teleoperator's action and adds it to the complementary data.
+
+        Args:
+            complementary_data: The incoming complementary data dictionary.
+
+        Returns:
+            A new dictionary with the teleoperator action added under the
+            `teleop_action` key.
+        """
+        new_complementary_data = dict(complementary_data)
+        new_complementary_data[TELEOP_ACTION_KEY] = self.teleop_device.get_action()
+        return new_complementary_data
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        return features
+
+
+@ProcessorStepRegistry.register("add_teleop_action_as_info")
+@dataclass
+class AddTeleopEventsAsInfoStep(InfoProcessorStep):
+    """
+    Adds teleoperator control events (e.g., terminate, success) to the transition's info.
+
+    This step extracts control events from teleoperators that support event-based
+    interaction, making these signals available to other parts of the system.
+
+    Attributes:
+        teleop_device: An instance of a teleoperator that implements the
+                       `HasTeleopEvents` protocol.
+    """
+
+    teleop_device: TeleopWithEvents
+
+    def __post_init__(self):
+        """Validates that the provided teleoperator supports events after initialization."""
+        _check_teleop_with_events(self.teleop_device)
+
+    def info(self, info: dict) -> dict:
+        """
+        Retrieves teleoperator events and updates the info dictionary.
+
+        Args:
+            info: The incoming info dictionary.
+
+        Returns:
+            A new dictionary including the teleoperator events.
+        """
+        new_info = dict(info)
+
+        teleop_events = self.teleop_device.get_teleop_events()
+        new_info.update(teleop_events)
+        return new_info
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        return features
+
+
+@ProcessorStepRegistry.register("image_crop_resize_processor")
+@dataclass
+class ImageCropResizeProcessorStep(ObservationProcessorStep):
+    """
+    Crops and/or resizes image observations.
+
+    This step iterates through all image keys in an observation dictionary and applies
+    the specified transformations. It handles device placement, moving tensors to the
+    CPU if necessary for operations not supported on certain accelerators like MPS.
+
+    Attributes:
+        crop_params_dict: A dictionary mapping image keys to cropping parameters
+                          (top, left, height, width).
+        resize_size: A tuple (height, width) to resize all images to.
+    """
+
+    crop_params_dict: dict[str, tuple[int, int, int, int]] | None = None
+    resize_size: tuple[int, int] | None = None
+
+    def observation(self, observation: dict) -> dict:
+        """
+        Applies cropping and resizing to all images in the observation dictionary.
+
+        Args:
+            observation: The observation dictionary, potentially containing image tensors.
+
+        Returns:
+            A new observation dictionary with transformed images.
+        """
+        if self.resize_size is None and not self.crop_params_dict:
+            return observation
+
+        new_observation = dict(observation)
+
+        # Process all image keys in the observation
+        for key in observation:
+            if "image" not in key:
+                continue
+
+            image = observation[key]
+            device = image.device
+            # NOTE (maractingi): No mps kernel for crop and resize, so we need to move to cpu
+            if device.type == "mps":
+                image = image.cpu()
+            # Crop if crop params are provided for this key
+            if self.crop_params_dict is not None and key in self.crop_params_dict:
+                crop_params = self.crop_params_dict[key]
+                image = F.crop(image, *crop_params)
+            if self.resize_size is not None:
+                image = F.resize(image, self.resize_size)
+                image = image.clamp(0.0, 1.0)
+            new_observation[key] = image.to(device)
+
+        return new_observation
+
+    def get_config(self) -> dict[str, Any]:
+        """
+        Returns the configuration of the step for serialization.
+
+        Returns:
+            A dictionary with the crop parameters and resize dimensions.
+        """
+        return {
+            "crop_params_dict": self.crop_params_dict,
+            "resize_size": self.resize_size,
+        }
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        """
+        Updates the image feature shapes in the policy features dictionary if resizing is applied.
+
+        Args:
+            features: The policy features dictionary.
+
+        Returns:
+            The updated policy features dictionary with new image shapes.
+        """
+        if self.resize_size is None:
+            return features
+        for key in features[PipelineFeatureType.OBSERVATION]:
+            if "image" in key:
+                nb_channel = features[PipelineFeatureType.OBSERVATION][key].shape[0]
+                features[PipelineFeatureType.OBSERVATION][key] = PolicyFeature(
+                    type=features[PipelineFeatureType.OBSERVATION][key].type,
+                    shape=(nb_channel, *self.resize_size),
+                )
+        return features
+
+
+@dataclass
+@ProcessorStepRegistry.register("time_limit_processor")
+class TimeLimitProcessorStep(TruncatedProcessorStep):
+    """
+    Tracks episode steps and enforces a time limit by truncating the episode.
+
+    Attributes:
+        max_episode_steps: The maximum number of steps allowed per episode.
+        current_step: The current step count for the active episode.
+    """
+
+    max_episode_steps: int
+    current_step: int = 0
+
+    def truncated(self, truncated: bool) -> bool:
+        """
+        Increments the step counter and sets the truncated flag if the time limit is reached.
+
+        Args:
+            truncated: The incoming truncated flag.
+
+        Returns:
+            True if the episode step limit is reached, otherwise the incoming value.
+        """
+        self.current_step += 1
+        if self.current_step >= self.max_episode_steps:
+            truncated = True
+        # TODO (steven): missing an else truncated = False?
+        return truncated
+
+    def get_config(self) -> dict[str, Any]:
+        """
+        Returns the configuration of the step for serialization.
+
+        Returns:
+            A dictionary containing the `max_episode_steps`.
+        """
+        return {
+            "max_episode_steps": self.max_episode_steps,
+        }
+
+    def reset(self) -> None:
+        """Resets the step counter, typically called at the start of a new episode."""
+        self.current_step = 0
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        return features
+
+
+@dataclass
+@ProcessorStepRegistry.register("gripper_penalty_processor")
+class GripperPenaltyProcessorStep(ComplementaryDataProcessorStep):
+    """
+    Applies a penalty for inefficient gripper usage.
+
+    This step penalizes actions that attempt to close an already closed gripper or
+    open an already open one, based on position thresholds.
+
+    Attributes:
+        penalty: The negative reward value to apply.
+        max_gripper_pos: The maximum position value for the gripper, used for normalization.
+    """
+
+    penalty: float = -0.01
+    max_gripper_pos: float = 30.0
+
+    def complementary_data(self, complementary_data: dict) -> dict:
+        """
+        Calculates the gripper penalty and adds it to the complementary data.
+
+        Args:
+            complementary_data: The incoming complementary data, which should contain
+                                raw joint positions.
+
+        Returns:
+            A new complementary data dictionary with the `discrete_penalty` key added.
+        """
+        action = self.transition.get(TransitionKey.ACTION)
+
+        current_gripper_pos = complementary_data.get("raw_joint_positions", None).get(GRIPPER_KEY, None)
+        if current_gripper_pos is None:
+            return complementary_data
+
+        gripper_action = action[f"{ACTION}.{GRIPPER_KEY}.pos"]
+        gripper_action_normalized = gripper_action / self.max_gripper_pos
+
+        # Normalize gripper state and action
+        gripper_state_normalized = current_gripper_pos / self.max_gripper_pos
+
+        # Calculate penalty boolean as in original
+        gripper_penalty_bool = (gripper_state_normalized < 0.5 and gripper_action_normalized > 0.5) or (
+            gripper_state_normalized > 0.75 and gripper_action_normalized < 0.5
+        )
+
+        gripper_penalty = self.penalty * int(gripper_penalty_bool)
+
+        # Create new complementary data with penalty info
+        new_complementary_data = dict(complementary_data)
+        new_complementary_data[DISCRETE_PENALTY_KEY] = gripper_penalty
+
+        return new_complementary_data
+
+    def get_config(self) -> dict[str, Any]:
+        """
+        Returns the configuration of the step for serialization.
+
+        Returns:
+            A dictionary containing the penalty value and max gripper position.
+        """
+        return {
+            "penalty": self.penalty,
+            "max_gripper_pos": self.max_gripper_pos,
+        }
+
+    def reset(self) -> None:
+        """Resets the processor's internal state."""
+        pass
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        return features
+
+
+@dataclass
+@ProcessorStepRegistry.register("intervention_action_processor")
+class InterventionActionProcessorStep(ProcessorStep):
+    """
+    Handles human intervention, overriding policy actions and managing episode termination.
+
+    When an intervention is detected (via teleoperator events in the `info` dict),
+    this step replaces the policy's action with the human's teleoperated action.
+    It also processes signals to terminate the episode or flag success.
+
+    Attributes:
+        use_gripper: Whether to include the gripper in the teleoperated action.
+        terminate_on_success: If True, automatically sets the `done` flag when a
+                              `success` event is received.
+    """
+
+    use_gripper: bool = False
+    terminate_on_success: bool = True
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        """
+        Processes the transition to handle interventions.
+
+        Args:
+            transition: The incoming environment transition.
+
+        Returns:
+            The modified transition, potentially with an overridden action, updated
+            reward, and termination status.
+        """
+        action = transition.get(TransitionKey.ACTION)
+        if action is None:
+            return transition
+
+        # Get intervention signals from complementary data
+        info = transition.get(TransitionKey.INFO, {})
+        complementary_data = transition.get(TransitionKey.COMPLEMENTARY_DATA, {})
+        teleop_action = complementary_data.get(TELEOP_ACTION_KEY, {})
+        is_intervention = info.get(TeleopEvents.IS_INTERVENTION, False)
+        terminate_episode = info.get(TeleopEvents.TERMINATE_EPISODE, False)
+        success = info.get(TeleopEvents.SUCCESS, False)
+        rerecord_episode = info.get(TeleopEvents.RERECORD_EPISODE, False)
+
+        new_transition = transition.copy()
+
+        # Override action if intervention is active
+        if is_intervention and teleop_action is not None:
+            if isinstance(teleop_action, dict):
+                # Convert teleop_action dict to tensor format
+                action_list = [
+                    teleop_action.get(f"{ACTION}.delta_x", 0.0),
+                    teleop_action.get(f"{ACTION}.delta_y", 0.0),
+                    teleop_action.get(f"{ACTION}.delta_z", 0.0),
+                ]
+                if self.use_gripper:
+                    action_list.append(teleop_action.get(GRIPPER_KEY, 1.0))
+            elif isinstance(teleop_action, np.ndarray):
+                action_list = teleop_action.tolist()
+            else:
+                action_list = teleop_action
+
+            teleop_action_tensor = torch.tensor(action_list, dtype=action.dtype, device=action.device)
+            new_transition[TransitionKey.ACTION] = teleop_action_tensor
+
+        # Handle episode termination
+        new_transition[TransitionKey.DONE] = bool(terminate_episode) or (
+            self.terminate_on_success and success
+        )
+        new_transition[TransitionKey.REWARD] = float(success)
+
+        # Update info with intervention metadata
+        info = new_transition.get(TransitionKey.INFO, {})
+        info[TeleopEvents.IS_INTERVENTION] = is_intervention
+        info[TeleopEvents.RERECORD_EPISODE] = rerecord_episode
+        info[TeleopEvents.SUCCESS] = success
+        new_transition[TransitionKey.INFO] = info
+
+        # Update complementary data with teleop action
+        complementary_data = new_transition.get(TransitionKey.COMPLEMENTARY_DATA, {})
+        complementary_data[TELEOP_ACTION_KEY] = new_transition.get(TransitionKey.ACTION)
+        new_transition[TransitionKey.COMPLEMENTARY_DATA] = complementary_data
+
+        return new_transition
+
+    def get_config(self) -> dict[str, Any]:
+        """
+        Returns the configuration of the step for serialization.
+
+        Returns:
+            A dictionary containing the step's configuration attributes.
+        """
+        return {
+            "use_gripper": self.use_gripper,
+            "terminate_on_success": self.terminate_on_success,
+        }
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        return features
+
+
+@dataclass
+@ProcessorStepRegistry.register("reward_classifier_processor")
+class RewardClassifierProcessorStep(ProcessorStep):
+    """
+    Applies a pretrained reward classifier to image observations to predict success.
+
+    This step uses a model to determine if the current state is successful, updating
+    the reward and potentially terminating the episode.
+
+    Attributes:
+        pretrained_path: Path to the pretrained reward classifier model.
+        device: The device to run the classifier on.
+        success_threshold: The probability threshold to consider a prediction as successful.
+        success_reward: The reward value to assign on success.
+        terminate_on_success: If True, terminates the episode upon successful classification.
+        reward_classifier: The loaded classifier model instance.
+    """
+
+    pretrained_path: str | None = None
+    device: str = "cpu"
+    success_threshold: float = 0.5
+    success_reward: float = 1.0
+    terminate_on_success: bool = True
+
+    reward_classifier: Any = None
+
+    def __post_init__(self):
+        """Initializes the reward classifier model after the dataclass is created."""
+        if self.pretrained_path is not None:
+            from lerobot.policies.sac.reward_model.modeling_classifier import Classifier
+
+            self.reward_classifier = Classifier.from_pretrained(self.pretrained_path)
+            self.reward_classifier.to(self.device)
+            self.reward_classifier.eval()
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        """
+        Processes a transition, applying the reward classifier to its image observations.
+
+        Args:
+            transition: The incoming environment transition.
+
+        Returns:
+            The modified transition with an updated reward and done flag based on the
+            classifier's prediction.
+        """
+        new_transition = transition.copy()
+        observation = new_transition.get(TransitionKey.OBSERVATION)
+        if observation is None or self.reward_classifier is None:
+            return new_transition
+
+        # Extract images from observation
+        images = {key: value for key, value in observation.items() if "image" in key}
+
+        if not images:
+            return new_transition
+
+        # Run reward classifier
+        start_time = time.perf_counter()
+        with torch.inference_mode():
+            success = self.reward_classifier.predict_reward(images, threshold=self.success_threshold)
+
+        classifier_frequency = 1 / (time.perf_counter() - start_time)
+
+        # Calculate reward and termination
+        reward = new_transition.get(TransitionKey.REWARD, 0.0)
+        terminated = new_transition.get(TransitionKey.DONE, False)
+
+        if math.isclose(success, 1, abs_tol=1e-2):
+            reward = self.success_reward
+            if self.terminate_on_success:
+                terminated = True
+
+        # Update transition
+        new_transition[TransitionKey.REWARD] = reward
+        new_transition[TransitionKey.DONE] = terminated
+
+        # Update info with classifier frequency
+        info = new_transition.get(TransitionKey.INFO, {})
+        info["reward_classifier_frequency"] = classifier_frequency
+        new_transition[TransitionKey.INFO] = info
+
+        return new_transition
+
+    def get_config(self) -> dict[str, Any]:
+        """
+        Returns the configuration of the step for serialization.
+
+        Returns:
+            A dictionary containing the step's configuration attributes.
+        """
+        return {
+            "device": self.device,
+            "success_threshold": self.success_threshold,
+            "success_reward": self.success_reward,
+            "terminate_on_success": self.terminate_on_success,
+        }
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        return features

src/lerobot/processor/joint_observations_processor.py

@@ -0,0 +1,211 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+from typing import Any
+
+import torch
+
+from lerobot.configs.types import PipelineFeatureType, PolicyFeature
+from lerobot.constants import OBS_STATE
+from lerobot.processor.pipeline import (
+    ObservationProcessorStep,
+    ProcessorStepRegistry,
+)
+from lerobot.robots import Robot
+
+
+@dataclass
+@ProcessorStepRegistry.register("joint_velocity_processor")
+class JointVelocityProcessorStep(ObservationProcessorStep):
+    """
+    Calculates and appends joint velocity information to the observation state.
+
+    This step computes the velocity of each joint by calculating the finite
+    difference between the current and the last observed joint positions. The
+    resulting velocity vector is then concatenated to the original state vector.
+
+    Attributes:
+        dt: The time step (delta time) in seconds between observations, used for
+            calculating velocity.
+        last_joint_positions: Stores the joint positions from the previous step
+                              to enable velocity calculation.
+    """
+
+    dt: float = 0.1
+
+    last_joint_positions: torch.Tensor | None = None
+
+    def observation(self, observation: dict) -> dict:
+        """
+        Computes joint velocities and adds them to the observation state.
+
+        Args:
+            observation: The input observation dictionary, expected to contain
+                         an `observation.state` key with joint positions.
+
+        Returns:
+            A new observation dictionary with the `observation.state` tensor
+            extended to include joint velocities.
+
+        Raises:
+            ValueError: If `observation.state` is not found in the observation.
+        """
+        # Get current joint positions (assuming they're in observation.state)
+        current_positions = observation.get(OBS_STATE)
+        if current_positions is None:
+            raise ValueError(f"{OBS_STATE} is not in observation")
+
+        # Initialize last joint positions if not already set
+        if self.last_joint_positions is None:
+            self.last_joint_positions = current_positions.clone()
+            joint_velocities = torch.zeros_like(current_positions)
+        else:
+            # Compute velocities
+            joint_velocities = (current_positions - self.last_joint_positions) / self.dt
+
+        self.last_joint_positions = current_positions.clone()
+
+        # Extend observation with velocities
+        extended_state = torch.cat([current_positions, joint_velocities], dim=-1)
+
+        # Create new observation dict
+        new_observation = dict(observation)
+        new_observation[OBS_STATE] = extended_state
+
+        return new_observation
+
+    def get_config(self) -> dict[str, Any]:
+        """
+        Returns the configuration of the step for serialization.
+
+        Returns:
+            A dictionary containing the time step `dt`.
+        """
+        return {
+            "dt": self.dt,
+        }
+
+    def reset(self) -> None:
+        """Resets the internal state, clearing the last known joint positions."""
+        self.last_joint_positions = None
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        """
+        Updates the `observation.state` feature to reflect the added velocities.
+
+        This method doubles the size of the first dimension of the `observation.state`
+        shape to account for the concatenation of position and velocity vectors.
+
+        Args:
+            features: The policy features dictionary.
+
+        Returns:
+            The updated policy features dictionary.
+        """
+        if OBS_STATE in features[PipelineFeatureType.OBSERVATION]:
+            original_feature = features[PipelineFeatureType.OBSERVATION][OBS_STATE]
+            # Double the shape to account for positions + velocities
+            new_shape = (original_feature.shape[0] * 2,) + original_feature.shape[1:]
+
+            features[PipelineFeatureType.OBSERVATION][OBS_STATE] = PolicyFeature(
+                type=original_feature.type, shape=new_shape
+            )
+        return features
+
+
+@dataclass
+@ProcessorStepRegistry.register("current_processor")
+class MotorCurrentProcessorStep(ObservationProcessorStep):
+    """
+    Reads motor currents from a robot and appends them to the observation state.
+
+    This step queries the robot's hardware interface to get the present current
+    for each motor and concatenates this information to the existing state vector.
+
+    Attributes:
+        robot: An instance of a `lerobot` Robot class that provides access to
+               the hardware bus.
+    """
+
+    robot: Robot | None = None
+
+    def observation(self, observation: dict) -> dict:
+        """
+        Fetches motor currents and adds them to the observation state.
+
+        Args:
+            observation: The input observation dictionary.
+
+        Returns:
+            A new observation dictionary with the `observation.state` tensor
+            extended to include motor currents.
+
+        Raises:
+            ValueError: If the `robot` attribute has not been set.
+        """
+        # Get current values from robot state
+        if self.robot is None:
+            raise ValueError("Robot is not set")
+
+        present_current_dict = self.robot.bus.sync_read("Present_Current")  # type: ignore[attr-defined]
+        motor_currents = torch.tensor(
+            [present_current_dict[name] for name in self.robot.bus.motors],  # type: ignore[attr-defined]
+            dtype=torch.float32,
+        ).unsqueeze(0)
+
+        current_state = observation.get(OBS_STATE)
+        if current_state is None:
+            return observation
+
+        extended_state = torch.cat([current_state, motor_currents], dim=-1)
+
+        # Create new observation dict
+        new_observation = dict(observation)
+        new_observation[OBS_STATE] = extended_state
+
+        return new_observation
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        """
+        Updates the `observation.state` feature to reflect the added motor currents.
+
+        This method increases the size of the first dimension of the `observation.state`
+        shape by the number of motors in the robot.
+
+        Args:
+            features: The policy features dictionary.
+
+        Returns:
+            The updated policy features dictionary.
+        """
+        if OBS_STATE in features[PipelineFeatureType.OBSERVATION] and self.robot is not None:
+            original_feature = features[PipelineFeatureType.OBSERVATION][OBS_STATE]
+            # Add motor current dimensions to the original state shape
+            num_motors = 0
+            if hasattr(self.robot, "bus") and hasattr(self.robot.bus, "motors"):  # type: ignore[attr-defined]
+                num_motors = len(self.robot.bus.motors)  # type: ignore[attr-defined]
+
+            if num_motors > 0:
+                new_shape = (original_feature.shape[0] + num_motors,) + original_feature.shape[1:]
+                features[PipelineFeatureType.OBSERVATION][OBS_STATE] = PolicyFeature(
+                    type=original_feature.type, shape=new_shape
+                )
+        return features

src/lerobot/processor/migrate_policy_normalization.py

@@ -0,0 +1,572 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+A generic script to migrate LeRobot policies with built-in normalization layers to the new
+pipeline-based processor system.
+
+This script performs the following steps:
+1.  Loads a pretrained policy model and its configuration from a local path or the
+    Hugging Face Hub.
+2.  Scans the model's state dictionary to extract normalization statistics (e.g., mean,
+    std, min, max) for all features.
+3.  Creates two new processor pipelines:
+    - A preprocessor that normalizes inputs (observations) and outputs (actions).
+    - A postprocessor that unnormalizes outputs (actions) for inference.
+4.  Removes the original normalization layers from the model's state dictionary,
+    creating a "clean" model.
+5.  Saves the new clean model, the preprocessor, the postprocessor, and a generated
+    model card to a new directory.
+6.  Optionally pushes all the new artifacts to the Hugging Face Hub.
+
+Usage:
+    python src/lerobot/processor/migrate_policy_normalization.py \
+        --pretrained-path lerobot/act_aloha_sim_transfer_cube_human \
+        --policy-type act \
+        --push-to-hub
+"""
+
+import argparse
+import importlib
+import json
+import os
+from copy import deepcopy
+from pathlib import Path
+from typing import Any
+
+import torch
+from huggingface_hub import hf_hub_download
+from safetensors.torch import load_file as load_safetensors
+
+from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
+
+from .batch_processor import AddBatchDimensionProcessorStep
+from .device_processor import DeviceProcessorStep
+from .normalize_processor import NormalizerProcessorStep, UnnormalizerProcessorStep
+from .pipeline import PolicyProcessorPipeline
+from .rename_processor import RenameObservationsProcessorStep
+
+# Policy type to class mapping
+POLICY_CLASSES = {
+    "act": "lerobot.policies.act.modeling_act.ACTPolicy",
+    "diffusion": "lerobot.policies.diffusion.modeling_diffusion.DiffusionPolicy",
+    "pi0": "lerobot.policies.pi0.modeling_pi0.PI0Policy",
+    "pi0fast": "lerobot.policies.pi0fast.modeling_pi0fast.PI0FASTPolicy",
+    "smolvla": "lerobot.policies.smolvla.modeling_smolvla.SmolVLAPolicy",
+    "tdmpc": "lerobot.policies.tdmpc.modeling_tdmpc.TDMPCPolicy",
+    "vqbet": "lerobot.policies.vqbet.modeling_vqbet.VQBeTPolicy",
+    "sac": "lerobot.policies.sac.modeling_sac.SACPolicy",
+    "classifier": "lerobot.policies.classifier.modeling_classifier.ClassifierPolicy",
+}
+
+
+def extract_normalization_stats(state_dict: dict[str, torch.Tensor]) -> dict[str, dict[str, torch.Tensor]]:
+    """
+    Scans a model's state_dict to find and extract normalization statistics.
+
+    This function identifies keys corresponding to normalization layers (e.g., those
+    for mean, std, min, max) based on a set of predefined patterns and organizes
+    them into a nested dictionary.
+
+    Args:
+        state_dict: The state dictionary of a pretrained policy model.
+
+    Returns:
+        A nested dictionary where outer keys are feature names (e.g.,
+        'observation.state') and inner keys are statistic types ('mean', 'std'),
+        mapping to their corresponding tensor values.
+    """
+    stats = {}
+
+    # Define patterns to match and their prefixes to remove
+    normalization_patterns = [
+        "normalize_inputs.buffer_",
+        "unnormalize_outputs.buffer_",
+        "normalize_targets.buffer_",
+        "normalize.",  # Must come after normalize_* patterns
+        "unnormalize.",  # Must come after unnormalize_* patterns
+        "input_normalizer.",
+        "output_normalizer.",
+    ]
+
+    # Process each key in state_dict
+    for key, tensor in state_dict.items():
+        # Try each pattern
+        for pattern in normalization_patterns:
+            if key.startswith(pattern):
+                # Extract the remaining part after the pattern
+                remaining = key[len(pattern) :]
+                parts = remaining.split(".")
+
+                # Need at least feature name and stat type
+                if len(parts) >= 2:
+                    # Last part is the stat type (mean, std, min, max, etc.)
+                    stat_type = parts[-1]
+                    # Everything else is the feature name
+                    feature_name = ".".join(parts[:-1]).replace("_", ".")
+
+                    # Add to stats
+                    if feature_name not in stats:
+                        stats[feature_name] = {}
+                    stats[feature_name][stat_type] = tensor.clone()
+
+                # Only process the first matching pattern
+                break
+
+    return stats
+
+
+def detect_features_and_norm_modes(
+    config: dict[str, Any], stats: dict[str, dict[str, torch.Tensor]]
+) -> tuple[dict[str, PolicyFeature], dict[FeatureType, NormalizationMode]]:
+    """
+    Infers policy features and normalization modes from the model config and stats.
+
+    This function first attempts to find feature definitions and normalization
+    mappings directly from the policy's configuration file. If this information is
+    not present, it infers it from the extracted normalization statistics, using
+    tensor shapes to determine feature shapes and the presence of specific stat
+    keys (e.g., 'mean'/'std' vs 'min'/'max') to determine the normalization mode.
+    It applies sensible defaults if inference is not possible.
+
+    Args:
+        config: The policy's configuration dictionary from `config.json`.
+        stats: The normalization statistics extracted from the model's state_dict.
+
+    Returns:
+        A tuple containing:
+        - A dictionary mapping feature names to `PolicyFeature` objects.
+        - A dictionary mapping `FeatureType` enums to `NormalizationMode` enums.
+    """
+    features = {}
+    norm_modes = {}
+
+    # First, check if there's a normalization_mapping in the config
+    if "normalization_mapping" in config:
+        print(f"Found normalization_mapping in config: {config['normalization_mapping']}")
+        # Extract normalization modes from config
+        for feature_name, mode_str in config["normalization_mapping"].items():
+            # Convert string to NormalizationMode enum
+            if mode_str == "mean_std":
+                mode = NormalizationMode.MEAN_STD
+            elif mode_str == "min_max":
+                mode = NormalizationMode.MIN_MAX
+            else:
+                print(f"Warning: Unknown normalization mode '{mode_str}' for feature '{feature_name}'")
+                continue
+
+            # Determine feature type from feature name
+            if "image" in feature_name or "visual" in feature_name:
+                feature_type = FeatureType.VISUAL
+            elif "state" in feature_name:
+                feature_type = FeatureType.STATE
+            elif "action" in feature_name:
+                feature_type = FeatureType.ACTION
+            else:
+                feature_type = FeatureType.STATE
+
+            norm_modes[feature_type] = mode
+
+    # Try to extract from config
+    if "features" in config:
+        for key, feature_config in config["features"].items():
+            shape = feature_config.get("shape", feature_config.get("dim"))
+            shape = (shape,) if isinstance(shape, int) else tuple(shape)
+
+            # Determine feature type
+            if "image" in key or "visual" in key:
+                feature_type = FeatureType.VISUAL
+            elif "state" in key:
+                feature_type = FeatureType.STATE
+            elif "action" in key:
+                feature_type = FeatureType.ACTION
+            else:
+                feature_type = FeatureType.STATE  # Default
+
+            features[key] = PolicyFeature(feature_type, shape)
+
+    # If no features in config, infer from stats
+    if not features:
+        for key, stat_dict in stats.items():
+            # Get shape from any stat tensor
+            tensor = next(iter(stat_dict.values()))
+            shape = tuple(tensor.shape)
+
+            # Determine feature type based on key
+            if "image" in key or "visual" in key or "pixels" in key:
+                feature_type = FeatureType.VISUAL
+            elif "state" in key or "joint" in key or "position" in key:
+                feature_type = FeatureType.STATE
+            elif "action" in key:
+                feature_type = FeatureType.ACTION
+            else:
+                feature_type = FeatureType.STATE
+
+            features[key] = PolicyFeature(feature_type, shape)
+
+    # If normalization modes weren't in config, determine based on available stats
+    if not norm_modes:
+        for key, stat_dict in stats.items():
+            if key in features:
+                if "mean" in stat_dict and "std" in stat_dict:
+                    feature_type = features[key].type
+                    if feature_type not in norm_modes:
+                        norm_modes[feature_type] = NormalizationMode.MEAN_STD
+                elif "min" in stat_dict and "max" in stat_dict:
+                    feature_type = features[key].type
+                    if feature_type not in norm_modes:
+                        norm_modes[feature_type] = NormalizationMode.MIN_MAX
+
+    # Default normalization modes if not detected
+    if FeatureType.VISUAL not in norm_modes:
+        norm_modes[FeatureType.VISUAL] = NormalizationMode.MEAN_STD
+    if FeatureType.STATE not in norm_modes:
+        norm_modes[FeatureType.STATE] = NormalizationMode.MIN_MAX
+    if FeatureType.ACTION not in norm_modes:
+        norm_modes[FeatureType.ACTION] = NormalizationMode.MEAN_STD
+
+    return features, norm_modes
+
+
+def remove_normalization_layers(state_dict: dict[str, torch.Tensor]) -> dict[str, torch.Tensor]:
+    """
+    Creates a new state_dict with all normalization-related layers removed.
+
+    This function filters the original state dictionary, excluding any keys that
+    match a set of predefined patterns associated with normalization modules.
+
+    Args:
+        state_dict: The original model state dictionary.
+
+    Returns:
+        A new state dictionary containing only the core model weights, without
+        any normalization parameters.
+    """
+    new_state_dict = {}
+
+    # Patterns to remove
+    remove_patterns = [
+        "normalize_inputs.",
+        "unnormalize_outputs.",
+        "normalize_targets.",  # Added pattern for target normalization
+        "normalize.",
+        "unnormalize.",
+        "input_normalizer.",
+        "output_normalizer.",
+        "normalizer.",
+    ]
+
+    for key, tensor in state_dict.items():
+        should_remove = any(pattern in key for pattern in remove_patterns)
+        if not should_remove:
+            new_state_dict[key] = tensor
+
+    return new_state_dict
+
+
+def convert_features_to_policy_features(features_dict: dict[str, dict]) -> dict[str, PolicyFeature]:
+    """
+    Converts a feature dictionary from the old config format to the new `PolicyFeature` format.
+
+    Args:
+        features_dict: The feature dictionary in the old format, where values are
+                       simple dictionaries (e.g., `{"shape": [7]}`).
+
+    Returns:
+        A dictionary mapping feature names to `PolicyFeature` dataclass objects.
+    """
+    converted_features = {}
+
+    for key, feature_dict in features_dict.items():
+        # Determine feature type based on key
+        if "image" in key or "visual" in key:
+            feature_type = FeatureType.VISUAL
+        elif "state" in key:
+            feature_type = FeatureType.STATE
+        elif "action" in key:
+            feature_type = FeatureType.ACTION
+        else:
+            feature_type = FeatureType.STATE
+
+        # Get shape from feature dict
+        shape = feature_dict.get("shape", feature_dict.get("dim"))
+        shape = (shape,) if isinstance(shape, int) else tuple(shape)
+
+        converted_features[key] = PolicyFeature(feature_type, shape)
+
+    return converted_features
+
+
+def load_model_from_hub(
+    repo_id: str, revision: str = None
+) -> tuple[dict[str, torch.Tensor], dict[str, Any], dict[str, Any]]:
+    """
+    Downloads and loads a model's state_dict and configs from the Hugging Face Hub.
+
+    Args:
+        repo_id: The repository ID on the Hub (e.g., 'lerobot/aloha').
+        revision: The specific git revision (branch, tag, or commit hash) to use.
+
+    Returns:
+        A tuple containing the model's state dictionary, the policy configuration,
+        and the training configuration.
+    """
+    # Download files.
+    safetensors_path = hf_hub_download(repo_id=repo_id, filename="model.safetensors", revision=revision)
+
+    config_path = hf_hub_download(repo_id=repo_id, filename="config.json", revision=revision)
+    train_config_path = hf_hub_download(repo_id=repo_id, filename="train_config.json", revision=revision)
+
+    # Load state_dict
+    state_dict = load_safetensors(safetensors_path)
+
+    # Load config
+    with open(config_path) as f:
+        config = json.load(f)
+
+    with open(train_config_path) as f:
+        train_config = json.load(f)
+
+    return state_dict, config, train_config
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Migrate policy models with normalization layers to new pipeline system"
+    )
+    parser.add_argument(
+        "--pretrained-path",
+        type=str,
+        required=True,
+        help="Path to pretrained model (hub repo or local directory)",
+    )
+    parser.add_argument(
+        "--output-dir",
+        type=str,
+        default=None,
+        help="Output directory for migrated model (default: same as pretrained-path)",
+    )
+    parser.add_argument("--push-to-hub", action="store_true", help="Push migrated model to hub")
+    parser.add_argument(
+        "--hub-repo-id",
+        type=str,
+        default=None,
+        help="Hub repository ID for pushing (default: same as pretrained-path)",
+    )
+    parser.add_argument("--revision", type=str, default=None, help="Revision of the model to load")
+    parser.add_argument("--private", action="store_true", help="Make the hub repository private")
+
+    args = parser.parse_args()
+
+    # Load model and config
+    print(f"Loading model from {args.pretrained_path}...")
+    if os.path.isdir(args.pretrained_path):
+        # Local directory
+        state_dict = load_safetensors(os.path.join(args.pretrained_path, "model.safetensors"))
+        with open(os.path.join(args.pretrained_path, "config.json")) as f:
+            config = json.load(f)
+        with open(os.path.join(args.pretrained_path, "train_config.json")) as f:
+            train_config = json.load(f)
+    else:
+        # Hub repository
+        state_dict, config, train_config = load_model_from_hub(args.pretrained_path, args.revision)
+
+    # Extract normalization statistics
+    print("Extracting normalization statistics...")
+    stats = extract_normalization_stats(state_dict)
+
+    print(f"Found normalization statistics for: {list(stats.keys())}")
+
+    # Detect input features and normalization modes
+    print("Detecting features and normalization modes...")
+    features, norm_map = detect_features_and_norm_modes(config, stats)
+
+    print(f"Detected features: {list(features.keys())}")
+    print(f"Normalization modes: {norm_map}")
+
+    # Remove normalization layers from state_dict
+    print("Removing normalization layers from model...")
+    new_state_dict = remove_normalization_layers(state_dict)
+
+    removed_keys = set(state_dict.keys()) - set(new_state_dict.keys())
+    if removed_keys:
+        print(f"Removed {len(removed_keys)} normalization layer keys")
+
+    # Determine output path
+    if args.output_dir:
+        output_dir = Path(args.output_dir)
+    else:
+        if os.path.isdir(args.pretrained_path):
+            output_dir = Path(args.pretrained_path).parent / f"{Path(args.pretrained_path).name}_migrated"
+        else:
+            output_dir = Path(f"./{args.pretrained_path.replace('/', '_')}_migrated")
+
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    # Clean up config - remove normalization_mapping field
+    cleaned_config = dict(config)
+    if "normalization_mapping" in cleaned_config:
+        print("Removing 'normalization_mapping' field from config")
+        del cleaned_config["normalization_mapping"]
+    policy_type = deepcopy(cleaned_config["type"])
+
+    del cleaned_config["type"]
+
+    # Instantiate the policy model with cleaned config and load the cleaned state dict
+    print(f"Instantiating {policy_type} policy model...")
+    policy_class_path = POLICY_CLASSES[policy_type]
+    module_path, class_name = policy_class_path.rsplit(".", 1)
+
+    module = importlib.import_module(module_path)
+    policy_class = getattr(module, class_name)
+
+    # Create config class instance
+    config_module_path = module_path.replace("modeling", "configuration")
+    config_module = importlib.import_module(config_module_path)
+    # Handle special cases for config class names
+    config_class_names = {
+        "act": "ACTConfig",
+        "diffusion": "DiffusionConfig",
+        "pi0": "PI0Config",
+        "pi0fast": "PI0FASTConfig",
+        "smolvla": "SmolVLAConfig",
+        "tdmpc": "TDMPCConfig",
+        "vqbet": "VQBeTConfig",
+        "sac": "SACConfig",
+        "classifier": "ClassifierConfig",
+    }
+    config_class_name = config_class_names.get(policy_type, f"{policy_type.upper()}Config")
+    config_class = getattr(config_module, config_class_name)
+
+    # Convert input_features and output_features to PolicyFeature objects - these are mandatory
+    if "input_features" not in cleaned_config:
+        raise ValueError("Missing mandatory 'input_features' in config")
+    if "output_features" not in cleaned_config:
+        raise ValueError("Missing mandatory 'output_features' in config")
+
+    cleaned_config["input_features"] = convert_features_to_policy_features(cleaned_config["input_features"])
+    cleaned_config["output_features"] = convert_features_to_policy_features(cleaned_config["output_features"])
+
+    # Create config instance from cleaned config dict
+    policy_config = config_class(**cleaned_config)
+
+    # Create policy instance - some policies expect dataset_stats
+    policy = policy_class(policy_config)
+
+    # Load the cleaned state dict
+    policy.load_state_dict(new_state_dict, strict=True)
+    print("Successfully loaded cleaned state dict into policy model")
+
+    # Now create preprocessor and postprocessor with cleaned_config available
+    print("Creating preprocessor and postprocessor...")
+    # The pattern from existing processor factories:
+    # - Preprocessor has two NormalizerProcessorSteps: one for input_features, one for output_features
+    # - Postprocessor has one UnnormalizerProcessorStep for output_features only
+
+    # Get features from cleaned_config (now they're PolicyFeature objects)
+    input_features = cleaned_config.get("input_features", {})
+    output_features = cleaned_config.get("output_features", {})
+
+    # Create preprocessor with two normalizers (following the pattern from processor factories)
+    preprocessor_steps = [
+        RenameObservationsProcessorStep(rename_map={}),
+        NormalizerProcessorStep(
+            features={**input_features, **output_features},
+            norm_map=norm_map,
+            stats=stats,
+        ),
+        AddBatchDimensionProcessorStep(),
+        DeviceProcessorStep(device=policy_config.device),
+    ]
+    preprocessor = PolicyProcessorPipeline(steps=preprocessor_steps, name="robot_preprocessor")
+
+    # Create postprocessor with unnormalizer for outputs only
+    postprocessor_steps = [
+        DeviceProcessorStep(device="cpu"),
+        UnnormalizerProcessorStep(features=output_features, norm_map=norm_map, stats=stats),
+    ]
+    postprocessor = PolicyProcessorPipeline(steps=postprocessor_steps, name="robot_postprocessor")
+
+    # Determine hub repo ID if pushing to hub
+    if args.push_to_hub:
+        if args.hub_repo_id:
+            hub_repo_id = args.hub_repo_id
+        else:
+            if not os.path.isdir(args.pretrained_path):
+                # Use same repo with "_migrated" suffix
+                hub_repo_id = f"{args.pretrained_path}_migrated"
+            else:
+                raise ValueError("--hub-repo-id must be specified when pushing local model to hub")
+    else:
+        hub_repo_id = None
+
+    # Save preprocessor and postprocessor to root directory
+    print(f"Saving preprocessor to {output_dir}...")
+    preprocessor.save_pretrained(output_dir)
+    if args.push_to_hub:
+        preprocessor.push_to_hub(repo_id=hub_repo_id, private=args.private)
+
+    print(f"Saving postprocessor to {output_dir}...")
+    postprocessor.save_pretrained(output_dir)
+    if args.push_to_hub:
+        postprocessor.push_to_hub(repo_id=hub_repo_id, private=args.private)
+
+    # Save model using the policy's save_pretrained method
+    print(f"Saving model to {output_dir}...")
+    policy.save_pretrained(
+        output_dir, push_to_hub=args.push_to_hub, repo_id=hub_repo_id, private=args.private
+    )
+
+    # Generate and save model card
+    print("Generating model card...")
+    # Get metadata from original config
+    dataset_repo_id = train_config.get("repo_id", "unknown")
+    license = config.get("license", "apache-2.0")
+
+    tags = config.get("tags", ["robotics", "lerobot", policy_type]) or ["robotics", "lerobot", policy_type]
+    tags = set(tags).union({"robotics", "lerobot", policy_type})
+    tags = list(tags)
+
+    # Generate model card
+    card = policy.generate_model_card(
+        dataset_repo_id=dataset_repo_id, model_type=policy_type, license=license, tags=tags
+    )
+
+    # Save model card locally
+    card.save(str(output_dir / "README.md"))
+    print(f"Model card saved to {output_dir / 'README.md'}")
+    # Push model card to hub if requested
+    if args.push_to_hub:
+        from huggingface_hub import HfApi
+
+        api = HfApi()
+        api.upload_file(
+            path_or_fileobj=str(output_dir / "README.md"),
+            path_in_repo="README.md",
+            repo_id=hub_repo_id,
+            repo_type="model",
+            commit_message="Add model card for migrated model",
+        )
+        print("Model card pushed to hub")
+
+    print("\nMigration complete!")
+    print(f"Migrated model saved to: {output_dir}")
+    if args.push_to_hub:
+        print(f"Successfully pushed to https://huggingface.co/{hub_repo_id}")
+
+
+if __name__ == "__main__":
+    main()

src/lerobot/processor/normalize_processor.py

@@ -1,67 +1,303 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 from __future__ import annotations
 
-from collections.abc import Mapping
+from copy import deepcopy
 from dataclasses import dataclass, field
 from typing import Any
 
-import numpy as np
 import torch
 from torch import Tensor
 
-from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
+from lerobot.configs.types import FeatureType, NormalizationMode, PipelineFeatureType, PolicyFeature
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.processor.pipeline import EnvTransition, ProcessorStepRegistry, TransitionKey
+
+from .converters import from_tensor_to_numpy, to_tensor
+from .core import EnvTransition, TransitionKey
+from .pipeline import PolicyProcessorPipeline, ProcessorStep, ProcessorStepRegistry
 
 
-def _convert_stats_to_tensors(stats: dict[str, dict[str, Any]]) -> dict[str, dict[str, Tensor]]:
-    """Convert numpy arrays and other types to torch tensors."""
-    tensor_stats: dict[str, dict[str, Tensor]] = {}
-    for key, sub in stats.items():
-        tensor_stats[key] = {}
-        for stat_name, value in sub.items():
-            if isinstance(value, np.ndarray):
-                tensor_val = torch.from_numpy(value.astype(np.float32))
-            elif isinstance(value, torch.Tensor):
-                tensor_val = value.to(dtype=torch.float32)
-            elif isinstance(value, (int, float, list, tuple)):
-                tensor_val = torch.tensor(value, dtype=torch.float32)
-            else:
-                raise TypeError(f"Unsupported type for stats['{key}']['{stat_name}']: {type(value)}")
-            tensor_stats[key][stat_name] = tensor_val
-    return tensor_stats
+@dataclass
+class _NormalizationMixin:
+    """
+    A mixin class providing core functionality for normalization and unnormalization.
+
+    This class manages normalization statistics (`stats`), converts them to tensors for
+    efficient computation, handles device placement, and implements the logic for
+    applying normalization transformations (mean/std and min/max). It is designed to
+    be inherited by concrete `ProcessorStep` implementations and should not be used
+    directly.
+
+    Attributes:
+        features: A dictionary mapping feature names to `PolicyFeature` objects, defining
+            the data structure to be processed.
+        norm_map: A dictionary mapping `FeatureType` to `NormalizationMode`, specifying
+            which normalization method to use for each type of feature.
+        stats: A dictionary containing the normalization statistics (e.g., mean, std,
+            min, max) for each feature.
+        device: The PyTorch device on which to store and perform tensor operations.
+        eps: A small epsilon value to prevent division by zero in normalization
+            calculations.
+        normalize_observation_keys: An optional set of keys to selectively apply
+            normalization to specific observation features.
+        _tensor_stats: An internal dictionary holding the normalization statistics as
+            PyTorch tensors.
+    """
+
+    features: dict[str, PolicyFeature]
+    norm_map: dict[FeatureType, NormalizationMode]
+    stats: dict[str, dict[str, Any]] | None = None
+    device: torch.device | str | None = None
+    dtype: torch.dtype | None = None
+    eps: float = 1e-8
+    normalize_observation_keys: set[str] | None = None
+
+    _tensor_stats: dict[str, dict[str, Tensor]] = field(default_factory=dict, init=False, repr=False)
+
+    def __post_init__(self):
+        """
+        Initializes the mixin after dataclass construction.
+
+        This method handles the robust deserialization of `features` and `norm_map`
+        from JSON-compatible formats (where enums become strings and tuples become
+        lists) and converts the provided `stats` dictionary into a dictionary of
+        tensors (`_tensor_stats`) on the specified device.
+        """
+        # Robust JSON deserialization handling (guard empty maps).
+        if self.features:
+            first_val = next(iter(self.features.values()))
+            if isinstance(first_val, dict):
+                reconstructed = {}
+                for key, ft_dict in self.features.items():
+                    reconstructed[key] = PolicyFeature(
+                        type=FeatureType(ft_dict["type"]), shape=tuple(ft_dict["shape"])
+                    )
+                self.features = reconstructed
+
+        if self.norm_map:
+            # if keys are strings (JSON), rebuild enum map
+            if all(isinstance(k, str) for k in self.norm_map.keys()):
+                reconstructed = {}
+                for ft_type_str, norm_mode_str in self.norm_map.items():
+                    reconstructed[FeatureType(ft_type_str)] = NormalizationMode(norm_mode_str)
+                self.norm_map = reconstructed
+
+        # Convert stats to tensors and move to the target device once during initialization.
+        self.stats = self.stats or {}
+        if self.dtype is None:
+            self.dtype = torch.float32
+        self._tensor_stats = to_tensor(self.stats, device=self.device, dtype=self.dtype)
+
+    def to(
+        self, device: torch.device | str | None = None, dtype: torch.dtype | None = None
+    ) -> _NormalizationMixin:
+        """
+        Moves the processor's normalization stats to the specified device.
+
+        Args:
+            device: The target PyTorch device.
+
+        Returns:
+            The instance of the class, allowing for method chaining.
+        """
+        if device is not None:
+            self.device = device
+        if dtype is not None:
+            self.dtype = dtype
+        self._tensor_stats = to_tensor(self.stats, device=self.device, dtype=self.dtype)
+        return self
+
+    def state_dict(self) -> dict[str, Tensor]:
+        """
+        Returns the normalization statistics as a flat state dictionary.
+
+        All tensors are moved to the CPU before being returned, which is standard practice
+        for saving state dictionaries.
+
+        Returns:
+            A flat dictionary mapping from `'feature_name.stat_name'` to the
+            corresponding statistics tensor on the CPU.
+        """
+        flat: dict[str, Tensor] = {}
+        for key, sub in self._tensor_stats.items():
+            for stat_name, tensor in sub.items():
+                flat[f"{key}.{stat_name}"] = tensor.cpu()  # Always save to CPU
+        return flat
+
+    def load_state_dict(self, state: dict[str, Tensor]) -> None:
+        """
+        Loads normalization statistics from a state dictionary.
+
+        The loaded tensors are moved to the processor's configured device.
+
+        Args:
+            state: A flat state dictionary with keys in the format
+                   `'feature_name.stat_name'`.
+        """
+        self._tensor_stats.clear()
+        for flat_key, tensor in state.items():
+            key, stat_name = flat_key.rsplit(".", 1)
+            # Load to the processor's configured device.
+            self._tensor_stats.setdefault(key, {})[stat_name] = tensor.to(
+                dtype=torch.float32, device=self.device
+            )
+
+        # Reconstruct the original stats dict from tensor stats for compatibility with to() method
+        # and other functions that rely on self.stats
+
+        self.stats = {}
+        for key, tensor_dict in self._tensor_stats.items():
+            self.stats[key] = {}
+            for stat_name, tensor in tensor_dict.items():
+                # Convert tensor back to python/numpy format
+                self.stats[key][stat_name] = from_tensor_to_numpy(tensor)
+
+    def get_config(self) -> dict[str, Any]:
+        """
+        Returns a serializable dictionary of the processor's configuration.
+
+        This method is used when saving the processor to disk, ensuring that its
+        configuration can be reconstructed later.
+
+        Returns:
+            A JSON-serializable dictionary containing the configuration.
+        """
+        config = {
+            "eps": self.eps,
+            "features": {
+                key: {"type": ft.type.value, "shape": ft.shape} for key, ft in self.features.items()
+            },
+            "norm_map": {ft_type.value: norm_mode.value for ft_type, norm_mode in self.norm_map.items()},
+        }
+        if self.normalize_observation_keys is not None:
+            config["normalize_observation_keys"] = sorted(self.normalize_observation_keys)
+        return config
+
+    def _normalize_observation(self, observation: dict[str, Any], inverse: bool) -> dict[str, Tensor]:
+        """
+        Applies (un)normalization to all relevant features in an observation dictionary.
+
+        Args:
+            observation: The observation dictionary to process.
+            inverse: If `True`, applies unnormalization; otherwise, applies normalization.
+
+        Returns:
+            A new observation dictionary with the transformed tensor values.
+        """
+        new_observation = dict(observation)
+        for key, feature in self.features.items():
+            if self.normalize_observation_keys is not None and key not in self.normalize_observation_keys:
+                continue
+            if feature.type != FeatureType.ACTION and key in new_observation:
+                # Convert to tensor but preserve original dtype for adaptation logic
+                tensor = torch.as_tensor(new_observation[key])
+                new_observation[key] = self._apply_transform(tensor, key, feature.type, inverse=inverse)
+        return new_observation
+
+    def _normalize_action(self, action: Any, inverse: bool) -> Tensor:
+        # Convert to tensor but preserve original dtype for adaptation logic
+        """
+        Applies (un)normalization to an action tensor.
+
+        Args:
+            action: The action tensor to process.
+            inverse: If `True`, applies unnormalization; otherwise, applies normalization.
+
+        Returns:
+            The transformed action tensor.
+        """
+        tensor = torch.as_tensor(action)
+        processed_action = self._apply_transform(tensor, "action", FeatureType.ACTION, inverse=inverse)
+        return processed_action
+
+    def _apply_transform(
+        self, tensor: Tensor, key: str, feature_type: FeatureType, *, inverse: bool = False
+    ) -> Tensor:
+        """
+        Core logic to apply a normalization or unnormalization transformation to a tensor.
+
+        This method selects the appropriate normalization mode (e.g., mean/std, min/max)
+        based on the feature type and applies the corresponding mathematical operation.
+
+        Args:
+            tensor: The input tensor to transform.
+            key: The feature key corresponding to the tensor.
+            feature_type: The `FeatureType` of the tensor.
+            inverse: If `True`, applies the inverse transformation (unnormalization).
+
+        Returns:
+            The transformed tensor.
+
+        Raises:
+            ValueError: If an unsupported normalization mode is encountered.
+        """
+        norm_mode = self.norm_map.get(feature_type, NormalizationMode.IDENTITY)
+        if norm_mode == NormalizationMode.IDENTITY or key not in self._tensor_stats:
+            return tensor
+
+        if norm_mode not in (NormalizationMode.MEAN_STD, NormalizationMode.MIN_MAX):
+            raise ValueError(f"Unsupported normalization mode: {norm_mode}")
+
+        # For Accelerate compatibility: Ensure stats are on the same device and dtype as the input tensor
+        if self._tensor_stats and key in self._tensor_stats:
+            first_stat = next(iter(self._tensor_stats[key].values()))
+            if first_stat.device != tensor.device or first_stat.dtype != tensor.dtype:
+                self.to(device=tensor.device, dtype=tensor.dtype)
+
+        stats = self._tensor_stats[key]
+
+        if norm_mode == NormalizationMode.MEAN_STD and "mean" in stats and "std" in stats:
+            mean, std = stats["mean"], stats["std"]
+            # Avoid division by zero by adding a small epsilon.
+            denom = std + self.eps
+            if inverse:
+                return tensor * std + mean
+            return (tensor - mean) / denom
+
+        if norm_mode == NormalizationMode.MIN_MAX and "min" in stats and "max" in stats:
+            min_val, max_val = stats["min"], stats["max"]
+            denom = max_val - min_val
+            # When min_val == max_val, substitute the denominator with a small epsilon
+            # to prevent division by zero. This consistently maps an input equal to
+            # min_val to -1, ensuring a stable transformation.
+            denom = torch.where(
+                denom == 0, torch.tensor(self.eps, device=tensor.device, dtype=tensor.dtype), denom
+            )
+            if inverse:
+                # Map from [-1, 1] back to [min, max]
+                return (tensor + 1) / 2 * denom + min_val
+            # Map from [min, max] to [-1, 1]
+            return 2 * (tensor - min_val) / denom - 1
+
+        # If necessary stats are missing, return input unchanged.
+        return tensor
 
 
 @dataclass
 @ProcessorStepRegistry.register(name="normalizer_processor")
-class NormalizerProcessor:
-    """Normalizes observations and actions in a single processor step.
-
-    This processor handles normalization of both observation and action tensors
-    using either mean/std normalization or min/max scaling to a [-1, 1] range.
-
-    For each tensor key in the stats dictionary, the processor will:
-    - Use mean/std normalization if those statistics are provided: (x - mean) / std
-    - Use min/max scaling if those statistics are provided: 2 * (x - min) / (max - min) - 1
-
-    The processor can be configured to normalize only specific keys by setting
-    the normalize_keys parameter.
+class NormalizerProcessorStep(_NormalizationMixin, ProcessorStep):
     """
+    A processor step that applies normalization to observations and actions in a transition.
 
-    # Features and normalisation map are mandatory to match the design of normalize.py
-    features: dict[str, PolicyFeature]
-    norm_map: dict[FeatureType, NormalizationMode]
-
-    # Pre-computed statistics coming from dataset.meta.stats for instance.
-    stats: dict[str, dict[str, Any]] | None = None
-
-    # Explicit subset of keys to normalise. If ``None`` every key (except
-    # "action") found in ``stats`` will be normalised. Using a ``set`` makes
-    # membership checks O(1).
-    normalize_keys: set[str] | None = None
-
-    eps: float = 1e-8
-
-    _tensor_stats: dict[str, dict[str, Tensor]] = field(default_factory=dict, init=False, repr=False)
+    This class uses the logic from `_NormalizationMixin` to perform forward normalization
+    (e.g., scaling data to have zero mean and unit variance, or to the range [-1, 1]).
+    It is typically used in the pre-processing pipeline before feeding data to a policy.
+    """
 
     @classmethod
     def from_lerobot_dataset(
@@ -70,158 +306,67 @@ class NormalizerProcessor:
         features: dict[str, PolicyFeature],
         norm_map: dict[FeatureType, NormalizationMode],
         *,
-        normalize_keys: set[str] | None = None,
+        normalize_observation_keys: set[str] | None = None,
         eps: float = 1e-8,
-    ) -> NormalizerProcessor:
-        """Factory helper that pulls statistics from a :class:`LeRobotDataset`.
-
-        The features and norm_map parameters are mandatory to match the design
-        pattern used in normalize.py.
+        device: torch.device | str | None = None,
+    ) -> NormalizerProcessorStep:
         """
+        Creates a `NormalizerProcessorStep` instance using statistics from a `LeRobotDataset`.
 
+        Args:
+            dataset: The dataset from which to extract normalization statistics.
+            features: The feature definition for the processor.
+            norm_map: The mapping from feature types to normalization modes.
+            normalize_observation_keys: An optional set of observation keys to normalize.
+            eps: A small epsilon value for numerical stability.
+            device: The target device for the processor.
+
+        Returns:
+            A new instance of `NormalizerProcessorStep`.
+        """
         return cls(
             features=features,
             norm_map=norm_map,
             stats=dataset.meta.stats,
-            normalize_keys=normalize_keys,
+            normalize_observation_keys=normalize_observation_keys,
             eps=eps,
+            device=device,
         )
 
-    def __post_init__(self):
-        # Handle deserialization from JSON config
-        if self.features and isinstance(list(self.features.values())[0], dict):
-            # Features came from JSON - need to reconstruct PolicyFeature objects
-            reconstructed_features = {}
-            for key, ft_dict in self.features.items():
-                reconstructed_features[key] = PolicyFeature(
-                    type=FeatureType(ft_dict["type"]), shape=tuple(ft_dict["shape"])
-                )
-            self.features = reconstructed_features
-
-        if self.norm_map and isinstance(list(self.norm_map.keys())[0], str):
-            # norm_map came from JSON - need to reconstruct enum keys and values
-            reconstructed_norm_map = {}
-            for ft_type_str, norm_mode_str in self.norm_map.items():
-                reconstructed_norm_map[FeatureType(ft_type_str)] = NormalizationMode(norm_mode_str)
-            self.norm_map = reconstructed_norm_map
-
-        # Convert statistics once so we avoid repeated numpy→Tensor conversions
-        # during runtime.
-        self.stats = self.stats or {}
-        self._tensor_stats = _convert_stats_to_tensors(self.stats)
-
-        # Ensure *normalize_keys* is a set for fast look-ups and compare by
-        # value later when returning the configuration.
-        if self.normalize_keys is not None and not isinstance(self.normalize_keys, set):
-            self.normalize_keys = set(self.normalize_keys)
-
-    def _normalize_obs(self, observation):
-        if observation is None:
-            return None
-
-        # Decide which keys should be normalised for this call.
-        if self.normalize_keys is not None:
-            keys_to_norm = self.normalize_keys
-        else:
-            # Use feature map to skip action keys.
-            keys_to_norm = {k for k, ft in self.features.items() if ft.type is not FeatureType.ACTION}
-
-        processed = dict(observation)
-        for key in keys_to_norm:
-            if key not in processed or key not in self._tensor_stats:
-                continue
-
-            orig_val = processed[key]
-            tensor = (
-                orig_val.to(dtype=torch.float32)
-                if isinstance(orig_val, torch.Tensor)
-                else torch.as_tensor(orig_val, dtype=torch.float32)
-            )
-            stats = {k: v.to(tensor.device) for k, v in self._tensor_stats[key].items()}
-
-            if "mean" in stats and "std" in stats:
-                mean, std = stats["mean"], stats["std"]
-                processed[key] = (tensor - mean) / (std + self.eps)
-            elif "min" in stats and "max" in stats:
-                min_val, max_val = stats["min"], stats["max"]
-                processed[key] = 2 * (tensor - min_val) / (max_val - min_val + self.eps) - 1
-        return processed
-
-    def _normalize_action(self, action):
-        if action is None or "action" not in self._tensor_stats:
-            return action
-
-        tensor = (
-            action.to(dtype=torch.float32)
-            if isinstance(action, torch.Tensor)
-            else torch.as_tensor(action, dtype=torch.float32)
-        )
-        stats = {k: v.to(tensor.device) for k, v in self._tensor_stats["action"].items()}
-        if "mean" in stats and "std" in stats:
-            mean, std = stats["mean"], stats["std"]
-            return (tensor - mean) / (std + self.eps)
-        if "min" in stats and "max" in stats:
-            min_val, max_val = stats["min"], stats["max"]
-            return 2 * (tensor - min_val) / (max_val - min_val + self.eps) - 1
-        raise ValueError("Action stats must contain either ('mean','std') or ('min','max')")
-
     def __call__(self, transition: EnvTransition) -> EnvTransition:
-        observation = self._normalize_obs(transition.get(TransitionKey.OBSERVATION))
-        action = self._normalize_action(transition.get(TransitionKey.ACTION))
-
-        # Create a new transition with normalized values
         new_transition = transition.copy()
-        new_transition[TransitionKey.OBSERVATION] = observation
-        new_transition[TransitionKey.ACTION] = action
+
+        # Handle observation normalization.
+        observation = new_transition.get(TransitionKey.OBSERVATION)
+        if observation is not None:
+            new_transition[TransitionKey.OBSERVATION] = self._normalize_observation(
+                observation, inverse=False
+            )
+
+        # Handle action normalization.
+        action = new_transition.get(TransitionKey.ACTION)
+        if action is not None:
+            new_transition[TransitionKey.ACTION] = self._normalize_action(action, inverse=False)
+
         return new_transition
 
-    def get_config(self) -> dict[str, Any]:
-        config = {
-            "eps": self.eps,
-            "features": {
-                key: {"type": ft.type.value, "shape": ft.shape} for key, ft in self.features.items()
-            },
-            "norm_map": {ft_type.value: norm_mode.value for ft_type, norm_mode in self.norm_map.items()},
-        }
-        if self.normalize_keys is not None:
-            # Serialise as a list for YAML / JSON friendliness
-            config["normalize_keys"] = sorted(self.normalize_keys)
-        return config
-
-    def state_dict(self) -> dict[str, Tensor]:
-        flat = {}
-        for key, sub in self._tensor_stats.items():
-            for stat_name, tensor in sub.items():
-                flat[f"{key}.{stat_name}"] = tensor
-        return flat
-
-    def load_state_dict(self, state: Mapping[str, Tensor]) -> None:
-        self._tensor_stats.clear()
-        for flat_key, tensor in state.items():
-            key, stat_name = flat_key.rsplit(".", 1)
-            self._tensor_stats.setdefault(key, {})[stat_name] = tensor
-
-    def reset(self):
-        pass
-
-    def feature_contract(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
         return features
 
 
 @dataclass
 @ProcessorStepRegistry.register(name="unnormalizer_processor")
-class UnnormalizerProcessor:
-    """Inverse normalisation for observations and actions.
-
-    Exactly mirrors :class:`NormalizerProcessor` but applies the inverse
-    transform.
+class UnnormalizerProcessorStep(_NormalizationMixin, ProcessorStep):
     """
+    A processor step that applies unnormalization to observations and actions.
 
-    features: dict[str, PolicyFeature]
-    norm_map: dict[FeatureType, NormalizationMode]
-    stats: dict[str, dict[str, Any]] | None = None
-
-    _tensor_stats: dict[str, dict[str, Tensor]] = field(default_factory=dict, init=False, repr=False)
+    This class inverts the normalization process, scaling data back to its original
+    range. It is typically used in the post-processing pipeline to convert a policy's
+    normalized action output into a format that can be executed by a robot or
+    environment.
+    """
 
     @classmethod
     def from_lerobot_dataset(
@@ -229,103 +374,67 @@ class UnnormalizerProcessor:
         dataset: LeRobotDataset,
         features: dict[str, PolicyFeature],
         norm_map: dict[FeatureType, NormalizationMode],
-    ) -> UnnormalizerProcessor:
-        return cls(features=features, norm_map=norm_map, stats=dataset.meta.stats)
+        *,
+        device: torch.device | str | None = None,
+    ) -> UnnormalizerProcessorStep:
+        """
+        Creates an `UnnormalizerProcessorStep` using statistics from a `LeRobotDataset`.
 
-    def __post_init__(self):
-        # Handle deserialization from JSON config
-        if self.features and isinstance(list(self.features.values())[0], dict):
-            # Features came from JSON - need to reconstruct PolicyFeature objects
-            reconstructed_features = {}
-            for key, ft_dict in self.features.items():
-                reconstructed_features[key] = PolicyFeature(
-                    type=FeatureType(ft_dict["type"]), shape=tuple(ft_dict["shape"])
-                )
-            self.features = reconstructed_features
+        Args:
+            dataset: The dataset from which to extract normalization statistics.
+            features: The feature definition for the processor.
+            norm_map: The mapping from feature types to normalization modes.
+            device: The target device for the processor.
 
-        if self.norm_map and isinstance(list(self.norm_map.keys())[0], str):
-            # norm_map came from JSON - need to reconstruct enum keys and values
-            reconstructed_norm_map = {}
-            for ft_type_str, norm_mode_str in self.norm_map.items():
-                reconstructed_norm_map[FeatureType(ft_type_str)] = NormalizationMode(norm_mode_str)
-            self.norm_map = reconstructed_norm_map
-
-        self.stats = self.stats or {}
-        self._tensor_stats = _convert_stats_to_tensors(self.stats)
-
-    def _unnormalize_obs(self, observation):
-        if observation is None:
-            return None
-        keys = [k for k, ft in self.features.items() if ft.type is not FeatureType.ACTION]
-        processed = dict(observation)
-        for key in keys:
-            if key not in processed or key not in self._tensor_stats:
-                continue
-            orig_val = processed[key]
-            tensor = (
-                orig_val.to(dtype=torch.float32)
-                if isinstance(orig_val, torch.Tensor)
-                else torch.as_tensor(orig_val, dtype=torch.float32)
-            )
-            stats = {k: v.to(tensor.device) for k, v in self._tensor_stats[key].items()}
-            if "mean" in stats and "std" in stats:
-                mean, std = stats["mean"], stats["std"]
-                processed[key] = tensor * std + mean
-            elif "min" in stats and "max" in stats:
-                min_val, max_val = stats["min"], stats["max"]
-                processed[key] = (tensor + 1) / 2 * (max_val - min_val) + min_val
-        return processed
-
-    def _unnormalize_action(self, action):
-        if action is None or "action" not in self._tensor_stats:
-            return action
-        tensor = (
-            action.to(dtype=torch.float32)
-            if isinstance(action, torch.Tensor)
-            else torch.as_tensor(action, dtype=torch.float32)
-        )
-        stats = {k: v.to(tensor.device) for k, v in self._tensor_stats["action"].items()}
-        if "mean" in stats and "std" in stats:
-            mean, std = stats["mean"], stats["std"]
-            return tensor * std + mean
-        if "min" in stats and "max" in stats:
-            min_val, max_val = stats["min"], stats["max"]
-            return (tensor + 1) / 2 * (max_val - min_val) + min_val
-        raise ValueError("Action stats must contain either ('mean','std') or ('min','max')")
+        Returns:
+            A new instance of `UnnormalizerProcessorStep`.
+        """
+        return cls(features=features, norm_map=norm_map, stats=dataset.meta.stats, device=device)
 
     def __call__(self, transition: EnvTransition) -> EnvTransition:
-        observation = self._unnormalize_obs(transition.get(TransitionKey.OBSERVATION))
-        action = self._unnormalize_action(transition.get(TransitionKey.ACTION))
-
-        # Create a new transition with unnormalized values
         new_transition = transition.copy()
-        new_transition[TransitionKey.OBSERVATION] = observation
-        new_transition[TransitionKey.ACTION] = action
+
+        # Handle observation unnormalization.
+        observation = new_transition.get(TransitionKey.OBSERVATION)
+        if observation is not None:
+            new_transition[TransitionKey.OBSERVATION] = self._normalize_observation(observation, inverse=True)
+
+        # Handle action unnormalization.
+        action = new_transition.get(TransitionKey.ACTION)
+        if action is not None:
+            new_transition[TransitionKey.ACTION] = self._normalize_action(action, inverse=True)
+
         return new_transition
 
-    def get_config(self) -> dict[str, Any]:
-        return {
-            "features": {
-                key: {"type": ft.type.value, "shape": ft.shape} for key, ft in self.features.items()
-            },
-            "norm_map": {ft_type.value: norm_mode.value for ft_type, norm_mode in self.norm_map.items()},
-        }
-
-    def state_dict(self) -> dict[str, Tensor]:
-        flat = {}
-        for key, sub in self._tensor_stats.items():
-            for stat_name, tensor in sub.items():
-                flat[f"{key}.{stat_name}"] = tensor
-        return flat
-
-    def load_state_dict(self, state: Mapping[str, Tensor]) -> None:
-        self._tensor_stats.clear()
-        for flat_key, tensor in state.items():
-            key, stat_name = flat_key.rsplit(".", 1)
-            self._tensor_stats.setdefault(key, {})[stat_name] = tensor
-
-    def reset(self):
-        pass
-
-    def feature_contract(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
         return features
+
+
+def hotswap_stats(
+    policy_processor: PolicyProcessorPipeline, stats: dict[str, dict[str, Any]]
+) -> PolicyProcessorPipeline:
+    """
+    Replaces normalization statistics in an existing `PolicyProcessorPipeline` instance.
+
+    This function creates a deep copy of the provided pipeline and updates the
+    statistics of any `NormalizerProcessorStep` or `UnnormalizerProcessorStep` it
+    contains. This is useful for adapting a trained policy to a new environment or
+    dataset with different data distributions without having to reconstruct the entire
+    pipeline.
+
+    Args:
+        policy_processor: The policy processor pipeline to modify.
+        stats: The new dictionary of normalization statistics to apply.
+
+    Returns:
+        A new `PolicyProcessorPipeline` instance with the updated statistics.
+    """
+    rp = deepcopy(policy_processor)
+    for step in rp.steps:
+        if isinstance(step, _NormalizationMixin):
+            step.stats = stats
+            # Re-initialize tensor_stats on the correct device.
+            step._tensor_stats = to_tensor(stats, device=step.device, dtype=step.dtype)
+    return rp

src/lerobot/processor/observation_processor.py

@@ -20,32 +20,54 @@ import numpy as np
 import torch
 from torch import Tensor
 
-from lerobot.configs.types import PolicyFeature
+from lerobot.configs.types import PipelineFeatureType, PolicyFeature
 from lerobot.constants import OBS_ENV_STATE, OBS_IMAGE, OBS_IMAGES, OBS_STATE
-from lerobot.processor.pipeline import ObservationProcessor, ProcessorStepRegistry
+
+from .pipeline import ObservationProcessorStep, ProcessorStepRegistry
 
 
 @dataclass
 @ProcessorStepRegistry.register(name="observation_processor")
-class VanillaObservationProcessor(ObservationProcessor):
+class VanillaObservationProcessorStep(ObservationProcessorStep):
     """
-    Processes environment observations into the LeRobot format by handling both images and states.
+    Processes standard Gymnasium observations into the LeRobot format.
 
-    Image processing:
-        - Converts channel-last (H, W, C) images to channel-first (C, H, W)
-        - Normalizes uint8 images ([0, 255]) to float32 ([0, 1])
-        - Adds a batch dimension if missing
-        - Supports single images and image dictionaries
+    This step handles both image and state data from a typical observation dictionary,
+    preparing it for use in a LeRobot policy.
 
-    State processing:
-        - Maps 'environment_state' to observation.environment_state
-        - Maps 'agent_pos' to observation.state
-        - Converts numpy arrays to tensors
-        - Adds a batch dimension if missing
+    **Image Processing:**
+    -   Converts channel-last (H, W, C), `uint8` images to channel-first (C, H, W),
+        `float32` tensors.
+    -   Normalizes pixel values from the [0, 255] range to [0, 1].
+    -   Adds a batch dimension if one is not already present.
+    -   Recognizes a single image under the key `"pixels"` and maps it to
+        `"observation.image"`.
+    -   Recognizes a dictionary of images under the key `"pixels"` and maps them
+        to `"observation.images.{camera_name}"`.
+
+    **State Processing:**
+    -   Maps the `"environment_state"` key to `"observation.environment_state"`.
+    -   Maps the `"agent_pos"` key to `"observation.state"`.
+    -   Converts NumPy arrays to PyTorch tensors.
+    -   Adds a batch dimension if one is not already present.
     """
 
     def _process_single_image(self, img: np.ndarray) -> Tensor:
-        """Process a single image array."""
+        """
+        Processes a single NumPy image array into a channel-first, normalized tensor.
+
+        Args:
+            img: A NumPy array representing the image, expected to be in channel-last
+                 (H, W, C) format with a `uint8` dtype.
+
+        Returns:
+            A `float32` PyTorch tensor in channel-first (B, C, H, W) format, with
+            pixel values normalized to the [0, 1] range.
+
+        Raises:
+            ValueError: If the input image does not appear to be in channel-last
+                        format or is not of `uint8` dtype.
+        """
         # Convert to tensor
         img_tensor = torch.from_numpy(img)
 
@@ -106,19 +128,32 @@ class VanillaObservationProcessor(ObservationProcessor):
     def observation(self, observation):
         return self._process_observation(observation)
 
-    def feature_contract(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
-        """Transforms feature keys to a standardized contract.
-
-        This method handles several renaming patterns:
-        - Exact matches (e.g., 'pixels' -> 'OBS_IMAGE').
-        - Prefixed exact matches (e.g., 'observation.pixels' -> 'OBS_IMAGE').
-        - Prefix matches (e.g., 'pixels.cam1' -> 'OBS_IMAGES.cam1').
-        - Prefixed prefix matches (e.g., 'observation.pixels.cam1' -> 'OBS_IMAGES.cam1').
-        - environment_state -> OBS_ENV_STATE,
-        - agent_pos -> OBS_STATE,
-        - observation.environment_state -> OBS_ENV_STATE,
-        - observation.agent_pos -> OBS_STATE
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
         """
+        Transforms feature keys from the Gym standard to the LeRobot standard.
+
+        This method standardizes the feature dictionary by renaming keys according
+        to LeRobot's conventions, ensuring that policies can be constructed correctly.
+        It handles various raw key formats, including those with an "observation." prefix.
+
+        **Renaming Rules:**
+        - `pixels` or `observation.pixels` -> `observation.image`
+        - `pixels.{cam}` or `observation.pixels.{cam}` -> `observation.images.{cam}`
+        - `environment_state` or `observation.environment_state` -> `observation.environment_state`
+        - `agent_pos` or `observation.agent_pos` -> `observation.state`
+
+        Args:
+            features: The policy features dictionary with Gym-style keys.
+
+        Returns:
+            The policy features dictionary with standardized LeRobot keys.
+        """
+        # Build a new features mapping keyed by the same FeatureType buckets
+        # We assume callers already placed features in the correct FeatureType.
+        new_features: dict[PipelineFeatureType, dict[str, PolicyFeature]] = {ft: {} for ft in features.keys()}
+
         exact_pairs = {
             "pixels": OBS_IMAGE,
             "environment_state": OBS_ENV_STATE,
@@ -129,29 +164,43 @@ class VanillaObservationProcessor(ObservationProcessor):
             "pixels.": f"{OBS_IMAGES}.",
         }
 
-        for key in list(features.keys()):
-            matched_prefix = False
-            for old_prefix, new_prefix in prefix_pairs.items():
-                prefixed_old = f"observation.{old_prefix}"
-                if key.startswith(prefixed_old):
-                    suffix = key[len(prefixed_old) :]
-                    features[f"{new_prefix}{suffix}"] = features.pop(key)
-                    matched_prefix = True
-                    break
+        # Iterate over all incoming feature buckets and normalize/move each entry
+        for src_ft, bucket in features.items():
+            for key, feat in list(bucket.items()):
+                handled = False
 
-                if key.startswith(old_prefix):
-                    suffix = key[len(old_prefix) :]
-                    features[f"{new_prefix}{suffix}"] = features.pop(key)
-                    matched_prefix = True
-                    break
-
-            if matched_prefix:
-                continue
-
-            for old, new in exact_pairs.items():
-                if key == old or key == f"observation.{old}":
-                    if key in features:
-                        features[new] = features.pop(key)
+                # Prefix-based rules (e.g. pixels.cam1 -> OBS_IMAGES.cam1)
+                for old_prefix, new_prefix in prefix_pairs.items():
+                    prefixed_old = f"observation.{old_prefix}"
+                    if key.startswith(prefixed_old):
+                        suffix = key[len(prefixed_old) :]
+                        new_key = f"{new_prefix}{suffix}"
+                        new_features[src_ft][new_key] = feat
+                        handled = True
                         break
 
-        return features
+                    if key.startswith(old_prefix):
+                        suffix = key[len(old_prefix) :]
+                        new_key = f"{new_prefix}{suffix}"
+                        new_features[src_ft][new_key] = feat
+                        handled = True
+                        break
+
+                if handled:
+                    continue
+
+                # Exact-name rules (pixels, environment_state, agent_pos)
+                for old, new in exact_pairs.items():
+                    if key == old or key == f"observation.{old}":
+                        new_key = new
+                        new_features[src_ft][new_key] = feat
+                        handled = True
+                        break
+
+                if handled:
+                    continue
+
+                # Default: keep key in the same source FeatureType bucket
+                new_features[src_ft][key] = feat
+
+        return new_features

src/lerobot/processor/rename_processor.py

@@ -13,20 +13,30 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+from copy import deepcopy
 from dataclasses import dataclass, field
 from typing import Any
 
-from lerobot.configs.types import PolicyFeature
-from lerobot.processor.pipeline import (
-    ObservationProcessor,
-    ProcessorStepRegistry,
-)
+from lerobot.configs.types import PipelineFeatureType, PolicyFeature
+
+from .pipeline import ObservationProcessorStep, ProcessorStepRegistry
 
 
 @dataclass
-@ProcessorStepRegistry.register(name="rename_processor")
-class RenameProcessor(ObservationProcessor):
-    """Rename processor that renames keys in the observation."""
+@ProcessorStepRegistry.register(name="rename_observations_processor")
+class RenameObservationsProcessorStep(ObservationProcessorStep):
+    """
+    A processor step that renames keys in an observation dictionary.
+
+    This step is useful for creating a standardized data interface by mapping keys
+    from an environment's format to the format expected by a LeRobot policy or
+    other downstream components.
+
+    Attributes:
+        rename_map: A dictionary mapping from old key names to new key names.
+                    Keys present in an observation that are not in this map will
+                    be kept with their original names.
+    """
 
     rename_map: dict[str, str] = field(default_factory=dict)
 
@@ -43,9 +53,41 @@ class RenameProcessor(ObservationProcessor):
     def get_config(self) -> dict[str, Any]:
         return {"rename_map": self.rename_map}
 
-    def feature_contract(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
         """Transforms:
         - Each key in the observation that appears in `rename_map` is renamed to its value.
         - Keys not in `rename_map` remain unchanged.
         """
-        return {self.rename_map.get(k, k): v for k, v in features.items()}
+        new_features: dict[PipelineFeatureType, dict[str, PolicyFeature]] = features.copy()
+        new_features[PipelineFeatureType.OBSERVATION] = {
+            self.rename_map.get(k, k): v for k, v in features[PipelineFeatureType.OBSERVATION].items()
+        }
+        return new_features
+
+
+def rename_stats(stats: dict[str, dict[str, Any]], rename_map: dict[str, str]) -> dict[str, dict[str, Any]]:
+    """
+    Renames the top-level keys in a statistics dictionary using a provided mapping.
+
+    This is a helper function typically used to keep normalization statistics
+    consistent with renamed observation or action features. It performs a defensive
+    deep copy to avoid modifying the original `stats` dictionary.
+
+    Args:
+        stats: A nested dictionary of statistics, where top-level keys are
+               feature names (e.g., `{"observation.state": {"mean": 0.5}}`).
+        rename_map: A dictionary mapping old feature names to new feature names.
+
+    Returns:
+        A new statistics dictionary with its top-level keys renamed. Returns an
+        empty dictionary if the input `stats` is empty.
+    """
+    if not stats:
+        return {}
+    renamed: dict[str, dict[str, Any]] = {}
+    for old_key, sub_stats in stats.items():
+        new_key = rename_map.get(old_key, old_key)
+        renamed[new_key] = deepcopy(sub_stats) if sub_stats is not None else {}
+    return renamed

src/lerobot/processor/tokenizer_processor.py

@@ -0,0 +1,273 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This script defines a processor for tokenizing natural language instructions from an environment transition.
+
+It uses a tokenizer from the Hugging Face `transformers` library to convert task descriptions (text) into
+token IDs and attention masks, which are then added to the observation dictionary.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import TYPE_CHECKING, Any
+
+import torch
+
+from lerobot.configs.types import FeatureType, PipelineFeatureType, PolicyFeature
+from lerobot.constants import OBS_LANGUAGE_ATTENTION_MASK, OBS_LANGUAGE_TOKENS
+from lerobot.utils.import_utils import _transformers_available
+
+from .core import EnvTransition, TransitionKey
+from .pipeline import ObservationProcessorStep, ProcessorStepRegistry
+
+# Conditional import for type checking and lazy loading
+if TYPE_CHECKING or _transformers_available:
+    from transformers import AutoTokenizer
+else:
+    AutoTokenizer = None
+
+
+@dataclass
+@ProcessorStepRegistry.register(name="tokenizer_processor")
+class TokenizerProcessorStep(ObservationProcessorStep):
+    """
+    Processor step to tokenize a natural language task description.
+
+    This step extracts a task string from the `complementary_data` of an `EnvTransition`,
+    tokenizes it using a Hugging Face `transformers` tokenizer, and adds the resulting
+    token IDs and attention mask to the `observation` dictionary.
+
+    Requires the `transformers` library to be installed.
+
+    Attributes:
+        tokenizer_name: The name of a pretrained tokenizer from the Hugging Face Hub (e.g., "bert-base-uncased").
+        tokenizer: A pre-initialized tokenizer object. If provided, `tokenizer_name` is ignored.
+        max_length: The maximum length to pad or truncate sequences to.
+        task_key: The key in `complementary_data` where the task string is stored.
+        padding_side: The side to pad on ('left' or 'right').
+        padding: The padding strategy ('max_length', 'longest', etc.).
+        truncation: Whether to truncate sequences longer than `max_length`.
+        input_tokenizer: The internal tokenizer instance, loaded during initialization.
+    """
+
+    tokenizer_name: str | None = None
+    tokenizer: Any | None = None  # Use `Any` for compatibility without a hard dependency
+    max_length: int = 512
+    task_key: str = "task"
+    padding_side: str = "right"
+    padding: str = "max_length"
+    truncation: bool = True
+
+    # Internal tokenizer instance (not part of the config)
+    input_tokenizer: Any = field(default=None, init=False, repr=False)
+
+    def __post_init__(self):
+        """
+        Initializes the tokenizer after the dataclass is created.
+
+        It checks for the availability of the `transformers` library and loads the tokenizer
+        either from a provided object or by name from the Hugging Face Hub.
+
+        Raises:
+            ImportError: If the `transformers` library is not installed.
+            ValueError: If neither `tokenizer` nor `tokenizer_name` is provided.
+        """
+        if not _transformers_available:
+            raise ImportError(
+                "The 'transformers' library is not installed. "
+                "Please install it with `pip install 'lerobot[transformers-dep]'` to use TokenizerProcessorStep."
+            )
+
+        if self.tokenizer is not None:
+            # Use provided tokenizer object directly
+            self.input_tokenizer = self.tokenizer
+        elif self.tokenizer_name is not None:
+            if AutoTokenizer is None:
+                raise ImportError("AutoTokenizer is not available")
+            self.input_tokenizer = AutoTokenizer.from_pretrained(self.tokenizer_name)
+        else:
+            raise ValueError(
+                "Either 'tokenizer' or 'tokenizer_name' must be provided. "
+                "Pass a tokenizer object directly or a tokenizer name to auto-load."
+            )
+
+    def get_task(self, transition: EnvTransition) -> list[str] | None:
+        """
+        Extracts the task description(s) from the transition's complementary data.
+
+        Args:
+            transition: The environment transition.
+
+        Returns:
+            A list of task strings, or None if the task key is not found or the value is None.
+        """
+        complementary_data = transition.get(TransitionKey.COMPLEMENTARY_DATA)
+        if complementary_data is None:
+            return None
+
+        if self.task_key not in complementary_data:
+            return None
+
+        task = complementary_data[self.task_key]
+        if task is None:
+            return None
+
+        # Standardize to a list of strings for the tokenizer
+        if isinstance(task, str):
+            return [task]
+        elif isinstance(task, list) and all(isinstance(t, str) for t in task):
+            return task
+
+        return None
+
+    def observation(self, observation: dict[str, Any]) -> dict[str, Any]:
+        """
+        Tokenizes the task description and adds it to the observation dictionary.
+
+        This method retrieves the task, tokenizes it, moves the resulting tensors to the
+        same device as other data in the transition, and updates the observation.
+
+        Args:
+            observation: The original observation dictionary.
+
+        Returns:
+            The updated observation dictionary including token IDs and an attention mask.
+        """
+        task = self.get_task(self.transition)
+        if task is None:
+            return observation
+
+        # Tokenize the task (this will create CPU tensors)
+        tokenized_prompt = self._tokenize_text(task)
+
+        # Detect the device from existing tensors in the transition to ensure consistency
+        target_device = self._detect_device(self.transition)
+
+        # Move new tokenized tensors to the detected device
+        if target_device is not None:
+            tokenized_prompt = {
+                k: v.to(target_device) if isinstance(v, torch.Tensor) else v
+                for k, v in tokenized_prompt.items()
+            }
+
+        # Create a new observation dict to avoid modifying the original in place
+        new_observation = dict(observation)
+
+        # Add tokenized data to the observation
+        new_observation[OBS_LANGUAGE_TOKENS] = tokenized_prompt["input_ids"]
+        new_observation[OBS_LANGUAGE_ATTENTION_MASK] = tokenized_prompt["attention_mask"].to(dtype=torch.bool)
+
+        return new_observation
+
+    def _detect_device(self, transition: EnvTransition) -> torch.device | None:
+        """
+        Detects the torch.device from existing tensors in the transition.
+
+        It checks tensors in the observation dictionary first, then the action tensor.
+
+        Args:
+            transition: The environment transition.
+
+        Returns:
+            The detected `torch.device`, or None if no tensors are found.
+        """
+        # Check observation tensors first (most likely place to find tensors)
+        observation = transition.get(TransitionKey.OBSERVATION)
+        if observation:
+            for value in observation.values():
+                if isinstance(value, torch.Tensor):
+                    return value.device
+
+        # Fallback to checking the action tensor
+        action = transition.get(TransitionKey.ACTION)
+        if isinstance(action, torch.Tensor):
+            return action.device
+
+        return None  # No tensors found, default will be CPU
+
+    def _tokenize_text(self, text: str | list[str]) -> dict[str, torch.Tensor]:
+        """
+        A wrapper around the tokenizer call.
+
+        Args:
+            text: A string or list of strings to tokenize.
+
+        Returns:
+            A dictionary containing tokenized 'input_ids' and 'attention_mask' as PyTorch tensors.
+        """
+        return self.input_tokenizer(
+            text,
+            max_length=self.max_length,
+            truncation=self.truncation,
+            padding=self.padding,
+            padding_side=self.padding_side,
+            return_tensors="pt",
+        )
+
+    def get_config(self) -> dict[str, Any]:
+        """
+        Returns the serializable configuration of the processor.
+
+        Note: The tokenizer object itself is not serialized. If the processor was initialized
+        with a tokenizer name, that name will be included in the config.
+
+        Returns:
+            A dictionary with the processor's configuration parameters.
+        """
+        config = {
+            "max_length": self.max_length,
+            "task_key": self.task_key,
+            "padding_side": self.padding_side,
+            "padding": self.padding,
+            "truncation": self.truncation,
+        }
+
+        # Only save tokenizer_name if it was used to create the tokenizer
+        if self.tokenizer_name is not None and self.tokenizer is None:
+            config["tokenizer_name"] = self.tokenizer_name
+
+        return config
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        """
+        Adds feature definitions for the language tokens and attention mask.
+
+        This updates the policy features dictionary to include the new data added to the
+        observation, ensuring downstream components are aware of their shape and type.
+
+        Args:
+            features: The dictionary of existing policy features.
+
+        Returns:
+            The updated dictionary of policy features.
+        """
+        # Add a feature for the token IDs if it doesn't already exist
+        if OBS_LANGUAGE_TOKENS not in features[PipelineFeatureType.OBSERVATION]:
+            features[PipelineFeatureType.OBSERVATION][OBS_LANGUAGE_TOKENS] = PolicyFeature(
+                type=FeatureType.LANGUAGE, shape=(self.max_length,)
+            )
+
+        # Add a feature for the attention mask if it doesn't already exist
+        if OBS_LANGUAGE_ATTENTION_MASK not in features[PipelineFeatureType.OBSERVATION]:
+            features[PipelineFeatureType.OBSERVATION][OBS_LANGUAGE_ATTENTION_MASK] = PolicyFeature(
+                type=FeatureType.LANGUAGE, shape=(self.max_length,)
+            )
+
+        return features

src/lerobot/record.py

@@ -18,7 +18,7 @@ Records a dataset. Actions for the robot can be either generated by teleoperatio
 Example:
 
 ```shell
-python -m lerobot.record \
+lerobot-record \
     --robot.type=so100_follower \
     --robot.port=/dev/tty.usbmodem58760431541 \
     --robot.cameras="{laptop: {type: opencv, camera_index: 0, width: 640, height: 480}}" \
@@ -36,7 +36,7 @@ python -m lerobot.record \
 
 Example recording with bimanual so100:
 ```shell
-python -m lerobot.record \
+lerobot-record \
   --robot.type=bi_so100_follower \
   --robot.left_arm_port=/dev/tty.usbmodem5A460851411 \
   --robot.right_arm_port=/dev/tty.usbmodem5A460812391 \
@@ -59,9 +59,10 @@ python -m lerobot.record \
 
 import logging
 import time
-from dataclasses import asdict, dataclass
+from dataclasses import asdict, dataclass, field
 from pathlib import Path
 from pprint import pformat
+from typing import Any
 
 from lerobot.cameras import (  # noqa: F401
     CameraConfig,  # noqa: F401
@@ -72,10 +73,25 @@ from lerobot.configs import parser
 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.datasets.image_writer import safe_stop_image_writer
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.datasets.utils import build_dataset_frame, hw_to_dataset_features
+from lerobot.datasets.utils import hw_to_dataset_features
 from lerobot.datasets.video_utils import VideoEncodingManager
-from lerobot.policies.factory import make_policy
+from lerobot.policies.factory import make_policy, make_pre_post_processors
 from lerobot.policies.pretrained import PreTrainedPolicy
+from lerobot.processor import (
+    EnvTransition,
+    IdentityProcessorStep,
+    PolicyProcessorPipeline,
+    RobotProcessorPipeline,
+    TransitionKey,
+)
+from lerobot.processor.converters import (
+    action_to_transition,
+    identity_transition,
+    observation_to_transition,
+    transition_to_action,
+    transition_to_dataset_frame,
+)
+from lerobot.processor.rename_processor import rename_stats
 from lerobot.robots import (  # noqa: F401
     Robot,
     RobotConfig,
@@ -149,6 +165,8 @@ class DatasetRecordConfig:
     # Number of episodes to record before batch encoding videos
     # Set to 1 for immediate encoding (default behavior), or higher for batched encoding
     video_encoding_batch_size: int = 1
+    # Rename map for the observation to override the image and state keys
+    rename_map: dict[str, str] = field(default_factory=dict)
 
     def __post_init__(self):
         if self.single_task is None:
@@ -187,6 +205,36 @@ class RecordConfig:
         return ["policy"]
 
 
+""" --------------- record_loop() data flow --------------------------
+       [ Robot ]
+           V
+     [ robot.get_observation() ] ---> raw_obs
+           V
+     [ robot_observation_processor ] ---> obs_transition
+           V
+     .-----( ACTION LOGIC )------------------.
+     V                                       V
+     [ From Teleoperator ]                   [ From Policy ]
+     |                                       |
+     |  [teleop.get_action] -> raw_action    |   [predict_action]
+     |          |                            |          |
+     |          V                            |          V
+     | [teleop_action_processor]             |          |
+     |          |                            |          |
+     '---> teleop_transition                 '---> policy_transition
+     |                                       |
+     '-------------------------.-------------'
+                               V
+                  [ robot_action_processor ] --> robot_action_to_send
+                               V
+                    [ robot.send_action() ] -- (Robot Executes)
+                               V
+        ( Transitions are merged & added to Dataset )
+                               V
+                  ( Rerun Log / Loop Wait )
+"""
+
+
 @safe_stop_image_writer
 def record_loop(
     robot: Robot,
@@ -195,21 +243,56 @@ def record_loop(
     dataset: LeRobotDataset | None = None,
     teleop: Teleoperator | list[Teleoperator] | None = None,
     policy: PreTrainedPolicy | None = None,
+    preprocessor: PolicyProcessorPipeline | None = None,
+    postprocessor: PolicyProcessorPipeline | None = None,
     control_time_s: int | None = None,
+    teleop_action_processor: RobotProcessorPipeline[EnvTransition] | None = None,  # runs after teleop
+    robot_action_processor: RobotProcessorPipeline[dict[str, Any]] | None = None,  # runs before robot
+    robot_observation_processor: RobotProcessorPipeline[EnvTransition] | None = None,  # runs after robot
     single_task: str | None = None,
     display_data: bool = False,
 ):
+    teleop_action_processor: RobotProcessorPipeline[EnvTransition] = (
+        teleop_action_processor
+        or RobotProcessorPipeline(
+            steps=[IdentityProcessorStep()], to_transition=action_to_transition, to_output=identity_transition
+        )
+    )
+    robot_action_processor: RobotProcessorPipeline[dict[str, Any]] = (
+        robot_action_processor
+        or RobotProcessorPipeline(
+            steps=[IdentityProcessorStep()],
+            to_transition=identity_transition,
+            to_output=transition_to_action,
+        )
+    )
+    robot_observation_processor: RobotProcessorPipeline[EnvTransition] = (
+        robot_observation_processor
+        or RobotProcessorPipeline(
+            steps=[IdentityProcessorStep()],
+            to_transition=observation_to_transition,
+            to_output=identity_transition,
+        )
+    )
+
     if dataset is not None and dataset.fps != fps:
         raise ValueError(f"The dataset fps should be equal to requested fps ({dataset.fps} != {fps}).")
 
     teleop_arm = teleop_keyboard = None
-    if isinstance(teleop, list):
+    if isinstance(teleop, list):  # For LeKiwi
         teleop_keyboard = next((t for t in teleop if isinstance(t, KeyboardTeleop)), None)
         teleop_arm = next(
             (
                 t
                 for t in teleop
-                if isinstance(t, (so100_leader.SO100Leader, so101_leader.SO101Leader, koch_leader.KochLeader))
+                if isinstance(
+                    t,
+                    (
+                        so100_leader.SO100Leader,
+                        so101_leader.SO101Leader,
+                        koch_leader.KochLeader,
+                    ),
+                )
             ),
             None,
         )
@@ -219,9 +302,20 @@ def record_loop(
                 "For multi-teleop, the list must contain exactly one KeyboardTeleop and one arm teleoperator. Currently only supported for LeKiwi robot."
             )
 
-    # if policy is given it needs cleaning up
-    if policy is not None:
+    # Reset policy and processor if they are provided
+    if policy is not None and preprocessor is not None and postprocessor is not None:
         policy.reset()
+        preprocessor.reset()
+        postprocessor.reset()
+
+    # Reset custom pipelines
+    teleop_action_processor.reset()
+    robot_action_processor.reset()
+    robot_observation_processor.reset()
+
+    policy_transition = None
+    teleop_transition = None
+    obs_transition = None
 
     timestamp = 0
     start_episode_t = time.perf_counter()
@@ -232,51 +326,91 @@ def record_loop(
             events["exit_early"] = False
             break
 
-        observation = robot.get_observation()
+        # Get robot observation
+        obs = robot.get_observation()
 
-        if policy is not None or dataset is not None:
-            observation_frame = build_dataset_frame(dataset.features, observation, prefix="observation")
+        # Applies a pipeline to the raw robot observation, default is IdentityProcessor
+        obs_transition = robot_observation_processor(obs)
+
+        # Get action from either policy or teleop
+        if policy is not None and preprocessor is not None and postprocessor is not None:
+            if dataset is not None:
+                observation_frame = transition_to_dataset_frame(
+                    obs_transition, dataset.features
+                )  # Convert the observation to the dataset format
 
-        if policy is not None:
             action_values = predict_action(
-                observation_frame,
-                policy,
-                get_safe_torch_device(policy.config.device),
-                policy.config.use_amp,
+                observation=observation_frame,
+                policy=policy,
+                device=get_safe_torch_device(policy.config.device),
+                preprocessor=preprocessor,
+                postprocessor=postprocessor,
+                use_amp=policy.config.use_amp,
                 task=single_task,
                 robot_type=robot.robot_type,
             )
-            action = {key: action_values[i].item() for i, key in enumerate(robot.action_features)}
-        elif policy is None and isinstance(teleop, Teleoperator):
-            action = teleop.get_action()
-        elif policy is None and isinstance(teleop, list):
-            # TODO(pepijn, steven): clean the record loop for use of multiple robots (possibly with pipeline)
+
+            action_names = dataset.features["action"]["names"]
+            policy_action = {f"action.{name}": float(action_values[i]) for i, name in enumerate(action_names)}
+            policy_transition = {
+                TransitionKey.ACTION: policy_action,
+                TransitionKey.COMPLEMENTARY_DATA: {},
+            }
+
+        elif isinstance(teleop, Teleoperator):
+            act = teleop.get_action()
+
+            # Applies a pipeline to the raw teleop action, default is IdentityProcessor
+            # TODO(Steven): This assumes that the processor passed by the user should have identity_transition as to_output.
+            teleop_transition = teleop_action_processor(act)
+
+        elif isinstance(teleop, list):
             arm_action = teleop_arm.get_action()
             arm_action = {f"arm_{k}": v for k, v in arm_action.items()}
-
             keyboard_action = teleop_keyboard.get_action()
             base_action = robot._from_keyboard_to_base_action(keyboard_action)
-
-            action = {**arm_action, **base_action} if len(base_action) > 0 else arm_action
+            act = {**arm_action, **base_action} if len(base_action) > 0 else arm_action
+            teleop_transition = teleop_action_processor(act)
         else:
             logging.info(
-                "No policy or teleoperator provided, skipping action generation."
+                "No policy or teleoperator provided, skipping action generation. "
                 "This is likely to happen when resetting the environment without a teleop device."
                 "The robot won't be at its rest position at the start of the next episode."
             )
             continue
 
-        # Action can eventually be clipped using `max_relative_target`,
-        # so action actually sent is saved in the dataset.
-        sent_action = robot.send_action(action)
+        # Applies a pipeline to the action, default is IdentityProcessor
+        # IMPORTANT: action_pipeline.to_output must return a dict suitable for robot.send_action()
+        if policy is not None and policy_transition is not None:
+            robot_action_to_send = robot_action_processor(policy_transition)
+        else:
+            robot_action_to_send = robot_action_processor(teleop_transition)
 
+        # Send action to robot
+        # Action can eventually be clipped using `max_relative_target`,
+        # so action actually sent is saved in the dataset. action = postprocessor.process(action)
+        # TODO(pepijn, adil): we should use a pipeline step to clip the action, so the sent action is the action that we input to the robot.
+        _ = robot.send_action(robot_action_to_send)
+
+        # Write to dataset
         if dataset is not None:
-            action_frame = build_dataset_frame(dataset.features, sent_action, prefix="action")
-            frame = {**observation_frame, **action_frame}
+            # If transition_to_dataset_frame is provided, use it to merge the transitions.
+            merged = []
+            if obs_transition is not None:  # The observation from the robot
+                merged.append(obs_transition)
+            if teleop_transition is not None:  # The action from teleop
+                merged.append(teleop_transition)
+            if policy_transition is not None:  # The action from policy
+                merged.append(policy_transition)
+            frame = transition_to_dataset_frame(
+                merged if len(merged) > 1 else merged[0], dataset.features
+            )  # Convert the observation to the dataset format
             dataset.add_frame(frame, task=single_task)
 
         if display_data:
-            log_rerun_data(observation, action)
+            log_rerun_data(
+                observation=obs_transition.get(TransitionKey.OBSERVATION), action=robot_action_to_send
+            )
 
         dt_s = time.perf_counter() - start_loop_t
         busy_wait(1 / fps - dt_s)
@@ -296,7 +430,15 @@ def record(cfg: RecordConfig) -> LeRobotDataset:
 
     action_features = hw_to_dataset_features(robot.action_features, "action", cfg.dataset.video)
     obs_features = hw_to_dataset_features(robot.observation_features, "observation", cfg.dataset.video)
-    dataset_features = {**action_features, **obs_features}
+
+    # Add next.* features that are generated during recording
+    transition_features = {
+        "next.reward": {"dtype": "float32", "shape": (1,), "names": None},
+        "next.done": {"dtype": "bool", "shape": (1,), "names": None},
+        "next.truncated": {"dtype": "bool", "shape": (1,), "names": None},
+    }
+
+    dataset_features = {**action_features, **obs_features, **transition_features}
 
     if cfg.resume:
         dataset = LeRobotDataset(
@@ -328,6 +470,18 @@ def record(cfg: RecordConfig) -> LeRobotDataset:
 
     # Load pretrained policy
     policy = None if cfg.policy is None else make_policy(cfg.policy, ds_meta=dataset.meta)
+    preprocessor = None
+    postprocessor = None
+    if cfg.policy is not None:
+        preprocessor, postprocessor = make_pre_post_processors(
+            policy_cfg=cfg.policy,
+            pretrained_path=cfg.policy.pretrained_path,
+            dataset_stats=rename_stats(dataset.meta.stats, cfg.dataset.rename_map),
+            preprocessor_overrides={
+                "device_processor": {"device": cfg.policy.device},
+                "rename_processor": {"rename_map": cfg.dataset.rename_map},
+            },
+        )
 
     robot.connect()
     if teleop is not None:
@@ -345,6 +499,8 @@ def record(cfg: RecordConfig) -> LeRobotDataset:
                 fps=cfg.dataset.fps,
                 teleop=teleop,
                 policy=policy,
+                preprocessor=preprocessor,
+                postprocessor=postprocessor,
                 dataset=dataset,
                 control_time_s=cfg.dataset.episode_time_s,
                 single_task=cfg.dataset.single_task,

src/lerobot/replay.py

@@ -18,7 +18,7 @@ Replays the actions of an episode from a dataset on a robot.
 Examples:
 
 ```shell
-python -m lerobot.replay \
+lerobot-replay \
     --robot.type=so100_follower \
     --robot.port=/dev/tty.usbmodem58760431541 \
     --robot.id=black \
@@ -28,7 +28,7 @@ python -m lerobot.replay \
 
 Example replay with bimanual so100:
 ```shell
-python -m lerobot.replay \
+lerobot-replay \
   --robot.type=bi_so100_follower \
   --robot.left_arm_port=/dev/tty.usbmodem5A460851411 \
   --robot.right_arm_port=/dev/tty.usbmodem5A460812391 \
@@ -45,9 +45,10 @@ from dataclasses import asdict, dataclass
 from pathlib import Path
 from pprint import pformat
 
-import draccus
-
+from lerobot.configs import parser
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.processor import IdentityProcessorStep, RobotProcessorPipeline
+from lerobot.processor.converters import action_to_transition, transition_to_action
 from lerobot.robots import (  # noqa: F401
     Robot,
     RobotConfig,
@@ -55,6 +56,7 @@ from lerobot.robots import (  # noqa: F401
     hope_jr,
     koch_follower,
     make_robot_from_config,
+    reachy2,
     so100_follower,
     so101_follower,
 )
@@ -83,13 +85,25 @@ class ReplayConfig:
     dataset: DatasetReplayConfig
     # Use vocal synthesis to read events.
     play_sounds: bool = True
+    # Optional processor for actions before sending to robot
+    robot_action_processor: RobotProcessorPipeline | None = None
 
 
-@draccus.wrap()
+@parser.wrap()
 def replay(cfg: ReplayConfig):
     init_logging()
     logging.info(pformat(asdict(cfg)))
 
+    # Initialize robot action processor with default if not provided
+    robot_action_processor = cfg.robot_action_processor or RobotProcessorPipeline(
+        steps=[IdentityProcessorStep()],
+        to_transition=action_to_transition,
+        to_output=transition_to_action,  # type: ignore[arg-type]
+    )
+
+    # Reset processor
+    robot_action_processor.reset()
+
     robot = make_robot_from_config(cfg.robot)
     dataset = LeRobotDataset(cfg.dataset.repo_id, root=cfg.dataset.root, episodes=[cfg.dataset.episode])
     actions = dataset.hf_dataset.select_columns("action")
@@ -104,7 +118,10 @@ def replay(cfg: ReplayConfig):
         for i, name in enumerate(dataset.features["action"]["names"]):
             action[name] = action_array[i]
 
-        robot.send_action(action)
+        # Process action through robot action processor
+        processed_action = robot_action_processor(action)
+
+        robot.send_action(processed_action)  # type: ignore[arg-type]
 
         dt_s = time.perf_counter() - start_episode_t
         busy_wait(1 / dataset.fps - dt_s)

src/lerobot/robots/bi_so100_follower/config_bi_so100_follower.py

@@ -29,10 +29,10 @@ class BiSO100FollowerConfig(RobotConfig):
 
     # Optional
     left_arm_disable_torque_on_disconnect: bool = True
-    left_arm_max_relative_target: int | None = None
+    left_arm_max_relative_target: float | dict[str, float] | None = None
     left_arm_use_degrees: bool = False
     right_arm_disable_torque_on_disconnect: bool = True
-    right_arm_max_relative_target: int | None = None
+    right_arm_max_relative_target: float | dict[str, float] | None = None
     right_arm_use_degrees: bool = False
 
     # cameras (shared between both arms)

src/lerobot/robots/hope_jr/config_hope_jr.py

@@ -44,8 +44,8 @@ class HopeJrArmConfig(RobotConfig):
     disable_torque_on_disconnect: bool = True
 
     # `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
-    # Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
-    # the number of motors in your follower arms.
-    max_relative_target: int | None = None
+    # Set this to a positive scalar to have the same value for all motors, or a dictionary that maps motor
+    # names to the max_relative_target value for that motor.
+    max_relative_target: float | dict[str, float] | None = None
 
     cameras: dict[str, CameraConfig] = field(default_factory=dict)

src/lerobot/robots/koch_follower/config_koch_follower.py

@@ -28,9 +28,9 @@ class KochFollowerConfig(RobotConfig):
     disable_torque_on_disconnect: bool = True
 
     # `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
-    # Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
-    # the number of motors in your follower arms.
-    max_relative_target: int | None = None
+    # Set this to a positive scalar to have the same value for all motors, or a dictionary that maps motor
+    # names to the max_relative_target value for that motor.
+    max_relative_target: float | dict[str, float] | None = None
 
     # cameras
     cameras: dict[str, CameraConfig] = field(default_factory=dict)

src/lerobot/robots/koch_follower/koch_follower.py

@@ -110,6 +110,7 @@ class KochFollower(Robot):
         return self.bus.is_calibrated
 
     def calibrate(self) -> None:
+        self.bus.disable_torque()
         if self.calibration:
             # Calibration file exists, ask user whether to use it or run new calibration
             user_input = input(
@@ -120,7 +121,6 @@ class KochFollower(Robot):
                 self.bus.write_calibration(self.calibration)
                 return
         logger.info(f"\nRunning calibration of {self}")
-        self.bus.disable_torque()
         for motor in self.bus.motors:
             self.bus.write("Operating_Mode", motor, OperatingMode.EXTENDED_POSITION.value)
 

src/lerobot/robots/lekiwi/config_lekiwi.py

@@ -39,9 +39,9 @@ class LeKiwiConfig(RobotConfig):
     disable_torque_on_disconnect: bool = True
 
     # `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
-    # Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
-    # the number of motors in your follower arms.
-    max_relative_target: int | None = None
+    # Set this to a positive scalar to have the same value for all motors, or a dictionary that maps motor
+    # names to the max_relative_target value for that motor.
+    max_relative_target: float | dict[str, float] | None = None
 
     cameras: dict[str, CameraConfig] = field(default_factory=lekiwi_cameras_config)
 

src/lerobot/robots/reachy2/__init__.py

@@ -0,0 +1,25 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .configuration_reachy2 import Reachy2RobotConfig
+from .robot_reachy2 import (
+    REACHY2_ANTENNAS_JOINTS,
+    REACHY2_L_ARM_JOINTS,
+    REACHY2_NECK_JOINTS,
+    REACHY2_R_ARM_JOINTS,
+    REACHY2_VEL,
+    Reachy2Robot,
+)

src/lerobot/robots/reachy2/configuration_reachy2.py

@@ -0,0 +1,107 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass, field
+
+from lerobot.cameras import CameraConfig
+from lerobot.cameras.configs import ColorMode
+from lerobot.cameras.reachy2_camera import Reachy2CameraConfig
+
+from ..config import RobotConfig
+
+
+@RobotConfig.register_subclass("reachy2")
+@dataclass
+class Reachy2RobotConfig(RobotConfig):
+    # `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
+    # Set this to a positive scalar to have the same value for all motors.
+    max_relative_target: float | None = None
+
+    # IP address of the Reachy 2 robot
+    ip_address: str | None = "localhost"
+
+    # If True, turn_off_smoothly() will be sent to the robot before disconnecting.
+    disable_torque_on_disconnect: bool = False
+
+    # Tag for external commands control
+    # Set to True if you use an external commands system to control the robot,
+    # such as the official teleoperation application: https://github.com/pollen-robotics/Reachy2Teleoperation
+    # If True, robot.send_action() will not send commands to the robot.
+    use_external_commands: bool = False
+
+    # Robot parts
+    # Set to False to not add the corresponding joints part to the robot list of joints.
+    # By default, all parts are set to True.
+    with_mobile_base: bool = True
+    with_l_arm: bool = True
+    with_r_arm: bool = True
+    with_neck: bool = True
+    with_antennas: bool = True
+
+    # Robot cameras
+    # Set to True if you want to use the corresponding cameras in the observations.
+    # By default, only the teleop cameras are used.
+    with_left_teleop_camera: bool = True
+    with_right_teleop_camera: bool = True
+    with_torso_camera: bool = False
+
+    cameras: dict[str, CameraConfig] = field(default_factory=dict)
+
+    def __post_init__(self) -> None:
+        # Add cameras with same ip_address as the robot
+        if self.with_left_teleop_camera:
+            self.cameras["teleop_left"] = Reachy2CameraConfig(
+                name="teleop",
+                image_type="left",
+                ip_address=self.ip_address,
+                fps=15,
+                width=640,
+                height=480,
+                color_mode=ColorMode.RGB,
+            )
+        if self.with_right_teleop_camera:
+            self.cameras["teleop_right"] = Reachy2CameraConfig(
+                name="teleop",
+                image_type="right",
+                ip_address=self.ip_address,
+                fps=15,
+                width=640,
+                height=480,
+                color_mode=ColorMode.RGB,
+            )
+        if self.with_torso_camera:
+            self.cameras["torso_rgb"] = Reachy2CameraConfig(
+                name="depth",
+                image_type="rgb",
+                ip_address=self.ip_address,
+                fps=15,
+                width=640,
+                height=480,
+                color_mode=ColorMode.RGB,
+            )
+
+        super().__post_init__()
+
+        if not (
+            self.with_mobile_base
+            or self.with_l_arm
+            or self.with_r_arm
+            or self.with_neck
+            or self.with_antennas
+        ):
+            raise ValueError(
+                "No Reachy2Robot part used.\n"
+                "At least one part of the robot must be set to True "
+                "(with_mobile_base, with_l_arm, with_r_arm, with_neck, with_antennas)"
+            )

src/lerobot/robots/reachy2/robot_reachy2.py

@@ -0,0 +1,230 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import time
+from typing import Any
+
+import numpy as np
+from reachy2_sdk import ReachySDK
+
+from lerobot.cameras.utils import make_cameras_from_configs
+
+from ..robot import Robot
+from ..utils import ensure_safe_goal_position
+from .configuration_reachy2 import Reachy2RobotConfig
+
+# {lerobot_keys: reachy2_sdk_keys}
+REACHY2_NECK_JOINTS = {
+    "neck_yaw.pos": "head.neck.yaw",
+    "neck_pitch.pos": "head.neck.pitch",
+    "neck_roll.pos": "head.neck.roll",
+}
+
+REACHY2_ANTENNAS_JOINTS = {
+    "l_antenna.pos": "head.l_antenna",
+    "r_antenna.pos": "head.r_antenna",
+}
+
+REACHY2_R_ARM_JOINTS = {
+    "r_shoulder_pitch.pos": "r_arm.shoulder.pitch",
+    "r_shoulder_roll.pos": "r_arm.shoulder.roll",
+    "r_elbow_yaw.pos": "r_arm.elbow.yaw",
+    "r_elbow_pitch.pos": "r_arm.elbow.pitch",
+    "r_wrist_roll.pos": "r_arm.wrist.roll",
+    "r_wrist_pitch.pos": "r_arm.wrist.pitch",
+    "r_wrist_yaw.pos": "r_arm.wrist.yaw",
+    "r_gripper.pos": "r_arm.gripper",
+}
+
+REACHY2_L_ARM_JOINTS = {
+    "l_shoulder_pitch.pos": "l_arm.shoulder.pitch",
+    "l_shoulder_roll.pos": "l_arm.shoulder.roll",
+    "l_elbow_yaw.pos": "l_arm.elbow.yaw",
+    "l_elbow_pitch.pos": "l_arm.elbow.pitch",
+    "l_wrist_roll.pos": "l_arm.wrist.roll",
+    "l_wrist_pitch.pos": "l_arm.wrist.pitch",
+    "l_wrist_yaw.pos": "l_arm.wrist.yaw",
+    "l_gripper.pos": "l_arm.gripper",
+}
+
+REACHY2_VEL = {
+    "mobile_base.vx": "vx",
+    "mobile_base.vy": "vy",
+    "mobile_base.vtheta": "vtheta",
+}
+
+
+class Reachy2Robot(Robot):
+    """
+    [Reachy 2](https://www.pollen-robotics.com/reachy/), by Pollen Robotics.
+    """
+
+    config_class = Reachy2RobotConfig
+    name = "reachy2"
+
+    def __init__(self, config: Reachy2RobotConfig):
+        super().__init__(config)
+
+        self.config = config
+        self.robot_type = self.config.type
+        self.use_external_commands = self.config.use_external_commands
+
+        self.reachy: None | ReachySDK = None
+        self.cameras = make_cameras_from_configs(config.cameras)
+
+        self.logs: dict[str, float] = {}
+
+        self.joints_dict: dict[str, str] = self._generate_joints_dict()
+
+    @property
+    def observation_features(self) -> dict[str, Any]:
+        return {**self.motors_features, **self.camera_features}
+
+    @property
+    def action_features(self) -> dict[str, type]:
+        return self.motors_features
+
+    @property
+    def camera_features(self) -> dict[str, tuple[int | None, int | None, int]]:
+        return {cam: (self.cameras[cam].height, self.cameras[cam].width, 3) for cam in self.cameras}
+
+    @property
+    def motors_features(self) -> dict[str, type]:
+        if self.config.with_mobile_base:
+            return {
+                **dict.fromkeys(
+                    self.joints_dict.keys(),
+                    float,
+                ),
+                **dict.fromkeys(
+                    REACHY2_VEL.keys(),
+                    float,
+                ),
+            }
+        else:
+            return dict.fromkeys(self.joints_dict.keys(), float)
+
+    @property
+    def is_connected(self) -> bool:
+        return self.reachy.is_connected() if self.reachy is not None else False
+
+    def connect(self, calibrate: bool = False) -> None:
+        self.reachy = ReachySDK(self.config.ip_address)
+        if not self.is_connected:
+            raise ConnectionError()
+
+        for cam in self.cameras.values():
+            cam.connect()
+
+        self.configure()
+
+    def configure(self) -> None:
+        if self.reachy is not None:
+            self.reachy.turn_on()
+            self.reachy.reset_default_limits()
+
+    @property
+    def is_calibrated(self) -> bool:
+        return True
+
+    def calibrate(self) -> None:
+        pass
+
+    def _generate_joints_dict(self) -> dict[str, str]:
+        joints = {}
+        if self.config.with_neck:
+            joints.update(REACHY2_NECK_JOINTS)
+        if self.config.with_l_arm:
+            joints.update(REACHY2_L_ARM_JOINTS)
+        if self.config.with_r_arm:
+            joints.update(REACHY2_R_ARM_JOINTS)
+        if self.config.with_antennas:
+            joints.update(REACHY2_ANTENNAS_JOINTS)
+        return joints
+
+    def _get_state(self) -> dict[str, float]:
+        if self.reachy is not None:
+            pos_dict = {k: self.reachy.joints[v].present_position for k, v in self.joints_dict.items()}
+            if not self.config.with_mobile_base:
+                return pos_dict
+            vel_dict = {k: self.reachy.mobile_base.odometry[v] for k, v in REACHY2_VEL.items()}
+            return {**pos_dict, **vel_dict}
+        else:
+            return {}
+
+    def get_observation(self) -> dict[str, np.ndarray]:
+        obs_dict: dict[str, Any] = {}
+
+        # Read Reachy 2 state
+        before_read_t = time.perf_counter()
+        obs_dict.update(self._get_state())
+        self.logs["read_pos_dt_s"] = time.perf_counter() - before_read_t
+
+        # Capture images from cameras
+        for cam_key, cam in self.cameras.items():
+            obs_dict[cam_key] = cam.async_read()
+
+        return obs_dict
+
+    def send_action(self, action: dict[str, Any]) -> dict[str, Any]:
+        if self.reachy is not None:
+            if not self.is_connected:
+                raise ConnectionError()
+
+            before_write_t = time.perf_counter()
+
+            vel = {}
+            goal_pos = {}
+            for key, val in action.items():
+                if key not in self.joints_dict:
+                    if key not in REACHY2_VEL:
+                        raise KeyError(f"Key '{key}' is not a valid motor key in Reachy 2.")
+                    else:
+                        vel[REACHY2_VEL[key]] = float(val)
+                else:
+                    if not self.use_external_commands and self.config.max_relative_target is not None:
+                        goal_pos[key] = float(val)
+                        goal_present_pos = {
+                            key: (
+                                goal_pos[key],
+                                self.reachy.joints[self.joints_dict[key]].present_position,
+                            )
+                        }
+                        safe_goal_pos = ensure_safe_goal_position(
+                            goal_present_pos, float(self.config.max_relative_target)
+                        )
+                        val = safe_goal_pos[key]
+                    self.reachy.joints[self.joints_dict[key]].goal_position = float(val)
+
+            if self.config.with_mobile_base:
+                self.reachy.mobile_base.set_goal_speed(vel["vx"], vel["vy"], vel["vtheta"])
+
+            # We don't send the goal positions if we control Reachy 2 externally
+            if not self.use_external_commands:
+                self.reachy.send_goal_positions()
+                if self.config.with_mobile_base:
+                    self.reachy.mobile_base.send_speed_command()
+
+            self.logs["write_pos_dt_s"] = time.perf_counter() - before_write_t
+        return action
+
+    def disconnect(self) -> None:
+        if self.reachy is not None:
+            for cam in self.cameras.values():
+                cam.disconnect()
+            if self.config.disable_torque_on_disconnect:
+                self.reachy.turn_off_smoothly()
+            self.reachy.disconnect()

src/lerobot/robots/so100_follower/__init__.py

@@ -14,6 +14,5 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from .config_so100_follower import SO100FollowerConfig, SO100FollowerEndEffectorConfig
+from .config_so100_follower import SO100FollowerConfig
 from .so100_follower import SO100Follower
-from .so100_follower_end_effector import SO100FollowerEndEffector

src/lerobot/robots/so100_follower/config_so100_follower.py

@@ -30,44 +30,12 @@ class SO100FollowerConfig(RobotConfig):
     disable_torque_on_disconnect: bool = True
 
     # `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
-    # Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
-    # the number of motors in your follower arms.
-    max_relative_target: int | None = None
+    # Set this to a positive scalar to have the same value for all motors, or a dictionary that maps motor
+    # names to the max_relative_target value for that motor.
+    max_relative_target: float | dict[str, float] | None = None
 
     # cameras
     cameras: dict[str, CameraConfig] = field(default_factory=dict)
 
     # Set to `True` for backward compatibility with previous policies/dataset
     use_degrees: bool = False
-
-
-@RobotConfig.register_subclass("so100_follower_end_effector")
-@dataclass
-class SO100FollowerEndEffectorConfig(SO100FollowerConfig):
-    """Configuration for the SO100FollowerEndEffector robot."""
-
-    # Path to URDF file for kinematics
-    # NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo:
-    # https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
-    urdf_path: str | None = None
-
-    # End-effector frame name in the URDF
-    target_frame_name: str = "gripper_frame_link"
-
-    # Default bounds for the end-effector position (in meters)
-    end_effector_bounds: dict[str, list[float]] = field(
-        default_factory=lambda: {
-            "min": [-1.0, -1.0, -1.0],  # min x, y, z
-            "max": [1.0, 1.0, 1.0],  # max x, y, z
-        }
-    )
-
-    max_gripper_pos: float = 50
-
-    end_effector_step_sizes: dict[str, float] = field(
-        default_factory=lambda: {
-            "x": 0.02,
-            "y": 0.02,
-            "z": 0.02,
-        }
-    )

src/lerobot/robots/so100_follower/robot_kinematic_processor.py

@@ -0,0 +1,490 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass, field
+
+import numpy as np
+
+from lerobot.configs.types import FeatureType, PipelineFeatureType, PolicyFeature
+from lerobot.constants import OBS_STATE
+from lerobot.model.kinematics import RobotKinematics
+from lerobot.processor import (
+    ActionProcessorStep,
+    ComplementaryDataProcessorStep,
+    EnvTransition,
+    ObservationProcessorStep,
+    ProcessorStep,
+    ProcessorStepRegistry,
+    TransitionKey,
+)
+from lerobot.robots.robot import Robot
+from lerobot.utils.rotation import Rotation
+
+
+@ProcessorStepRegistry.register("ee_reference_and_delta")
+@dataclass
+class EEReferenceAndDelta(ActionProcessorStep):
+    """
+    Computes a target end-effector pose from a relative delta command.
+
+    This step takes a desired change in position and orientation (`target_*`) and applies it to a
+    reference end-effector pose to calculate an absolute target pose. The reference pose is derived
+    from the current robot joint positions using forward kinematics.
+
+    The processor can operate in two modes:
+    1.  `use_latched_reference=True`: The reference pose is "latched" or saved at the moment the action
+        is first enabled. Subsequent commands are relative to this fixed reference.
+    2.  `use_latched_reference=False`: The reference pose is updated to the robot's current pose at
+        every step.
+
+    Attributes:
+        kinematics: The robot's kinematic model for forward kinematics.
+        end_effector_step_sizes: A dictionary scaling the input delta commands.
+        motor_names: A list of motor names required for forward kinematics.
+        use_latched_reference: If True, latch the reference pose on enable; otherwise, always use the
+            current pose as the reference.
+        reference_ee_pose: Internal state storing the latched reference pose.
+        _prev_enabled: Internal state to detect the rising edge of the enable signal.
+        _command_when_disabled: Internal state to hold the last command while disabled.
+    """
+
+    kinematics: RobotKinematics
+    end_effector_step_sizes: dict
+    motor_names: list[str]
+    use_latched_reference: bool = (
+        True  # If True, latch reference on enable; if False, always use current pose
+    )
+
+    reference_ee_pose: np.ndarray | None = field(default=None, init=False, repr=False)
+    _prev_enabled: bool = field(default=False, init=False, repr=False)
+    _command_when_disabled: np.ndarray | None = field(default=None, init=False, repr=False)
+
+    def action(self, action):
+        new_action = action.copy()
+        comp = self.transition.get(TransitionKey.COMPLEMENTARY_DATA)
+
+        # Get joint positions from complimentary data
+        raw = comp.get("raw_joint_positions", None)
+        if raw is None:
+            raise ValueError(
+                "raw_joint_positions is not in complementary data and is required for EEReferenceAndDelta"
+            )
+
+        if "reference_joint_positions" in comp:
+            q = comp["reference_joint_positions"]
+        else:
+            q = np.array([float(raw[n]) for n in self.motor_names], dtype=float)
+
+        # Current pose from FK on measured joints
+        t_curr = self.kinematics.forward_kinematics(q)
+
+        enabled = bool(new_action.pop("enabled", 0))
+        tx = float(new_action.pop("target_x", 0.0))
+        ty = float(new_action.pop("target_y", 0.0))
+        tz = float(new_action.pop("target_z", 0.0))
+        wx = float(new_action.pop("target_wx", 0.0))
+        wy = float(new_action.pop("target_wy", 0.0))
+        wz = float(new_action.pop("target_wz", 0.0))
+
+        desired = None
+
+        if enabled:
+            ref = t_curr
+            if self.use_latched_reference:
+                # Latched reference mode: latch reference at the rising edge
+                if not self._prev_enabled or self.reference_ee_pose is None:
+                    self.reference_ee_pose = t_curr.copy()
+                ref = self.reference_ee_pose if self.reference_ee_pose is not None else t_curr
+
+            delta_p = np.array(
+                [
+                    tx * self.end_effector_step_sizes["x"],
+                    ty * self.end_effector_step_sizes["y"],
+                    tz * self.end_effector_step_sizes["z"],
+                ],
+                dtype=float,
+            )
+            r_abs = Rotation.from_rotvec([wx, wy, wz]).as_matrix()
+            desired = np.eye(4, dtype=float)
+            desired[:3, :3] = ref[:3, :3] @ r_abs
+            desired[:3, 3] = ref[:3, 3] + delta_p
+
+            self._command_when_disabled = desired.copy()
+        else:
+            # While disabled, keep sending the same command to avoid drift.
+            if self._command_when_disabled is None:
+                # If we've never had an enabled command yet, freeze current FK pose once.
+                self._command_when_disabled = t_curr.copy()
+            desired = self._command_when_disabled.copy()
+
+        # Write action fields
+        pos = desired[:3, 3]
+        tw = Rotation.from_matrix(desired[:3, :3]).as_rotvec()
+        new_action["ee.x"] = float(pos[0])
+        new_action["ee.y"] = float(pos[1])
+        new_action["ee.z"] = float(pos[2])
+        new_action["ee.wx"] = float(tw[0])
+        new_action["ee.wy"] = float(tw[1])
+        new_action["ee.wz"] = float(tw[2])
+
+        self._prev_enabled = enabled
+        return new_action
+
+    def reset(self):
+        """Resets the internal state of the processor."""
+        self._prev_enabled = False
+        self.reference_ee_pose = None
+        self._command_when_disabled = None
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        features[PipelineFeatureType.ACTION].pop("enabled", None)
+        features[PipelineFeatureType.ACTION].pop("target_x", None)
+        features[PipelineFeatureType.ACTION].pop("target_y", None)
+        features[PipelineFeatureType.ACTION].pop("target_z", None)
+        features[PipelineFeatureType.ACTION].pop("target_wx", None)
+        features[PipelineFeatureType.ACTION].pop("target_wy", None)
+        features[PipelineFeatureType.ACTION].pop("target_wz", None)
+
+        features[PipelineFeatureType.ACTION]["ee.x"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        features[PipelineFeatureType.ACTION]["ee.y"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        features[PipelineFeatureType.ACTION]["ee.z"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        features[PipelineFeatureType.ACTION]["ee.wx"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        features[PipelineFeatureType.ACTION]["ee.wy"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        features[PipelineFeatureType.ACTION]["ee.wz"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        return features
+
+
+@ProcessorStepRegistry.register("ee_bounds_and_safety")
+@dataclass
+class EEBoundsAndSafety(ActionProcessorStep):
+    """
+    Clips the end-effector pose to predefined bounds and checks for unsafe jumps.
+
+    This step ensures that the target end-effector pose remains within a safe operational workspace.
+    It also moderates the command to prevent large, sudden movements between consecutive steps.
+
+    Attributes:
+        end_effector_bounds: A dictionary with "min" and "max" keys for position clipping.
+        max_ee_step_m: The maximum allowed change in position (in meters) between steps.
+        max_ee_twist_step_rad: The maximum allowed change in orientation (in radians) between steps.
+        _last_pos: Internal state storing the last commanded position.
+        _last_twist: Internal state storing the last commanded orientation.
+    """
+
+    end_effector_bounds: dict
+    max_ee_step_m: float = 0.05
+    max_ee_twist_step_rad: float = 0.20
+    _last_pos: np.ndarray | None = field(default=None, init=False, repr=False)
+    _last_twist: np.ndarray | None = field(default=None, init=False, repr=False)
+
+    def action(self, act: dict) -> dict:
+        x = act.get("ee.x", None)
+        y = act.get("ee.y", None)
+        z = act.get("ee.z", None)
+        wx = act.get("ee.wx", None)
+        wy = act.get("ee.wy", None)
+        wz = act.get("ee.wz", None)
+
+        if None in (x, y, z, wx, wy, wz):
+            raise ValueError(
+                "Missing required end-effector pose components: x, y, z, wx, wy, wz must all be present in action"
+            )
+
+        pos = np.array([x, y, z], dtype=float)
+        twist = np.array([wx, wy, wz], dtype=float)
+
+        # Clip position
+        pos = np.clip(pos, self.end_effector_bounds["min"], self.end_effector_bounds["max"])
+
+        # Check for jumps in position
+        if self._last_pos is not None:
+            dpos = pos - self._last_pos
+            n = float(np.linalg.norm(dpos))
+            if n > self.max_ee_step_m and n > 0:
+                pos = self._last_pos + dpos * (self.max_ee_step_m / n)
+                raise ValueError(f"EE jump {n:.3f}m > {self.max_ee_step_m}m")
+
+        self._last_pos = pos
+        self._last_twist = twist
+
+        act["ee.x"] = float(pos[0])
+        act["ee.y"] = float(pos[1])
+        act["ee.z"] = float(pos[2])
+        act["ee.wx"] = float(twist[0])
+        act["ee.wy"] = float(twist[1])
+        act["ee.wz"] = float(twist[2])
+        return act
+
+    def reset(self):
+        """Resets the last known position and orientation."""
+        self._last_pos = None
+        self._last_twist = None
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        return features
+
+
+@ProcessorStepRegistry.register("inverse_kinematics_ee_to_joints")
+@dataclass
+class InverseKinematicsEEToJoints(ProcessorStep):
+    """
+    Computes desired joint positions from a target end-effector pose using inverse kinematics (IK).
+
+    This step translates a Cartesian command (position and orientation of the end-effector) into
+    the corresponding joint-space commands for each motor.
+
+    Attributes:
+        kinematics: The robot's kinematic model for inverse kinematics.
+        motor_names: A list of motor names for which to compute joint positions.
+        q_curr: Internal state storing the last joint positions, used as an initial guess for the IK solver.
+        initial_guess_current_joints: If True, use the robot's current joint state as the IK guess.
+            If False, use the solution from the previous step.
+    """
+
+    kinematics: RobotKinematics
+    motor_names: list[str]
+    q_curr: np.ndarray | None = field(default=None, init=False, repr=False)
+    initial_guess_current_joints: bool = True
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        new_transition = transition.copy()
+        act = new_transition.get(TransitionKey.ACTION) or {}
+        comp = new_transition.get(TransitionKey.COMPLEMENTARY_DATA) or {}
+
+        x = act.get("ee.x", None)
+        y = act.get("ee.y", None)
+        z = act.get("ee.z", None)
+        wx = act.get("ee.wx", None)
+        wy = act.get("ee.wy", None)
+        wz = act.get("ee.wz", None)
+
+        if None in (x, y, z, wx, wy, wz):
+            return new_transition
+
+        # Get joint positions from complimentary data
+        raw = comp.get("raw_joint_positions", None)
+        if raw is None:
+            raise ValueError(
+                "raw_joint_positions is not in complementary data and is required for EEReferenceAndDelta"
+            )
+
+        if self.initial_guess_current_joints:  # Use current joints as initial guess
+            self.q_curr = np.array([float(raw[n]) for n in self.motor_names], dtype=float)
+        else:  # Use previous ik solution as initial guess
+            if self.q_curr is None:
+                self.q_curr = np.array([float(raw[n]) for n in self.motor_names], dtype=float)
+
+        # Build desired 4x4 transform from pos + rotvec (twist)
+        t_des = np.eye(4, dtype=float)
+        t_des[:3, :3] = Rotation.from_rotvec([wx, wy, wz]).as_matrix()
+        t_des[:3, 3] = [x, y, z]
+
+        # Compute inverse kinematics
+        q_target = self.kinematics.inverse_kinematics(self.q_curr, t_des)
+        self.q_curr = q_target
+
+        new_act = dict(act)
+        for i, name in enumerate(self.motor_names):
+            if name == "gripper":
+                # TODO(pepijn): Investigate if this is correct
+                # Do we want an observation key in the action field?
+                new_act["gripper.pos"] = float(raw["gripper"])
+            else:
+                new_act[f"{name}.pos"] = float(q_target[i])
+        new_transition[TransitionKey.ACTION] = new_act
+        if not self.initial_guess_current_joints:
+            new_transition[TransitionKey.COMPLEMENTARY_DATA]["reference_joint_positions"] = q_target
+        return new_transition
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        features[PipelineFeatureType.ACTION]["gripper.pos"] = PolicyFeature(
+            type=FeatureType.ACTION, shape=(1,)
+        )
+        for name in self.motor_names:
+            features[PipelineFeatureType.ACTION][f"{name}.pos"] = PolicyFeature(
+                type=FeatureType.ACTION, shape=(1,)
+            )
+
+        return features
+
+    def reset(self):
+        """Resets the initial guess for the IK solver."""
+        self.q_curr = None
+
+
+@ProcessorStepRegistry.register("gripper_velocity_to_joint")
+@dataclass
+class GripperVelocityToJoint(ProcessorStep):
+    """
+    Converts a gripper velocity command into a target gripper joint position.
+
+    This step integrates a normalized velocity command over time to produce a position command,
+    taking the current gripper position as a starting point. It also supports a discrete mode
+    where integer actions map to open, close, or no-op.
+
+    Attributes:
+        motor_names: A list of motor names, which must include 'gripper'.
+        speed_factor: A scaling factor to convert the normalized velocity command to a position change.
+        clip_min: The minimum allowed gripper joint position.
+        clip_max: The maximum allowed gripper joint position.
+        discrete_gripper: If True, treat the input action as discrete (0: open, 1: close, 2: stay).
+    """
+
+    motor_names: list[str]
+    speed_factor: float = 20.0
+    clip_min: float = 0.0
+    clip_max: float = 100.0
+    discrete_gripper: bool = False
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        new_transition = transition.copy()
+        obs = new_transition.get(TransitionKey.OBSERVATION) or {}
+        act = new_transition.get(TransitionKey.ACTION) or {}
+        comp = new_transition.get(TransitionKey.COMPLEMENTARY_DATA) or {}
+
+        if "gripper" not in act:
+            raise ValueError("Required action key 'gripper' not found in transition")
+
+        if "gripper" not in self.motor_names:
+            raise ValueError(
+                f"Required motor name 'gripper' not found in self.motor_names={self.motor_names}"
+            )
+
+        if self.discrete_gripper:
+            # Discrete gripper actions are in [0, 1, 2]
+            # 0: open, 1: close, 2: stay
+            # We need to shift them to [-1, 0, 1] and then scale them to clip_max
+            gripper_action = act.get("gripper", 1.0)
+            gripper_action = gripper_action - 1.0
+            gripper_action *= self.clip_max
+            act["gripper"] = gripper_action
+
+        # Get current gripper position from complementary data
+        raw = comp.get("raw_joint_positions") or {}
+        curr_pos = float(raw.get("gripper"))
+
+        # Compute desired gripper velocity
+        u = float(act.get("gripper", 0.0))
+        delta = u * float(self.speed_factor)
+        gripper_pos = float(np.clip(curr_pos + delta, self.clip_min, self.clip_max))
+
+        new_act = dict(act)
+        new_act["gripper.pos"] = gripper_pos
+        new_act.pop("gripper", None)
+        new_transition[TransitionKey.ACTION] = new_act
+
+        obs[f"{OBS_STATE}.gripper.pos"] = curr_pos
+        new_transition[TransitionKey.OBSERVATION] = obs
+        return new_transition
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        features[PipelineFeatureType.ACTION].pop("gripper", None)
+        features[PipelineFeatureType.ACTION]["gripper.pos"] = PolicyFeature(
+            type=FeatureType.ACTION, shape=(1,)
+        )
+        features[PipelineFeatureType.OBSERVATION][f"{OBS_STATE}.gripper.pos"] = PolicyFeature(
+            type=FeatureType.STATE, shape=(1,)
+        )
+
+        return features
+
+
+@ProcessorStepRegistry.register("forward_kinematics_joints_to_ee")
+@dataclass
+class ForwardKinematicsJointsToEE(ObservationProcessorStep):
+    """
+    Computes the end-effector pose from joint positions using forward kinematics (FK).
+
+    This step is typically used to add the robot's Cartesian pose to the observation space,
+    which can be useful for visualization or as an input to a policy.
+
+    Attributes:
+        kinematics: The robot's kinematic model.
+        motor_names: A list of motor names whose joint positions are used for FK.
+    """
+
+    kinematics: RobotKinematics
+    motor_names: list[str]
+
+    def observation(self, obs: dict) -> dict:
+        if not all(f"{OBS_STATE}.{n}.pos" in obs for n in self.motor_names):
+            raise ValueError(f"Missing required joint positions for motors: {self.motor_names}")
+
+        q = np.array([obs[f"{OBS_STATE}.{n}.pos"] for n in self.motor_names], dtype=float)
+        t = self.kinematics.forward_kinematics(q)
+        pos = t[:3, 3]
+        tw = Rotation.from_matrix(t[:3, :3]).as_rotvec()
+
+        obs[f"{OBS_STATE}.ee.x"] = float(pos[0])
+        obs[f"{OBS_STATE}.ee.y"] = float(pos[1])
+        obs[f"{OBS_STATE}.ee.z"] = float(pos[2])
+        obs[f"{OBS_STATE}.ee.wx"] = float(tw[0])
+        obs[f"{OBS_STATE}.ee.wy"] = float(tw[1])
+        obs[f"{OBS_STATE}.ee.wz"] = float(tw[2])
+        return obs
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        # We specify the dataset features of this step that we want to be stored in the dataset
+        for k in ["x", "y", "z", "wx", "wy", "wz"]:
+            features[PipelineFeatureType.OBSERVATION][f"{OBS_STATE}.ee.{k}"] = PolicyFeature(
+                type=FeatureType.STATE, shape=(1,)
+            )
+        return features
+
+
+@ProcessorStepRegistry.register("add_robot_observation")
+@dataclass
+class AddRobotObservationAsComplimentaryData(ComplementaryDataProcessorStep):
+    """
+    Reads the robot's current observation and adds it to the transition's complementary data.
+
+    This step acts as a bridge to the physical robot, injecting its real-time sensor readings
+    (like raw joint positions) into the data processing pipeline. This data is then available
+    for other processing steps.
+
+    Attributes:
+        robot: An instance of a `Robot` class used to get observations from hardware.
+    """
+
+    robot: Robot
+
+    def complementary_data(self, comp: dict | None) -> dict:
+        new_comp = dict(comp)
+        obs = (
+            self.robot.get_observation()
+        )  # todo(steven): why not self.trtansition.get(TransitionKey.OBSERVATION)?
+
+        new_comp["raw_joint_positions"] = {
+            k.removesuffix(".pos"): float(v)
+            for k, v in obs.items()
+            if isinstance(k, str) and k.endswith(".pos")
+        }
+        return new_comp
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        return features

src/lerobot/robots/so100_follower/so100_follower.py

@@ -161,6 +161,11 @@ class SO100Follower(Robot):
                 self.bus.write("I_Coefficient", motor, 0)
                 self.bus.write("D_Coefficient", motor, 32)
 
+                if motor == "gripper":
+                    self.bus.write("Max_Torque_Limit", motor, 500)  # 50% of max torque to avoid burnout
+                    self.bus.write("Protection_Current", motor, 250)  # 50% of max current to avoid burnout
+                    self.bus.write("Overload_Torque", motor, 25)  # 25% torque when overloaded
+
     def setup_motors(self) -> None:
         for motor in reversed(self.bus.motors):
             input(f"Connect the controller board to the '{motor}' motor only and press enter.")

src/lerobot/robots/so100_follower/so100_follower_end_effector.py

@@ -1,200 +0,0 @@
-# !/usr/bin/env python
-
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-import logging
-import time
-from typing import Any
-
-import numpy as np
-
-from lerobot.cameras import make_cameras_from_configs
-from lerobot.errors import DeviceNotConnectedError
-from lerobot.model.kinematics import RobotKinematics
-from lerobot.motors import Motor, MotorNormMode
-from lerobot.motors.feetech import FeetechMotorsBus
-
-from . import SO100Follower
-from .config_so100_follower import SO100FollowerEndEffectorConfig
-
-logger = logging.getLogger(__name__)
-
-
-class SO100FollowerEndEffector(SO100Follower):
-    """
-    SO100Follower robot with end-effector space control.
-
-    This robot inherits from SO100Follower but transforms actions from
-    end-effector space to joint space before sending them to the motors.
-    """
-
-    config_class = SO100FollowerEndEffectorConfig
-    name = "so100_follower_end_effector"
-
-    def __init__(self, config: SO100FollowerEndEffectorConfig):
-        super().__init__(config)
-        self.bus = FeetechMotorsBus(
-            port=self.config.port,
-            motors={
-                "shoulder_pan": Motor(1, "sts3215", MotorNormMode.DEGREES),
-                "shoulder_lift": Motor(2, "sts3215", MotorNormMode.DEGREES),
-                "elbow_flex": Motor(3, "sts3215", MotorNormMode.DEGREES),
-                "wrist_flex": Motor(4, "sts3215", MotorNormMode.DEGREES),
-                "wrist_roll": Motor(5, "sts3215", MotorNormMode.DEGREES),
-                "gripper": Motor(6, "sts3215", MotorNormMode.RANGE_0_100),
-            },
-            calibration=self.calibration,
-        )
-
-        self.cameras = make_cameras_from_configs(config.cameras)
-
-        self.config = config
-
-        # Initialize the kinematics module for the so100 robot
-        if self.config.urdf_path is None:
-            raise ValueError(
-                "urdf_path must be provided in the configuration for end-effector control. "
-                "Please set urdf_path in your SO100FollowerEndEffectorConfig."
-            )
-
-        self.kinematics = RobotKinematics(
-            urdf_path=self.config.urdf_path,
-            target_frame_name=self.config.target_frame_name,
-        )
-
-        # Store the bounds for end-effector position
-        self.end_effector_bounds = self.config.end_effector_bounds
-
-        self.current_ee_pos = None
-        self.current_joint_pos = None
-
-    @property
-    def action_features(self) -> dict[str, Any]:
-        """
-        Define action features for end-effector control.
-        Returns dictionary with dtype, shape, and names.
-        """
-        return {
-            "dtype": "float32",
-            "shape": (4,),
-            "names": {"delta_x": 0, "delta_y": 1, "delta_z": 2, "gripper": 3},
-        }
-
-    def send_action(self, action: dict[str, Any]) -> dict[str, Any]:
-        """
-        Transform action from end-effector space to joint space and send to motors.
-
-        Args:
-            action: Dictionary with keys 'delta_x', 'delta_y', 'delta_z' for end-effector control
-                   or a numpy array with [delta_x, delta_y, delta_z]
-
-        Returns:
-            The joint-space action that was sent to the motors
-        """
-
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")
-
-        # Convert action to numpy array if not already
-        if isinstance(action, dict):
-            if all(k in action for k in ["delta_x", "delta_y", "delta_z"]):
-                delta_ee = np.array(
-                    [
-                        action["delta_x"] * self.config.end_effector_step_sizes["x"],
-                        action["delta_y"] * self.config.end_effector_step_sizes["y"],
-                        action["delta_z"] * self.config.end_effector_step_sizes["z"],
-                    ],
-                    dtype=np.float32,
-                )
-                if "gripper" not in action:
-                    action["gripper"] = [1.0]
-                action = np.append(delta_ee, action["gripper"])
-            else:
-                logger.warning(
-                    f"Expected action keys 'delta_x', 'delta_y', 'delta_z', got {list(action.keys())}"
-                )
-                action = np.zeros(4, dtype=np.float32)
-
-        if self.current_joint_pos is None:
-            # Read current joint positions
-            current_joint_pos = self.bus.sync_read("Present_Position")
-            self.current_joint_pos = np.array([current_joint_pos[name] for name in self.bus.motors])
-
-        # Calculate current end-effector position using forward kinematics
-        if self.current_ee_pos is None:
-            self.current_ee_pos = self.kinematics.forward_kinematics(self.current_joint_pos)
-
-        # Set desired end-effector position by adding delta
-        desired_ee_pos = np.eye(4)
-        desired_ee_pos[:3, :3] = self.current_ee_pos[:3, :3]  # Keep orientation
-
-        # Add delta to position and clip to bounds
-        desired_ee_pos[:3, 3] = self.current_ee_pos[:3, 3] + action[:3]
-        if self.end_effector_bounds is not None:
-            desired_ee_pos[:3, 3] = np.clip(
-                desired_ee_pos[:3, 3],
-                self.end_effector_bounds["min"],
-                self.end_effector_bounds["max"],
-            )
-
-        # Compute inverse kinematics to get joint positions
-        target_joint_values_in_degrees = self.kinematics.inverse_kinematics(
-            self.current_joint_pos, desired_ee_pos
-        )
-
-        # Create joint space action dictionary
-        joint_action = {
-            f"{key}.pos": target_joint_values_in_degrees[i] for i, key in enumerate(self.bus.motors.keys())
-        }
-
-        # Handle gripper separately if included in action
-        # Gripper delta action is in the range 0 - 2,
-        # We need to shift the action to the range -1, 1 so that we can expand it to -Max_gripper_pos, Max_gripper_pos
-        joint_action["gripper.pos"] = np.clip(
-            self.current_joint_pos[-1] + (action[-1] - 1) * self.config.max_gripper_pos,
-            5,
-            self.config.max_gripper_pos,
-        )
-
-        self.current_ee_pos = desired_ee_pos.copy()
-        self.current_joint_pos = target_joint_values_in_degrees.copy()
-        self.current_joint_pos[-1] = joint_action["gripper.pos"]
-
-        # Send joint space action to parent class
-        return super().send_action(joint_action)
-
-    def get_observation(self) -> dict[str, Any]:
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")
-
-        # Read arm position
-        start = time.perf_counter()
-        obs_dict = self.bus.sync_read("Present_Position")
-        obs_dict = {f"{motor}.pos": val for motor, val in obs_dict.items()}
-        dt_ms = (time.perf_counter() - start) * 1e3
-        logger.debug(f"{self} read state: {dt_ms:.1f}ms")
-
-        # Capture images from cameras
-        for cam_key, cam in self.cameras.items():
-            start = time.perf_counter()
-            obs_dict[cam_key] = cam.async_read()
-            dt_ms = (time.perf_counter() - start) * 1e3
-            logger.debug(f"{self} read {cam_key}: {dt_ms:.1f}ms")
-
-        return obs_dict
-
-    def reset(self):
-        self.current_ee_pos = None
-        self.current_joint_pos = None

src/lerobot/robots/so101_follower/config_so101_follower.py

@@ -30,9 +30,9 @@ class SO101FollowerConfig(RobotConfig):
     disable_torque_on_disconnect: bool = True
 
     # `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
-    # Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
-    # the number of motors in your follower arms.
-    max_relative_target: int | None = None
+    # Set this to a positive scalar to have the same value for all motors, or a dictionary that maps motor
+    # names to the max_relative_target value for that motor.
+    max_relative_target: float | dict[str, float] | None = None
 
     # cameras
     cameras: dict[str, CameraConfig] = field(default_factory=dict)