refactor(tests): update processor test assertions to reflect new preprocessor and postprocessor names (#1869)

- Changed assertions in multiple processor test files to verify the updated names from "robot_preprocessor" and "robot_postprocessor" to "policy_preprocessor" and "policy_postprocessor" for consistency with recent refactoring.

.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

@@ -39,6 +39,8 @@
 - 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}}" \

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/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

@@ -1,7 +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_processor
+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
@@ -46,7 +46,7 @@ listener, events = init_keyboard_listener()
 if not robot.is_connected:
     raise ValueError("Robot is not connected!")
 
-preprocessor, postprocessor = make_processor(
+preprocessor, postprocessor = make_pre_post_processors(
     policy_cfg=policy,
     pretrained_path=HF_MODEL_ID,
     dataset_stats=dataset.meta.stats,

examples/phone_to_so100/evaluate.py

@@ -17,15 +17,15 @@
 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
-from lerobot.datasets.utils import merge_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_processor
+from lerobot.policies.factory import make_pre_post_processors
+from lerobot.processor import RobotProcessorPipeline
 from lerobot.processor.converters import (
-    to_output_robot_action,
-    to_transition_robot_observation,
+    observation_to_transition,
+    transition_to_robot_action,
 )
-from lerobot.processor.pipeline import RobotProcessor
 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 (
@@ -65,7 +65,7 @@ kinematics_solver = RobotKinematics(
 )
 
 # Build pipeline to convert ee pose action to joint action
-robot_ee_to_joints = RobotProcessor(
+robot_ee_to_joints_processor = RobotProcessorPipeline(
     steps=[
         AddRobotObservationAsComplimentaryData(robot=robot),
         InverseKinematicsEEToJoints(
@@ -75,21 +75,21 @@ robot_ee_to_joints = RobotProcessor(
         ),
     ],
     to_transition=lambda tr: tr,
-    to_output=to_output_robot_action,
+    to_output=transition_to_robot_action,
 )
 
 # Build pipeline to convert joint observation to ee pose observation
-robot_joints_to_ee_pose = RobotProcessor(
+robot_joints_to_ee_pose_processor = RobotProcessorPipeline(
     steps=[
         ForwardKinematicsJointsToEE(kinematics=kinematics_solver, motor_names=list(robot.bus.motors.keys()))
     ],
-    to_transition=to_transition_robot_observation,
+    to_transition=observation_to_transition,
     to_output=lambda tr: tr,
 )
 
 # Build dataset action and gripper features
 action_ee_and_gripper = aggregate_pipeline_dataset_features(
-    pipeline=robot_ee_to_joints,
+    pipeline=robot_ee_to_joints_processor,
     initial_features={},
     use_videos=True,
     patterns=["action.ee", "action.gripper.pos", "observation.state.gripper.pos"],
@@ -97,13 +97,13 @@ action_ee_and_gripper = aggregate_pipeline_dataset_features(
 
 # Build dataset observation features
 obs_ee = aggregate_pipeline_dataset_features(
-    pipeline=robot_joints_to_ee_pose,
+    pipeline=robot_joints_to_ee_pose_processor,
     initial_features=robot.observation_features,
     use_videos=True,
     patterns=["observation.state.ee"],
 )  # Get all ee observation features
 
-dataset_features = merge_features(obs_ee, action_ee_and_gripper)
+dataset_features = combine_feature_dicts(obs_ee, action_ee_and_gripper)
 
 print("All dataset features: ", dataset_features)
 
@@ -127,7 +127,7 @@ robot.connect()
 episode_idx = 0
 
 policy = ACTPolicy.from_pretrained(HF_MODEL_ID)
-preprocessor, postprocessor = make_processor(
+preprocessor, postprocessor = make_pre_post_processors(
     policy_cfg=policy,
     pretrained_path=HF_MODEL_ID,
     dataset_stats=dataset.meta.stats,
@@ -147,8 +147,8 @@ for episode_idx in range(NUM_EPISODES):
         control_time_s=EPISODE_TIME_SEC,
         single_task=TASK_DESCRIPTION,
         display_data=True,
-        robot_action_processor=robot_ee_to_joints,
-        robot_observation_processor=robot_joints_to_ee_pose,
+        robot_action_processor=robot_ee_to_joints_processor,
+        robot_observation_processor=robot_joints_to_ee_pose_processor,
     )
     dataset.save_episode()
 

examples/phone_to_so100/record.py

@@ -18,14 +18,14 @@
 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
-from lerobot.datasets.utils import merge_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 (
-    to_output_robot_action,
-    to_transition_robot_observation,
-    to_transition_teleop_action,
+    action_to_transition,
+    observation_to_transition,
+    transition_to_robot_action,
 )
-from lerobot.processor.pipeline import RobotProcessor
 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 (
@@ -38,8 +38,8 @@ from lerobot.robots.so100_follower.robot_kinematic_processor import (
 )
 from lerobot.robots.so100_follower.so100_follower import SO100Follower
 from lerobot.teleoperators.phone.config_phone import PhoneConfig, PhoneOS
-from lerobot.teleoperators.phone.phone import Phone
 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
@@ -73,7 +73,7 @@ kinematics_solver = RobotKinematics(
 )
 
 # Build pipeline to convert phone action to ee pose action
-phone_to_robot_ee_pose = RobotProcessor(
+phone_to_robot_ee_pose_processor = RobotProcessorPipeline(
     steps=[
         MapPhoneActionToRobotAction(platform=teleop_config.phone_os),
         AddRobotObservationAsComplimentaryData(robot=robot),
@@ -88,12 +88,12 @@ phone_to_robot_ee_pose = RobotProcessor(
             max_ee_twist_step_rad=0.50,
         ),
     ],
-    to_transition=to_transition_teleop_action,
+    to_transition=action_to_transition,
     to_output=lambda tr: tr,
 )
 
 # Build pipeline to convert ee pose action to joint action
-robot_ee_to_joints = RobotProcessor(
+robot_ee_to_joints_processor = RobotProcessorPipeline(
     steps=[
         InverseKinematicsEEToJoints(
             kinematics=kinematics_solver,
@@ -106,21 +106,21 @@ robot_ee_to_joints = RobotProcessor(
         ),
     ],
     to_transition=lambda tr: tr,
-    to_output=to_output_robot_action,
+    to_output=transition_to_robot_action,
 )
 
 # Build pipeline to convert joint observation to ee pose observation
-robot_joints_to_ee_pose = RobotProcessor(
+robot_joints_to_ee_pose = RobotProcessorPipeline(
     steps=[
         ForwardKinematicsJointsToEE(kinematics=kinematics_solver, motor_names=list(robot.bus.motors.keys()))
     ],
-    to_transition=to_transition_robot_observation,
+    to_transition=observation_to_transition,
     to_output=lambda tr: tr,
 )
 
 # Build dataset ee action features
 action_ee = aggregate_pipeline_dataset_features(
-    pipeline=phone_to_robot_ee_pose,
+    pipeline=phone_to_robot_ee_pose_processor,
     initial_features=phone.action_features,
     use_videos=True,
     patterns=["action.ee"],
@@ -128,7 +128,7 @@ action_ee = aggregate_pipeline_dataset_features(
 
 # Get gripper pos action features
 gripper = aggregate_pipeline_dataset_features(
-    pipeline=robot_ee_to_joints,
+    pipeline=robot_ee_to_joints_processor,
     initial_features={},
     use_videos=True,
     patterns=["action.gripper.pos", "observation.state.gripper.pos"],
@@ -142,7 +142,7 @@ observation_ee = aggregate_pipeline_dataset_features(
     patterns=["observation.state.ee"],
 )
 
-dataset_features = merge_features(action_ee, gripper, observation_ee)
+dataset_features = combine_feature_dicts(action_ee, gripper, observation_ee)
 
 print("All dataset features: ", dataset_features)
 
@@ -177,8 +177,8 @@ while episode_idx < NUM_EPISODES and not events["stop_recording"]:
         control_time_s=EPISODE_TIME_SEC,
         single_task=TASK_DESCRIPTION,
         display_data=True,
-        teleop_action_processor=phone_to_robot_ee_pose,
-        robot_action_processor=robot_ee_to_joints,
+        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,
     )
 
@@ -193,8 +193,8 @@ while episode_idx < NUM_EPISODES and not events["stop_recording"]:
             control_time_s=RESET_TIME_SEC,
             single_task=TASK_DESCRIPTION,
             display_data=True,
-            teleop_action_processor=phone_to_robot_ee_pose,
-            robot_action_processor=robot_ee_to_joints,
+            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,
         )
 

examples/phone_to_so100/replay.py

@@ -19,8 +19,8 @@ import time
 
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.model.kinematics import RobotKinematics
-from lerobot.processor.converters import to_output_robot_action
-from lerobot.processor.pipeline import RobotProcessor
+from lerobot.processor import RobotProcessorPipeline
+from lerobot.processor.converters import action_to_transition, transition_to_robot_action
 from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
 from lerobot.robots.so100_follower.robot_kinematic_processor import (
     AddRobotObservationAsComplimentaryData,
@@ -49,33 +49,8 @@ kinematics_solver = RobotKinematics(
     joint_names=list(robot.bus.motors.keys()),
 )
 
-
-# This method converts the action from the dataset to a transition for pipeline
-def action_to_transition(action: dict):
-    act = {}
-
-    # EE pose
-    for k in ("ee.x", "ee.y", "ee.z", "ee.wx", "ee.wy", "ee.wz"):
-        if k in action:
-            act[f"action.{k}"] = float(action[k])
-
-    # Gripper: your dataset has absolute position
-    if "gripper.pos" in action:
-        act["action.gripper.pos"] = float(action["gripper.pos"])
-
-    return {
-        "observation": None,
-        "action": act,
-        "reward": None,
-        "done": False,
-        "truncated": False,
-        "info": {},
-        "complementary_data": {},
-    }
-
-
 # Build pipeline to convert ee pose action to joint action
-robot_ee_to_joints = RobotProcessor(
+robot_ee_to_joints_processor = RobotProcessorPipeline(
     steps=[
         AddRobotObservationAsComplimentaryData(robot=robot),
         InverseKinematicsEEToJoints(
@@ -85,10 +60,10 @@ robot_ee_to_joints = RobotProcessor(
         ),
     ],
     to_transition=action_to_transition,
-    to_output=to_output_robot_action,
+    to_output=transition_to_robot_action,
 )
 
-robot_ee_to_joints.reset()
+robot_ee_to_joints_processor.reset()
 
 log_say(f"Replaying episode {EPISODE_IDX}")
 for idx in range(dataset.num_frames):
@@ -98,7 +73,7 @@ for idx in range(dataset.num_frames):
         name: float(actions[idx]["action"][i]) for i, name in enumerate(dataset.features["action"]["names"])
     }
 
-    joint_action = robot_ee_to_joints(ee_action)
+    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))

examples/phone_to_so100/teleoperate.py

@@ -16,8 +16,8 @@
 import time
 
 from lerobot.model.kinematics import RobotKinematics
-from lerobot.processor import RobotProcessor
-from lerobot.processor.converters import to_output_robot_action, to_transition_teleop_action
+from lerobot.processor import RobotProcessorPipeline
+from lerobot.processor.converters import action_to_transition, transition_to_robot_action
 from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
 from lerobot.robots.so100_follower.robot_kinematic_processor import (
     AddRobotObservationAsComplimentaryData,
@@ -28,8 +28,8 @@ from lerobot.robots.so100_follower.robot_kinematic_processor import (
 )
 from lerobot.robots.so100_follower.so100_follower import SO100Follower
 from lerobot.teleoperators.phone.config_phone import PhoneConfig, PhoneOS
-from lerobot.teleoperators.phone.phone import Phone
 from lerobot.teleoperators.phone.phone_processor import MapPhoneActionToRobotAction
+from lerobot.teleoperators.phone.teleop_phone import Phone
 
 # Initialize the robot and teleoperator
 robot_config = SO100FollowerConfig(
@@ -48,8 +48,8 @@ kinematics_solver = RobotKinematics(
     joint_names=list(robot.bus.motors.keys()),
 )
 
-# Build pipeline to convert phone action to ee pose action
-phone_to_robot_ee_pose = RobotProcessor(
+# 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),
@@ -63,14 +63,6 @@ phone_to_robot_ee_pose = RobotProcessor(
             max_ee_step_m=0.10,
             max_ee_twist_step_rad=0.50,
         ),
-    ],
-    to_transition=to_transition_teleop_action,
-    to_output=lambda tr: tr,
-)
-
-# Build pipeline to convert ee pose action to joint action
-robot_ee_to_joints = RobotProcessor(
-    steps=[
         InverseKinematicsEEToJoints(
             kinematics=kinematics_solver,
             motor_names=list(robot.bus.motors.keys()),
@@ -80,8 +72,8 @@ robot_ee_to_joints = RobotProcessor(
             speed_factor=20.0,
         ),
     ],
-    to_transition=lambda tr: tr,
-    to_output=to_output_robot_action,
+    to_transition=action_to_transition,
+    to_output=transition_to_robot_action,
 )
 
 robot.connect()
@@ -89,19 +81,11 @@ teleop_device.connect()
 
 print("Starting teleop loop. Move your phone to teleoperate the robot.")
 while True:
-    phone_obs = teleop_device.get_action()
-    if not phone_obs:
-        time.sleep(0.01)
-        continue
-
     # Get teleop observation
     phone_obs = teleop_device.get_action()
 
-    # Phone to EE pose transition
-    ee_transition = phone_to_robot_ee_pose(phone_obs)
-
-    # EE pose to Joints transition
-    joint_action = robot_ee_to_joints(ee_transition)
+    # Phone -> EE pose -> Joints transition
+    joint_action = phone_to_robot_joints_processor(phone_obs)
 
     if joint_action:
         robot.send_action(joint_action)

pyproject.toml

@@ -73,7 +73,6 @@ dependencies = [
     "pynput>=1.7.7",
     "pyserial>=3.5",
     "wandb>=0.20.0",
-    "scipy>=1.15.2",
 
     "torch>=2.2.1,<2.8.0", # TODO: Bumb dependency
     "torchcodec>=0.2.1,<0.6.0; sys_platform != 'win32' and (sys_platform != 'linux' or (platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')) and (sys_platform != 'darwin' or platform_machine != 'x86_64')", # TODO: Bumb dependency

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/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/constants.py

@@ -24,6 +24,11 @@ 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"
@@ -40,6 +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

@@ -15,12 +15,13 @@
 from collections.abc import Sequence
 from typing import Any
 
+from lerobot.constants import ACTION, OBS_IMAGES, OBS_STATE
 from lerobot.datasets.utils import hw_to_dataset_features
-from lerobot.processor.pipeline import RobotProcessor
+from lerobot.processor import DataProcessorPipeline
 
 
 def aggregate_pipeline_dataset_features(
-    pipeline: RobotProcessor,
+    pipeline: DataProcessorPipeline,
     initial_features: dict[str, Any],
     *,
     use_videos: bool = True,
@@ -59,26 +60,26 @@ def aggregate_pipeline_dataset_features(
 
     # Go over every feature from the pipeline and merge:
     for full_key, ty in all_features.items():
-        if full_key.startswith("action."):
+        if full_key.startswith(f"{ACTION}."):
             # action.<feat>
             if not keep(full_key):
                 continue
-            name = full_key[len("action.") :]
-            hw.setdefault("action", {})[name] = ty
+            name = full_key[len(f"{ACTION}.") :]
+            hw.setdefault(ACTION, {})[name] = ty
 
-        elif full_key.startswith("observation.state."):
+        elif full_key.startswith(f"{OBS_STATE}."):
             # observation.state.<feat>
             if not keep(full_key):
                 continue
-            name = full_key[len("observation.state.") :]
+            name = full_key[len(f"{OBS_STATE}.") :]
             hw.setdefault("observation", {})[name] = ty
 
-        elif full_key.startswith("observation.images."):
+        elif full_key.startswith(f"{OBS_IMAGES}."):
             # observation.images.<cam>
             # images obey ONLY the use_videos flag, not patterns
             if not use_videos:
                 continue
-            name = full_key[len("observation.images.") :]
+            name = full_key[len(f"{OBS_IMAGES}.") :]
             hw.setdefault("observation", {})[name] = ty
 
         else:
@@ -86,8 +87,8 @@ def aggregate_pipeline_dataset_features(
             continue
 
     out: dict[str, dict] = {}
-    if "action" in hw:
-        out.update(hw_to_dataset_features(hw["action"], "action", use_videos))
+    if ACTION in hw:
+        out.update(hw_to_dataset_features(hw[ACTION], ACTION, use_videos))
     if "observation" in hw:
         out.update(hw_to_dataset_features(hw["observation"], "observation", use_videos))
 

src/lerobot/datasets/utils.py

@@ -470,7 +470,7 @@ def dataset_to_policy_features(features: dict[str, dict]) -> dict[str, PolicyFea
     return policy_features
 
 
-def merge_features(*dicts: dict) -> dict:
+def combine_feature_dicts(*dicts: dict) -> dict:
     """
     Merge LeRobot grouped feature dicts.
 

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/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/act/modeling_act.py

@@ -287,7 +287,7 @@ class ACT(nn.Module):
                                 └───────────────────────┘
     """
 
-    def __init__(self, config: ACTConfig, dataset_stats=None):
+    def __init__(self, config: ACTConfig):
         # BERT style VAE encoder with input tokens [cls, robot_state, *action_sequence].
         # The cls token forms parameters of the latent's distribution (like this [*means, *log_variances]).
         super().__init__()

src/lerobot/policies/act/processor_act.py

@@ -15,36 +15,56 @@
 # 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 (
-    DeviceProcessor,
-    NormalizerProcessor,
-    RenameProcessor,
-    RobotProcessor,
-    ToBatchProcessor,
-    UnnormalizerProcessor,
+    AddBatchDimensionProcessorStep,
+    DeviceProcessorStep,
+    NormalizerProcessorStep,
+    PolicyProcessorPipeline,
+    ProcessorKwargs,
+    RenameProcessorStep,
+    UnnormalizerProcessorStep,
 )
 
 
-def make_act_processor(
-    config: ACTConfig, dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None
-) -> tuple[RobotProcessor, RobotProcessor]:
+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]:
+    if preprocessor_kwargs is None:
+        preprocessor_kwargs = {}
+    if postprocessor_kwargs is None:
+        postprocessor_kwargs = {}
+
     input_steps = [
-        RenameProcessor(rename_map={}),
-        NormalizerProcessor(
+        RenameProcessorStep(rename_map={}),
+        NormalizerProcessorStep(
             features={**config.input_features, **config.output_features},
             norm_map=config.normalization_mapping,
             stats=dataset_stats,
         ),
-        ToBatchProcessor(),
-        DeviceProcessor(device=config.device),
+        AddBatchDimensionProcessorStep(),
+        DeviceProcessorStep(device=config.device),
     ]
     output_steps = [
-        DeviceProcessor(device="cpu"),
-        UnnormalizerProcessor(
+        DeviceProcessorStep(device="cpu"),
+        UnnormalizerProcessorStep(
             features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
         ),
     ]
-    return RobotProcessor(steps=input_steps, name="robot_preprocessor"), RobotProcessor(
-        steps=output_steps, name="robot_postprocessor"
+
+    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/processor_diffusion.py

@@ -16,36 +16,55 @@
 # 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 (
-    DeviceProcessor,
-    NormalizerProcessor,
-    RenameProcessor,
-    RobotProcessor,
-    ToBatchProcessor,
-    UnnormalizerProcessor,
+    AddBatchDimensionProcessorStep,
+    DeviceProcessorStep,
+    NormalizerProcessorStep,
+    PolicyProcessorPipeline,
+    ProcessorKwargs,
+    RenameProcessorStep,
+    UnnormalizerProcessorStep,
 )
 
 
-def make_diffusion_processor(
-    config: DiffusionConfig, dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None
-) -> tuple[RobotProcessor, RobotProcessor]:
+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]:
+    if preprocessor_kwargs is None:
+        preprocessor_kwargs = {}
+    if postprocessor_kwargs is None:
+        postprocessor_kwargs = {}
+
     input_steps = [
-        RenameProcessor(rename_map={}),
-        NormalizerProcessor(
+        RenameProcessorStep(rename_map={}),
+        NormalizerProcessorStep(
             features={**config.input_features, **config.output_features},
             norm_map=config.normalization_mapping,
             stats=dataset_stats,
         ),
-        ToBatchProcessor(),
-        DeviceProcessor(device=config.device),
+        AddBatchDimensionProcessorStep(),
+        DeviceProcessorStep(device=config.device),
     ]
     output_steps = [
-        DeviceProcessor(device="cpu"),
-        UnnormalizerProcessor(
+        DeviceProcessorStep(device="cpu"),
+        UnnormalizerProcessorStep(
             features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
         ),
     ]
-    return RobotProcessor(steps=input_steps, name="robot_preprocessor"), RobotProcessor(
-        steps=output_steps, name="robot_postprocessor"
+    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

@@ -17,14 +17,14 @@
 from __future__ import annotations
 
 import logging
-from typing import Any, TypedDict, cast
+from typing import Any, TypedDict
 
 import torch
-from torch import nn
 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
@@ -39,7 +39,7 @@ 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.pipeline import RobotProcessor
+from lerobot.processor import PolicyProcessorPipeline, ProcessorKwargs
 
 
 def get_policy_class(name: str) -> type[PreTrainedPolicy]:
@@ -115,13 +115,15 @@ class ProcessorConfigKwargs(TypedDict, total=False):
     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_processor(
+def make_pre_post_processors(
     policy_cfg: PreTrainedConfig,
     pretrained_path: str | None = None,
     **kwargs: Unpack[ProcessorConfigKwargs],
-) -> tuple[RobotProcessor, RobotProcessor]:
+) -> tuple[PolicyProcessorPipeline, PolicyProcessorPipeline]:
     """Make a processor instance for a given policy type.
 
     This function creates the appropriate processor configuration based on the policy type.
@@ -140,84 +142,120 @@ def make_processor(
         NotImplementedError: If the policy type doesn't have a processor implemented.
     """
     if pretrained_path:
-        # Load a pretrained processor
-        # TODO(azouitine): Handle this case.
+        # Extract preprocessor and postprocessor kwargs
+        preprocessor_kwargs = kwargs.get("preprocessor_kwargs", {})
+        postprocessor_kwargs = kwargs.get("postprocessor_kwargs", {})
+
         return (
-            RobotProcessor.from_pretrained(
+            PolicyProcessorPipeline.from_pretrained(
                 pretrained_model_name_or_path=pretrained_path,
-                config_filename=kwargs.get("preprocessor_config_filename", "robot_preprocessor.json"),
+                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"),
             ),
-            RobotProcessor.from_pretrained(
+            PolicyProcessorPipeline.from_pretrained(
                 pretrained_model_name_or_path=pretrained_path,
-                config_filename=kwargs.get("postprocessor_config_filename", "robot_postprocessor.json"),
+                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 policy_cfg.type == "tdmpc":
-        from lerobot.policies.tdmpc.configuration_tdmpc import TDMPCConfig
-        from lerobot.policies.tdmpc.processor_tdmpc import make_tdmpc_processor
+    if isinstance(policy_cfg, TDMPCConfig):
+        from lerobot.policies.tdmpc.processor_tdmpc import make_tdmpc_pre_post_processors
 
-        processors = make_tdmpc_processor(
-            config=cast(TDMPCConfig, policy_cfg), dataset_stats=kwargs.get("dataset_stats")
+        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 policy_cfg.type == "diffusion":
-        from lerobot.policies.diffusion.processor_diffusion import make_diffusion_processor
+    elif isinstance(policy_cfg, DiffusionConfig):
+        from lerobot.policies.diffusion.processor_diffusion import make_diffusion_pre_post_processors
 
-        processors = make_diffusion_processor(
-            cast(DiffusionConfig, policy_cfg), dataset_stats=kwargs.get("dataset_stats")
+        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 policy_cfg.type == "act":
-        from lerobot.policies.act.processor_act import make_act_processor
+    elif isinstance(policy_cfg, ACTConfig):
+        from lerobot.policies.act.processor_act import make_act_pre_post_processors
 
-        processors = make_act_processor(
-            config=cast(ACTConfig, policy_cfg), dataset_stats=kwargs.get("dataset_stats")
+        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 policy_cfg.type == "vqbet":
-        from lerobot.policies.vqbet.processor_vqbet import make_vqbet_processor
+    elif isinstance(policy_cfg, VQBeTConfig):
+        from lerobot.policies.vqbet.processor_vqbet import make_vqbet_pre_post_processors
 
-        processors = make_vqbet_processor(
-            config=cast(VQBeTConfig, policy_cfg), dataset_stats=kwargs.get("dataset_stats")
+        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 policy_cfg.type == "pi0":
-        from lerobot.policies.pi0.processor_pi0 import make_pi0_processor
+    elif isinstance(policy_cfg, PI0Config):
+        from lerobot.policies.pi0.processor_pi0 import make_pi0_pre_post_processors
 
-        processors = make_pi0_processor(
-            config=cast(PI0Config, policy_cfg), dataset_stats=kwargs.get("dataset_stats")
+        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 policy_cfg.type == "pi0fast":
-        from lerobot.policies.pi0fast.processor_pi0fast import make_pi0fast_processor
+    elif isinstance(policy_cfg, PI0FASTConfig):
+        from lerobot.policies.pi0fast.processor_pi0fast import make_pi0fast_pre_post_processors
 
-        processors = make_pi0fast_processor(
-            cast(PI0Config, policy_cfg), dataset_stats=kwargs.get("dataset_stats")
+        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 policy_cfg.type == "sac":
-        from lerobot.policies.sac.processor_sac import make_sac_processor
+    elif isinstance(policy_cfg, SACConfig):
+        from lerobot.policies.sac.processor_sac import make_sac_pre_post_processors
 
-        processors = make_sac_processor(
-            cast(SACConfig, policy_cfg), dataset_stats=kwargs.get("dataset_stats")
+        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 policy_cfg.type == "reward_classifier":
+    elif isinstance(policy_cfg, RewardClassifierConfig):
         from lerobot.policies.sac.reward_model.processor_classifier import make_classifier_processor
 
         processors = make_classifier_processor(
-            cast(RewardClassifierConfig, policy_cfg), dataset_stats=kwargs.get("dataset_stats")
+            config=policy_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+            preprocessor_kwargs=kwargs.get("preprocessor_kwargs"),
+            postprocessor_kwargs=kwargs.get("postprocessor_kwargs"),
         )
 
-    elif policy_cfg.type == "smolvla":
-        from lerobot.policies.smolvla.processor_smolvla import make_smolvla_processor
+    elif isinstance(policy_cfg, SmolVLAConfig):
+        from lerobot.policies.smolvla.processor_smolvla import make_smolvla_pre_post_processors
 
-        processors = make_smolvla_processor(
-            cast(SmolVLAConfig, policy_cfg), dataset_stats=kwargs.get("dataset_stats")
+        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:
@@ -297,7 +335,7 @@ 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
 ```

src/lerobot/policies/pi0/processor_pi0.py

@@ -14,107 +14,105 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import Any
 
 import torch
 
 from lerobot.configs.types import 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 (
-    DeviceProcessor,
-    NormalizerProcessor,
-    RobotProcessor,
-    ToBatchProcessor,
-    TokenizerProcessor,
-    UnnormalizerProcessor,
-)
-from lerobot.processor.pipeline import (
-    EnvTransition,
+    AddBatchDimensionProcessorStep,
+    ComplementaryDataProcessorStep,
+    DeviceProcessorStep,
+    NormalizerProcessorStep,
+    PolicyProcessorPipeline,
+    ProcessorKwargs,
     ProcessorStep,
     ProcessorStepRegistry,
-    TransitionKey,
+    RenameProcessorStep,
+    TokenizerProcessorStep,
+    UnnormalizerProcessorStep,
 )
-from lerobot.processor.rename_processor import RenameProcessor
 
 
 @ProcessorStepRegistry.register(name="pi0_new_line_processor")
-class Pi0NewLineProcessor(ProcessorStep):
+class Pi0NewLineProcessor(ComplementaryDataProcessorStep):
     """Add a new line to the end of the task if it doesn't have one.
     This is required for the PaliGemma tokenizer.
     """
 
-    def __call__(self, transition: EnvTransition) -> EnvTransition:
-        # Check if complementary_data exists
-        complementary_data = transition.get(TransitionKey.COMPLEMENTARY_DATA)
-        if complementary_data is None or "task" not in complementary_data:
-            return transition
+    def complementary_data(self, complementary_data):
+        if "task" not in complementary_data:
+            return complementary_data
 
         task = complementary_data["task"]
         if task is None:
-            return transition
+            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"):
-                complementary_data["task"] = f"{task}\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
-            complementary_data["task"] = [t if t.endswith("\n") else f"{t}\n" for t in task]
+            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 transition
+        return new_complementary_data
 
     def transform_features(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
-        """Add tokenized task features to the features."""
         return features
 
-    def state_dict(self) -> dict[str, torch.Tensor]:
-        """Return state dictionary (empty for this processor)."""
-        return {}
 
-    def load_state_dict(self, state: dict[str, torch.Tensor]) -> None:
-        """Load state dictionary (no-op for this processor)."""
-        pass
+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]:
+    if preprocessor_kwargs is None:
+        preprocessor_kwargs = {}
+    if postprocessor_kwargs is None:
+        postprocessor_kwargs = {}
 
-    def reset(self) -> None:
-        """Reset processor state (no-op for this processor)."""
-        pass
-
-    def get_config(self) -> dict[str, Any]:
-        """Return configuration for serialization."""
-        return {}
-
-
-def make_pi0_processor(
-    config: PI0Config, dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None
-) -> tuple[RobotProcessor, RobotProcessor]:
     # Add remaining processors
     input_steps: list[ProcessorStep] = [
-        RenameProcessor(rename_map={}),  # To mimic the same processor as pretrained one
-        NormalizerProcessor(
+        RenameProcessorStep(rename_map={}),  # To mimic the same processor as pretrained one
+        NormalizerProcessorStep(
             features={**config.input_features, **config.output_features},
             norm_map=config.normalization_mapping,
             stats=dataset_stats,
         ),
-        ToBatchProcessor(),
+        AddBatchDimensionProcessorStep(),
         Pi0NewLineProcessor(),  # Add newlines before tokenization for PaliGemma
-        TokenizerProcessor(
+        TokenizerProcessorStep(
             tokenizer_name="google/paligemma-3b-pt-224",
             max_length=config.tokenizer_max_length,
             padding_side="right",
             padding="max_length",
         ),
-        DeviceProcessor(device=config.device),
+        DeviceProcessorStep(device=config.device),
     ]
 
     output_steps: list[ProcessorStep] = [
-        DeviceProcessor(device="cpu"),
-        UnnormalizerProcessor(
+        DeviceProcessorStep(device="cpu"),
+        UnnormalizerProcessorStep(
             features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
         ),
     ]
 
-    return RobotProcessor(steps=input_steps, name="robot_preprocessor"), RobotProcessor(
-        steps=output_steps, name="robot_postprocessor"
+    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
 ```

src/lerobot/policies/pi0fast/processor_pi0fast.py

@@ -16,36 +16,55 @@
 
 import torch
 
+from lerobot.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
 from lerobot.policies.pi0.configuration_pi0 import PI0Config
 from lerobot.processor import (
-    DeviceProcessor,
-    NormalizerProcessor,
-    RenameProcessor,
-    RobotProcessor,
-    ToBatchProcessor,
-    UnnormalizerProcessor,
+    AddBatchDimensionProcessorStep,
+    DeviceProcessorStep,
+    NormalizerProcessorStep,
+    PolicyProcessorPipeline,
+    ProcessorKwargs,
+    RenameProcessorStep,
+    UnnormalizerProcessorStep,
 )
 
 
-def make_pi0fast_processor(
-    config: PI0Config, dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None
-) -> tuple[RobotProcessor, RobotProcessor]:
+def make_pi0fast_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]:
+    if preprocessor_kwargs is None:
+        preprocessor_kwargs = {}
+    if postprocessor_kwargs is None:
+        postprocessor_kwargs = {}
+
     input_steps = [
-        RenameProcessor(rename_map={}),  # To mimic the same processor as pretrained one
-        NormalizerProcessor(
+        RenameProcessorStep(rename_map={}),  # To mimic the same processor as pretrained one
+        NormalizerProcessorStep(
             features={**config.input_features, **config.output_features},
             norm_map=config.normalization_mapping,
             stats=dataset_stats,
         ),
-        ToBatchProcessor(),
-        DeviceProcessor(device=config.device),
+        AddBatchDimensionProcessorStep(),
+        DeviceProcessorStep(device=config.device),
     ]
     output_steps = [
-        DeviceProcessor(device="cpu"),
-        UnnormalizerProcessor(
+        DeviceProcessorStep(device="cpu"),
+        UnnormalizerProcessorStep(
             features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
         ),
     ]
-    return RobotProcessor(steps=input_steps, name="robot_preprocessor"), RobotProcessor(
-        steps=output_steps, name="robot_postprocessor"
+    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/processor_sac.py

@@ -17,36 +17,55 @@
 
 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 (
-    DeviceProcessor,
-    NormalizerProcessor,
-    RenameProcessor,
-    RobotProcessor,
-    ToBatchProcessor,
-    UnnormalizerProcessor,
+    AddBatchDimensionProcessorStep,
+    DeviceProcessorStep,
+    NormalizerProcessorStep,
+    PolicyProcessorPipeline,
+    ProcessorKwargs,
+    RenameProcessorStep,
+    UnnormalizerProcessorStep,
 )
 
 
-def make_sac_processor(
-    config: SACConfig, dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None
-) -> tuple[RobotProcessor, RobotProcessor]:
+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]:
+    if preprocessor_kwargs is None:
+        preprocessor_kwargs = {}
+    if postprocessor_kwargs is None:
+        postprocessor_kwargs = {}
+
     input_steps = [
-        RenameProcessor(rename_map={}),
-        NormalizerProcessor(
+        RenameProcessorStep(rename_map={}),
+        NormalizerProcessorStep(
             features={**config.input_features, **config.output_features},
             norm_map=config.normalization_mapping,
             stats=dataset_stats,
         ),
-        ToBatchProcessor(),
-        DeviceProcessor(device=config.device),
+        AddBatchDimensionProcessorStep(),
+        DeviceProcessorStep(device=config.device),
     ]
     output_steps = [
-        DeviceProcessor(device="cpu"),
-        UnnormalizerProcessor(
+        DeviceProcessorStep(device="cpu"),
+        UnnormalizerProcessorStep(
             features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
         ),
     ]
-    return RobotProcessor(steps=input_steps, name="robot_preprocessor"), RobotProcessor(
-        steps=output_steps, name="robot_postprocessor"
+    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/processor_classifier.py

@@ -17,26 +17,45 @@ import torch
 
 from lerobot.policies.sac.reward_model.configuration_classifier import RewardClassifierConfig
 from lerobot.processor import (
-    DeviceProcessor,
-    IdentityProcessor,
-    NormalizerProcessor,
-    RobotProcessor,
+    DeviceProcessorStep,
+    IdentityProcessorStep,
+    NormalizerProcessorStep,
+    PolicyProcessorPipeline,
+    ProcessorKwargs,
 )
 
 
 def make_classifier_processor(
-    config: RewardClassifierConfig, dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None
-) -> tuple[RobotProcessor, RobotProcessor]:
+    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]:
+    if preprocessor_kwargs is None:
+        preprocessor_kwargs = {}
+    if postprocessor_kwargs is None:
+        postprocessor_kwargs = {}
+
     input_steps = [
-        NormalizerProcessor(
+        NormalizerProcessorStep(
             features=config.input_features, norm_map=config.normalization_mapping, stats=dataset_stats
         ),
-        NormalizerProcessor(
+        NormalizerProcessorStep(
             features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
         ),
-        DeviceProcessor(device=config.device),
+        DeviceProcessorStep(device=config.device),
     ]
-    output_steps = [DeviceProcessor(device="cpu"), IdentityProcessor()]
-    return RobotProcessor(steps=input_steps, name="classifier_preprocessor"), RobotProcessor(
-        steps=output_steps, name="classifier_postprocessor"
+    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 \

src/lerobot/policies/smolvla/processor_smolvla.py

@@ -13,97 +13,99 @@
 # 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 typing import Any
 
 import torch
 
 from lerobot.configs.types import 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 (
-    DeviceProcessor,
-    NormalizerProcessor,
-    RenameProcessor,
-    RobotProcessor,
-    ToBatchProcessor,
-    TokenizerProcessor,
-    UnnormalizerProcessor,
+    AddBatchDimensionProcessorStep,
+    ComplementaryDataProcessorStep,
+    DeviceProcessorStep,
+    NormalizerProcessorStep,
+    PolicyProcessorPipeline,
+    ProcessorKwargs,
+    ProcessorStepRegistry,
+    RenameProcessorStep,
+    TokenizerProcessorStep,
+    UnnormalizerProcessorStep,
 )
-from lerobot.processor.pipeline import EnvTransition, ProcessorStep, ProcessorStepRegistry, TransitionKey
 
 
-def make_smolvla_processor(
-    config: SmolVLAConfig, dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None
-) -> tuple[RobotProcessor, RobotProcessor]:
+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]:
+    if preprocessor_kwargs is None:
+        preprocessor_kwargs = {}
+    if postprocessor_kwargs is None:
+        postprocessor_kwargs = {}
+
     input_steps = [
-        RenameProcessor(rename_map={}),  # To mimic the same processor as pretrained one
-        NormalizerProcessor(
+        RenameProcessorStep(rename_map={}),  # To mimic the same processor as pretrained one
+        NormalizerProcessorStep(
             features={**config.input_features, **config.output_features},
             norm_map=config.normalization_mapping,
             stats=dataset_stats,
         ),
-        ToBatchProcessor(),
+        AddBatchDimensionProcessorStep(),
         SmolVLANewLineProcessor(),
-        TokenizerProcessor(
+        TokenizerProcessorStep(
             tokenizer_name=config.vlm_model_name,
             padding=config.pad_language_to,
             padding_side="right",
             max_length=config.tokenizer_max_length,
         ),
-        DeviceProcessor(device=config.device),
+        DeviceProcessorStep(device=config.device),
     ]
     output_steps = [
-        DeviceProcessor(device="cpu"),
-        UnnormalizerProcessor(
+        DeviceProcessorStep(device="cpu"),
+        UnnormalizerProcessorStep(
             features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
         ),
     ]
-    return RobotProcessor(steps=input_steps, name="robot_preprocessor"), RobotProcessor(
-        steps=output_steps, name="robot_postprocessor"
+    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(ProcessorStep):
+class SmolVLANewLineProcessor(ComplementaryDataProcessorStep):
     """Add a new line to the end of the task if it doesn't have one."""
 
-    def __call__(self, transition: EnvTransition) -> EnvTransition:
-        # Check if complementary_data exists
-        complementary_data = transition.get(TransitionKey.COMPLEMENTARY_DATA)
-        if complementary_data is None or "task" not in complementary_data:
-            return transition
+    def complementary_data(self, complementary_data):
+        if "task" not in complementary_data:
+            return complementary_data
 
         task = complementary_data["task"]
         if task is None:
-            return transition
+            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"):
-                complementary_data["task"] = f"{task}\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
-            complementary_data["task"] = [t if t.endswith("\n") else f"{t}\n" for t in task]
+            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 transition
+        return new_complementary_data
 
     def transform_features(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
-        """Adds nothing to the features."""
         return features
-
-    def state_dict(self) -> dict[str, torch.Tensor]:
-        """Return state dictionary (empty for this processor)."""
-        return {}
-
-    def load_state_dict(self, state: dict[str, torch.Tensor]) -> None:
-        """Load state dictionary (no-op for this processor)."""
-        pass
-
-    def reset(self) -> None:
-        """Reset processor state (no-op for this processor)."""
-        pass
-
-    def get_config(self) -> dict[str, Any]:
-        """Return configuration for serialization."""
-        return {}

src/lerobot/policies/tdmpc/processor_tdmpc.py

@@ -16,36 +16,55 @@
 # 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 (
-    DeviceProcessor,
-    NormalizerProcessor,
-    RenameProcessor,
-    RobotProcessor,
-    ToBatchProcessor,
-    UnnormalizerProcessor,
+    AddBatchDimensionProcessorStep,
+    DeviceProcessorStep,
+    NormalizerProcessorStep,
+    PolicyProcessorPipeline,
+    ProcessorKwargs,
+    RenameProcessorStep,
+    UnnormalizerProcessorStep,
 )
 
 
-def make_tdmpc_processor(
-    config: TDMPCConfig, dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None
-) -> tuple[RobotProcessor, RobotProcessor]:
+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]:
+    if preprocessor_kwargs is None:
+        preprocessor_kwargs = {}
+    if postprocessor_kwargs is None:
+        postprocessor_kwargs = {}
+
     input_steps = [
-        RenameProcessor(rename_map={}),
-        NormalizerProcessor(
+        RenameProcessorStep(rename_map={}),
+        NormalizerProcessorStep(
             features={**config.input_features, **config.output_features},
             norm_map=config.normalization_mapping,
             stats=dataset_stats,
         ),
-        ToBatchProcessor(),
-        DeviceProcessor(device=config.device),
+        AddBatchDimensionProcessorStep(),
+        DeviceProcessorStep(device=config.device),
     ]
     output_steps = [
-        DeviceProcessor(device="cpu"),
-        UnnormalizerProcessor(
+        DeviceProcessorStep(device="cpu"),
+        UnnormalizerProcessorStep(
             features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
         ),
     ]
-    return RobotProcessor(steps=input_steps, name="robot_preprocessor"), RobotProcessor(
-        steps=output_steps, name="robot_postprocessor"
+    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/processor_vqbet.py

@@ -17,36 +17,55 @@
 # 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 (
-    DeviceProcessor,
-    NormalizerProcessor,
-    RenameProcessor,
-    RobotProcessor,
-    ToBatchProcessor,
-    UnnormalizerProcessor,
+    AddBatchDimensionProcessorStep,
+    DeviceProcessorStep,
+    NormalizerProcessorStep,
+    PolicyProcessorPipeline,
+    ProcessorKwargs,
+    RenameProcessorStep,
+    UnnormalizerProcessorStep,
 )
 
 
-def make_vqbet_processor(
-    config: VQBeTConfig, dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None
-) -> tuple[RobotProcessor, RobotProcessor]:
+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]:
+    if preprocessor_kwargs is None:
+        preprocessor_kwargs = {}
+    if postprocessor_kwargs is None:
+        postprocessor_kwargs = {}
+
     input_steps = [
-        RenameProcessor(rename_map={}),  # Let the possibility to the user to rename the keys
-        NormalizerProcessor(
+        RenameProcessorStep(rename_map={}),  # Let the possibility to the user to rename the keys
+        NormalizerProcessorStep(
             features={**config.input_features, **config.output_features},
             norm_map=config.normalization_mapping,
             stats=dataset_stats,
         ),
-        ToBatchProcessor(),
-        DeviceProcessor(device=config.device),
+        AddBatchDimensionProcessorStep(),
+        DeviceProcessorStep(device=config.device),
     ]
     output_steps = [
-        DeviceProcessor(device="cpu"),
-        UnnormalizerProcessor(
+        DeviceProcessorStep(device="cpu"),
+        UnnormalizerProcessorStep(
             features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
         ),
     ]
-    return RobotProcessor(steps=input_steps, name="robot_preprocessor"), RobotProcessor(
-        steps=output_steps, name="robot_postprocessor"
+    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

@@ -14,46 +14,90 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from .batch_processor import ToBatchProcessor
-from .device_processor import DeviceProcessor
-from .normalize_processor import NormalizerProcessor, UnnormalizerProcessor, hotswap_stats
-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 .tokenizer_processor import TokenizerProcessor
+from .rename_processor import RenameProcessorStep
+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",
+    "GripperPenaltyProcessorStep",
     "hotswap_stats",
-    "ObservationProcessor",
+    "IdentityProcessorStep",
+    "ImageCropResizeProcessorStep",
+    "InfoProcessorStep",
+    "InterventionActionProcessorStep",
+    "JointVelocityProcessorStep",
+    "MapDeltaActionToRobotActionStep",
+    "MapTensorToDeltaActionDictStep",
+    "merge_transitions",
+    "MotorCurrentProcessorStep",
+    "NormalizerProcessorStep",
+    "Numpy2TorchActionProcessorStep",
+    "ObservationProcessorStep",
+    "PolicyProcessorPipeline",
+    "ProcessorKwargs",
     "ProcessorStep",
     "ProcessorStepRegistry",
-    "RenameProcessor",
-    "RewardProcessor",
-    "RobotProcessor",
-    "ToBatchProcessor",
-    "TokenizerProcessor",
+    "RenameProcessorStep",
+    "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

@@ -11,20 +11,99 @@
 # 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
+from dataclasses import dataclass, field
 
-import torch
 from torch import Tensor
 
 from lerobot.configs.types import PolicyFeature
 from lerobot.constants import OBS_ENV_STATE, OBS_IMAGE, OBS_IMAGES, OBS_STATE
-from lerobot.processor.pipeline import EnvTransition, ProcessorStepRegistry, TransitionKey
+
+from .core import EnvTransition
+from .pipeline import (
+    ActionProcessorStep,
+    ComplementaryDataProcessorStep,
+    ObservationProcessorStep,
+    ProcessorStep,
+    ProcessorStepRegistry,
+)
+
+
+@dataclass
+@ProcessorStepRegistry.register(name="to_batch_processor_action")
+class AddBatchDimensionActionStep(ActionProcessorStep):
+    """Process action component in-place, adding batch dimension if needed."""
+
+    def action(self, action):
+        if not isinstance(action, Tensor) or action.dim() != 1:
+            return action
+
+        return action.unsqueeze(0)
+
+    def transform_features(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
+        return features
+
+
+@dataclass
+@ProcessorStepRegistry.register(name="to_batch_processor_observation")
+class AddBatchDimensionObservationStep(ObservationProcessorStep):
+    """Process observation component in-place, adding batch dimensions where needed."""
+
+    def observation(self, observation):
+        # 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[str, PolicyFeature]) -> dict[str, PolicyFeature]:
+        return features
+
+
+@dataclass
+@ProcessorStepRegistry.register(name="to_batch_processor_complementary_data")
+class AddBatchDimensionComplementaryDataStep(ComplementaryDataProcessorStep):
+    """Process complementary data in-place, handling task field batching."""
+
+    def complementary_data(self, complementary_data):
+        # 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[str, PolicyFeature]) -> dict[str, PolicyFeature]:
+        return features
 
 
 @dataclass
 @ProcessorStepRegistry.register(name="to_batch_processor")
-class ToBatchProcessor:
+class AddBatchDimensionProcessorStep(ProcessorStep):
     """Processor that adds batch dimensions to observations and actions when needed.
 
     This processor ensures that observations and actions have proper batch dimensions for model processing:
@@ -59,81 +138,22 @@ class ToBatchProcessor:
         ```
     """
 
+    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:
-        self._process_observation(transition)
-        self._process_action(transition)
-        self._process_complementary_data(transition)
+        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 _process_observation(self, transition: EnvTransition) -> None:
-        """Process observation component in-place, adding batch dimensions where needed."""
-        observation = transition.get(TransitionKey.OBSERVATION)
-        if observation is None:
-            return
-
-        # 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)
-
-    def _process_action(self, transition: EnvTransition) -> None:
-        """Process action component in-place, adding batch dimension if needed."""
-        action = transition.get(TransitionKey.ACTION)
-        if action is not None and isinstance(action, Tensor) and action.dim() == 1:
-            transition[TransitionKey.ACTION] = action.unsqueeze(0)
-
-    def _process_complementary_data(self, transition: EnvTransition) -> None:
-        """Process complementary data in-place, handling task field batching."""
-        complementary_data = transition.get(TransitionKey.COMPLEMENTARY_DATA)
-        if complementary_data is None:
-            return
-
-        # 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)
-
-    def get_config(self) -> dict[str, Any]:
-        """Return configuration for serialization."""
-        return {}
-
-    def state_dict(self) -> dict[str, torch.Tensor]:
-        """Return state dictionary (empty for this processor)."""
-        return {}
-
-    def load_state_dict(self, state: dict[str, torch.Tensor]) -> None:
-        """Load state dictionary (no-op for this processor)."""
-        pass
-
-    def reset(self) -> None:
-        """Reset processor state (no-op for this processor)."""
-        pass
-
     def transform_features(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
+        # NOTE: We ignore the batch dimension when transforming features
         return features

src/lerobot/processor/converters.py

@@ -16,31 +16,131 @@
 
 from __future__ import annotations
 
-from collections.abc import Iterable, Sequence
+from collections.abc import Sequence
 from copy import deepcopy
+from functools import singledispatch
 from typing import Any
 
 import numpy as np
 import torch
-from scipy.spatial.transform import Rotation
 
-from .pipeline import EnvTransition, TransitionKey
+from lerobot.constants import ACTION, DONE, OBS_IMAGES, OBS_STATE, REWARD, TRUNCATED
+from lerobot.utils.rotation import Rotation
+
+from .core import EnvTransition, TransitionKey
 
 
-def _to_tensor(x: torch.Tensor | np.ndarray | Sequence[int | float]):
-    if isinstance(x, torch.Tensor):
-        return x
-    if isinstance(x, np.ndarray):
-        # Keep images (uint8 HWC) and python objects as-is
-        if x.dtype == np.uint8 or x.dtype == np.object_:
-            return x
-        # Scalars/arrays to float32 tensor
-        return torch.as_tensor(x, dtype=torch.float32)
-    # Anything else to float32 tensor
-    return torch.as_tensor(x, dtype=torch.float32)
+@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:
+        PyTorch tensor.
+
+    Raises:
+        TypeError: If the input type is not supported.
+    """
+    raise TypeError(f"Unsupported type for tensor conversion: {type(value)}")
 
 
-def _from_tensor(x: Any):
+@to_tensor.register(torch.Tensor)
+def _(value: torch.Tensor, *, dtype=torch.float32, device=None, **kwargs) -> torch.Tensor:
+    """Handle 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 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 (torch.from_numpy handles contiguity automatically)
+    tensor = torch.from_numpy(value)
+
+    # Apply dtype 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 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 sequences (lists, tuples)."""
+    return torch.tensor(value, dtype=dtype, device=device)
+
+
+@to_tensor.register(dict)
+def _(value: dict, *, device=None, **kwargs) -> dict:
+    """Handle dictionaries by recursively converting 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(x: torch.Tensor | Any) -> np.ndarray | float | int | Any:
+    """Convert tensor to numpy/scalar if needed."""
     if isinstance(x, torch.Tensor):
         return x.item() if x.numel() == 1 else x.detach().cpu().numpy()
     return x
@@ -53,28 +153,87 @@ def _is_image(arr: Any) -> bool:
 def _split_obs_to_state_and_images(obs: dict[str, Any]) -> tuple[dict[str, Any], dict[str, Any]]:
     state, images = {}, {}
     for k, v in obs.items():
-        if _is_image(v):
+        if "image" in k.lower() or _is_image(v):
             images[k] = v
         else:
             state[k] = v
     return state, images
 
 
-def make_obs_act_transition(
-    *, obs: dict[str, Any] | None = None, act: dict[str, Any] | None = None
+# ============================================================================
+# Private Helper Functions (Common Logic)
+# ============================================================================
+
+
+def _extract_complementary_data(batch: dict[str, Any]) -> dict[str, Any]:
+    """Extract complementary data (pad flags, task, index, task_index)."""
+    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 other taking precedence."""
+    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 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:
+        Complete EnvTransition dictionary.
+    """
     return {
-        TransitionKey.OBSERVATION: {} if obs is None else obs,
-        TransitionKey.ACTION: {} if act is None else act,
-        TransitionKey.INFO: {},
-        TransitionKey.COMPLEMENTARY_DATA: {},
-        TransitionKey.REWARD: None,
-        TransitionKey.DONE: None,
-        TransitionKey.TRUNCATED: None,
+        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 to_transition_teleop_action(action: dict[str, Any]) -> EnvTransition:
+def action_to_transition(action: dict[str, Any]) -> EnvTransition:  # action_to_transition
     """
     Convert a raw teleop action dict into an EnvTransition under the ACTION TransitionKey.
     """
@@ -82,17 +241,17 @@ def to_transition_teleop_action(action: dict[str, Any]) -> EnvTransition:
     for k, v in action.items():
         # Check if the value is a type that should not be converted to a tensor.
         if isinstance(v, (Rotation, dict)):
-            act_dict[f"action.{k}"] = v
+            act_dict[f"{ACTION}.{k}"] = v
             continue
 
         arr = np.array(v) if np.isscalar(v) else v
-        act_dict[f"action.{k}"] = _to_tensor(arr)
+        act_dict[f"{ACTION}.{k}"] = to_tensor(arr)
 
-    return make_obs_act_transition(act=act_dict)
+    return create_transition(observation={}, action=act_dict)
 
 
 # TODO(Adil, Pepijn): Overtime we can maybe add these converters to pipeline.py itself
-def to_transition_robot_observation(observation: dict[str, Any]) -> EnvTransition:
+def observation_to_transition(observation: dict[str, Any]) -> EnvTransition:
     """
     Convert a raw robot observation dict into an EnvTransition under the OBSERVATION TransitionKey.
     """
@@ -101,92 +260,87 @@ def to_transition_robot_observation(observation: dict[str, Any]) -> EnvTransitio
     obs_dict: dict[str, Any] = {}
     for k, v in state.items():
         arr = np.array(v) if np.isscalar(v) else v
-        obs_dict[f"observation.state.{k}"] = _to_tensor(arr)
+        obs_dict[f"{OBS_STATE}.{k}"] = to_tensor(arr)
 
     for cam, img in images.items():
-        obs_dict[f"observation.images.{cam}"] = img
+        obs_dict[f"{OBS_IMAGES}.{cam}"] = img
 
-    return make_obs_act_transition(obs=obs_dict)
+    return create_transition(observation=obs_dict, action={})
 
 
-def to_output_robot_action(transition: EnvTransition) -> dict[str, Any]:
+def transition_to_robot_action(transition: EnvTransition) -> dict[str, Any]:
     """
     Converts a EnvTransition under the ACTION TransitionKey to a dict with keys ending in '.pos' for raw robot actions.
     """
     out: dict[str, Any] = {}
     action_dict = transition.get(TransitionKey.ACTION) or {}
 
+    if action_dict is None:
+        return out
+
     for k, v in action_dict.items():
-        if isinstance(k, str) and k.startswith("action.") and k.endswith((".pos", ".vel")):
-            out_key = k[len("action.") :]  # Strip the 'action.' prefix.
+        if isinstance(k, str) and k.startswith(f"{ACTION}.") and k.endswith((".pos", ".vel")):
+            out_key = k[len(f"{ACTION}.") :]  # Strip the 'action.' prefix.
             out[out_key] = float(v)
 
     return out
 
 
-def to_dataset_frame(
-    transitions_or_transition: EnvTransition | Iterable[EnvTransition], features: dict[str, dict]
-) -> dict[str, any]:
-    """
-    Converts a single EnvTransition or an iterable of them into a flat,
-    dataset-friendly dictionary for training or evaluation, according to
-    the provided `features` spec.
+def merge_transitions(transitions: Sequence[EnvTransition] | EnvTransition) -> EnvTransition:
+    """Merge multiple transitions or return single transition.
 
     Args:
-        transitions_or_transition: Either a single EnvTransition dict
-            or an iterable of them (which will be merged).
-        features (dict[str, dict]):
-            A feature specification dictionary:
-              - 'action': dict with 'names': list of action feature names
-              - 'observation.state': dict with 'names': list of state feature names
-              - keys starting with 'observation.images.' are passed through
+        transitions: Either a single transition or iterable of transitions.
 
     Returns:
-        batch (dict[str, any]): Flat dictionary containing:
-          - numpy arrays for "observation.state" and "action"
-          - any image tensors defined in features
-          - next.{reward,done,truncated}
-          - info dict
-          - *_is_pad flags and task from complementary_data
+        Merged EnvTransition.
     """
-    action_names = features.get("action", {}).get("names", [])
-    obs_state_names = features.get("observation.state", {}).get("names", [])
-    image_keys = [k for k in features if k.startswith("observation.images.")]
 
-    def _merge(base: EnvTransition, other: EnvTransition) -> EnvTransition:
-        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
+    if not isinstance(transitions, Sequence):  # Single transition
+        return transitions
 
-    def _ensure_transition(obj) -> EnvTransition:
-        # single transition
-        if isinstance(obj, dict) and any(isinstance(k, TransitionKey) for k in obj):
-            return obj
-        # iterable of transitions
-        if isinstance(obj, Iterable):
-            items = list(obj)
-            if not items:
-                return {}
-            acc = items[0]
-            for t in items[1:]:
-                acc = _merge(acc, t)
-            return acc
-        raise TypeError("Expected EnvTransition or iterable of them")
+    items = list(transitions)
+    if not items:
+        raise ValueError("merge_transitions() requires a non-empty sequence of transitions")
 
-    tr = _ensure_transition(transitions_or_transition)
+    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 a single EnvTransition or an iterable of them into a flat, dataset-friendly dictionary for training or evaluation.
+
+    Processes transitions according to the provided feature specification and returns
+    data in the format expected by machine learning models and datasets.
+
+    Args:
+        transitions_or_transition: Either a single EnvTransition dict or an iterable of them
+            (which will be merged using merge_transitions).
+        features: Feature specification dictionary with the following structure:
+            - 'action': dict with 'names': list of action feature names
+            - 'observation.state': dict with 'names': list of state feature names
+            - keys starting with 'observation.images.' are passed through as-is
+
+    Returns:
+        Flat dictionary containing:
+        - numpy arrays for "observation.state" and "action" (vectorized from feature names)
+        - any image tensors defined in features (passed through unchanged)
+        - next.{reward,done,truncated} scalar values
+        - info dict
+        - *_is_pad flags and task from complementary_data
+    """
+    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] = {}
+    batch: dict[str, Any] = {}
 
     # Images passthrough
     for k in image_keys:
@@ -195,21 +349,36 @@ def to_dataset_frame(
 
     # Observation.state vector
     if obs_state_names:
-        vals = [_from_tensor(obs.get(f"observation.state.{n}", 0.0)) for n in obs_state_names]
-        batch["observation.state"] = np.asarray(vals, dtype=np.float32)
+        vals = [_from_tensor(obs.get(f"{OBS_STATE}.{n}", 0.0)) for n in obs_state_names]
+        batch[OBS_STATE] = np.asarray(vals, dtype=np.float32)
 
     # Action vector
     if action_names:
-        vals = [_from_tensor(act.get(f"action.{n}", 0.0)) for n in action_names]
-        batch["action"] = np.asarray(vals, dtype=np.float32)
+        vals = [_from_tensor(act.get(f"{ACTION}.{n}", 0.0)) for n in action_names]
+        batch[ACTION] = np.asarray(vals, dtype=np.float32)
 
-    # Next.* fields
+    # Add transition metadata
     if tr.get(TransitionKey.REWARD) is not None:
-        batch["next.reward"] = _from_tensor(tr[TransitionKey.REWARD])
+        reward_val = _from_tensor(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:
-        batch["next.done"] = _from_tensor(tr[TransitionKey.DONE])
+        done_val = _from_tensor(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:
-        batch["next.truncated"] = _from_tensor(tr[TransitionKey.TRUNCATED])
+        truncated_val = _from_tensor(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
 
     # Complementary data flags and task
     comp = tr.get(TransitionKey.COMPLEMENTARY_DATA) or {}
@@ -223,3 +392,90 @@ def to_dataset_frame(
             batch["task"] = comp["task"]
 
     return batch
+
+
+def batch_to_transition(batch: dict[str, Any]) -> EnvTransition:
+    """Convert a batch dict coming from LeRobot replay/dataset code into an EnvTransition dictionary.
+
+    The function maps well known keys to the EnvTransition structure. Missing keys are
+    filled with sane defaults (None or 0.0/False).
+
+    Keys recognised (case-sensitive):
+    * "observation.*" (keys starting with "observation." are grouped into observation dict)
+    * "action"
+    * "next.reward"
+    * "next.done"
+    * "next.truncated"
+    * "info"
+    * "_is_pad" patterns (padding flags)
+    * "task", "index", "task_index" (complementary data)
+
+    Additional keys are ignored so that existing dataloaders can carry extra
+    metadata without breaking the processor.
+
+    Args:
+        batch: Batch dictionary from datasets or dataloaders containing the above keys.
+
+    Returns:
+        EnvTransition dictionary with properly structured transition data.
+    """
+
+    # Validate input type
+    if not isinstance(batch, dict):
+        raise ValueError(f"EnvTransition must be a dictionary. Got {type(batch).__name__}")
+
+    # Extract observation 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]:
+    """Inverse of batch_to_transition. Returns a dict with canonical field names used throughout LeRobot.
+
+    Converts an EnvTransition back to the batch format expected by datasets, dataloaders,
+    and other LeRobot components.
+
+    Output format:
+    * "action": Action data from transition
+    * "next.reward": Reward value (defaults to 0.0)
+    * "next.done": Done flag (defaults to False)
+    * "next.truncated": Truncated flag (defaults to False)
+    * "info": Info dictionary (defaults to {})
+    * Flattened observation keys (e.g., "observation.state", "observation.images.cam1")
+    * Complementary data fields ("task", "index", "task_index", padding flags)
+
+    Args:
+        transition: EnvTransition dictionary to convert.
+
+    Returns:
+        Batch dictionary with canonical LeRobot field names suitable for dataloaders.
+    """
+    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 dict
+    observation = transition.get(TransitionKey.OBSERVATION)
+    if isinstance(observation, dict):
+        batch.update(observation)
+
+    return batch

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 +1,145 @@
+# !/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, PolicyFeature
+from lerobot.constants import ACTION
+
+from .pipeline import ActionProcessorStep, ProcessorStepRegistry
+
+
+@ProcessorStepRegistry.register("map_tensor_to_delta_action_dict")
+@dataclass
+class MapTensorToDeltaActionDictStep(ActionProcessorStep):
+    """
+    Map a tensor to a delta action dictionary.
+    """
+
+    use_gripper: bool = True
+
+    def action(self, action: Tensor) -> dict:
+        if action.dim() > 1:
+            action = action.squeeze(0)
+
+        # TODO (maractingi): add rotation
+        delta_action = {
+            f"{ACTION}.delta_x": action[0],
+            f"{ACTION}.delta_y": action[1],
+            f"{ACTION}.delta_z": action[2],
+        }
+        if self.use_gripper:
+            delta_action[f"{ACTION}.gripper"] = action[3]
+        return delta_action
+
+    def transform_features(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
+        features[f"{ACTION}.delta_x"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        features[f"{ACTION}.delta_y"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        features[f"{ACTION}.delta_z"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        if self.use_gripper:
+            features[f"{ACTION}.gripper"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        return features
+
+
+@ProcessorStepRegistry.register("map_delta_action_to_robot_action")
+@dataclass
+class MapDeltaActionToRobotActionStep(ActionProcessorStep):
+    """
+    Map delta actions from teleoperators (gamepad, keyboard) to robot target actions
+    for use with inverse kinematics processors.
+
+    Expected input ACTION keys:
+    {
+        "action.delta_x": float,
+        "action.delta_y": float,
+        "action.delta_z": float,
+        "action.gripper": float (optional),
+    }
+
+    Output ACTION keys:
+    {
+        "action.enabled": bool,
+        "action.target_x": float,
+        "action.target_y": float,
+        "action.target_z": float,
+        "action.target_wx": float,
+        "action.target_wy": float,
+        "action.target_wz": float,
+        "action.gripper": float,
+    }
+    """
+
+    # 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(f"{ACTION}.delta_x", 0.0)
+        delta_y = action.pop(f"{ACTION}.delta_y", 0.0)
+        delta_z = action.pop(f"{ACTION}.delta_z", 0.0)
+        gripper = action.pop(f"{ACTION}.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 = {
+            f"{ACTION}.enabled": enabled,
+            f"{ACTION}.target_x": scaled_delta_x,
+            f"{ACTION}.target_y": scaled_delta_y,
+            f"{ACTION}.target_z": scaled_delta_z,
+            f"{ACTION}.target_wx": target_wx,
+            f"{ACTION}.target_wy": target_wy,
+            f"{ACTION}.target_wz": target_wz,
+            f"{ACTION}.gripper": float(gripper),
+        }
+
+        return action
+
+    def transform_features(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
+        """Transform features to match output format."""
+        features.pop(f"{ACTION}.delta_x", None)
+        features.pop(f"{ACTION}.delta_y", None)
+        features.pop(f"{ACTION}.delta_z", None)
+        features.pop(f"{ACTION}.gripper", None)
+
+        features[f"{ACTION}.enabled"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        features[f"{ACTION}.target_x"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        features[f"{ACTION}.target_y"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        features[f"{ACTION}.target_z"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        features[f"{ACTION}.target_wx"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        features[f"{ACTION}.target_wy"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        features[f"{ACTION}.target_wz"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        features[f"{ACTION}.gripper"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        return features

src/lerobot/processor/device_processor.py

@@ -19,13 +19,15 @@ from typing import Any
 import torch
 
 from lerobot.configs.types import PolicyFeature
-from lerobot.processor.pipeline import EnvTransition, ProcessorStepRegistry, TransitionKey
 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:
+class DeviceProcessorStep(ProcessorStep):
     """Processes transitions by moving tensors to the specified device and optionally converting float dtypes.
 
     This processor ensures that all tensors in the transition are moved to the
@@ -36,39 +38,54 @@ class DeviceProcessor:
 
     device: str = "cpu"
     float_dtype: str | None = None
-    _device: torch.device | 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)
-        self.device = self._device.type
+        self.tensor_device: torch.device = get_safe_torch_device(self.device)
+        self.device = self.tensor_device.type  # cuda might have changed to cuda:1
         self.non_blocking = "cuda" in str(self.device)
 
         # Validate and convert float_dtype string to torch dtype
         if self.float_dtype is not None:
-            dtype_mapping = {
-                "float16": torch.float16,
-                "float32": torch.float32,
-                "float64": torch.float64,
-                "bfloat16": torch.bfloat16,
-                "half": torch.float16,
-                "float": torch.float32,
-                "double": torch.float64,
-            }
-
-            if self.float_dtype not in dtype_mapping:
-                available_dtypes = list(dtype_mapping.keys())
+            if self.float_dtype not in self.DTYPE_MAPPING:
                 raise ValueError(
-                    f"Invalid float_dtype '{self.float_dtype}'. Available options: {available_dtypes}"
+                    f"Invalid float_dtype '{self.float_dtype}'. Available options: {list(self.DTYPE_MAPPING.keys())}"
                 )
 
-            self._target_float_dtype = dtype_mapping[self.float_dtype]
+            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:
-        """Process a tensor by moving to device and optionally converting float dtype."""
-        # Move to device first
-        tensor = tensor.to(self.device, non_blocking=self.non_blocking)
+        """Process a tensor by moving to device and optionally converting float dtype.
+
+        If the tensor is already on a GPU and we're configured for a GPU, it preserves
+        that GPU placement (useful for multi-GPU training with Accelerate).
+        Otherwise, it moves to the configured device.
+        """
+        # 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():
@@ -77,51 +94,35 @@ class DeviceProcessor:
         return tensor
 
     def __call__(self, transition: EnvTransition) -> EnvTransition:
-        # Create a copy of the transition
         new_transition = transition.copy()
 
-        # Process observation tensors
-        observation = transition.get(TransitionKey.OBSERVATION)
-        if observation is not None:
-            new_observation = {
-                k: self._process_tensor(v) 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] = self._process_tensor(action)
+        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] = self._process_tensor(reward)
+        # Process simple 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] = self._process_tensor(done)
-
-        # Process truncated tensor
-        truncated = transition.get(TransitionKey.TRUNCATED)
-        if truncated is not None and isinstance(truncated, torch.Tensor):
-            new_transition[TransitionKey.TRUNCATED] = self._process_tensor(truncated)
-
-        # Process complementary data tensors
-        complementary_data = transition.get(TransitionKey.COMPLEMENTARY_DATA)
-        if complementary_data is not None:
-            new_complementary_data = {}
-
-            # Process all items in complementary_data
-            for key, value in complementary_data.items():
-                if isinstance(value, torch.Tensor):
-                    new_complementary_data[key] = self._process_tensor(value)
-                else:
-                    new_complementary_data[key] = value
-
-            new_transition[TransitionKey.COMPLEMENTARY_DATA] = new_complementary_data
+        # Process dictionary-like tensors
+        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
 
@@ -129,17 +130,5 @@ class DeviceProcessor:
         """Return configuration for serialization."""
         return {"device": self.device, "float_dtype": self.float_dtype}
 
-    def state_dict(self) -> dict[str, torch.Tensor]:
-        """Return state dictionary (empty for this processor)."""
-        return {}
-
-    def load_state_dict(self, state: dict[str, torch.Tensor]) -> None:
-        """Load state dictionary (no-op for this processor)."""
-        pass
-
-    def reset(self) -> None:
-        """Reset processor state (no-op for this processor)."""
-        pass
-
     def transform_features(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
         return features

src/lerobot/processor/gym_action_processor.py

@@ -0,0 +1,72 @@
+#! /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,
+
+from dataclasses import dataclass
+
+import numpy as np
+import torch
+
+from lerobot.configs.types import PolicyFeature
+
+from .converters import to_tensor
+from .pipeline import ActionProcessorStep, ProcessorStepRegistry
+
+
+@ProcessorStepRegistry.register("torch2numpy_action_processor")
+@dataclass
+class Torch2NumpyActionProcessorStep(ActionProcessorStep):
+    """Convert PyTorch tensor actions to NumPy arrays."""
+
+    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[str, PolicyFeature]) -> dict[str, PolicyFeature]:
+        return features
+
+
+@ProcessorStepRegistry.register("numpy2torch_action_processor")
+@dataclass
+class Numpy2TorchActionProcessorStep(ActionProcessorStep):
+    """Convert NumPy array action to PyTorch tensor."""
+
+    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[str, PolicyFeature]) -> dict[str, PolicyFeature]:
+        return features

src/lerobot/processor/hil_processor.py

@@ -0,0 +1,382 @@
+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 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 only defines the additional get_teleop_events() method,
+    avoiding duplication of the entire Teleoperator interface.
+    """
+
+    def get_teleop_events(self) -> dict[str, Any]:
+        """Get extra control events from the teleoperator.
+
+        Returns:
+            Dictionary containing control events such as:
+                - is_intervention: bool - Whether 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 get_teleop_events."""
+    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):
+    """Add teleoperator action to transition complementary data."""
+
+    teleop_device: Teleoperator
+
+    def complementary_data(self, complementary_data: dict) -> dict:
+        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[str, PolicyFeature]) -> dict[str, PolicyFeature]:
+        return features
+
+
+@ProcessorStepRegistry.register("add_teleop_action_as_info")
+@dataclass
+class AddTeleopEventsAsInfoStep(InfoProcessorStep):
+    """Add teleoperator control events to transition info.
+
+    This processor step extracts control events from teleoperators that support
+    event-based interaction (intervention detection, episode termination, etc.).
+
+    Works with any teleoperator that inherits from Teleoperator and implements the
+    get_teleop_events() method, including custom user-defined teleoperators.
+
+    Built-in compatible teleoperators:
+        - GamepadTeleop: Uses gamepad buttons for control events
+        - KeyboardEndEffectorTeleop: Uses keyboard keys for control events
+    """
+
+    teleop_device: TeleopWithEvents
+
+    def __post_init__(self):
+        """Validate that the teleoperator supports events."""
+        _check_teleop_with_events(self.teleop_device)
+
+    def info(self, info: dict) -> dict:
+        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[str, PolicyFeature]) -> dict[str, PolicyFeature]:
+        return features
+
+
+@ProcessorStepRegistry.register("image_crop_resize_processor")
+@dataclass
+class ImageCropResizeProcessorStep(ObservationProcessorStep):
+    """Crop and resize image observations."""
+
+    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:
+        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]:
+        return {
+            "crop_params_dict": self.crop_params_dict,
+            "resize_size": self.resize_size,
+        }
+
+    def transform_features(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
+        if self.resize_size is None:
+            return features
+        for key in features:
+            if "image" in key:
+                features[key] = PolicyFeature(type=features[key].type, shape=self.resize_size)
+        return features
+
+
+@dataclass
+@ProcessorStepRegistry.register("time_limit_processor")
+class TimeLimitProcessorStep(TruncatedProcessorStep):
+    """Track episode steps and enforce time limits."""
+
+    max_episode_steps: int
+    current_step: int = 0
+
+    def truncated(self, truncated):
+        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]:
+        return {
+            "max_episode_steps": self.max_episode_steps,
+        }
+
+    def reset(self) -> None:
+        self.current_step = 0
+
+    def transform_features(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
+        return features
+
+
+@dataclass
+@ProcessorStepRegistry.register("gripper_penalty_processor")
+class GripperPenaltyProcessorStep(ComplementaryDataProcessorStep):
+    """Apply penalty for inappropriate gripper usage."""
+
+    penalty: float = -0.01
+    max_gripper_pos: float = 30.0
+
+    def complementary_data(self, complementary_data):
+        """Calculate gripper penalty and add to complementary data."""
+        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]:
+        return {
+            "penalty": self.penalty,
+            "max_gripper_pos": self.max_gripper_pos,
+        }
+
+    def reset(self) -> None:
+        """Reset the processor state."""
+        self.last_gripper_state = None
+
+    def transform_features(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
+        return features
+
+
+@dataclass
+@ProcessorStepRegistry.register("intervention_action_processor")
+class InterventionActionProcessorStep(ProcessorStep):
+    """Handle human intervention actions and episode termination."""
+
+    use_gripper: bool = False
+    terminate_on_success: bool = True
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        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]:
+        return {
+            "use_gripper": self.use_gripper,
+            "terminate_on_success": self.terminate_on_success,
+        }
+
+    def transform_features(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
+        return features
+
+
+@dataclass
+@ProcessorStepRegistry.register("reward_classifier_processor")
+class RewardClassifierProcessorStep(ProcessorStep):
+    """Apply reward classification to image observations."""
+
+    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):
+        """Initialize the reward classifier after dataclass initialization."""
+        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:
+        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]:
+        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[str, PolicyFeature]) -> dict[str, PolicyFeature]:
+        return features

src/lerobot/processor/joint_observations_processor.py

@@ -0,0 +1,109 @@
+from dataclasses import dataclass
+from typing import Any
+
+import torch
+
+from lerobot.configs.types import 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):
+    """Add joint velocity information to observations."""
+
+    dt: float = 0.1
+
+    last_joint_positions: torch.Tensor | None = None
+
+    def observation(self, observation: dict) -> dict:
+        # Get current joint positions (assuming they're in observation.state)
+        current_positions = observation.get(OBS_STATE)
+        if current_positions is None:
+            # TODO(steven): if we get here, then the transform_features method will not hold
+            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]:
+        return {
+            "dt": self.dt,
+        }
+
+    def reset(self) -> None:
+        self.last_joint_positions = None
+
+    def transform_features(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
+        if OBS_STATE in features:
+            original_feature = features[OBS_STATE]
+            # Double the shape to account for positions + velocities
+            new_shape = (original_feature.shape[0] * 2,) + original_feature.shape[1:]
+
+            features[OBS_STATE] = PolicyFeature(type=original_feature.type, shape=new_shape)
+        return features
+
+
+@dataclass
+@ProcessorStepRegistry.register("current_processor")
+class MotorCurrentProcessorStep(ObservationProcessorStep):
+    """Add motor current information to observations."""
+
+    robot: Robot | None = None
+
+    def observation(self, observation: dict) -> dict:
+        # 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[str, PolicyFeature]) -> dict[str, PolicyFeature]:
+        if OBS_STATE in features and self.robot is not None:
+            original_feature = features[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[OBS_STATE] = PolicyFeature(type=original_feature.type, shape=new_shape)
+        return features

src/lerobot/processor/migrate_policy_normalization.py

@@ -46,11 +46,12 @@ 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 lerobot.processor.batch_processor import ToBatchProcessor
-from lerobot.processor.device_processor import DeviceProcessor
-from lerobot.processor.normalize_processor import NormalizerProcessor, UnnormalizerProcessor
-from lerobot.processor.pipeline import RobotProcessor
-from lerobot.processor.rename_processor import RenameProcessor
+
+from .batch_processor import AddBatchDimensionProcessorStep
+from .device_processor import DeviceProcessorStep
+from .normalize_processor import NormalizerProcessorStep, UnnormalizerProcessorStep
+from .pipeline import PolicyProcessorPipeline
+from .rename_processor import RenameProcessorStep
 
 # Policy type to class mapping
 POLICY_CLASSES = {
@@ -403,8 +404,8 @@ def main():
     # Now create preprocessor and postprocessor with cleaned_config available
     print("Creating preprocessor and postprocessor...")
     # The pattern from existing processor factories:
-    # - Preprocessor has two NormalizerProcessors: one for input_features, one for output_features
-    # - Postprocessor has one UnnormalizerProcessor for output_features only
+    # - 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", {})
@@ -412,23 +413,23 @@ def main():
 
     # Create preprocessor with two normalizers (following the pattern from processor factories)
     preprocessor_steps = [
-        RenameProcessor(rename_map={}),
-        NormalizerProcessor(
+        RenameProcessorStep(rename_map={}),
+        NormalizerProcessorStep(
             features={**input_features, **output_features},
             norm_map=norm_map,
             stats=stats,
         ),
-        ToBatchProcessor(),
-        DeviceProcessor(device=policy_config.device),
+        AddBatchDimensionProcessorStep(),
+        DeviceProcessorStep(device=policy_config.device),
     ]
-    preprocessor = RobotProcessor(steps=preprocessor_steps, name="robot_preprocessor")
+    preprocessor = PolicyProcessorPipeline(steps=preprocessor_steps, name="robot_preprocessor")
 
     # Create postprocessor with unnormalizer for outputs only
     postprocessor_steps = [
-        DeviceProcessor(device="cpu"),
-        UnnormalizerProcessor(features=output_features, norm_map=norm_map, stats=stats),
+        DeviceProcessorStep(device="cpu"),
+        UnnormalizerProcessorStep(features=output_features, norm_map=norm_map, stats=stats),
     ]
-    postprocessor = RobotProcessor(steps=postprocessor_steps, name="robot_postprocessor")
+    postprocessor = PolicyProcessorPipeline(steps=postprocessor_steps, name="robot_postprocessor")
 
     # Determine hub repo ID if pushing to hub
     if args.push_to_hub:

src/lerobot/processor/normalize_processor.py

@@ -1,232 +1,84 @@
 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.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.processor.pipeline import EnvTransition, ProcessorStepRegistry, RobotProcessor, TransitionKey
 
-
-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
+from .converters import to_tensor
+from .core import EnvTransition, TransitionKey
+from .pipeline import PolicyProcessorPipeline, ProcessorStep, ProcessorStepRegistry
 
 
 @dataclass
-@ProcessorStepRegistry.register(name="normalizer_processor")
-class NormalizerProcessor:
-    """Normalizes observations and actions in a single processor step.
+class _NormalizationMixin:
+    """
+    A mixin class providing core functionality for normalization and unnormalization.
 
-    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.
+    This class manages normalization statistics, their conversion to tensors, device placement,
+    and the application of normalization transformations. It is designed to be inherited by
+    concrete ProcessorStep implementations.
     """
 
-    # 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
-
+    device: torch.device | str | 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)
 
-    @classmethod
-    def from_lerobot_dataset(
-        cls,
-        dataset: LeRobotDataset,
-        features: dict[str, PolicyFeature],
-        norm_map: dict[FeatureType, NormalizationMode],
-        *,
-        normalize_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.
-        """
-
-        return cls(
-            features=features,
-            norm_map=norm_map,
-            stats=dataset.meta.stats,
-            normalize_keys=normalize_keys,
-            eps=eps,
-        )
-
     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
+        # 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 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
+        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 statistics once so we avoid repeated numpy→Tensor conversions
-        # during runtime.
+        # Convert stats to tensors and move to the target device once during initialization.
         self.stats = self.stats or {}
-        self._tensor_stats = _convert_stats_to_tensors(self.stats)
+        self._tensor_stats = to_tensor(self.stats, device=self.device)
 
-        # 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 to(self, device: torch.device | str) -> _NormalizationMixin:
+        """Moves the processor's normalization stats to the specified device and returns self."""
+        self.device = device
+        self._tensor_stats = to_tensor(self.stats, device=self.device)
+        return self
 
-    def _normalize_obs(self, observation, normalized_info):
-        if observation is None:
-            return None
+    def state_dict(self) -> dict[str, Tensor]:
+        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
 
-        # 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.features:
-                continue
-
-            # Check the normalization mode for this feature type
-            feature = self.features[key]
-            norm_mode = self.norm_map.get(feature.type, NormalizationMode.IDENTITY)
-
-            # Skip normalization if mode is IDENTITY
-            if norm_mode is NormalizationMode.IDENTITY:
-                normalized_info[key] = "IDENTITY"
-                continue
-
-            # Skip if no stats available for this key
-            if 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)
+    def load_state_dict(self, state: dict[str, Tensor]) -> None:
+        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
             )
-            stats = {k: v.to(tensor.device) for k, v in self._tensor_stats[key].items()}
-
-            if norm_mode is NormalizationMode.MEAN_STD:
-                if "mean" in stats and "std" in stats:
-                    mean, std = stats["mean"], stats["std"]
-                    processed[key] = (tensor - mean) / (std + self.eps)
-                    normalized_info[key] = "MEAN_STD"
-            elif norm_mode is NormalizationMode.MIN_MAX:
-                if "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
-                    normalized_info[key] = "MIN_MAX"
-            else:
-                raise ValueError(f"Unsupported normalization mode: {norm_mode}")
-
-        return processed
-
-    def _normalize_action(self, action, normalized_info):
-        if action is None:
-            return action
-
-        # Check the normalization mode for actions
-        norm_mode = self.norm_map.get(FeatureType.ACTION, NormalizationMode.IDENTITY)
-
-        # Skip normalization if mode is IDENTITY
-        if norm_mode is NormalizationMode.IDENTITY:
-            normalized_info["action"] = "IDENTITY"
-            return action
-
-        # Skip if no stats available for actions
-        if "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 norm_mode is NormalizationMode.MEAN_STD:
-            if "mean" in stats and "std" in stats:
-                mean, std = stats["mean"], stats["std"]
-                normalized_info["action"] = "MEAN_STD"
-                return (tensor - mean) / (std + self.eps)
-        elif norm_mode is NormalizationMode.MIN_MAX:
-            if "min" in stats and "max" in stats:
-                min_val, max_val = stats["min"], stats["max"]
-                normalized_info["action"] = "MIN_MAX"
-                return 2 * (tensor - min_val) / (max_val - min_val + self.eps) - 1
-        else:
-            raise ValueError(f"Unsupported normalization mode: {norm_mode}")
-
-        # If we reach here, the required stats for the normalization mode are not available
-        raise ValueError(f"Action stats must contain appropriate values for {norm_mode} normalization")
-
-    def __call__(self, transition: EnvTransition) -> EnvTransition:
-        # Track what was normalized
-        normalized_info = {}
-
-        observation = self._normalize_obs(transition.get(TransitionKey.OBSERVATION), normalized_info)
-        action = self._normalize_action(transition.get(TransitionKey.ACTION), normalized_info)
-
-        # Create a new transition with normalized values
-        new_transition = transition.copy()
-        new_transition[TransitionKey.OBSERVATION] = observation
-        new_transition[TransitionKey.ACTION] = action
-
-        # Add normalization info to complementary data
-        if normalized_info:
-            comp_data = new_transition.get(TransitionKey.COMPLEMENTARY_DATA, {})
-            comp_data = {} if comp_data is None else dict(comp_data)
-            comp_data["normalized_keys"] = normalized_info
-            new_transition[TransitionKey.COMPLEMENTARY_DATA] = comp_data
-
-        return new_transition
 
     def get_config(self) -> dict[str, Any]:
         config = {
@@ -236,45 +88,87 @@ class NormalizerProcessor:
             },
             "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)
+        if self.normalize_observation_keys is not None:
+            config["normalize_observation_keys"] = sorted(self.normalize_observation_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 _normalize_observation(self, observation: dict[str, Any], inverse: bool) -> dict[str, Tensor]:
+        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:
+                tensor = torch.as_tensor(new_observation[key], dtype=torch.float32)
+                new_observation[key] = self._apply_transform(tensor, key, feature.type, inverse=inverse)
+        return new_observation
 
-    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 _normalize_action(self, action: Any, inverse: bool) -> Tensor:
+        tensor = torch.as_tensor(action, dtype=torch.float32)
+        processed_action = self._apply_transform(tensor, "action", FeatureType.ACTION, inverse=inverse)
+        return processed_action
 
-    def reset(self):
-        pass
+    def _apply_transform(
+        self, tensor: Tensor, key: str, feature_type: FeatureType, *, inverse: bool = False
+    ) -> Tensor:
+        """Core logic to apply normalization or unnormalization."""
+        norm_mode = self.norm_map.get(feature_type, NormalizationMode.IDENTITY)
+        if norm_mode == NormalizationMode.IDENTITY or key not in self._tensor_stats:
+            return tensor
 
-    def transform_features(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
-        return features
+        if norm_mode not in (NormalizationMode.MEAN_STD, NormalizationMode.MIN_MAX):
+            raise ValueError(f"Unsupported normalization mode: {norm_mode}")
+
+        # Ensure input tensor is on the same device as the stats.
+        if self.device and tensor.device != self.device:
+            tensor = tensor.to(self.device)
+
+        # For Accelerate compatibility: move stats to match input tensor device
+        input_device = tensor.device
+        stats = self._tensor_stats[key]
+        tensor = tensor.to(dtype=torch.float32)
+
+        # Move stats to input device if needed
+        stats_device = next(iter(stats.values())).device
+        if stats_device != input_device:
+            stats = to_tensor({key: self._tensor_stats[key]}, device=input_device)[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=input_device, dtype=torch.float32), 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="unnormalizer_processor")
-class UnnormalizerProcessor:
-    """Inverse normalisation for observations and actions.
-
-    Exactly mirrors :class:`NormalizerProcessor` but applies the inverse
-    transform.
+@ProcessorStepRegistry.register(name="normalizer_processor")
+class NormalizerProcessorStep(_NormalizationMixin, ProcessorStep):
     """
+    A processor that applies normalization to observations and actions in a transition.
 
-    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 directly implements the normalization logic for both observation and action
+    components of an `EnvTransition`, using statistics (mean/std or min/max) provided at
+    initialization.
+    """
 
     @classmethod
     def from_lerobot_dataset(
@@ -282,194 +176,97 @@ 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)
-
-    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
-
-        self.stats = self.stats or {}
-        self._tensor_stats = _convert_stats_to_tensors(self.stats)
-
-    def _unnormalize_obs(self, observation, unnormalized_info):
-        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.features:
-                continue
-
-            # Check the normalization mode for this feature type
-            feature = self.features[key]
-            norm_mode = self.norm_map.get(feature.type, NormalizationMode.IDENTITY)
-
-            # Skip unnormalization if mode is IDENTITY
-            if norm_mode is NormalizationMode.IDENTITY:
-                unnormalized_info[key] = "IDENTITY"
-                continue
-
-            # Skip if no stats available for this key
-            if 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 norm_mode is NormalizationMode.MEAN_STD:
-                if "mean" in stats and "std" in stats:
-                    mean, std = stats["mean"], stats["std"]
-                    processed[key] = tensor * std + mean
-                    unnormalized_info[key] = "MEAN_STD"
-            elif norm_mode is NormalizationMode.MIN_MAX:
-                if "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
-                    unnormalized_info[key] = "MIN_MAX"
-            else:
-                raise ValueError(f"Unsupported normalization mode: {norm_mode}")
-
-        return processed
-
-    def _unnormalize_action(self, action, unnormalized_info):
-        if action is None:
-            return action
-
-        # Check the normalization mode for actions
-        norm_mode = self.norm_map.get(FeatureType.ACTION, NormalizationMode.IDENTITY)
-
-        # Skip unnormalization if mode is IDENTITY
-        if norm_mode is NormalizationMode.IDENTITY:
-            unnormalized_info["action"] = "IDENTITY"
-            return action
-
-        # Skip if no stats available for actions
-        if "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)
+        *,
+        normalize_observation_keys: set[str] | None = None,
+        eps: float = 1e-8,
+        device: torch.device | str | None = None,
+    ) -> NormalizerProcessorStep:
+        return cls(
+            features=features,
+            norm_map=norm_map,
+            stats=dataset.meta.stats,
+            normalize_observation_keys=normalize_observation_keys,
+            eps=eps,
+            device=device,
         )
-        stats = {k: v.to(tensor.device) for k, v in self._tensor_stats["action"].items()}
-
-        if norm_mode is NormalizationMode.MEAN_STD:
-            if "mean" in stats and "std" in stats:
-                mean, std = stats["mean"], stats["std"]
-                unnormalized_info["action"] = "MEAN_STD"
-                return tensor * std + mean
-        elif norm_mode is NormalizationMode.MIN_MAX:
-            if "min" in stats and "max" in stats:
-                min_val, max_val = stats["min"], stats["max"]
-                unnormalized_info["action"] = "MIN_MAX"
-                return (tensor + 1) / 2 * (max_val - min_val) + min_val
-        else:
-            raise ValueError(f"Unsupported normalization mode: {norm_mode}")
-
-        # If we reach here, the required stats for the normalization mode are not available
-        raise ValueError(f"Action stats must contain appropriate values for {norm_mode} normalization")
 
     def __call__(self, transition: EnvTransition) -> EnvTransition:
-        # Track what was unnormalized
-        unnormalized_info = {}
-
-        observation = self._unnormalize_obs(transition.get(TransitionKey.OBSERVATION), unnormalized_info)
-        action = self._unnormalize_action(transition.get(TransitionKey.ACTION), unnormalized_info)
-
-        # Create a new transition with unnormalized values
         new_transition = transition.copy()
-        new_transition[TransitionKey.OBSERVATION] = observation
-        new_transition[TransitionKey.ACTION] = action
 
-        # Add unnormalization info to complementary data
-        if unnormalized_info:
-            comp_data = new_transition.get(TransitionKey.COMPLEMENTARY_DATA, {})
-            comp_data = {} if comp_data is None else dict(comp_data)
-            comp_data["unnormalized_keys"] = unnormalized_info
-            new_transition[TransitionKey.COMPLEMENTARY_DATA] = comp_data
+        # 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]:
-        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 transform_features(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
         return features
 
 
-def hotswap_stats(robot_processor: RobotProcessor, stats: dict[str, dict[str, Any]]) -> RobotProcessor:
-    robot_processor = deepcopy(robot_processor)
-    for step in robot_processor.steps:
-        if isinstance(step, NormalizerProcessor) or isinstance(step, UnnormalizerProcessor):
-            step: NormalizerProcessor | UnnormalizerProcessor
-            step.stats = stats
-            step._tensor_stats = _convert_stats_to_tensors(stats)
-    return robot_processor
-
-
-def rename_stats(stats: dict[str, dict[str, Any]], rename_map: dict[str, str]) -> dict[str, dict[str, Any]]:
-    """Rename keys in the stats dictionary according to the provided mapping.
-
-    Args:
-        stats: The statistics dictionary with structure {feature_key: {stat_name: value}}
-        rename_map: Dictionary mapping old key names to new key names
-
-    Returns:
-        A new stats dictionary with renamed keys
-
-    Example:
-        >>> stats = {"observation.state": {"mean": 0.0, "std": 1.0}, "action": {"mean": 0.5, "std": 0.5}}
-        >>> rename_map = {"observation.state": "observation.robot_state"}
-        >>> new_stats = rename_stats(stats, rename_map)
-        >>> # new_stats will have "observation.robot_state" instead of "observation.state"
+@dataclass
+@ProcessorStepRegistry.register(name="unnormalizer_processor")
+class UnnormalizerProcessorStep(_NormalizationMixin, ProcessorStep):
     """
-    renamed_stats = {}
+    A processor that applies unnormalization (the inverse of normalization) to
+    observations and actions in a transition.
 
-    for old_key, sub_stats in stats.items():
-        # Use the new key if it exists in the rename map, otherwise keep the old key
-        new_key = rename_map.get(old_key, old_key)
-        renamed_stats[new_key] = deepcopy(sub_stats)
+    This is typically used to transform actions from a normalized policy output back into
+    the original scale for execution in an environment.
+    """
 
-    return renamed_stats
+    @classmethod
+    def from_lerobot_dataset(
+        cls,
+        dataset: LeRobotDataset,
+        features: dict[str, PolicyFeature],
+        norm_map: dict[FeatureType, NormalizationMode],
+        *,
+        device: torch.device | str | None = None,
+    ) -> UnnormalizerProcessorStep:
+        return cls(features=features, norm_map=norm_map, stats=dataset.meta.stats, device=device)
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        new_transition = transition.copy()
+
+        # 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 transform_features(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
+        return features
+
+
+def hotswap_stats(
+    policy_processor: PolicyProcessorPipeline, stats: dict[str, dict[str, Any]]
+) -> PolicyProcessorPipeline:
+    """
+    Replaces normalization statistics in a PolicyProcessor pipeline.
+
+    This function creates a deep copy of the provided `PolicyProcessorPipeline` and updates the
+    statistics of any `NormalizerProcessorStep` or `UnnormalizerProcessorStep` steps within it.
+    It's useful for adapting a trained policy to a new environment or dataset with
+    different data distributions.
+    """
+    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)
+    return rp

src/lerobot/processor/observation_processor.py

@@ -22,12 +22,13 @@ from torch import Tensor
 
 from lerobot.configs.types import 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.
 

src/lerobot/processor/rename_processor.py

@@ -13,19 +13,18 @@
 # 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 .pipeline import ObservationProcessorStep, ProcessorStepRegistry
 
 
 @dataclass
 @ProcessorStepRegistry.register(name="rename_processor")
-class RenameProcessor(ObservationProcessor):
+class RenameProcessorStep(ObservationProcessorStep):
     """Rename processor that renames keys in the observation."""
 
     rename_map: dict[str, str] = field(default_factory=dict)
@@ -49,3 +48,14 @@ class RenameProcessor(ObservationProcessor):
         - Keys not in `rename_map` remain unchanged.
         """
         return {self.rename_map.get(k, k): v for k, v in features.items()}
+
+
+def rename_stats(stats: dict[str, dict[str, Any]], rename_map: dict[str, str]) -> dict[str, dict[str, Any]]:
+    """Rename keys in the stats dictionary according to rename_map (defensive copy)."""
+    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

@@ -10,10 +10,12 @@ from typing import TYPE_CHECKING, Any
 import torch
 
 from lerobot.configs.types import FeatureType, PolicyFeature
-from lerobot.constants import OBS_LANGUAGE
-from lerobot.processor.pipeline import EnvTransition, ProcessorStepRegistry, TransitionKey
+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
+
 if TYPE_CHECKING or _transformers_available:
     from transformers import AutoTokenizer
 else:
@@ -22,7 +24,7 @@ else:
 
 @dataclass
 @ProcessorStepRegistry.register(name="tokenizer_processor")
-class TokenizerProcessor:
+class TokenizerProcessorStep(ObservationProcessorStep):
     """Tokenizes text tasks in complementary data using a huggingface tokenizer.
 
     This processor handles tokenization of task strings found in the complementary_data
@@ -46,7 +48,7 @@ class TokenizerProcessor:
     Examples:
         Using tokenizer name (auto-loaded):
         ```python
-        processor = TokenizerProcessor(tokenizer_name="bert-base-uncased", max_length=128)
+        processor = TokenizerProcessorStep(tokenizer_name="bert-base-uncased", max_length=128)
         ```
 
         Using custom tokenizer object:
@@ -54,7 +56,7 @@ class TokenizerProcessor:
         from transformers import AutoTokenizer
 
         custom_tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
-        processor = TokenizerProcessor(tokenizer=custom_tokenizer, max_length=128)
+        processor = TokenizerProcessorStep(tokenizer=custom_tokenizer, max_length=128)
         ```
     """
 
@@ -67,23 +69,23 @@ class TokenizerProcessor:
     truncation: bool = True
 
     # Internal tokenizer instance (not serialized)
-    _tokenizer: Any = field(default=None, init=False, repr=False)
+    input_tokenizer: Any = field(default=None, init=False, repr=False)
 
     def __post_init__(self):
         """Initialize the tokenizer from the provided tokenizer or tokenizer name."""
         if not _transformers_available:
             raise ImportError(
                 "The 'transformers' library is not installed. "
-                "Please install it with `pip install 'lerobot[transformers-dep]'` to use TokenizerProcessor."
+                "Please install it with `pip install 'lerobot[transformers-dep]'` to use TokenizerProcessorStep."
             )
 
         if self.tokenizer is not None:
             # Use provided tokenizer object directly
-            self._tokenizer = self.tokenizer
+            self.input_tokenizer = self.tokenizer
         elif self.tokenizer_name is not None:
             if AutoTokenizer is None:
                 raise ImportError("AutoTokenizer is not available")
-            self._tokenizer = AutoTokenizer.from_pretrained(self.tokenizer_name)
+            self.input_tokenizer = AutoTokenizer.from_pretrained(self.tokenizer_name)
         else:
             raise ValueError(
                 "Either 'tokenizer' or 'tokenizer_name' must be provided. "
@@ -118,7 +120,7 @@ class TokenizerProcessor:
 
         return None
 
-    def __call__(self, transition: EnvTransition) -> EnvTransition:
+    def observation(self, observation):
         """Process the transition by tokenizing the task text.
 
         Args:
@@ -130,28 +132,57 @@ class TokenizerProcessor:
         Raises:
             ValueError: If tokenizer initialization failed.
         """
-        task = self.get_task(transition)
+        task = self.get_task(self.transition)
         if task is None:
-            return transition
+            return observation
 
-        # Tokenize the task
+        # Tokenize the task (creates CPU tensors)
         tokenized_prompt = self._tokenize_text(task)
 
+        # Detect device from existing tensors in the transition
+        target_device = self._detect_device(self.transition)
+
+        # Move tokenized tensors to match the device of other data
+        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()
+            }
+
         # Get or create observation dict
-        observation = transition.get(TransitionKey.OBSERVATION)
-        if observation is None:
-            observation = {}
-        else:
-            observation = dict(observation)  # Make a copy
+        new_observation = dict(observation)
 
         # Add tokenized data to observation
-        observation[f"{OBS_LANGUAGE}.tokens"] = tokenized_prompt["input_ids"]
-        observation[f"{OBS_LANGUAGE}.attention_mask"] = tokenized_prompt["attention_mask"].to(
-            dtype=torch.bool
-        )
+        new_observation[OBS_LANGUAGE_TOKENS] = tokenized_prompt["input_ids"]
+        new_observation[OBS_LANGUAGE_ATTENTION_MASK] = tokenized_prompt["attention_mask"].to(dtype=torch.bool)
 
-        transition[TransitionKey.OBSERVATION.value] = observation  # type: ignore[misc]
-        return transition
+        return new_observation
+
+    def _detect_device(self, transition: EnvTransition) -> torch.device | None:
+        """Detect device from existing tensors in the transition.
+
+        This allows the tokenized tensors to match the device of other data,
+        which is especially important for multi-GPU training with Accelerate.
+
+        Args:
+            transition: The transition to search for existing tensors.
+
+        Returns:
+            The device of the first tensor found, or None if no tensors exist.
+        """
+        # Check observation tensors first (most likely to exist)
+        observation = transition.get(TransitionKey.OBSERVATION)
+        if observation:
+            for value in observation.values():
+                if isinstance(value, torch.Tensor):
+                    return value.device
+
+        # Check action tensor
+        action = transition.get(TransitionKey.ACTION)
+        if isinstance(action, torch.Tensor):
+            return action.device
+
+        return None  # No tensors found, keep on CPU
 
     def _tokenize_text(self, text: str | list[str]) -> dict[str, torch.Tensor]:
         """Tokenize text using the configured tokenizer.
@@ -162,7 +193,7 @@ class TokenizerProcessor:
         Returns:
             Dictionary containing tokenized output with keys like 'input_ids', 'attention_mask'.
         """
-        return self._tokenizer(
+        return self.input_tokenizer(
             text,
             max_length=self.max_length,
             truncation=self.truncation,
@@ -186,23 +217,12 @@ class TokenizerProcessor:
         }
 
         # Only include tokenizer_name if it was used (not when tokenizer object was provided)
-        if self.tokenizer_name is not None:
+        # TODO(steven): Consider saving the name of the _tokenizer if it was loaded
+        if self.tokenizer_name is not None and self.tokenizer is None:
             config["tokenizer_name"] = self.tokenizer_name
 
         return config
 
-    def state_dict(self) -> dict[str, torch.Tensor]:
-        """Return state dictionary (empty for this processor)."""
-        return {}
-
-    def load_state_dict(self, state: dict[str, torch.Tensor]) -> None:
-        """Load state dictionary (no-op for this processor)."""
-        pass
-
-    def reset(self) -> None:
-        """Reset processor state (no-op for this processor)."""
-        pass
-
     def transform_features(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
         """Add tokenized task features to the feature contract.
 
@@ -214,13 +234,13 @@ class TokenizerProcessor:
         """
         # Add features for tokenized output if they don't exist
         # Standard tokenizer output includes tokens and attention_mask
-        tokens_key = f"{OBS_LANGUAGE}.tokens"
-        attention_mask_key = f"{OBS_LANGUAGE}.attention_mask"
 
-        if tokens_key not in features:
-            features[tokens_key] = PolicyFeature(type=FeatureType.LANGUAGE, shape=(self.max_length,))
+        if OBS_LANGUAGE_TOKENS not in features:
+            features[OBS_LANGUAGE_TOKENS] = PolicyFeature(type=FeatureType.LANGUAGE, shape=(self.max_length,))
 
-        if attention_mask_key not in features:
-            features[attention_mask_key] = PolicyFeature(type=FeatureType.LANGUAGE, shape=(self.max_length,))
+        if OBS_LANGUAGE_ATTENTION_MASK not in features:
+            features[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 \
@@ -74,17 +74,21 @@ from lerobot.datasets.image_writer import safe_stop_image_writer
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.datasets.utils import hw_to_dataset_features
 from lerobot.datasets.video_utils import VideoEncodingManager
-from lerobot.policies.factory import make_policy, make_processor
+from lerobot.policies.factory import make_policy, make_pre_post_processors
 from lerobot.policies.pretrained import PreTrainedPolicy
-from lerobot.processor import RobotProcessor
-from lerobot.processor.converters import (
-    to_dataset_frame,
-    to_output_robot_action,
-    to_transition_robot_observation,
-    to_transition_teleop_action,
+from lerobot.processor import (
+    IdentityProcessorStep,
+    PolicyProcessorPipeline,
+    RobotProcessorPipeline,
+    TransitionKey,
 )
-from lerobot.processor.normalize_processor import rename_stats
-from lerobot.processor.pipeline import IdentityProcessor, TransitionKey
+from lerobot.processor.converters import (
+    action_to_transition,
+    observation_to_transition,
+    transition_to_dataset_frame,
+    transition_to_robot_action,
+)
+from lerobot.processor.rename_processor import rename_stats
 from lerobot.robots import (  # noqa: F401
     Robot,
     RobotConfig,
@@ -236,23 +240,25 @@ def record_loop(
     dataset: LeRobotDataset | None = None,
     teleop: Teleoperator | list[Teleoperator] | None = None,
     policy: PreTrainedPolicy | None = None,
-    preprocessor: RobotProcessor | None = None,
-    postprocessor: RobotProcessor | None = None,
+    preprocessor: PolicyProcessorPipeline | None = None,
+    postprocessor: PolicyProcessorPipeline | None = None,
     control_time_s: int | None = None,
-    teleop_action_processor: RobotProcessor | None = None,  # runs after teleop
-    robot_action_processor: RobotProcessor | None = None,  # runs before robot
-    robot_observation_processor: RobotProcessor | None = None,  # runs after robot
+    teleop_action_processor: RobotProcessorPipeline | None = None,  # runs after teleop
+    robot_action_processor: RobotProcessorPipeline | None = None,  # runs before robot
+    robot_observation_processor: RobotProcessorPipeline | None = None,  # runs after robot
     single_task: str | None = None,
     display_data: bool = False,
 ):
-    teleop_action_processor = teleop_action_processor or RobotProcessor(
-        steps=[IdentityProcessor()], to_transition=to_transition_teleop_action, to_output=lambda tr: tr
+    teleop_action_processor = teleop_action_processor or RobotProcessorPipeline(
+        steps=[IdentityProcessorStep()], to_transition=action_to_transition, to_output=lambda tr: tr
     )
-    robot_action_processor = robot_action_processor or RobotProcessor(
-        steps=[IdentityProcessor()], to_transition=lambda tr: tr, to_output=to_output_robot_action
+    robot_action_processor = robot_action_processor or RobotProcessorPipeline(
+        steps=[IdentityProcessorStep()], to_transition=lambda tr: tr, to_output=transition_to_robot_action
     )
-    robot_observation_processor = robot_observation_processor or RobotProcessor(
-        steps=[IdentityProcessor()], to_transition=to_transition_robot_observation, to_output=lambda tr: tr
+    robot_observation_processor = robot_observation_processor or RobotProcessorPipeline(
+        steps=[IdentityProcessorStep()],
+        to_transition=observation_to_transition,
+        to_output=lambda tr: tr,
     )
 
     if dataset is not None and dataset.fps != fps:
@@ -308,7 +314,7 @@ def record_loop(
         # 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 = to_dataset_frame(
+                observation_frame = transition_to_dataset_frame(
                     obs_transition, dataset.features
                 )  # Convert the observation to the dataset format
 
@@ -346,13 +352,14 @@ def record_loop(
         else:
             logging.info(
                 "No policy or teleoperator provided, skipping action generation. "
-                "This is likely to happen during environment reset."
+                "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."
             )
-            # Still continue to next loop to respect timing
+            continue
 
         # 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_transition is not None:
+        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)
@@ -365,7 +372,7 @@ def record_loop(
 
         # Write to dataset
         if dataset is not None:
-            # If to_dataset_frame is provided, use it to merge the transitions.
+            # 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)
@@ -373,7 +380,7 @@ def record_loop(
                 merged.append(teleop_transition)
             if policy_transition is not None:  # The action from policy
                 merged.append(policy_transition)
-            frame = to_dataset_frame(
+            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)
@@ -399,7 +406,15 @@ 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(
@@ -434,7 +449,7 @@ def record(cfg: RecordConfig) -> LeRobotDataset:
     preprocessor = None
     postprocessor = None
     if cfg.policy is not None:
-        preprocessor, postprocessor = make_processor(
+        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),
@@ -510,5 +525,9 @@ def record(cfg: RecordConfig) -> LeRobotDataset:
     return dataset
 
 
-if __name__ == "__main__":
+def main():
     record()
+
+
+if __name__ == "__main__":
+    main()

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_robot_action
 from lerobot.robots import (  # noqa: F401
     Robot,
     RobotConfig,
@@ -83,13 +84,25 @@ 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_robot_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 +117,10 @@ 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/so100_follower/robot_kinematic_processor.py

@@ -17,24 +17,26 @@
 from dataclasses import dataclass, field
 
 import numpy as np
-from scipy.spatial.transform import Rotation
 
-from lerobot.configs.types import PolicyFeature
+from lerobot.configs.types import FeatureType, PolicyFeature
+from lerobot.constants import ACTION, OBS_STATE
 from lerobot.model.kinematics import RobotKinematics
-from lerobot.processor.pipeline import (
-    ActionProcessor,
-    ComplementaryDataProcessor,
+from lerobot.processor import (
+    ActionProcessorStep,
+    ComplementaryDataProcessorStep,
     EnvTransition,
-    ObservationProcessor,
+    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:
+class EEReferenceAndDelta(ActionProcessorStep):
     """
     Compute the desired end-effector pose from the target pose and the current pose.
 
@@ -53,14 +55,17 @@ class EEReferenceAndDelta:
     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 __call__(self, transition: EnvTransition) -> EnvTransition:
-        act = transition.get(TransitionKey.ACTION) or {}
-        comp = transition.get(TransitionKey.COMPLEMENTARY_DATA) or {}
+    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)
@@ -69,27 +74,31 @@ class EEReferenceAndDelta:
                 "raw_joint_positions is not in complementary data and is required for EEReferenceAndDelta"
             )
 
-        q = np.array([float(raw[n]) for n in self.motor_names], dtype=float)
+        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(act.pop("action.enabled", 0))
-        tx = float(act.pop("action.target_x", 0.0))
-        ty = float(act.pop("action.target_y", 0.0))
-        tz = float(act.pop("action.target_z", 0.0))
-        wx = float(act.pop("action.target_wx", 0.0))
-        wy = float(act.pop("action.target_wy", 0.0))
-        wz = float(act.pop("action.target_wz", 0.0))
+        enabled = bool(new_action.pop(f"{ACTION}.enabled", 0))
+        tx = float(new_action.pop(f"{ACTION}.target_x", 0.0))
+        ty = float(new_action.pop(f"{ACTION}.target_y", 0.0))
+        tz = float(new_action.pop(f"{ACTION}.target_z", 0.0))
+        wx = float(new_action.pop(f"{ACTION}.target_wx", 0.0))
+        wy = float(new_action.pop(f"{ACTION}.target_wy", 0.0))
+        wz = float(new_action.pop(f"{ACTION}.target_wz", 0.0))
 
         desired = None
 
         if enabled:
-            # Latch a reference at the rising edge; also be defensive if None
-            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
+            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(
                 [
@@ -100,7 +109,6 @@ class EEReferenceAndDelta:
                 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
@@ -116,28 +124,42 @@ class EEReferenceAndDelta:
         # Write action fields
         pos = desired[:3, 3]
         tw = Rotation.from_matrix(desired[:3, :3]).as_rotvec()
-        act.update(
-            {
-                "action.ee.x": float(pos[0]),
-                "action.ee.y": float(pos[1]),
-                "action.ee.z": float(pos[2]),
-                "action.ee.wx": float(tw[0]),
-                "action.ee.wy": float(tw[1]),
-                "action.ee.wz": float(tw[2]),
-            }
-        )
+        new_action[f"{ACTION}.ee.x"] = float(pos[0])
+        new_action[f"{ACTION}.ee.y"] = float(pos[1])
+        new_action[f"{ACTION}.ee.z"] = float(pos[2])
+        new_action[f"{ACTION}.ee.wx"] = float(tw[0])
+        new_action[f"{ACTION}.ee.wy"] = float(tw[1])
+        new_action[f"{ACTION}.ee.wz"] = float(tw[2])
 
         self._prev_enabled = enabled
-        transition[TransitionKey.ACTION] = act
-        return transition
+        return new_action
+
+    def reset(self):
+        self._prev_enabled = False
+        self.reference_ee_pose = None
+        self._command_when_disabled = None
 
     def transform_features(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
+        features.pop(f"{ACTION}.enabled", None)
+        features.pop(f"{ACTION}.target_x", None)
+        features.pop(f"{ACTION}.target_y", None)
+        features.pop(f"{ACTION}.target_z", None)
+        features.pop(f"{ACTION}.target_wx", None)
+        features.pop(f"{ACTION}.target_wy", None)
+        features.pop(f"{ACTION}.target_wz", None)
+
+        features[f"{ACTION}.ee.x"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        features[f"{ACTION}.ee.y"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        features[f"{ACTION}.ee.z"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        features[f"{ACTION}.ee.wx"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        features[f"{ACTION}.ee.wy"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        features[f"{ACTION}.ee.wz"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
         return features
 
 
 @ProcessorStepRegistry.register("ee_bounds_and_safety")
 @dataclass
-class EEBoundsAndSafety(ActionProcessor):
+class EEBoundsAndSafety(ActionProcessorStep):
     """
     Clip the end-effector pose to the bounds and check for jumps.
 
@@ -156,17 +178,20 @@ class EEBoundsAndSafety(ActionProcessor):
     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 | None) -> dict:
-        x = act.pop("action.ee.x", None)
-        y = act.pop("action.ee.y", None)
-        z = act.pop("action.ee.z", None)
-        wx = act.pop("action.ee.wx", None)
-        wy = act.pop("action.ee.wy", None)
-        wz = act.pop("action.ee.wz", None)
+    def action(self, act: dict) -> dict:
+        x = act.get(f"{ACTION}.ee.x", None)
+        y = act.get(f"{ACTION}.ee.y", None)
+        z = act.get(f"{ACTION}.ee.z", None)
+        wx = act.get(f"{ACTION}.ee.wx", None)
+        wy = act.get(f"{ACTION}.ee.wy", None)
+        wz = act.get(f"{ACTION}.ee.wz", None)
 
         if None in (x, y, z, wx, wy, wz):
-            return act
+            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)
@@ -185,35 +210,27 @@ class EEBoundsAndSafety(ActionProcessor):
         self._last_pos = pos
         self._last_twist = twist
 
-        act.update(
-            {
-                "action.ee.x": float(pos[0]),
-                "action.ee.y": float(pos[1]),
-                "action.ee.z": float(pos[2]),
-                "action.ee.wx": float(twist[0]),
-                "action.ee.wy": float(twist[1]),
-                "action.ee.wz": float(twist[2]),
-            }
-        )
+        act[f"{ACTION}.ee.x"] = float(pos[0])
+        act[f"{ACTION}.ee.y"] = float(pos[1])
+        act[f"{ACTION}.ee.z"] = float(pos[2])
+        act[f"{ACTION}.ee.wx"] = float(twist[0])
+        act[f"{ACTION}.ee.wy"] = float(twist[1])
+        act[f"{ACTION}.ee.wz"] = float(twist[2])
         return act
 
     def reset(self):
         self._last_pos = None
+        self._last_twist = None
 
     def transform_features(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
-        # Because this is last step we specify the dataset features of this step that we want to be stored in the dataset
-        features["action.ee.x"] = float
-        features["action.ee.y"] = float
-        features["action.ee.z"] = float
-        features["action.ee.wx"] = float
-        features["action.ee.wy"] = float
-        features["action.ee.wz"] = float
+        # check if features as f"{ACTION}.ee.{x,y,z,wx,wy,wz}"
+
         return features
 
 
 @ProcessorStepRegistry.register("inverse_kinematics_ee_to_joints")
 @dataclass
-class InverseKinematicsEEToJoints:
+class InverseKinematicsEEToJoints(ProcessorStep):
     """
     Compute the desired joint positions from the desired end-effector pose.
 
@@ -238,30 +255,19 @@ class InverseKinematicsEEToJoints:
     initial_guess_current_joints: bool = True
 
     def __call__(self, transition: EnvTransition) -> EnvTransition:
-        act = transition.get(TransitionKey.ACTION) or {}
-        comp = transition.get(TransitionKey.COMPLEMENTARY_DATA) or {}
+        new_transition = transition.copy()
+        act = new_transition.get(TransitionKey.ACTION) or {}
+        comp = new_transition.get(TransitionKey.COMPLEMENTARY_DATA) or {}
 
-        x = act.get("action.ee.x", None)
-        y = act.get("action.ee.y", None)
-        z = act.get("action.ee.z", None)
-        wx = act.get("action.ee.wx", None)
-        wy = act.get("action.ee.wy", None)
-        wz = act.get("action.ee.wz", None)
+        x = act.get(f"{ACTION}.ee.x", None)
+        y = act.get(f"{ACTION}.ee.y", None)
+        z = act.get(f"{ACTION}.ee.z", None)
+        wx = act.get(f"{ACTION}.ee.wx", None)
+        wy = act.get(f"{ACTION}.ee.wy", None)
+        wz = act.get(f"{ACTION}.ee.wz", None)
 
         if None in (x, y, z, wx, wy, wz):
-            # Nothing to do; restore what we popped and return
-            act.update(
-                {
-                    "action.ee.x": x,
-                    "action.ee.y": y,
-                    "action.ee.z": z,
-                    "action.ee.wx": wx,
-                    "action.ee.wy": wy,
-                    "action.ee.wz": wz,
-                }
-            )
-            transition[TransitionKey.ACTION] = act
-            return transition
+            return new_transition
 
         # Get joint positions from complimentary data
         raw = comp.get("raw_joint_positions", None)
@@ -288,23 +294,21 @@ class InverseKinematicsEEToJoints:
         new_act = dict(act)
         for i, name in enumerate(self.motor_names):
             if name == "gripper":
-                new_act["observation.state.gripper.pos"] = float(raw["gripper"])
+                # TODO(pepijn): Investigate if this is correct
+                # Do we want an observation key in the action field?
+                new_act[f"{ACTION}.gripper.pos"] = float(raw["gripper"])
             else:
-                new_act[f"action.{name}.pos"] = float(q_target[i])
-        transition[TransitionKey.ACTION] = new_act
-        return transition
+                new_act[f"{ACTION}.{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[str, PolicyFeature]) -> dict[str, PolicyFeature]:
-        # We specify the dataset features of this step that we want to be stored in the dataset
-        features["action.ee.x"] = float
-        features["action.ee.y"] = float
-        features["action.ee.z"] = float
-        features["action.ee.wx"] = float
-        features["action.ee.wy"] = float
-        features["action.ee.wz"] = float
+        features[f"{ACTION}.gripper.pos"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        for name in self.motor_names:
+            features[f"{ACTION}.{name}.pos"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
 
-        features["observation.state.gripper.pos"] = float
-        features["action.gripper.pos"] = float
         return features
 
     def reset(self):
@@ -313,7 +317,7 @@ class InverseKinematicsEEToJoints:
 
 @ProcessorStepRegistry.register("gripper_velocity_to_joint")
 @dataclass
-class GripperVelocityToJoint:
+class GripperVelocityToJoint(ProcessorStep):
     """
     Convert the gripper velocity to a joint velocity.
 
@@ -332,49 +336,60 @@ class GripperVelocityToJoint:
     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:
-        obs = transition.get(TransitionKey.OBSERVATION) or {}
-        act = transition.get(TransitionKey.ACTION) or {}
-        comp = transition.get(TransitionKey.COMPLEMENTARY_DATA) or {}
+        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 "action.gripper" not in act:
-            return transition
+        if f"{ACTION}.gripper" not in act:
+            raise ValueError(f"Required action key '{ACTION}.gripper' not found in transition")
 
         if "gripper" not in self.motor_names:
-            new_act = dict(act)
-            new_act.pop("action.gripper", None)
-            transition[TransitionKey.ACTION] = new_act
-            return transition
+            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(f"{ACTION}.gripper", 1.0)
+            gripper_action = gripper_action - 1.0
+            gripper_action *= self.clip_max
+            act[f"{ACTION}.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("action.gripper", 0.0))
+        u = float(act.get(f"{ACTION}.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["action.gripper.pos"] = gripper_pos
-        new_act.pop("action.gripper", None)
-        transition[TransitionKey.ACTION] = new_act
+        new_act[f"{ACTION}.gripper.pos"] = gripper_pos
+        new_act.pop(f"{ACTION}.gripper", None)
+        new_transition[TransitionKey.ACTION] = new_act
 
-        obs.update({"observation.state.gripper.pos": curr_pos})
-        transition[TransitionKey.OBSERVATION] = obs
-        return transition
+        obs[f"{OBS_STATE}.gripper.pos"] = curr_pos
+        new_transition[TransitionKey.OBSERVATION] = obs
+        return new_transition
 
     def transform_features(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
-        # We specify the dataset features of this step that we want to be stored in the dataset
-        features["observation.state.gripper.pos"] = float
-        features["action.gripper.pos"] = float
+        features.pop(f"{ACTION}.gripper", None)
+        features[f"{ACTION}.gripper.pos"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        features[f"{OBS_STATE}.gripper.pos"] = PolicyFeature(type=FeatureType.STATE, shape=(1,))
+
         return features
 
 
 @ProcessorStepRegistry.register("forward_kinematics_joints_to_ee")
 @dataclass
-class ForwardKinematicsJointsToEE(ObservationProcessor):
+class ForwardKinematicsJointsToEE(ObservationProcessorStep):
     """
     Compute the end-effector pose from the joint positions.
 
@@ -392,37 +407,33 @@ class ForwardKinematicsJointsToEE(ObservationProcessor):
     kinematics: RobotKinematics
     motor_names: list[str]
 
-    def observation(self, obs: dict | None) -> dict:
-        if not all(f"observation.state.{n}.pos" in obs for n in self.motor_names):
-            return obs
+    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"observation.state.{n}.pos"] for n in self.motor_names], dtype=float)
+        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.update(
-            {
-                "observation.state.ee.x": float(pos[0]),
-                "observation.state.ee.y": float(pos[1]),
-                "observation.state.ee.z": float(pos[2]),
-                "observation.state.ee.wx": float(tw[0]),
-                "observation.state.ee.wy": float(tw[1]),
-                "observation.state.ee.wz": float(tw[2]),
-            }
-        )
+        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[str, PolicyFeature]) -> 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[f"observation.state.ee.{k}"] = float
+            features[f"{OBS_STATE}.ee.{k}"] = PolicyFeature(type=FeatureType.STATE, shape=(1,))
         return features
 
 
 @ProcessorStepRegistry.register("add_robot_observation")
 @dataclass
-class AddRobotObservationAsComplimentaryData(ComplementaryDataProcessor):
+class AddRobotObservationAsComplimentaryData(ComplementaryDataProcessorStep):
     """
     Read the robot's current observation and insert it into the transition as complementary data.
 
@@ -433,15 +444,17 @@ class AddRobotObservationAsComplimentaryData(ComplementaryDataProcessor):
     robot: Robot
 
     def complementary_data(self, comp: dict | None) -> dict:
-        comp = {} if comp is None else dict(comp)
-        obs = self.robot.get_observation()
+        new_comp = dict(comp)
+        obs = (
+            self.robot.get_observation()
+        )  # todo(steven): why not self.trtansition.get(TransitionKey.OBSERVATION)?
 
-        comp["raw_joint_positions"] = {
+        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 comp
+        return new_comp
 
     def transform_features(self, features: dict[str, PolicyFeature]) -> 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/so101_follower/so101_follower.py

@@ -157,6 +157,13 @@ class SO101Follower(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 the max torque limit 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/utils.py

@@ -29,10 +29,6 @@ def make_robot_from_config(config: RobotConfig) -> Robot:
         from .so100_follower import SO100Follower
 
         return SO100Follower(config)
-    elif config.type == "so100_follower_end_effector":
-        from .so100_follower import SO100FollowerEndEffector
-
-        return SO100FollowerEndEffector(config)
     elif config.type == "so101_follower":
         from .so101_follower import SO101Follower
 

src/lerobot/robots/viperx/README.md

@@ -141,10 +141,10 @@ python lerobot/scripts/control_robot.py \
 
 ## Train a policy
 
-To train a policy to control your robot, use the [`python -m lerobot.scripts.train`](../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`](../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}/aloha_test \
   --policy.type=act \
   --output_dir=outputs/train/act_aloha_test \

src/lerobot/scripts/eval.py

@@ -21,7 +21,7 @@ You want to evaluate a model from the hub (eg: https://huggingface.co/lerobot/di
 for 10 episodes.
 
 ```
-python -m lerobot.scripts.eval \
+lerobot-eval \
     --policy.path=lerobot/diffusion_pusht \
     --env.type=pusht \
     --eval.batch_size=10 \
@@ -32,7 +32,7 @@ python -m lerobot.scripts.eval \
 
 OR, you want to evaluate a model checkpoint from the LeRobot training script for 10 episodes.
 ```
-python -m lerobot.scripts.eval \
+lerobot-eval \
     --policy.path=outputs/train/diffusion_pusht/checkpoints/005000/pretrained_model \
     --env.type=pusht \
     --eval.batch_size=10 \

src/lerobot/scripts/rl/actor.py

@@ -62,9 +62,16 @@ from lerobot.configs import parser
 from lerobot.configs.train import TrainRLServerPipelineConfig
 from lerobot.policies.factory import make_policy
 from lerobot.policies.sac.modeling_sac import SACPolicy
+from lerobot.processor import TransitionKey
 from lerobot.robots import so100_follower  # noqa: F401
-from lerobot.scripts.rl.gym_manipulator import make_robot_env
+from lerobot.scripts.rl.gym_manipulator import (
+    create_transition,
+    make_processors,
+    make_robot_env,
+    step_env_and_process_transition,
+)
 from lerobot.teleoperators import gamepad, so101_leader  # noqa: F401
+from lerobot.teleoperators.utils import TeleopEvents
 from lerobot.transport import services_pb2, services_pb2_grpc
 from lerobot.transport.utils import (
     bytes_to_state_dict,
@@ -91,7 +98,6 @@ from lerobot.utils.utils import (
 
 ACTOR_SHUTDOWN_TIMEOUT = 30
 
-
 #################################################
 # Main entry point #
 #################################################
@@ -236,7 +242,8 @@ def act_with_policy(
 
     logging.info("make_env online")
 
-    online_env = make_robot_env(cfg=cfg.env)
+    online_env, teleop_device = make_robot_env(cfg=cfg.env)
+    env_processor, action_processor = make_processors(online_env, teleop_device, cfg.env, cfg.policy.device)
 
     set_seed(cfg.seed)
     device = get_safe_torch_device(cfg.policy.device, log=True)
@@ -257,6 +264,12 @@ def act_with_policy(
     assert isinstance(policy, nn.Module)
 
     obs, info = online_env.reset()
+    env_processor.reset()
+    action_processor.reset()
+
+    # Process initial observation
+    transition = create_transition(observation=obs, info=info)
+    transition = env_processor(transition)
 
     # NOTE: For the moment we will solely handle the case of a single environment
     sum_reward_episode = 0
@@ -274,45 +287,71 @@ def act_with_policy(
             logging.info("[ACTOR] Shutting down act_with_policy")
             return
 
-        if interaction_step >= cfg.policy.online_step_before_learning:
-            # Time policy inference and check if it meets FPS requirement
-            with policy_timer:
-                action = policy.select_action(batch=obs)
-            policy_fps = policy_timer.fps_last
+        observation = {
+            k: v for k, v in transition[TransitionKey.OBSERVATION].items() if k in cfg.policy.input_features
+        }
 
-            log_policy_frequency_issue(policy_fps=policy_fps, cfg=cfg, interaction_step=interaction_step)
+        # Time policy inference and check if it meets FPS requirement
+        with policy_timer:
+            # Extract observation from transition for policy
+            action = policy.select_action(batch=observation)
+        policy_fps = policy_timer.fps_last
 
-        else:
-            action = online_env.action_space.sample()
+        log_policy_frequency_issue(policy_fps=policy_fps, cfg=cfg, interaction_step=interaction_step)
 
-        next_obs, reward, done, truncated, info = online_env.step(action)
+        # Use the new step function
+        new_transition = step_env_and_process_transition(
+            env=online_env,
+            transition=transition,
+            action=action,
+            env_processor=env_processor,
+            action_processor=action_processor,
+        )
+
+        # Extract values from processed transition
+        next_observation = {
+            k: v
+            for k, v in new_transition[TransitionKey.OBSERVATION].items()
+            if k in cfg.policy.input_features
+        }
+
+        # Teleop action is the action that was executed in the environment
+        # It is either the action from the teleop device or the action from the policy
+        executed_action = new_transition[TransitionKey.COMPLEMENTARY_DATA]["teleop_action"]
+
+        reward = new_transition[TransitionKey.REWARD]
+        done = new_transition.get(TransitionKey.DONE, False)
+        truncated = new_transition.get(TransitionKey.TRUNCATED, False)
 
         sum_reward_episode += float(reward)
-        # Increment total steps counter for intervention rate
         episode_total_steps += 1
 
-        # NOTE: We override the action if the intervention is True, because the action applied is the intervention action
-        if "is_intervention" in info and info["is_intervention"]:
-            # NOTE: The action space for demonstration before hand is with the full action space
-            # but sometimes for example we want to deactivate the gripper
-            action = info["action_intervention"]
+        # Check for intervention from transition info
+        intervention_info = new_transition[TransitionKey.INFO]
+        if intervention_info.get(TeleopEvents.IS_INTERVENTION, False):
             episode_intervention = True
-            # Increment intervention steps counter
             episode_intervention_steps += 1
 
+        complementary_info = {
+            "discrete_penalty": torch.tensor(
+                [new_transition[TransitionKey.COMPLEMENTARY_DATA].get("discrete_penalty", 0.0)]
+            ),
+        }
+        # Create transition for learner (convert to old format)
         list_transition_to_send_to_learner.append(
             Transition(
-                state=obs,
-                action=action,
+                state=observation,
+                action=executed_action,
                 reward=reward,
-                next_state=next_obs,
+                next_state=next_observation,
                 done=done,
-                truncated=truncated,  # TODO: (azouitine) Handle truncation properly
-                complementary_info=info,
+                truncated=truncated,
+                complementary_info=complementary_info,
             )
         )
-        # assign obs to the next obs and continue the rollout
-        obs = next_obs
+
+        # Update transition for next iteration
+        transition = new_transition
 
         if done or truncated:
             logging.info(f"[ACTOR] Global step {interaction_step}: Episode reward: {sum_reward_episode}")
@@ -347,12 +386,20 @@ def act_with_policy(
                 )
             )
 
-            # Reset intervention counters
+            # Reset intervention counters and environment
             sum_reward_episode = 0.0
             episode_intervention = False
             episode_intervention_steps = 0
             episode_total_steps = 0
+
+            # Reset environment and processors
             obs, info = online_env.reset()
+            env_processor.reset()
+            action_processor.reset()
+
+            # Process initial observation
+            transition = create_transition(observation=obs, info=info)
+            transition = env_processor(transition)
 
         if cfg.env.fps is not None:
             dt_time = time.perf_counter() - start_time

src/lerobot/scripts/rl/learner.py

@@ -75,6 +75,7 @@ from lerobot.policies.sac.modeling_sac import SACPolicy
 from lerobot.robots import so100_follower  # noqa: F401
 from lerobot.scripts.rl import learner_service
 from lerobot.teleoperators import gamepad, so101_leader  # noqa: F401
+from lerobot.teleoperators.utils import TeleopEvents
 from lerobot.transport import services_pb2_grpc
 from lerobot.transport.utils import (
     MAX_MESSAGE_SIZE,
@@ -1174,7 +1175,7 @@ def process_transitions(
 
             # Add to offline buffer if it's an intervention
             if dataset_repo_id is not None and transition.get("complementary_info", {}).get(
-                "is_intervention"
+                TeleopEvents.IS_INTERVENTION
             ):
                 offline_replay_buffer.add(**transition)
 

src/lerobot/scripts/server/robot_client.py

@@ -302,11 +302,6 @@ class RobotClient:
 
                     self.logger.debug(f"Current latest action: {latest_action}")
 
-                    # Get queue state before changes
-                    old_size, old_timesteps = self._inspect_action_queue()
-                    if not old_timesteps:
-                        old_timesteps = [latest_action]  # queue was empty
-
                     # Get queue state before changes
                     old_size, old_timesteps = self._inspect_action_queue()
                     if not old_timesteps:

src/lerobot/scripts/train.py

@@ -26,12 +26,13 @@ from torch.optim import Optimizer
 
 from lerobot.configs import parser
 from lerobot.configs.train import TrainPipelineConfig
+from lerobot.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
 from lerobot.datasets.factory import make_dataset
 from lerobot.datasets.sampler import EpisodeAwareSampler
 from lerobot.datasets.utils import cycle
 from lerobot.envs.factory import make_env
 from lerobot.optim.factory import make_optimizer_and_scheduler
-from lerobot.policies.factory import make_policy, make_processor
+from lerobot.policies.factory import make_policy, make_pre_post_processors
 from lerobot.policies.pretrained import PreTrainedPolicy
 from lerobot.policies.utils import get_device_from_parameters
 from lerobot.scripts.eval import eval_policy
@@ -140,7 +141,7 @@ def train(cfg: TrainPipelineConfig):
         cfg=cfg.policy,
         ds_meta=dataset.meta,
     )
-    preprocessor, postprocessor = make_processor(
+    preprocessor, postprocessor = make_pre_post_processors(
         policy_cfg=cfg.policy, pretrained_path=cfg.policy.pretrained_path, dataset_stats=dataset.meta.stats
     )
 
@@ -152,6 +153,12 @@ def train(cfg: TrainPipelineConfig):
 
     if cfg.resume:
         step, optimizer, lr_scheduler = load_training_state(cfg.checkpoint_path, optimizer, lr_scheduler)
+        preprocessor.from_pretrained(
+            cfg.policy.pretrained_path, config_filename=f"{POLICY_PREPROCESSOR_DEFAULT_NAME}.json"
+        )
+        postprocessor.from_pretrained(
+            cfg.policy.pretrained_path, config_filename=f"{POLICY_POSTPROCESSOR_DEFAULT_NAME}.json"
+        )
 
     num_learnable_params = sum(p.numel() for p in policy.parameters() if p.requires_grad)
     num_total_params = sum(p.numel() for p in policy.parameters())
@@ -209,10 +216,6 @@ def train(cfg: TrainPipelineConfig):
         batch = preprocessor(batch)
         train_tracker.dataloading_s = time.perf_counter() - start_time
 
-        for key in batch:
-            if isinstance(batch[key], torch.Tensor):
-                batch[key] = batch[key].to(device, non_blocking=device.type == "cuda")
-
         train_tracker, output_dict = update_policy(
             train_tracker,
             policy,
@@ -244,7 +247,9 @@ def train(cfg: TrainPipelineConfig):
         if cfg.save_checkpoint and is_saving_step:
             logging.info(f"Checkpoint policy after step {step}")
             checkpoint_dir = get_step_checkpoint_dir(cfg.output_dir, cfg.steps, step)
-            save_checkpoint(checkpoint_dir, step, cfg, policy, optimizer, lr_scheduler, preprocessor)
+            save_checkpoint(
+                checkpoint_dir, step, cfg, policy, optimizer, lr_scheduler, preprocessor, postprocessor
+            )
             update_last_checkpoint(checkpoint_dir)
             if wandb_logger:
                 wandb_logger.log_policy(checkpoint_dir)
@@ -288,10 +293,8 @@ def train(cfg: TrainPipelineConfig):
 
     if cfg.policy.push_to_hub:
         policy.push_model_to_hub(cfg)
-        if preprocessor:
-            preprocessor.push_to_hub(cfg.policy.repo_id)
-        if postprocessor:
-            postprocessor.push_to_hub(cfg.policy.repo_id)
+        preprocessor.push_to_hub(cfg.policy.repo_id)
+        postprocessor.push_to_hub(cfg.policy.repo_id)
 
 
 def main():

src/lerobot/setup_motors.py

@@ -18,7 +18,7 @@ Helper to set motor ids and baudrate.
 Example:
 
 ```shell
-python -m lerobot.setup_motors \
+lerobot-setup-motors \
     --teleop.type=so100_leader \
     --teleop.port=/dev/tty.usbmodem575E0031751
 ```

src/lerobot/teleoperate.py

@@ -18,7 +18,7 @@ Simple script to control a robot from teleoperation.
 Example:
 
 ```shell
-python -m lerobot.teleoperate \
+lerobot-teleoperate \
     --robot.type=so101_follower \
     --robot.port=/dev/tty.usbmodem58760431541 \
     --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 1920, height: 1080, fps: 30}}" \
@@ -32,7 +32,7 @@ python -m lerobot.teleoperate \
 Example teleoperation with bimanual so100:
 
 ```shell
-python -m lerobot.teleoperate \
+lerobot-teleoperate \
   --robot.type=bi_so100_follower \
   --robot.left_arm_port=/dev/tty.usbmodem5A460851411 \
   --robot.right_arm_port=/dev/tty.usbmodem5A460812391 \
@@ -56,11 +56,17 @@ import time
 from dataclasses import asdict, dataclass
 from pprint import pformat
 
-import draccus
 import rerun as rr
 
 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig  # noqa: F401
 from lerobot.cameras.realsense.configuration_realsense import RealSenseCameraConfig  # noqa: F401
+from lerobot.configs import parser
+from lerobot.processor import IdentityProcessorStep, RobotProcessorPipeline
+from lerobot.processor.converters import (
+    action_to_transition,
+    observation_to_transition,
+    transition_to_robot_action,
+)
 from lerobot.robots import (  # noqa: F401
     Robot,
     RobotConfig,
@@ -97,39 +103,84 @@ class TeleoperateConfig:
     teleop_time_s: float | None = None
     # Display all cameras on screen
     display_data: bool = False
+    # Optional processors for data transformation
+    teleop_action_processor: RobotProcessorPipeline | None = None  # runs after teleop
+    robot_action_processor: RobotProcessorPipeline | None = None  # runs before robot
+    robot_observation_processor: RobotProcessorPipeline | None = None  # runs after robot
 
 
 def teleop_loop(
-    teleop: Teleoperator, robot: Robot, fps: int, display_data: bool = False, duration: float | None = None
+    teleop: Teleoperator,
+    robot: Robot,
+    fps: int,
+    display_data: bool = False,
+    duration: float | None = None,
+    teleop_action_processor: RobotProcessorPipeline | None = None,
+    robot_action_processor: RobotProcessorPipeline | None = None,
+    robot_observation_processor: RobotProcessorPipeline | None = None,
 ):
+    # Initialize processors with defaults if not provided
+    teleop_action_processor = teleop_action_processor or RobotProcessorPipeline(
+        steps=[IdentityProcessorStep()], to_transition=action_to_transition, to_output=lambda tr: tr
+    )
+    robot_action_processor = robot_action_processor or RobotProcessorPipeline(
+        steps=[IdentityProcessorStep()],
+        to_transition=lambda tr: tr,
+        to_output=transition_to_robot_action,  # type: ignore[arg-type]
+    )
+    robot_observation_processor = robot_observation_processor or RobotProcessorPipeline(
+        steps=[IdentityProcessorStep()],
+        to_transition=observation_to_transition,
+        to_output=lambda tr: tr,
+    )
+
+    # Reset processors
+    teleop_action_processor.reset()
+    robot_action_processor.reset()
+    robot_observation_processor.reset()
+
     display_len = max(len(key) for key in robot.action_features)
     start = time.perf_counter()
+
     while True:
         loop_start = time.perf_counter()
-        action = teleop.get_action()
-        if display_data:
-            observation = robot.get_observation()
-            log_rerun_data(observation=observation, action=action)
 
-        robot.send_action(action)
+        # Get teleop action
+        raw_action = teleop.get_action()
+
+        # Process teleop action through pipeline
+        teleop_transition = teleop_action_processor(raw_action)
+
+        # Process action for robot through pipeline
+        robot_action_to_send = robot_action_processor(teleop_transition)
+
+        # Send processed action to robot (robot_action_processor.to_output should return dict[str, Any])
+        robot.send_action(robot_action_to_send)  # type: ignore[arg-type]
+
+        if display_data:
+            # Get robot observation
+            obs = robot.get_observation()
+            # Process robot observation through pipeline
+            obs_transition = robot_observation_processor(obs)
+            log_rerun_data([obs_transition, teleop_transition])
+
+            print("\n" + "-" * (display_len + 10))
+            print(f"{'NAME':<{display_len}} | {'NORM':>7}")
+            # Display the final robot action that was sent
+            for motor, value in robot_action_to_send.items():
+                print(f"{motor:<{display_len}} | {value:>7.2f}")
+            move_cursor_up(len(robot_action_to_send) + 5)
+
         dt_s = time.perf_counter() - loop_start
         busy_wait(1 / fps - dt_s)
-
         loop_s = time.perf_counter() - loop_start
-
-        print("\n" + "-" * (display_len + 10))
-        print(f"{'NAME':<{display_len}} | {'NORM':>7}")
-        for motor, value in action.items():
-            print(f"{motor:<{display_len}} | {value:>7.2f}")
         print(f"\ntime: {loop_s * 1e3:.2f}ms ({1 / loop_s:.0f} Hz)")
 
         if duration is not None and time.perf_counter() - start >= duration:
             return
 
-        move_cursor_up(len(action) + 5)
 
-
-@draccus.wrap()
+@parser.wrap()
 def teleoperate(cfg: TeleoperateConfig):
     init_logging()
     logging.info(pformat(asdict(cfg)))
@@ -143,7 +194,16 @@ def teleoperate(cfg: TeleoperateConfig):
     robot.connect()
 
     try:
-        teleop_loop(teleop, robot, cfg.fps, display_data=cfg.display_data, duration=cfg.teleop_time_s)
+        teleop_loop(
+            teleop=teleop,
+            robot=robot,
+            fps=cfg.fps,
+            display_data=cfg.display_data,
+            duration=cfg.teleop_time_s,
+            teleop_action_processor=cfg.teleop_action_processor,
+            robot_action_processor=cfg.robot_action_processor,
+            robot_observation_processor=cfg.robot_observation_processor,
+        )
     except KeyboardInterrupt:
         pass
     finally:

src/lerobot/teleoperators/__init__.py

@@ -16,4 +16,4 @@
 
 from .config import TeleoperatorConfig
 from .teleoperator import Teleoperator
-from .utils import make_teleoperator_from_config
+from .utils import TeleopEvents, make_teleoperator_from_config

src/lerobot/teleoperators/gamepad/gamepad_utils.py

@@ -16,6 +16,8 @@
 
 import logging
 
+from ..utils import TeleopEvents
+
 
 class InputController:
     """Base class for input controllers that generate motion deltas."""
@@ -134,10 +136,10 @@ class KeyboardController(InputController):
                     return False
                 elif key == keyboard.Key.enter:
                     self.key_states["success"] = True
-                    self.episode_end_status = "success"
+                    self.episode_end_status = TeleopEvents.SUCCESS
                 elif key == keyboard.Key.backspace:
                     self.key_states["failure"] = True
-                    self.episode_end_status = "failure"
+                    self.episode_end_status = TeleopEvents.FAILURE
             except AttributeError:
                 pass
 
@@ -255,13 +257,13 @@ class GamepadController(InputController):
         for event in pygame.event.get():
             if event.type == pygame.JOYBUTTONDOWN:
                 if event.button == 3:
-                    self.episode_end_status = "success"
+                    self.episode_end_status = TeleopEvents.SUCCESS
                 # A button (1) for failure
                 elif event.button == 1:
-                    self.episode_end_status = "failure"
+                    self.episode_end_status = TeleopEvents.FAILURE
                 # X button (0) for rerecord
                 elif event.button == 0:
-                    self.episode_end_status = "rerecord_episode"
+                    self.episode_end_status = TeleopEvents.RERECORD_EPISODE
 
                 # RB button (6) for closing gripper
                 elif event.button == 6:
@@ -451,11 +453,11 @@ class GamepadControllerHID(InputController):
                 # Check if X/Square button (bit 5) is pressed for failure
                 # Check if A/Cross button (bit 4) is pressed for rerecording
                 if buttons & 1 << 7:
-                    self.episode_end_status = "success"
+                    self.episode_end_status = TeleopEvents.SUCCESS
                 elif buttons & 1 << 5:
-                    self.episode_end_status = "failure"
+                    self.episode_end_status = TeleopEvents.FAILURE
                 elif buttons & 1 << 4:
-                    self.episode_end_status = "rerecord_episode"
+                    self.episode_end_status = TeleopEvents.RERECORD_EPISODE
                 else:
                     self.episode_end_status = None
 

src/lerobot/teleoperators/gamepad/teleop_gamepad.py

@@ -21,6 +21,7 @@ from typing import Any
 import numpy as np
 
 from ..teleoperator import Teleoperator
+from ..utils import TeleopEvents
 from .configuration_gamepad import GamepadTeleopConfig
 
 
@@ -93,9 +94,9 @@ class GamepadTeleop(Teleoperator):
         gamepad_action = np.array([delta_x, delta_y, delta_z], dtype=np.float32)
 
         action_dict = {
-            "delta_x": gamepad_action[0],
-            "delta_y": gamepad_action[1],
-            "delta_z": gamepad_action[2],
+            "action.delta_x": gamepad_action[0],
+            "action.delta_y": gamepad_action[1],
+            "action.delta_z": gamepad_action[2],
         }
 
         # Default gripper action is to stay
@@ -107,6 +108,48 @@ class GamepadTeleop(Teleoperator):
 
         return action_dict
 
+    def get_teleop_events(self) -> dict[str, Any]:
+        """
+        Get extra control events from the gamepad such as intervention status,
+        episode termination, success indicators, etc.
+
+        Returns:
+            Dictionary containing:
+                - is_intervention: bool - Whether 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
+        """
+        if self.gamepad is None:
+            return {
+                TeleopEvents.IS_INTERVENTION: False,
+                TeleopEvents.TERMINATE_EPISODE: False,
+                TeleopEvents.SUCCESS: False,
+                TeleopEvents.RERECORD_EPISODE: False,
+            }
+
+        # Update gamepad state to get fresh inputs
+        self.gamepad.update()
+
+        # Check if intervention is active
+        is_intervention = self.gamepad.should_intervene()
+
+        # Get episode end status
+        episode_end_status = self.gamepad.get_episode_end_status()
+        terminate_episode = episode_end_status in [
+            TeleopEvents.RERECORD_EPISODE,
+            TeleopEvents.FAILURE,
+        ]
+        success = episode_end_status == TeleopEvents.SUCCESS
+        rerecord_episode = episode_end_status == TeleopEvents.RERECORD_EPISODE
+
+        return {
+            TeleopEvents.IS_INTERVENTION: is_intervention,
+            TeleopEvents.TERMINATE_EPISODE: terminate_episode,
+            TeleopEvents.SUCCESS: success,
+            TeleopEvents.RERECORD_EPISODE: rerecord_episode,
+        }
+
     def disconnect(self) -> None:
         """Disconnect from the gamepad."""
         if self.gamepad is not None:

src/lerobot/teleoperators/keyboard/teleop_keyboard.py

@@ -24,6 +24,7 @@ from typing import Any
 from lerobot.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
 
 from ..teleoperator import Teleoperator
+from ..utils import TeleopEvents
 from .configuration_keyboard import KeyboardEndEffectorTeleopConfig, KeyboardTeleopConfig
 
 PYNPUT_AVAILABLE = True
@@ -167,25 +168,15 @@ class KeyboardEndEffectorTeleop(KeyboardTeleop):
             return {
                 "dtype": "float32",
                 "shape": (4,),
-                "names": {"delta_x": 0, "delta_y": 1, "delta_z": 2, "gripper": 3},
+                "names": {"action.delta_x": 0, "action.delta_y": 1, "action.delta_z": 2, "action.gripper": 3},
             }
         else:
             return {
                 "dtype": "float32",
                 "shape": (3,),
-                "names": {"delta_x": 0, "delta_y": 1, "delta_z": 2},
+                "names": {"action.delta_x": 0, "action.delta_y": 1, "action.delta_z": 2},
             }
 
-    def _on_press(self, key):
-        if hasattr(key, "char"):
-            key = key.char
-        self.event_queue.put((key, True))
-
-    def _on_release(self, key):
-        if hasattr(key, "char"):
-            key = key.char
-        self.event_queue.put((key, False))
-
     def get_action(self) -> dict[str, Any]:
         if not self.is_connected:
             raise DeviceNotConnectedError(
@@ -226,12 +217,75 @@ class KeyboardEndEffectorTeleop(KeyboardTeleop):
         self.current_pressed.clear()
 
         action_dict = {
-            "delta_x": delta_x,
-            "delta_y": delta_y,
-            "delta_z": delta_z,
+            "action.delta_x": delta_x,
+            "action.delta_y": delta_y,
+            "action.delta_z": delta_z,
         }
 
         if self.config.use_gripper:
             action_dict["gripper"] = gripper_action
 
         return action_dict
+
+    def get_teleop_events(self) -> dict[str, Any]:
+        """
+        Get extra control events from the keyboard such as intervention status,
+        episode termination, success indicators, etc.
+
+        Keyboard mappings:
+        - Any movement keys pressed = intervention active
+        - 's' key = success (terminate episode successfully)
+        - 'r' key = rerecord episode (terminate and rerecord)
+        - 'q' key = quit episode (terminate without success)
+
+        Returns:
+            Dictionary containing:
+                - is_intervention: bool - Whether 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
+        """
+        if not self.is_connected:
+            return {
+                TeleopEvents.IS_INTERVENTION: False,
+                TeleopEvents.TERMINATE_EPISODE: False,
+                TeleopEvents.SUCCESS: False,
+                TeleopEvents.RERECORD_EPISODE: False,
+            }
+
+        # Check if any movement keys are currently pressed (indicates intervention)
+        movement_keys = [
+            keyboard.Key.up,
+            keyboard.Key.down,
+            keyboard.Key.left,
+            keyboard.Key.right,
+            keyboard.Key.shift,
+            keyboard.Key.shift_r,
+            keyboard.Key.ctrl_r,
+            keyboard.Key.ctrl_l,
+        ]
+        is_intervention = any(self.current_pressed.get(key, False) for key in movement_keys)
+
+        # Check for episode control commands from misc_keys_queue
+        terminate_episode = False
+        success = False
+        rerecord_episode = False
+
+        # Process any pending misc keys
+        while not self.misc_keys_queue.empty():
+            key = self.misc_keys_queue.get_nowait()
+            if key == "s":
+                success = True
+            elif key == "r":
+                terminate_episode = True
+                rerecord_episode = True
+            elif key == "q":
+                terminate_episode = True
+                success = False
+
+        return {
+            TeleopEvents.IS_INTERVENTION: is_intervention,
+            TeleopEvents.TERMINATE_EPISODE: terminate_episode,
+            TeleopEvents.SUCCESS: success,
+            TeleopEvents.RERECORD_EPISODE: rerecord_episode,
+        }

src/lerobot/teleoperators/phone/phone.py

@@ -1,246 +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.
-
-# Docs:
-# hebi: https://docs.hebi.us/tools.html#mobile-io
-# teleop: https://github.com/SpesRobotics/teleop
-
-import logging
-import threading
-import time
-
-import hebi
-import numpy as np
-from scipy.spatial.transform import Rotation
-from teleop import Teleop
-
-from lerobot.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
-from lerobot.teleoperators.phone.config_phone import PhoneConfig, PhoneOS
-from lerobot.teleoperators.teleoperator import Teleoperator
-
-logger = logging.getLogger(__name__)
-
-
-class Phone(Teleoperator):
-    """
-    Phone-based teleoperator using ARKit (iOS via HEBI Mobile I/O App) or the teleop Python package (Android via WebXR API).
-    For HEBI Mobile I/O we also expose 8 analog (a1-a8) and 8 digital (b1-b8) inputs.
-
-    Press and hold **B1** to enable teleoperation. While enabled, the first B1 press
-    captures a reference pose and rotation, when disabled and pressed again the position is reapplied.
-    """
-
-    config_class = PhoneConfig
-    name = "phone"
-
-    def __init__(self, config: PhoneConfig):
-        super().__init__(config)
-        self.config = config
-        self._group = None
-        self._teleop = None
-        self._teleop_thread = None
-        self._latest_pose = None
-        self._latest_message = None
-        self._enabled: bool = False
-        self._calib_pos: np.ndarray | None = None
-        self._calib_rot_inv: Rotation | None = None
-
-    @property
-    def is_connected(self) -> bool:
-        return (self.config.phone_os == PhoneOS.IOS and self._group is not None) or (
-            self.config.phone_os == PhoneOS.ANDROID and self._teleop is not None
-        )
-
-    def connect(self) -> None:
-        if self.is_connected:
-            raise DeviceAlreadyConnectedError(f"{self} already connected")
-
-        if self.config.phone_os == PhoneOS.IOS:
-            logger.info("Connecting to IPhone, make sure to open the HEBI Mobile I/O app.")
-            lookup = hebi.Lookup()
-            time.sleep(2.0)
-            group = lookup.get_group_from_names(["HEBI"], ["mobileIO"])
-            if group is None:
-                raise RuntimeError("Mobile I/O not found — check name/family settings in the app.")
-            self._group = group
-            logger.info(f"{self} connected to HEBI group with {group.size} module(s).")
-        elif self.config.phone_os == PhoneOS.ANDROID:
-            logger.info("Starting teleop stream for Android...")
-            self._teleop = Teleop()
-            self._teleop.subscribe(self._android_callback)
-            self._teleop_thread = threading.Thread(target=self._teleop.run, daemon=True)
-            self._teleop_thread.start()
-            logger.info(f"{self} connected, teleop stream started.")
-        else:
-            raise ValueError(f"Invalid config phone_os: {self.config.phone_os}")
-
-        self.calibrate()
-
-    def calibrate(self) -> None:
-        print(
-            "Hold the phone so that: top edge points forward in same direction as the robot (robot +x) and screen points up (robot +z)"
-        )
-        if self.config.phone_os == PhoneOS.IOS:
-            print("Press and hold B1 in the HEBI Mobile I/O app to capture this pose...\n")
-        else:
-            print("Touch and move on the WebXR page to capture this pose...\n")
-
-        pos, rot = self._wait_for_capture_trigger()
-        self._calib_pos = pos.copy()
-        self._calib_rot_inv = rot.inv()
-        self._enabled = False
-        print("Calibration done\n")
-
-    def _reapply_position_calibration(self, pos: np.ndarray) -> None:
-        self._calib_pos = pos.copy()
-
-    @property
-    def is_calibrated(self) -> bool:
-        return (self._calib_pos is not None) and (self._calib_rot_inv is not None)
-
-    @property
-    def action_features(self) -> dict[str, type]:
-        return {
-            "phone.pos": np.ndarray,  # shape (3,)
-            "phone.rot": Rotation,  # scipy.spatial.transform.Rotation
-            "phone.raw_inputs": dict,  # analogs/buttons or webXR meta
-            "phone.enabled": bool,
-        }
-
-    def _wait_for_capture_trigger(self) -> tuple[np.ndarray, Rotation]:
-        """Wait trigger for calibration: iOS: B1. Android: 'move'."""
-        while True:
-            ok, pos, rot, pose = self._read_current_pose()
-            if not ok:
-                time.sleep(0.01)
-                continue
-
-            if self.config.phone_os == PhoneOS.IOS:
-                io = getattr(pose, "io", None)
-                b = getattr(io, "b", None) if io is not None else None
-                b1 = False
-                if b is not None:
-                    b1 = bool(b.get_int(1))
-                if b1:
-                    return pos, rot
-            else:
-                msg = self._latest_message or {}
-                if bool(msg.get("move", False)):
-                    return pos, rot
-
-            time.sleep(0.01)
-
-    def _read_current_pose(self) -> tuple[bool, np.ndarray | None, Rotation | None, object | None]:
-        if self.config.phone_os == PhoneOS.IOS:
-            fbk = self._group.get_next_feedback()
-            pose = fbk[0]
-            ar_pos = getattr(pose, "ar_position", None)
-            ar_quat = getattr(pose, "ar_orientation", None)
-            if ar_pos is None or ar_quat is None:
-                return False, None, None, None
-            quat_xyzw = np.concatenate((ar_quat[1:], [ar_quat[0]]))  # wxyz to xyzw
-            rot = Rotation.from_quat(quat_xyzw)
-            pos = ar_pos - rot.apply(self.config.camera_offset)
-            return True, pos, rot, pose
-        else:
-            p = self._latest_pose
-            if p is None:
-                return False, None, None, None
-            rot = Rotation.from_matrix(p[:3, :3])
-            pos = p[:3, 3] - rot.apply(self.config.camera_offset)
-            pose = self._latest_pose
-            return True, pos, rot, pose
-
-    @property
-    def feedback_features(self) -> dict[str, type]:
-        # No haptic or other feedback implemented yet
-        pass
-
-    def configure(self) -> None:
-        # No additional configuration required for phone teleop
-        pass
-
-    def _android_callback(self, pose: np.ndarray, message: dict) -> None:
-        self._latest_pose = pose
-        self._latest_message = message
-        time.sleep(0.001)  # 1ms delay to avoid race condition
-
-    def get_action(self) -> dict:
-        ok, raw_pos, raw_rot, pose = self._read_current_pose()
-        if not ok or not self.is_calibrated:
-            return {}
-
-        # Collect raw inputs (B1 / analogs on iOS, move/scale on Android)
-        raw_inputs: dict[str, float | int | bool] = {}
-        if self.config.phone_os == PhoneOS.IOS:
-            io = getattr(pose, "io", None)
-            if io is not None:
-                bank_a, bank_b = io.a, io.b
-                if bank_a:
-                    for ch in range(1, 9):
-                        if bank_a.has_float(ch):
-                            raw_inputs[f"a{ch}"] = float(bank_a.get_float(ch))
-                if bank_b:
-                    for ch in range(1, 9):
-                        if bank_b.has_int(ch):
-                            raw_inputs[f"b{ch}"] = int(bank_b.get_int(ch))
-                        elif hasattr(bank_b, "has_bool") and bank_b.has_bool(ch):
-                            raw_inputs[f"b{ch}"] = int(bank_b.get_bool(ch))
-        else:
-            msg = self._latest_message or {}
-            raw_inputs["move"] = bool(msg.get("move", False))
-            raw_inputs["scale"] = float(msg.get("scale", 1.0))
-            raw_inputs["reservedButtonA"] = bool(msg.get("reservedButtonA", False))
-            raw_inputs["reservedButtonB"] = bool(msg.get("reservedButtonB", False))
-
-        if self.config.phone_os == PhoneOS.IOS:
-            enable = bool(raw_inputs.get("b1", 0))
-        else:
-            enable = bool(raw_inputs.get("move", False))
-
-        # Rising edge then re-capture calibration immediately from current raw pose
-        if enable and not self._enabled:
-            self._reapply_position_calibration(raw_pos)
-
-        # Apply calibration
-        pos_cal = self._calib_rot_inv.apply(raw_pos - self._calib_pos)
-        rot_cal = self._calib_rot_inv * raw_rot
-
-        self._enabled = enable
-
-        return {
-            "phone.pos": pos_cal,
-            "phone.rot": rot_cal,
-            "phone.raw_inputs": raw_inputs,
-            "phone.enabled": self._enabled,
-        }
-
-    def send_feedback(self, feedback: dict[str, float]) -> None:
-        # We could add haptic feedback (vibrations) here, but it's not implemented yet
-        raise NotImplementedError
-
-    def disconnect(self) -> None:
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")
-
-        if self.config.phone_os == PhoneOS.IOS:
-            self._group = None
-        else:
-            self._teleop = None
-            if self._teleop_thread and self._teleop_thread.is_alive():
-                self._teleop_thread.join(timeout=1.0)
-                self._teleop_thread = None
-                self._latest_pose = None

src/lerobot/teleoperators/phone/phone_processor.py

@@ -16,14 +16,15 @@
 
 from dataclasses import dataclass, field
 
-from lerobot.configs.types import PolicyFeature
-from lerobot.processor.pipeline import ActionProcessor, ProcessorStepRegistry
+from lerobot.configs.types import FeatureType, PolicyFeature
+from lerobot.constants import ACTION
+from lerobot.processor import ActionProcessorStep, ProcessorStepRegistry
 from lerobot.teleoperators.phone.config_phone import PhoneOS
 
 
 @ProcessorStepRegistry.register("map_phone_action_to_robot_action")
 @dataclass
-class MapPhoneActionToRobotAction(ActionProcessor):
+class MapPhoneActionToRobotAction(ActionProcessorStep):
     """
     Map calibrated phone pose (actions) to the inputs for robot actions
 
@@ -46,15 +47,15 @@ class MapPhoneActionToRobotAction(ActionProcessor):
     platform: PhoneOS
     _enabled_prev: bool = field(default=False, init=False, repr=False)
 
-    def action(self, act: dict | None) -> dict:
+    def action(self, act: dict) -> dict:
         # Pop them from the action
-        enabled = act.pop("action.phone.enabled", 0)
-        pos = act.pop("action.phone.pos", None)
-        rot = act.pop("action.phone.rot", None)
-        inputs = act.pop("action.phone.raw_inputs", {})
+        enabled = bool(act.pop(f"{ACTION}.phone.enabled", 0))
+        pos = act.pop(f"{ACTION}.phone.pos", None)
+        rot = act.pop(f"{ACTION}.phone.rot", None)
+        inputs = act.pop(f"{ACTION}.phone.raw_inputs", {})
 
         if pos is None or rot is None:
-            return act
+            raise ValueError("pos and rot must be present in action")
 
         rotvec = rot.as_rotvec()  # Absolute orientation as rotvec
 
@@ -69,19 +70,28 @@ class MapPhoneActionToRobotAction(ActionProcessor):
             )  # Positive if a is pressed, negative if b is pressed, 0 if both or neither are pressed
 
         # For some actions we need to invert the axis
-        act.update(
-            {
-                "action.enabled": enabled,
-                "action.target_x": -pos[1] if enabled else 0.0,
-                "action.target_y": pos[0] if enabled else 0.0,
-                "action.target_z": pos[2] if enabled else 0.0,
-                "action.target_wx": rotvec[1] if enabled else 0.0,
-                "action.target_wy": rotvec[0] if enabled else 0.0,
-                "action.target_wz": -rotvec[2] if enabled else 0.0,
-                "action.gripper": gripper,  # Still send gripper action when disabled
-            }
-        )
+        act[f"{ACTION}.enabled"] = enabled
+        act[f"{ACTION}.target_x"] = -pos[1] if enabled else 0.0
+        act[f"{ACTION}.target_y"] = pos[0] if enabled else 0.0
+        act[f"{ACTION}.target_z"] = pos[2] if enabled else 0.0
+        act[f"{ACTION}.target_wx"] = rotvec[1] if enabled else 0.0
+        act[f"{ACTION}.target_wy"] = rotvec[0] if enabled else 0.0
+        act[f"{ACTION}.target_wz"] = -rotvec[2] if enabled else 0.0
+        act[f"{ACTION}.gripper"] = gripper  # Still send gripper action when disabled
         return act
 
     def transform_features(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
+        features.pop(f"{ACTION}.phone.enabled", None)
+        features.pop(f"{ACTION}.phone.pos", None)
+        features.pop(f"{ACTION}.phone.rot", None)
+        features.pop(f"{ACTION}.phone.raw_inputs", None)
+
+        features[f"{ACTION}.enabled"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        features[f"{ACTION}.target_x"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        features[f"{ACTION}.target_y"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        features[f"{ACTION}.target_z"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        features[f"{ACTION}.target_wx"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        features[f"{ACTION}.target_wy"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        features[f"{ACTION}.target_wz"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+        features[f"{ACTION}.gripper"] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
         return features

src/lerobot/teleoperators/phone/teleop_phone.py

@@ -0,0 +1,358 @@
+#!/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.
+
+# Docs:
+# hebi: https://docs.hebi.us/tools.html#mobile-io
+# teleop: https://github.com/SpesRobotics/teleop
+
+import logging
+import threading
+import time
+
+import hebi
+import numpy as np
+from teleop import Teleop
+
+from lerobot.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+from lerobot.teleoperators.phone.config_phone import PhoneConfig, PhoneOS
+from lerobot.teleoperators.teleoperator import Teleoperator
+from lerobot.utils.rotation import Rotation
+
+logger = logging.getLogger(__name__)
+
+
+class BasePhone:
+    _enabled: bool = False
+    _calib_pos: np.ndarray | None = None
+    _calib_rot_inv: Rotation | None = None
+
+    def _reapply_position_calibration(self, pos: np.ndarray) -> None:
+        self._calib_pos = pos.copy()
+
+    @property
+    def is_calibrated(self) -> bool:
+        return (self._calib_pos is not None) and (self._calib_rot_inv is not None)
+
+    @property
+    def action_features(self) -> dict[str, type]:
+        return {
+            "phone.pos": np.ndarray,  # shape (3,)
+            "phone.rot": Rotation,  # scipy.spatial.transform.Rotation
+            "phone.raw_inputs": dict,  # analogs/buttons or webXR meta
+            "phone.enabled": bool,
+        }
+
+    @property
+    def feedback_features(self) -> dict[str, type]:
+        # No haptic or other feedback implemented yet
+        pass
+
+    def configure(self) -> None:
+        # No additional configuration required for phone teleop
+        pass
+
+    def send_feedback(self, feedback: dict[str, float]) -> None:
+        # We could add haptic feedback (vibrations) here, but it's not implemented yet
+        raise NotImplementedError
+
+
+class IOSPhone(BasePhone, Teleoperator):
+    name = "ios_phone"
+
+    def __init__(self, config: PhoneConfig):
+        super().__init__(config)
+        self.config = config
+        self._group = None
+
+    @property
+    def is_connected(self) -> bool:
+        return self._group is not None
+
+    def connect(self) -> None:
+        if self.is_connected:
+            raise DeviceAlreadyConnectedError(f"{self} already connected")
+
+        logger.info("Connecting to IPhone, make sure to open the HEBI Mobile I/O app.")
+        lookup = hebi.Lookup()
+        time.sleep(2.0)
+        group = lookup.get_group_from_names(["HEBI"], ["mobileIO"])
+        if group is None:
+            raise RuntimeError("Mobile I/O not found — check name/family settings in the app.")
+        self._group = group
+        logger.info(f"{self} connected to HEBI group with {group.size} module(s).")
+
+        self.calibrate()
+
+    def calibrate(self) -> None:
+        print(
+            "Hold the phone so that: top edge points forward in same direction as the robot (robot +x) and screen points up (robot +z)"
+        )
+        print("Press and hold B1 in the HEBI Mobile I/O app to capture this pose...\n")
+        position, rotation = self._wait_for_capture_trigger()
+        self._calib_pos = position.copy()
+        self._calib_rot_inv = rotation.inv()
+        self._enabled = False
+        print("Calibration done\n")
+
+    def _wait_for_capture_trigger(self) -> tuple[np.ndarray, Rotation]:
+        """Wait trigger for calibration: iOS: B1. Android: 'move'."""
+        while True:
+            has_pose, position, rotation, fb_pose = self._read_current_pose()
+            if not has_pose:
+                time.sleep(0.01)
+                continue
+
+            io = getattr(fb_pose, "io", None)
+            button_b = getattr(io, "b", None) if io is not None else None
+            button_b1_pressed = False
+            if button_b is not None:
+                button_b1_pressed = bool(button_b.get_int(1))
+            if button_b1_pressed:
+                return position, rotation
+
+            time.sleep(0.01)
+
+    def _read_current_pose(self) -> tuple[bool, np.ndarray | None, Rotation | None, object | None]:
+        fbk = self._group.get_next_feedback()
+        pose = fbk[0]
+        ar_pos = getattr(pose, "ar_position", None)
+        ar_quat = getattr(pose, "ar_orientation", None)
+        if ar_pos is None or ar_quat is None:
+            return False, None, None, None
+        # HEBI provides orientation in w, x, y, z format.
+        # Scipy's Rotation expects x, y, z, w.
+        quat_xyzw = np.concatenate((ar_quat[1:], [ar_quat[0]]))  # wxyz to xyzw
+        rot = Rotation.from_quat(quat_xyzw)
+        pos = ar_pos - rot.apply(self.config.camera_offset)
+        return True, pos, rot, pose
+
+    def get_action(self) -> dict:
+        has_pose, raw_position, raw_rotation, fb_pose = self._read_current_pose()
+        if not has_pose or not self.is_calibrated:
+            return {}
+
+        # Collect raw inputs (B1 / analogs on iOS, move/scale on Android)
+        raw_inputs: dict[str, float | int | bool] = {}
+        io = getattr(fb_pose, "io", None)
+        if io is not None:
+            bank_a, bank_b = io.a, io.b
+            if bank_a:
+                for ch in range(1, 9):
+                    if bank_a.has_float(ch):
+                        raw_inputs[f"a{ch}"] = float(bank_a.get_float(ch))
+            if bank_b:
+                for ch in range(1, 9):
+                    if bank_b.has_int(ch):
+                        raw_inputs[f"b{ch}"] = int(bank_b.get_int(ch))
+                    elif hasattr(bank_b, "has_bool") and bank_b.has_bool(ch):
+                        raw_inputs[f"b{ch}"] = int(bank_b.get_bool(ch))
+
+        enable = bool(raw_inputs.get("b1", 0))
+
+        # Rising edge then re-capture calibration immediately from current raw pose
+        if enable and not self._enabled:
+            self._reapply_position_calibration(raw_position)
+
+        # Apply calibration
+        pos_cal = self._calib_rot_inv.apply(raw_position - self._calib_pos)
+        rot_cal = self._calib_rot_inv * raw_rotation
+
+        self._enabled = enable
+
+        return {
+            "phone.pos": pos_cal,
+            "phone.rot": rot_cal,
+            "phone.raw_inputs": raw_inputs,
+            "phone.enabled": self._enabled,
+        }
+
+    def disconnect(self) -> None:
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+
+        self._group = None
+
+
+class AndroidPhone(BasePhone, Teleoperator):
+    name = "android_phone"
+
+    def __init__(self, config: PhoneConfig):
+        super().__init__(config)
+        self.config = config
+        self._teleop = None
+        self._teleop_thread = None
+        self._latest_pose = None
+        self._latest_message = None
+        self._android_lock = threading.Lock()
+
+    @property
+    def is_connected(self) -> bool:
+        return self._teleop is not None
+
+    def connect(self) -> None:
+        if self.is_connected:
+            raise DeviceAlreadyConnectedError(f"{self} already connected")
+
+        logger.info("Starting teleop stream for Android...")
+        self._teleop = Teleop()
+        self._teleop.subscribe(self._android_callback)
+        self._teleop_thread = threading.Thread(target=self._teleop.run, daemon=True)
+        self._teleop_thread.start()
+        logger.info(f"{self} connected, teleop stream started.")
+
+        self.calibrate()
+
+    def calibrate(self) -> None:
+        print(
+            "Hold the phone so that: top edge points forward in same direction as the robot (robot +x) and screen points up (robot +z)"
+        )
+        print("Touch and move on the WebXR page to capture this pose...\n")
+
+        pos, rot = self._wait_for_capture_trigger()
+        self._calib_pos = pos.copy()
+        self._calib_rot_inv = rot.inv()
+        self._enabled = False
+        print("Calibration done\n")
+
+    def _wait_for_capture_trigger(self) -> tuple[np.ndarray, Rotation]:
+        """Wait trigger for calibration: iOS: B1. Android: 'move'."""
+        while True:
+            with self._android_lock:
+                msg = self._latest_message or {}
+
+            if bool(msg.get("move", False)):
+                ok, pos, rot, _pose = self._read_current_pose()
+                if ok:
+                    return pos, rot
+
+            time.sleep(0.01)
+
+    def _read_current_pose(self) -> tuple[bool, np.ndarray | None, Rotation | None, object | None]:
+        with self._android_lock:
+            if self._latest_pose is None:
+                return False, None, None, None
+            p = self._latest_pose.copy()
+            pose = self._latest_pose
+        rot = Rotation.from_matrix(p[:3, :3])
+        pos = p[:3, 3] - rot.apply(self.config.camera_offset)
+        return True, pos, rot, pose
+
+    def _android_callback(self, pose: np.ndarray, message: dict) -> None:
+        with self._android_lock:
+            self._latest_pose = pose
+            self._latest_message = message
+
+    def get_action(self) -> dict:
+        ok, raw_pos, raw_rot, pose = self._read_current_pose()
+        if not ok or not self.is_calibrated:
+            return {}
+
+        # Collect raw inputs (B1 / analogs on iOS, move/scale on Android)
+        raw_inputs: dict[str, float | int | bool] = {}
+        msg = self._latest_message or {}
+        raw_inputs["move"] = bool(msg.get("move", False))
+        raw_inputs["scale"] = float(msg.get("scale", 1.0))
+        raw_inputs["reservedButtonA"] = bool(msg.get("reservedButtonA", False))
+        raw_inputs["reservedButtonB"] = bool(msg.get("reservedButtonB", False))
+
+        enable = bool(raw_inputs.get("move", False))
+
+        # Rising edge then re-capture calibration immediately from current raw pose
+        if enable and not self._enabled:
+            self._reapply_position_calibration(raw_pos)
+
+        # Apply calibration
+        pos_cal = self._calib_rot_inv.apply(raw_pos - self._calib_pos)
+        rot_cal = self._calib_rot_inv * raw_rot
+
+        self._enabled = enable
+
+        return {
+            "phone.pos": pos_cal,
+            "phone.rot": rot_cal,
+            "phone.raw_inputs": raw_inputs,
+            "phone.enabled": self._enabled,
+        }
+
+    def disconnect(self) -> None:
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+
+        self._teleop = None
+        if self._teleop_thread and self._teleop_thread.is_alive():
+            self._teleop_thread.join(timeout=1.0)
+            self._teleop_thread = None
+            self._latest_pose = None
+
+
+class Phone(Teleoperator):
+    """
+    Phone-based teleoperator using ARKit (iOS via HEBI Mobile I/O App) or the teleop Python package (Android via WebXR API).
+    For HEBI Mobile I/O we also expose 8 analog (a1-a8) and 8 digital (b1-b8) inputs.
+
+    Press and hold **B1** to enable teleoperation. While enabled, the first B1 press
+    captures a reference pose and rotation, when disabled and pressed again the position is reapplied.
+    """
+
+    config_class = PhoneConfig
+    name = "phone"
+
+    def __init__(self, config: PhoneConfig):
+        super().__init__(config)
+        self.config = config
+
+        self._phone_impl: Teleoperator
+
+        if self.config.phone_os == PhoneOS.IOS:
+            self._phone_impl = IOSPhone(config)
+        elif self.config.phone_os == PhoneOS.ANDROID:
+            self._phone_impl = AndroidPhone(config)
+        else:
+            raise ValueError(f"Invalid config phone_os: {self.config.phone_os}")
+
+    @property
+    def is_connected(self) -> bool:
+        return self._phone_impl.is_connected
+
+    def connect(self) -> None:
+        return self._phone_impl.connect()
+
+    def calibrate(self) -> None:
+        return self._phone_impl.calibrate()
+
+    @property
+    def is_calibrated(self) -> bool:
+        return self._phone_impl.is_calibrated
+
+    @property
+    def action_features(self) -> dict[str, type]:
+        return self._phone_impl.action_features
+
+    @property
+    def feedback_features(self) -> dict[str, type]:
+        return self._phone_impl.feedback_features
+
+    def configure(self) -> None:
+        return self._phone_impl.configure()
+
+    def get_action(self) -> dict:
+        return self._phone_impl.get_action()
+
+    def send_feedback(self, feedback: dict[str, float]) -> None:
+        return self._phone_impl.send_feedback(feedback)
+
+    def disconnect(self) -> None:
+        return self._phone_impl.disconnect()

src/lerobot/teleoperators/utils.py

@@ -12,10 +12,22 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from enum import Enum
+
 from .config import TeleoperatorConfig
 from .teleoperator import Teleoperator
 
 
+class TeleopEvents(Enum):
+    """Shared constants for teleoperator events across teleoperators."""
+
+    SUCCESS = "success"
+    FAILURE = "failure"
+    RERECORD_EPISODE = "rerecord_episode"
+    IS_INTERVENTION = "is_intervention"
+    TERMINATE_EPISODE = "terminate_episode"
+
+
 def make_teleoperator_from_config(config: TeleoperatorConfig) -> Teleoperator:
     if config.type == "keyboard":
         from .keyboard import KeyboardTeleop

src/lerobot/templates/lerobot_modelcard_template.md

@@ -44,7 +44,7 @@ Below is the short version on how to train and run inference/eval:
 ### Train from scratch
 
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
   --dataset.repo_id=${HF_USER}/<dataset> \
   --policy.type=act \
   --output_dir=outputs/train/<desired_policy_repo_id> \
@@ -59,7 +59,7 @@ _Writes checkpoints to `outputs/train/<desired_policy_repo_id>/checkpoints/`._
 ### Evaluate the policy/run inference
 
 ```bash
-python -m lerobot.record \
+lerobot-record \
   --robot.type=so100_follower \
   --dataset.repo_id=<hf_user>/eval_<dataset> \
   --policy.path=<hf_user>/<desired_policy_repo_id> \

src/lerobot/utils/control_utils.py

@@ -31,7 +31,7 @@ from termcolor import colored
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.datasets.utils import DEFAULT_FEATURES
 from lerobot.policies.pretrained import PreTrainedPolicy
-from lerobot.processor import RobotProcessor, TransitionKey
+from lerobot.processor import PolicyProcessorPipeline, TransitionKey
 from lerobot.robots import Robot
 
 
@@ -102,8 +102,8 @@ def predict_action(
     observation: dict[str, np.ndarray],
     policy: PreTrainedPolicy,
     device: torch.device,
-    preprocessor: RobotProcessor,
-    postprocessor: RobotProcessor,
+    preprocessor: PolicyProcessorPipeline,
+    postprocessor: PolicyProcessorPipeline,
     use_amp: bool,
     task: str | None = None,
     robot_type: str | None = None,

src/lerobot/utils/robot_utils.py

@@ -17,10 +17,9 @@ import time
 
 
 def busy_wait(seconds):
-    if platform.system() == "Darwin":
-        # On Mac, `time.sleep` is not accurate and we need to use this while loop trick,
+    if platform.system() == "Darwin" or platform.system() == "Windows":
+        # On Mac and Windows, `time.sleep` is not accurate and we need to use this while loop trick,
         # but it consumes CPU cycles.
-        # TODO(rcadene): find an alternative: from python 11, time.sleep is precise
         end_time = time.perf_counter() + seconds
         while time.perf_counter() < end_time:
             pass

src/lerobot/utils/rotation.py

@@ -0,0 +1,174 @@
+#!/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.
+
+"""Custom rotation utilities to replace scipy.spatial.transform.Rotation."""
+
+import numpy as np
+
+
+class Rotation:
+    """
+    Custom rotation class that provides a subset of scipy.spatial.transform.Rotation functionality.
+
+    Supports conversions between rotation vectors, rotation matrices, and quaternions.
+    """
+
+    def __init__(self, quat: np.ndarray) -> None:
+        """Initialize rotation from quaternion [x, y, z, w]."""
+        self._quat = np.asarray(quat, dtype=float)
+        # Normalize quaternion
+        norm = np.linalg.norm(self._quat)
+        if norm > 0:
+            self._quat = self._quat / norm
+
+    @classmethod
+    def from_rotvec(cls, rotvec: np.ndarray) -> "Rotation":
+        """
+        Create rotation from rotation vector using Rodrigues' formula.
+
+        Args:
+            rotvec: Rotation vector [x, y, z] where magnitude is angle in radians
+
+        Returns:
+            Rotation instance
+        """
+        rotvec = np.asarray(rotvec, dtype=float)
+        angle = np.linalg.norm(rotvec)
+
+        if angle < 1e-8:
+            # For very small angles, use identity quaternion
+            quat = np.array([0.0, 0.0, 0.0, 1.0])
+        else:
+            axis = rotvec / angle
+            half_angle = angle / 2.0
+            sin_half = np.sin(half_angle)
+            cos_half = np.cos(half_angle)
+
+            # Quaternion [x, y, z, w]
+            quat = np.array([axis[0] * sin_half, axis[1] * sin_half, axis[2] * sin_half, cos_half])
+
+        return cls(quat)
+
+    @classmethod
+    def from_matrix(cls, matrix: np.ndarray) -> "Rotation":
+        """
+        Create rotation from 3x3 rotation matrix.
+
+        Args:
+            matrix: 3x3 rotation matrix
+
+        Returns:
+            Rotation instance
+        """
+        matrix = np.asarray(matrix, dtype=float)
+
+        # Shepherd's method for converting rotation matrix to quaternion
+        trace = np.trace(matrix)
+
+        if trace > 0:
+            s = np.sqrt(trace + 1.0) * 2  # s = 4 * qw
+            qw = 0.25 * s
+            qx = (matrix[2, 1] - matrix[1, 2]) / s
+            qy = (matrix[0, 2] - matrix[2, 0]) / s
+            qz = (matrix[1, 0] - matrix[0, 1]) / s
+        elif matrix[0, 0] > matrix[1, 1] and matrix[0, 0] > matrix[2, 2]:
+            s = np.sqrt(1.0 + matrix[0, 0] - matrix[1, 1] - matrix[2, 2]) * 2  # s = 4 * qx
+            qw = (matrix[2, 1] - matrix[1, 2]) / s
+            qx = 0.25 * s
+            qy = (matrix[0, 1] + matrix[1, 0]) / s
+            qz = (matrix[0, 2] + matrix[2, 0]) / s
+        elif matrix[1, 1] > matrix[2, 2]:
+            s = np.sqrt(1.0 + matrix[1, 1] - matrix[0, 0] - matrix[2, 2]) * 2  # s = 4 * qy
+            qw = (matrix[0, 2] - matrix[2, 0]) / s
+            qx = (matrix[0, 1] + matrix[1, 0]) / s
+            qy = 0.25 * s
+            qz = (matrix[1, 2] + matrix[2, 1]) / s
+        else:
+            s = np.sqrt(1.0 + matrix[2, 2] - matrix[0, 0] - matrix[1, 1]) * 2  # s = 4 * qz
+            qw = (matrix[1, 0] - matrix[0, 1]) / s
+            qx = (matrix[0, 2] + matrix[2, 0]) / s
+            qy = (matrix[1, 2] + matrix[2, 1]) / s
+            qz = 0.25 * s
+
+        quat = np.array([qx, qy, qz, qw])
+        return cls(quat)
+
+    @classmethod
+    def from_quat(cls, quat: np.ndarray) -> "Rotation":
+        """
+        Create rotation from quaternion.
+
+        Args:
+            quat: Quaternion [x, y, z, w] or [w, x, y, z] (specify convention in docstring)
+                  This implementation expects [x, y, z, w] format
+
+        Returns:
+            Rotation instance
+        """
+        return cls(quat)
+
+    def as_matrix(self) -> np.ndarray:
+        """
+        Convert rotation to 3x3 rotation matrix.
+
+        Returns:
+            3x3 rotation matrix
+        """
+        qx, qy, qz, qw = self._quat
+
+        # Compute rotation matrix from quaternion
+        return np.array(
+            [
+                [1 - 2 * (qy * qy + qz * qz), 2 * (qx * qy - qz * qw), 2 * (qx * qz + qy * qw)],
+                [2 * (qx * qy + qz * qw), 1 - 2 * (qx * qx + qz * qz), 2 * (qy * qz - qx * qw)],
+                [2 * (qx * qz - qy * qw), 2 * (qy * qz + qx * qw), 1 - 2 * (qx * qx + qy * qy)],
+            ],
+            dtype=float,
+        )
+
+    def as_rotvec(self) -> np.ndarray:
+        """
+        Convert rotation to rotation vector.
+
+        Returns:
+            Rotation vector [x, y, z] where magnitude is angle in radians
+        """
+        qx, qy, qz, qw = self._quat
+
+        # Ensure qw is positive for unique representation
+        if qw < 0:
+            qx, qy, qz, qw = -qx, -qy, -qz, -qw
+
+        # Compute angle and axis
+        angle = 2.0 * np.arccos(np.clip(abs(qw), 0.0, 1.0))
+        sin_half_angle = np.sqrt(1.0 - qw * qw)
+
+        if sin_half_angle < 1e-8:
+            # For very small angles, use linearization: rotvec ≈ 2 * [qx, qy, qz]
+            return 2.0 * np.array([qx, qy, qz])
+
+        # Extract axis and scale by angle
+        axis = np.array([qx, qy, qz]) / sin_half_angle
+        return angle * axis
+
+    def as_quat(self) -> np.ndarray:
+        """
+        Get quaternion representation.
+
+        Returns:
+            Quaternion [x, y, z, w]
+        """
+        return self._quat.copy()

src/lerobot/utils/train_utils.py

@@ -32,6 +32,7 @@ from lerobot.datasets.utils import load_json, write_json
 from lerobot.optim.optimizers import load_optimizer_state, save_optimizer_state
 from lerobot.optim.schedulers import load_scheduler_state, save_scheduler_state
 from lerobot.policies.pretrained import PreTrainedPolicy
+from lerobot.processor import PolicyProcessorPipeline
 from lerobot.utils.random_utils import load_rng_state, save_rng_state
 
 
@@ -74,7 +75,8 @@ def save_checkpoint(
     policy: PreTrainedPolicy,
     optimizer: Optimizer,
     scheduler: LRScheduler | None = None,
-    preprocessor=None,
+    preprocessor: PolicyProcessorPipeline | None = None,
+    postprocessor: PolicyProcessorPipeline | None = None,
 ) -> None:
     """This function creates the following directory structure:
 
@@ -105,6 +107,8 @@ def save_checkpoint(
     cfg.save_pretrained(pretrained_dir)
     if preprocessor is not None:
         preprocessor.save_pretrained(pretrained_dir)
+    if postprocessor is not None:
+        postprocessor.save_pretrained(pretrained_dir)
     save_training_state(checkpoint_dir, step, optimizer, scheduler)
 
 

src/lerobot/utils/visualization_utils.py

@@ -19,7 +19,7 @@ from typing import Any
 import numpy as np
 import rerun as rr
 
-from lerobot.processor.pipeline import EnvTransition, TransitionKey
+from lerobot.processor import EnvTransition, TransitionKey
 
 
 def _init_rerun(session_name: str = "lerobot_control_loop") -> None:

tests/artifacts/policies/aloha_sim_insertion_human_act_/actions.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f3e4c8e85e146b043fd4e4984947c2a6f01627f174a19f18b5914cf690579d77
+oid sha256:ee0c29d3782aa1cadcf4dc6ed767d9460ff00fff9fc70b460502340b832eefcc
 size 5104

tests/artifacts/policies/aloha_sim_insertion_human_act_/param_stats.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9b5f557e30aead3731c38cbd85af8c706395d8689a918ad88805b5a886245603
-size 33400
+oid sha256:ea76e6711959fd3f905ec2bdc306f488920f00ec99421e4870d05f6205eb323e
+size 31672