fix

- Use normalization processor as default example
- Add section on transform features
- Add section on overrides.

.github/PULL_REQUEST_TEMPLATE.md

@@ -30,7 +30,7 @@ pytest -sx tests/test_stuff.py::test_something
 ```
 
 ```bash
-lerobot-train --some.option=true
+python -m lerobot.scripts.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-cpu:latest
-  DOCKER_IMAGE_NAME_GPU: huggingface/lerobot-gpu:latest
+  DOCKER_IMAGE_NAME_CPU: huggingface/lerobot-gpu:latest
+  DOCKER_IMAGE_NAME_GPU: huggingface/lerobot-cpu: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:
-	lerobot-train \
+	python -m lerobot.scripts.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:
-	lerobot-train \
+	python -m lerobot.scripts.train \
 		--config_path=tests/outputs/act/checkpoints/000002/pretrained_model/train_config.json \
 		--resume=true
 
 test-act-ete-eval:
-	lerobot-eval \
+	python -m lerobot.scripts.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:
-	lerobot-train \
+	python -m lerobot.scripts.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:
-	lerobot-eval \
+	python -m lerobot.scripts.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:
-	lerobot-train \
+	python -m lerobot.scripts.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:
-	lerobot-eval \
+	python -m lerobot.scripts.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:
-	lerobot-train \
+	python -m lerobot.scripts.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:
-	lerobot-eval \
+	python -m lerobot.scripts.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/nightly.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/nighty.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/)
@@ -233,7 +233,7 @@ Under the hood, the `LeRobotDataset` format makes use of several ways to seriali
 
 Here are the important details and internal structure organization of a typical `LeRobotDataset` instantiated with `dataset = LeRobotDataset("lerobot/aloha_static_coffee")`. The exact features will change from dataset to dataset but not the main aspects:
 
-````
+```
 dataset attributes:
   ├ hf_dataset: a Hugging Face dataset (backed by Arrow/parquet). Typical features example:
   │  ├ observation.images.cam_high (VideoFrame):
@@ -246,30 +246,20 @@ dataset attributes:
   │  ├ timestamp (float32): timestamp in the episode
   │  ├ next.done (bool): indicates the end of an episode ; True for the last frame in each episode
   │  └ index (int64): general index in the whole dataset
-  ├ meta: a LeRobotDatasetMetadata object containing:
-  │  ├ info: a dictionary of metadata on the dataset
-  │  │  ├ codebase_version (str): this is to keep track of the codebase version the dataset was created with
-  │  │  ├ fps (int): frame per second the dataset is recorded/synchronized to
-  │  │  ├ features (dict): all features contained in the dataset with their shapes and types
-  │  │  ├ total_episodes (int): total number of episodes in the dataset
-  │  │  ├ total_frames (int): total number of frames in the dataset
-  │  │  ├ robot_type (str): robot type used for recording
-  │  │  ├ data_path (str): formattable string for the parquet files
-  │  │  └ video_path (str): formattable string for the video files (if using videos)
-  │  ├ episodes: a DataFrame containing episode metadata with columns:
-  │  │  ├ episode_index (int): index of the episode
-  │  │  ├ tasks (list): list of tasks for this episode
-  │  │  ├ length (int): number of frames in this episode
-  │  │  ├ dataset_from_index (int): start index of this episode in the dataset
-  │  │  └ dataset_to_index (int): end index of this episode in the dataset
-  │  ├ stats: a dictionary of statistics (max, mean, min, std) for each feature in the dataset, for instance
-  │  │  ├ observation.images.front_cam: {'max': tensor with same number of dimensions (e.g. `(c, 1, 1)` for images, `(c,)` for states), etc.}
-  │  │  └ ...
-  │  └ tasks: a DataFrame containing task information with task names as index and task_index as values
-  ├ root (Path): local directory where the dataset is stored
-  ├ image_transforms (Callable): optional image transformations to apply to visual modalities
-  └ delta_timestamps (dict): optional delta timestamps for temporal queries
-decoding videos (e.g., 'pyav', 'torchcodec')
+  ├ episode_data_index: contains 2 tensors with the start and end indices of each episode
+  │  ├ from (1D int64 tensor): first frame index for each episode — shape (num episodes,) starts with 0
+  │  └ to: (1D int64 tensor): last frame index for each episode — shape (num episodes,)
+  ├ stats: a dictionary of statistics (max, mean, min, std) for each feature in the dataset, for instance
+  │  ├ observation.images.cam_high: {'max': tensor with same number of dimensions (e.g. `(c, 1, 1)` for images, `(c,)` for states), etc.}
+  │  ...
+  ├ info: a dictionary of metadata on the dataset
+  │  ├ codebase_version (str): this is to keep track of the codebase version the dataset was created with
+  │  ├ fps (float): frame per second the dataset is recorded/synchronized to
+  │  ├ video (bool): indicates if frames are encoded in mp4 video files to save space or stored as png files
+  │  └ encoding (dict): if video, this documents the main options that were used with ffmpeg to encode the videos
+  ├ videos_dir (Path): where the mp4 videos or png images are stored/accessed
+  └ camera_keys (list of string): the keys to access camera features in the item returned by the dataset (e.g. `["observation.images.cam_high", ...]`)
+```
 
 A `LeRobotDataset` is serialised using several widespread file formats for each of its parts, namely:
 
@@ -286,22 +276,22 @@ 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
-lerobot-eval \
+python -m lerobot.scripts.eval \
     --policy.path=lerobot/diffusion_pusht \
     --env.type=pusht \
     --eval.batch_size=10 \
     --eval.n_episodes=10 \
     --policy.use_amp=false \
     --policy.device=cuda
-````
+```
 
 Note: After training your own policy, you can re-evaluate the checkpoints with:
 
 ```bash
-lerobot-eval --policy.path={OUTPUT_DIR}/checkpoints/last/pretrained_model
+python -m lerobot.scripts.eval --policy.path={OUTPUT_DIR}/checkpoints/last/pretrained_model
 ```
 
-See `lerobot-eval --help` for more instructions.
+See `python -m lerobot.scripts.eval --help` for more instructions.
 
 ### Train your own policy
 
@@ -313,7 +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 `lerobot-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 `python -m lerobot.scripts.eval --help` for more instructions.
 
 #### Reproduce state-of-the-art (SOTA)
 
@@ -321,7 +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
-lerobot-train --config_path=lerobot/diffusion_pusht
+python -m lerobot.scripts.train --config_path=lerobot/diffusion_pusht
 ```
 
 reproduces SOTA results for Diffusion Policy on the PushT task.

benchmarks/video/run_video_benchmark.py

@@ -108,8 +108,7 @@ def save_decoded_frames(
 
 
 def save_first_episode(imgs_dir: Path, dataset: LeRobotDataset) -> None:
-    episode_index = 0
-    ep_num_images = dataset.meta.episodes["length"][episode_index]
+    ep_num_images = dataset.episode_data_index["to"][0].item()
     if imgs_dir.exists() and len(list(imgs_dir.glob("frame_*.png"))) == ep_num_images:
         return
 
@@ -266,8 +265,7 @@ def benchmark_encoding_decoding(
             overwrite=True,
         )
 
-    episode_index = 0
-    ep_num_images = dataset.meta.episodes["length"][episode_index]
+    ep_num_images = dataset.episode_data_index["to"][0].item()
     width, height = tuple(dataset[0][dataset.meta.camera_keys[0]].shape[-2:])
     num_pixels = width * height
     video_size_bytes = video_path.stat().st_size

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-dev ffmpeg \
+    build-essential git curl libglib2.0-0 libegl1-mesa 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

@@ -20,19 +20,20 @@
   - local: async
     title: Use Async Inference
   title: "Tutorials"
-- sections:
-  - local: lerobot-dataset-v3
-    title: Using LeRobotDataset
-  - local: porting_datasets_v3
-    title: Porting Large Datasets
-  title: "Datasets"
 - sections:
   - local: smolvla
     title: Finetune SmolVLA
   title: "Policies"
+
+- sections:
+  - local: introduction_processors
+    title: Introduction to Robot Processors
+  - local: implement_your_own_processor
+    title: Implement your own processor
+  - local: processors_robots_teleop
+    title: Processors for Robots and Teleoperators
+  title: "Robot Processors"
 - sections:
-  - local: hope_jr
-    title: Hope Jr
   - local: so101
     title: SO-101
   - local: so100
@@ -41,14 +42,16 @@
     title: Koch v1.1
   - local: lekiwi
     title: LeKiwi
-  - local: reachy2
-    title: Reachy 2
+  - local: hope_jr
+    title: Hope Jr
   title: "Robots"
+- sections:
+  - local: phone_teleop
+    title: Phone
+  title: "Teleoperators"
 - 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
-lerobot-find-cameras opencv # or realsense for Intel Realsense cameras
+python -m 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

@@ -1,71 +0,0 @@
-# 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

@@ -412,7 +412,7 @@ Example configuration for training the [reward classifier](https://huggingface.c
 To train the classifier, use the `train.py` script with your configuration:
 
 ```bash
-lerobot-train --config_path path/to/reward_classifier_train_config.json
+python -m lerobot.scripts.train --config_path path/to/reward_classifier_train_config.json
 ```
 
 **Deploying and Testing the Model**
@@ -458,7 +458,7 @@ The reward classifier will automatically provide rewards based on the visual inp
 3. **Train the classifier**:
 
    ```bash
-   lerobot-train --config_path src/lerobot/configs/reward_classifier_train_config.json
+   python -m lerobot.scripts.train --config_path src/lerobot/configs/reward_classifier_train_config.json
    ```
 
 4. **Test the classifier**:

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
-lerobot-find-port
+python -m 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
-lerobot-calibrate \
+python -m 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
-lerobot-calibrate \
+python -m 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
-lerobot-calibrate \
+python -m 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
-lerobot-calibrate \
+python -m 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
-lerobot-teleoperate \
+python -m lerobot.teleoperate \
     --robot.type=hope_jr_hand \
     --robot.port=/dev/tty.usbmodem58760432281 \
     --robot.id=blue \
@@ -194,7 +194,7 @@ lerobot-teleoperate \
 ### Arm
 
 ```bash
-lerobot-teleoperate \
+python -m 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
-lerobot-record \
+python -m lerobot.record \
     --robot.type=hope_jr_hand \
     --robot.port=/dev/tty.usbmodem58760432281 \
     --robot.id=right \
@@ -236,7 +236,7 @@ lerobot-record \
 ### Replay
 
 ```bash
-lerobot-replay \
+python -m lerobot.replay \
     --robot.type=hope_jr_hand \
     --robot.port=/dev/tty.usbmodem58760432281 \
     --robot.id=right \
@@ -248,7 +248,7 @@ lerobot-replay \
 ### Train
 
 ```bash
-lerobot-train \
+python -m lerobot.scripts.train \
   --dataset.repo_id=nepyope/hand_record_test_with_video_data \
   --policy.type=act \
   --output_dir=outputs/train/hopejr_hand \
@@ -263,7 +263,7 @@ lerobot-train \
 This training run can be viewed as an example [here](https://wandb.ai/tino/lerobot/runs/rp0k8zvw?nw=nwusertino).
 
 ```bash
-lerobot-record \
+python -m 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
-lerobot-teleoperate \
+python -m 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
-lerobot-teleoperate \
+python -m 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
-lerobot-record \
+python -m 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
-lerobot-replay \
+python -m 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 [`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:
+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:
 
 ```bash
-lerobot-train \
+python -m lerobot.scripts.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
-lerobot-train \
+python -m lerobot.scripts.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
-lerobot-record  \
+python -m lerobot.record  \
   --robot.type=so100_follower \
   --robot.port=/dev/ttyACM1 \
   --robot.cameras="{ up: {type: opencv, index_or_path: /dev/video10, width: 640, height: 480, fps: 30}, side: {type: intelrealsense, serial_number_or_name: 233522074606, width: 640, height: 480, fps: 30}}" \
@@ -519,11 +519,14 @@ from lerobot.utils.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import _init_rerun
 from lerobot.record import record_loop
+from lerobot.policies.factory import make_processor
 
 NUM_EPISODES = 5
 FPS = 30
 EPISODE_TIME_SEC = 60
 TASK_DESCRIPTION = "My task description"
+HF_MODEL_ID = "<hf_username>/<model_repo_id>"
+HF_DATASET_ID = "<hf_username>/<eval_dataset_repo_id>"
 
 # Create the robot configuration
 camera_config = {"front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS)}
@@ -535,7 +538,7 @@ robot_config = SO100FollowerConfig(
 robot = SO100Follower(robot_config)
 
 # Initialize the policy
-policy = ACTPolicy.from_pretrained("<hf_username>/<my_policy_repo_id>")
+policy = ACTPolicy.from_pretrained(HF_MODEL_ID)
 
 # Configure the dataset features
 action_features = hw_to_dataset_features(robot.action_features, "action")
@@ -544,7 +547,7 @@ dataset_features = {**action_features, **obs_features}
 
 # Create the dataset
 dataset = LeRobotDataset.create(
-    repo_id="<hf_username>/eval_<dataset_repo_id>",
+    repo_id=HF_DATASET_ID,
     fps=FPS,
     features=dataset_features,
     robot_type=robot.name,
@@ -559,6 +562,12 @@ _init_rerun(session_name="recording")
 # Connect the robot
 robot.connect()
 
+preprocessor, postprocessor = make_processor(
+    policy_cfg=policy,
+    pretrained_path=HF_MODEL_ID,
+    dataset_stats=dataset.meta.stats,
+)
+
 for episode_idx in range(NUM_EPISODES):
     log_say(f"Running inference, recording eval episode {episode_idx + 1} of {NUM_EPISODES}")
 
@@ -568,6 +577,8 @@ for episode_idx in range(NUM_EPISODES):
         events=events,
         fps=FPS,
         policy=policy,
+        preprocessor=preprocessor,
+        postprocessor=postprocessor,
         dataset=dataset,
         control_time_s=EPISODE_TIME_SEC,
         single_task=TASK_DESCRIPTION,

docs/source/il_sim.mdx

@@ -96,10 +96,10 @@ 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 [`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:
+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:
 
 ```bash
-lerobot-train \
+python -m lerobot.scripts.train \
   --dataset.repo_id=${HF_USER}/il_gym \
   --policy.type=act \
   --output_dir=outputs/train/il_sim_test \

docs/source/implement_your_own_processor.mdx

@@ -0,0 +1,323 @@
+# Implement your own Robot Processor
+
+In this tutorial, you'll learn how to implement your own Robot Processor.
+It begins by exploring the need for a custom processor, then uses the Normalization processors as the running example to explain how to implement, configure, and serialize a processor. Finally, it lists all helper processors that ship with LeRobot.
+
+## Why would you need a custom processor?
+
+In most cases, when reading raw data from a sensor like the camera and robot motor encoders,
+you will need to process this data to transform it into a format that is compatible to use with the policies in LeRobot.
+For example, raw images are encoded with `uint8` and the values are in the range `[0, 255]`.
+To use these images with the policies, you will need to cast them to `float32` and normalize them to the range `[0, 1]`.
+
+For example, in LeRobot's `VanillaObservationProcessor`, raw images come from the environment as numpy arrays with `uint8` values in range `[0, 255]` and in channel-last format `(H, W, C)`. The processor transforms them into PyTorch tensors with `float32` values in range `[0, 1]` and channel-first format `(C, H, W)`:
+
+```python
+# Input: numpy array with shape (480, 640, 3) and dtype uint8
+raw_image = env_observation["pixels"]  # Values in [0, 255]
+
+# After processing: torch tensor with shape (1, 3, 480, 640) and dtype float32
+processed_image = processor(transition)["observation"]["observation.image"]  # Values in [0, 1]
+```
+
+On the other hand, when a model returns a certain action to be executed on the robot, it is often that one has to post-process this action to make it compatible to run on the robot.
+For example, the model might return joint positions values that range from `[-1, 1]` and one would need to scale them to the ranges of the minimum and maximum joint angle positions of the robot.
+
+In LeRobot, this normalization workflow is handled by the `NormalizerProcessor` (for inputs) and the `UnnormalizerProcessor` (for outputs). These processors are heavily used by policies (e.g., Pi0, SmolVLA) and integrate tightly with the `RobotProcessor`'s `get_config`, `state_dict`, and `load_state_dict` APIs.
+
+For instance, `UnnormalizerProcessor` converts model outputs in `[-1, 1]` back to actual robot joint ranges:
+
+```python
+# Input: model action with normalized values in [-1, 1]
+normalized_action = torch.tensor([-0.5, 0.8, -1.0, 0.2])  # Model output
+
+# After post-processing: real joint positions in robot's native ranges
+# Example: joints range from [-180.0, 180.0]
+real_action = unnormalizer(transition)["action"]
+# real action after post-processing: [ -90.,  144., -180.,   36.]
+```
+
+The unnormalizer uses the dataset statistics to convert back:
+
+```python
+# For MIN_MAX normalization: action = (normalized + 1) * (max - min) / 2 + min
+real_action = (normalized_action + 1) * (max_val - min_val) / 2 + min_val
+```
+
+All these situations point us towards the need for a mechanism to preprocess the data before being passed to the policies and then post-process the action that are returned to be executed on the robot.
+
+To that end, LeRobot provides a pipeline mechanism to implement a sequence of processing steps for the input data and the output action.
+
+## How to implement your own processor?
+
+We'll use the `NormalizerProcessor` as a concrete running example because it is central to most policies and demonstrates configuration and state serialization cleanly.
+
+Prepare the sequence of processing steps necessary for your problem. A processor step is a class that implements the following methods:
+
+- `__call__`: implements the processing step for the input transition.
+- `get_config`: gets the configuration of the processor step.
+- `state_dict`: gets the state of the processor step.
+- `load_state_dict`: loads the state of the processor step.
+- `reset`: resets the state of the processor step.
+- `feature_contract`: displays the modification to the feature space during the processor step.
+
+### Implement the `__call__` method
+
+The `__call__` method is the core of your processor step. It takes an `EnvTransition` and returns a modified `EnvTransition`. Here's how the `NormalizerProcessor` conceptually works (simplified):
+
+```python
+from dataclasses import dataclass
+import torch
+from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
+from lerobot.processor.pipeline import EnvTransition, TransitionKey
+
+@dataclass
+class NormalizerProcessor:
+    features: dict[str, PolicyFeature]
+    norm_map: dict[FeatureType, NormalizationMode]
+    stats: dict[str, dict[str, torch.Tensor]]
+    eps: float = 1e-8
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        normalized_info = {}
+
+        obs = transition.get(TransitionKey.OBSERVATION)
+        act = transition.get(TransitionKey.ACTION)
+
+        new_obs = self._normalize_observation(obs, normalized_info)
+        new_act = self._normalize_action(act, normalized_info)
+
+        new_transition = transition.copy()
+        new_transition[TransitionKey.OBSERVATION] = new_obs
+        new_transition[TransitionKey.ACTION] = new_act
+
+        # Record what was normalized into complementary_data
+        if normalized_info:
+            comp = new_transition.get(TransitionKey.COMPLEMENTARY_DATA) or {}
+            comp = dict(comp)
+            comp["normalized_keys"] = normalized_info
+            new_transition[TransitionKey.COMPLEMENTARY_DATA] = comp
+
+        return new_transition
+```
+
+See the full implementation in `src/lerobot/processor/normalize_processor.py` for details on mean/std and min/max modes and key selection.
+
+**Key principles:**
+
+- Always check if required data exists before processing
+- Return unchanged transition if no processing is needed
+- Use `transition.copy()` to avoid side effects
+- Only modify the specific keys your processor handles
+
+**Tip**: For observation-only processors, you can inherit from `ObservationProcessor` to avoid writing `__call__` boilerplate. The normalizer is mixed (observations and actions), so it implements `__call__` directly.
+
+### Configuration and State Management
+
+Processors support serialization through three methods that separate configuration from tensor state. This is especially important for normalization processors, which carry dataset statistics (tensors) in their state, and hyperparameters in their config:
+
+```python
+from dataclasses import dataclass, field
+from typing import Any
+import torch
+from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
+
+@dataclass
+class NormalizerProcessor:
+    features: dict[str, PolicyFeature]
+    norm_map: dict[FeatureType, NormalizationMode]
+    eps: float = 1e-8
+    _tensor_stats: dict[str, dict[str, torch.Tensor]] = field(default_factory=dict, init=False, repr=False)
+
+    def get_config(self) -> dict[str, Any]:
+        """JSON-serializable configuration (no tensors)."""
+        return {
+            "eps": self.eps,
+            "features": {k: {"type": v.type.value, "shape": v.shape} for k, v in self.features.items()},
+            "norm_map": {ft.value: nm.value for ft, nm in self.norm_map.items()},
+        }
+
+    def state_dict(self) -> dict[str, torch.Tensor]:
+        """Tensor state only (e.g., dataset statistics)."""
+        flat: dict[str, torch.Tensor] = {}
+        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: dict[str, torch.Tensor]) -> None:
+        """Restore tensor state at runtime."""
+        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
+```
+
+**Usage:**
+
+```python
+# Save (e.g., inside a policy)
+config = processor.get_config()
+tensors = processor.state_dict()
+
+# Restore (e.g., loading a pretrained policy)
+new_processor = NormalizerProcessor(**config)
+new_processor.load_state_dict(tensors)
+```
+
+### Transform features
+
+The `transform_features` method defines how your processor transforms feature names and shapes. This is crucial for policy configuration and debugging.
+
+Normalization typically preserves the feature keys and shapes, so `NormalizerProcessor.transform_features` returns the input features unchanged. When your processor renames or reshapes, implement this method to reflect the mapping for downstream components. For example, a simple rename processor:
+
+```python
+def transform_features(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
+    # Simple renaming
+    if "pixels" in features:
+        features["observation.image"] = features.pop("pixels")
+
+    # Pattern-based renaming
+    for key in list(features.keys()):
+        if key.startswith("env_state."):
+            suffix = key[len("env_state."):]
+            features[f"observation.{suffix}"] = features.pop(key)
+
+    return features
+```
+
+**Key principles:**
+
+- Use `features.pop(old_key)` to remove and get the old feature
+- Use `features[new_key] = old_feature` to add the renamed feature
+- Always return the modified features dictionary
+- Document transformations clearly in the docstring
+
+### Example of usage from the codebase
+
+`transform_features` is used by `RobotProcessor` to derive the dataset/policy feature contract from an initial feature set by applying each step's transformation. You can see concrete examples in the codebase:
+
+- Phone teleoperation record pipeline (`examples/phone_so100_record.py`): processors like `ForwardKinematicsJointsToEE`, `GripperVelocityToJoint`, and `EEBoundsAndSafety` implement `transform_features` to declare which action/observation keys should be materialized in the dataset.
+- SO100 follower kinematics (`src/lerobot/robots/so100_follower/robot_kinematic_processor.py`): each processor's `transform_features` method adds or refines feature keys such as `observation.state.ee.{x,y,z,wx,wy,wz}` or `action.gripper.pos`.
+- Rename and tokenizer processors (`src/lerobot/processor/rename_processor.py`, `src/lerobot/processor/tokenizer_processor.py`): demonstrate key renaming and adding language token features to the contract.
+
+In practice, you will often aggregate features by running `RobotProcessor.transform_features(...)` with your initial features to compute the final contract before recording or training.
+
+## Helper Classes
+
+LeRobot provides pre-built processor classes for common transformations. Below is a comprehensive list of registered processors in the codebase.
+
+### Core processors (observations, actions, normalization)
+
+- **`VanillaObservationProcessor`** (`observation_processor`): Images and state processing to LeRobot format.
+- **`NormalizerProcessor`** (`normalizer_processor`): Normalize observations/actions (mean/std or min/max to [-1, 1]).
+- **`UnnormalizerProcessor`** (`unnormalizer_processor`): Inverse of the normalizer for model outputs.
+- **`DeviceProcessor`** (`device_processor`): Move tensors to a specific device (CPU/GPU) and optional float dtype.
+- **`ToBatchProcessor`** (`to_batch_processor`): Add batch dimension to observations/actions when missing.
+- **`RenameProcessor`** (`rename_processor`): Rename observation keys using a mapping dictionary.
+- **`TokenizerProcessor`** (`tokenizer_processor`): Tokenize language tasks into `observation.language.*` tensors.
+
+### Teleoperation mapping processors
+
+- **`MapDeltaActionToRobotAction`** (`map_delta_action_to_robot_action`): Map teleop deltas (e.g., gamepad) to `action.target_*` fields.
+- **`MapPhoneActionToRobotAction`** (`map_phone_action_to_robot_action`): Map calibrated phone pose/buttons to `action.target_*` and gripper.
+
+### Robot kinematics processors (SO100 follower example)
+
+- **`EEReferenceAndDelta`** (`ee_reference_and_delta`): Compute desired EE pose from target deltas and current pose.
+- **`EEBoundsAndSafety`** (`ee_bounds_and_safety`): Clip EE pose to bounds and check for jumps.
+- **`InverseKinematicsEEToJoints`** (`inverse_kinematics_ee_to_joints`): Convert EE pose to joint targets via IK.
+- **`GripperVelocityToJoint`** (`gripper_velocity_to_joint`): Convert gripper velocity input to joint position command.
+- **`ForwardKinematicsJointsToEE`** (`forward_kinematics_joints_to_ee`): Compute EE pose features from joint positions via FK.
+- **`AddRobotObservationAsComplimentaryData`** (`add_robot_observation`): Read robot observation and insert `raw_joint_positions` into complementary data.
+
+### Policy-specific utility processors
+
+- **`Pi0NewLineProcessor`** (`pi0_new_line_processor`): Ensure text tasks end with a newline (Pi0 tokenizer compatibility).
+- **`SmolVLANewLineProcessor`** (`smolvla_new_line_processor`): Ensure text tasks end with a newline (SmolVLA tokenizer compatibility).
+
+### Usage Example
+
+```python
+from lerobot.processor import NormalizerProcessor, DeviceProcessor, RobotProcessor, ToBatchProcessor
+
+# Create a processing pipeline (typical policy preprocessor)
+steps = [
+    NormalizerProcessor(features=features, norm_map=norm_map, stats=stats),
+    ToBatchProcessor(),
+    DeviceProcessor(device="cuda"),
+]
+
+# Use in RobotProcessor
+processor = RobotProcessor(steps=steps)
+processed_transition = processor(raw_transition)
+```
+
+### Using overrides
+
+You can override step parameters at load-time using `overrides`. This is handy for non-serializable objects or site-specific settings. It works both in policy factories and with `RobotProcessor.from_pretrained(...)`.
+
+Example: during policy evaluation on the robot, override the device and rename map.
+Use this to run a policy trained on CUDA on a CPU-only robot, or to remap camera keys when the robot uses different names than the dataset.
+
+```437:445:src/lerobot/record.py
+preprocessor, postprocessor = make_processor(
+    policy_cfg=cfg.policy,
+    pretrained_path=cfg.policy.pretrained_path,
+    dataset_stats=rename_stats(dataset.meta.stats, cfg.dataset.rename_map),
+    preprocessor_overrides={
+        "device_processor": {"device": cfg.policy.device},
+        "rename_processor": {"rename_map": cfg.dataset.rename_map},
+    },
+)
+```
+
+Direct usage with `from_pretrained`:
+
+```python
+from lerobot.processor import RobotProcessor
+
+processor = RobotProcessor.from_pretrained(
+    "username/my-processor",
+    overrides={
+        "device_processor": {"device": "cuda:0"},  # registry name for registered steps
+        "CustomStep": {"param": 42},               # class name for non-registered steps
+    },
+)
+```
+
+## Best Practices
+
+- **Keep processors atomic** - One transformation per processor for reusability and debugging
+- **Use dataclasses** - Clean initialization with `@dataclass`
+- **Always register processors** - Use `@ProcessorStepRegistry.register("name")` for discoverability
+- **Check for None** - Always validate required data exists before processing
+- **Use copy() for safety** - Avoid side effects with `transition.copy()`
+- **Separate config and state** - JSON-serializable config vs tensor state_dict
+- **Use base classes** - Inherit from `ObservationProcessor` for observation-only processing
+
+```python
+@ProcessorStepRegistry.register("my_processor")
+@dataclass
+class MyProcessor(ObservationProcessor):
+    threshold: float = 0.5
+
+    def observation(self, observation):
+        if observation is None:
+            return observation
+        # Your processing logic here
+        return processed_observation
+```
+
+## Conclusion
+
+You now have all the tools to implement custom processors in LeRobot! The key steps are:
+
+1. **Define your processor** as a dataclass with the required methods (`__call__`, `get_config`, `state_dict`, `load_state_dict`, `reset`, `feature_contract`)
+2. **Register it** using `@ProcessorStepRegistry.register("name")` for discoverability
+3. **Integrate it** into a `RobotProcessor` pipeline with other processing steps
+4. **Use base classes** like `ObservationProcessor` when possible to reduce boilerplate
+
+The processor system is designed to be modular and composable, allowing you to build complex data processing pipelines from simple, focused components. Whether you're preprocessing sensor data for training or post-processing model outputs for robot execution, custom processors give you the flexibility to handle any data transformation your robotics application requires. Policies like Pi0 and SmolVLA use the same normalization processors described above, so your understanding here will transfer directly when wiring policy preprocessors and postprocessors.
+
+Start simple, test thoroughly, and leverage the existing helper classes to build robust data processing pipelines for your robot learning workflows.

docs/source/introduction_processors.mdx

@@ -0,0 +1,991 @@
+# Introduction to Processors
+
+In robotics, there's a fundamental mismatch between the data that robots and humans produce and what machine learning models expect. This creates several translation challenges:
+
+**Raw Robot Data → Model Input:**
+
+- Robots output raw sensor data (camera images, joint positions, force readings) that need normalization, batching, and device placement before models can process them
+- Language instructions from humans ("pick up the red cube") must be tokenized into numerical representations
+- Different robots use different coordinate systems and units that need standardization
+
+**Model Output → Robot Commands:**
+
+- Models might output end-effector positions, but robots need joint-space commands
+- Teleoperators (like gamepads) produce relative movements (delta positions), but robots expect absolute commands
+- Model predictions are often normalized and need to be converted back to real-world scales
+
+**Cross-Domain Translation:**
+
+- Training data from one robot setup needs adaptation for deployment on different hardware
+- Models trained with specific camera configurations must work with new camera arrangements
+- Datasets with different naming conventions need harmonization
+
+**That's where processors come in.** They serve as the universal translators that bridge these gaps, ensuring seamless data flow from sensors to models to actuators.
+
+Processors are the data transformation backbone of LeRobot. They handle all the preprocessing and postprocessing steps needed to convert raw environment data into model-ready inputs and vice versa. This guide will walk you through everything you need to know about processors - from basic concepts to advanced usage patterns.
+
+## What are Processors?
+
+In robotics, data comes in many forms - images from cameras, joint positions from sensors, text instructions from users, and more. Each type of data requires specific transformations before a model can use it effectively. Models need this data to be:
+
+- **Normalized**: Scaled to appropriate ranges for neural network processing
+- **Batched**: Organized with proper dimensions for batch processing
+- **Tokenized**: Text converted to numerical representations
+- **Device-placed**: Moved to the right hardware (CPU/GPU)
+- **Type-converted**: Cast to appropriate data types
+
+Processors handle these transformations through composable, reusable steps that can be chained together into pipelines. Think of them as a modular assembly line where each station performs a specific transformation on your data.
+
+## Core Concepts
+
+### EnvTransition: The Universal Data Container
+
+The `EnvTransition` is the fundamental data structure that flows through all processors. It's a typed dictionary that represents a complete robot-environment interaction:
+
+```python
+from lerobot.processor.pipeline import TransitionKey, EnvTransition
+
+# Example transition from a robot collecting data
+transition: EnvTransition = {
+    TransitionKey.OBSERVATION: {
+        "observation.images.camera0": camera0_image_tensor,  # Shape: (H, W, C)
+        "observation.images.camera1": camera1_image_tensor,  # Shape: (H, W, C)
+        "observation.state": joint_positions_tensor,         # Shape: (7,) for 7-DOF arm
+        "observation.environment_state": env_state_tensor    # Shape: (3,) for object position
+    },
+    TransitionKey.ACTION: action_tensor,                    # Shape: (7,) for joint velocities
+    TransitionKey.REWARD: 0.0,                             # Scalar reward signal
+    TransitionKey.DONE: False,                             # Episode termination flag
+    TransitionKey.TRUNCATED: False,                        # Episode truncation flag
+    TransitionKey.INFO: {"success": False},                # Additional metadata
+    TransitionKey.COMPLEMENTARY_DATA: {
+        "task": "pick up the red cube",                    # Language instruction
+        "task_index": 0,                                   # Task identifier
+        "index": 42                                        # Frame index
+    }
+}
+```
+
+Each key in the transition has a specific purpose:
+
+- **OBSERVATION**: All sensor data (images, states, proprioception)
+- **ACTION**: The action to execute or that was executed
+- **REWARD**: Reinforcement learning signal
+- **DONE/TRUNCATED**: Episode boundary indicators
+- **INFO**: Arbitrary metadata
+- **COMPLEMENTARY_DATA**: Task descriptions, indices, padding flags, inter-step data (e.g., you need to compute the velocities and then use this velocity to clip the action)
+
+### ProcessorStep: The Building Block Interface
+
+A `ProcessorStep` is a single transformation unit that processes transitions. It's a protocol (interface) that any processor step must implement:
+
+```python
+from lerobot.processor.pipeline import ProcessorStep, EnvTransition
+from lerobot.configs.types import PolicyFeature
+from typing import Any
+import torch
+
+class MyProcessorStep:
+    """Example processor step interface - all methods must be implemented."""
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        """Transform the transition - this is the main processing logic."""
+        raise NotImplementedError
+
+    def feature_contract(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
+        """Declare how this step transforms feature shapes/types."""
+        raise NotImplementedError
+
+    def get_config(self) -> dict[str, Any]:
+        """Return JSON-serializable configuration for saving/loading."""
+        raise NotImplementedError
+
+    def state_dict(self) -> dict[str, torch.Tensor]:
+        """Return any learnable parameters (tensors only)."""
+        raise NotImplementedError
+
+    def load_state_dict(self, state: dict[str, torch.Tensor]) -> None:
+        """Load learnable parameters from saved state."""
+        raise NotImplementedError
+
+    def reset(self) -> None:
+        """Reset any internal state between episodes."""
+        raise NotImplementedError
+```
+
+### RobotProcessor: The Pipeline Orchestrator
+
+The `RobotProcessor` chains multiple `ProcessorStep` instances together, executing them sequentially. It provides automatic format conversion to handle both batch dictionaries (from datasets) and EnvTransition dictionaries:
+
+```python
+from lerobot.processor.pipeline import RobotProcessor, _default_batch_to_transition, _default_transition_to_batch
+
+# Create a processing pipeline
+processor = RobotProcessor(
+    steps=[
+        step1,  # First transformation
+        step2,  # Second transformation
+        step3   # Third transformation
+    ],
+    name="my_preprocessing_pipeline",
+
+    # Optional: Custom converters for input/output formats
+    to_transition=_default_batch_to_transition,  # How to convert batch dict → EnvTransition
+    to_output=_default_transition_to_batch       # How to convert EnvTransition → output format
+)
+
+# The processor automatically handles different input formats:
+# 1. If input is a batch dict (from dataset), converts to EnvTransition
+# 2. Passes through each step sequentially
+# 3. Converts back to original format (or custom output format)
+
+# Example with batch dict input (common in training)
+batch_dict = {"observation.state": tensor, "action": tensor}
+output = processor(batch_dict)  # Automatically converted to/from EnvTransition
+
+# Example with EnvTransition input (common in inference)
+transition = {TransitionKey.OBSERVATION: {...}, TransitionKey.ACTION: ...}
+output = processor(transition)  # Stays as EnvTransition throughout
+```
+
+The `to_transition` and `to_output` converters enable seamless integration with existing codebases.
+By default, they handle the standard LeRobot batch format, but you can customize them for different data structures.
+
+### Additional Converter Functions
+
+LeRobot provides several specialized converter functions for common robotics scenarios:
+
+```python
+from lerobot.processor.converters import (
+    to_transition_teleop_action,
+    to_transition_robot_observation,
+    to_output_robot_action,
+    to_dataset_frame
+)
+```
+
+**`to_transition_teleop_action`** - Converts teleoperation device actions to EnvTransitions:
+
+```python
+# Use case: Phone, gamepad, or other teleop device control
+phone_action = {"x": 0.1, "y": -0.2, "gripper": 0.8}
+transition = to_transition_teleop_action(phone_action)
+# Creates: {ACTION: {"action.x": 0.1, "action.y": -0.2, "action.gripper": 0.8}, ...}
+```
+
+**`to_transition_robot_observation`** - Converts robot sensor data to EnvTransitions:
+
+```python
+# Use case: Live robot observation during inference
+robot_obs = {
+    "joint_1": 0.5, "joint_2": -0.3,  # joint positions
+    "camera_0": image_array            # camera images
+}
+transition = to_transition_robot_observation(robot_obs)
+# Creates: {OBSERVATION: {"observation.state.joint_1": 0.5, "observation.images.camera_0": image, ...}}
+```
+
+**`to_output_robot_action`** - Extracts robot-executable actions from EnvTransitions:
+
+```python
+# Use case: Converting model outputs back to robot commands
+model_transition = {ACTION: {"action.joint_1": 0.2, "action.joint_2": 0.1}}
+robot_action = to_output_robot_action(model_transition)
+# Returns: {"joint_1": 0.2, "joint_2": 0.1} - ready for robot.send_action()
+```
+
+**`to_dataset_frame`** - Converts transitions to dataset-compatible format:
+
+```python
+# Use case: Saving processed data or creating training batches
+features = {
+    "action": {"names": ["joint_1", "joint_2"]},
+    "observation.state": {"names": ["joint_1", "joint_2"]},
+    "observation.images.camera0": {...}
+}
+batch = to_dataset_frame(transition, features)
+# Returns: {"action": [0.2, 0.1], "observation.state": [0.5, -0.3], ...}
+```
+
+These converters are particularly useful when integrating with real robots, as shown in the examples:
+
+```python
+# Example from phone_so100_teleop.py - Real robot teleoperation
+phone_to_robot_ee_pose = RobotProcessor(
+    steps=[...],
+    to_transition=to_transition_teleop_action,  # Phone → EnvTransition
+    to_output=lambda tr: tr                     # Keep as EnvTransition
+)
+
+# Example from phone_so100_eval.py - Robot action execution
+robot_ee_to_joints = RobotProcessor(
+    steps=[...],
+    to_transition=lambda tr: tr,                # Already EnvTransition
+    to_output=to_output_robot_action           # EnvTransition → Robot action
+)
+
+# Example from phone_so100_record.py - Dataset recording
+robot_joints_to_ee_pose = RobotProcessor(
+    steps=[...],
+    to_transition=to_transition_robot_observation,  # Robot obs → EnvTransition
+    to_output=lambda tr: tr                         # Keep as EnvTransition for dataset
+)
+```
+
+### Data Format Conversion
+
+Different data sources have different formats, but processors need a unified `EnvTransition` structure internally.
+The default converters handle LeRobot datasets, but you can customize them:
+
+```python
+# Default: LeRobot batch format
+lerobot_batch = {
+    "observation.state": torch.tensor(...),
+    "action": torch.tensor(...),
+    "next.reward": torch.tensor(...),
+    "task": ["pick cube", ...]
+}
+# → Converts to EnvTransition → Processes → Converts back
+
+# Custom: Live robot data
+robot_data = {
+    "cameras": {"wrist_cam": np.array(...)},
+    "joint_positions": np.array(...),
+    "gripper_state": 0.5
+}
+
+def robot_to_transition(data: dict) -> EnvTransition:
+    return {
+        TransitionKey.OBSERVATION: {
+            "observation.images.wrist": torch.from_numpy(data["cameras"]["wrist_cam"]),
+            "observation.state": torch.from_numpy(data["joint_positions"])
+        },
+        TransitionKey.ACTION: None,
+        # ... other fields with defaults
+    }
+
+# Use custom converter
+processor = RobotProcessor(
+    steps=[...],
+    to_transition=robot_to_transition,
+    to_output=lambda transition: transition  # Keep as EnvTransition
+)
+```
+
+**When to customize:** Live robot data, Gymnasium environments, legacy datasets, or any non-LeRobot format.
+
+## Common Processor Steps
+
+LeRobot provides a rich set of pre-built processor steps for common transformations.
+Let's explore each in detail:
+
+### Data Normalization
+
+Normalization is crucial for neural network training and inference.
+The `NormalizerProcessor` handles both mean-std normalization and min-max scaling:
+
+```python
+from lerobot.processor.normalize_processor import NormalizerProcessor, UnnormalizerProcessor
+from lerobot.configs.types import PolicyFeature, FeatureType, NormalizationMode
+
+# Define what features exist in your data
+features = {
+    "observation.images.camera0": PolicyFeature(
+        type=FeatureType.IMAGE,
+        shape=(224, 224, 3)
+    ),
+    "observation.state": PolicyFeature(
+        type=FeatureType.STATE,
+        shape=(7,)
+    ),
+    "action": PolicyFeature(
+        type=FeatureType.ACTION,
+        shape=(7,)
+    )
+}
+
+# Define normalization strategy per feature type
+norm_map = {
+    FeatureType.IMAGE: NormalizationMode.MEAN_STD,    # Images: (x - mean) / std
+    FeatureType.STATE: NormalizationMode.MIN_MAX,     # States: scale to [-1, 1]
+    FeatureType.ACTION: NormalizationMode.MIN_MAX     # Actions: scale to [-1, 1]
+}
+
+# Create normalizer with dataset statistics
+normalizer = NormalizerProcessor(
+    features=features,
+    norm_map=norm_map,
+    stats=dataset.meta.stats,  # Contains mean, std, min, max per feature
+    normalize_keys={"observation.state", "action"}  # Optional: only normalize specific keys
+)
+
+# For postprocessing: inverse transformation
+unnormalizer = UnnormalizerProcessor(
+    features=features,
+    norm_map=norm_map,
+    stats=dataset.meta.stats
+)
+
+# The normalizer automatically:
+# - Detects which normalization to apply based on feature type
+# - Handles device placement of statistics tensors
+# - Skips keys not in stats or not in normalize_keys
+# - Adds metadata about what was normalized
+```
+
+### Device Management
+
+The `DeviceProcessor` ensures tensors are on the right device with the right dtype:
+
+```python
+from lerobot.processor.device_processor import DeviceProcessor
+
+# Basic GPU placement
+gpu_processor = DeviceProcessor(device="cuda:0")
+
+# Advanced: GPU with half-precision for inference
+efficient_processor = DeviceProcessor(
+    device="cuda:0",
+    float_dtype="float16"  # Convert float32 -> float16 for memory efficiency
+)
+
+# The processor:
+# - Moves all tensors to specified device
+# - Preserves non-tensor data unchanged
+# - Optionally converts float dtypes while preserving int/bool types
+# - Uses non_blocking transfers for CUDA devices
+# - Handles nested structures (observations, complementary_data)
+
+# Supported float dtypes:
+# "float16" / "half": 16-bit floating point
+# "float32" / "float": 32-bit floating point (default)
+# "float64" / "double": 64-bit floating point
+# "bfloat16": Brain floating point (better for training)
+```
+
+### Batch Processing
+
+Models expect batched inputs, but robot interactions often produce unbatched data:
+
+```python
+from lerobot.processor.batch_processor import ToBatchProcessor
+
+batch_processor = ToBatchProcessor()
+
+# Automatically adds batch dimensions where needed:
+# State: (7,) -> (1, 7)
+# Image: (224, 224, 3) -> (1, 224, 224, 3)
+# Action: (4,) -> (1, 4)
+# Task: "pick_cube" -> ["pick_cube"]
+# Already batched: (1, 7) -> (1, 7) [unchanged]
+
+# The processor intelligently:
+# - Detects tensor dimensionality
+# - Adds batch dim to 1D states/actions
+# - Adds batch dim to 3D images
+# - Wraps string tasks in lists
+# - Preserves already-batched data
+
+# Example usage in inference:
+single_observation = robot.get_observation()  # Unbatched
+batched_input = batch_processor({"observation": single_observation})
+model_output = model(batched_input)  # Model expects batch dim
+```
+
+### Text Tokenization
+
+For language-conditioned policies, text instructions must be tokenized:
+
+```python
+from lerobot.processor.tokenizer_processor import TokenizerProcessor
+from transformers import AutoTokenizer
+
+# Option 1: Auto-load tokenizer by name
+tokenizer_proc = TokenizerProcessor(
+    tokenizer_name="google/paligemma-3b-pt-224",
+    max_length=128,
+    task_key="task",           # Where to find text in complementary_data
+    padding="max_length",       # Pad to max_length
+    padding_side="right",
+    truncation=True            # Truncate if longer than max_length
+)
+
+# Option 2: Provide custom tokenizer
+custom_tokenizer = AutoTokenizer.from_pretrained("microsoft/DialoGPT-medium")
+custom_proc = TokenizerProcessor(
+    tokenizer=custom_tokenizer,
+    max_length=256,
+    padding_side="left"        # For autoregressive models
+)
+
+# The processor:
+# - Extracts task text from complementary_data
+# - Tokenizes using HuggingFace tokenizer
+# - Adds tokens and attention_mask to observations
+# - Handles both single strings and lists of strings
+# - Preserves original task in complementary_data
+
+# Output structure:
+# observation["observation.language.tokens"] = tensor([101, 2032, ...])
+# observation["observation.language.attention_mask"] = tensor([1, 1, 0, ...])
+```
+
+### Key Renaming
+
+Different datasets and models may use different naming conventions.
+The `RenameProcessor` solves this mismatch:
+
+**Why is this useful?**
+
+- When loading a model trained on a different dataset with different key names
+- When using foundation models that expect specific key naming conventions
+- When standardizing datasets from different sources
+- When adapting legacy code to new naming standards
+
+```python
+from lerobot.processor.rename_processor import RenameProcessor
+
+# Example 1: Dataset uses "top"/"wrist", model expects "camera0"/"camera1"
+rename_proc = RenameProcessor(
+    rename_map={
+        "observation.images.top": "observation.images.camera0",
+        "observation.images.wrist": "observation.images.camera1",
+    }
+)
+
+# Example 2: Foundation model compatibility
+# Your dataset: "observation.state", Foundation model: "proprio"
+foundation_rename = RenameProcessor(
+    rename_map={
+        "observation.state": "proprio",
+        "observation.images.main": "rgb",
+    }
+)
+
+# Example 3: Standardizing multiple datasets
+standardize_rename = RenameProcessor(
+    rename_map={
+        # Different robots might use different names
+        "observation.joint_positions": "observation.state",
+        "observation.gripper_state": "observation.end_effector",
+        "observation.arm_camera": "observation.images.wrist",
+    }
+)
+```
+
+## Building Complete Pipelines
+
+Let's build a real-world preprocessing and postprocessing pipeline for a vision-based
+manipulation policy:
+
+```python
+# Consolidated imports
+from lerobot.processor import (
+    RobotProcessor,
+    NormalizerProcessor,
+    UnnormalizerProcessor,
+    DeviceProcessor,
+    ToBatchProcessor,
+    TokenizerProcessor,
+    RenameProcessor
+)
+
+# Step 1: Define the preprocessing pipeline
+preprocessor = RobotProcessor(
+    steps=[
+        # 1. Standardize naming from dataset
+        RenameProcessor(
+            rename_map={
+                "observation.images.top": "observation.images.camera0",
+                "observation.images.wrist": "observation.images.camera1"
+            }
+        ),
+
+        # 2. Add batch dimensions for model
+        ToBatchProcessor(),
+
+        # 3. Tokenize language instructions if present
+        TokenizerProcessor(
+            tokenizer_name="google/paligemma-3b-pt-224",
+            max_length=64,
+            task_key="task"
+        ),
+
+        # 4. Normalize numerical data
+        NormalizerProcessor(
+            features=policy_features,
+            norm_map={
+                FeatureType.IMAGE: NormalizationMode.MEAN_STD,
+                FeatureType.STATE: NormalizationMode.MIN_MAX,
+                FeatureType.ACTION: NormalizationMode.MIN_MAX
+            },
+            stats=dataset.meta.stats
+        ),
+
+        # 5. Move to GPU and convert to half precision
+        DeviceProcessor(
+            device="cuda:0",
+            float_dtype="float16"
+        )
+    ],
+    name="robot_preprocessor"
+)
+
+# Step 2: Define the postprocessing pipeline
+postprocessor = RobotProcessor(
+    steps=[
+        # 1. Move back to CPU for robot hardware
+        DeviceProcessor(device="cpu"),
+
+        # 2. Denormalize actions to original scale
+        UnnormalizerProcessor(
+            features=policy_features,
+            norm_map={
+                FeatureType.ACTION: NormalizationMode.MIN_MAX
+            },
+            stats=dataset.meta.stats
+        )
+    ],
+    name="robot_postprocessor"
+)
+```
+
+## Using Processors in Practice
+
+### Training Loop Integration
+
+Here's how processors integrate into a training loop using the policy's forward method:
+
+```python
+from torch.utils.data import DataLoader
+
+# Create dataset and dataloader
+dataset = LeRobotDataset(repo_id="your_dataset")
+dataloader = DataLoader(dataset, batch_size=32, shuffle=True)
+
+# Initialize model and processors
+model = YourPolicy.from_pretrained("your_model")
+preprocessor = RobotProcessor.from_pretrained(
+    "your_model",
+    config_filename="robot_preprocessor.json"
+)
+
+# Training loop
+for epoch in range(num_epochs):
+    for batch in dataloader:
+        # Preprocess batch
+        processed_batch = preprocessor(batch)
+
+        # Forward pass - returns loss and optional metrics
+        loss, metrics = model.forward(processed_batch)
+
+        # Backward pass
+        optimizer.zero_grad()
+        loss.backward()
+        optimizer.step()
+
+        # Log metrics if available
+        if metrics:
+            wandb.log(metrics)
+```
+
+### Inference Pipeline
+
+For deployment, processors ensure consistent data handling with real robots:
+
+```python
+# Load model and processors
+policy = YourPolicy.from_pretrained("path/to/model")
+preprocessor = RobotProcessor.from_pretrained(
+    "path/to/model",
+    config_filename="robot_preprocessor.json"
+)
+postprocessor = RobotProcessor.from_pretrained(
+    "path/to/model",
+    config_filename="robot_postprocessor.json"
+)
+
+# Connect to robot
+robot = make_robot_from_config(robot_config)
+robot.connect()
+
+# Inference loop
+policy.eval()
+# Reset the policy and processors
+policy.reset()
+preprocessor.reset()
+postprocessor.reset()
+
+with torch.no_grad():
+    while not done:
+        # Get observation from robot
+        observation = robot.get_observation()
+
+        # Build dataset-compatible frame
+        observation_frame = build_dataset_frame(
+            dataset.features,
+            observation,
+            prefix="observation"
+        )
+
+        # Add task instruction to complementary data
+        observation_frame["task"] = "pick up the red cube"
+
+        # Preprocess for model
+        model_input = preprocessor(observation_frame)
+
+        # Run policy
+        raw_action = policy.select_action(model_input)
+
+        # Postprocess action
+        action_transition = {TransitionKey.ACTION: raw_action}
+        processed = postprocessor(action_transition)
+        action = processed[TransitionKey.ACTION]
+
+        # Convert to robot action format
+        robot_action = {
+            key: action[i].item()
+            for i, key in enumerate(robot.action_features)
+        }
+
+        # Execute on robot
+        robot.send_action(robot_action)
+```
+
+## Saving and Loading Processors
+
+Processors can be persisted and shared just like models, making them portable across different
+environments and ensuring reproducibility:
+
+### Local Save/Load
+
+```python
+# Save processor configuration and state
+preprocessor.save_pretrained(
+    "./my_robot_processor",
+    config_filename="preprocessor.json"  # Optional custom name
+)
+
+# The save creates:
+# my_robot_processor/
+# ├── preprocessor.json                    # Configuration
+# ├── preprocessor_step_0_normalizer.safetensors  # Step 0 state (stats)
+# └── preprocessor_step_1_device.safetensors     # Step 1 state (if any)
+
+# Load processor
+loaded = RobotProcessor.from_pretrained(
+    "./my_robot_processor",
+    config_filename="preprocessor.json"
+)
+```
+
+### HuggingFace Hub Integration
+
+The HuggingFace Hub provides a centralized place to share and version your processors.
+This is particularly useful for sharing preprocessing configurations with models,
+ensuring that anyone who downloads your model can reproduce your exact preprocessing pipeline.
+It also enables versioning and collaboration on preprocessing strategies.
+
+```python
+# Save to HuggingFace Hub
+preprocessor.save_pretrained("username/my-robot-policy")
+
+# Load from Hub with automatic download
+hub_processor = RobotProcessor.from_pretrained(
+    "username/my-robot-policy",
+    config_filename="robot_preprocessor.json",
+    revision="main",  # Optional: specific revision
+    cache_dir="./cache"  # Optional: local cache directory
+)
+
+# The Hub integration provides:
+# - Automatic versioning with git
+# - Public or private sharing
+# - Download caching for efficiency
+# - Integration with model repositories
+```
+
+### Loading with Overrides
+
+Sometimes you need to modify loaded processors for new environments or datasets.
+The override mechanism allows you to update specific processor configurations without modifying
+the saved files:
+
+```python
+# Load processor with configuration overrides
+processor = RobotProcessor.from_pretrained(
+    "./saved_processor",
+    overrides={
+        # Change device for different hardware
+        "device_processor": {"device": "cuda:1"},
+
+        # Update statistics for new dataset
+        "normalizer_processor": {"stats": new_dataset.meta.stats},
+
+        # Provide non-serializable objects (like tokenizers)
+        "tokenizer_processor": {"tokenizer": custom_tokenizer}
+    }
+)
+
+# Common override scenarios:
+# 1. Adapting to different hardware (GPU availability)
+# 2. Fine-tuning on new datasets with different statistics
+# 3. Providing runtime dependencies that can't be serialized
+# 4. Testing variations without creating new saved configs
+```
+
+## Creating Custom Processor Steps
+
+Build your own processor steps for specialized transformations.
+The key is implementing the required interface:
+
+### Basic Custom Step with Registration
+
+The registration mechanism allows your custom processors to be saved and loaded by name rather
+than by module path.
+This makes them more portable and easier to share:
+
+```python
+from dataclasses import dataclass
+from lerobot.processor.pipeline import ProcessorStepRegistry, ObservationProcessor
+
+# The @register decorator adds your processor to the global registry
+# Use a unique name, preferably namespaced to avoid conflicts
+@dataclass
+@ProcessorStepRegistry.register("my_company/gaussian_noise")
+class GaussianNoiseProcessor(ObservationProcessor):
+    """Add Gaussian noise to observations for robustness training."""
+
+    noise_std: float = 0.01
+    training_only: bool = True
+    is_training: bool = True
+
+    def observation(self, observation):
+        """Add noise to observation tensors."""
+        if not self.is_training and self.training_only:
+            return observation
+
+        noisy_obs = {}
+        for key, value in observation.items():
+            if isinstance(value, torch.Tensor) and "image" not in key:
+                # Add noise to non-image observations
+                noise = torch.randn_like(value) * self.noise_std
+                noisy_obs[key] = value + noise
+            else:
+                noisy_obs[key] = value
+
+        return noisy_obs
+
+    def get_config(self):
+        return {
+            "noise_std": self.noise_std,
+            "training_only": self.training_only,
+            "is_training": self.is_training
+        }
+
+# Why register?
+# 1. Enables saving by name: config saves "my_company/gaussian_noise" instead of full module path
+# 2. More portable: Others can use your processor without your exact module structure
+# 3. Version-safe: Module refactoring won't break saved configs
+# 4. Cleaner configs: JSON shows readable names instead of long import paths
+```
+
+### Using Base Classes for Common Patterns
+
+LeRobot provides base classes like `ObservationProcessor`, `ActionProcessor`, etc., that handle
+the boilerplate of extracting and reinserting specific components:
+
+```python
+from lerobot.processor import ActionProcessor
+
+@dataclass
+@ProcessorStepRegistry.register("my_company/action_clipper")
+class ActionClipProcessor(ActionProcessor):
+    """Clip actions to safe ranges."""
+
+    min_value: float = -1.0
+    max_value: float = 1.0
+
+    def action(self, action):
+        """Process only the action component."""
+        # No need to handle transition dict - base class does it
+        return torch.clamp(action, self.min_value, self.max_value)
+
+    def get_config(self):
+        return {"min_value": self.min_value, "max_value": self.max_value}
+```
+
+For more advanced processor patterns including stateful processors, see [Implement Your Own Processor](implement_your_own_processor.mdx).
+
+## Advanced Features
+
+### Debugging with Hooks
+
+Processors support hooks for monitoring and debugging without modifying the pipeline code:
+
+```python
+# Define monitoring hooks
+def log_shapes(step_idx: int, transition: EnvTransition):
+    """Log tensor shapes after each step."""
+    obs = transition.get(TransitionKey.OBSERVATION)
+    if obs:
+        print(f"Step {step_idx} shapes:")
+        for key, value in obs.items():
+            if isinstance(value, torch.Tensor):
+                print(f"  {key}: {value.shape}")
+
+def check_nans(step_idx: int, transition: EnvTransition):
+    """Check for NaN values."""
+    obs = transition.get(TransitionKey.OBSERVATION)
+    if obs:
+        for key, value in obs.items():
+            if isinstance(value, torch.Tensor) and torch.isnan(value).any():
+                print(f"Warning: NaN detected in {key} at step {step_idx}")
+
+# Register hooks
+processor.register_after_step_hook(log_shapes)
+processor.register_after_step_hook(check_nans)
+
+# Process data - hooks will be called after each step
+output = processor(input_data)
+
+# Remove hooks when done debugging
+processor.unregister_after_step_hook(log_shapes)
+processor.unregister_after_step_hook(check_nans)
+```
+
+### Step-by-Step Inspection
+
+Use `step_through()` for detailed debugging of the transformation pipeline:
+
+```python
+# Inspect data at each transformation stage
+for i, intermediate in enumerate(processor.step_through(data)):
+    print(f"\n=== After step {i} ===")
+
+    # Check observation shapes
+    obs = intermediate.get(TransitionKey.OBSERVATION)
+    if obs:
+        for key, value in obs.items():
+            if isinstance(value, torch.Tensor):
+                print(f"{key}: shape={value.shape}, "
+                      f"dtype={value.dtype}, "
+                      f"device={value.device}, "
+                      f"range=[{value.min():.3f}, {value.max():.3f}]")
+
+    # Check action if present
+    action = intermediate.get(TransitionKey.ACTION)
+    if action is not None and isinstance(action, torch.Tensor):
+        print(f"action: shape={action.shape}, range=[{action.min():.3f}, {action.max():.3f}]")
+```
+
+### Pipeline Slicing
+
+Extract subsets of a pipeline for testing or creating variations:
+
+```python
+# Get specific steps
+first_three_steps = processor[:3]  # Returns new RobotProcessor
+middle_step = processor[2]         # Returns single ProcessorStep
+
+# Test individual steps
+test_input = {...}
+step_output = processor[0](test_input)  # Test first step only
+
+# Create variations
+variant_processor = RobotProcessor(
+    steps=processor.steps[:-1] + [new_final_step],
+    name="variant"
+)
+```
+
+## Best Practices and Tips
+
+### 1. Order Matters
+
+The sequence of processors is crucial. Follow this general order:
+
+```python
+# Preprocessing: Raw → Model-ready
+1. Rename (standardize keys)
+2. Batch (add dimensions)
+3. Tokenize (text → tokens)
+4. Normalize (scale values)
+5. Device (move to GPU)
+
+# Postprocessing: Model → Robot-ready
+1. Device (move to CPU)
+2. Unnormalize (restore scale)
+3. Unbatch (remove dimensions if needed)
+```
+
+### 2. Registration Best Practices
+
+```python
+# Always register custom steps for better portability
+@ProcessorStepRegistry.register("my_company/special_processor")
+class SpecialProcessor:
+    ...
+
+# Use namespaced names to avoid conflicts
+# Good: "my_company/augmentation"
+# Bad: "augmentation" (too generic)
+
+# Check registered processors
+print(ProcessorStepRegistry.list())  # See all registered processors
+```
+
+### 3. Common Pitfalls and Solutions
+
+**Tensor Device Mismatch:**
+
+```python
+# Problem: RuntimeError: Expected all tensors on same device
+# Solution: Ensure DeviceProcessor is in pipeline
+preprocessor = RobotProcessor(
+    steps=[
+        NormalizerProcessor(...),
+        DeviceProcessor(device="cuda")  # Add this
+    ]
+)
+```
+
+**Missing Statistics:**
+
+```python
+# Problem: NormalizerProcessor has no stats
+# Solution 1: Compute stats from dataset
+from lerobot.datasets.compute_stats import compute_stats
+stats = compute_stats(dataset)
+
+# Solution 2: Load with overrides
+processor = RobotProcessor.from_pretrained(
+    "model_path",
+    overrides={"normalizer_processor": {"stats": dataset.meta.stats}}
+)
+```
+
+## Next Steps
+
+Now that you understand processors, explore these topics:
+
+- [**Implement Your Own Processor**](implement_your_own_processor.mdx) - Deep dive into creating custom processors with advanced features like stateful processing
+- [**Policy Documentation**](policies.mdx) - Learn how different policies use processors
+- [**Dataset Documentation**](datasets.mdx) - Understand the data format that processors transform
+- [**Training Guide**](training.mdx) - See processors in action during model training
+- [**Evaluation Guide**](evaluation.mdx) - Learn about processor usage during policy evaluation
+
+## Summary
+
+Processors are the unsung heroes of robotics pipelines, handling the critical transformations between raw sensor data and model-ready tensors. By understanding and effectively using processors, you can:
+
+- Build robust, reusable data pipelines
+- Share preprocessing configurations across projects
+- Debug data transformations systematically
+- Ensure consistency between training and deployment
+- Create custom transformations for specialized tasks
+
+Remember: good preprocessing is often the difference between a model that works in theory
+and one that works in practice!
+The modular pipeline approach ensures your transformations are testable, reproducible,
+and portable across different robots and environments.

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
-lerobot-find-port
+python -m 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
-lerobot-setup-motors \
+python -m 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
-lerobot-setup-motors \
+python -m 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
-lerobot-calibrate \
+python -m 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
-lerobot-calibrate \
+python -m 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
-lerobot-find-port
+python -m 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
-lerobot-setup-motors \
+python -m 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
-lerobot-calibrate \
+python -m 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
-lerobot-calibrate \
+python -m 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/lerobot-dataset-v3.mdx

@@ -1,169 +0,0 @@
-# LeRobotDataset v3.0
-
-`LeRobotDataset v3.0` is a standardized format for robot learning data. It provides unified access to multi-modal time-series data, sensorimotor signals and multi‑camera video, as well as rich metadata for indexing, search, and visualization on the Hugging Face Hub.
-
-This docs will guide you to:
-
-- Understand the v3.0 design and directory layout
-- Record a dataset and push it to the Hub
-- Load datasets for training with `LeRobotDataset`
-- Stream datasets without downloading using `StreamingLeRobotDataset`
-- Migrate existing `v2.1` datasets to `v3.0`
-
-## What’s new in `v3`
-
-- **File-based storage**: Many episodes per Parquet/MP4 file (v2 used one file per episode).
-- **Relational metadata**: Episode boundaries and lookups are resolved through metadata, not filenames.
-- **Hub-native streaming**: Consume datasets directly from the Hub with `StreamingLeRobotDataset`.
-- **Lower file-system pressure**: Fewer, larger files ⇒ faster initialization and fewer issues at scale.
-- **Unified organization**: Clean directory layout with consistent path templates across data and videos.
-
-## Installation
-
-`LeRobotDataset v3.0` will be included in `lerobot >= 0.4.0`.
-
-Until that stable release, you can use the main branch by following the [build from source instructions](./installation#from-source).
-
-## Record a dataset
-
-Run the command below to record a dataset with the SO-101 and push to the Hub:
-
-```bash
-lerobot-record \
-  --robot.type=so101_follower \
-  --robot.port=/dev/tty.usbmodem585A0076841 \
-  --robot.id=my_awesome_follower_arm \
-  --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 1920, height: 1080, fps: 30}}" \
-  --teleop.type=so101_leader \
-  --teleop.port=/dev/tty.usbmodem58760431551 \
-  --teleop.id=my_awesome_leader_arm \
-  --display_data=true \
-  --dataset.repo_id=${HF_USER}/record-test \
-  --dataset.num_episodes=5 \
-  --dataset.single_task="Grab the black cube"
-```
-
-See the [recording guide](./il_robots#record-a-dataset) for more details.
-
-## Format design
-
-A core v3 principle is **decoupling storage from the user API**: data is stored efficiently (few large files), while the public API exposes intuitive episode-level access.
-
-`v3` has three pillars:
-
-1. **Tabular data**: Low‑dimensional, high‑frequency signals (states, actions, timestamps) stored in **Apache Parquet**. Access is memory‑mapped or streamed via the `datasets` stack.
-2. **Visual data**: Camera frames concatenated and encoded into **MP4**. Frames from the same episode are grouped; videos are sharded per camera for practical sizes.
-3. **Metadata**: JSON/Parquet records describing schema (feature names, dtypes, shapes), frame rates, normalization stats, and **episode segmentation** (start/end offsets into shared Parquet/MP4 files).
-
-> To scale to millions of episodes, tabular rows and video frames from multiple episodes are **concatenated** into larger files. Episode‑specific views are reconstructed **via metadata**, not file boundaries.
-
-<div style="display:flex; justify-content:center; gap:12px; flex-wrap:wrap;">
-  <figure style="margin:0; text-align:center;">
-    <img
-      src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobotdataset-v3/asset1datasetv3.png"
-      alt="LeRobotDataset v3 diagram"
-      width="220"
-    />
-    <figcaption style="font-size:0.9em; color:#666;">
-      From episode‑based to file‑based datasets
-    </figcaption>
-  </figure>
-</div>
-
-### Directory layout (simplified)
-
-- **`meta/info.json`**: canonical schema (features, shapes/dtypes), FPS, codebase version, and **path templates** to locate data/video shards.
-- **`meta/stats.json`**: global feature statistics (mean/std/min/max) used for normalization; exposed as `dataset.meta.stats`.
-- **`meta/tasks.jsonl`**: natural‑language task descriptions mapped to integer IDs for task‑conditioned policies.
-- **`meta/episodes/`**: per‑episode records (lengths, tasks, offsets) stored as **chunked Parquet** for scalability.
-- **`data/`**: frame‑by‑frame **Parquet** shards; each file typically contains **many episodes**.
-- **`videos/`**: **MP4** shards per camera; each file typically contains **many episodes**.
-
-## Load a dataset for training
-
-`LeRobotDataset` returns Python dictionaries of PyTorch tensors and integrates with `torch.utils.data.DataLoader`. Here is a code example showing its use:
-
-```python
-import torch
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-
-repo_id = "yaak-ai/L2D-v3"
-
-# 1) Load from the Hub (cached locally)
-dataset = LeRobotDataset(repo_id)
-
-# 2) Random access by index
-sample = dataset[100]
-print(sample)
-# {
-#   'observation.state': tensor([...]),
-#   'action': tensor([...]),
-#   'observation.images.front_left': tensor([C, H, W]),
-#   'timestamp': tensor(1.234),
-#   ...
-# }
-
-# 3) Temporal windows via delta_timestamps (seconds relative to t)
-delta_timestamps = {
-    "observation.images.front_left": [-0.2, -0.1, 0.0]  # 0.2s and 0.1s before current frame
-}
-
-dataset = LeRobotDataset(repo_id, delta_timestamps=delta_timestamps)
-
-# Accessing an index now returns a stack for the specified key(s)
-sample = dataset[100]
-print(sample["observation.images.front_left"].shape)  # [T, C, H, W], where T=3
-
-# 4) Wrap with a DataLoader for training
-batch_size = 16
-data_loader = torch.utils.data.DataLoader(dataset, batch_size=batch_size)
-
-device = "cuda" if torch.cuda.is_available() else "cpu"
-for batch in data_loader:
-    observations = batch["observation.state"].to(device)
-    actions = batch["action"].to(device)
-    images = batch["observation.images.front_left"].to(device)
-    # model.forward(batch)
-```
-
-## Stream a dataset (no downloads)
-
-Use `StreamingLeRobotDataset` to iterate directly from the Hub without local copies. This allows to stream large datasets without the need to downloading them onto disk or loading them onto memory, and is a key feature of the new dataset format.
-
-```python
-from lerobot.datasets.streaming_dataset import StreamingLeRobotDataset
-
-repo_id = "yaak-ai/L2D-v3"
-dataset = StreamingLeRobotDataset(repo_id)  # streams directly from the Hub
-```
-
-<div style="display:flex; justify-content:center; gap:12px; flex-wrap:wrap;">
-  <figure style="margin:0; text-align:center;">
-    <img
-      src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobotdataset-v3/streaming-lerobot.png"
-      alt="StreamingLeRobotDataset"
-      width="520"
-    />
-    <figcaption style="font-size:0.9em; color:#666;">
-      Stream directly from the Hub for on‑the‑fly training.
-    </figcaption>
-  </figure>
-</div>
-
-## Migrate `v2.1` → `v3.0`
-
-A converter aggregates per‑episode files into larger shards and writes episode offsets/metadata. Convert your dataset using the instructions below.
-
-```bash
-# Pre-release build with v3 support:
-pip install "https://github.com/huggingface/lerobot/archive/33cad37054c2b594ceba57463e8f11ee374fa93c.zip"
-
-# Convert an existing v2.1 dataset hosted on the Hub:
-python -m lerobot.datasets.v30.convert_dataset_v21_to_v30 --repo-id=<HF_USER/DATASET_ID>
-```
-
-**What it does**
-
-- Aggregates parquet files: `episode-0000.parquet`, `episode-0001.parquet`, … → **`file-0000.parquet`**, …
-- Aggregates mp4 files: `episode-0000.mp4`, `episode-0001.mp4`, … → **`file-0000.mp4`**, …
-- Updates `meta/episodes/*` (chunked Parquet) with per‑episode lengths, tasks, and byte/frame offsets.

docs/source/phone_teleop.mdx

@@ -0,0 +1,195 @@
+# Phone
+
+Use your phone (iOS or Android) to control your robot.
+
+**In this guide you'll learn:**
+
+- How to connect an iOS/Android phone
+- How phone pose is mapped to robot end‑effector (EE) targets
+- How to tweak safety limits, gripper control, and IK settings
+
+To use phone to control your robot, install the relevant dependencies with:
+
+```bash
+pip install lerobot[phone]
+```
+
+## Get started
+
+### Supported platforms
+
+- iOS: Uses the HEBI Mobile I/O app (ARKit pose + buttons). Download the app first, open it and the examples will discover it on your network and stream the phone pose and inputs.
+- Android: Uses the `teleop` package (WebXR). When you start the Python process, it prints a local URL. Open the link on your phone, tap Start, then use Move to stream pose.
+
+Links:
+
+- Android WebXR library: [`teleop` on PyPI](https://pypi.org/project/teleop/)
+- iOS app: [HEBI Mobile I/O](https://docs.hebi.us/tools.html#mobile-io)
+
+### Phone orientation and controls
+
+- Orientation: hold the phone with the screen facing up and the top edge pointing in the same direction as the robot gripper. This ensures calibration aligns the phone’s frame with the robot frame so motion feels natural.
+- Enable/disable:
+  - iOS: Hold `B1` to enable teleoperation, release to stop. The first press captures a reference pose.
+  - Android: Press and hold the `Move` button, release to stop. The first press captures a reference pose.
+- Gripper control:
+  - iOS: Analog input `A3` controls the gripper as velocity input.
+  - Android: Buttons `A` and `B` act like increment/decrement (A opens, B closes). You can tune velocity in the `GripperVelocityToJoint` step.
+
+### Step 1: Choose the platform
+
+Modify the examples to use `PhoneOS.IOS` or `PhoneOS.ANDROID` in `PhoneConfig`. The API is identical across platforms, only the input source differs. All examples are under `examples/` and have `phone_so100_*.py` variants.
+
+Teleoperation example:
+
+```36:43:examples/phone_so100_teleop.py
+from lerobot.teleoperators.phone.config_phone import PhoneConfig, PhoneOS
+
+teleop_config = PhoneConfig(phone_os=PhoneOS.IOS)  # or PhoneOS.ANDROID
+teleop_device = Phone(teleop_config)
+```
+
+### Step 2: Connect and calibrate
+
+When `Phone(teleop_config)` is created and `connect()` is called, calibration is prompted automatically. Hold the phone in the orientation described above, then:
+
+- iOS: press and hold `B1` to capture the reference pose.
+- Android: press `Move` button on the WebXR page to capture the reference pose.
+
+Why calibrate? We capture the current pose so subsequent poses are expressed in a robot aligned frame. When you again press the button to enable control, the position is recaptured to avoid drift when your phone is repositioned while it was disabled.
+
+### Step 3: Run an example
+
+Run on of the examples scripts to teleoperate, record a dataset, replay a dataset or evaluate a policy.
+
+All scripts assume you configured your robot (e.g., SO-100 follower) and set the correct serial port.
+
+- Android: after starting the script, open the printed local URL on your phone, tap Start, then press and hold Move.
+- iOS: open HEBI Mobile I/O first; B1 enables motion. A3 controls the gripper.
+
+You can customize mapping or safety limits by editing the processor steps shown in the examples.
+
+You can also remap inputs (e.g., use a different analog input) or adapt the pipeline to other robots (e.g., LeKiwi) by modifying the input and kinematics steps. More about this in the [Processors for Robots and Teleoperators](./processors_robots_teleop.mdx) guide.
+
+- Run this example to teleoperate:
+
+  ```bash
+  python examples/phone_so100_teleop.py
+  ```
+
+- Run this example to record a dataset, which saves absolute end effector observations and actions:
+
+  ```bash
+  python examples/phone_so100_record.py
+  ```
+
+- Run this example to replay recorded episodes:
+
+  ```bash
+  python examples/phone_so100_replay.py
+  ```
+
+- Run this example to evaluate a pretrained policy:
+
+  ```bash
+  python examples/phone_so100_eval.py
+  ```
+
+### Important pipeline steps and options
+
+- Kinematics are used in multiple steps. We use [Placo](https://github.com/Rhoban/placo) which is a wrapper around Pinocchio for handling our kinematics. We construct the kinematics object by passing the robot's URDF and target frame. We set `target_frame_name` to the gripper frame.
+
+  ```44:49:examples/phone_so100_teleop.py
+  RobotKinematics(
+      urdf_path="./src/lerobot/teleoperators/sim/so101_new_calib.urdf",
+      target_frame_name="gripper_frame_link",
+      joint_names=list(robot.bus.motors.keys()),
+  )
+  ```
+
+- The `MapPhoneActionToRobotAction` step converts the calibrated phone pose and inputs into target deltas and gripper commands, below is shown what the step outputs.
+
+  ```72:83:src/lerobot/teleoperators/phone/phone_processor.py
+  # Map calibrated phone pose to robot targets (enabled gates the motion)
+  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,
+      }
+  )
+  ```
+
+- The `EEReferenceAndDelta` step converts target deltas to an absolute desired EE pose, storing a reference on enable, the `end_effector_step_sizes` are the step sizes for the EE pose and can be modified to change the motion speed.
+
+  ```56:65:examples/phone_so100_teleop.py
+  EEReferenceAndDelta(
+      kinematics=kinematics_solver,
+      end_effector_step_sizes={"x": 0.5, "y": 0.5, "z": 0.5},
+      motor_names=list(robot.bus.motors.keys()),
+  )
+  ```
+
+- The `EEBoundsAndSafety` step clamps EE motion to a workspace and checks for large ee step jumps to ensure safety. The `end_effector_bounds` are the bounds for the EE pose and can be modified to change the workspace. The `max_ee_step_m` and `max_ee_twist_step_rad` are the step limits for the EE pose and can be modified to change the safety limits.
+
+  ```61:66:examples/phone_so100_teleop.py
+  EEBoundsAndSafety(
+      end_effector_bounds={"min": [-1.0, -1.0, -1.0], "max": [1.0, 1.0, 1.0]},
+      max_ee_step_m=0.10,
+      max_ee_twist_step_rad=0.50,
+  )
+  ```
+
+- The `GripperVelocityToJoint` step turns a velocity‑like gripper input into absolute gripper position using the current measured state. The `speed_factor` is the factor by which the velocity is multiplied.
+
+  ```78:81:examples/phone_so100_teleop.py
+  GripperVelocityToJoint(
+      motor_names=list(robot.bus.motors.keys()),
+      speed_factor=20.0,
+  )
+  ```
+
+#### Different IK initial guesses
+
+We use different IK initial guesses in the kinematic steps. As initial guess either the current measured joints or the previous IK solution is used.
+
+- Closed loop (used in record/eval): sets `initial_guess_current_joints=True` so IK starts from the measured joints each frame.
+
+  ```71:76:examples/phone_so100_eval.py
+  InverseKinematicsEEToJoints(
+      kinematics=kinematics_solver,
+      motor_names=list(robot.bus.motors.keys()),
+      initial_guess_current_joints=True,  # closed loop
+  )
+  ```
+
+- Open loop (used in replay): sets `initial_guess_current_joints=False` so IK continues from the previous IK solution rather than the measured state. This preserves action stability when we replay without feedback.
+
+  ```80:86:examples/phone_so100_replay.py
+  InverseKinematicsEEToJoints(
+      kinematics=kinematics_solver,
+      motor_names=list(robot.bus.motors.keys()),
+      initial_guess_current_joints=False,  # open loop
+  )
+  ```
+
+### Pipeline steps explained
+
+- MapPhoneActionToRobotAction: converts calibrated phone pose and inputs into target deltas and a gripper command. Motion is gated by an enable signal (B1 on iOS, Move on Android).
+- AddRobotObservationAsComplimentaryData: reads current robot joints and inserts them under `complementary_data.raw_joint_positions` for FK/IK steps to use.
+- EEReferenceAndDelta: latches a reference EE pose on enable and combines it with target deltas to produce an absolute desired EE pose each frame. When disabled, it keeps sending the last commanded pose.
+- EEBoundsAndSafety: clamps the EE pose to a workspace and rate‑limits jumps for safety. Also declares `action.ee.*` features.
+- InverseKinematicsEEToJoints: turns an EE pose into joint positions with IK. `initial_guess_current_joints=True` is recommended for closed‑loop control; set `False` for open‑loop replay for stability.
+- GripperVelocityToJoint: integrates a velocity‑like gripper input into an absolute gripper position using the current measured state.
+- ForwardKinematicsJointsToEE: computes `observation.state.ee.*` from observed joints for logging and training on EE state.
+
+### Troubleshooting
+
+- iOS not discovered: ensure HEBI Mobile I/O is open and your laptop/phone are on the same network.
+- Android URL not reachable: check local you used `https` instead of `http`, use the exact IP printed by the script and allow your browser to enter and ignore the certificate issue.
+- Motion feels inverted: adjust the sign flips in `MapPhoneActionToRobotAction` or swap axes to match your setup.

docs/source/porting_datasets_v3.mdx

@@ -1,321 +0,0 @@
-# Porting Large Datasets to LeRobot Dataset v3.0
-
-This tutorial explains how to port large-scale robotic datasets to the LeRobot Dataset v3.0 format. We'll use the **DROID 1.0.1** dataset as our primary example, which demonstrates handling multi-terabyte datasets with thousands of shards across SLURM clusters.
-
-## File Organization: v2.1 vs v3.0
-
-Dataset v3.0 fundamentally changes how data is organized and stored:
-
-**v2.1 Structure (Episode-based)**:
-
-```
-dataset/
-├── data/chunk-000/episode_000000.parquet
-├── data/chunk-000/episode_000001.parquet
-├── videos/chunk-000/camera/episode_000000.mp4
-└── meta/episodes.jsonl
-```
-
-**v3.0 Structure (File-based)**:
-
-```
-dataset/
-├── data/chunk-000/file-000.parquet        # Multiple episodes per file
-├── videos/camera/chunk-000/file-000.mp4   # Consolidated video chunks
-└── meta/episodes/chunk-000/file-000.parquet  # Structured metadata
-```
-
-This transition from individual episode files to file-based chunks dramatically improves performance and reduces storage overhead.
-
-## What's New in Dataset v3.0
-
-Dataset v3.0 introduces significant improvements for handling large datasets:
-
-### 🏗️ **Enhanced File Organization**
-
-- **File-based structure**: Episodes are now grouped into chunked files rather than individual episode files
-- **Configurable file sizes**: for data and video files
-- **Improved storage efficiency**: Better compression and reduced overhead
-
-### 📊 **Modern Metadata Management**
-
-- **Parquet-based metadata**: Replaced JSON Lines with efficient parquet format
-- **Structured episode access**: Direct pandas DataFrame access via `dataset.meta.episodes`
-- **Per-episode statistics**: Enhanced statistics tracking at episode level
-
-### 🚀 **Performance Enhancements**
-
-- **Memory-mapped access**: Improved RAM usage through PyArrow memory mapping
-- **Faster loading**: Significantly reduced dataset initialization time
-- **Better scalability**: Designed for datasets with millions of episodes
-
-## Prerequisites
-
-Before porting large datasets, ensure you have:
-
-- **LeRobot installed** with v3.0 support. Follow our [Installation Guide](./installation).
-- **Sufficient storage**: Raw datasets can be very large (e.g., DROID requires 2TB)
-- **Cluster access** (recommended for large datasets): SLURM or similar job scheduler
-- **Dataset-specific dependencies**: For DROID, you'll need TensorFlow Dataset utilities
-
-## Understanding the DROID Dataset
-
-[DROID 1.0.1](https://droid-dataset.github.io/droid/the-droid-dataset) is an excellent example of a large-scale robotic dataset:
-
-- **Size**: 1.7TB (RLDS format), 8.7TB (raw data)
-- **Structure**: 2048 pre-defined TensorFlow dataset shards
-- **Content**: 76,000+ robot manipulation trajectories from Franka Emika Panda robots
-- **Scope**: Real-world manipulation tasks across multiple environments and objects
-- **Format**: Originally in TensorFlow Records/RLDS format, requiring conversion to LeRobot format
-- **Hosting**: Google Cloud Storage with public access via `gsutil`
-
-The dataset contains diverse manipulation demonstrations with:
-
-- Multiple camera views (wrist camera, exterior cameras)
-- Natural language task descriptions
-- Robot proprioceptive state and actions
-- Success/failure annotations
-
-### DROID Features Schema
-
-```python
-DROID_FEATURES = {
-    # Episode markers
-    "is_first": {"dtype": "bool", "shape": (1,)},
-    "is_last": {"dtype": "bool", "shape": (1,)},
-    "is_terminal": {"dtype": "bool", "shape": (1,)},
-
-    # Language instructions
-    "language_instruction": {"dtype": "string", "shape": (1,)},
-    "language_instruction_2": {"dtype": "string", "shape": (1,)},
-    "language_instruction_3": {"dtype": "string", "shape": (1,)},
-
-    # Robot state
-    "observation.state.gripper_position": {"dtype": "float32", "shape": (1,)},
-    "observation.state.cartesian_position": {"dtype": "float32", "shape": (6,)},
-    "observation.state.joint_position": {"dtype": "float32", "shape": (7,)},
-
-    # Camera observations
-    "observation.images.wrist_left": {"dtype": "image"},
-    "observation.images.exterior_1_left": {"dtype": "image"},
-    "observation.images.exterior_2_left": {"dtype": "image"},
-
-    # Actions
-    "action.gripper_position": {"dtype": "float32", "shape": (1,)},
-    "action.cartesian_position": {"dtype": "float32", "shape": (6,)},
-    "action.joint_position": {"dtype": "float32", "shape": (7,)},
-
-    # Standard LeRobot format
-    "observation.state": {"dtype": "float32", "shape": (8,)},  # joints + gripper
-    "action": {"dtype": "float32", "shape": (8,)},  # joints + gripper
-}
-```
-
-## Approach 1: Single Computer Porting
-
-### Step 1: Install Dependencies
-
-For DROID specifically:
-
-```bash
-pip install tensorflow
-pip install tensorflow_datasets
-```
-
-For other datasets, install the appropriate readers for your source format.
-
-### Step 2: Download Raw Data
-
-Download DROID from Google Cloud Storage using `gsutil`:
-
-```bash
-# Install Google Cloud SDK if not already installed
-# https://cloud.google.com/sdk/docs/install
-
-# Download the full RLDS dataset (1.7TB)
-gsutil -m cp -r gs://gresearch/robotics/droid/1.0.1 /your/data/
-
-# Or download just the 100-episode sample (2GB) for testing
-gsutil -m cp -r gs://gresearch/robotics/droid_100 /your/data/
-```
-
-> [!WARNING]
-> Large datasets require substantial time and storage:
->
-> - **Full DROID (1.7TB)**: Several days to download depending on bandwidth
-> - **Processing time**: 7+ days for local porting of full dataset
-> - **Upload time**: 3+ days to push to Hugging Face Hub
-> - **Local storage**: ~400GB for processed LeRobot format
-
-### Step 3: Port the Dataset
-
-```bash
-python examples/port_datasets/port_droid.py \
-    --raw-dir /your/data/droid/1.0.1 \
-    --repo-id your_id/droid_1.0.1 \
-    --push-to-hub
-```
-
-### Development and Testing
-
-For development, you can port a single shard:
-
-```bash
-python examples/port_datasets/port_droid.py \
-    --raw-dir /your/data/droid/1.0.1 \
-    --repo-id your_id/droid_1.0.1_test \
-    --num-shards 2048 \
-    --shard-index 0
-```
-
-This approach works for smaller datasets or testing, but large datasets require cluster computing.
-
-## Approach 2: SLURM Cluster Porting (Recommended)
-
-For large datasets like DROID, parallel processing across multiple nodes dramatically reduces processing time.
-
-### Step 1: Install Cluster Dependencies
-
-```bash
-pip install datatrove  # Hugging Face's distributed processing library
-```
-
-### Step 2: Configure Your SLURM Environment
-
-Find your partition information:
-
-```bash
-sinfo --format="%R"  # List available partitions
-sinfo -N -p your_partition -h -o "%N cpus=%c mem=%m"  # Check resources
-```
-
-Choose a **CPU partition** - no GPU needed for dataset porting.
-
-### Step 3: Launch Parallel Porting Jobs
-
-```bash
-python examples/port_datasets/slurm_port_shards.py \
-    --raw-dir /your/data/droid/1.0.1 \
-    --repo-id your_id/droid_1.0.1 \
-    --logs-dir /your/logs \
-    --job-name port_droid \
-    --partition your_partition \
-    --workers 2048 \
-    --cpus-per-task 8 \
-    --mem-per-cpu 1950M
-```
-
-#### Parameter Guidelines
-
-- **`--workers`**: Number of parallel jobs (max 2048 for DROID's shard count)
-- **`--cpus-per-task`**: 8 CPUs recommended for frame encoding parallelization
-- **`--mem-per-cpu`**: ~16GB total RAM (8×1950M) for loading raw frames
-
-> [!TIP]
-> Start with fewer workers (e.g., 100) to test your cluster configuration before launching thousands of jobs.
-
-### Step 4: Monitor Progress
-
-Check running jobs:
-
-```bash
-squeue -u $USER
-```
-
-Monitor overall progress:
-
-```bash
-jobs_status /your/logs
-```
-
-Inspect individual job logs:
-
-```bash
-less /your/logs/port_droid/slurm_jobs/JOB_ID_WORKER_ID.out
-```
-
-Debug failed jobs:
-
-```bash
-failed_logs /your/logs/port_droid
-```
-
-### Step 5: Aggregate Shards
-
-Once all porting jobs complete:
-
-```bash
-python examples/port_datasets/slurm_aggregate_shards.py \
-    --repo-id your_id/droid_1.0.1 \
-    --logs-dir /your/logs \
-    --job-name aggr_droid \
-    --partition your_partition \
-    --workers 2048 \
-    --cpus-per-task 8 \
-    --mem-per-cpu 1950M
-```
-
-### Step 6: Upload to Hub
-
-```bash
-python examples/port_datasets/slurm_upload.py \
-    --repo-id your_id/droid_1.0.1 \
-    --logs-dir /your/logs \
-    --job-name upload_droid \
-    --partition your_partition \
-    --workers 50 \
-    --cpus-per-task 4 \
-    --mem-per-cpu 1950M
-```
-
-> [!NOTE]
-> Upload uses fewer workers (50) since it's network-bound rather than compute-bound.
-
-## Dataset v3.0 File Structure
-
-Your completed dataset will have this modern structure:
-
-```
-dataset/
-├── meta/
-│   ├── episodes/
-│   │   └── chunk-000/
-│   │       └── file-000.parquet    # Episode metadata
-│   ├── tasks.parquet               # Task definitions
-│   ├── stats.json                  # Aggregated statistics
-│   └── info.json                   # Dataset information
-├── data/
-│   └── chunk-000/
-│       └── file-000.parquet        # Consolidated episode data
-└── videos/
-    └── camera_key/
-        └── chunk-000/
-            └── file-000.mp4        # Consolidated video files
-```
-
-This replaces the old episode-per-file structure with efficient, optimally-sized chunks.
-
-## Migrating from Dataset v2.1
-
-If you have existing datasets in v2.1 format, use the migration tool:
-
-```bash
-python src/lerobot/datasets/v30/convert_dataset_v21_to_v30.py \
-    --repo-id your_id/existing_dataset
-```
-
-This automatically:
-
-- Converts file structure to v3.0 format
-- Migrates metadata from JSON Lines to parquet
-- Aggregates statistics and creates per-episode stats
-- Updates version information
-
-## Performance Benefits
-
-Dataset v3.0 provides significant improvements for large datasets:
-
-- **Faster loading**: 3-5x reduction in initialization time
-- **Memory efficiency**: Better RAM usage through memory mapping
-- **Scalable processing**: Handles millions of episodes efficiently
-- **Storage optimization**: Reduced file count and improved compression

docs/source/processors_robots_teleop.mdx

@@ -0,0 +1,148 @@
+# Processors for Robots and Teleoperators
+
+This guide shows how to build and modify processing pipelines that connect teleoperators (e.g., phone) to robots and datasets. Pipelines standardize conversions between different action/observation spaces so you can swap teleops and robots without rewriting glue code.
+
+We use the Phone to SO‑100 follower examples for concreteness, but the same patterns apply to other robots.
+
+**What you'll learn**
+
+- Absolute vs. relative EE control: What each means, trade‑offs, and how to choose for your task.
+- Three-pipeline pattern: How to map teleop actions → dataset actions → robot commands, and robot observations → dataset observations.
+- Adapters (`to_transition` / `to_output`): How these convert raw dicts to `EnvTransition` and back to reduce boilerplate.
+- Dataset feature contracts: How steps declare features via `transform_features(...)`, and how to aggregate/merge them for recording.
+- Choosing a representation: When to store joints, absolute EE poses, or relative EE deltas—and how that affects training.
+- Pipeline customization guidance: How to swap robots/URDFs safely and tune bounds, step sizes, and options like IK initialization.
+
+### Absolute vs relative EE control
+
+The examples in this guide use absolute end effector (EE) poses because they are easy to reason about. In practice, relative EE deltas or joint position are often preferred as learning features.
+
+You can choose what you save and learn from the teleop and robot action spaces, joints, absolute EE, or relative EE by using/implementing the right steps (and `transform_features()`) in your pipelines.
+
+## Three pipelines
+
+We often compose three pipelines. Depending on your setup, some can be empty if action and observation spaces already match.
+Each of these pipelines handle different conversions between different action and observation spaces. Below is a quick explanation of each pipeline.
+
+1. Pipeline 1: Teleop action space → dataset action space (phone pose → EE targets)
+2. Pipeline 2: Dataset action space → robot command space (EE targets → joints)
+3. Pipeline 3: Robot observation space → dataset observation space (joints → EE pose)
+
+Below is an example of the three pipelines that we use in the phone to SO-100 follower examples:
+
+```69:90:examples/phone_so100_record.py
+phone_to_robot_ee_pose = RobotProcessor(  # teleop -> dataset action
+    steps=[MapPhoneActionToRobotAction(platform=teleop_config.phone_os),
+           AddRobotObservationAsComplimentaryData(robot=robot),
+            EEReferenceAndDelta(kinematics=kinematics_solver,
+                               end_effector_step_sizes={"x": 0.5, "y": 0.5, "z": 0.5},
+                               motor_names=list(robot.bus.motors.keys())),
+           EEBoundsAndSafety(end_effector_bounds={"min": [-1, -1, -1], "max": [1, 1, 1]},
+                             max_ee_step_m=0.20, max_ee_twist_step_rad=0.50)],
+    to_transition=to_transition_teleop_action,
+    to_output=lambda tr: tr,
+)
+
+robot_ee_to_joints = RobotProcessor(      # dataset action -> robot
+    steps=[InverseKinematicsEEToJoints(kinematics=kinematics_solver,
+                                       motor_names=list(robot.bus.motors.keys()),
+                                       initial_guess_current_joints=True),
+           GripperVelocityToJoint(motor_names=list(robot.bus.motors.keys()), speed_factor=20.0)],
+    to_transition=lambda tr: tr,
+    to_output=to_output_robot_action,
+)
+
+robot_joints_to_ee_pose = RobotProcessor( # robot obs -> dataset obs
+    steps=[ForwardKinematicsJointsToEE(kinematics=kinematics_solver,
+                                       motor_names=list(robot.bus.motors.keys()))],
+    to_transition=to_transition_robot_observation,
+    to_output=lambda tr: tr,
+)
+```
+
+## Why to_transition / to_output
+
+To convert from robot/teleoperator to pipeline and back, we use the `to_transition` and `to_output` pipeline adapters.
+They standardize conversions to reduce boilerplate code, and form the bridge between the robot and teleoperators raw dicts and the pipeline’s `EnvTransition` format.
+In the phone to SO-100 follower examples we use the following adapters:
+
+- `to_transition_teleop_action`: transforms the teleop action dict to a pipeline transition (puts keys under `action.*`, converts scalars/arrays to tensors, keeps objects like `Rotation` intact)
+- `to_output_robot_action`: transforms the pipeline transition to a robot action dict (extracts keys ending with `.pos`/`.vel` and strips `action.` prefix)
+- `to_transition_robot_observation`: transforms the robot observation dict to a pipeline transition (splits state vs images; stores state under `observation.state.*` and images under `observation.images.*`)
+
+See `src/lerobot/processor/converters.py` for more details.
+
+## Dataset feature contracts
+
+Dataset features are the keys saved in the dataset. Each step can declare what its dataset features are via `transform_features(...)`. We can then aggregate features per pipeline with `aggregate_pipeline_dataset_features()` and merge multiple groups with `merge_features(...)`.
+
+Below is and example of how we declare features with the `transform_features` method in the phone to SO-100 follower examples:
+
+```203:211:src/lerobot/robots/so100_follower/robot_kinematic_processor.py
+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
+    return features
+```
+
+Tip: declare features at the last step that produces them (e.g., `EEBoundsAndSafety` declares `action.ee.*`, `ForwardKinematicsJointsToEE` declares `observation.state.ee.*`).
+
+Below is an example of how we aggregate and merge features in the phone to SO-100 follower examples:
+
+```121:145:examples/phone_so100_record.py
+action_ee = aggregate_pipeline_dataset_features(
+    pipeline=phone_to_robot_ee_pose,
+    initial_features=phone.action_features,
+    use_videos=True,
+    patterns=["action.ee"],
+)
+
+gripper = aggregate_pipeline_dataset_features(
+    pipeline=robot_ee_to_joints,
+    initial_features={},
+    use_videos=True,
+    patterns=["action.gripper.pos", "observation.state.gripper.pos"],
+)
+
+observation_ee = aggregate_pipeline_dataset_features(
+    pipeline=robot_joints_to_ee_pose,
+    initial_features=robot.observation_features,
+    use_videos=True,
+    patterns=["observation.state.ee"],
+)
+
+dataset_features = merge_features(action_ee, gripper, observation_ee)
+```
+
+How it works:
+
+- `aggregate_pipeline_dataset_features(...)`: applies `transform_features` across the pipeline and filters by patterns (images included when `use_videos=True`).
+- `merge_features(...)`: combine multiple feature dicts.
+- Recording uses `to_dataset_frame(...)` to build frames consistent with `dataset.features` before we call `add_frame(...)` to add the frame to the dataset.
+
+## Guidance when customizing robot pipelines
+
+You can store any of the following features as your action/observation space:
+
+- Joint positions
+- Absolute EE poses
+- Relative EE deltas
+- Other features: joint velocity, etc.
+
+Pick what you want to use for your policy action and observation space and configure/modify the pipelines and steps accordingly.
+
+### Different robots
+
+- Swap `RobotKinematics` URDF and `motor_names`. Ensure `target_frame_name` points to your gripper/wrist.
+
+### Safety first
+
+- When changing pipelines, start with tight bounds, implement safety steps when working with real robots.
+- Its advised to start with simulation first and then move to real robots.
+
+Hope this guide helps you get started with customizing your robot pipelines, If you run into any issues at any point, jump into our [Discord community](https://discord.com/invite/s3KuuzsPFb) for support.

docs/source/reachy2.mdx

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

docs/source/smolvla.mdx

@@ -54,7 +54,7 @@ If you don't have a gpu device, you can train using our notebook on [![Google Co
 Pass your dataset to the training script using `--dataset.repo_id`. If you want to test your installation, run the following command where we use one of the datasets we collected for the [SmolVLA Paper](https://huggingface.co/papers/2506.01844).
 
 ```bash
-cd lerobot && lerobot-train \
+cd lerobot && python -m lerobot.scripts.train \
   --policy.path=lerobot/smolvla_base \
   --dataset.repo_id=${HF_USER}/mydataset \
   --batch_size=64 \
@@ -73,7 +73,7 @@ cd lerobot && lerobot-train \
 Fine-tuning is an art. For a complete overview of the options for finetuning, run
 
 ```bash
-lerobot-train --help
+python -m lerobot.scripts.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
-lerobot-record \
+python -m 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
-lerobot-find-port
+python -m 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
-lerobot-setup-motors \
+python -m 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
-lerobot-setup-motors \
+python -m 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
-lerobot-calibrate \
+python -m 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
-lerobot-calibrate \
+python -m 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
-lerobot-find-port
+python -m 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
-lerobot-setup-motors \
+python -m 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
-lerobot-setup-motors \
+python -m 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
-lerobot-calibrate \
+python -m 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
-lerobot-calibrate \
+python -m 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/1_load_lerobot_dataset.py

@@ -92,11 +92,11 @@ print(dataset.hf_dataset)
 # LeRobot datasets also subclasses PyTorch datasets so you can do everything you know and love from working
 # with the latter, like iterating through the dataset.
 # The __getitem__ iterates over the frames of the dataset. Since our datasets are also structured by
-# episodes, you can access the frame indices of any episode using dataset.meta.episodes. Here, we access
+# episodes, you can access the frame indices of any episode using the episode_data_index. Here, we access
 # frame indices associated to the first episode:
 episode_index = 0
-from_idx = dataset.meta.episodes["dataset_from_index"][episode_index]
-to_idx = dataset.meta.episodes["dataset_to_index"][episode_index]
+from_idx = dataset.episode_data_index["from"][episode_index].item()
+to_idx = dataset.episode_data_index["to"][episode_index].item()
 
 # Then we grab all the image frames from the first camera:
 camera_key = dataset.meta.camera_keys[0]

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
-lerobot-train \
+python -m lerobot.scripts.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
-lerobot-train \
+python -m lerobot.scripts.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
-lerobot-train \
+python -m lerobot.scripts.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
-lerobot-train \
+python -m lerobot.scripts.train \
     --config_path=outputs/train/act_aloha_transfer/checkpoints/last/pretrained_model/ \
     --output_dir=outputs/train/act_aloha_transfer_2
 ```
@@ -137,7 +137,7 @@ lerobot-train \
 Similarly to Hydra, we can still override some parameters in the CLI if we want to, e.g.:
 
 ```bash
-lerobot-train \
+python -m lerobot.scripts.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 @@ lerobot-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
-lerobot-train --config_path=lerobot/diffusion_pusht
+python -m lerobot.scripts.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
-lerobot-train \
+python -m lerobot.scripts.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
-lerobot-train \
+python -m lerobot.scripts.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
-lerobot-train \
+python -m lerobot.scripts.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
-lerobot-train \
+python -m lerobot.scripts.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
-lerobot-train \
+python -m lerobot.scripts.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 @@ lerobot-train \
 #### Train a policy from scratch - config file + CLI
 
 ```bash
-lerobot-train \
+python -m lerobot.scripts.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 @@ lerobot-train \
 #### Resume/continue a training run
 
 ```bash
-lerobot-train \
+python -m lerobot.scripts.train \
     --config_path=checkpoint/pretrained_model/ \
     --resume=true \
     --steps=200000  # <- you can change some training parameters
@@ -296,7 +296,7 @@ lerobot-train \
 #### Fine-tuning
 
 ```bash
-lerobot-train \
+python -m lerobot.scripts.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/5_train_with_streaming.py

@@ -1,116 +0,0 @@
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-"""This script demonstrates how to train a Diffusion Policy on the PushT environment,
-using a dataset processed in streaming mode.
-
-Once you have trained a model with this script, you can try to evaluate it on
-examples/2_evaluate_pretrained_policy.py
-"""
-
-from pathlib import Path
-
-import torch
-
-from lerobot.configs.types import FeatureType
-from lerobot.constants import ACTION
-from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
-from lerobot.datasets.streaming_dataset import StreamingLeRobotDataset
-from lerobot.datasets.utils import dataset_to_policy_features
-from lerobot.policies.act.configuration_act import ACTConfig
-from lerobot.policies.act.modeling_act import ACTPolicy
-
-
-def main():
-    # Create a directory to store the training checkpoint.
-    output_directory = Path("outputs/train/example_streaming_dataset")
-    output_directory.mkdir(parents=True, exist_ok=True)
-
-    # Selects the "best" device available
-    device = (
-        torch.device("cuda")
-        if torch.cuda.is_available()
-        else torch.device("mps")
-        if torch.backends.mps.is_available()
-        else torch.device("cpu")
-    )
-    print(f"Using device: {device}")
-
-    training_steps = 10
-    log_freq = 1
-
-    dataset_id = (
-        "aractingi/droid_1.0.1"  # 26M frames! Would require 4TB of disk space if installed locally (:
-    )
-    dataset_metadata = LeRobotDatasetMetadata(dataset_id)
-    features = dataset_to_policy_features(dataset_metadata.features)
-    output_features = {key: ft for key, ft in features.items() if ft.type is FeatureType.ACTION}
-    input_features = {key: ft for key, ft in features.items() if key not in output_features}
-
-    # We can now instantiate our policy with this config and the dataset stats.
-    cfg = ACTConfig(input_features=input_features, output_features=output_features)
-    policy = ACTPolicy(cfg, dataset_stats=dataset_metadata.stats)
-    policy.train()
-    policy.to(device)
-
-    # Delta timestamps are used to (1) augment frames used during training and (2) supervise the policy.
-    # Here, we use delta-timestamps to only provide ground truth actions for supervision
-    delta_timestamps = {
-        ACTION: [t / dataset_metadata.fps for t in range(cfg.n_action_steps)],
-    }
-
-    # Instantiating the training dataset in streaming mode allows to not consume up memory as the data is fetched
-    # iteratively rather than being load into memory all at once. Retrieved frames are shuffled across epochs
-    dataset = StreamingLeRobotDataset(dataset_id, delta_timestamps=delta_timestamps, tolerance_s=1e-3)
-
-    optimizer = torch.optim.Adam(policy.parameters(), lr=1e-4)
-    dataloader = torch.utils.data.DataLoader(
-        dataset,
-        num_workers=4,
-        batch_size=16,
-        pin_memory=device.type != "cpu",
-        drop_last=True,
-        prefetch_factor=2,  # loads batches with multiprocessing while policy trains
-    )
-
-    # Run training loop.
-    step = 0
-    done = False
-    while not done:
-        for batch in dataloader:
-            batch = {
-                k: (v.type(torch.float32) if isinstance(v, torch.Tensor) and v.dtype != torch.bool else v)
-                for k, v in batch.items()
-            }
-            batch = {k: (v.to(device) if isinstance(v, torch.Tensor) else v) for k, v in batch.items()}
-
-            # batch = {k: (v.to(device) if isinstance(v, torch.Tensor) else v) for k, v in batch.items()}
-            loss, _ = policy.forward(batch)
-            loss.backward()
-            optimizer.step()
-            optimizer.zero_grad()
-
-            if step % log_freq == 0:
-                print(f"step: {step} loss: {loss.item():.3f}")
-            step += 1
-            if step >= training_steps:
-                done = True
-                break
-
-    # Save a policy checkpoint.
-    policy.save_pretrained(output_directory)
-
-
-if __name__ == "__main__":
-    main()

examples/backward_compatibility/replay.py

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

examples/lekiwi/evaluate.py

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

examples/lekiwi/teleoperate.py

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

examples/phone_so100_eval.py

@@ -0,0 +1,158 @@
+# !/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features
+from lerobot.datasets.utils import merge_features
+from lerobot.model.kinematics import RobotKinematics
+from lerobot.policies.act.modeling_act import ACTPolicy
+from lerobot.policies.factory import make_processor
+from lerobot.processor.converters import (
+    to_output_robot_action,
+    to_transition_robot_observation,
+)
+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 (
+    AddRobotObservationAsComplimentaryData,
+    ForwardKinematicsJointsToEE,
+    InverseKinematicsEEToJoints,
+)
+from lerobot.robots.so100_follower.so100_follower import SO100Follower
+from lerobot.utils.control_utils import init_keyboard_listener
+from lerobot.utils.utils import log_say
+from lerobot.utils.visualization_utils import _init_rerun
+
+NUM_EPISODES = 5
+FPS = 30
+EPISODE_TIME_SEC = 60
+TASK_DESCRIPTION = "My task description"
+HF_MODEL_ID = "<hf_username>/<model_repo_id>"
+HF_DATASET_ID = "<hf_username>/<dataset_repo_id>"
+
+# Initialize the robot with degrees
+camera_config = {"front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS)}
+robot_config = SO100FollowerConfig(
+    port="/dev/tty.usbmodem58760434471",
+    id="my_awesome_follower_arm",
+    cameras=camera_config,
+    use_degrees=True,
+)
+
+# Initialize the robot
+robot = SO100Follower(robot_config)
+
+# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+kinematics_solver = RobotKinematics(
+    urdf_path="./src/lerobot/teleoperators/sim/so101_new_calib.urdf",
+    target_frame_name="gripper_frame_link",
+    joint_names=list(robot.bus.motors.keys()),
+)
+
+# Build pipeline to convert ee pose action to joint action
+robot_ee_to_joints = RobotProcessor(
+    steps=[
+        AddRobotObservationAsComplimentaryData(robot=robot),
+        InverseKinematicsEEToJoints(
+            kinematics=kinematics_solver,
+            motor_names=list(robot.bus.motors.keys()),
+            initial_guess_current_joints=True,
+        ),
+    ],
+    to_transition=lambda tr: tr,
+    to_output=to_output_robot_action,
+)
+
+# Build pipeline to convert joint observation to ee pose observation
+robot_joints_to_ee_pose = RobotProcessor(
+    steps=[
+        ForwardKinematicsJointsToEE(kinematics=kinematics_solver, motor_names=list(robot.bus.motors.keys()))
+    ],
+    to_transition=to_transition_robot_observation,
+    to_output=lambda tr: tr,
+)
+
+# Build dataset action and gripper features
+action_ee_and_gripper = aggregate_pipeline_dataset_features(
+    pipeline=robot_ee_to_joints,
+    initial_features={},
+    use_videos=True,
+    patterns=["action.ee", "action.gripper.pos", "observation.state.gripper.pos"],
+)  # Get all ee action features + gripper pos action features
+
+# Build dataset observation features
+obs_ee = aggregate_pipeline_dataset_features(
+    pipeline=robot_joints_to_ee_pose,
+    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)
+
+print("All dataset features: ", dataset_features)
+
+# Create the dataset
+dataset = LeRobotDataset.create(
+    repo_id=HF_DATASET_ID,
+    fps=FPS,
+    features=dataset_features,
+    robot_type=robot.name,
+    use_videos=True,
+    image_writer_threads=4,
+)
+
+# Initialize the keyboard listener and rerun visualization
+_, events = init_keyboard_listener()
+_init_rerun(session_name="recording_phone")
+
+# Connect the robot and teleoperator
+robot.connect()
+
+episode_idx = 0
+
+policy = ACTPolicy.from_pretrained(HF_MODEL_ID)
+preprocessor, postprocessor = make_processor(
+    policy_cfg=policy,
+    pretrained_path=HF_MODEL_ID,
+    dataset_stats=dataset.meta.stats,
+)
+
+for episode_idx in range(NUM_EPISODES):
+    log_say(f"Running inference, recording eval episode {episode_idx + 1} of {NUM_EPISODES}")
+
+    record_loop(
+        robot=robot,
+        events=events,
+        fps=FPS,
+        policy=policy,
+        preprocessor=preprocessor,
+        postprocessor=postprocessor,
+        dataset=dataset,
+        control_time_s=EPISODE_TIME_SEC,
+        single_task=TASK_DESCRIPTION,
+        display_data=True,
+        robot_action_processor=robot_ee_to_joints,
+        robot_observation_processor=robot_joints_to_ee_pose,
+    )
+    dataset.save_episode()
+
+# Clean up
+log_say("Stop recording")
+robot.disconnect()
+dataset.push_to_hub()

examples/phone_so100_record.py

@@ -0,0 +1,215 @@
+# !/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features
+from lerobot.datasets.utils import merge_features
+from lerobot.model.kinematics import RobotKinematics
+from lerobot.processor.converters import (
+    to_output_robot_action,
+    to_transition_robot_observation,
+    to_transition_teleop_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 (
+    AddRobotObservationAsComplimentaryData,
+    EEBoundsAndSafety,
+    EEReferenceAndDelta,
+    ForwardKinematicsJointsToEE,
+    GripperVelocityToJoint,
+    InverseKinematicsEEToJoints,
+)
+from lerobot.robots.so100_follower.so100_follower import SO100Follower
+from lerobot.teleoperators.phone.config_phone import PhoneConfig, PhoneOS
+from lerobot.teleoperators.phone.phone import Phone
+from lerobot.teleoperators.phone.phone_processor import MapPhoneActionToRobotAction
+from lerobot.utils.control_utils import init_keyboard_listener
+from lerobot.utils.utils import log_say
+from lerobot.utils.visualization_utils import _init_rerun
+
+NUM_EPISODES = 10
+FPS = 30
+EPISODE_TIME_SEC = 60
+RESET_TIME_SEC = 30
+TASK_DESCRIPTION = "My task description"
+HF_REPO_ID = "<hf_username>/<dataset_repo_id>"
+
+# Initialize the robot and teleoperator
+camera_config = {"front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS)}
+robot_config = SO100FollowerConfig(
+    port="/dev/tty.usbmodem58760434471",
+    id="my_awesome_follower_arm",
+    cameras=camera_config,
+    use_degrees=True,
+)
+teleop_config = PhoneConfig(phone_os=PhoneOS.IOS)  # or PhoneOS.ANDROID
+
+# Initialize the robot and teleoperator
+robot = SO100Follower(robot_config)
+phone = Phone(teleop_config)
+
+# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+kinematics_solver = RobotKinematics(
+    urdf_path="./src/lerobot/teleoperators/sim/so101_new_calib.urdf",
+    target_frame_name="gripper_frame_link",
+    joint_names=list(robot.bus.motors.keys()),
+)
+
+# Build pipeline to convert phone action to ee pose action
+phone_to_robot_ee_pose = RobotProcessor(
+    steps=[
+        MapPhoneActionToRobotAction(platform=teleop_config.phone_os),
+        AddRobotObservationAsComplimentaryData(robot=robot),
+        EEReferenceAndDelta(
+            kinematics=kinematics_solver,
+            end_effector_step_sizes={"x": 0.5, "y": 0.5, "z": 0.5},
+            motor_names=list(robot.bus.motors.keys()),
+        ),
+        EEBoundsAndSafety(
+            end_effector_bounds={"min": [-1.0, -1.0, -1.0], "max": [1.0, 1.0, 1.0]},
+            max_ee_step_m=0.20,
+            max_ee_twist_step_rad=0.50,
+        ),
+    ],
+    to_transition=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()),
+            initial_guess_current_joints=True,
+        ),
+        GripperVelocityToJoint(
+            motor_names=list(robot.bus.motors.keys()),
+            speed_factor=20.0,
+        ),
+    ],
+    to_transition=lambda tr: tr,
+    to_output=to_output_robot_action,
+)
+
+# Build pipeline to convert joint observation to ee pose observation
+robot_joints_to_ee_pose = RobotProcessor(
+    steps=[
+        ForwardKinematicsJointsToEE(kinematics=kinematics_solver, motor_names=list(robot.bus.motors.keys()))
+    ],
+    to_transition=to_transition_robot_observation,
+    to_output=lambda tr: tr,
+)
+
+# Build dataset ee action features
+action_ee = aggregate_pipeline_dataset_features(
+    pipeline=phone_to_robot_ee_pose,
+    initial_features=phone.action_features,
+    use_videos=True,
+    patterns=["action.ee"],
+)
+
+# Get gripper pos action features
+gripper = aggregate_pipeline_dataset_features(
+    pipeline=robot_ee_to_joints,
+    initial_features={},
+    use_videos=True,
+    patterns=["action.gripper.pos", "observation.state.gripper.pos"],
+)
+
+# Build dataset ee observation features
+observation_ee = aggregate_pipeline_dataset_features(
+    pipeline=robot_joints_to_ee_pose,
+    initial_features=robot.observation_features,
+    use_videos=True,
+    patterns=["observation.state.ee"],
+)
+
+dataset_features = merge_features(action_ee, gripper, observation_ee)
+
+print("All dataset features: ", dataset_features)
+
+# Create the dataset
+dataset = LeRobotDataset.create(
+    repo_id=HF_REPO_ID,
+    fps=FPS,
+    features=dataset_features,
+    robot_type=robot.name,
+    use_videos=True,
+    image_writer_threads=4,
+)
+
+# Initialize the keyboard listener and rerun visualization
+_, events = init_keyboard_listener()
+_init_rerun(session_name="recording_phone")
+
+# Connect the robot and teleoperator
+robot.connect()
+phone.connect()
+
+episode_idx = 0
+while episode_idx < NUM_EPISODES and not events["stop_recording"]:
+    log_say(f"Recording episode {episode_idx + 1} of {NUM_EPISODES}")
+
+    record_loop(
+        robot=robot,
+        events=events,
+        fps=FPS,
+        teleop=phone,
+        dataset=dataset,
+        control_time_s=EPISODE_TIME_SEC,
+        single_task=TASK_DESCRIPTION,
+        display_data=True,
+        teleop_action_processor=phone_to_robot_ee_pose,
+        robot_action_processor=robot_ee_to_joints,
+        robot_observation_processor=robot_joints_to_ee_pose,
+    )
+
+    # Reset the environment if not stopping or re-recording
+    if not events["stop_recording"] and (episode_idx < NUM_EPISODES - 1 or events["rerecord_episode"]):
+        log_say("Reset the environment")
+        record_loop(
+            robot=robot,
+            events=events,
+            fps=FPS,
+            teleop=phone,
+            control_time_s=RESET_TIME_SEC,
+            single_task=TASK_DESCRIPTION,
+            display_data=True,
+            teleop_action_processor=phone_to_robot_ee_pose,
+            robot_action_processor=robot_ee_to_joints,
+            robot_observation_processor=robot_joints_to_ee_pose,
+        )
+
+    if events["rerecord_episode"]:
+        log_say("Re-recording episode")
+        events["rerecord_episode"] = False
+        events["exit_early"] = False
+        dataset.clear_episode_buffer()
+        continue
+
+    dataset.save_episode()
+    episode_idx += 1
+
+# Clean up
+log_say("Stop recording")
+robot.disconnect()
+phone.disconnect()
+dataset.push_to_hub()

examples/phone_so100_replay.py

@@ -0,0 +1,106 @@
+# !/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import time
+
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.model.kinematics import RobotKinematics
+from lerobot.processor.converters import to_output_robot_action
+from lerobot.processor.pipeline import RobotProcessor
+from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
+from lerobot.robots.so100_follower.robot_kinematic_processor import (
+    AddRobotObservationAsComplimentaryData,
+    InverseKinematicsEEToJoints,
+)
+from lerobot.robots.so100_follower.so100_follower import SO100Follower
+from lerobot.utils.robot_utils import busy_wait
+from lerobot.utils.utils import log_say
+
+EPISODE_IDX = 0
+HF_REPO_ID = "<hf_username>/<dataset_repo_id>"
+
+robot_config = SO100FollowerConfig(
+    port="/dev/tty.usbmodem58760434471", id="my_awesome_follower_arm", use_degrees=True
+)
+robot = SO100Follower(robot_config)
+robot.connect()
+
+dataset = LeRobotDataset(HF_REPO_ID, episodes=[EPISODE_IDX])
+actions = dataset.hf_dataset.select_columns("action")
+
+# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+kinematics_solver = RobotKinematics(
+    urdf_path="./src/lerobot/teleoperators/sim/so101_new_calib.urdf",
+    target_frame_name="gripper_frame_link",
+    joint_names=list(robot.bus.motors.keys()),
+)
+
+
+# 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(
+    steps=[
+        AddRobotObservationAsComplimentaryData(robot=robot),
+        InverseKinematicsEEToJoints(
+            kinematics=kinematics_solver,
+            motor_names=list(robot.bus.motors.keys()),
+            initial_guess_current_joints=False,  # Because replay is open loop
+        ),
+    ],
+    to_transition=action_to_transition,
+    to_output=to_output_robot_action,
+)
+
+robot_ee_to_joints.reset()
+
+log_say(f"Replaying episode {EPISODE_IDX}")
+for idx in range(dataset.num_frames):
+    t0 = time.perf_counter()
+
+    ee_action = {
+        name: float(actions[idx]["action"][i]) for i, name in enumerate(dataset.features["action"]["names"])
+    }
+
+    joint_action = robot_ee_to_joints(ee_action)
+    action_sent = robot.send_action(joint_action)
+
+    busy_wait(1.0 / dataset.fps - (time.perf_counter() - t0))
+
+robot.disconnect()

examples/phone_so100_teleop.py

@@ -0,0 +1,109 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specif
+
+import time
+
+from lerobot.model.kinematics import RobotKinematics
+from lerobot.processor import RobotProcessor
+from lerobot.processor.converters import to_output_robot_action, to_transition_teleop_action
+from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
+from lerobot.robots.so100_follower.robot_kinematic_processor import (
+    AddRobotObservationAsComplimentaryData,
+    EEBoundsAndSafety,
+    EEReferenceAndDelta,
+    GripperVelocityToJoint,
+    InverseKinematicsEEToJoints,
+)
+from lerobot.robots.so100_follower.so100_follower import SO100Follower
+from lerobot.teleoperators.phone.config_phone import PhoneConfig, PhoneOS
+from lerobot.teleoperators.phone.phone import Phone
+from lerobot.teleoperators.phone.phone_processor import MapPhoneActionToRobotAction
+
+# Initialize the robot and teleoperator
+robot_config = SO100FollowerConfig(
+    port="/dev/tty.usbmodem58760434471", id="my_awesome_follower_arm", use_degrees=True
+)
+teleop_config = PhoneConfig(phone_os=PhoneOS.IOS)  # or PhoneOS.ANDROID
+
+# Initialize the robot and teleoperator
+robot = SO100Follower(robot_config)
+teleop_device = Phone(teleop_config)
+
+# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+kinematics_solver = RobotKinematics(
+    urdf_path="./src/lerobot/teleoperators/sim/so101_new_calib.urdf",
+    target_frame_name="gripper_frame_link",
+    joint_names=list(robot.bus.motors.keys()),
+)
+
+# Build pipeline to convert phone action to ee pose action
+phone_to_robot_ee_pose = RobotProcessor(
+    steps=[
+        MapPhoneActionToRobotAction(platform=teleop_config.phone_os),
+        AddRobotObservationAsComplimentaryData(robot=robot),
+        EEReferenceAndDelta(
+            kinematics=kinematics_solver,
+            end_effector_step_sizes={"x": 0.5, "y": 0.5, "z": 0.5},
+            motor_names=list(robot.bus.motors.keys()),
+        ),
+        EEBoundsAndSafety(
+            end_effector_bounds={"min": [-1.0, -1.0, -1.0], "max": [1.0, 1.0, 1.0]},
+            max_ee_step_m=0.10,
+            max_ee_twist_step_rad=0.50,
+        ),
+    ],
+    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()),
+        ),
+        GripperVelocityToJoint(
+            motor_names=list(robot.bus.motors.keys()),
+            speed_factor=20.0,
+        ),
+    ],
+    to_transition=lambda tr: tr,
+    to_output=to_output_robot_action,
+)
+
+robot.connect()
+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)
+
+    if joint_action:
+        robot.send_action(joint_action)
+
+    time.sleep(0.01)

examples/port_datasets/display_error_files.py

@@ -1,85 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-import argparse
-import json
-from pathlib import Path
-
-
-def find_missing_workers(completions_dir, world_size):
-    """Find workers that are not completed and returns their indices."""
-    full = list(range(world_size))
-
-    completed = []
-    for path in completions_dir.glob("*"):
-        if path.name in [".", ".."]:
-            continue
-        index = path.name.lstrip("0")
-        index = 0 if index == "" else int(index)
-        completed.append(index)
-
-    missing_workers = set(full) - set(completed)
-    return missing_workers
-
-
-def find_output_files(slurm_dir, worker_indices):
-    """Find output files associated to worker indices, and return tuples
-    of (worker index, output file path)
-    """
-    out_files = []
-    for path in slurm_dir.glob("*.out"):
-        _, worker_id = path.name.replace(".out", "").split("_")
-        worker_id = int(worker_id)
-        if worker_id in worker_indices:
-            out_files.append((worker_id, path))
-    return out_files
-
-
-def display_error_files(logs_dir, job_name):
-    executor_path = Path(logs_dir) / job_name / "executor.json"
-    completions_dir = Path(logs_dir) / job_name / "completions"
-
-    with open(executor_path) as f:
-        executor = json.load(f)
-
-    missing_workers = find_missing_workers(completions_dir, executor["world_size"])
-
-    for missing in sorted(missing_workers)[::-1]:
-        print(missing)
-
-
-def main():
-    parser = argparse.ArgumentParser()
-
-    parser.add_argument(
-        "--logs-dir",
-        type=str,
-        help="Path to logs directory for `datatrove`.",
-    )
-    parser.add_argument(
-        "--job-name",
-        type=str,
-        default="port_droid",
-        help="Job name used in slurm, and name of the directory created inside the provided logs directory.",
-    )
-
-    args = parser.parse_args()
-
-    display_error_files(**vars(args))
-
-
-if __name__ == "__main__":
-    main()

examples/port_datasets/port_droid.py

@@ -1,430 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-import argparse
-import logging
-import time
-from pathlib import Path
-
-import numpy as np
-import tensorflow_datasets as tfds
-
-from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
-from lerobot.utils.utils import get_elapsed_time_in_days_hours_minutes_seconds
-
-DROID_SHARDS = 2048
-DROID_FPS = 15
-DROID_ROBOT_TYPE = "Franka"
-
-# Dataset schema slightly adapted from: https://droid-dataset.github.io/droid/the-droid-dataset.html#-dataset-schema
-DROID_FEATURES = {
-    # true on first step of the episode
-    "is_first": {
-        "dtype": "bool",
-        "shape": (1,),
-        "names": None,
-    },
-    # true on last step of the episode
-    "is_last": {
-        "dtype": "bool",
-        "shape": (1,),
-        "names": None,
-    },
-    # true on last step of the episode if it is a terminal step, True for demos
-    "is_terminal": {
-        "dtype": "bool",
-        "shape": (1,),
-        "names": None,
-    },
-    # language_instruction is also stored as "task" to follow LeRobot standard
-    "language_instruction": {
-        "dtype": "string",
-        "shape": (1,),
-        "names": None,
-    },
-    "language_instruction_2": {
-        "dtype": "string",
-        "shape": (1,),
-        "names": None,
-    },
-    "language_instruction_3": {
-        "dtype": "string",
-        "shape": (1,),
-        "names": None,
-    },
-    "observation.state.gripper_position": {
-        "dtype": "float32",
-        "shape": (1,),
-        "names": {
-            "axes": ["gripper"],
-        },
-    },
-    "observation.state.cartesian_position": {
-        "dtype": "float32",
-        "shape": (6,),
-        "names": {
-            "axes": ["x", "y", "z", "roll", "pitch", "yaw"],
-        },
-    },
-    "observation.state.joint_position": {
-        "dtype": "float32",
-        "shape": (7,),
-        "names": {
-            "axes": ["joint_0", "joint_1", "joint_2", "joint_3", "joint_4", "joint_5", "joint_6"],
-        },
-    },
-    # Add this new feature to follow LeRobot standard of using joint position + gripper
-    "observation.state": {
-        "dtype": "float32",
-        "shape": (8,),
-        "names": {
-            "axes": ["joint_0", "joint_1", "joint_2", "joint_3", "joint_4", "joint_5", "joint_6", "gripper"],
-        },
-    },
-    # Initially called wrist_image_left
-    "observation.images.wrist_left": {
-        "dtype": "video",
-        "shape": (180, 320, 3),
-        "names": [
-            "height",
-            "width",
-            "channels",
-        ],
-    },
-    # Initially called exterior_image_1_left
-    "observation.images.exterior_1_left": {
-        "dtype": "video",
-        "shape": (180, 320, 3),
-        "names": [
-            "height",
-            "width",
-            "channels",
-        ],
-    },
-    # Initially called exterior_image_2_left
-    "observation.images.exterior_2_left": {
-        "dtype": "video",
-        "shape": (180, 320, 3),
-        "names": [
-            "height",
-            "width",
-            "channels",
-        ],
-    },
-    "action.gripper_position": {
-        "dtype": "float32",
-        "shape": (1,),
-        "names": {
-            "axes": ["gripper"],
-        },
-    },
-    "action.gripper_velocity": {
-        "dtype": "float32",
-        "shape": (1,),
-        "names": {
-            "axes": ["gripper"],
-        },
-    },
-    "action.cartesian_position": {
-        "dtype": "float32",
-        "shape": (6,),
-        "names": {
-            "axes": ["x", "y", "z", "roll", "pitch", "yaw"],
-        },
-    },
-    "action.cartesian_velocity": {
-        "dtype": "float32",
-        "shape": (6,),
-        "names": {
-            "axes": ["x", "y", "z", "roll", "pitch", "yaw"],
-        },
-    },
-    "action.joint_position": {
-        "dtype": "float32",
-        "shape": (7,),
-        "names": {
-            "axes": ["joint_0", "joint_1", "joint_2", "joint_3", "joint_4", "joint_5", "joint_6"],
-        },
-    },
-    "action.joint_velocity": {
-        "dtype": "float32",
-        "shape": (7,),
-        "names": {
-            "axes": ["joint_0", "joint_1", "joint_2", "joint_3", "joint_4", "joint_5", "joint_6"],
-        },
-    },
-    # This feature was called "action" in RLDS dataset and consists of [6x joint velocities, 1x gripper position]
-    "action.original": {
-        "dtype": "float32",
-        "shape": (7,),
-        "names": {
-            "axes": ["x", "y", "z", "roll", "pitch", "yaw", "gripper"],
-        },
-    },
-    # Add this new feature to follow LeRobot standard of using joint position + gripper
-    "action": {
-        "dtype": "float32",
-        "shape": (8,),
-        "names": {
-            "axes": ["joint_0", "joint_1", "joint_2", "joint_3", "joint_4", "joint_5", "joint_6", "gripper"],
-        },
-    },
-    "discount": {
-        "dtype": "float32",
-        "shape": (1,),
-        "names": None,
-    },
-    "reward": {
-        "dtype": "float32",
-        "shape": (1,),
-        "names": None,
-    },
-    # Meta data that are the same for all frames in the episode
-    "task_category": {
-        "dtype": "string",
-        "shape": (1,),
-        "names": None,
-    },
-    "building": {
-        "dtype": "string",
-        "shape": (1,),
-        "names": None,
-    },
-    "collector_id": {
-        "dtype": "string",
-        "shape": (1,),
-        "names": None,
-    },
-    "date": {
-        "dtype": "string",
-        "shape": (1,),
-        "names": None,
-    },
-    "camera_extrinsics.wrist_left": {
-        "dtype": "float32",
-        "shape": (6,),
-        "names": {
-            "axes": ["x", "y", "z", "roll", "pitch", "yaw"],
-        },
-    },
-    "camera_extrinsics.exterior_1_left": {
-        "dtype": "float32",
-        "shape": (6,),
-        "names": {
-            "axes": ["x", "y", "z", "roll", "pitch", "yaw"],
-        },
-    },
-    "camera_extrinsics.exterior_2_left": {
-        "dtype": "float32",
-        "shape": (6,),
-        "names": {
-            "axes": ["x", "y", "z", "roll", "pitch", "yaw"],
-        },
-    },
-    "is_episode_successful": {
-        "dtype": "bool",
-        "shape": (1,),
-        "names": None,
-    },
-}
-
-
-def is_episode_successful(tf_episode_metadata):
-    # Adapted from: https://github.com/droid-dataset/droid_policy_learning/blob/dd1020eb20d981f90b5ff07dc80d80d5c0cb108b/robomimic/utils/rlds_utils.py#L8
-    return "/success/" in tf_episode_metadata["file_path"].numpy().decode()
-
-
-def generate_lerobot_frames(tf_episode):
-    m = tf_episode["episode_metadata"]
-    frame_meta = {
-        "task_category": m["building"].numpy().decode(),
-        "building": m["building"].numpy().decode(),
-        "collector_id": m["collector_id"].numpy().decode(),
-        "date": m["date"].numpy().decode(),
-        "camera_extrinsics.wrist_left": m["extrinsics_wrist_cam"].numpy(),
-        "camera_extrinsics.exterior_1_left": m["extrinsics_exterior_cam_1"].numpy(),
-        "camera_extrinsics.exterior_2_left": m["extrinsics_exterior_cam_2"].numpy(),
-        "is_episode_successful": np.array([is_episode_successful(m)]),
-    }
-    for f in tf_episode["steps"]:
-        # Dataset schema slightly adapted from: https://droid-dataset.github.io/droid/the-droid-dataset.html#-dataset-schema
-        frame = {
-            "is_first": np.array([f["is_first"].numpy()]),
-            "is_last": np.array([f["is_last"].numpy()]),
-            "is_terminal": np.array([f["is_terminal"].numpy()]),
-            "language_instruction": f["language_instruction"].numpy().decode(),
-            "language_instruction_2": f["language_instruction_2"].numpy().decode(),
-            "language_instruction_3": f["language_instruction_3"].numpy().decode(),
-            "observation.state.gripper_position": f["observation"]["gripper_position"].numpy(),
-            "observation.state.cartesian_position": f["observation"]["cartesian_position"].numpy(),
-            "observation.state.joint_position": f["observation"]["joint_position"].numpy(),
-            "observation.images.wrist_left": f["observation"]["wrist_image_left"].numpy(),
-            "observation.images.exterior_1_left": f["observation"]["exterior_image_1_left"].numpy(),
-            "observation.images.exterior_2_left": f["observation"]["exterior_image_2_left"].numpy(),
-            "action.gripper_position": f["action_dict"]["gripper_position"].numpy(),
-            "action.gripper_velocity": f["action_dict"]["gripper_velocity"].numpy(),
-            "action.cartesian_position": f["action_dict"]["cartesian_position"].numpy(),
-            "action.cartesian_velocity": f["action_dict"]["cartesian_velocity"].numpy(),
-            "action.joint_position": f["action_dict"]["joint_position"].numpy(),
-            "action.joint_velocity": f["action_dict"]["joint_velocity"].numpy(),
-            "discount": np.array([f["discount"].numpy()]),
-            "reward": np.array([f["reward"].numpy()]),
-            "action.original": f["action"].numpy(),
-        }
-
-        # language_instruction is also stored as "task" to follow LeRobot standard
-        frame["task"] = frame["language_instruction"]
-
-        # Add this new feature to follow LeRobot standard of using joint position + gripper
-        frame["observation.state"] = np.concatenate(
-            [frame["observation.state.joint_position"], frame["observation.state.gripper_position"]]
-        )
-        frame["action"] = np.concatenate([frame["action.joint_position"], frame["action.gripper_position"]])
-
-        # Meta data that are the same for all frames in the episode
-        frame.update(frame_meta)
-
-        # Cast fp64 to fp32
-        for key in frame:
-            if isinstance(frame[key], np.ndarray) and frame[key].dtype == np.float64:
-                frame[key] = frame[key].astype(np.float32)
-
-        yield frame
-
-
-def port_droid(
-    raw_dir: Path,
-    repo_id: str,
-    push_to_hub: bool = False,
-    num_shards: int | None = None,
-    shard_index: int | None = None,
-):
-    dataset_name = raw_dir.parent.name
-    version = raw_dir.name
-    data_dir = raw_dir.parent.parent
-
-    builder = tfds.builder(f"{dataset_name}/{version}", data_dir=data_dir, version="")
-
-    if num_shards is not None:
-        tfds_num_shards = builder.info.splits["train"].num_shards
-        if tfds_num_shards != DROID_SHARDS:
-            raise ValueError(
-                f"Number of shards of Droid dataset is expected to be {DROID_SHARDS} but is {tfds_num_shards}."
-            )
-        if num_shards != tfds_num_shards:
-            raise ValueError(
-                f"We only shard over the fixed number of shards provided by tensorflow dataset ({tfds_num_shards}), but {num_shards} shards provided instead."
-            )
-        if shard_index >= tfds_num_shards:
-            raise ValueError(
-                f"Shard index is greater than the num of shards ({shard_index} >= {num_shards})."
-            )
-
-        raw_dataset = builder.as_dataset(split=f"train[{shard_index}shard]")
-    else:
-        raw_dataset = builder.as_dataset(split="train")
-
-    lerobot_dataset = LeRobotDataset.create(
-        repo_id=repo_id,
-        robot_type=DROID_ROBOT_TYPE,
-        fps=DROID_FPS,
-        features=DROID_FEATURES,
-    )
-
-    start_time = time.time()
-    num_episodes = raw_dataset.cardinality().numpy().item()
-    logging.info(f"Number of episodes {num_episodes}")
-
-    for episode_index, episode in enumerate(raw_dataset):
-        elapsed_time = time.time() - start_time
-        d, h, m, s = get_elapsed_time_in_days_hours_minutes_seconds(elapsed_time)
-
-        logging.info(
-            f"{episode_index} / {num_episodes} episodes processed (after {d} days, {h} hours, {m} minutes, {s:.3f} seconds)"
-        )
-
-        for frame in generate_lerobot_frames(episode):
-            lerobot_dataset.add_frame(frame)
-
-        lerobot_dataset.save_episode()
-        logging.info("Save_episode")
-
-    if push_to_hub:
-        lerobot_dataset.push_to_hub(
-            # Add openx tag, since it belongs to the openx collection of datasets
-            tags=["openx"],
-            private=False,
-        )
-
-
-def validate_dataset(repo_id):
-    """Sanity check that ensure meta data can be loaded and all files are present."""
-    meta = LeRobotDatasetMetadata(repo_id)
-
-    if meta.total_episodes == 0:
-        raise ValueError("Number of episodes is 0.")
-
-    for ep_idx in range(meta.total_episodes):
-        data_path = meta.root / meta.get_data_file_path(ep_idx)
-
-        if not data_path.exists():
-            raise ValueError(f"Parquet file is missing in: {data_path}")
-
-        for vid_key in meta.video_keys:
-            vid_path = meta.root / meta.get_video_file_path(ep_idx, vid_key)
-            if not vid_path.exists():
-                raise ValueError(f"Video file is missing in: {vid_path}")
-
-
-def main():
-    parser = argparse.ArgumentParser()
-
-    parser.add_argument(
-        "--raw-dir",
-        type=Path,
-        required=True,
-        help="Directory containing input raw datasets (e.g. `path/to/dataset` or `path/to/dataset/version).",
-    )
-    parser.add_argument(
-        "--repo-id",
-        type=str,
-        help="Repositery identifier on Hugging Face: a community or a user name `/` the name of the dataset, required when push-to-hub is True",
-    )
-    parser.add_argument(
-        "--push-to-hub",
-        action="store_true",
-        help="Upload to hub.",
-    )
-    parser.add_argument(
-        "--num-shards",
-        type=int,
-        default=None,
-        help="Number of shards. Can be either None to load the full dataset, or 2048 to load one of the 2048 tensorflow dataset files.",
-    )
-    parser.add_argument(
-        "--shard-index",
-        type=int,
-        default=None,
-        help="Index of the shard. Can be either None to load the full dataset, or in [0,2047] to load one of the 2048 tensorflow dataset files.",
-    )
-
-    args = parser.parse_args()
-
-    port_droid(**vars(args))
-
-
-if __name__ == "__main__":
-    main()

examples/port_datasets/slurm_aggregate_shards.py

@@ -1,148 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-import argparse
-import logging
-from pathlib import Path
-
-from datatrove.executor import LocalPipelineExecutor
-from datatrove.executor.slurm import SlurmPipelineExecutor
-from datatrove.pipeline.base import PipelineStep
-from port_datasets.droid_rlds.port_droid import DROID_SHARDS
-
-from lerobot.datasets.aggregate import aggregate_datasets
-from lerobot.utils.utils import init_logging
-
-
-class AggregateDatasets(PipelineStep):
-    def __init__(
-        self,
-        repo_ids: list[str],
-        aggregated_repo_id: str,
-    ):
-        super().__init__()
-        self.repo_ids = repo_ids
-        self.aggr_repo_id = aggregated_repo_id
-
-    def run(self, data=None, rank: int = 0, world_size: int = 1):
-        init_logging()
-
-        # Since aggregate_datasets already handles parallel processing internally,
-        # we only need one worker to run the entire aggregation
-        if rank == 0:
-            logging.info(f"Starting aggregation of {len(self.repo_ids)} datasets into {self.aggr_repo_id}")
-            aggregate_datasets(self.repo_ids, self.aggr_repo_id)
-            logging.info("Aggregation complete!")
-        else:
-            logging.info(f"Worker {rank} skipping - only worker 0 performs aggregation")
-
-
-def make_aggregate_executor(
-    repo_ids, repo_id, job_name, logs_dir, workers, partition, cpus_per_task, mem_per_cpu, slurm=True
-):
-    kwargs = {
-        "pipeline": [
-            AggregateDatasets(repo_ids, repo_id),
-        ],
-        "logging_dir": str(logs_dir / job_name),
-    }
-
-    if slurm:
-        # For aggregation, we only need 1 task since aggregate_datasets handles everything
-        kwargs.update(
-            {
-                "job_name": job_name,
-                "tasks": 1,  # Only need 1 task for aggregation
-                "workers": 1,  # Only need 1 worker
-                "time": "08:00:00",
-                "partition": partition,
-                "cpus_per_task": cpus_per_task,
-                "sbatch_args": {"mem-per-cpu": mem_per_cpu},
-            }
-        )
-        executor = SlurmPipelineExecutor(**kwargs)
-    else:
-        kwargs.update(
-            {
-                "tasks": 1,
-                "workers": 1,
-            }
-        )
-        executor = LocalPipelineExecutor(**kwargs)
-
-    return executor
-
-
-def main():
-    parser = argparse.ArgumentParser()
-
-    parser.add_argument(
-        "--repo-id",
-        type=str,
-        help="Repository identifier on Hugging Face: a community or a user name `/` the name of the dataset, required when push-to-hub is True.",
-    )
-    parser.add_argument(
-        "--logs-dir",
-        type=Path,
-        help="Path to logs directory for `datatrove`.",
-    )
-    parser.add_argument(
-        "--job-name",
-        type=str,
-        default="aggr_droid",
-        help="Job name used in slurm, and name of the directory created inside the provided logs directory.",
-    )
-    parser.add_argument(
-        "--slurm",
-        type=int,
-        default=1,
-        help="Launch over slurm. Use `--slurm 0` to launch sequentially (useful to debug).",
-    )
-    parser.add_argument(
-        "--workers",
-        type=int,
-        default=1,  # Changed default to 1 since aggregation doesn't need multiple workers
-        help="Number of slurm workers. For aggregation, this should be 1.",
-    )
-    parser.add_argument(
-        "--partition",
-        type=str,
-        help="Slurm partition. Ideally a CPU partition. No need for GPU partition.",
-    )
-    parser.add_argument(
-        "--cpus-per-task",
-        type=int,
-        default=8,
-        help="Number of cpus that each slurm worker will use.",
-    )
-    parser.add_argument(
-        "--mem-per-cpu",
-        type=str,
-        default="1950M",
-        help="Memory per cpu that each worker will use.",
-    )
-
-    args = parser.parse_args()
-    kwargs = vars(args)
-    kwargs["slurm"] = kwargs.pop("slurm") == 1
-
-    repo_ids = [f"{args.repo_id}_world_{DROID_SHARDS}_rank_{rank}" for rank in range(DROID_SHARDS)]
-    aggregate_executor = make_aggregate_executor(repo_ids, **kwargs)
-    aggregate_executor.run()
-
-
-if __name__ == "__main__":
-    main()

examples/port_datasets/slurm_port_shards.py

@@ -1,162 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-import argparse
-from pathlib import Path
-
-from datatrove.executor import LocalPipelineExecutor
-from datatrove.executor.slurm import SlurmPipelineExecutor
-from datatrove.pipeline.base import PipelineStep
-from port_datasets.droid_rlds.port_droid import DROID_SHARDS
-
-
-class PortDroidShards(PipelineStep):
-    def __init__(
-        self,
-        raw_dir: Path | str,
-        repo_id: str = None,
-    ):
-        super().__init__()
-        self.raw_dir = Path(raw_dir)
-        self.repo_id = repo_id
-
-    def run(self, data=None, rank: int = 0, world_size: int = 1):
-        from datasets.utils.tqdm import disable_progress_bars
-        from port_datasets.droid_rlds.port_droid import port_droid, validate_dataset
-
-        from lerobot.utils.utils import init_logging
-
-        init_logging()
-        disable_progress_bars()
-
-        shard_repo_id = f"{self.repo_id}_world_{world_size}_rank_{rank}"
-
-        try:
-            validate_dataset(shard_repo_id)
-            return
-        except Exception:
-            pass  # nosec B110 - Dataset doesn't exist yet, continue with porting
-
-        port_droid(
-            self.raw_dir,
-            shard_repo_id,
-            push_to_hub=False,
-            num_shards=world_size,
-            shard_index=rank,
-        )
-
-        validate_dataset(shard_repo_id)
-
-
-def make_port_executor(
-    raw_dir, repo_id, job_name, logs_dir, workers, partition, cpus_per_task, mem_per_cpu, slurm=True
-):
-    kwargs = {
-        "pipeline": [
-            PortDroidShards(raw_dir, repo_id),
-        ],
-        "logging_dir": str(logs_dir / job_name),
-    }
-
-    if slurm:
-        kwargs.update(
-            {
-                "job_name": job_name,
-                "tasks": DROID_SHARDS,
-                "workers": workers,
-                "time": "08:00:00",
-                "partition": partition,
-                "cpus_per_task": cpus_per_task,
-                "sbatch_args": {"mem-per-cpu": mem_per_cpu},
-            }
-        )
-        executor = SlurmPipelineExecutor(**kwargs)
-    else:
-        kwargs.update(
-            {
-                "tasks": 1,
-                "workers": 1,
-            }
-        )
-        executor = LocalPipelineExecutor(**kwargs)
-
-    return executor
-
-
-def main():
-    parser = argparse.ArgumentParser()
-
-    parser.add_argument(
-        "--raw-dir",
-        type=Path,
-        required=True,
-        help="Directory containing input raw datasets (e.g. `path/to/dataset` or `path/to/dataset/version).",
-    )
-    parser.add_argument(
-        "--repo-id",
-        type=str,
-        help="Repositery identifier on Hugging Face: a community or a user name `/` the name of the dataset, required when push-to-hub is True.",
-    )
-    parser.add_argument(
-        "--logs-dir",
-        type=Path,
-        help="Path to logs directory for `datatrove`.",
-    )
-    parser.add_argument(
-        "--job-name",
-        type=str,
-        default="port_droid",
-        help="Job name used in slurm, and name of the directory created inside the provided logs directory.",
-    )
-    parser.add_argument(
-        "--slurm",
-        type=int,
-        default=1,
-        help="Launch over slurm. Use `--slurm 0` to launch sequentially (useful to debug).",
-    )
-    parser.add_argument(
-        "--workers",
-        type=int,
-        default=2048,
-        help="Number of slurm workers. It should be less than the maximum number of shards.",
-    )
-    parser.add_argument(
-        "--partition",
-        type=str,
-        help="Slurm partition. Ideally a CPU partition. No need for GPU partition.",
-    )
-    parser.add_argument(
-        "--cpus-per-task",
-        type=int,
-        default=8,
-        help="Number of cpus that each slurm worker will use.",
-    )
-    parser.add_argument(
-        "--mem-per-cpu",
-        type=str,
-        default="1950M",
-        help="Memory per cpu that each worker will use.",
-    )
-
-    args = parser.parse_args()
-    kwargs = vars(args)
-    kwargs["slurm"] = kwargs.pop("slurm") == 1
-    port_executor = make_port_executor(**kwargs)
-    port_executor.run()
-
-
-if __name__ == "__main__":
-    main()

examples/port_datasets/slurm_upload.py

@@ -1,281 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-import argparse
-import logging
-import os
-from pathlib import Path
-
-from datatrove.executor import LocalPipelineExecutor
-from datatrove.executor.slurm import SlurmPipelineExecutor
-from datatrove.pipeline.base import PipelineStep
-from huggingface_hub import HfApi
-from huggingface_hub.constants import REPOCARD_NAME
-from port_datasets.droid_rlds.port_droid import DROID_SHARDS
-
-from lerobot.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDatasetMetadata
-from lerobot.datasets.utils import create_lerobot_dataset_card
-from lerobot.utils.utils import init_logging
-
-
-class UploadDataset(PipelineStep):
-    def __init__(
-        self,
-        repo_id: str,
-        branch: str | None = None,
-        revision: str | None = None,
-        tags: list | None = None,
-        license: str | None = "apache-2.0",
-        private: bool = False,
-        distant_repo_id: str | None = None,
-        **card_kwargs,
-    ):
-        super().__init__()
-        self.repo_id = repo_id
-        self.distant_repo_id = self.repo_id if distant_repo_id is None else distant_repo_id
-        self.branch = branch
-        self.tags = tags
-        self.license = license
-        self.private = private
-        self.card_kwargs = card_kwargs
-        self.revision = revision if revision else CODEBASE_VERSION
-
-        if os.environ.get("HF_HUB_ENABLE_HF_TRANSFER", "0") != "1":
-            logging.warning(
-                'HF_HUB_ENABLE_HF_TRANSFER is not set to "1". Install hf_transfer and set the env '
-                "variable for faster uploads:\npip install hf-transfer\nexport HF_HUB_ENABLE_HF_TRANSFER=1"
-            )
-
-        self.create_repo()
-
-    def create_repo(self):
-        logging.info(f"Loading meta data from {self.repo_id}...")
-        meta = LeRobotDatasetMetadata(self.repo_id)
-
-        logging.info(f"Creating repo {self.distant_repo_id}...")
-        hub_api = HfApi()
-        hub_api.create_repo(
-            repo_id=self.distant_repo_id,
-            private=self.private,
-            repo_type="dataset",
-            exist_ok=True,
-        )
-        if self.branch:
-            hub_api.create_branch(
-                repo_id=self.distant_repo_id,
-                branch=self.branch,
-                revision=self.revision,
-                repo_type="dataset",
-                exist_ok=True,
-            )
-
-        if not hub_api.file_exists(
-            self.distant_repo_id, REPOCARD_NAME, repo_type="dataset", revision=self.branch
-        ):
-            card = create_lerobot_dataset_card(
-                tags=self.tags, dataset_info=meta.info, license=self.license, **self.card_kwargs
-            )
-            card.push_to_hub(repo_id=self.distant_repo_id, repo_type="dataset", revision=self.branch)
-
-            hub_api.create_tag(self.distant_repo_id, tag=CODEBASE_VERSION, repo_type="dataset")
-
-        def list_files_recursively(directory):
-            base_path = Path(directory)
-            return [str(file.relative_to(base_path)) for file in base_path.rglob("*") if file.is_file()]
-
-        logging.info(f"Listing all local files from {self.repo_id}...")
-        self.file_paths = list_files_recursively(meta.root)
-        self.file_paths = sorted(self.file_paths)
-
-    def create_chunks(self, lst, n):
-        from itertools import islice
-
-        it = iter(lst)
-        return [list(islice(it, size)) for size in [len(lst) // n + (i < len(lst) % n) for i in range(n)]]
-
-    def create_commits(self, additions):
-        import logging
-        import math
-        import random
-        import time
-
-        from huggingface_hub import create_commit
-        from huggingface_hub.utils import HfHubHTTPError
-
-        FILES_BETWEEN_COMMITS = 10  # noqa: N806
-        BASE_DELAY = 0.1  # noqa: N806
-        MAX_RETRIES = 12  # noqa: N806
-
-        # Split the files into smaller chunks for faster commit
-        # and avoiding "A commit has happened since" error
-        num_chunks = math.ceil(len(additions) / FILES_BETWEEN_COMMITS)
-        chunks = self.create_chunks(additions, num_chunks)
-
-        for chunk in chunks:
-            retries = 0
-            while True:
-                try:
-                    create_commit(
-                        self.distant_repo_id,
-                        repo_type="dataset",
-                        operations=chunk,
-                        commit_message=f"DataTrove upload ({len(chunk)} files)",
-                        revision=self.branch,
-                    )
-                    # TODO: every 100 chunks super_squach_commits()
-                    logging.info("create_commit completed!")
-                    break
-                except HfHubHTTPError as e:
-                    if "A commit has happened since" in e.server_message:
-                        if retries >= MAX_RETRIES:
-                            logging.error(f"Failed to create commit after {MAX_RETRIES=}. Giving up.")
-                            raise e
-                        logging.info("Commit creation race condition issue. Waiting...")
-                        time.sleep(BASE_DELAY * 2**retries + random.uniform(0, 2))
-                        retries += 1
-                    else:
-                        raise e
-
-    def run(self, data=None, rank: int = 0, world_size: int = 1):
-        import logging
-
-        from datasets.utils.tqdm import disable_progress_bars
-        from huggingface_hub import CommitOperationAdd, preupload_lfs_files
-
-        from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
-        from lerobot.utils.utils import init_logging
-
-        init_logging()
-        disable_progress_bars()
-
-        chunks = self.create_chunks(self.file_paths, world_size)
-        file_paths = chunks[rank]
-
-        if len(file_paths) == 0:
-            raise ValueError(file_paths)
-
-        logging.info("Pre-uploading LFS files...")
-        for i, path in enumerate(file_paths):
-            logging.info(f"{i}: {path}")
-
-        meta = LeRobotDatasetMetadata(self.repo_id)
-        additions = [
-            CommitOperationAdd(path_in_repo=path, path_or_fileobj=meta.root / path) for path in file_paths
-        ]
-        preupload_lfs_files(
-            repo_id=self.distant_repo_id, repo_type="dataset", additions=additions, revision=self.branch
-        )
-
-        logging.info("Creating commits...")
-        self.create_commits(additions)
-        logging.info("Done!")
-
-
-def make_upload_executor(
-    repo_id, job_name, logs_dir, workers, partition, cpus_per_task, mem_per_cpu, slurm=True
-):
-    kwargs = {
-        "pipeline": [
-            UploadDataset(repo_id),
-        ],
-        "logging_dir": str(logs_dir / job_name),
-    }
-
-    if slurm:
-        kwargs.update(
-            {
-                "job_name": job_name,
-                "tasks": DROID_SHARDS,
-                "workers": workers,
-                "time": "08:00:00",
-                "partition": partition,
-                "cpus_per_task": cpus_per_task,
-                "sbatch_args": {"mem-per-cpu": mem_per_cpu},
-            }
-        )
-        executor = SlurmPipelineExecutor(**kwargs)
-    else:
-        kwargs.update(
-            {
-                "tasks": DROID_SHARDS,
-                "workers": 1,
-            }
-        )
-        executor = LocalPipelineExecutor(**kwargs)
-
-    return executor
-
-
-def main():
-    parser = argparse.ArgumentParser()
-
-    parser.add_argument(
-        "--repo-id",
-        type=str,
-        help="Repositery identifier on Hugging Face: a community or a user name `/` the name of the dataset, required when push-to-hub is True.",
-    )
-    parser.add_argument(
-        "--logs-dir",
-        type=Path,
-        help="Path to logs directory for `datatrove`.",
-    )
-    parser.add_argument(
-        "--job-name",
-        type=str,
-        default="upload_droid",
-        help="Job name used in slurm, and name of the directory created inside the provided logs directory.",
-    )
-    parser.add_argument(
-        "--slurm",
-        type=int,
-        default=1,
-        help="Launch over slurm. Use `--slurm 0` to launch sequentially (useful to debug).",
-    )
-    parser.add_argument(
-        "--workers",
-        type=int,
-        default=50,
-        help="Number of slurm workers. It should be less than the maximum number of shards.",
-    )
-    parser.add_argument(
-        "--partition",
-        type=str,
-        help="Slurm partition. Ideally a CPU partition. No need for GPU partition.",
-    )
-    parser.add_argument(
-        "--cpus-per-task",
-        type=int,
-        default=8,
-        help="Number of cpus that each slurm worker will use.",
-    )
-    parser.add_argument(
-        "--mem-per-cpu",
-        type=str,
-        default="1950M",
-        help="Memory per cpu that each worker will use.",
-    )
-
-    init_logging()
-
-    args = parser.parse_args()
-    kwargs = vars(args)
-    kwargs["slurm"] = kwargs.pop("slurm") == 1
-    upload_executor = make_upload_executor(**kwargs)
-    upload_executor.run()
-
-
-if __name__ == "__main__":
-    main()

pyproject.toml

@@ -73,6 +73,7 @@ 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
@@ -84,6 +85,7 @@ dependencies = [
 
     # Support dependencies
     "deepdiff>=7.0.1,<9.0.0",
+    "flask>=3.0.3,<4.0.0",
     "imageio[ffmpeg]>=2.34.0,<3.0.0",
     "termcolor>=2.4.0,<4.0.0",
 ]
@@ -94,7 +96,7 @@ dependencies = [
 # Common
 pygame-dep = ["pygame>=2.5.1"]
 placo-dep = ["placo>=0.9.6"]
-transformers-dep = ["transformers>=4.50.3,<4.52.0"] # TODO: Bumb dependency
+transformers-dep = ["transformers<=4.52.0"]
 grpcio-dep = ["grpcio==1.73.1", "protobuf==6.31.0"]
 
 # Motors
@@ -105,12 +107,12 @@ dynamixel = ["dynamixel-sdk>=3.7.31"]
 gamepad = ["lerobot[pygame-dep]", "hidapi>=0.14.0"]
 hopejr = ["lerobot[feetech]", "lerobot[pygame-dep]"]
 lekiwi = ["lerobot[feetech]", "pyzmq>=26.2.1"]
-reachy2 = ["reachy2_sdk>=1.0.14"]
 kinematics = ["lerobot[placo-dep]"]
 intelrealsense = [
     "pyrealsense2>=2.55.1.6486 ; sys_platform != 'darwin'",
     "pyrealsense2-macosx>=2.54 ; sys_platform == 'darwin'",
 ]
+phone = ["hebi-py>=2.8.0", "teleop>=0.1.0"]
 # stretch = [
 #     "hello-robot-stretch-body>=0.7.27 ; sys_platform == 'linux'",
 #     "pyrender @ git+https://github.com/mmatl/pyrender.git ; sys_platform == 'linux'",
@@ -141,7 +143,6 @@ all = [
     "lerobot[gamepad]",
     "lerobot[hopejr]",
     "lerobot[lekiwi]",
-    "lerobot[reachy2]",
     "lerobot[kinematics]",
     "lerobot[intelrealsense]",
     "lerobot[pi0]",
@@ -153,7 +154,8 @@ all = [
     "lerobot[video_benchmark]",
     "lerobot[aloha]",
     "lerobot[pusht]",
-    "lerobot[xarm]"
+    "lerobot[xarm]",
+    "lerobot[phone]",
 ]
 
 [project.scripts]

scripts/convert_videos_to_images.py

@@ -0,0 +1,74 @@
+#!/usr/bin/env python
+
+"""
+Convert video dataset to image dataset for faster training.
+This pre-extracts all frames from MP4 files to PNG images.
+"""
+
+import argparse
+from pathlib import Path
+import logging
+import shutil
+
+def convert_dataset_videos_to_images(repo_id: str, root: str | None = None):
+    """Convert all videos in a LeRobot dataset to individual image files."""
+    from lerobot.datasets.lerobot_dataset import LeRobotDataset
+    from lerobot.datasets.video_utils import decode_video_frames
+    import torch
+    
+    # Load dataset
+    dataset = LeRobotDataset(repo_id, root=root, download_videos=True)
+    
+    total_frames_processed = 0
+    
+    for ep_idx in range(dataset.meta.total_episodes):
+        logging.info(f"Processing episode {ep_idx}/{dataset.meta.total_episodes}")
+        
+        for vid_key in dataset.meta.video_keys:
+            video_path = dataset.root / dataset.meta.get_video_file_path(ep_idx, vid_key)
+            
+            if not video_path.exists():
+                logging.warning(f"Video not found: {video_path}")
+                continue
+                
+            # Create image directory  
+            img_dir = dataset.root / f"images/chunk-{dataset.meta.get_episode_chunk(ep_idx)}/{vid_key}"
+            img_dir.mkdir(parents=True, exist_ok=True)
+            
+            # Decode all frames from video
+            # Get episode length to decode all frames
+            ep_length = dataset.meta.episodes[ep_idx]["length"]
+            timestamps = [i / dataset.fps for i in range(ep_length)]
+            
+            try:
+                frames = decode_video_frames(video_path, timestamps, dataset.tolerance_s, dataset.video_backend)
+                
+                # Save each frame as PNG
+                for i, frame in enumerate(frames.squeeze(0)):
+                    img_path = img_dir / f"episode_{ep_idx:06d}_{i:06d}.png"
+                    # Convert tensor to PIL and save
+                    import torchvision.transforms as T
+                    to_pil = T.ToPILImage()
+                    pil_frame = to_pil(frame)
+                    pil_frame.save(img_path)
+                    
+                total_frames_processed += len(frames.squeeze(0))
+                logging.info(f"  Extracted {len(frames.squeeze(0))} frames to {img_dir}")
+                
+            except Exception as e:
+                logging.error(f"Failed to process {video_path}: {e}")
+                continue
+    
+    logging.info(f"Conversion complete! Processed {total_frames_processed} total frames")
+    logging.info(f"You can now use download_videos=False to use the extracted images")
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Convert LeRobot video dataset to images")
+    parser.add_argument("repo_id", help="Dataset repo ID (e.g., 'kenmacken/record-test-2')")
+    parser.add_argument("--root", help="Local root directory", default=None)
+    
+    args = parser.parse_args()
+    
+    logging.basicConfig(level=logging.INFO)
+    
+    convert_dataset_videos_to_images(args.repo_id, args.root)

src/lerobot/calibrate.py

@@ -18,7 +18,7 @@ Helper to recalibrate your device (robot or teleoperator).
 Example:
 
 ```shell
-lerobot-calibrate \
+python -m 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
-    lerobot-find-cameras opencv
+    python -m lerobot.find_cameras opencv
     ```
 
     The camera's default settings (FPS, resolution, color mode) are used unless
@@ -165,7 +165,8 @@ class OpenCVCamera(Camera):
             self.videocapture.release()
             self.videocapture = None
             raise ConnectionError(
-                f"Failed to open {self}.Run `lerobot-find-cameras opencv` to find available cameras."
+                f"Failed to open {self}."
+                f"Run `python -m lerobot.find_cameras opencv` to find available cameras."
             )
 
         self._configure_capture_settings()

src/lerobot/cameras/reachy2_camera/configuration_reachy2_camera.py

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

src/lerobot/cameras/reachy2_camera/reachy2_camera.py

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

src/lerobot/cameras/realsense/camera_realsense.py

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

src/lerobot/cameras/utils.py

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

src/lerobot/configs/default.py

@@ -33,11 +33,12 @@ class DatasetConfig:
     # Root directory where the dataset will be stored (e.g. 'dataset/path').
     root: str | None = None
     episodes: list[int] | None = None
+    # Percentage of dataset to use (0-100). If set, overrides episodes parameter.
+    percentage: float | None = None
     image_transforms: ImageTransformsConfig = field(default_factory=ImageTransformsConfig)
     revision: str | None = None
     use_imagenet_stats: bool = True
     video_backend: str = field(default_factory=get_safe_default_codec)
-    streaming: bool = False
 
 
 @dataclass

src/lerobot/configs/policies.py

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

src/lerobot/configs/types.py

@@ -24,6 +24,7 @@ class FeatureType(str, Enum):
     ENV = "ENV"
     ACTION = "ACTION"
     REWARD = "REWARD"
+    LANGUAGE = "LANGUAGE"
 
 
 class NormalizationMode(str, Enum):

src/lerobot/constants.py

@@ -21,6 +21,7 @@ OBS_ENV_STATE = "observation.environment_state"
 OBS_STATE = "observation.state"
 OBS_IMAGE = "observation.image"
 OBS_IMAGES = "observation.images"
+OBS_LANGUAGE = "observation.language"
 ACTION = "action"
 REWARD = "next.reward"
 
@@ -52,8 +53,3 @@ HF_LEROBOT_HOME = Path(os.getenv("HF_LEROBOT_HOME", default_cache_path)).expandu
 # calibration dir
 default_calibration_path = HF_LEROBOT_HOME / "calibration"
 HF_LEROBOT_CALIBRATION = Path(os.getenv("HF_LEROBOT_CALIBRATION", default_calibration_path)).expanduser()
-
-
-# streaming datasets
-LOOKBACK_BACKTRACKTABLE = 100
-LOOKAHEAD_BACKTRACKTABLE = 100

src/lerobot/datasets/aggregate.py

@@ -1,502 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2025 The HuggingFace Inc. team.
-# All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-import logging
-import shutil
-from pathlib import Path
-
-import pandas as pd
-import tqdm
-
-from lerobot.datasets.compute_stats import aggregate_stats
-from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
-from lerobot.datasets.utils import (
-    DEFAULT_CHUNK_SIZE,
-    DEFAULT_DATA_FILE_SIZE_IN_MB,
-    DEFAULT_DATA_PATH,
-    DEFAULT_EPISODES_PATH,
-    DEFAULT_VIDEO_FILE_SIZE_IN_MB,
-    DEFAULT_VIDEO_PATH,
-    get_parquet_file_size_in_mb,
-    get_video_size_in_mb,
-    to_parquet_with_hf_images,
-    update_chunk_file_indices,
-    write_info,
-    write_stats,
-    write_tasks,
-)
-from lerobot.datasets.video_utils import concatenate_video_files
-
-
-def validate_all_metadata(all_metadata: list[LeRobotDatasetMetadata]):
-    """Validates that all dataset metadata have consistent properties.
-
-    Ensures all datasets have the same fps, robot_type, and features to guarantee
-    compatibility when aggregating them into a single dataset.
-
-    Args:
-        all_metadata: List of LeRobotDatasetMetadata objects to validate.
-
-    Returns:
-        tuple: A tuple containing (fps, robot_type, features) from the first metadata.
-
-    Raises:
-        ValueError: If any metadata has different fps, robot_type, or features
-                   than the first metadata in the list.
-    """
-
-    fps = all_metadata[0].fps
-    robot_type = all_metadata[0].robot_type
-    features = all_metadata[0].features
-
-    for meta in tqdm.tqdm(all_metadata, desc="Validate all meta data"):
-        if fps != meta.fps:
-            raise ValueError(f"Same fps is expected, but got fps={meta.fps} instead of {fps}.")
-        if robot_type != meta.robot_type:
-            raise ValueError(
-                f"Same robot_type is expected, but got robot_type={meta.robot_type} instead of {robot_type}."
-            )
-        if features != meta.features:
-            raise ValueError(
-                f"Same features is expected, but got features={meta.features} instead of {features}."
-            )
-
-    return fps, robot_type, features
-
-
-def update_data_df(df, src_meta, dst_meta):
-    """Updates a data DataFrame with new indices and task mappings for aggregation.
-
-    Adjusts episode indices, frame indices, and task indices to account for
-    previously aggregated data in the destination dataset.
-
-    Args:
-        df: DataFrame containing the data to be updated.
-        src_meta: Source dataset metadata.
-        dst_meta: Destination dataset metadata.
-
-    Returns:
-        pd.DataFrame: Updated DataFrame with adjusted indices.
-    """
-
-    def _update(row):
-        row["episode_index"] = row["episode_index"] + dst_meta.info["total_episodes"]
-        row["index"] = row["index"] + dst_meta.info["total_frames"]
-        task = src_meta.tasks.iloc[row["task_index"]].name
-        row["task_index"] = dst_meta.tasks.loc[task].task_index.item()
-        return row
-
-    return df.apply(_update, axis=1)
-
-
-def update_meta_data(
-    df,
-    dst_meta,
-    meta_idx,
-    data_idx,
-    videos_idx,
-):
-    """Updates metadata DataFrame with new chunk, file, and timestamp indices.
-
-    Adjusts all indices and timestamps to account for previously aggregated
-    data and videos in the destination dataset.
-
-    Args:
-        df: DataFrame containing the metadata to be updated.
-        dst_meta: Destination dataset metadata.
-        meta_idx: Dictionary containing current metadata chunk and file indices.
-        data_idx: Dictionary containing current data chunk and file indices.
-        videos_idx: Dictionary containing current video indices and timestamps.
-
-    Returns:
-        pd.DataFrame: Updated DataFrame with adjusted indices and timestamps.
-    """
-
-    def _update(row):
-        row["meta/episodes/chunk_index"] = row["meta/episodes/chunk_index"] + meta_idx["chunk"]
-        row["meta/episodes/file_index"] = row["meta/episodes/file_index"] + meta_idx["file"]
-        row["data/chunk_index"] = row["data/chunk_index"] + data_idx["chunk"]
-        row["data/file_index"] = row["data/file_index"] + data_idx["file"]
-        for key, video_idx in videos_idx.items():
-            row[f"videos/{key}/chunk_index"] = row[f"videos/{key}/chunk_index"] + video_idx["chunk"]
-            row[f"videos/{key}/file_index"] = row[f"videos/{key}/file_index"] + video_idx["file"]
-            row[f"videos/{key}/from_timestamp"] = (
-                row[f"videos/{key}/from_timestamp"] + video_idx["latest_duration"]
-            )
-            row[f"videos/{key}/to_timestamp"] = (
-                row[f"videos/{key}/to_timestamp"] + video_idx["latest_duration"]
-            )
-
-        row["dataset_from_index"] = row["dataset_from_index"] + dst_meta.info["total_frames"]
-        row["dataset_to_index"] = row["dataset_to_index"] + dst_meta.info["total_frames"]
-        row["episode_index"] = row["episode_index"] + dst_meta.info["total_episodes"]
-        return row
-
-    return df.apply(_update, axis=1)
-
-
-def aggregate_datasets(
-    repo_ids: list[str],
-    aggr_repo_id: str,
-    roots: list[Path] | None = None,
-    aggr_root: Path | None = None,
-    data_files_size_in_mb: float | None = None,
-    video_files_size_in_mb: float | None = None,
-    chunk_size: int | None = None,
-):
-    """Aggregates multiple LeRobot datasets into a single unified dataset.
-
-    This is the main function that orchestrates the aggregation process by:
-    1. Loading and validating all source dataset metadata
-    2. Creating a new destination dataset with unified tasks
-    3. Aggregating videos, data, and metadata from all source datasets
-    4. Finalizing the aggregated dataset with proper statistics
-
-    Args:
-        repo_ids: List of repository IDs for the datasets to aggregate.
-        aggr_repo_id: Repository ID for the aggregated output dataset.
-        roots: Optional list of root paths for the source datasets.
-        aggr_root: Optional root path for the aggregated dataset.
-        data_files_size_in_mb: Maximum size for data files in MB (defaults to DEFAULT_DATA_FILE_SIZE_IN_MB)
-        video_files_size_in_mb: Maximum size for video files in MB (defaults to DEFAULT_VIDEO_FILE_SIZE_IN_MB)
-        chunk_size: Maximum number of files per chunk (defaults to DEFAULT_CHUNK_SIZE)
-    """
-    logging.info("Start aggregate_datasets")
-
-    if data_files_size_in_mb is None:
-        data_files_size_in_mb = DEFAULT_DATA_FILE_SIZE_IN_MB
-    if video_files_size_in_mb is None:
-        video_files_size_in_mb = DEFAULT_VIDEO_FILE_SIZE_IN_MB
-    if chunk_size is None:
-        chunk_size = DEFAULT_CHUNK_SIZE
-
-    all_metadata = (
-        [LeRobotDatasetMetadata(repo_id) for repo_id in repo_ids]
-        if roots is None
-        else [
-            LeRobotDatasetMetadata(repo_id, root=root) for repo_id, root in zip(repo_ids, roots, strict=False)
-        ]
-    )
-    fps, robot_type, features = validate_all_metadata(all_metadata)
-    video_keys = [key for key in features if features[key]["dtype"] == "video"]
-
-    dst_meta = LeRobotDatasetMetadata.create(
-        repo_id=aggr_repo_id,
-        fps=fps,
-        robot_type=robot_type,
-        features=features,
-        root=aggr_root,
-    )
-
-    logging.info("Find all tasks")
-    unique_tasks = pd.concat([m.tasks for m in all_metadata]).index.unique()
-    dst_meta.tasks = pd.DataFrame({"task_index": range(len(unique_tasks))}, index=unique_tasks)
-
-    meta_idx = {"chunk": 0, "file": 0}
-    data_idx = {"chunk": 0, "file": 0}
-    videos_idx = {
-        key: {"chunk": 0, "file": 0, "latest_duration": 0, "episode_duration": 0} for key in video_keys
-    }
-
-    dst_meta.episodes = {}
-
-    for src_meta in tqdm.tqdm(all_metadata, desc="Copy data and videos"):
-        videos_idx = aggregate_videos(src_meta, dst_meta, videos_idx, video_files_size_in_mb, chunk_size)
-        data_idx = aggregate_data(src_meta, dst_meta, data_idx, data_files_size_in_mb, chunk_size)
-
-        meta_idx = aggregate_metadata(src_meta, dst_meta, meta_idx, data_idx, videos_idx)
-
-        dst_meta.info["total_episodes"] += src_meta.total_episodes
-        dst_meta.info["total_frames"] += src_meta.total_frames
-
-    finalize_aggregation(dst_meta, all_metadata)
-    logging.info("Aggregation complete.")
-
-
-def aggregate_videos(src_meta, dst_meta, videos_idx, video_files_size_in_mb, chunk_size):
-    """Aggregates video chunks from a source dataset into the destination dataset.
-
-    Handles video file concatenation and rotation based on file size limits.
-    Creates new video files when size limits are exceeded.
-
-    Args:
-        src_meta: Source dataset metadata.
-        dst_meta: Destination dataset metadata.
-        videos_idx: Dictionary tracking video chunk and file indices.
-        video_files_size_in_mb: Maximum size for video files in MB (defaults to DEFAULT_VIDEO_FILE_SIZE_IN_MB)
-        chunk_size: Maximum number of files per chunk (defaults to DEFAULT_CHUNK_SIZE)
-
-    Returns:
-        dict: Updated videos_idx with current chunk and file indices.
-    """
-    for key, video_idx in videos_idx.items():
-        unique_chunk_file_pairs = {
-            (chunk, file)
-            for chunk, file in zip(
-                src_meta.episodes[f"videos/{key}/chunk_index"],
-                src_meta.episodes[f"videos/{key}/file_index"],
-                strict=False,
-            )
-        }
-        unique_chunk_file_pairs = sorted(unique_chunk_file_pairs)
-
-        chunk_idx = video_idx["chunk"]
-        file_idx = video_idx["file"]
-
-        for src_chunk_idx, src_file_idx in unique_chunk_file_pairs:
-            src_path = src_meta.root / DEFAULT_VIDEO_PATH.format(
-                video_key=key,
-                chunk_index=src_chunk_idx,
-                file_index=src_file_idx,
-            )
-
-            dst_path = dst_meta.root / DEFAULT_VIDEO_PATH.format(
-                video_key=key,
-                chunk_index=chunk_idx,
-                file_index=file_idx,
-            )
-
-            # If a new file is created, we don't want to increment the latest_duration
-            update_latest_duration = False
-
-            if not dst_path.exists():
-                # First write to this destination file
-                dst_path.parent.mkdir(parents=True, exist_ok=True)
-                shutil.copy(str(src_path), str(dst_path))
-                continue  # not accumulating further, already copied the file in place
-
-            # Check file sizes before appending
-            src_size = get_video_size_in_mb(src_path)
-            dst_size = get_video_size_in_mb(dst_path)
-
-            if dst_size + src_size >= video_files_size_in_mb:
-                # Rotate to a new chunk/file
-                chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, chunk_size)
-                dst_path = dst_meta.root / DEFAULT_VIDEO_PATH.format(
-                    video_key=key,
-                    chunk_index=chunk_idx,
-                    file_index=file_idx,
-                )
-                dst_path.parent.mkdir(parents=True, exist_ok=True)
-                shutil.copy(str(src_path), str(dst_path))
-            else:
-                # Get the timestamps shift for this video
-                timestamps_shift_s = dst_meta.info["total_frames"] / dst_meta.info["fps"]
-
-                # Append to existing video file
-                concatenate_video_files(
-                    [dst_path, src_path],
-                    dst_path,
-                )
-                # Update the latest_duration when appending (shifts timestamps!)
-                update_latest_duration = not update_latest_duration
-
-        # Update the videos_idx with the final chunk and file indices for this key
-        videos_idx[key]["chunk"] = chunk_idx
-        videos_idx[key]["file"] = file_idx
-
-        if update_latest_duration:
-            videos_idx[key]["latest_duration"] += timestamps_shift_s
-
-    return videos_idx
-
-
-def aggregate_data(src_meta, dst_meta, data_idx, data_files_size_in_mb, chunk_size):
-    """Aggregates data chunks from a source dataset into the destination dataset.
-
-    Reads source data files, updates indices to match the aggregated dataset,
-    and writes them to the destination with proper file rotation.
-
-    Args:
-        src_meta: Source dataset metadata.
-        dst_meta: Destination dataset metadata.
-        data_idx: Dictionary tracking data chunk and file indices.
-
-    Returns:
-        dict: Updated data_idx with current chunk and file indices.
-    """
-    unique_chunk_file_ids = {
-        (c, f)
-        for c, f in zip(
-            src_meta.episodes["data/chunk_index"], src_meta.episodes["data/file_index"], strict=False
-        )
-    }
-
-    unique_chunk_file_ids = sorted(unique_chunk_file_ids)
-
-    for src_chunk_idx, src_file_idx in unique_chunk_file_ids:
-        src_path = src_meta.root / DEFAULT_DATA_PATH.format(
-            chunk_index=src_chunk_idx, file_index=src_file_idx
-        )
-        df = pd.read_parquet(src_path)
-        df = update_data_df(df, src_meta, dst_meta)
-
-        data_idx = append_or_create_parquet_file(
-            df,
-            src_path,
-            data_idx,
-            data_files_size_in_mb,
-            chunk_size,
-            DEFAULT_DATA_PATH,
-            contains_images=len(dst_meta.image_keys) > 0,
-            aggr_root=dst_meta.root,
-        )
-
-    return data_idx
-
-
-def aggregate_metadata(src_meta, dst_meta, meta_idx, data_idx, videos_idx):
-    """Aggregates metadata from a source dataset into the destination dataset.
-
-    Reads source metadata files, updates all indices and timestamps,
-    and writes them to the destination with proper file rotation.
-
-    Args:
-        src_meta: Source dataset metadata.
-        dst_meta: Destination dataset metadata.
-        meta_idx: Dictionary tracking metadata chunk and file indices.
-        data_idx: Dictionary tracking data chunk and file indices.
-        videos_idx: Dictionary tracking video indices and timestamps.
-
-    Returns:
-        dict: Updated meta_idx with current chunk and file indices.
-    """
-    chunk_file_ids = {
-        (c, f)
-        for c, f in zip(
-            src_meta.episodes["meta/episodes/chunk_index"],
-            src_meta.episodes["meta/episodes/file_index"],
-            strict=False,
-        )
-    }
-
-    chunk_file_ids = sorted(chunk_file_ids)
-    for chunk_idx, file_idx in chunk_file_ids:
-        src_path = src_meta.root / DEFAULT_EPISODES_PATH.format(chunk_index=chunk_idx, file_index=file_idx)
-        df = pd.read_parquet(src_path)
-        df = update_meta_data(
-            df,
-            dst_meta,
-            meta_idx,
-            data_idx,
-            videos_idx,
-        )
-
-        for k in videos_idx:
-            videos_idx[k]["latest_duration"] += videos_idx[k]["episode_duration"]
-
-        meta_idx = append_or_create_parquet_file(
-            df,
-            src_path,
-            meta_idx,
-            DEFAULT_DATA_FILE_SIZE_IN_MB,
-            DEFAULT_CHUNK_SIZE,
-            DEFAULT_EPISODES_PATH,
-            contains_images=False,
-            aggr_root=dst_meta.root,
-        )
-
-    return meta_idx
-
-
-def append_or_create_parquet_file(
-    df: pd.DataFrame,
-    src_path: Path,
-    idx: dict[str, int],
-    max_mb: float,
-    chunk_size: int,
-    default_path: str,
-    contains_images: bool = False,
-    aggr_root: Path = None,
-):
-    """Appends data to an existing parquet file or creates a new one based on size constraints.
-
-    Manages file rotation when size limits are exceeded to prevent individual files
-    from becoming too large. Handles both regular parquet files and those containing images.
-
-    Args:
-        df: DataFrame to write to the parquet file.
-        src_path: Path to the source file (used for size estimation).
-        idx: Dictionary containing current 'chunk' and 'file' indices.
-        max_mb: Maximum allowed file size in MB before rotation.
-        chunk_size: Maximum number of files per chunk before incrementing chunk index.
-        default_path: Format string for generating file paths.
-        contains_images: Whether the data contains images requiring special handling.
-        aggr_root: Root path for the aggregated dataset.
-
-    Returns:
-        dict: Updated index dictionary with current chunk and file indices.
-    """
-    dst_path = aggr_root / default_path.format(chunk_index=idx["chunk"], file_index=idx["file"])
-
-    if not dst_path.exists():
-        dst_path.parent.mkdir(parents=True, exist_ok=True)
-        if contains_images:
-            to_parquet_with_hf_images(df, dst_path)
-        else:
-            df.to_parquet(dst_path)
-        return idx
-
-    src_size = get_parquet_file_size_in_mb(src_path)
-    dst_size = get_parquet_file_size_in_mb(dst_path)
-
-    if dst_size + src_size >= max_mb:
-        idx["chunk"], idx["file"] = update_chunk_file_indices(idx["chunk"], idx["file"], chunk_size)
-        new_path = aggr_root / default_path.format(chunk_index=idx["chunk"], file_index=idx["file"])
-        new_path.parent.mkdir(parents=True, exist_ok=True)
-        final_df = df
-        target_path = new_path
-    else:
-        existing_df = pd.read_parquet(dst_path)
-        final_df = pd.concat([existing_df, df], ignore_index=True)
-        target_path = dst_path
-
-    if contains_images:
-        to_parquet_with_hf_images(final_df, target_path)
-    else:
-        final_df.to_parquet(target_path)
-
-    return idx
-
-
-def finalize_aggregation(aggr_meta, all_metadata):
-    """Finalizes the dataset aggregation by writing summary files and statistics.
-
-    Writes the tasks file, info file with total counts and splits, and
-    aggregated statistics from all source datasets.
-
-    Args:
-        aggr_meta: Aggregated dataset metadata.
-        all_metadata: List of all source dataset metadata objects.
-    """
-    logging.info("write tasks")
-    write_tasks(aggr_meta.tasks, aggr_meta.root)
-
-    logging.info("write info")
-    aggr_meta.info.update(
-        {
-            "total_tasks": len(aggr_meta.tasks),
-            "total_episodes": sum(m.total_episodes for m in all_metadata),
-            "total_frames": sum(m.total_frames for m in all_metadata),
-            "splits": {"train": f"0:{sum(m.total_episodes for m in all_metadata)}"},
-        }
-    )
-    write_info(aggr_meta.info, aggr_meta.root)
-
-    logging.info("write stats")
-    aggr_meta.stats = aggregate_stats([m.stats for m in all_metadata])
-    write_stats(aggr_meta.stats, aggr_meta.root)

src/lerobot/datasets/backward_compatibility.py

@@ -14,13 +14,33 @@
 
 import packaging.version
 
-V30_MESSAGE = """
+V2_MESSAGE = """
 The dataset you requested ({repo_id}) is in {version} format.
 
-We introduced a new format since v3.0 which is not backward compatible with v2.1.
-Please, update your dataset to the new format using this command:
+We introduced a new format since v2.0 which is not backward compatible with v1.x.
+Please, use our conversion script. Modify the following command with your own task description:
 ```
-python -m lerobot.datasets.v30.convert_dataset_v21_to_v30 --repo-id={repo_id}
+python -m lerobot.datasets.v2.convert_dataset_v1_to_v2 \\
+    --repo-id {repo_id} \\
+    --single-task "TASK DESCRIPTION."  # <---- /!\\ Replace TASK DESCRIPTION /!\\
+```
+
+A few examples to replace TASK DESCRIPTION: "Pick up the blue cube and place it into the bin.", "Insert the
+peg into the socket.", "Slide open the ziploc bag.", "Take the elevator to the 1st floor.", "Open the top
+cabinet, store the pot inside it then close the cabinet.", "Push the T-shaped block onto the T-shaped
+target.", "Grab the spray paint on the shelf and place it in the bin on top of the robot dog.", "Fold the
+sweatshirt.", ...
+
+If you encounter a problem, contact LeRobot maintainers on [Discord](https://discord.com/invite/s3KuuzsPFb)
+or open an [issue on GitHub](https://github.com/huggingface/lerobot/issues/new/choose).
+"""
+
+V21_MESSAGE = """
+The dataset you requested ({repo_id}) is in {version} format.
+While current version of LeRobot is backward-compatible with it, the version of your dataset still uses global
+stats instead of per-episode stats. Update your dataset stats to the new format using this command:
+```
+python -m lerobot.datasets.v21.convert_dataset_v20_to_v21 --repo-id={repo_id}
 ```
 
 If you encounter a problem, contact LeRobot maintainers on [Discord](https://discord.com/invite/s3KuuzsPFb)
@@ -38,12 +58,7 @@ class CompatibilityError(Exception): ...
 
 class BackwardCompatibilityError(CompatibilityError):
     def __init__(self, repo_id: str, version: packaging.version.Version):
-        if version.major == 2 and version.minor == 1:
-            message = V30_MESSAGE.format(repo_id=repo_id, version=version)
-        else:
-            raise NotImplementedError(
-                "Contact the maintainer on [Discord](https://discord.com/invite/s3KuuzsPFb)."
-            )
+        message = V2_MESSAGE.format(repo_id=repo_id, version=version)
         super().__init__(message)
 
 

src/lerobot/datasets/factory.py

@@ -13,7 +13,6 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-import logging
 from pprint import pformat
 
 import torch
@@ -25,7 +24,6 @@ from lerobot.datasets.lerobot_dataset import (
     LeRobotDatasetMetadata,
     MultiLeRobotDataset,
 )
-from lerobot.datasets.streaming_dataset import StreamingLeRobotDataset
 from lerobot.datasets.transforms import ImageTransforms
 
 IMAGENET_STATS = {
@@ -88,26 +86,29 @@ def make_dataset(cfg: TrainPipelineConfig) -> LeRobotDataset | MultiLeRobotDatas
             cfg.dataset.repo_id, root=cfg.dataset.root, revision=cfg.dataset.revision
         )
         delta_timestamps = resolve_delta_timestamps(cfg.policy, ds_meta)
-        if not cfg.dataset.streaming:
-            dataset = LeRobotDataset(
-                cfg.dataset.repo_id,
-                root=cfg.dataset.root,
-                episodes=cfg.dataset.episodes,
-                delta_timestamps=delta_timestamps,
-                image_transforms=image_transforms,
-                revision=cfg.dataset.revision,
-                video_backend=cfg.dataset.video_backend,
-            )
-        else:
-            dataset = StreamingLeRobotDataset(
-                cfg.dataset.repo_id,
-                root=cfg.dataset.root,
-                episodes=cfg.dataset.episodes,
-                delta_timestamps=delta_timestamps,
-                image_transforms=image_transforms,
-                revision=cfg.dataset.revision,
-                max_num_shards=cfg.num_workers,
+
+        # Handle percentage parameter
+        episodes = cfg.dataset.episodes
+        if cfg.dataset.percentage is not None:
+            # Calculate episodes based on percentage
+            total_episodes = ds_meta.total_episodes
+            num_episodes_to_use = max(1, int(total_episodes * cfg.dataset.percentage / 100))
+            episodes = list(range(num_episodes_to_use))
+            import logging
+
+            logging.info(
+                f"Using {cfg.dataset.percentage}% of dataset: {num_episodes_to_use}/{total_episodes} episodes"
             )
+
+        dataset = LeRobotDataset(
+            cfg.dataset.repo_id,
+            root=cfg.dataset.root,
+            episodes=episodes,
+            delta_timestamps=delta_timestamps,
+            image_transforms=image_transforms,
+            revision=cfg.dataset.revision,
+            video_backend=cfg.dataset.video_backend,
+        )
     else:
         raise NotImplementedError("The MultiLeRobotDataset isn't supported for now.")
         dataset = MultiLeRobotDataset(

src/lerobot/datasets/online_buffer.py

@@ -337,11 +337,13 @@ def compute_sampler_weights(
     if len(offline_dataset) > 0:
         offline_data_mask_indices = []
         for start_index, end_index in zip(
-            offline_dataset.meta.episodes["dataset_from_index"],
-            offline_dataset.meta.episodes["dataset_to_index"],
+            offline_dataset.episode_data_index["from"],
+            offline_dataset.episode_data_index["to"],
             strict=True,
         ):
-            offline_data_mask_indices.extend(range(start_index, end_index - offline_drop_n_last_frames))
+            offline_data_mask_indices.extend(
+                range(start_index.item(), end_index.item() - offline_drop_n_last_frames)
+            )
         offline_data_mask = torch.zeros(len(offline_dataset), dtype=torch.bool)
         offline_data_mask[torch.tensor(offline_data_mask_indices)] = True
         weights.append(

src/lerobot/datasets/pipeline_features.py

@@ -0,0 +1,94 @@
+# 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 collections.abc import Sequence
+from typing import Any
+
+from lerobot.datasets.utils import hw_to_dataset_features
+from lerobot.processor.pipeline import RobotProcessor
+
+
+def aggregate_pipeline_dataset_features(
+    pipeline: RobotProcessor,
+    initial_features: dict[str, Any],
+    *,
+    use_videos: bool = True,
+    patterns: Sequence[str] | None = None,
+) -> dict[str, dict]:
+    """
+    Aggregates the pipeline's features and returns a features dict ready for the dataset,
+    filtered to only those keys matching any of the given patterns (for action/state only).
+
+    - `initial_features`: raw camera specs, e.g. {"front": (h,w,c), ...}
+    - `use_videos`: whether to treat image features as video streams
+    - `patterns`: regexes to filter action & state features; images are included
+                  whenever use_videos=True, regardless of patterns.
+    """
+    import re
+
+    # Gather everything the pipeline features specifies, seeded with hardware cams:
+    all_features = pipeline.transform_features(initial_features)
+
+    # Helper to decide which action/state keys survive the `patterns` filter:
+    def keep(key: str) -> bool:
+        if patterns is None:
+            return True
+        return any(re.search(pat, key) for pat in patterns)
+
+    # Start with hardware dict, injecting initial cameras if videos are ON:
+    hw: dict[str, dict[str, Any]] = {}
+    if use_videos:
+        cams = {
+            name: shape
+            for name, shape in initial_features.items()
+            if isinstance(shape, tuple) and len(shape) == 3
+        }
+        if cams:
+            hw["observation"] = dict(cams)
+
+    # Go over every feature from the pipeline and merge:
+    for full_key, ty in all_features.items():
+        if full_key.startswith("action."):
+            # action.<feat>
+            if not keep(full_key):
+                continue
+            name = full_key[len("action.") :]
+            hw.setdefault("action", {})[name] = ty
+
+        elif full_key.startswith("observation.state."):
+            # observation.state.<feat>
+            if not keep(full_key):
+                continue
+            name = full_key[len("observation.state.") :]
+            hw.setdefault("observation", {})[name] = ty
+
+        elif full_key.startswith("observation.images."):
+            # observation.images.<cam>
+            # images obey ONLY the use_videos flag, not patterns
+            if not use_videos:
+                continue
+            name = full_key[len("observation.images.") :]
+            hw.setdefault("observation", {})[name] = ty
+
+        else:
+            # anything else (e.g. policy-only features) is ignored here
+            continue
+
+    out: dict[str, dict] = {}
+    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))
+
+    return out

src/lerobot/datasets/sampler.py

@@ -21,8 +21,7 @@ import torch
 class EpisodeAwareSampler:
     def __init__(
         self,
-        dataset_from_indices: list[int],
-        dataset_to_indices: list[int],
+        episode_data_index: dict,
         episode_indices_to_use: list | None = None,
         drop_n_first_frames: int = 0,
         drop_n_last_frames: int = 0,
@@ -31,8 +30,7 @@ class EpisodeAwareSampler:
         """Sampler that optionally incorporates episode boundary information.
 
         Args:
-            dataset_from_indices: List of indices containing the start of each episode in the dataset.
-            dataset_to_indices: List of indices containing the end of each episode in the dataset.
+            episode_data_index: Dictionary with keys 'from' and 'to' containing the start and end indices of each episode.
             episode_indices_to_use: List of episode indices to use. If None, all episodes are used.
                                     Assumes that episodes are indexed from 0 to N-1.
             drop_n_first_frames: Number of frames to drop from the start of each episode.
@@ -41,10 +39,12 @@ class EpisodeAwareSampler:
         """
         indices = []
         for episode_idx, (start_index, end_index) in enumerate(
-            zip(dataset_from_indices, dataset_to_indices, strict=True)
+            zip(episode_data_index["from"], episode_data_index["to"], strict=True)
         ):
             if episode_indices_to_use is None or episode_idx in episode_indices_to_use:
-                indices.extend(range(start_index + drop_n_first_frames, end_index - drop_n_last_frames))
+                indices.extend(
+                    range(start_index.item() + drop_n_first_frames, end_index.item() - drop_n_last_frames)
+                )
 
         self.indices = indices
         self.shuffle = shuffle

src/lerobot/datasets/streaming_dataset.py

@@ -1,535 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-from collections.abc import Callable, Generator, Iterator
-from pathlib import Path
-
-import datasets
-import numpy as np
-import torch
-from datasets import load_dataset
-
-from lerobot.constants import HF_LEROBOT_HOME, LOOKAHEAD_BACKTRACKTABLE, LOOKBACK_BACKTRACKTABLE
-from lerobot.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDatasetMetadata
-from lerobot.datasets.utils import (
-    Backtrackable,
-    LookAheadError,
-    LookBackError,
-    check_version_compatibility,
-    find_float_index,
-    get_delta_indices,
-    is_float_in_list,
-    item_to_torch,
-    safe_shard,
-)
-from lerobot.datasets.video_utils import (
-    VideoDecoderCache,
-    decode_video_frames_torchcodec,
-)
-
-
-class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
-    """LeRobotDataset with streaming capabilities.
-
-    This class extends LeRobotDataset to add streaming functionality, allowing data to be streamed
-    rather than loaded entirely into memory. This is especially useful for large datasets that may
-    not fit in memory or when you want to quickly explore a dataset without downloading it completely.
-
-    The key innovation is using a Backtrackable iterator that maintains a bounded buffer of recent
-    items, allowing us to access previous frames for delta timestamps without loading the entire
-    dataset into memory.
-
-    Example:
-        Basic usage:
-        ```python
-        from lerobot.common.datasets.streaming_dataset import StreamingLeRobotDataset
-
-        # Create a streaming dataset with delta timestamps
-        delta_timestamps = {
-            "observation.image": [-1.0, -0.5, 0.0],  # 1 sec ago, 0.5 sec ago, current
-            "action": [0.0, 0.1, 0.2],  # current, 0.1 sec future, 0.2 sec future
-        }
-
-        dataset = StreamingLeRobotDataset(
-            repo_id="your-dataset-repo-id",
-            delta_timestamps=delta_timestamps,
-            streaming=True,
-            buffer_size=1000,
-        )
-
-        # Iterate over the dataset
-        for i, item in enumerate(dataset):
-            print(f"Sample {i}: Episode {item['episode_index']} Frame {item['frame_index']}")
-            # item will contain stacked frames according to delta_timestamps
-            if i >= 10:
-                break
-        ```
-    """
-
-    def __init__(
-        self,
-        repo_id: str,
-        root: str | Path | None = None,
-        episodes: list[int] | None = None,
-        image_transforms: Callable | None = None,
-        delta_timestamps: dict[list[float]] | None = None,
-        tolerance_s: float = 1e-4,
-        revision: str | None = None,
-        force_cache_sync: bool = False,
-        streaming: bool = True,
-        buffer_size: int = 1000,
-        max_num_shards: int = 16,
-        seed: int = 42,
-        rng: np.random.Generator | None = None,
-        shuffle: bool = True,
-    ):
-        """Initialize a StreamingLeRobotDataset.
-
-        Args:
-            repo_id (str): This is the repo id that will be used to fetch the dataset.
-            root (Path | None, optional): Local directory to use for downloading/writing files.
-            episodes (list[int] | None, optional): If specified, this will only load episodes specified by
-                their episode_index in this list.
-            image_transforms (Callable | None, optional): Transform to apply to image data.
-            tolerance_s (float, optional): Tolerance in seconds for timestamp matching.
-            revision (str, optional): Git revision id (branch name, tag, or commit hash).
-            force_cache_sync (bool, optional): Flag to sync and refresh local files first.
-            streaming (bool, optional): Whether to stream the dataset or load it all. Defaults to True.
-            buffer_size (int, optional): Buffer size for shuffling when streaming. Defaults to 1000.
-            max_num_shards (int, optional): Number of shards to re-shard the input dataset into. Defaults to 16.
-            seed (int, optional): Reproducibility random seed.
-            rng (np.random.Generator | None, optional): Random number generator.
-            shuffle (bool, optional): Whether to shuffle the dataset across exhaustions. Defaults to True.
-        """
-        super().__init__()
-        self.repo_id = repo_id
-        self.root = Path(root) if root else HF_LEROBOT_HOME / repo_id
-        self.streaming_from_local = root is not None
-
-        self.image_transforms = image_transforms
-        self.episodes = episodes
-        self.tolerance_s = tolerance_s
-        self.revision = revision if revision else CODEBASE_VERSION
-        self.seed = seed
-        self.rng = rng if rng is not None else np.random.default_rng(seed)
-        self.shuffle = shuffle
-
-        self.streaming = streaming
-        self.buffer_size = buffer_size
-
-        # We cache the video decoders to avoid re-initializing them at each frame (avoiding a ~10x slowdown)
-        self.video_decoder_cache = None
-
-        self.root.mkdir(exist_ok=True, parents=True)
-
-        # Load metadata
-        self.meta = LeRobotDatasetMetadata(
-            self.repo_id, self.root, self.revision, force_cache_sync=force_cache_sync
-        )
-        # Check version
-        check_version_compatibility(self.repo_id, self.meta._version, CODEBASE_VERSION)
-
-        self.delta_timestamps = None
-        self.delta_indices = None
-
-        if delta_timestamps is not None:
-            self._validate_delta_timestamp_keys(delta_timestamps)  # raises ValueError if invalid
-            self.delta_timestamps = delta_timestamps
-            self.delta_indices = get_delta_indices(self.delta_timestamps, self.fps)
-
-        self.hf_dataset: datasets.IterableDataset = load_dataset(
-            self.repo_id if not self.streaming_from_local else str(self.root),
-            split="train",
-            streaming=self.streaming,
-            data_files="data/*/*.parquet",
-            revision=self.revision,
-        )
-
-        self.num_shards = min(self.hf_dataset.num_shards, max_num_shards)
-
-    @property
-    def num_frames(self):
-        return self.meta.total_frames
-
-    @property
-    def num_episodes(self):
-        return self.meta.total_episodes
-
-    @property
-    def fps(self):
-        return self.meta.fps
-
-    @staticmethod
-    def _iter_random_indices(
-        rng: np.random.Generator, buffer_size: int, random_batch_size=100
-    ) -> Iterator[int]:
-        while True:
-            yield from (int(i) for i in rng.integers(0, buffer_size, size=random_batch_size))
-
-    @staticmethod
-    def _infinite_generator_over_elements(rng: np.random.Generator, elements: list[int]) -> Iterator[int]:
-        while True:
-            yield rng.choice(elements)
-
-    # TODO(fracapuano): Implement multi-threaded prefetching to accelerate data loading.
-    # The current sequential iteration is a bottleneck. A producer-consumer pattern
-    # could be used with a ThreadPoolExecutor to run `make_frame` (especially video decoding)
-    # in parallel, feeding a queue from which this iterator will yield processed items.
-    def __iter__(self) -> Iterator[dict[str, torch.Tensor]]:
-        if self.video_decoder_cache is None:
-            self.video_decoder_cache = VideoDecoderCache()
-
-        # keep the same seed across exhaustions if shuffle is False, otherwise shuffle data across exhaustions
-        rng = np.random.default_rng(self.seed) if not self.shuffle else self.rng
-
-        buffer_indices_generator = self._iter_random_indices(rng, self.buffer_size)
-
-        idx_to_backtrack_dataset = {
-            idx: self._make_backtrackable_dataset(safe_shard(self.hf_dataset, idx, self.num_shards))
-            for idx in range(self.num_shards)
-        }
-
-        # This buffer is populated while iterating on the dataset's shards
-        # the logic is to add 2 levels of randomness:
-        # (1) sample one shard at random from the ones available, and
-        # (2) sample one frame from the shard sampled at (1)
-        frames_buffer = []
-        while available_shards := list(idx_to_backtrack_dataset.keys()):
-            shard_key = next(self._infinite_generator_over_elements(rng, available_shards))
-            backtrack_dataset = idx_to_backtrack_dataset[shard_key]  # selects which shard to iterate on
-
-            try:
-                for frame in self.make_frame(backtrack_dataset):
-                    if len(frames_buffer) == self.buffer_size:
-                        i = next(buffer_indices_generator)  # samples a element from the buffer
-                        yield frames_buffer[i]
-                        frames_buffer[i] = frame
-                    else:
-                        frames_buffer.append(frame)
-                    break  # random shard sampled, switch shard
-            except (
-                RuntimeError,
-                StopIteration,
-            ):  # NOTE: StopIteration inside a generator throws a RuntimeError since python 3.7
-                del idx_to_backtrack_dataset[shard_key]  # Remove exhausted shard, onto another shard
-
-        # Once shards are all exhausted, shuffle the buffer and yield the remaining frames
-        rng.shuffle(frames_buffer)
-        yield from frames_buffer
-
-    def _get_window_steps(
-        self, delta_timestamps: dict[str, list[float]] | None = None, dynamic_bounds: bool = False
-    ) -> tuple[int, int]:
-        if delta_timestamps is None:
-            return 1, 1
-
-        if not dynamic_bounds:
-            # Fix the windows
-            lookback = LOOKBACK_BACKTRACKTABLE
-            lookahead = LOOKAHEAD_BACKTRACKTABLE
-        else:
-            # Dynamically adjust the windows based on the given delta_timesteps
-            all_timestamps = sum(delta_timestamps.values(), [])
-            lookback = min(all_timestamps) * self.fps
-            lookahead = max(all_timestamps) * self.fps
-
-            # When lookback is >=0 it means no negative timesteps have been provided
-            lookback = 0 if lookback >= 0 else (lookback * -1)
-
-        return lookback, lookahead
-
-    def _make_backtrackable_dataset(self, dataset: datasets.IterableDataset) -> Backtrackable:
-        lookback, lookahead = self._get_window_steps(self.delta_timestamps)
-        return Backtrackable(dataset, history=lookback, lookahead=lookahead)
-
-    def _make_timestamps_from_indices(
-        self, start_ts: float, indices: dict[str, list[int]] | None = None
-    ) -> dict[str, list[float]]:
-        if indices is not None:
-            return {
-                key: (
-                    start_ts + torch.tensor(indices[key]) / self.fps
-                ).tolist()  # NOTE: why not delta_timestamps directly?
-                for key in self.delta_timestamps
-            }
-        else:
-            return dict.fromkeys(self.meta.video_keys, [start_ts])
-
-    def _make_padding_camera_frame(self, camera_key: str):
-        """Variable-shape padding frame for given camera keys, given in (H, W, C)"""
-        return torch.zeros(self.meta.info["features"][camera_key]["shape"]).permute(-1, 0, 1)
-
-    def _get_video_frame_padding_mask(
-        self,
-        video_frames: dict[str, torch.Tensor],
-        query_timestamps: dict[str, list[float]],
-        original_timestamps: dict[str, list[float]],
-    ) -> dict[str, torch.BoolTensor]:
-        padding_mask = {}
-
-        for video_key, timestamps in original_timestamps.items():
-            if video_key not in video_frames:
-                continue  # only padding on video keys that are available
-            frames = []
-            mask = []
-            padding_frame = self._make_padding_camera_frame(video_key)
-            for ts in timestamps:
-                if is_float_in_list(ts, query_timestamps[video_key]):
-                    idx = find_float_index(ts, query_timestamps[video_key])
-                    frames.append(video_frames[video_key][idx, :])
-                    mask.append(False)
-                else:
-                    frames.append(padding_frame)
-                    mask.append(True)
-
-            padding_mask[f"{video_key}_is_pad"] = torch.BoolTensor(mask)
-
-        return padding_mask
-
-    def make_frame(
-        self, dataset_iterator: Backtrackable, previous_dataset_iterator: Backtrackable | None = None
-    ) -> Generator:
-        """Makes a frame starting from a dataset iterator"""
-        item = next(dataset_iterator)
-        item = item_to_torch(item)
-
-        updates = []  # list of "updates" to apply to the item retrieved from hf_dataset (w/o camera features)
-
-        # Get episode index from the item
-        ep_idx = item["episode_index"]
-
-        # "timestamp" restarts from 0 for each episode, whereas we need a global timestep within the single .mp4 file (given by index/fps)
-        current_ts = item["index"] / self.fps
-
-        episode_boundaries_ts = {
-            key: (
-                self.meta.episodes[ep_idx][f"videos/{key}/from_timestamp"],
-                self.meta.episodes[ep_idx][f"videos/{key}/to_timestamp"],
-            )
-            for key in self.meta.video_keys
-        }
-
-        # Apply delta querying logic if necessary
-        if self.delta_indices is not None:
-            query_result, padding = self._get_delta_frames(dataset_iterator, item)
-            updates.append(query_result)
-            updates.append(padding)
-
-        # Load video frames, when needed
-        if len(self.meta.video_keys) > 0:
-            original_timestamps = self._make_timestamps_from_indices(current_ts, self.delta_indices)
-
-            # Some timestamps might not result available considering the episode's boundaries
-            query_timestamps = self._get_query_timestamps(
-                current_ts, self.delta_indices, episode_boundaries_ts
-            )
-            video_frames = self._query_videos(query_timestamps, ep_idx)
-
-            if self.image_transforms is not None:
-                image_keys = self.meta.camera_keys
-                for cam in image_keys:
-                    video_frames[cam] = self.image_transforms(video_frames[cam])
-
-            updates.append(video_frames)
-
-            if self.delta_indices is not None:
-                # We always return the same number of frames. Unavailable frames are padded.
-                padding_mask = self._get_video_frame_padding_mask(
-                    video_frames, query_timestamps, original_timestamps
-                )
-                updates.append(padding_mask)
-
-        result = item.copy()
-        for update in updates:
-            result.update(update)
-
-        result["task"] = self.meta.tasks.iloc[item["task_index"]].name
-
-        yield result
-
-    def _get_query_timestamps(
-        self,
-        current_ts: float,
-        query_indices: dict[str, list[int]] | None = None,
-        episode_boundaries_ts: dict[str, tuple[float, float]] | None = None,
-    ) -> dict[str, list[float]]:
-        query_timestamps = {}
-        keys_to_timestamps = self._make_timestamps_from_indices(current_ts, query_indices)
-        for key in self.meta.video_keys:
-            if query_indices is not None and key in query_indices:
-                timestamps = keys_to_timestamps[key]
-                # Clamp out timesteps outside of episode boundaries
-                query_timestamps[key] = torch.clamp(
-                    torch.tensor(timestamps), *episode_boundaries_ts[key]
-                ).tolist()
-
-            else:
-                query_timestamps[key] = [current_ts]
-
-        return query_timestamps
-
-    def _query_videos(self, query_timestamps: dict[str, list[float]], ep_idx: int) -> dict:
-        """Note: When using data workers (e.g. DataLoader with num_workers>0), do not call this function
-        in the main process (e.g. by using a second Dataloader with num_workers=0). It will result in a
-        Segmentation Fault. This probably happens because a memory reference to the video loader is created in
-        the main process and a subprocess fails to access it.
-        """
-
-        item = {}
-        for video_key, query_ts in query_timestamps.items():
-            root = self.meta.url_root if self.streaming and not self.streaming_from_local else self.root
-            video_path = f"{root}/{self.meta.get_video_file_path(ep_idx, video_key)}"
-            frames = decode_video_frames_torchcodec(
-                video_path, query_ts, self.tolerance_s, decoder_cache=self.video_decoder_cache
-            )
-
-            item[video_key] = frames.squeeze(0) if len(query_ts) == 1 else frames
-
-        return item
-
-    def _get_delta_frames(self, dataset_iterator: Backtrackable, current_item: dict):
-        # TODO(fracapuano): Modularize this function, refactor the code
-        """Get frames with delta offsets using the backtrackable iterator.
-
-        Args:
-            current_item (dict): Current item from the iterator.
-            ep_idx (int): Episode index.
-
-        Returns:
-            tuple: (query_result, padding) - frames at delta offsets and padding info.
-        """
-        current_episode_idx = current_item["episode_index"]
-
-        # Prepare results
-        query_result = {}
-        padding = {}
-
-        for key, delta_indices in self.delta_indices.items():
-            if key in self.meta.video_keys:
-                continue  # visual frames are decoded separately
-
-            target_frames = []
-            is_pad = []
-
-            # Create a results dictionary to store frames in processing order, then reconstruct original order for stacking
-            delta_results = {}
-
-            # Separate and sort deltas by difficulty (easier operations first)
-            negative_deltas = sorted([d for d in delta_indices if d < 0], reverse=True)  # [-1, -2, -3, ...]
-            positive_deltas = sorted([d for d in delta_indices if d > 0])  # [1, 2, 3, ...]
-            zero_deltas = [d for d in delta_indices if d == 0]
-
-            # Process zero deltas (current frame)
-            for delta in zero_deltas:
-                delta_results[delta] = (
-                    current_item[key],
-                    False,
-                )
-
-            # Process negative deltas in order of increasing difficulty
-            lookback_failed = False
-
-            last_successful_frame = current_item[key]
-
-            for delta in negative_deltas:
-                if lookback_failed:
-                    delta_results[delta] = (last_successful_frame, True)
-                    continue
-
-                try:
-                    steps_back = abs(delta)
-                    if dataset_iterator.can_peek_back(steps_back):
-                        past_item = dataset_iterator.peek_back(steps_back)
-                        past_item = item_to_torch(past_item)
-
-                        if past_item["episode_index"] == current_episode_idx:
-                            delta_results[delta] = (past_item[key], False)
-                            last_successful_frame = past_item[key]
-
-                        else:
-                            raise LookBackError("Retrieved frame is from different episode!")
-                    else:
-                        raise LookBackError("Cannot go back further than the history buffer!")
-
-                except LookBackError:
-                    delta_results[delta] = (last_successful_frame, True)
-                    lookback_failed = True  # All subsequent negative deltas will also fail
-
-            # Process positive deltas in order of increasing difficulty
-            lookahead_failed = False
-            last_successful_frame = current_item[key]
-
-            for delta in positive_deltas:
-                if lookahead_failed:
-                    delta_results[delta] = (last_successful_frame, True)
-                    continue
-
-                try:
-                    if dataset_iterator.can_peek_ahead(delta):
-                        future_item = dataset_iterator.peek_ahead(delta)
-                        future_item = item_to_torch(future_item)
-
-                        if future_item["episode_index"] == current_episode_idx:
-                            delta_results[delta] = (future_item[key], False)
-                            last_successful_frame = future_item[key]
-
-                        else:
-                            raise LookAheadError("Retrieved frame is from different episode!")
-                    else:
-                        raise LookAheadError("Cannot go ahead further than the lookahead buffer!")
-
-                except LookAheadError:
-                    delta_results[delta] = (last_successful_frame, True)
-                    lookahead_failed = True  # All subsequent positive deltas will also fail
-
-            # Reconstruct original order for stacking
-            for delta in delta_indices:
-                frame, is_padded = delta_results[delta]
-
-                # add batch dimension for stacking
-                target_frames.append(frame)  # frame.unsqueeze(0))
-                is_pad.append(is_padded)
-
-            # Stack frames and add to results
-            if target_frames:
-                query_result[key] = torch.stack(target_frames)
-                padding[f"{key}_is_pad"] = torch.BoolTensor(is_pad)
-
-        return query_result, padding
-
-    def _validate_delta_timestamp_keys(self, delta_timestamps: dict[list[float]]) -> None:
-        """
-        Validate that all keys in delta_timestamps correspond to actual features in the dataset.
-
-        Raises:
-            ValueError: If any delta timestamp key doesn't correspond to a dataset feature.
-        """
-        if delta_timestamps is None:
-            return
-
-        # Get all available feature keys from the dataset metadata
-        available_features = set(self.meta.features.keys())
-
-        # Get all keys from delta_timestamps
-        delta_keys = set(delta_timestamps.keys())
-
-        # Find any keys that don't correspond to features
-        invalid_keys = delta_keys - available_features
-
-        if invalid_keys:
-            raise ValueError(
-                f"The following delta_timestamp keys do not correspond to dataset features: {invalid_keys}. "
-                f"Available features are: {sorted(available_features)}"
-            )

src/lerobot/datasets/utils.py

@@ -17,57 +17,43 @@ import contextlib
 import importlib.resources
 import json
 import logging
-from collections import deque
-from collections.abc import Iterable, Iterator
+from collections.abc import Iterator
+from itertools import accumulate
 from pathlib import Path
 from pprint import pformat
-from typing import Any, Deque, Generic, TypeVar
+from types import SimpleNamespace
+from typing import Any
 
 import datasets
+import jsonlines
 import numpy as np
 import packaging.version
-import pandas
-import pandas as pd
-import pyarrow.parquet as pq
 import torch
-from datasets import Dataset, concatenate_datasets
 from datasets.table import embed_table_storage
 from huggingface_hub import DatasetCard, DatasetCardData, HfApi
 from huggingface_hub.errors import RevisionNotFoundError
 from PIL import Image as PILImage
 from torchvision import transforms
 
-from lerobot.configs.types import FeatureType, PolicyFeature
+from lerobot.configs.types import DictLike, FeatureType, PolicyFeature
 from lerobot.datasets.backward_compatibility import (
-    FUTURE_MESSAGE,
+    V21_MESSAGE,
     BackwardCompatibilityError,
     ForwardCompatibilityError,
 )
 from lerobot.utils.utils import is_valid_numpy_dtype_string
 
-DEFAULT_CHUNK_SIZE = 1000  # Max number of files per chunk
-DEFAULT_DATA_FILE_SIZE_IN_MB = 100  # Max size per file
-DEFAULT_VIDEO_FILE_SIZE_IN_MB = 500  # Max size per file
+DEFAULT_CHUNK_SIZE = 1000  # Max number of episodes per chunk
 
 INFO_PATH = "meta/info.json"
+EPISODES_PATH = "meta/episodes.jsonl"
 STATS_PATH = "meta/stats.json"
+EPISODES_STATS_PATH = "meta/episodes_stats.jsonl"
+TASKS_PATH = "meta/tasks.jsonl"
 
-EPISODES_DIR = "meta/episodes"
-DATA_DIR = "data"
-VIDEO_DIR = "videos"
-
-CHUNK_FILE_PATTERN = "chunk-{chunk_index:03d}/file-{file_index:03d}"
-DEFAULT_TASKS_PATH = "meta/tasks.parquet"
-DEFAULT_EPISODES_PATH = EPISODES_DIR + "/" + CHUNK_FILE_PATTERN + ".parquet"
-DEFAULT_DATA_PATH = DATA_DIR + "/" + CHUNK_FILE_PATTERN + ".parquet"
-DEFAULT_VIDEO_PATH = VIDEO_DIR + "/{video_key}/" + CHUNK_FILE_PATTERN + ".mp4"
-DEFAULT_IMAGE_PATH = "images/{image_key}/episode-{episode_index:06d}/frame-{frame_index:06d}.png"
-
-LEGACY_EPISODES_PATH = "meta/episodes.jsonl"
-LEGACY_EPISODES_STATS_PATH = "meta/episodes_stats.jsonl"
-LEGACY_TASKS_PATH = "meta/tasks.jsonl"
-LEGACY_DEFAULT_VIDEO_PATH = "videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4"
-LEGACY_DEFAULT_PARQUET_PATH = "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet"
+DEFAULT_VIDEO_PATH = "videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4"
+DEFAULT_PARQUET_PATH = "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet"
+DEFAULT_IMAGE_PATH = "images/{image_key}/episode_{episode_index:06d}/frame_{frame_index:06d}.png"
 
 DATASET_CARD_TEMPLATE = """
 ---
@@ -87,67 +73,6 @@ DEFAULT_FEATURES = {
     "task_index": {"dtype": "int64", "shape": (1,), "names": None},
 }
 
-T = TypeVar("T")
-
-
-def get_parquet_file_size_in_mb(parquet_path: str | Path) -> float:
-    metadata = pq.read_metadata(parquet_path)
-    total_uncompressed_size = 0
-    for row_group in range(metadata.num_row_groups):
-        rg_metadata = metadata.row_group(row_group)
-        for column in range(rg_metadata.num_columns):
-            col_metadata = rg_metadata.column(column)
-            total_uncompressed_size += col_metadata.total_uncompressed_size
-    return total_uncompressed_size / (1024**2)
-
-
-def get_hf_dataset_size_in_mb(hf_ds: Dataset) -> int:
-    return hf_ds.data.nbytes // (1024**2)
-
-
-def get_hf_dataset_cache_dir(hf_ds: Dataset) -> Path | None:
-    if hf_ds.cache_files is None or len(hf_ds.cache_files) == 0:
-        return None
-    return Path(hf_ds.cache_files[0]["filename"]).parents[2]
-
-
-def update_chunk_file_indices(chunk_idx: int, file_idx: int, chunks_size: int) -> tuple[int, int]:
-    if file_idx == chunks_size - 1:
-        file_idx = 0
-        chunk_idx += 1
-    else:
-        file_idx += 1
-    return chunk_idx, file_idx
-
-
-def load_nested_dataset(pq_dir: Path, features: datasets.Features | None = None) -> Dataset:
-    """Find parquet files in provided directory {pq_dir}/chunk-xxx/file-xxx.parquet
-    Convert parquet files to pyarrow memory mapped in a cache folder for efficient RAM usage
-    Concatenate all pyarrow references to return HF Dataset format
-
-    Args:
-        pq_dir: Directory containing parquet files
-        features: Optional features schema to ensure consistent loading of complex types like images
-    """
-    paths = sorted(pq_dir.glob("*/*.parquet"))
-    if len(paths) == 0:
-        raise FileNotFoundError(f"Provided directory does not contain any parquet file: {pq_dir}")
-
-    # TODO(rcadene): set num_proc to accelerate conversion to pyarrow
-    datasets = [Dataset.from_parquet(str(path), features=features) for path in paths]
-    return concatenate_datasets(datasets)
-
-
-def get_parquet_num_frames(parquet_path: str | Path) -> int:
-    metadata = pq.read_metadata(parquet_path)
-    return metadata.num_rows
-
-
-def get_video_size_in_mb(mp4_path: Path) -> float:
-    file_size_bytes = mp4_path.stat().st_size
-    file_size_mb = file_size_bytes / (1024**2)
-    return file_size_mb
-
 
 def flatten_dict(d: dict, parent_key: str = "", sep: str = "/") -> dict:
     """Flatten a nested dictionary structure by collapsing nested keys into one key with a separator.
@@ -157,7 +82,6 @@ def flatten_dict(d: dict, parent_key: str = "", sep: str = "/") -> dict:
     >>> dct = {"a": {"b": 1, "c": {"d": 2}}, "e": 3}`
     >>> print(flatten_dict(dct))
     {"a/b": 1, "a/c/d": 2, "e": 3}
-    ```
     """
     items = []
     for k, v in d.items():
@@ -182,13 +106,23 @@ def unflatten_dict(d: dict, sep: str = "/") -> dict:
     return outdict
 
 
+def get_nested_item(obj: DictLike, flattened_key: str, sep: str = "/") -> Any:
+    split_keys = flattened_key.split(sep)
+    getter = obj[split_keys[0]]
+    if len(split_keys) == 1:
+        return getter
+
+    for key in split_keys[1:]:
+        getter = getter[key]
+
+    return getter
+
+
 def serialize_dict(stats: dict[str, torch.Tensor | np.ndarray | dict]) -> dict:
     serialized_dict = {}
     for key, value in flatten_dict(stats).items():
         if isinstance(value, (torch.Tensor, np.ndarray)):
             serialized_dict[key] = value.tolist()
-        elif isinstance(value, list) and isinstance(value[0], (int, float, list)):
-            serialized_dict[key] = value
         elif isinstance(value, np.generic):
             serialized_dict[key] = value.item()
         elif isinstance(value, (int, float)):
@@ -218,7 +152,24 @@ def write_json(data: dict, fpath: Path) -> None:
         json.dump(data, f, indent=4, ensure_ascii=False)
 
 
-def write_info(info: dict, local_dir: Path) -> None:
+def load_jsonlines(fpath: Path) -> list[Any]:
+    with jsonlines.open(fpath, "r") as reader:
+        return list(reader)
+
+
+def write_jsonlines(data: dict, fpath: Path) -> None:
+    fpath.parent.mkdir(exist_ok=True, parents=True)
+    with jsonlines.open(fpath, "w") as writer:
+        writer.write_all(data)
+
+
+def append_jsonlines(data: dict, fpath: Path) -> None:
+    fpath.parent.mkdir(exist_ok=True, parents=True)
+    with jsonlines.open(fpath, "a") as writer:
+        writer.write(data)
+
+
+def write_info(info: dict, local_dir: Path):
     write_json(info, local_dir / INFO_PATH)
 
 
@@ -229,68 +180,65 @@ def load_info(local_dir: Path) -> dict:
     return info
 
 
-def write_stats(stats: dict, local_dir: Path) -> None:
+def write_stats(stats: dict, local_dir: Path):
     serialized_stats = serialize_dict(stats)
     write_json(serialized_stats, local_dir / STATS_PATH)
 
 
-def cast_stats_to_numpy(stats: dict) -> dict[str, dict[str, np.ndarray]]:
+def cast_stats_to_numpy(stats) -> dict[str, dict[str, np.ndarray]]:
     stats = {key: np.array(value) for key, value in flatten_dict(stats).items()}
     return unflatten_dict(stats)
 
 
-def load_stats(local_dir: Path) -> dict[str, dict[str, np.ndarray]] | None:
+def load_stats(local_dir: Path) -> dict[str, dict[str, np.ndarray]]:
     if not (local_dir / STATS_PATH).exists():
         return None
     stats = load_json(local_dir / STATS_PATH)
     return cast_stats_to_numpy(stats)
 
 
-def write_tasks(tasks: pandas.DataFrame, local_dir: Path) -> None:
-    path = local_dir / DEFAULT_TASKS_PATH
-    path.parent.mkdir(parents=True, exist_ok=True)
-    tasks.to_parquet(path)
+def write_task(task_index: int, task: dict, local_dir: Path):
+    task_dict = {
+        "task_index": task_index,
+        "task": task,
+    }
+    append_jsonlines(task_dict, local_dir / TASKS_PATH)
 
 
-def load_tasks(local_dir: Path) -> pandas.DataFrame:
-    tasks = pd.read_parquet(local_dir / DEFAULT_TASKS_PATH)
-    return tasks
+def load_tasks(local_dir: Path) -> tuple[dict, dict]:
+    tasks = load_jsonlines(local_dir / TASKS_PATH)
+    tasks = {item["task_index"]: item["task"] for item in sorted(tasks, key=lambda x: x["task_index"])}
+    task_to_task_index = {task: task_index for task_index, task in tasks.items()}
+    return tasks, task_to_task_index
 
 
-def write_episodes(episodes: Dataset, local_dir: Path) -> None:
-    """Write episode metadata to a parquet file in the LeRobot v3.0 format.
-    This function writes episode-level metadata to a single parquet file.
-    Used primarily during dataset conversion (v2.1 → v3.0) and in test fixtures.
-
-    Args:
-        episodes: HuggingFace Dataset containing episode metadata
-        local_dir: Root directory where the dataset will be stored
-    """
-    episode_size_mb = get_hf_dataset_size_in_mb(episodes)
-    if episode_size_mb > DEFAULT_DATA_FILE_SIZE_IN_MB:
-        raise NotImplementedError(
-            f"Episodes dataset is too large ({episode_size_mb} MB) to write to a single file. "
-            f"The current limit is {DEFAULT_DATA_FILE_SIZE_IN_MB} MB. "
-            "This function only supports single-file episode metadata. "
-        )
-
-    fpath = local_dir / DEFAULT_EPISODES_PATH.format(chunk_index=0, file_index=0)
-    fpath.parent.mkdir(parents=True, exist_ok=True)
-    episodes.to_parquet(fpath)
+def write_episode(episode: dict, local_dir: Path):
+    append_jsonlines(episode, local_dir / EPISODES_PATH)
 
 
-def load_episodes(local_dir: Path) -> datasets.Dataset:
-    episodes = load_nested_dataset(local_dir / EPISODES_DIR)
-    # Select episode features/columns containing references to episode data and videos
-    # (e.g. tasks, dataset_from_index, dataset_to_index, data/chunk_index, data/file_index, etc.)
-    # This is to speedup access to these data, instead of having to load episode stats.
-    episodes = episodes.select_columns([key for key in episodes.features if not key.startswith("stats/")])
-    return episodes
+def load_episodes(local_dir: Path) -> dict:
+    episodes = load_jsonlines(local_dir / EPISODES_PATH)
+    return {item["episode_index"]: item for item in sorted(episodes, key=lambda x: x["episode_index"])}
+
+
+def write_episode_stats(episode_index: int, episode_stats: dict, local_dir: Path):
+    # We wrap episode_stats in a dictionary since `episode_stats["episode_index"]`
+    # is a dictionary of stats and not an integer.
+    episode_stats = {"episode_index": episode_index, "stats": serialize_dict(episode_stats)}
+    append_jsonlines(episode_stats, local_dir / EPISODES_STATS_PATH)
+
+
+def load_episodes_stats(local_dir: Path) -> dict:
+    episodes_stats = load_jsonlines(local_dir / EPISODES_STATS_PATH)
+    return {
+        item["episode_index"]: cast_stats_to_numpy(item["stats"])
+        for item in sorted(episodes_stats, key=lambda x: x["episode_index"])
+    }
 
 
 def backward_compatible_episodes_stats(
     stats: dict[str, dict[str, np.ndarray]], episodes: list[int]
-) -> dict[int, dict[str, dict[str, np.ndarray]]]:
+) -> dict[str, dict[str, np.ndarray]]:
     return dict.fromkeys(episodes, stats)
 
 
@@ -306,7 +254,7 @@ def load_image_as_numpy(
     return img_array
 
 
-def hf_transform_to_torch(items_dict: dict[str, list[Any]]) -> dict[str, list[torch.Tensor | str]]:
+def hf_transform_to_torch(items_dict: dict[torch.Tensor | None]):
     """Get a transform function that convert items from Hugging Face dataset (pyarrow)
     to torch tensors. Importantly, images are converted from PIL, which corresponds to
     a channel last representation (h w c) of uint8 type, to a torch image representation
@@ -351,7 +299,7 @@ def check_version_compatibility(
     if v_check.major < v_current.major and enforce_breaking_major:
         raise BackwardCompatibilityError(repo_id, v_check)
     elif v_check.minor < v_current.minor:
-        logging.warning(FUTURE_MESSAGE.format(repo_id=repo_id, version=v_check))
+        logging.warning(V21_MESSAGE.format(repo_id=repo_id, version=v_check))
 
 
 def get_repo_versions(repo_id: str) -> list[packaging.version.Version]:
@@ -522,15 +470,56 @@ def dataset_to_policy_features(features: dict[str, dict]) -> dict[str, PolicyFea
     return policy_features
 
 
+def merge_features(*dicts: dict) -> dict:
+    """
+    Merge LeRobot grouped feature dicts.
+
+    - For 1D numeric specs (dtype not image/video/string) with "names": we merge the names and recompute the shape.
+    - For others (observation.images.*), last one wins (if they are identical).
+    """
+    out: dict = {}
+    for d in dicts:
+        for key, value in d.items():
+            if not isinstance(value, dict):
+                out[key] = value
+                continue
+
+            dtype = value.get("dtype")
+            shape = value.get("shape")
+            is_vector = (
+                dtype not in ("image", "video", "string")
+                and isinstance(shape, tuple)
+                and len(shape) == 1
+                and "names" in value
+            )
+
+            if is_vector:
+                # Initialize or retrieve the accumulating dict for this feature key
+                target = out.setdefault(key, {"dtype": dtype, "names": [], "shape": (0,)})
+                # Ensure consistent data types across merged entries
+                if "dtype" in target and dtype != target["dtype"]:
+                    raise ValueError(f"dtype mismatch for '{key}': {target['dtype']} vs {dtype}")
+
+                # Merge feature names: append only new ones to preserve order without duplicates
+                seen = set(target["names"])
+                for n in value["names"]:
+                    if n not in seen:
+                        target["names"].append(n)
+                        seen.add(n)
+                # Recompute the shape to reflect the updated number of features
+                target["shape"] = (len(target["names"]),)
+            else:
+                # For images/videos and non-1D entries: override with the latest definition
+                out[key] = value
+    return out
+
+
 def create_empty_dataset_info(
     codebase_version: str,
     fps: int,
     features: dict,
     use_videos: bool,
     robot_type: str | None = None,
-    chunks_size: int | None = None,
-    data_files_size_in_mb: int | None = None,
-    video_files_size_in_mb: int | None = None,
 ) -> dict:
     return {
         "codebase_version": codebase_version,
@@ -538,17 +527,104 @@ def create_empty_dataset_info(
         "total_episodes": 0,
         "total_frames": 0,
         "total_tasks": 0,
-        "chunks_size": chunks_size or DEFAULT_CHUNK_SIZE,
-        "data_files_size_in_mb": data_files_size_in_mb or DEFAULT_DATA_FILE_SIZE_IN_MB,
-        "video_files_size_in_mb": video_files_size_in_mb or DEFAULT_VIDEO_FILE_SIZE_IN_MB,
+        "total_videos": 0,
+        "total_chunks": 0,
+        "chunks_size": DEFAULT_CHUNK_SIZE,
         "fps": fps,
         "splits": {},
-        "data_path": DEFAULT_DATA_PATH,
+        "data_path": DEFAULT_PARQUET_PATH,
         "video_path": DEFAULT_VIDEO_PATH if use_videos else None,
         "features": features,
     }
 
 
+def get_episode_data_index(
+    episode_dicts: dict[dict], episodes: list[int] | None = None
+) -> dict[str, torch.Tensor]:
+    episode_lengths = {ep_idx: ep_dict["length"] for ep_idx, ep_dict in episode_dicts.items()}
+    if episodes is not None:
+        episode_lengths = {ep_idx: episode_lengths[ep_idx] for ep_idx in episodes}
+
+    cumulative_lengths = list(accumulate(episode_lengths.values()))
+    return {
+        "from": torch.LongTensor([0] + cumulative_lengths[:-1]),
+        "to": torch.LongTensor(cumulative_lengths),
+    }
+
+
+def check_timestamps_sync(
+    timestamps: np.ndarray,
+    episode_indices: np.ndarray,
+    episode_data_index: dict[str, np.ndarray],
+    fps: int,
+    tolerance_s: float,
+    raise_value_error: bool = True,
+) -> bool:
+    """
+    This check is to make sure that each timestamp is separated from the next by (1/fps) +/- tolerance
+    to account for possible numerical error.
+
+    Args:
+        timestamps (np.ndarray): Array of timestamps in seconds.
+        episode_indices (np.ndarray): Array indicating the episode index for each timestamp.
+        episode_data_index (dict[str, np.ndarray]): A dictionary that includes 'to',
+            which identifies indices for the end of each episode.
+        fps (int): Frames per second. Used to check the expected difference between consecutive timestamps.
+        tolerance_s (float): Allowed deviation from the expected (1/fps) difference.
+        raise_value_error (bool): Whether to raise a ValueError if the check fails.
+
+    Returns:
+        bool: True if all checked timestamp differences lie within tolerance, False otherwise.
+
+    Raises:
+        ValueError: If the check fails and `raise_value_error` is True.
+    """
+    if timestamps.shape != episode_indices.shape:
+        raise ValueError(
+            "timestamps and episode_indices should have the same shape. "
+            f"Found {timestamps.shape=} and {episode_indices.shape=}."
+        )
+
+    # Consecutive differences
+    diffs = np.diff(timestamps)
+    within_tolerance = np.abs(diffs - (1.0 / fps)) <= tolerance_s
+
+    # Mask to ignore differences at the boundaries between episodes
+    mask = np.ones(len(diffs), dtype=bool)
+    ignored_diffs = episode_data_index["to"][:-1] - 1  # indices at the end of each episode
+    mask[ignored_diffs] = False
+    filtered_within_tolerance = within_tolerance[mask]
+
+    # Check if all remaining diffs are within tolerance
+    if not np.all(filtered_within_tolerance):
+        # Track original indices before masking
+        original_indices = np.arange(len(diffs))
+        filtered_indices = original_indices[mask]
+        outside_tolerance_filtered_indices = np.nonzero(~filtered_within_tolerance)[0]
+        outside_tolerance_indices = filtered_indices[outside_tolerance_filtered_indices]
+
+        outside_tolerances = []
+        for idx in outside_tolerance_indices:
+            entry = {
+                "timestamps": [timestamps[idx], timestamps[idx + 1]],
+                "diff": diffs[idx],
+                "episode_index": episode_indices[idx].item()
+                if hasattr(episode_indices[idx], "item")
+                else episode_indices[idx],
+            }
+            outside_tolerances.append(entry)
+
+        if raise_value_error:
+            raise ValueError(
+                f"""One or several timestamps unexpectedly violate the tolerance inside episode range.
+                This might be due to synchronization issues during data collection.
+                \n{pformat(outside_tolerances)}"""
+            )
+        return False
+
+    return True
+
+
 def check_delta_timestamps(
     delta_timestamps: dict[str, list[float]], fps: int, tolerance_s: float, raise_value_error: bool = True
 ) -> bool:
@@ -587,7 +663,7 @@ def get_delta_indices(delta_timestamps: dict[str, list[float]], fps: int) -> dic
     return delta_indices
 
 
-def cycle(iterable: Any) -> Iterator[Any]:
+def cycle(iterable):
     """The equivalent of itertools.cycle, but safe for Pytorch dataloaders.
 
     See https://github.com/pytorch/pytorch/issues/23900 for information on why itertools.cycle is not safe.
@@ -600,7 +676,7 @@ def cycle(iterable: Any) -> Iterator[Any]:
             iterator = iter(iterable)
 
 
-def create_branch(repo_id: str, *, branch: str, repo_type: str | None = None) -> None:
+def create_branch(repo_id, *, branch: str, repo_type: str | None = None) -> None:
     """Create a branch on a existing Hugging Face repo. Delete the branch if it already
     exists before creating it.
     """
@@ -653,28 +729,76 @@ def create_lerobot_dataset_card(
     )
 
 
-def validate_frame(frame: dict, features: dict) -> None:
+class IterableNamespace(SimpleNamespace):
+    """
+    A namespace object that supports both dictionary-like iteration and dot notation access.
+    Automatically converts nested dictionaries into IterableNamespaces.
+
+    This class extends SimpleNamespace to provide:
+    - Dictionary-style iteration over keys
+    - Access to items via both dot notation (obj.key) and brackets (obj["key"])
+    - Dictionary-like methods: items(), keys(), values()
+    - Recursive conversion of nested dictionaries
+
+    Args:
+        dictionary: Optional dictionary to initialize the namespace
+        **kwargs: Additional keyword arguments passed to SimpleNamespace
+
+    Examples:
+        >>> data = {"name": "Alice", "details": {"age": 25}}
+        >>> ns = IterableNamespace(data)
+        >>> ns.name
+        'Alice'
+        >>> ns.details.age
+        25
+        >>> list(ns.keys())
+        ['name', 'details']
+        >>> for key, value in ns.items():
+        ...     print(f"{key}: {value}")
+        name: Alice
+        details: IterableNamespace(age=25)
+    """
+
+    def __init__(self, dictionary: dict[str, Any] = None, **kwargs):
+        super().__init__(**kwargs)
+        if dictionary is not None:
+            for key, value in dictionary.items():
+                if isinstance(value, dict):
+                    setattr(self, key, IterableNamespace(value))
+                else:
+                    setattr(self, key, value)
+
+    def __iter__(self) -> Iterator[str]:
+        return iter(vars(self))
+
+    def __getitem__(self, key: str) -> Any:
+        return vars(self)[key]
+
+    def items(self):
+        return vars(self).items()
+
+    def values(self):
+        return vars(self).values()
+
+    def keys(self):
+        return vars(self).keys()
+
+
+def validate_frame(frame: dict, features: dict):
     expected_features = set(features) - set(DEFAULT_FEATURES)
     actual_features = set(frame)
 
-    # task is a special required field that's not part of regular features
-    if "task" not in actual_features:
-        raise ValueError("Feature mismatch in `frame` dictionary:\nMissing features: {'task'}\n")
+    error_message = validate_features_presence(actual_features, expected_features)
 
-    # Remove task from actual_features for regular feature validation
-    actual_features_for_validation = actual_features - {"task"}
-
-    error_message = validate_features_presence(actual_features_for_validation, expected_features)
-
-    common_features = actual_features_for_validation & expected_features
-    for name in common_features:
+    common_features = actual_features & expected_features
+    for name in common_features - {"task"}:
         error_message += validate_feature_dtype_and_shape(name, features[name], frame[name])
 
     if error_message:
         raise ValueError(error_message)
 
 
-def validate_features_presence(actual_features: set[str], expected_features: set[str]) -> str:
+def validate_features_presence(actual_features: set[str], expected_features: set[str]):
     error_message = ""
     missing_features = expected_features - actual_features
     extra_features = actual_features - expected_features
@@ -689,9 +813,7 @@ def validate_features_presence(actual_features: set[str], expected_features: set
     return error_message
 
 
-def validate_feature_dtype_and_shape(
-    name: str, feature: dict, value: np.ndarray | PILImage.Image | str
-) -> str:
+def validate_feature_dtype_and_shape(name: str, feature: dict, value: np.ndarray | PILImage.Image | str):
     expected_dtype = feature["dtype"]
     expected_shape = feature["shape"]
     if is_valid_numpy_dtype_string(expected_dtype):
@@ -706,7 +828,7 @@ def validate_feature_dtype_and_shape(
 
 def validate_feature_numpy_array(
     name: str, expected_dtype: str, expected_shape: list[int], value: np.ndarray
-) -> str:
+):
     error_message = ""
     if isinstance(value, np.ndarray):
         actual_dtype = value.dtype
@@ -723,9 +845,7 @@ def validate_feature_numpy_array(
     return error_message
 
 
-def validate_feature_image_or_video(
-    name: str, expected_shape: list[str], value: np.ndarray | PILImage.Image
-) -> str:
+def validate_feature_image_or_video(name: str, expected_shape: list[str], value: np.ndarray | PILImage.Image):
     # Note: The check of pixels range ([0,1] for float and [0,255] for uint8) is done by the image writer threads.
     error_message = ""
     if isinstance(value, np.ndarray):
@@ -741,13 +861,13 @@ def validate_feature_image_or_video(
     return error_message
 
 
-def validate_feature_string(name: str, value: str) -> str:
+def validate_feature_string(name: str, value: str):
     if not isinstance(value, str):
         return f"The feature '{name}' is expected to be of type 'str', but type '{type(value)}' provided instead.\n"
     return ""
 
 
-def validate_episode_buffer(episode_buffer: dict, total_episodes: int, features: dict) -> None:
+def validate_episode_buffer(episode_buffer: dict, total_episodes: int, features: dict):
     if "size" not in episode_buffer:
         raise ValueError("size key not found in episode_buffer")
 
@@ -771,238 +891,3 @@ def validate_episode_buffer(episode_buffer: dict, total_episodes: int, features:
             f"In episode_buffer not in features: {buffer_keys - set(features)}"
             f"In features not in episode_buffer: {set(features) - buffer_keys}"
         )
-
-
-def to_parquet_with_hf_images(df: pandas.DataFrame, path: Path) -> None:
-    """This function correctly writes to parquet a panda DataFrame that contains images encoded by HF dataset.
-    This way, it can be loaded by HF dataset and correctly formatted images are returned.
-    """
-    # TODO(qlhoest): replace this weird synthax by `df.to_parquet(path)` only
-    datasets.Dataset.from_dict(df.to_dict(orient="list")).to_parquet(path)
-
-
-def item_to_torch(item: dict) -> dict:
-    """Convert all items in a dictionary to PyTorch tensors where appropriate.
-
-    This function is used to convert an item from a streaming dataset to PyTorch tensors.
-
-    Args:
-        item (dict): Dictionary of items from a dataset.
-
-    Returns:
-        dict: Dictionary with all tensor-like items converted to torch.Tensor.
-    """
-    for key, val in item.items():
-        if isinstance(val, (np.ndarray, list)) and key not in ["task"]:
-            # Convert numpy arrays and lists to torch tensors
-            item[key] = torch.tensor(val)
-    return item
-
-
-def is_float_in_list(target, float_list, threshold=1e-6):
-    return any(abs(target - x) <= threshold for x in float_list)
-
-
-def find_float_index(target, float_list, threshold=1e-6):
-    for i, x in enumerate(float_list):
-        if abs(target - x) <= threshold:
-            return i
-    return -1
-
-
-class LookBackError(Exception):
-    """
-    Exception raised when trying to look back in the history of a Backtrackable object.
-    """
-
-    pass
-
-
-class LookAheadError(Exception):
-    """
-    Exception raised when trying to look ahead in the future of a Backtrackable object.
-    """
-
-    pass
-
-
-class Backtrackable(Generic[T]):
-    """
-    Wrap any iterator/iterable so you can step back up to `history` items
-    and look ahead up to `lookahead` items.
-
-    This is useful for streaming datasets where you need to access previous and future items
-    but can't load the entire dataset into memory.
-
-    Example:
-    -------
-    ```python
-    ds = load_dataset("c4", "en", streaming=True, split="train")
-    rev = Backtrackable(ds, history=3, lookahead=2)
-
-    x0 = next(rev)  # forward
-    x1 = next(rev)
-    x2 = next(rev)
-
-    # Look ahead
-    x3_peek = rev.peek_ahead(1)  # next item without moving cursor
-    x4_peek = rev.peek_ahead(2)  # two items ahead
-
-    # Look back
-    x1_again = rev.peek_back(1)  # previous item without moving cursor
-    x0_again = rev.peek_back(2)  # two items back
-
-    # Move backward
-    x1_back = rev.prev()  # back one step
-    next(rev)  # returns x2, continues forward from where we were
-    ```
-    """
-
-    __slots__ = ("_source", "_back_buf", "_ahead_buf", "_cursor", "_history", "_lookahead")
-
-    def __init__(self, iterable: Iterable[T], *, history: int = 1, lookahead: int = 0):
-        if history < 1:
-            raise ValueError("history must be >= 1")
-        if lookahead <= 0:
-            raise ValueError("lookahead must be > 0")
-
-        self._source: Iterator[T] = iter(iterable)
-        self._back_buf: Deque[T] = deque(maxlen=history)
-        self._ahead_buf: Deque[T] = deque(maxlen=lookahead) if lookahead > 0 else deque()
-        self._cursor: int = 0
-        self._history = history
-        self._lookahead = lookahead
-
-    def __iter__(self) -> "Backtrackable[T]":
-        return self
-
-    def __next__(self) -> T:
-        # If we've stepped back, consume from back buffer first
-        if self._cursor < 0:  # -1 means "last item", etc.
-            self._cursor += 1
-            return self._back_buf[self._cursor]
-
-        # If we have items in the ahead buffer, use them first
-        item = self._ahead_buf.popleft() if self._ahead_buf else next(self._source)
-
-        # Add current item to back buffer and reset cursor
-        self._back_buf.append(item)
-        self._cursor = 0
-        return item
-
-    def prev(self) -> T:
-        """
-        Step one item back in history and return it.
-        Raises IndexError if already at the oldest buffered item.
-        """
-        if len(self._back_buf) + self._cursor <= 1:
-            raise LookBackError("At start of history")
-
-        self._cursor -= 1
-        return self._back_buf[self._cursor]
-
-    def peek_back(self, n: int = 1) -> T:
-        """
-        Look `n` items back (n=1 == previous item) without moving the cursor.
-        """
-        if n < 0 or n + 1 > len(self._back_buf) + self._cursor:
-            raise LookBackError("peek_back distance out of range")
-
-        return self._back_buf[self._cursor - (n + 1)]
-
-    def peek_ahead(self, n: int = 1) -> T:
-        """
-        Look `n` items ahead (n=1 == next item) without moving the cursor.
-        Fills the ahead buffer if necessary.
-        """
-        if n < 1:
-            raise LookAheadError("peek_ahead distance must be 1 or more")
-        elif n > self._lookahead:
-            raise LookAheadError("peek_ahead distance exceeds lookahead limit")
-
-        # Fill ahead buffer if we don't have enough items
-        while len(self._ahead_buf) < n:
-            try:
-                item = next(self._source)
-                self._ahead_buf.append(item)
-
-            except StopIteration as err:
-                raise LookAheadError("peek_ahead: not enough items in source") from err
-
-        return self._ahead_buf[n - 1]
-
-    def history(self) -> list[T]:
-        """
-        Return a copy of the buffered history (most recent last).
-        The list length ≤ `history` argument passed at construction.
-        """
-        if self._cursor == 0:
-            return list(self._back_buf)
-
-        # When cursor<0, slice so the order remains chronological
-        return list(self._back_buf)[: self._cursor or None]
-
-    def lookahead_buffer(self) -> list[T]:
-        """
-        Return a copy of the current lookahead buffer.
-        """
-        return list(self._ahead_buf)
-
-    def can_peek_back(self, steps: int = 1) -> bool:
-        """
-        Check if we can go back `steps` items without raising an IndexError.
-        """
-        return steps <= len(self._back_buf) + self._cursor
-
-    def can_peek_ahead(self, steps: int = 1) -> bool:
-        """
-        Check if we can peek ahead `steps` items.
-        This may involve trying to fill the ahead buffer.
-        """
-        if self._lookahead > 0 and steps > self._lookahead:
-            return False
-
-        # Try to fill ahead buffer to check if we can peek that far
-        try:
-            while len(self._ahead_buf) < steps:
-                if self._lookahead > 0 and len(self._ahead_buf) >= self._lookahead:
-                    return False
-                item = next(self._source)
-                self._ahead_buf.append(item)
-            return True
-        except StopIteration:
-            return False
-
-    def reset_cursor(self) -> None:
-        """
-        Reset cursor to the most recent position (equivalent to calling next()
-        until you're back to the latest item).
-        """
-        self._cursor = 0
-
-    def clear_ahead_buffer(self) -> None:
-        """
-        Clear the ahead buffer, discarding any pre-fetched items.
-        """
-        self._ahead_buf.clear()
-
-    def switch_source_iterable(self, new_source: Iterable[T]) -> None:
-        """
-        Switch the source of the backtrackable to a new iterable, keeping the history.
-
-        This is useful when iterating over a sequence of datasets. The history from the
-        previous source is kept, but the lookahead buffer is cleared. The cursor is reset
-        to the present.
-        """
-        self._source = iter(new_source)
-        self.clear_ahead_buffer()
-        self.reset_cursor()
-
-
-def safe_shard(dataset: datasets.IterableDataset, index: int, num_shards: int) -> datasets.Dataset:
-    """
-    Safe shards the dataset.
-    """
-    shard_idx = min(dataset.num_shards, index + 1) - 1
-
-    return dataset.shard(num_shards, index=shard_idx)

src/lerobot/datasets/v2/batch_convert_dataset_v1_to_v2.py

@@ -0,0 +1,884 @@
+#!/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.
+
+"""
+This script is for internal use to convert all datasets under the 'lerobot' hub user account to v2.
+
+Note: Since the original Aloha datasets don't use shadow motors, you need to comment those out in
+lerobot/configs/robot/aloha.yaml before running this script.
+"""
+
+import traceback
+from pathlib import Path
+from textwrap import dedent
+
+from lerobot import available_datasets
+from lerobot.datasets.v2.convert_dataset_v1_to_v2 import convert_dataset
+from lerobot.robots.aloha.configuration_aloha import AlohaRobotConfig
+
+LOCAL_DIR = Path("data/")
+
+# spellchecker:off
+ALOHA_MOBILE_INFO = {
+    "robot_config": AlohaRobotConfig(),
+    "license": "mit",
+    "url": "https://mobile-aloha.github.io/",
+    "paper": "https://huggingface.co/papers/2401.02117",
+    "citation_bibtex": dedent(r"""
+        @inproceedings{fu2024mobile,
+            author    = {Fu, Zipeng and Zhao, Tony Z. and Finn, Chelsea},
+            title     = {Mobile ALOHA: Learning Bimanual Mobile Manipulation with Low-Cost Whole-Body Teleoperation},
+            booktitle = {arXiv},
+            year      = {2024},
+        }""").lstrip(),
+}
+ALOHA_STATIC_INFO = {
+    "robot_config": AlohaRobotConfig(),
+    "license": "mit",
+    "url": "https://tonyzhaozh.github.io/aloha/",
+    "paper": "https://huggingface.co/papers/2304.13705",
+    "citation_bibtex": dedent(r"""
+        @article{Zhao2023LearningFB,
+            title={Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware},
+            author={Tony Zhao and Vikash Kumar and Sergey Levine and Chelsea Finn},
+            journal={RSS},
+            year={2023},
+            volume={abs/2304.13705},
+            url={https://huggingface.co/papers/2304.13705}
+        }""").lstrip(),
+}
+PUSHT_INFO = {
+    "license": "mit",
+    "url": "https://diffusion-policy.cs.columbia.edu/",
+    "paper": "https://huggingface.co/papers/2303.04137",
+    "citation_bibtex": dedent(r"""
+        @article{chi2024diffusionpolicy,
+            author = {Cheng Chi and Zhenjia Xu and Siyuan Feng and Eric Cousineau and Yilun Du and Benjamin Burchfiel and Russ Tedrake and Shuran Song},
+            title ={Diffusion Policy: Visuomotor Policy Learning via Action Diffusion},
+            journal = {The International Journal of Robotics Research},
+            year = {2024},
+        }""").lstrip(),
+}
+XARM_INFO = {
+    "license": "mit",
+    "url": "https://www.nicklashansen.com/td-mpc/",
+    "paper": "https://huggingface.co/papers/2203.04955",
+    "citation_bibtex": dedent(r"""
+        @inproceedings{Hansen2022tdmpc,
+            title={Temporal Difference Learning for Model Predictive Control},
+            author={Nicklas Hansen and Xiaolong Wang and Hao Su},
+            booktitle={ICML},
+            year={2022}
+        }
+    """),
+}
+UNITREEH_INFO = {
+    "license": "apache-2.0",
+}
+
+DATASETS = {
+    "aloha_mobile_cabinet": {
+        "single_task": "Open the top cabinet, store the pot inside it then close the cabinet.",
+        **ALOHA_MOBILE_INFO,
+    },
+    "aloha_mobile_chair": {
+        "single_task": "Push the chairs in front of the desk to place them against it.",
+        **ALOHA_MOBILE_INFO,
+    },
+    "aloha_mobile_elevator": {
+        "single_task": "Take the elevator to the 1st floor.",
+        **ALOHA_MOBILE_INFO,
+    },
+    "aloha_mobile_shrimp": {
+        "single_task": "Sauté the raw shrimp on both sides, then serve it in the bowl.",
+        **ALOHA_MOBILE_INFO,
+    },
+    "aloha_mobile_wash_pan": {
+        "single_task": "Pick up the pan, rinse it in the sink and then place it in the drying rack.",
+        **ALOHA_MOBILE_INFO,
+    },
+    "aloha_mobile_wipe_wine": {
+        "single_task": "Pick up the wet cloth on the faucet and use it to clean the spilled wine on the table and underneath the glass.",
+        **ALOHA_MOBILE_INFO,
+    },
+    "aloha_static_battery": {
+        "single_task": "Place the battery into the slot of the remote controller.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_candy": {"single_task": "Pick up the candy and unwrap it.", **ALOHA_STATIC_INFO},
+    "aloha_static_coffee": {
+        "single_task": "Place the coffee capsule inside the capsule container, then place the cup onto the center of the cup tray, then push the 'Hot Water' and 'Travel Mug' buttons.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_coffee_new": {
+        "single_task": "Place the coffee capsule inside the capsule container, then place the cup onto the center of the cup tray.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_cups_open": {
+        "single_task": "Pick up the plastic cup and open its lid.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_fork_pick_up": {
+        "single_task": "Pick up the fork and place it on the plate.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_pingpong_test": {
+        "single_task": "Transfer one of the two balls in the right glass into the left glass, then transfer it back to the right glass.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_pro_pencil": {
+        "single_task": "Pick up the pencil with the right arm, hand it over to the left arm then place it back onto the table.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_screw_driver": {
+        "single_task": "Pick up the screwdriver with the right arm, hand it over to the left arm then place it into the cup.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_tape": {
+        "single_task": "Cut a small piece of tape from the tape dispenser then place it on the cardboard box's edge.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_thread_velcro": {
+        "single_task": "Pick up the velcro cable tie with the left arm, then insert the end of the velcro tie into the other end's loop with the right arm.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_towel": {
+        "single_task": "Pick up a piece of paper towel and place it on the spilled liquid.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_vinh_cup": {
+        "single_task": "Pick up the plastic cup with the right arm, then pop its lid open with the left arm.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_vinh_cup_left": {
+        "single_task": "Pick up the plastic cup with the left arm, then pop its lid open with the right arm.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_ziploc_slide": {"single_task": "Slide open the ziploc bag.", **ALOHA_STATIC_INFO},
+    "aloha_sim_insertion_scripted": {"single_task": "Insert the peg into the socket.", **ALOHA_STATIC_INFO},
+    "aloha_sim_insertion_scripted_image": {
+        "single_task": "Insert the peg into the socket.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_sim_insertion_human": {"single_task": "Insert the peg into the socket.", **ALOHA_STATIC_INFO},
+    "aloha_sim_insertion_human_image": {
+        "single_task": "Insert the peg into the socket.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_sim_transfer_cube_scripted": {
+        "single_task": "Pick up the cube with the right arm and transfer it to the left arm.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_sim_transfer_cube_scripted_image": {
+        "single_task": "Pick up the cube with the right arm and transfer it to the left arm.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_sim_transfer_cube_human": {
+        "single_task": "Pick up the cube with the right arm and transfer it to the left arm.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_sim_transfer_cube_human_image": {
+        "single_task": "Pick up the cube with the right arm and transfer it to the left arm.",
+        **ALOHA_STATIC_INFO,
+    },
+    "pusht": {"single_task": "Push the T-shaped block onto the T-shaped target.", **PUSHT_INFO},
+    "pusht_image": {"single_task": "Push the T-shaped block onto the T-shaped target.", **PUSHT_INFO},
+    "unitreeh1_fold_clothes": {"single_task": "Fold the sweatshirt.", **UNITREEH_INFO},
+    "unitreeh1_rearrange_objects": {"single_task": "Put the object into the bin.", **UNITREEH_INFO},
+    "unitreeh1_two_robot_greeting": {
+        "single_task": "Greet the other robot with a high five.",
+        **UNITREEH_INFO,
+    },
+    "unitreeh1_warehouse": {
+        "single_task": "Grab the spray paint on the shelf and place it in the bin on top of the robot dog.",
+        **UNITREEH_INFO,
+    },
+    "xarm_lift_medium": {"single_task": "Pick up the cube and lift it.", **XARM_INFO},
+    "xarm_lift_medium_image": {"single_task": "Pick up the cube and lift it.", **XARM_INFO},
+    "xarm_lift_medium_replay": {"single_task": "Pick up the cube and lift it.", **XARM_INFO},
+    "xarm_lift_medium_replay_image": {"single_task": "Pick up the cube and lift it.", **XARM_INFO},
+    "xarm_push_medium": {"single_task": "Push the cube onto the target.", **XARM_INFO},
+    "xarm_push_medium_image": {"single_task": "Push the cube onto the target.", **XARM_INFO},
+    "xarm_push_medium_replay": {"single_task": "Push the cube onto the target.", **XARM_INFO},
+    "xarm_push_medium_replay_image": {"single_task": "Push the cube onto the target.", **XARM_INFO},
+    "umi_cup_in_the_wild": {
+        "single_task": "Put the cup on the plate.",
+        "license": "apache-2.0",
+    },
+    "asu_table_top": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "paper": "https://link.springer.com/article/10.1007/s10514-023-10129-1",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{zhou2023modularity,
+                title={Modularity through Attention: Efficient Training and Transfer of Language-Conditioned Policies for Robot Manipulation},
+                author={Zhou, Yifan and Sonawani, Shubham and Phielipp, Mariano and Stepputtis, Simon and Amor, Heni},
+                booktitle={Conference on Robot Learning},
+                pages={1684--1695},
+                year={2023},
+                organization={PMLR}
+            }
+            @article{zhou2023learning,
+                title={Learning modular language-conditioned robot policies through attention},
+                author={Zhou, Yifan and Sonawani, Shubham and Phielipp, Mariano and Ben Amor, Heni and Stepputtis, Simon},
+                journal={Autonomous Robots},
+                pages={1--21},
+                year={2023},
+                publisher={Springer}
+            }""").lstrip(),
+    },
+    "austin_buds_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://ut-austin-rpl.github.io/BUDS-website/",
+        "paper": "https://huggingface.co/papers/2109.13841",
+        "citation_bibtex": dedent(r"""
+            @article{zhu2022bottom,
+                title={Bottom-Up Skill Discovery From Unsegmented Demonstrations for Long-Horizon Robot Manipulation},
+                author={Zhu, Yifeng and Stone, Peter and Zhu, Yuke},
+                journal={IEEE Robotics and Automation Letters},
+                volume={7},
+                number={2},
+                pages={4126--4133},
+                year={2022},
+                publisher={IEEE}
+            }""").lstrip(),
+    },
+    "austin_sailor_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://ut-austin-rpl.github.io/sailor/",
+        "paper": "https://huggingface.co/papers/2210.11435",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{nasiriany2022sailor,
+                title={Learning and Retrieval from Prior Data for Skill-based Imitation Learning},
+                author={Soroush Nasiriany and Tian Gao and Ajay Mandlekar and Yuke Zhu},
+                booktitle={Conference on Robot Learning (CoRL)},
+                year={2022}
+            }""").lstrip(),
+    },
+    "austin_sirius_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://ut-austin-rpl.github.io/sirius/",
+        "paper": "https://huggingface.co/papers/2211.08416",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{liu2022robot,
+                title = {Robot Learning on the Job: Human-in-the-Loop Autonomy and Learning During Deployment},
+                author = {Huihan Liu and Soroush Nasiriany and Lance Zhang and Zhiyao Bao and Yuke Zhu},
+                booktitle = {Robotics: Science and Systems (RSS)},
+                year = {2023}
+            }""").lstrip(),
+    },
+    "berkeley_autolab_ur5": {
+        "tasks_col": "language_instruction",
+        "license": "cc-by-4.0",
+        "url": "https://sites.google.com/view/berkeley-ur5/home",
+        "citation_bibtex": dedent(r"""
+            @misc{BerkeleyUR5Website,
+                title = {Berkeley {UR5} Demonstration Dataset},
+                author = {Lawrence Yunliang Chen and Simeon Adebola and Ken Goldberg},
+                howpublished = {https://sites.google.com/view/berkeley-ur5/home},
+            }""").lstrip(),
+    },
+    "berkeley_cable_routing": {
+        "tasks_col": "language_instruction",
+        "license": "cc-by-4.0",
+        "url": "https://sites.google.com/view/cablerouting/home",
+        "paper": "https://huggingface.co/papers/2307.08927",
+        "citation_bibtex": dedent(r"""
+            @article{luo2023multistage,
+                author    = {Jianlan Luo and Charles Xu and Xinyang Geng and Gilbert Feng and Kuan Fang and Liam Tan and Stefan Schaal and Sergey Levine},
+                title     = {Multi-Stage Cable Routing through Hierarchical Imitation Learning},
+                journal   = {arXiv pre-print},
+                year      = {2023},
+                url       = {https://huggingface.co/papers/2307.08927},
+            }""").lstrip(),
+    },
+    "berkeley_fanuc_manipulation": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://sites.google.com/berkeley.edu/fanuc-manipulation",
+        "citation_bibtex": dedent(r"""
+            @article{fanuc_manipulation2023,
+                title={Fanuc Manipulation: A Dataset for Learning-based Manipulation with FANUC Mate 200iD Robot},
+                author={Zhu, Xinghao and Tian, Ran and Xu, Chenfeng and Ding, Mingyu and Zhan, Wei and Tomizuka, Masayoshi},
+                year={2023},
+            }""").lstrip(),
+    },
+    "berkeley_gnm_cory_hall": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "paper": "https://huggingface.co/papers/1709.10489",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{kahn2018self,
+                title={Self-supervised deep reinforcement learning with generalized computation graphs for robot navigation},
+                author={Kahn, Gregory and Villaflor, Adam and Ding, Bosen and Abbeel, Pieter and Levine, Sergey},
+                booktitle={2018 IEEE international conference on robotics and automation (ICRA)},
+                pages={5129--5136},
+                year={2018},
+                organization={IEEE}
+            }""").lstrip(),
+    },
+    "berkeley_gnm_recon": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://sites.google.com/view/recon-robot",
+        "paper": "https://huggingface.co/papers/2104.05859",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{shah2021rapid,
+                title={Rapid Exploration for Open-World Navigation with Latent Goal Models},
+                author={Dhruv Shah and Benjamin Eysenbach and Nicholas Rhinehart and Sergey Levine},
+                booktitle={5th Annual Conference on Robot Learning },
+                year={2021},
+                url={https://openreview.net/forum?id=d_SWJhyKfVw}
+            }""").lstrip(),
+    },
+    "berkeley_gnm_sac_son": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://sites.google.com/view/SACSoN-review",
+        "paper": "https://huggingface.co/papers/2306.01874",
+        "citation_bibtex": dedent(r"""
+            @article{hirose2023sacson,
+                title={SACSoN: Scalable Autonomous Data Collection for Social Navigation},
+                author={Hirose, Noriaki and Shah, Dhruv and Sridhar, Ajay and Levine, Sergey},
+                journal={arXiv preprint arXiv:2306.01874},
+                year={2023}
+            }""").lstrip(),
+    },
+    "berkeley_mvp": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "paper": "https://huggingface.co/papers/2203.06173",
+        "citation_bibtex": dedent(r"""
+            @InProceedings{Radosavovic2022,
+                title = {Real-World Robot Learning with Masked Visual Pre-training},
+                author = {Ilija Radosavovic and Tete Xiao and Stephen James and Pieter Abbeel and Jitendra Malik and Trevor Darrell},
+                booktitle = {CoRL},
+                year = {2022}
+            }""").lstrip(),
+    },
+    "berkeley_rpt": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "paper": "https://huggingface.co/papers/2306.10007",
+        "citation_bibtex": dedent(r"""
+            @article{Radosavovic2023,
+                title={Robot Learning with Sensorimotor Pre-training},
+                author={Ilija Radosavovic and Baifeng Shi and Letian Fu and Ken Goldberg and Trevor Darrell and Jitendra Malik},
+                year={2023},
+                journal={arXiv:2306.10007}
+            }""").lstrip(),
+    },
+    "cmu_franka_exploration_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://human-world-model.github.io/",
+        "paper": "https://huggingface.co/papers/2308.10901",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{mendonca2023structured,
+                title={Structured World Models from Human Videos},
+                author={Mendonca, Russell  and Bahl, Shikhar and Pathak, Deepak},
+                journal={RSS},
+                year={2023}
+            }""").lstrip(),
+    },
+    "cmu_play_fusion": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://play-fusion.github.io/",
+        "paper": "https://huggingface.co/papers/2312.04549",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{chen2023playfusion,
+                title={PlayFusion: Skill Acquisition via Diffusion from Language-Annotated Play},
+                author={Chen, Lili and Bahl, Shikhar and Pathak, Deepak},
+                booktitle={CoRL},
+                year={2023}
+            }""").lstrip(),
+    },
+    "cmu_stretch": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://robo-affordances.github.io/",
+        "paper": "https://huggingface.co/papers/2304.08488",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{bahl2023affordances,
+                title={Affordances from Human Videos as a Versatile Representation for Robotics},
+                author={Bahl, Shikhar and Mendonca, Russell and Chen, Lili and Jain, Unnat and Pathak, Deepak},
+                booktitle={CVPR},
+                year={2023}
+            }
+                @article{mendonca2023structured,
+                title={Structured World Models from Human Videos},
+                author={Mendonca, Russell and Bahl, Shikhar and Pathak, Deepak},
+                journal={CoRL},
+                year={2023}
+            }""").lstrip(),
+    },
+    "columbia_cairlab_pusht_real": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://diffusion-policy.cs.columbia.edu/",
+        "paper": "https://huggingface.co/papers/2303.04137",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{chi2023diffusionpolicy,
+                title={Diffusion Policy: Visuomotor Policy Learning via Action Diffusion},
+                author={Chi, Cheng and Feng, Siyuan and Du, Yilun and Xu, Zhenjia and Cousineau, Eric and Burchfiel, Benjamin and Song, Shuran},
+                booktitle={Proceedings of Robotics: Science and Systems (RSS)},
+                year={2023}
+            }""").lstrip(),
+    },
+    "conq_hose_manipulation": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://sites.google.com/view/conq-hose-manipulation-dataset/home",
+        "citation_bibtex": dedent(r"""
+            @misc{ConqHoseManipData,
+                author={Peter Mitrano and Dmitry Berenson},
+                title={Conq Hose Manipulation Dataset, v1.15.0},
+                year={2024},
+                howpublished={https://sites.google.com/view/conq-hose-manipulation-dataset}
+            }""").lstrip(),
+    },
+    "dlr_edan_shared_control": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "paper": "https://ieeexplore.ieee.org/document/9341156",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{vogel_edan_2020,
+                title = {EDAN - an EMG-Controlled Daily Assistant to Help People with Physical Disabilities},
+                language = {en},
+                booktitle = {2020 {IEEE}/{RSJ} {International} {Conference} on {Intelligent} {Robots} and {Systems} ({IROS})},
+                author = {Vogel, Jörn and Hagengruber, Annette and Iskandar, Maged and Quere, Gabriel and Leipscher, Ulrike and Bustamante, Samuel and Dietrich, Alexander and Hoeppner, Hannes and Leidner, Daniel and Albu-Schäffer, Alin},
+                year = {2020}
+            }
+            @inproceedings{quere_shared_2020,
+                address = {Paris, France},
+                title = {Shared {Control} {Templates} for {Assistive} {Robotics}},
+                language = {en},
+                booktitle = {2020 {IEEE} {International} {Conference} on {Robotics} and {Automation} ({ICRA})},
+                author = {Quere, Gabriel and Hagengruber, Annette and Iskandar, Maged and Bustamante, Samuel and Leidner, Daniel and Stulp, Freek and Vogel, Joern},
+                year = {2020},
+                pages = {7},
+            }""").lstrip(),
+    },
+    "dlr_sara_grid_clamp": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "paper": "https://www.researchsquare.com/article/rs-3289569/v1",
+        "citation_bibtex": dedent(r"""
+            @article{padalkar2023guided,
+                title={A guided reinforcement learning approach using shared control templates for learning manipulation skills in the real world},
+                author={Padalkar, Abhishek and Quere, Gabriel and Raffin, Antonin and Silv{\'e}rio, Jo{\~a}o and Stulp, Freek},
+                journal={Research square preprint rs-3289569/v1},
+                year={2023}
+            }""").lstrip(),
+    },
+    "dlr_sara_pour": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "paper": "https://elib.dlr.de/193739/1/padalkar2023rlsct.pdf",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{padalkar2023guiding,
+                title={Guiding Reinforcement Learning with Shared Control Templates},
+                author={Padalkar, Abhishek and Quere, Gabriel and Steinmetz, Franz and Raffin, Antonin and Nieuwenhuisen, Matthias and Silv{\'e}rio, Jo{\~a}o and Stulp, Freek},
+                booktitle={40th IEEE International Conference on Robotics and Automation, ICRA 2023},
+                year={2023},
+                organization={IEEE}
+            }""").lstrip(),
+    },
+    "droid_100": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://droid-dataset.github.io/",
+        "paper": "https://huggingface.co/papers/2403.12945",
+        "citation_bibtex": dedent(r"""
+            @article{khazatsky2024droid,
+                title   = {DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset},
+                author  = {Alexander Khazatsky and Karl Pertsch and Suraj Nair and Ashwin Balakrishna and Sudeep Dasari and Siddharth Karamcheti and Soroush Nasiriany and Mohan Kumar Srirama and Lawrence Yunliang Chen and Kirsty Ellis and Peter David Fagan and Joey Hejna and Masha Itkina and Marion Lepert and Yecheng Jason Ma and Patrick Tree Miller and Jimmy Wu and Suneel Belkhale and Shivin Dass and Huy Ha and Arhan Jain and Abraham Lee and Youngwoon Lee and Marius Memmel and Sungjae Park and Ilija Radosavovic and Kaiyuan Wang and Albert Zhan and Kevin Black and Cheng Chi and Kyle Beltran Hatch and Shan Lin and Jingpei Lu and Jean Mercat and Abdul Rehman and Pannag R Sanketi and Archit Sharma and Cody Simpson and Quan Vuong and Homer Rich Walke and Blake Wulfe and Ted Xiao and Jonathan Heewon Yang and Arefeh Yavary and Tony Z. Zhao and Christopher Agia and Rohan Baijal and Mateo Guaman Castro and Daphne Chen and Qiuyu Chen and Trinity Chung and Jaimyn Drake and Ethan Paul Foster and Jensen Gao and David Antonio Herrera and Minho Heo and Kyle Hsu and Jiaheng Hu and Donovon Jackson and Charlotte Le and Yunshuang Li and Kevin Lin and Roy Lin and Zehan Ma and Abhiram Maddukuri and Suvir Mirchandani and Daniel Morton and Tony Nguyen and Abigail O'Neill and Rosario Scalise and Derick Seale and Victor Son and Stephen Tian and Emi Tran and Andrew E. Wang and Yilin Wu and Annie Xie and Jingyun Yang and Patrick Yin and Yunchu Zhang and Osbert Bastani and Glen Berseth and Jeannette Bohg and Ken Goldberg and Abhinav Gupta and Abhishek Gupta and Dinesh Jayaraman and Joseph J Lim and Jitendra Malik and Roberto Martín-Martín and Subramanian Ramamoorthy and Dorsa Sadigh and Shuran Song and Jiajun Wu and Michael C. Yip and Yuke Zhu and Thomas Kollar and Sergey Levine and Chelsea Finn},
+                year    = {2024},
+            }""").lstrip(),
+    },
+    "fmb": {
+        "tasks_col": "language_instruction",
+        "license": "cc-by-4.0",
+        "url": "https://functional-manipulation-benchmark.github.io/",
+        "paper": "https://huggingface.co/papers/2401.08553",
+        "citation_bibtex": dedent(r"""
+            @article{luo2024fmb,
+                title={FMB: a Functional Manipulation Benchmark for Generalizable Robotic Learning},
+                author={Luo, Jianlan and Xu, Charles and Liu, Fangchen and Tan, Liam and Lin, Zipeng and Wu, Jeffrey and Abbeel, Pieter and Levine, Sergey},
+                journal={arXiv preprint arXiv:2401.08553},
+                year={2024}
+            }""").lstrip(),
+    },
+    "iamlab_cmu_pickup_insert": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://openreview.net/forum?id=WuBv9-IGDUA",
+        "paper": "https://huggingface.co/papers/2401.14502",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{saxena2023multiresolution,
+                title={Multi-Resolution Sensing for Real-Time Control with Vision-Language Models},
+                author={Saumya Saxena and Mohit Sharma and Oliver Kroemer},
+                booktitle={7th Annual Conference on Robot Learning},
+                year={2023},
+                url={https://openreview.net/forum?id=WuBv9-IGDUA}
+            }""").lstrip(),
+    },
+    "imperialcollege_sawyer_wrist_cam": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+    },
+    "jaco_play": {
+        "tasks_col": "language_instruction",
+        "license": "cc-by-4.0",
+        "url": "https://github.com/clvrai/clvr_jaco_play_dataset",
+        "citation_bibtex": dedent(r"""
+            @software{dass2023jacoplay,
+                author = {Dass, Shivin and Yapeter, Jullian and Zhang, Jesse and Zhang, Jiahui
+                            and Pertsch, Karl and Nikolaidis, Stefanos and Lim, Joseph J.},
+                title = {CLVR Jaco Play Dataset},
+                url = {https://github.com/clvrai/clvr_jaco_play_dataset},
+                version = {1.0.0},
+                year = {2023}
+            }""").lstrip(),
+    },
+    "kaist_nonprehensile": {
+        "tasks_col": "language_instruction",
+        "license": "cc-by-4.0",
+        "url": "https://github.com/JaeHyung-Kim/rlds_dataset_builder",
+        "citation_bibtex": dedent(r"""
+            @article{kimpre,
+                title={Pre-and post-contact policy decomposition for non-prehensile manipulation with zero-shot sim-to-real transfer},
+                author={Kim, Minchan and Han, Junhyek and Kim, Jaehyung and Kim, Beomjoon},
+                booktitle={2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
+                year={2023},
+                organization={IEEE}
+            }""").lstrip(),
+    },
+    "nyu_door_opening_surprising_effectiveness": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://jyopari.github.io/VINN/",
+        "paper": "https://huggingface.co/papers/2112.01511",
+        "citation_bibtex": dedent(r"""
+            @misc{pari2021surprising,
+                title={The Surprising Effectiveness of Representation Learning for Visual Imitation},
+                author={Jyothish Pari and Nur Muhammad Shafiullah and Sridhar Pandian Arunachalam and Lerrel Pinto},
+                year={2021},
+                eprint={2112.01511},
+                archivePrefix={arXiv},
+                primaryClass={cs.RO}
+            }""").lstrip(),
+    },
+    "nyu_franka_play_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://play-to-policy.github.io/",
+        "paper": "https://huggingface.co/papers/2210.10047",
+        "citation_bibtex": dedent(r"""
+            @article{cui2022play,
+                title   = {From Play to Policy: Conditional Behavior Generation from Uncurated Robot Data},
+                author  = {Cui, Zichen Jeff and Wang, Yibin and Shafiullah, Nur Muhammad Mahi and Pinto, Lerrel},
+                journal = {arXiv preprint arXiv:2210.10047},
+                year    = {2022}
+            }""").lstrip(),
+    },
+    "nyu_rot_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://rot-robot.github.io/",
+        "paper": "https://huggingface.co/papers/2206.15469",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{haldar2023watch,
+                title={Watch and match: Supercharging imitation with regularized optimal transport},
+                author={Haldar, Siddhant and Mathur, Vaibhav and Yarats, Denis and Pinto, Lerrel},
+                booktitle={Conference on Robot Learning},
+                pages={32--43},
+                year={2023},
+                organization={PMLR}
+            }""").lstrip(),
+    },
+    "roboturk": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://roboturk.stanford.edu/dataset_real.html",
+        "paper": "PAPER",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{mandlekar2019scaling,
+                title={Scaling robot supervision to hundreds of hours with roboturk: Robotic manipulation dataset through human reasoning and dexterity},
+                author={Mandlekar, Ajay and Booher, Jonathan and Spero, Max and Tung, Albert and Gupta, Anchit and Zhu, Yuke and Garg, Animesh and Savarese, Silvio and Fei-Fei, Li},
+                booktitle={2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
+                pages={1048--1055},
+                year={2019},
+                organization={IEEE}
+            }""").lstrip(),
+    },
+    "stanford_hydra_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://sites.google.com/view/hydra-il-2023",
+        "paper": "https://huggingface.co/papers/2306.17237",
+        "citation_bibtex": dedent(r"""
+            @article{belkhale2023hydra,
+                title={HYDRA: Hybrid Robot Actions for Imitation Learning},
+                author={Belkhale, Suneel and Cui, Yuchen and Sadigh, Dorsa},
+                journal={arxiv},
+                year={2023}
+            }""").lstrip(),
+    },
+    "stanford_kuka_multimodal_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://sites.google.com/view/visionandtouch",
+        "paper": "https://huggingface.co/papers/1810.10191",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{lee2019icra,
+                title={Making sense of vision and touch: Self-supervised learning of multimodal representations for contact-rich tasks},
+                author={Lee, Michelle A and Zhu, Yuke and Srinivasan, Krishnan and Shah, Parth and Savarese, Silvio and Fei-Fei, Li and  Garg, Animesh and Bohg, Jeannette},
+                booktitle={2019 IEEE International Conference on Robotics and Automation (ICRA)},
+                year={2019},
+                url={https://huggingface.co/papers/1810.10191}
+            }""").lstrip(),
+    },
+    "stanford_robocook": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://hshi74.github.io/robocook/",
+        "paper": "https://huggingface.co/papers/2306.14447",
+        "citation_bibtex": dedent(r"""
+            @article{shi2023robocook,
+                title={RoboCook: Long-Horizon Elasto-Plastic Object Manipulation with Diverse Tools},
+                author={Shi, Haochen and Xu, Huazhe and Clarke, Samuel and Li, Yunzhu and Wu, Jiajun},
+                journal={arXiv preprint arXiv:2306.14447},
+                year={2023}
+            }""").lstrip(),
+    },
+    "taco_play": {
+        "tasks_col": "language_instruction",
+        "license": "cc-by-4.0",
+        "url": "https://www.kaggle.com/datasets/oiermees/taco-robot",
+        "paper": "https://huggingface.co/papers/2209.08959, https://huggingface.co/papers/2210.01911",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{rosete2022tacorl,
+                author = {Erick Rosete-Beas and Oier Mees and Gabriel Kalweit and Joschka Boedecker and Wolfram Burgard},
+                title = {Latent Plans for Task Agnostic Offline Reinforcement Learning},
+                journal = {Proceedings of the 6th Conference on Robot Learning (CoRL)},
+                year = {2022}
+            }
+            @inproceedings{mees23hulc2,
+                title={Grounding  Language  with  Visual  Affordances  over  Unstructured  Data},
+                author={Oier Mees and Jessica Borja-Diaz and Wolfram Burgard},
+                booktitle = {Proceedings of the IEEE International Conference on Robotics and Automation (ICRA)},
+                year={2023},
+                address = {London, UK}
+            }""").lstrip(),
+    },
+    "tokyo_u_lsmo": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "URL",
+        "paper": "https://huggingface.co/papers/2107.05842",
+        "citation_bibtex": dedent(r"""
+            @Article{Osa22,
+                author  = {Takayuki Osa},
+                journal = {The International Journal of Robotics Research},
+                title   = {Motion Planning by Learning the Solution Manifold in Trajectory Optimization},
+                year    = {2022},
+                number  = {3},
+                pages   = {291--311},
+                volume  = {41},
+            }""").lstrip(),
+    },
+    "toto": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://toto-benchmark.org/",
+        "paper": "https://huggingface.co/papers/2306.00942",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{zhou2023train,
+                author={Zhou, Gaoyue and Dean, Victoria and Srirama, Mohan Kumar and Rajeswaran, Aravind and Pari, Jyothish and Hatch, Kyle and Jain, Aryan and Yu, Tianhe and Abbeel, Pieter and Pinto, Lerrel and Finn, Chelsea and Gupta, Abhinav},
+                booktitle={2023 IEEE International Conference on Robotics and Automation (ICRA)},
+                title={Train Offline, Test Online: A Real Robot Learning Benchmark},
+                year={2023},
+            }""").lstrip(),
+    },
+    "ucsd_kitchen_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "citation_bibtex": dedent(r"""
+            @ARTICLE{ucsd_kitchens,
+                author = {Ge Yan, Kris Wu, and Xiaolong Wang},
+                title = {{ucsd kitchens Dataset}},
+                year = {2023},
+                month = {August}
+            }""").lstrip(),
+    },
+    "ucsd_pick_and_place_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://owmcorl.github.io/#",
+        "paper": "https://huggingface.co/papers/2310.16029",
+        "citation_bibtex": dedent(r"""
+            @preprint{Feng2023Finetuning,
+                title={Finetuning Offline World Models in the Real World},
+                author={Yunhai Feng, Nicklas Hansen, Ziyan Xiong, Chandramouli Rajagopalan, Xiaolong Wang},
+                year={2023}
+            }""").lstrip(),
+    },
+    "uiuc_d3field": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://robopil.github.io/d3fields/",
+        "paper": "https://huggingface.co/papers/2309.16118",
+        "citation_bibtex": dedent(r"""
+            @article{wang2023d3field,
+                title={D^3Field: Dynamic 3D Descriptor Fields for Generalizable Robotic Manipulation},
+                author={Wang, Yixuan and Li, Zhuoran and Zhang, Mingtong and Driggs-Campbell, Katherine and Wu, Jiajun and Fei-Fei, Li and Li, Yunzhu},
+                journal={arXiv preprint arXiv:},
+                year={2023},
+            }""").lstrip(),
+    },
+    "usc_cloth_sim": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://uscresl.github.io/dmfd/",
+        "paper": "https://huggingface.co/papers/2207.10148",
+        "citation_bibtex": dedent(r"""
+            @article{salhotra2022dmfd,
+                author={Salhotra, Gautam and Liu, I-Chun Arthur and Dominguez-Kuhne, Marcus and Sukhatme, Gaurav S.},
+                journal={IEEE Robotics and Automation Letters},
+                title={Learning Deformable Object Manipulation From Expert Demonstrations},
+                year={2022},
+                volume={7},
+                number={4},
+                pages={8775-8782},
+                doi={10.1109/LRA.2022.3187843}
+            }""").lstrip(),
+    },
+    "utaustin_mutex": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://ut-austin-rpl.github.io/MUTEX/",
+        "paper": "https://huggingface.co/papers/2309.14320",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{shah2023mutex,
+                title={{MUTEX}: Learning Unified Policies from Multimodal Task Specifications},
+                author={Rutav Shah and Roberto Mart{\'\i}n-Mart{\'\i}n and Yuke Zhu},
+                booktitle={7th Annual Conference on Robot Learning},
+                year={2023},
+                url={https://openreview.net/forum?id=PwqiqaaEzJ}
+            }""").lstrip(),
+    },
+    "utokyo_pr2_opening_fridge": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "citation_bibtex": dedent(r"""
+            @misc{oh2023pr2utokyodatasets,
+                author={Jihoon Oh and Naoaki Kanazawa and Kento Kawaharazuka},
+                title={X-Embodiment U-Tokyo PR2 Datasets},
+                year={2023},
+                url={https://github.com/ojh6404/rlds_dataset_builder},
+            }""").lstrip(),
+    },
+    "utokyo_pr2_tabletop_manipulation": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "citation_bibtex": dedent(r"""
+            @misc{oh2023pr2utokyodatasets,
+                author={Jihoon Oh and Naoaki Kanazawa and Kento Kawaharazuka},
+                title={X-Embodiment U-Tokyo PR2 Datasets},
+                year={2023},
+                url={https://github.com/ojh6404/rlds_dataset_builder},
+            }""").lstrip(),
+    },
+    "utokyo_saytap": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://saytap.github.io/",
+        "paper": "https://huggingface.co/papers/2306.07580",
+        "citation_bibtex": dedent(r"""
+            @article{saytap2023,
+                author = {Yujin Tang and Wenhao Yu and Jie Tan and Heiga Zen and Aleksandra Faust and
+                Tatsuya Harada},
+                title  = {SayTap: Language to Quadrupedal Locomotion},
+                eprint = {arXiv:2306.07580},
+                url    = {https://saytap.github.io},
+                note   = {https://saytap.github.io},
+                year   = {2023}
+            }""").lstrip(),
+    },
+    "utokyo_xarm_bimanual": {
+        "tasks_col": "language_instruction",
+        "license": "cc-by-4.0",
+        "citation_bibtex": dedent(r"""
+            @misc{matsushima2023weblab,
+                title={Weblab xArm Dataset},
+                author={Tatsuya Matsushima and Hiroki Furuta and Yusuke Iwasawa and Yutaka Matsuo},
+                year={2023},
+            }""").lstrip(),
+    },
+    "utokyo_xarm_pick_and_place": {
+        "tasks_col": "language_instruction",
+        "license": "cc-by-4.0",
+        "citation_bibtex": dedent(r"""
+            @misc{matsushima2023weblab,
+                title={Weblab xArm Dataset},
+                author={Tatsuya Matsushima and Hiroki Furuta and Yusuke Iwasawa and Yutaka Matsuo},
+                year={2023},
+            }""").lstrip(),
+    },
+    "viola": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://ut-austin-rpl.github.io/VIOLA/",
+        "paper": "https://huggingface.co/papers/2210.11339",
+        "citation_bibtex": dedent(r"""
+            @article{zhu2022viola,
+                title={VIOLA: Imitation Learning for Vision-Based Manipulation with Object Proposal Priors},
+                author={Zhu, Yifeng and Joshi, Abhishek and Stone, Peter and Zhu, Yuke},
+                journal={6th Annual Conference on Robot Learning (CoRL)},
+                year={2022}
+            }""").lstrip(),
+    },
+}
+# spellchecker:on
+
+
+def batch_convert():
+    status = {}
+    logfile = LOCAL_DIR / "conversion_log.txt"
+    assert set(DATASETS) == {id_.split("/")[1] for id_ in available_datasets}
+    for num, (name, kwargs) in enumerate(DATASETS.items()):
+        repo_id = f"lerobot/{name}"
+        print(f"\nConverting {repo_id} ({num}/{len(DATASETS)})")
+        print("---------------------------------------------------------")
+        try:
+            convert_dataset(repo_id, LOCAL_DIR, **kwargs)
+            status = f"{repo_id}: success."
+            with open(logfile, "a") as file:
+                file.write(status + "\n")
+        except Exception:
+            status = f"{repo_id}: failed\n    {traceback.format_exc()}"
+            with open(logfile, "a") as file:
+                file.write(status + "\n")
+            continue
+
+
+if __name__ == "__main__":
+    batch_convert()

src/lerobot/datasets/v2/convert_dataset_v1_to_v2.py

@@ -0,0 +1,687 @@
+#!/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.
+
+"""
+This script will help you convert any LeRobot dataset already pushed to the hub from codebase version 1.6 to
+2.0. You will be required to provide the 'tasks', which is a short but accurate description in plain English
+for each of the task performed in the dataset. This will allow to easily train models with task-conditioning.
+
+We support 3 different scenarios for these tasks (see instructions below):
+    1. Single task dataset: all episodes of your dataset have the same single task.
+    2. Single task episodes: the episodes of your dataset each contain a single task but they can differ from
+      one episode to the next.
+    3. Multi task episodes: episodes of your dataset may each contain several different tasks.
+
+
+Can you can also provide a robot config .yaml file (not mandatory) to this script via the option
+'--robot-config' so that it writes information about the robot (robot type, motors names) this dataset was
+recorded with. For now, only Aloha/Koch type robots are supported with this option.
+
+
+# 1. Single task dataset
+If your dataset contains a single task, you can simply provide it directly via the CLI with the
+'--single-task' option.
+
+Examples:
+
+```bash
+python -m lerobot.datasets.v2.convert_dataset_v1_to_v2 \
+    --repo-id lerobot/aloha_sim_insertion_human_image \
+    --single-task "Insert the peg into the socket." \
+    --robot-config lerobot/configs/robot/aloha.yaml \
+    --local-dir data
+```
+
+```bash
+python -m lerobot.datasets.v2.convert_dataset_v1_to_v2 \
+    --repo-id aliberts/koch_tutorial \
+    --single-task "Pick the Lego block and drop it in the box on the right." \
+    --robot-config lerobot/configs/robot/koch.yaml \
+    --local-dir data
+```
+
+
+# 2. Single task episodes
+If your dataset is a multi-task dataset, you have two options to provide the tasks to this script:
+
+- If your dataset already contains a language instruction column in its parquet file, you can simply provide
+  this column's name with the '--tasks-col' arg.
+
+    Example:
+
+    ```bash
+    python -m lerobot.datasets.v2.convert_dataset_v1_to_v2 \
+        --repo-id lerobot/stanford_kuka_multimodal_dataset \
+        --tasks-col "language_instruction" \
+        --local-dir data
+    ```
+
+- If your dataset doesn't contain a language instruction, you should provide the path to a .json file with the
+  '--tasks-path' arg. This file should have the following structure where keys correspond to each
+  episode_index in the dataset, and values are the language instruction for that episode.
+
+    Example:
+
+    ```json
+    {
+        "0": "Do something",
+        "1": "Do something else",
+        "2": "Do something",
+        "3": "Go there",
+        ...
+    }
+    ```
+
+# 3. Multi task episodes
+If you have multiple tasks per episodes, your dataset should contain a language instruction column in its
+parquet file, and you must provide this column's name with the '--tasks-col' arg.
+
+Example:
+
+```bash
+python -m lerobot.datasets.v2.convert_dataset_v1_to_v2 \
+    --repo-id lerobot/stanford_kuka_multimodal_dataset \
+    --tasks-col "language_instruction" \
+    --local-dir data
+```
+"""
+
+import argparse
+import contextlib
+import filecmp
+import json
+import logging
+import math
+import shutil
+import subprocess
+import tempfile
+from pathlib import Path
+
+import datasets
+import pyarrow.compute as pc
+import pyarrow.parquet as pq
+import torch
+from datasets import Dataset
+from huggingface_hub import HfApi
+from huggingface_hub.errors import EntryNotFoundError, HfHubHTTPError
+from safetensors.torch import load_file
+
+from lerobot.datasets.utils import (
+    DEFAULT_CHUNK_SIZE,
+    DEFAULT_PARQUET_PATH,
+    DEFAULT_VIDEO_PATH,
+    EPISODES_PATH,
+    INFO_PATH,
+    STATS_PATH,
+    TASKS_PATH,
+    create_branch,
+    create_lerobot_dataset_card,
+    flatten_dict,
+    get_safe_version,
+    load_json,
+    unflatten_dict,
+    write_json,
+    write_jsonlines,
+)
+from lerobot.datasets.video_utils import (
+    VideoFrame,  # noqa: F401
+    get_image_pixel_channels,
+    get_video_info,
+)
+from lerobot.robots import RobotConfig
+
+V16 = "v1.6"
+V20 = "v2.0"
+
+GITATTRIBUTES_REF = "aliberts/gitattributes_reference"
+V1_VIDEO_FILE = "{video_key}_episode_{episode_index:06d}.mp4"
+V1_INFO_PATH = "meta_data/info.json"
+V1_STATS_PATH = "meta_data/stats.safetensors"
+
+
+def parse_robot_config(robot_cfg: RobotConfig) -> tuple[str, dict]:
+    if robot_cfg.type in ["aloha", "koch"]:
+        state_names = [
+            f"{arm}_{motor}" if len(robot_cfg.follower_arms) > 1 else motor
+            for arm in robot_cfg.follower_arms
+            for motor in robot_cfg.follower_arms[arm].motors
+        ]
+        action_names = [
+            # f"{arm}_{motor}" for arm in ["left", "right"] for motor in robot_cfg["leader_arms"][arm]["motors"]
+            f"{arm}_{motor}" if len(robot_cfg.leader_arms) > 1 else motor
+            for arm in robot_cfg.leader_arms
+            for motor in robot_cfg.leader_arms[arm].motors
+        ]
+    # elif robot_cfg["robot_type"] == "stretch3": TODO
+    else:
+        raise NotImplementedError(
+            "Please provide robot_config={'robot_type': ..., 'names': ...} directly to convert_dataset()."
+        )
+
+    return {
+        "robot_type": robot_cfg.type,
+        "names": {
+            "observation.state": state_names,
+            "observation.effort": state_names,
+            "action": action_names,
+        },
+    }
+
+
+def convert_stats_to_json(v1_dir: Path, v2_dir: Path) -> None:
+    safetensor_path = v1_dir / V1_STATS_PATH
+    stats = load_file(safetensor_path)
+    serialized_stats = {key: value.tolist() for key, value in stats.items()}
+    serialized_stats = unflatten_dict(serialized_stats)
+
+    json_path = v2_dir / STATS_PATH
+    json_path.parent.mkdir(exist_ok=True, parents=True)
+    with open(json_path, "w") as f:
+        json.dump(serialized_stats, f, indent=4)
+
+    # Sanity check
+    with open(json_path) as f:
+        stats_json = json.load(f)
+
+    stats_json = flatten_dict(stats_json)
+    stats_json = {key: torch.tensor(value) for key, value in stats_json.items()}
+    for key in stats:
+        torch.testing.assert_close(stats_json[key], stats[key])
+
+
+def get_features_from_hf_dataset(
+    dataset: Dataset, robot_config: RobotConfig | None = None
+) -> dict[str, list]:
+    robot_config = parse_robot_config(robot_config)
+    features = {}
+    for key, ft in dataset.features.items():
+        if isinstance(ft, datasets.Value):
+            dtype = ft.dtype
+            shape = (1,)
+            names = None
+        if isinstance(ft, datasets.Sequence):
+            assert isinstance(ft.feature, datasets.Value)
+            dtype = ft.feature.dtype
+            shape = (ft.length,)
+            motor_names = (
+                robot_config["names"][key] if robot_config else [f"motor_{i}" for i in range(ft.length)]
+            )
+            assert len(motor_names) == shape[0]
+            names = {"motors": motor_names}
+        elif isinstance(ft, datasets.Image):
+            dtype = "image"
+            image = dataset[0][key]  # Assuming first row
+            channels = get_image_pixel_channels(image)
+            shape = (image.height, image.width, channels)
+            names = ["height", "width", "channels"]
+        elif ft._type == "VideoFrame":
+            dtype = "video"
+            shape = None  # Add shape later
+            names = ["height", "width", "channels"]
+
+        features[key] = {
+            "dtype": dtype,
+            "shape": shape,
+            "names": names,
+        }
+
+    return features
+
+
+def add_task_index_by_episodes(dataset: Dataset, tasks_by_episodes: dict) -> tuple[Dataset, list[str]]:
+    df = dataset.to_pandas()
+    tasks = list(set(tasks_by_episodes.values()))
+    tasks_to_task_index = {task: task_idx for task_idx, task in enumerate(tasks)}
+    episodes_to_task_index = {ep_idx: tasks_to_task_index[task] for ep_idx, task in tasks_by_episodes.items()}
+    df["task_index"] = df["episode_index"].map(episodes_to_task_index).astype(int)
+
+    features = dataset.features
+    features["task_index"] = datasets.Value(dtype="int64")
+    dataset = Dataset.from_pandas(df, features=features, split="train")
+    return dataset, tasks
+
+
+def add_task_index_from_tasks_col(
+    dataset: Dataset, tasks_col: str
+) -> tuple[Dataset, dict[str, list[str]], list[str]]:
+    df = dataset.to_pandas()
+
+    # HACK: This is to clean some of the instructions in our version of Open X datasets
+    prefix_to_clean = "tf.Tensor(b'"
+    suffix_to_clean = "', shape=(), dtype=string)"
+    df[tasks_col] = df[tasks_col].str.removeprefix(prefix_to_clean).str.removesuffix(suffix_to_clean)
+
+    # Create task_index col
+    tasks_by_episode = df.groupby("episode_index")[tasks_col].unique().apply(lambda x: x.tolist()).to_dict()
+    tasks = df[tasks_col].unique().tolist()
+    tasks_to_task_index = {task: idx for idx, task in enumerate(tasks)}
+    df["task_index"] = df[tasks_col].map(tasks_to_task_index).astype(int)
+
+    # Build the dataset back from df
+    features = dataset.features
+    features["task_index"] = datasets.Value(dtype="int64")
+    dataset = Dataset.from_pandas(df, features=features, split="train")
+    dataset = dataset.remove_columns(tasks_col)
+
+    return dataset, tasks, tasks_by_episode
+
+
+def split_parquet_by_episodes(
+    dataset: Dataset,
+    total_episodes: int,
+    total_chunks: int,
+    output_dir: Path,
+) -> list:
+    table = dataset.data.table
+    episode_lengths = []
+    for ep_chunk in range(total_chunks):
+        ep_chunk_start = DEFAULT_CHUNK_SIZE * ep_chunk
+        ep_chunk_end = min(DEFAULT_CHUNK_SIZE * (ep_chunk + 1), total_episodes)
+        chunk_dir = "/".join(DEFAULT_PARQUET_PATH.split("/")[:-1]).format(episode_chunk=ep_chunk)
+        (output_dir / chunk_dir).mkdir(parents=True, exist_ok=True)
+        for ep_idx in range(ep_chunk_start, ep_chunk_end):
+            ep_table = table.filter(pc.equal(table["episode_index"], ep_idx))
+            episode_lengths.insert(ep_idx, len(ep_table))
+            output_file = output_dir / DEFAULT_PARQUET_PATH.format(
+                episode_chunk=ep_chunk, episode_index=ep_idx
+            )
+            pq.write_table(ep_table, output_file)
+
+    return episode_lengths
+
+
+def move_videos(
+    repo_id: str,
+    video_keys: list[str],
+    total_episodes: int,
+    total_chunks: int,
+    work_dir: Path,
+    clean_gittatributes: Path,
+    branch: str = "main",
+) -> None:
+    """
+    HACK: Since HfApi() doesn't provide a way to move files directly in a repo, this function will run git
+    commands to fetch git lfs video files references to move them into subdirectories without having to
+    actually download them.
+    """
+    _lfs_clone(repo_id, work_dir, branch)
+
+    videos_moved = False
+    video_files = [str(f.relative_to(work_dir)) for f in work_dir.glob("videos*/*.mp4")]
+    if len(video_files) == 0:
+        video_files = [str(f.relative_to(work_dir)) for f in work_dir.glob("videos*/*/*/*.mp4")]
+        videos_moved = True  # Videos have already been moved
+
+    assert len(video_files) == total_episodes * len(video_keys)
+
+    lfs_untracked_videos = _get_lfs_untracked_videos(work_dir, video_files)
+
+    current_gittatributes = work_dir / ".gitattributes"
+    if not filecmp.cmp(current_gittatributes, clean_gittatributes, shallow=False):
+        fix_gitattributes(work_dir, current_gittatributes, clean_gittatributes)
+
+    if lfs_untracked_videos:
+        fix_lfs_video_files_tracking(work_dir, video_files)
+
+    if videos_moved:
+        return
+
+    video_dirs = sorted(work_dir.glob("videos*/"))
+    for ep_chunk in range(total_chunks):
+        ep_chunk_start = DEFAULT_CHUNK_SIZE * ep_chunk
+        ep_chunk_end = min(DEFAULT_CHUNK_SIZE * (ep_chunk + 1), total_episodes)
+        for vid_key in video_keys:
+            chunk_dir = "/".join(DEFAULT_VIDEO_PATH.split("/")[:-1]).format(
+                episode_chunk=ep_chunk, video_key=vid_key
+            )
+            (work_dir / chunk_dir).mkdir(parents=True, exist_ok=True)
+
+            for ep_idx in range(ep_chunk_start, ep_chunk_end):
+                target_path = DEFAULT_VIDEO_PATH.format(
+                    episode_chunk=ep_chunk, video_key=vid_key, episode_index=ep_idx
+                )
+                video_file = V1_VIDEO_FILE.format(video_key=vid_key, episode_index=ep_idx)
+                if len(video_dirs) == 1:
+                    video_path = video_dirs[0] / video_file
+                else:
+                    for dir in video_dirs:
+                        if (dir / video_file).is_file():
+                            video_path = dir / video_file
+                            break
+
+                video_path.rename(work_dir / target_path)
+
+    commit_message = "Move video files into chunk subdirectories"
+    subprocess.run(["git", "add", "."], cwd=work_dir, check=True)
+    subprocess.run(["git", "commit", "-m", commit_message], cwd=work_dir, check=True)
+    subprocess.run(["git", "push"], cwd=work_dir, check=True)
+
+
+def fix_lfs_video_files_tracking(work_dir: Path, lfs_untracked_videos: list[str]) -> None:
+    """
+    HACK: This function fixes the tracking by git lfs which was not properly set on some repos. In that case,
+    there's no other option than to download the actual files and reupload them with lfs tracking.
+    """
+    for i in range(0, len(lfs_untracked_videos), 100):
+        files = lfs_untracked_videos[i : i + 100]
+        try:
+            subprocess.run(["git", "rm", "--cached", *files], cwd=work_dir, capture_output=True, check=True)
+        except subprocess.CalledProcessError as e:
+            print("git rm --cached ERROR:")
+            print(e.stderr)
+        subprocess.run(["git", "add", *files], cwd=work_dir, check=True)
+
+    commit_message = "Track video files with git lfs"
+    subprocess.run(["git", "commit", "-m", commit_message], cwd=work_dir, check=True)
+    subprocess.run(["git", "push"], cwd=work_dir, check=True)
+
+
+def fix_gitattributes(work_dir: Path, current_gittatributes: Path, clean_gittatributes: Path) -> None:
+    shutil.copyfile(clean_gittatributes, current_gittatributes)
+    subprocess.run(["git", "add", ".gitattributes"], cwd=work_dir, check=True)
+    subprocess.run(["git", "commit", "-m", "Fix .gitattributes"], cwd=work_dir, check=True)
+    subprocess.run(["git", "push"], cwd=work_dir, check=True)
+
+
+def _lfs_clone(repo_id: str, work_dir: Path, branch: str) -> None:
+    subprocess.run(["git", "lfs", "install"], cwd=work_dir, check=True)
+    repo_url = f"https://huggingface.co/datasets/{repo_id}"
+    env = {"GIT_LFS_SKIP_SMUDGE": "1"}  # Prevent downloading LFS files
+    subprocess.run(
+        ["git", "clone", "--branch", branch, "--single-branch", "--depth", "1", repo_url, str(work_dir)],
+        check=True,
+        env=env,
+    )
+
+
+def _get_lfs_untracked_videos(work_dir: Path, video_files: list[str]) -> list[str]:
+    lfs_tracked_files = subprocess.run(
+        ["git", "lfs", "ls-files", "-n"], cwd=work_dir, capture_output=True, text=True, check=True
+    )
+    lfs_tracked_files = set(lfs_tracked_files.stdout.splitlines())
+    return [f for f in video_files if f not in lfs_tracked_files]
+
+
+def get_videos_info(repo_id: str, local_dir: Path, video_keys: list[str], branch: str) -> dict:
+    # Assumes first episode
+    video_files = [
+        DEFAULT_VIDEO_PATH.format(episode_chunk=0, video_key=vid_key, episode_index=0)
+        for vid_key in video_keys
+    ]
+    hub_api = HfApi()
+    hub_api.snapshot_download(
+        repo_id=repo_id, repo_type="dataset", local_dir=local_dir, revision=branch, allow_patterns=video_files
+    )
+    videos_info_dict = {}
+    for vid_key, vid_path in zip(video_keys, video_files, strict=True):
+        videos_info_dict[vid_key] = get_video_info(local_dir / vid_path)
+
+    return videos_info_dict
+
+
+def convert_dataset(
+    repo_id: str,
+    local_dir: Path,
+    single_task: str | None = None,
+    tasks_path: Path | None = None,
+    tasks_col: Path | None = None,
+    robot_config: RobotConfig | None = None,
+    test_branch: str | None = None,
+    **card_kwargs,
+):
+    v1 = get_safe_version(repo_id, V16)
+    v1x_dir = local_dir / V16 / repo_id
+    v20_dir = local_dir / V20 / repo_id
+    v1x_dir.mkdir(parents=True, exist_ok=True)
+    v20_dir.mkdir(parents=True, exist_ok=True)
+
+    hub_api = HfApi()
+    hub_api.snapshot_download(
+        repo_id=repo_id, repo_type="dataset", revision=v1, local_dir=v1x_dir, ignore_patterns="videos*/"
+    )
+    branch = "main"
+    if test_branch:
+        branch = test_branch
+        create_branch(repo_id=repo_id, branch=test_branch, repo_type="dataset")
+
+    metadata_v1 = load_json(v1x_dir / V1_INFO_PATH)
+    dataset = datasets.load_dataset("parquet", data_dir=v1x_dir / "data", split="train")
+    features = get_features_from_hf_dataset(dataset, robot_config)
+    video_keys = [key for key, ft in features.items() if ft["dtype"] == "video"]
+
+    if single_task and "language_instruction" in dataset.column_names:
+        logging.warning(
+            "'single_task' provided but 'language_instruction' tasks_col found. Using 'language_instruction'.",
+        )
+        single_task = None
+        tasks_col = "language_instruction"
+
+    # Episodes & chunks
+    episode_indices = sorted(dataset.unique("episode_index"))
+    total_episodes = len(episode_indices)
+    assert episode_indices == list(range(total_episodes))
+    total_videos = total_episodes * len(video_keys)
+    total_chunks = total_episodes // DEFAULT_CHUNK_SIZE
+    if total_episodes % DEFAULT_CHUNK_SIZE != 0:
+        total_chunks += 1
+
+    # Tasks
+    if single_task:
+        tasks_by_episodes = dict.fromkeys(episode_indices, single_task)
+        dataset, tasks = add_task_index_by_episodes(dataset, tasks_by_episodes)
+        tasks_by_episodes = {ep_idx: [task] for ep_idx, task in tasks_by_episodes.items()}
+    elif tasks_path:
+        tasks_by_episodes = load_json(tasks_path)
+        tasks_by_episodes = {int(ep_idx): task for ep_idx, task in tasks_by_episodes.items()}
+        dataset, tasks = add_task_index_by_episodes(dataset, tasks_by_episodes)
+        tasks_by_episodes = {ep_idx: [task] for ep_idx, task in tasks_by_episodes.items()}
+    elif tasks_col:
+        dataset, tasks, tasks_by_episodes = add_task_index_from_tasks_col(dataset, tasks_col)
+    else:
+        raise ValueError
+
+    assert set(tasks) == {task for ep_tasks in tasks_by_episodes.values() for task in ep_tasks}
+    tasks = [{"task_index": task_idx, "task": task} for task_idx, task in enumerate(tasks)]
+    write_jsonlines(tasks, v20_dir / TASKS_PATH)
+    features["task_index"] = {
+        "dtype": "int64",
+        "shape": (1,),
+        "names": None,
+    }
+
+    # Videos
+    if video_keys:
+        assert metadata_v1.get("video", False)
+        dataset = dataset.remove_columns(video_keys)
+        clean_gitattr = Path(
+            hub_api.hf_hub_download(
+                repo_id=GITATTRIBUTES_REF, repo_type="dataset", local_dir=local_dir, filename=".gitattributes"
+            )
+        ).absolute()
+        with tempfile.TemporaryDirectory() as tmp_video_dir:
+            move_videos(
+                repo_id, video_keys, total_episodes, total_chunks, Path(tmp_video_dir), clean_gitattr, branch
+            )
+        videos_info = get_videos_info(repo_id, v1x_dir, video_keys=video_keys, branch=branch)
+        for key in video_keys:
+            features[key]["shape"] = (
+                videos_info[key].pop("video.height"),
+                videos_info[key].pop("video.width"),
+                videos_info[key].pop("video.channels"),
+            )
+            features[key]["video_info"] = videos_info[key]
+            assert math.isclose(videos_info[key]["video.fps"], metadata_v1["fps"], rel_tol=1e-3)
+            if "encoding" in metadata_v1:
+                assert videos_info[key]["video.pix_fmt"] == metadata_v1["encoding"]["pix_fmt"]
+    else:
+        assert metadata_v1.get("video", 0) == 0
+        videos_info = None
+
+    # Split data into 1 parquet file by episode
+    episode_lengths = split_parquet_by_episodes(dataset, total_episodes, total_chunks, v20_dir)
+
+    if robot_config is not None:
+        robot_type = robot_config.type
+        repo_tags = [robot_type]
+    else:
+        robot_type = "unknown"
+        repo_tags = None
+
+    # Episodes
+    episodes = [
+        {"episode_index": ep_idx, "tasks": tasks_by_episodes[ep_idx], "length": episode_lengths[ep_idx]}
+        for ep_idx in episode_indices
+    ]
+    write_jsonlines(episodes, v20_dir / EPISODES_PATH)
+
+    # Assemble metadata v2.0
+    metadata_v2_0 = {
+        "codebase_version": V20,
+        "robot_type": robot_type,
+        "total_episodes": total_episodes,
+        "total_frames": len(dataset),
+        "total_tasks": len(tasks),
+        "total_videos": total_videos,
+        "total_chunks": total_chunks,
+        "chunks_size": DEFAULT_CHUNK_SIZE,
+        "fps": metadata_v1["fps"],
+        "splits": {"train": f"0:{total_episodes}"},
+        "data_path": DEFAULT_PARQUET_PATH,
+        "video_path": DEFAULT_VIDEO_PATH if video_keys else None,
+        "features": features,
+    }
+    write_json(metadata_v2_0, v20_dir / INFO_PATH)
+    convert_stats_to_json(v1x_dir, v20_dir)
+    card = create_lerobot_dataset_card(tags=repo_tags, dataset_info=metadata_v2_0, **card_kwargs)
+
+    with contextlib.suppress(EntryNotFoundError, HfHubHTTPError):
+        hub_api.delete_folder(repo_id=repo_id, path_in_repo="data", repo_type="dataset", revision=branch)
+
+    with contextlib.suppress(EntryNotFoundError, HfHubHTTPError):
+        hub_api.delete_folder(repo_id=repo_id, path_in_repo="meta_data", repo_type="dataset", revision=branch)
+
+    with contextlib.suppress(EntryNotFoundError, HfHubHTTPError):
+        hub_api.delete_folder(repo_id=repo_id, path_in_repo="meta", repo_type="dataset", revision=branch)
+
+    hub_api.upload_folder(
+        repo_id=repo_id,
+        path_in_repo="data",
+        folder_path=v20_dir / "data",
+        repo_type="dataset",
+        revision=branch,
+    )
+    hub_api.upload_folder(
+        repo_id=repo_id,
+        path_in_repo="meta",
+        folder_path=v20_dir / "meta",
+        repo_type="dataset",
+        revision=branch,
+    )
+
+    card.push_to_hub(repo_id=repo_id, repo_type="dataset", revision=branch)
+
+    if not test_branch:
+        create_branch(repo_id=repo_id, branch=V20, repo_type="dataset")
+
+
+def make_robot_config(robot_type: str, **kwargs) -> RobotConfig:
+    if robot_type == "aloha":
+        raise NotImplementedError  # TODO
+
+    elif robot_type == "koch_follower":
+        from lerobot.robots.koch_follower import KochFollowerConfig
+
+        return KochFollowerConfig(**kwargs)
+    elif robot_type == "so100_follower":
+        from lerobot.robots.so100_follower import SO100FollowerConfig
+
+        return SO100FollowerConfig(**kwargs)
+    elif robot_type == "stretch":
+        from lerobot.robots.stretch3 import Stretch3RobotConfig
+
+        return Stretch3RobotConfig(**kwargs)
+    elif robot_type == "lekiwi":
+        from lerobot.robots.lekiwi import LeKiwiConfig
+
+        return LeKiwiConfig(**kwargs)
+    else:
+        raise ValueError(f"Robot type '{robot_type}' is not available.")
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    task_args = parser.add_mutually_exclusive_group(required=True)
+
+    parser.add_argument(
+        "--repo-id",
+        type=str,
+        required=True,
+        help="Repository identifier on Hugging Face: a community or a user name `/` the name of the dataset (e.g. `lerobot/pusht`, `cadene/aloha_sim_insertion_human`).",
+    )
+    task_args.add_argument(
+        "--single-task",
+        type=str,
+        help="A short but accurate description of the single task performed in the dataset.",
+    )
+    task_args.add_argument(
+        "--tasks-col",
+        type=str,
+        help="The name of the column containing language instructions",
+    )
+    task_args.add_argument(
+        "--tasks-path",
+        type=Path,
+        help="The path to a .json file containing one language instruction for each episode_index",
+    )
+    parser.add_argument(
+        "--robot",
+        type=str,
+        default=None,
+        help="Robot config used for the dataset during conversion (e.g. 'koch', 'aloha', 'so100', etc.)",
+    )
+    parser.add_argument(
+        "--local-dir",
+        type=Path,
+        default=None,
+        help="Local directory to store the dataset during conversion. Defaults to /tmp/lerobot_dataset_v2",
+    )
+    parser.add_argument(
+        "--license",
+        type=str,
+        default="apache-2.0",
+        help="Repo license. Must be one of https://huggingface.co/docs/hub/repositories-licenses. Defaults to mit.",
+    )
+    parser.add_argument(
+        "--test-branch",
+        type=str,
+        default=None,
+        help="Repo branch to test your conversion first (e.g. 'v2.0.test')",
+    )
+
+    args = parser.parse_args()
+    if not args.local_dir:
+        args.local_dir = Path("/tmp/lerobot_dataset_v2")
+
+    if args.robot is not None:
+        robot_config = make_robot_config(args.robot)
+
+    del args.robot
+
+    convert_dataset(**vars(args), robot_config=robot_config)
+
+
+if __name__ == "__main__":
+    main()

src/lerobot/datasets/v21/_remove_language_instruction.py

@@ -0,0 +1,87 @@
+# 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 logging
+import traceback
+from pathlib import Path
+
+from datasets import get_dataset_config_info
+from huggingface_hub import HfApi
+
+from lerobot import available_datasets
+from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
+from lerobot.datasets.utils import INFO_PATH, write_info
+from lerobot.datasets.v21.convert_dataset_v20_to_v21 import V20, SuppressWarnings
+
+LOCAL_DIR = Path("data/")
+
+hub_api = HfApi()
+
+
+def fix_dataset(repo_id: str) -> str:
+    if not hub_api.revision_exists(repo_id, V20, repo_type="dataset"):
+        return f"{repo_id}: skipped (not in {V20})."
+
+    dataset_info = get_dataset_config_info(repo_id, "default")
+    with SuppressWarnings():
+        lerobot_metadata = LeRobotDatasetMetadata(repo_id, revision=V20, force_cache_sync=True)
+
+    meta_features = {key for key, ft in lerobot_metadata.features.items() if ft["dtype"] != "video"}
+    parquet_features = set(dataset_info.features)
+
+    diff_parquet_meta = parquet_features - meta_features
+    diff_meta_parquet = meta_features - parquet_features
+
+    if diff_parquet_meta:
+        raise ValueError(f"In parquet not in info.json: {parquet_features - meta_features}")
+
+    if not diff_meta_parquet:
+        return f"{repo_id}: skipped (no diff)"
+
+    if diff_meta_parquet:
+        logging.warning(f"In info.json not in parquet: {meta_features - parquet_features}")
+        assert diff_meta_parquet == {"language_instruction"}
+        lerobot_metadata.features.pop("language_instruction")
+        write_info(lerobot_metadata.info, lerobot_metadata.root)
+        commit_info = hub_api.upload_file(
+            path_or_fileobj=lerobot_metadata.root / INFO_PATH,
+            path_in_repo=INFO_PATH,
+            repo_id=repo_id,
+            repo_type="dataset",
+            revision=V20,
+            commit_message="Remove 'language_instruction'",
+            create_pr=True,
+        )
+        return f"{repo_id}: success - PR: {commit_info.pr_url}"
+
+
+def batch_fix():
+    status = {}
+    LOCAL_DIR.mkdir(parents=True, exist_ok=True)
+    logfile = LOCAL_DIR / "fix_features_v20.txt"
+    for num, repo_id in enumerate(available_datasets):
+        print(f"\nConverting {repo_id} ({num}/{len(available_datasets)})")
+        print("---------------------------------------------------------")
+        try:
+            status = fix_dataset(repo_id)
+        except Exception:
+            status = f"{repo_id}: failed\n    {traceback.format_exc()}"
+
+        logging.info(status)
+        with open(logfile, "a") as file:
+            file.write(status + "\n")
+
+
+if __name__ == "__main__":
+    batch_fix()

src/lerobot/datasets/v21/batch_convert_dataset_v20_to_v21.py

@@ -0,0 +1,54 @@
+#!/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.
+
+"""
+This script is for internal use to convert all datasets under the 'lerobot' hub user account to v2.1.
+"""
+
+import traceback
+from pathlib import Path
+
+from huggingface_hub import HfApi
+
+from lerobot import available_datasets
+from lerobot.datasets.v21.convert_dataset_v20_to_v21 import V21, convert_dataset
+
+LOCAL_DIR = Path("data/")
+
+
+def batch_convert():
+    status = {}
+    LOCAL_DIR.mkdir(parents=True, exist_ok=True)
+    logfile = LOCAL_DIR / "conversion_log_v21.txt"
+    hub_api = HfApi()
+    for num, repo_id in enumerate(available_datasets):
+        print(f"\nConverting {repo_id} ({num}/{len(available_datasets)})")
+        print("---------------------------------------------------------")
+        try:
+            if hub_api.revision_exists(repo_id, V21, repo_type="dataset"):
+                status = f"{repo_id}: success (already in {V21})."
+            else:
+                convert_dataset(repo_id)
+                status = f"{repo_id}: success."
+        except Exception:
+            status = f"{repo_id}: failed\n    {traceback.format_exc()}"
+
+        with open(logfile, "a") as file:
+            file.write(status + "\n")
+
+
+if __name__ == "__main__":
+    batch_convert()

src/lerobot/datasets/v21/convert_dataset_v20_to_v21.py

@@ -0,0 +1,137 @@
+# 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.
+
+"""
+This script converts a LeRobot dataset already pushed to the Hub from codebase version 2.0 to 2.1.
+It downloads metadata from a SOURCE dataset repo, computes/validates per-episode stats, updates
+the codebase version in `info.json`, and uploads the result to a DESTINATION dataset repo.
+It will:
+
+- Generate per-episodes stats and writes them in `episodes_stats.jsonl`
+- Check consistency between these new stats and the old ones.
+- Remove the deprecated `stats.json`.
+- Update codebase_version in `info.json`.
+- Push this new version to the destination repo/branch and tag it with the current codebase version.
+
+Usage:
+
+```bash
+python -m lerobot.datasets.v21.convert_dataset_v20_to_v21 \
+    --source-repo-id=namespace/source_dataset \
+    --dest-repo-id=namespace/destination_dataset \
+    --branch=main
+```
+
+"""
+
+import argparse
+import logging
+
+from huggingface_hub import HfApi
+
+from lerobot.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDataset
+from lerobot.datasets.utils import EPISODES_STATS_PATH, STATS_PATH, load_stats, write_info
+from lerobot.datasets.v21.convert_stats import check_aggregate_stats, convert_stats
+
+V20 = "v2.0"
+V21 = "v2.1"
+
+
+class SuppressWarnings:
+    def __enter__(self):
+        self.previous_level = logging.getLogger().getEffectiveLevel()
+        logging.getLogger().setLevel(logging.ERROR)
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        logging.getLogger().setLevel(self.previous_level)
+
+
+def convert_dataset(
+    source_repo_id: str,
+    dest_repo_id: str,
+    branch: str | None = None,
+    num_workers: int = 4,
+):
+    # Download metadata from the source repo at v2.0
+    with SuppressWarnings():
+        dataset = LeRobotDataset(source_repo_id, revision=V20, force_cache_sync=True)
+
+    # Ensure we recompute fresh episodes stats
+    if (dataset.root / EPISODES_STATS_PATH).is_file():
+        (dataset.root / EPISODES_STATS_PATH).unlink()
+
+    # Compute and validate stats
+    convert_stats(dataset, num_workers=num_workers)
+    ref_stats = load_stats(dataset.root)
+    check_aggregate_stats(dataset, ref_stats)
+
+    # Update codebase version in info.json
+    dataset.meta.info["codebase_version"] = CODEBASE_VERSION
+    write_info(dataset.meta.info, dataset.root)
+
+    # Remove deprecated stats.json locally so it won't be uploaded
+    if (dataset.root / STATS_PATH).is_file():
+        (dataset.root / STATS_PATH).unlink()
+
+    # Push only meta/ to destination repo
+    hub_api = HfApi()
+    hub_api.create_repo(repo_id=dest_repo_id, private=False, repo_type="dataset", exist_ok=True)
+    if branch:
+        hub_api.create_branch(repo_id=dest_repo_id, branch=branch, repo_type="dataset", exist_ok=True)
+
+    hub_api.upload_folder(
+        repo_id=dest_repo_id,
+        folder_path=str(dataset.root),
+        repo_type="dataset",
+        revision=branch,
+        allow_patterns="meta/",
+    )
+
+    # Ensure old stats.json is deleted on destination
+    if hub_api.file_exists(repo_id=dest_repo_id, filename=STATS_PATH, revision=branch, repo_type="dataset"):
+        hub_api.delete_file(path_in_repo=STATS_PATH, repo_id=dest_repo_id, revision=branch, repo_type="dataset")
+
+    # Tag destination with current codebase version
+    hub_api.create_tag(dest_repo_id, tag=CODEBASE_VERSION, revision=branch, repo_type="dataset")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--source-repo-id",
+        type=str,
+        required=True,
+        help="Source dataset repo id to download from (must be v2.0).",
+    )
+    parser.add_argument(
+        "--dest-repo-id",
+        type=str,
+        required=True,
+        help="Destination dataset repo id to upload the converted metadata to.",
+    )
+    parser.add_argument(
+        "--branch",
+        type=str,
+        default=None,
+        help="Repo branch to push your dataset. Defaults to the main branch.",
+    )
+    parser.add_argument(
+        "--num-workers",
+        type=int,
+        default=4,
+        help="Number of workers for parallelizing stats compute. Defaults to 4.",
+    )
+
+    args = parser.parse_args()
+    convert_dataset(**vars(args))

src/lerobot/datasets/v21/convert_stats.py

@@ -0,0 +1,99 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from concurrent.futures import ThreadPoolExecutor, as_completed
+
+import numpy as np
+from tqdm import tqdm
+
+from lerobot.datasets.compute_stats import aggregate_stats, get_feature_stats, sample_indices
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.utils import write_episode_stats
+
+
+def sample_episode_video_frames(dataset: LeRobotDataset, episode_index: int, ft_key: str) -> np.ndarray:
+    ep_len = dataset.meta.episodes[episode_index]["length"]
+    sampled_indices = sample_indices(ep_len)
+    query_timestamps = dataset._get_query_timestamps(0.0, {ft_key: sampled_indices})
+    video_frames = dataset._query_videos(query_timestamps, episode_index)
+    return video_frames[ft_key].numpy()
+
+
+def convert_episode_stats(dataset: LeRobotDataset, ep_idx: int):
+    ep_start_idx = dataset.episode_data_index["from"][ep_idx]
+    ep_end_idx = dataset.episode_data_index["to"][ep_idx]
+    ep_data = dataset.hf_dataset.select(range(ep_start_idx, ep_end_idx))
+
+    ep_stats = {}
+    for key, ft in dataset.features.items():
+        if ft["dtype"] == "video":
+            # We sample only for videos
+            ep_ft_data = sample_episode_video_frames(dataset, ep_idx, key)
+        else:
+            ep_ft_data = np.array(ep_data[key])
+
+        axes_to_reduce = (0, 2, 3) if ft["dtype"] in ["image", "video"] else 0
+        keepdims = True if ft["dtype"] in ["image", "video"] else ep_ft_data.ndim == 1
+        ep_stats[key] = get_feature_stats(ep_ft_data, axis=axes_to_reduce, keepdims=keepdims)
+
+        if ft["dtype"] in ["image", "video"]:  # remove batch dim
+            ep_stats[key] = {
+                k: v if k == "count" else np.squeeze(v, axis=0) for k, v in ep_stats[key].items()
+            }
+
+    dataset.meta.episodes_stats[ep_idx] = ep_stats
+
+
+def convert_stats(dataset: LeRobotDataset, num_workers: int = 0):
+    assert dataset.episodes is None
+    print("Computing episodes stats")
+    total_episodes = dataset.meta.total_episodes
+    if num_workers > 0:
+        with ThreadPoolExecutor(max_workers=num_workers) as executor:
+            futures = {
+                executor.submit(convert_episode_stats, dataset, ep_idx): ep_idx
+                for ep_idx in range(total_episodes)
+            }
+            for future in tqdm(as_completed(futures), total=total_episodes):
+                future.result()
+    else:
+        for ep_idx in tqdm(range(total_episodes)):
+            convert_episode_stats(dataset, ep_idx)
+
+    for ep_idx in tqdm(range(total_episodes)):
+        write_episode_stats(ep_idx, dataset.meta.episodes_stats[ep_idx], dataset.root)
+
+
+def check_aggregate_stats(
+    dataset: LeRobotDataset,
+    reference_stats: dict[str, dict[str, np.ndarray]],
+    video_rtol_atol: tuple[float] = (1e-2, 1e-2),
+    default_rtol_atol: tuple[float] = (5e-6, 6e-5),
+):
+    """Verifies that the aggregated stats from episodes_stats are close to reference stats."""
+    agg_stats = aggregate_stats(list(dataset.meta.episodes_stats.values()))
+    for key, ft in dataset.features.items():
+        # These values might need some fine-tuning
+        if ft["dtype"] == "video":
+            # to account for image sub-sampling
+            rtol, atol = video_rtol_atol
+        else:
+            rtol, atol = default_rtol_atol
+
+        for stat, val in agg_stats[key].items():
+            if key in reference_stats and stat in reference_stats[key]:
+                err_msg = f"feature='{key}' stats='{stat}'"
+                np.testing.assert_allclose(
+                    val, reference_stats[key][stat], rtol=rtol, atol=atol, err_msg=err_msg
+                )

src/lerobot/datasets/v30/convert_dataset_v21_to_v30.py

@@ -1,500 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-"""
-This script will help you convert any LeRobot dataset already pushed to the hub from codebase version 2.1 to
-3.0. It will:
-
-- Generate per-episodes stats and writes them in `episodes_stats.jsonl`
-- Check consistency between these new stats and the old ones.
-- Remove the deprecated `stats.json`.
-- Update codebase_version in `info.json`.
-- Push this new version to the hub on the 'main' branch and tags it with "v3.0".
-
-Usage:
-
-```bash
-python src/lerobot/datasets/v30/convert_dataset_v21_to_v30.py \
-    --repo-id=lerobot/pusht
-```
-
-"""
-
-import argparse
-import shutil
-from pathlib import Path
-from typing import Any
-
-import jsonlines
-import pandas as pd
-import pyarrow as pa
-import tqdm
-from datasets import Dataset, Features, Image
-from huggingface_hub import HfApi, snapshot_download
-from requests import HTTPError
-
-from lerobot.constants import HF_LEROBOT_HOME
-from lerobot.datasets.compute_stats import aggregate_stats
-from lerobot.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDataset
-from lerobot.datasets.utils import (
-    DEFAULT_CHUNK_SIZE,
-    DEFAULT_DATA_FILE_SIZE_IN_MB,
-    DEFAULT_DATA_PATH,
-    DEFAULT_VIDEO_FILE_SIZE_IN_MB,
-    DEFAULT_VIDEO_PATH,
-    LEGACY_EPISODES_PATH,
-    LEGACY_EPISODES_STATS_PATH,
-    LEGACY_TASKS_PATH,
-    cast_stats_to_numpy,
-    flatten_dict,
-    get_parquet_file_size_in_mb,
-    get_parquet_num_frames,
-    get_video_size_in_mb,
-    load_info,
-    update_chunk_file_indices,
-    write_episodes,
-    write_info,
-    write_stats,
-    write_tasks,
-)
-from lerobot.datasets.video_utils import concatenate_video_files, get_video_duration_in_s
-
-V21 = "v2.1"
-
-
-"""
--------------------------
-OLD
-data/chunk-000/episode_000000.parquet
-
-NEW
-data/chunk-000/file_000.parquet
--------------------------
-OLD
-videos/chunk-000/CAMERA/episode_000000.mp4
-
-NEW
-videos/chunk-000/file_000.mp4
--------------------------
-OLD
-episodes.jsonl
-{"episode_index": 1, "tasks": ["Put the blue block in the green bowl"], "length": 266}
-
-NEW
-meta/episodes/chunk-000/episodes_000.parquet
-episode_index | video_chunk_index | video_file_index | data_chunk_index | data_file_index | tasks | length
--------------------------
-OLD
-tasks.jsonl
-{"task_index": 1, "task": "Put the blue block in the green bowl"}
-
-NEW
-meta/tasks/chunk-000/file_000.parquet
-task_index | task
--------------------------
-OLD
-episodes_stats.jsonl
-
-NEW
-meta/episodes_stats/chunk-000/file_000.parquet
-episode_index | mean | std | min | max
--------------------------
-UPDATE
-meta/info.json
--------------------------
-"""
-
-
-def load_jsonlines(fpath: Path) -> list[Any]:
-    with jsonlines.open(fpath, "r") as reader:
-        return list(reader)
-
-
-def legacy_load_episodes(local_dir: Path) -> dict:
-    episodes = load_jsonlines(local_dir / LEGACY_EPISODES_PATH)
-    return {item["episode_index"]: item for item in sorted(episodes, key=lambda x: x["episode_index"])}
-
-
-def legacy_load_episodes_stats(local_dir: Path) -> dict:
-    episodes_stats = load_jsonlines(local_dir / LEGACY_EPISODES_STATS_PATH)
-    return {
-        item["episode_index"]: cast_stats_to_numpy(item["stats"])
-        for item in sorted(episodes_stats, key=lambda x: x["episode_index"])
-    }
-
-
-def legacy_load_tasks(local_dir: Path) -> tuple[dict, dict]:
-    tasks = load_jsonlines(local_dir / LEGACY_TASKS_PATH)
-    tasks = {item["task_index"]: item["task"] for item in sorted(tasks, key=lambda x: x["task_index"])}
-    task_to_task_index = {task: task_index for task_index, task in tasks.items()}
-    return tasks, task_to_task_index
-
-
-def convert_tasks(root, new_root):
-    tasks, _ = legacy_load_tasks(root)
-    task_indices = tasks.keys()
-    task_strings = tasks.values()
-    df_tasks = pd.DataFrame({"task_index": task_indices}, index=task_strings)
-    write_tasks(df_tasks, new_root)
-
-
-def concat_data_files(paths_to_cat, new_root, chunk_idx, file_idx, image_keys):
-    # TODO(rcadene): to save RAM use Dataset.from_parquet(file) and concatenate_datasets
-    dataframes = [pd.read_parquet(file) for file in paths_to_cat]
-    # Concatenate all DataFrames along rows
-    concatenated_df = pd.concat(dataframes, ignore_index=True)
-
-    path = new_root / DEFAULT_DATA_PATH.format(chunk_index=chunk_idx, file_index=file_idx)
-    path.parent.mkdir(parents=True, exist_ok=True)
-
-    if len(image_keys) > 0:
-        schema = pa.Schema.from_pandas(concatenated_df)
-        features = Features.from_arrow_schema(schema)
-        for key in image_keys:
-            features[key] = Image()
-        schema = features.arrow_schema
-    else:
-        schema = None
-
-    concatenated_df.to_parquet(path, index=False, schema=schema)
-
-
-def convert_data(root: Path, new_root: Path, data_file_size_in_mb: int):
-    data_dir = root / "data"
-    ep_paths = sorted(data_dir.glob("*/*.parquet"))
-
-    image_keys = get_image_keys(root)
-
-    ep_idx = 0
-    chunk_idx = 0
-    file_idx = 0
-    size_in_mb = 0
-    num_frames = 0
-    paths_to_cat = []
-    episodes_metadata = []
-    for ep_path in ep_paths:
-        ep_size_in_mb = get_parquet_file_size_in_mb(ep_path)
-        ep_num_frames = get_parquet_num_frames(ep_path)
-        ep_metadata = {
-            "episode_index": ep_idx,
-            "data/chunk_index": chunk_idx,
-            "data/file_index": file_idx,
-            "dataset_from_index": num_frames,
-            "dataset_to_index": num_frames + ep_num_frames,
-        }
-        size_in_mb += ep_size_in_mb
-        num_frames += ep_num_frames
-        episodes_metadata.append(ep_metadata)
-        ep_idx += 1
-
-        if size_in_mb < data_file_size_in_mb:
-            paths_to_cat.append(ep_path)
-            continue
-
-        if paths_to_cat:
-            concat_data_files(paths_to_cat, new_root, chunk_idx, file_idx, image_keys)
-
-        # Reset for the next file
-        size_in_mb = ep_size_in_mb
-        num_frames = ep_num_frames
-        paths_to_cat = [ep_path]
-
-        chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, DEFAULT_CHUNK_SIZE)
-
-    # Write remaining data if any
-    if paths_to_cat:
-        concat_data_files(paths_to_cat, new_root, chunk_idx, file_idx, image_keys)
-
-    return episodes_metadata
-
-
-def get_video_keys(root):
-    info = load_info(root)
-    features = info["features"]
-    video_keys = [key for key, ft in features.items() if ft["dtype"] == "video"]
-    return video_keys
-
-
-def get_image_keys(root):
-    info = load_info(root)
-    features = info["features"]
-    image_keys = [key for key, ft in features.items() if ft["dtype"] == "image"]
-    return image_keys
-
-
-def convert_videos(root: Path, new_root: Path, video_file_size_in_mb: int):
-    video_keys = get_video_keys(root)
-    if len(video_keys) == 0:
-        return None
-
-    video_keys = sorted(video_keys)
-
-    eps_metadata_per_cam = []
-    for camera in video_keys:
-        eps_metadata = convert_videos_of_camera(root, new_root, camera, video_file_size_in_mb)
-        eps_metadata_per_cam.append(eps_metadata)
-
-    num_eps_per_cam = [len(eps_cam_map) for eps_cam_map in eps_metadata_per_cam]
-    if len(set(num_eps_per_cam)) != 1:
-        raise ValueError(f"All cams dont have same number of episodes ({num_eps_per_cam}).")
-
-    episods_metadata = []
-    num_cameras = len(video_keys)
-    num_episodes = num_eps_per_cam[0]
-    for ep_idx in range(num_episodes):
-        # Sanity check
-        ep_ids = [eps_metadata_per_cam[cam_idx][ep_idx]["episode_index"] for cam_idx in range(num_cameras)]
-        ep_ids += [ep_idx]
-        if len(set(ep_ids)) != 1:
-            raise ValueError(f"All episode indices need to match ({ep_ids}).")
-
-        ep_dict = {}
-        for cam_idx in range(num_cameras):
-            ep_dict.update(eps_metadata_per_cam[cam_idx][ep_idx])
-        episods_metadata.append(ep_dict)
-
-    return episods_metadata
-
-
-def convert_videos_of_camera(root: Path, new_root: Path, video_key: str, video_file_size_in_mb: int):
-    # Access old paths to mp4
-    videos_dir = root / "videos"
-    ep_paths = sorted(videos_dir.glob(f"*/{video_key}/*.mp4"))
-
-    ep_idx = 0
-    chunk_idx = 0
-    file_idx = 0
-    size_in_mb = 0
-    duration_in_s = 0.0
-    paths_to_cat = []
-    episodes_metadata = []
-    for ep_path in tqdm.tqdm(ep_paths, desc=f"convert videos of {video_key}"):
-        ep_size_in_mb = get_video_size_in_mb(ep_path)
-        ep_duration_in_s = get_video_duration_in_s(ep_path)
-
-        # Check if adding this episode would exceed the limit
-        if size_in_mb + ep_size_in_mb >= video_file_size_in_mb and len(paths_to_cat) > 0:
-            # Size limit would be exceeded, save current accumulation WITHOUT this episode
-            concatenate_video_files(
-                paths_to_cat,
-                new_root
-                / DEFAULT_VIDEO_PATH.format(video_key=video_key, chunk_index=chunk_idx, file_index=file_idx),
-            )
-
-            # Update episodes metadata for the file we just saved
-            for i, _ in enumerate(paths_to_cat):
-                past_ep_idx = ep_idx - len(paths_to_cat) + i
-                episodes_metadata[past_ep_idx][f"videos/{video_key}/chunk_index"] = chunk_idx
-                episodes_metadata[past_ep_idx][f"videos/{video_key}/file_index"] = file_idx
-
-            # Move to next file and start fresh with current episode
-            chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, DEFAULT_CHUNK_SIZE)
-            size_in_mb = 0
-            duration_in_s = 0.0
-            paths_to_cat = []
-
-        # Add current episode metadata
-        ep_metadata = {
-            "episode_index": ep_idx,
-            f"videos/{video_key}/chunk_index": chunk_idx,  # Will be updated when file is saved
-            f"videos/{video_key}/file_index": file_idx,  # Will be updated when file is saved
-            f"videos/{video_key}/from_timestamp": duration_in_s,
-            f"videos/{video_key}/to_timestamp": duration_in_s + ep_duration_in_s,
-        }
-        episodes_metadata.append(ep_metadata)
-
-        # Add current episode to accumulation
-        paths_to_cat.append(ep_path)
-        size_in_mb += ep_size_in_mb
-        duration_in_s += ep_duration_in_s
-        ep_idx += 1
-
-    # Write remaining videos if any
-    if paths_to_cat:
-        concatenate_video_files(
-            paths_to_cat,
-            new_root
-            / DEFAULT_VIDEO_PATH.format(video_key=video_key, chunk_index=chunk_idx, file_index=file_idx),
-        )
-
-        # Update episodes metadata for the final file
-        for i, _ in enumerate(paths_to_cat):
-            past_ep_idx = ep_idx - len(paths_to_cat) + i
-            episodes_metadata[past_ep_idx][f"videos/{video_key}/chunk_index"] = chunk_idx
-            episodes_metadata[past_ep_idx][f"videos/{video_key}/file_index"] = file_idx
-
-    return episodes_metadata
-
-
-def generate_episode_metadata_dict(
-    episodes_legacy_metadata, episodes_metadata, episodes_stats, episodes_videos=None
-):
-    num_episodes = len(episodes_metadata)
-    episodes_legacy_metadata_vals = list(episodes_legacy_metadata.values())
-    episodes_stats_vals = list(episodes_stats.values())
-    episodes_stats_keys = list(episodes_stats.keys())
-
-    for i in range(num_episodes):
-        ep_legacy_metadata = episodes_legacy_metadata_vals[i]
-        ep_metadata = episodes_metadata[i]
-        ep_stats = episodes_stats_vals[i]
-
-        ep_ids_set = {
-            ep_legacy_metadata["episode_index"],
-            ep_metadata["episode_index"],
-            episodes_stats_keys[i],
-        }
-
-        if episodes_videos is None:
-            ep_video = {}
-        else:
-            ep_video = episodes_videos[i]
-            ep_ids_set.add(ep_video["episode_index"])
-
-        if len(ep_ids_set) != 1:
-            raise ValueError(f"Number of episodes is not the same ({ep_ids_set}).")
-
-        ep_dict = {**ep_metadata, **ep_video, **ep_legacy_metadata, **flatten_dict({"stats": ep_stats})}
-        ep_dict["meta/episodes/chunk_index"] = 0
-        ep_dict["meta/episodes/file_index"] = 0
-        yield ep_dict
-
-
-def convert_episodes_metadata(root, new_root, episodes_metadata, episodes_video_metadata=None):
-    episodes_legacy_metadata = legacy_load_episodes(root)
-    episodes_stats = legacy_load_episodes_stats(root)
-
-    num_eps_set = {len(episodes_legacy_metadata), len(episodes_metadata)}
-    if episodes_video_metadata is not None:
-        num_eps_set.add(len(episodes_video_metadata))
-
-    if len(num_eps_set) != 1:
-        raise ValueError(f"Number of episodes is not the same ({num_eps_set}).")
-
-    ds_episodes = Dataset.from_generator(
-        lambda: generate_episode_metadata_dict(
-            episodes_legacy_metadata, episodes_metadata, episodes_stats, episodes_video_metadata
-        )
-    )
-    write_episodes(ds_episodes, new_root)
-
-    stats = aggregate_stats(list(episodes_stats.values()))
-    write_stats(stats, new_root)
-
-
-def convert_info(root, new_root, data_file_size_in_mb, video_file_size_in_mb):
-    info = load_info(root)
-    info["codebase_version"] = "v3.0"
-    del info["total_chunks"]
-    del info["total_videos"]
-    info["data_files_size_in_mb"] = data_file_size_in_mb
-    info["video_files_size_in_mb"] = video_file_size_in_mb
-    info["data_path"] = DEFAULT_DATA_PATH
-    info["video_path"] = DEFAULT_VIDEO_PATH
-    info["fps"] = float(info["fps"])
-    for key in info["features"]:
-        if info["features"][key]["dtype"] == "video":
-            # already has fps in video_info
-            continue
-        info["features"][key]["fps"] = info["fps"]
-    write_info(info, new_root)
-
-
-def convert_dataset(
-    repo_id: str,
-    branch: str | None = None,
-    data_file_size_in_mb: int | None = None,
-    video_file_size_in_mb: int | None = None,
-):
-    root = HF_LEROBOT_HOME / repo_id
-    old_root = HF_LEROBOT_HOME / f"{repo_id}_old"
-    new_root = HF_LEROBOT_HOME / f"{repo_id}_v30"
-
-    if data_file_size_in_mb is None:
-        data_file_size_in_mb = DEFAULT_DATA_FILE_SIZE_IN_MB
-    if video_file_size_in_mb is None:
-        video_file_size_in_mb = DEFAULT_VIDEO_FILE_SIZE_IN_MB
-
-    if old_root.is_dir() and root.is_dir():
-        shutil.rmtree(str(root))
-        shutil.move(str(old_root), str(root))
-
-    if new_root.is_dir():
-        shutil.rmtree(new_root)
-
-    snapshot_download(
-        repo_id,
-        repo_type="dataset",
-        revision=V21,
-        local_dir=root,
-    )
-
-    convert_info(root, new_root, data_file_size_in_mb, video_file_size_in_mb)
-    convert_tasks(root, new_root)
-    episodes_metadata = convert_data(root, new_root, data_file_size_in_mb)
-    episodes_videos_metadata = convert_videos(root, new_root, video_file_size_in_mb)
-    convert_episodes_metadata(root, new_root, episodes_metadata, episodes_videos_metadata)
-
-    shutil.move(str(root), str(old_root))
-    shutil.move(str(new_root), str(root))
-
-    hub_api = HfApi()
-    try:
-        hub_api.delete_tag(repo_id, tag=CODEBASE_VERSION, repo_type="dataset")
-    except HTTPError as e:
-        print(f"tag={CODEBASE_VERSION} probably doesn't exist. Skipping exception ({e})")
-        pass
-    hub_api.delete_files(
-        delete_patterns=["data/chunk*/episode_*", "meta/*.jsonl", "videos/chunk*"],
-        repo_id=repo_id,
-        revision=branch,
-        repo_type="dataset",
-    )
-    hub_api.create_tag(repo_id, tag=CODEBASE_VERSION, revision=branch, repo_type="dataset")
-
-    LeRobotDataset(repo_id).push_to_hub()
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser.add_argument(
-        "--repo-id",
-        type=str,
-        required=True,
-        help="Repository identifier on Hugging Face: a community or a user name `/` the name of the dataset "
-        "(e.g. `lerobot/pusht`, `cadene/aloha_sim_insertion_human`).",
-    )
-    parser.add_argument(
-        "--branch",
-        type=str,
-        default=None,
-        help="Repo branch to push your dataset. Defaults to the main branch.",
-    )
-    parser.add_argument(
-        "--data-file-size-in-mb",
-        type=int,
-        default=None,
-        help="File size in MB. Defaults to 100 for data and 500 for videos.",
-    )
-    parser.add_argument(
-        "--video-file-size-in-mb",
-        type=int,
-        default=None,
-        help="File size in MB. Defaults to 100 for data and 500 for videos.",
-    )
-
-    args = parser.parse_args()
-    convert_dataset(**vars(args))

src/lerobot/datasets/video_utils.py

@@ -17,15 +17,12 @@ import glob
 import importlib
 import logging
 import shutil
-import tempfile
 import warnings
 from dataclasses import dataclass, field
 from pathlib import Path
-from threading import Lock
 from typing import Any, ClassVar
 
 import av
-import fsspec
 import pyarrow as pa
 import torch
 import torchvision
@@ -171,68 +168,15 @@ def decode_video_frames_torchvision(
     return closest_frames
 
 
-class VideoDecoderCache:
-    """Thread-safe cache for video decoders to avoid expensive re-initialization."""
-
-    def __init__(self):
-        self._cache: dict[str, tuple[Any, Any]] = {}
-        self._lock = Lock()
-
-    def get_decoder(self, video_path: str):
-        """Get a cached decoder or create a new one."""
-        if importlib.util.find_spec("torchcodec"):
-            from torchcodec.decoders import VideoDecoder
-        else:
-            raise ImportError("torchcodec is required but not available.")
-
-        video_path = str(video_path)
-
-        with self._lock:
-            if video_path not in self._cache:
-                file_handle = fsspec.open(video_path).__enter__()
-                decoder = VideoDecoder(file_handle, seek_mode="approximate")
-                self._cache[video_path] = (decoder, file_handle)
-
-            return self._cache[video_path][0]
-
-    def clear(self):
-        """Clear the cache and close file handles."""
-        with self._lock:
-            for _, file_handle in self._cache.values():
-                file_handle.close()
-            self._cache.clear()
-
-    def size(self) -> int:
-        """Return the number of cached decoders."""
-        with self._lock:
-            return len(self._cache)
-
-
-class FrameTimestampError(ValueError):
-    """Helper error to indicate the retrieved timestamps exceed the queried ones"""
-
-    pass
-
-
-_default_decoder_cache = VideoDecoderCache()
-
-
 def decode_video_frames_torchcodec(
     video_path: Path | str,
     timestamps: list[float],
     tolerance_s: float,
+    device: str = "cpu",
     log_loaded_timestamps: bool = False,
-    decoder_cache: VideoDecoderCache | None = None,
 ) -> torch.Tensor:
     """Loads frames associated with the requested timestamps of a video using torchcodec.
 
-    Args:
-        video_path: Path to the video file.
-        timestamps: List of timestamps to extract frames.
-        tolerance_s: Allowed deviation in seconds for frame retrieval.
-        log_loaded_timestamps: Whether to log loaded timestamps.
-        decoder_cache: Optional decoder cache instance. Uses default if None.
-
     Note: Setting device="cuda" outside the main process, e.g. in data loader workers, will lead to CUDA initialization errors.
 
     Note: Video benefits from inter-frame compression. Instead of storing every frame individually,
@@ -241,24 +185,27 @@ def decode_video_frames_torchcodec(
     and all subsequent frames until reaching the requested frame. The number of key frames in a video
     can be adjusted during encoding to take into account decoding time and video size in bytes.
     """
-    if decoder_cache is None:
-        decoder_cache = _default_decoder_cache
 
-    # Use cached decoder instead of creating new one each time
-    decoder = decoder_cache.get_decoder(str(video_path))
+    if importlib.util.find_spec("torchcodec"):
+        from torchcodec.decoders import VideoDecoder
+    else:
+        raise ImportError("torchcodec is required but not available.")
 
-    loaded_ts = []
+    # initialize video decoder
+    decoder = VideoDecoder(video_path, device=device, seek_mode="approximate")
     loaded_frames = []
-
+    loaded_ts = []
     # get metadata for frame information
     metadata = decoder.metadata
     average_fps = metadata.average_fps
+
     # convert timestamps to frame indices
     frame_indices = [round(ts * average_fps) for ts in timestamps]
+
     # retrieve frames based on indices
     frames_batch = decoder.get_frames_at(indices=frame_indices)
 
-    for frame, pts in zip(frames_batch.data, frames_batch.pts_seconds, strict=True):
+    for frame, pts in zip(frames_batch.data, frames_batch.pts_seconds, strict=False):
         loaded_frames.append(frame)
         loaded_ts.append(pts.item())
         if log_loaded_timestamps:
@@ -289,14 +236,10 @@ def decode_video_frames_torchcodec(
     if log_loaded_timestamps:
         logging.info(f"{closest_ts=}")
 
-    # convert to float32 in [0,1] range
-    closest_frames = (closest_frames / 255.0).type(torch.float32)
-
-    if not len(timestamps) == len(closest_frames):
-        raise FrameTimestampError(
-            f"Retrieved timestamps differ from queried {set(closest_frames) - set(timestamps)}"
-        )
+    # convert to float32 in [0,1] range (channel first)
+    closest_frames = closest_frames.type(torch.float32) / 255
 
+    assert len(timestamps) == len(closest_frames)
     return closest_frames
 
 
@@ -320,11 +263,7 @@ def encode_video_frames(
     video_path = Path(video_path)
     imgs_dir = Path(imgs_dir)
 
-    if video_path.exists() and not overwrite:
-        logging.warning(f"Video file already exists: {video_path}. Skipping encoding.")
-        return
-
-    video_path.parent.mkdir(parents=True, exist_ok=True)
+    video_path.parent.mkdir(parents=True, exist_ok=overwrite)
 
     # Encoders/pixel formats incompatibility check
     if (vcodec == "libsvtav1" or vcodec == "hevc") and pix_fmt == "yuv444p":
@@ -334,9 +273,9 @@ def encode_video_frames(
         pix_fmt = "yuv420p"
 
     # Get input frames
-    template = "frame-" + ("[0-9]" * 6) + ".png"
+    template = "frame_" + ("[0-9]" * 6) + ".png"
     input_list = sorted(
-        glob.glob(str(imgs_dir / template)), key=lambda x: int(x.split("-")[-1].split(".")[0])
+        glob.glob(str(imgs_dir / template)), key=lambda x: int(x.split("_")[-1].split(".")[0])
     )
 
     # Define video output frame size (assuming all input frames are the same size)
@@ -361,7 +300,7 @@ def encode_video_frames(
 
     # Set logging level
     if log_level is not None:
-        # "While less efficient, it is generally preferable to modify logging with Python's logging"
+        # "While less efficient, it is generally preferable to modify logging with Python’s logging"
         logging.getLogger("libav").setLevel(log_level)
 
     # Create and open output file (overwrite by default)
@@ -392,89 +331,6 @@ def encode_video_frames(
         raise OSError(f"Video encoding did not work. File not found: {video_path}.")
 
 
-def concatenate_video_files(
-    input_video_paths: list[Path | str], output_video_path: Path, overwrite: bool = True
-):
-    """
-    Concatenate multiple video files into a single video file using pyav.
-
-    This function takes a list of video input file paths and concatenates them into a single
-    output video file. It uses ffmpeg's concat demuxer with stream copy mode for fast
-    concatenation without re-encoding.
-
-    Args:
-        input_video_paths: Ordered list of input video file paths to concatenate.
-        output_video_path: Path to the output video file.
-        overwrite: Whether to overwrite the output video file if it already exists. Default is True.
-
-    Note:
-        - Creates a temporary directory for intermediate files that is cleaned up after use.
-        - Uses ffmpeg's concat demuxer which requires all input videos to have the same
-          codec, resolution, and frame rate for proper concatenation.
-    """
-
-    output_video_path = Path(output_video_path)
-
-    if output_video_path.exists() and not overwrite:
-        logging.warning(f"Video file already exists: {output_video_path}. Skipping concatenation.")
-        return
-
-    output_video_path.parent.mkdir(parents=True, exist_ok=True)
-
-    if len(input_video_paths) == 0:
-        raise FileNotFoundError("No input video paths provided.")
-
-    # Create a temporary .ffconcat file to list the input video paths
-    with tempfile.NamedTemporaryFile(mode="w", suffix=".ffconcat", delete=False) as tmp_concatenate_file:
-        tmp_concatenate_file.write("ffconcat version 1.0\n")
-        for input_path in input_video_paths:
-            tmp_concatenate_file.write(f"file '{str(input_path)}'\n")
-        tmp_concatenate_file.flush()
-        tmp_concatenate_path = tmp_concatenate_file.name
-
-    # Create input and output containers
-    input_container = av.open(
-        tmp_concatenate_path, mode="r", format="concat", options={"safe": "0"}
-    )  # safe = 0 allows absolute paths as well as relative paths
-
-    tmp_output_video_path = tempfile.NamedTemporaryFile(suffix=".mp4", delete=False).name
-    output_container = av.open(
-        tmp_output_video_path, mode="w", options={"movflags": "faststart"}
-    )  # faststart is to move the metadata to the beginning of the file to speed up loading
-
-    # Replicate input streams in output container
-    stream_map = {}
-    for input_stream in input_container.streams:
-        if input_stream.type in ("video", "audio", "subtitle"):  # only copy compatible streams
-            stream_map[input_stream.index] = output_container.add_stream_from_template(
-                template=input_stream, opaque=True
-            )
-            stream_map[
-                input_stream.index
-            ].time_base = (
-                input_stream.time_base
-            )  # set the time base to the input stream time base (missing in the codec context)
-
-    # Demux + remux packets (no re-encode)
-    for packet in input_container.demux():
-        # Skip packets from un-mapped streams
-        if packet.stream.index not in stream_map:
-            continue
-
-        # Skip demux flushing packets
-        if packet.dts is None:
-            continue
-
-        output_stream = stream_map[packet.stream.index]
-        packet.stream = output_stream
-        output_container.mux(packet)
-
-    input_container.close()
-    output_container.close()
-    shutil.move(tmp_output_video_path, output_video_path)
-    Path(tmp_concatenate_path).unlink()
-
-
 @dataclass
 class VideoFrame:
     # TODO(rcadene, lhoestq): move to Hugging Face `datasets` repo
@@ -598,28 +454,6 @@ def get_image_pixel_channels(image: Image):
         raise ValueError("Unknown format")
 
 
-def get_video_duration_in_s(video_path: Path | str) -> float:
-    """
-    Get the duration of a video file in seconds using PyAV.
-
-    Args:
-        video_path: Path to the video file.
-
-    Returns:
-        Duration of the video in seconds.
-    """
-    with av.open(str(video_path)) as container:
-        # Get the first video stream
-        video_stream = container.streams.video[0]
-        # Calculate duration: stream.duration * stream.time_base gives duration in seconds
-        if video_stream.duration is not None:
-            duration = float(video_stream.duration * video_stream.time_base)
-        else:
-            # Fallback to container duration if stream duration is not available
-            duration = float(container.duration / av.time_base)
-    return duration
-
-
 class VideoEncodingManager:
     """
     Context manager that ensures proper video encoding and data cleanup even if exceptions occur.
@@ -653,7 +487,7 @@ class VideoEncodingManager:
                 f"Encoding remaining {self.dataset.episodes_since_last_encoding} episodes, "
                 f"from episode {start_ep} to {end_ep - 1}"
             )
-            self.dataset._batch_save_episode_video(start_ep, end_ep)
+            self.dataset.batch_encode_videos(start_ep, end_ep)
 
         # Clean up episode images if recording was interrupted
         if exc_type is not None:

src/lerobot/find_cameras.py

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

src/lerobot/find_port.py

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

src/lerobot/motors/dynamixel/tables.py

@@ -107,8 +107,6 @@ 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
-    lerobot-find-port.py
+    python -m 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 `lerobot-find-port`\n"
+                "\nTry running `python -m lerobot.find_port`\n"
             ) from e
 
     @abc.abstractmethod

src/lerobot/policies/__init__.py

@@ -15,6 +15,19 @@
 from .act.configuration_act import ACTConfig as ACTConfig
 from .diffusion.configuration_diffusion import DiffusionConfig as DiffusionConfig
 from .pi0.configuration_pi0 import PI0Config as PI0Config
+from .pi0.processor_pi0 import Pi0NewLineProcessor
+from .rlearn.configuration_rlearn import RLearNConfig as RLearNConfig
 from .smolvla.configuration_smolvla import SmolVLAConfig as SmolVLAConfig
+from .smolvla.processor_smolvla import SmolVLANewLineProcessor
 from .tdmpc.configuration_tdmpc import TDMPCConfig as TDMPCConfig
 from .vqbet.configuration_vqbet import VQBeTConfig as VQBeTConfig
+
+__all__ = [
+    "ACTConfig",
+    "DiffusionConfig",
+    "PI0Config",
+    "SmolVLAConfig",
+    "TDMPCConfig",
+    "VQBeTConfig",
+    "RLearNConfig",
+]

src/lerobot/policies/act/modeling_act.py

@@ -35,7 +35,6 @@ from torchvision.ops.misc import FrozenBatchNorm2d
 
 from lerobot.constants import ACTION, OBS_IMAGES
 from lerobot.policies.act.configuration_act import ACTConfig
-from lerobot.policies.normalize import Normalize, Unnormalize
 from lerobot.policies.pretrained import PreTrainedPolicy
 
 
@@ -51,27 +50,16 @@ class ACTPolicy(PreTrainedPolicy):
     def __init__(
         self,
         config: ACTConfig,
-        dataset_stats: dict[str, dict[str, Tensor]] | None = None,
     ):
         """
         Args:
             config: Policy configuration class instance or None, in which case the default instantiation of
                     the configuration class is used.
-            dataset_stats: Dataset statistics to be used for normalization. If not passed here, it is expected
-                that they will be passed with a call to `load_state_dict` before the policy is used.
         """
         super().__init__(config)
         config.validate_features()
         self.config = config
 
-        self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
-        self.normalize_targets = Normalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-        self.unnormalize_outputs = Unnormalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-
         self.model = ACT(config)
 
         if config.temporal_ensemble_coeff is not None:
@@ -137,23 +125,19 @@ class ACTPolicy(PreTrainedPolicy):
         """Predict a chunk of actions given environment observations."""
         self.eval()
 
-        batch = self.normalize_inputs(batch)
         if self.config.image_features:
             batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
             batch[OBS_IMAGES] = [batch[key] for key in self.config.image_features]
 
         actions = self.model(batch)[0]
-        actions = self.unnormalize_outputs({ACTION: actions})[ACTION]
         return actions
 
     def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict]:
         """Run the batch through the model and compute the loss for training or validation."""
-        batch = self.normalize_inputs(batch)
         if self.config.image_features:
             batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
             batch[OBS_IMAGES] = [batch[key] for key in self.config.image_features]
 
-        batch = self.normalize_targets(batch)
         actions_hat, (mu_hat, log_sigma_x2_hat) = self.model(batch)
 
         l1_loss = (
@@ -303,7 +287,7 @@ class ACT(nn.Module):
                                 └───────────────────────┘
     """
 
-    def __init__(self, config: ACTConfig):
+    def __init__(self, config: ACTConfig, dataset_stats=None):
         # 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

@@ -0,0 +1,50 @@
+#!/usr/bin/env python
+
+# Copyright 2024 Tony Z. Zhao and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import torch
+
+from lerobot.policies.act.configuration_act import ACTConfig
+from lerobot.processor import (
+    DeviceProcessor,
+    NormalizerProcessor,
+    RenameProcessor,
+    RobotProcessor,
+    ToBatchProcessor,
+    UnnormalizerProcessor,
+)
+
+
+def make_act_processor(
+    config: ACTConfig, dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None
+) -> tuple[RobotProcessor, RobotProcessor]:
+    input_steps = [
+        RenameProcessor(rename_map={}),
+        NormalizerProcessor(
+            features={**config.input_features, **config.output_features},
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+        ),
+        ToBatchProcessor(),
+        DeviceProcessor(device=config.device),
+    ]
+    output_steps = [
+        DeviceProcessor(device="cpu"),
+        UnnormalizerProcessor(
+            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"
+    )

src/lerobot/policies/diffusion/modeling_diffusion.py

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

src/lerobot/policies/diffusion/processor_diffusion.py

@@ -0,0 +1,51 @@
+#!/usr/bin/env python
+
+# Copyright 2024 Columbia Artificial Intelligence, Robotics Lab,
+# and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import torch
+
+from lerobot.policies.diffusion.configuration_diffusion import DiffusionConfig
+from lerobot.processor import (
+    DeviceProcessor,
+    NormalizerProcessor,
+    RenameProcessor,
+    RobotProcessor,
+    ToBatchProcessor,
+    UnnormalizerProcessor,
+)
+
+
+def make_diffusion_processor(
+    config: DiffusionConfig, dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None
+) -> tuple[RobotProcessor, RobotProcessor]:
+    input_steps = [
+        RenameProcessor(rename_map={}),
+        NormalizerProcessor(
+            features={**config.input_features, **config.output_features},
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+        ),
+        ToBatchProcessor(),
+        DeviceProcessor(device=config.device),
+    ]
+    output_steps = [
+        DeviceProcessor(device="cpu"),
+        UnnormalizerProcessor(
+            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"
+    )

src/lerobot/policies/factory.py

@@ -14,9 +14,14 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import logging
+from __future__ import annotations
 
+import logging
+from typing import Any, TypedDict, cast
+
+import torch
 from torch import nn
+from typing_extensions import Unpack
 
 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.configs.types import FeatureType
@@ -29,14 +34,16 @@ from lerobot.policies.diffusion.configuration_diffusion import DiffusionConfig
 from lerobot.policies.pi0.configuration_pi0 import PI0Config
 from lerobot.policies.pi0fast.configuration_pi0fast import PI0FASTConfig
 from lerobot.policies.pretrained import PreTrainedPolicy
+from lerobot.policies.rlearn.configuration_rlearn import RLearNConfig
 from lerobot.policies.sac.configuration_sac import SACConfig
 from lerobot.policies.sac.reward_model.configuration_classifier import RewardClassifierConfig
 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
 
 
-def get_policy_class(name: str) -> PreTrainedPolicy:
+def get_policy_class(name: str) -> type[PreTrainedPolicy]:
     """Get the policy's class and config class given a name (matching the policy class' `name` attribute)."""
     if name == "tdmpc":
         from lerobot.policies.tdmpc.modeling_tdmpc import TDMPCPolicy
@@ -74,6 +81,10 @@ def get_policy_class(name: str) -> PreTrainedPolicy:
         from lerobot.policies.smolvla.modeling_smolvla import SmolVLAPolicy
 
         return SmolVLAPolicy
+    elif name == "rlearn":
+        from lerobot.policies.rlearn.modeling_rlearn import RLearNPolicy
+
+        return RLearNPolicy
     else:
         raise NotImplementedError(f"Policy with name {name} is not implemented.")
 
@@ -97,14 +108,143 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
         return SmolVLAConfig(**kwargs)
     elif policy_type == "reward_classifier":
         return RewardClassifierConfig(**kwargs)
+    elif policy_type == "rlearn":
+        return RLearNConfig(**kwargs)
     else:
         raise ValueError(f"Policy type '{policy_type}' is not available.")
 
 
+class ProcessorConfigKwargs(TypedDict, total=False):
+    """Keyword arguments for the processor config."""
+
+    preprocessor_config_filename: str | None
+    postprocessor_config_filename: str | None
+    preprocessor_overrides: dict[str, Any] | None
+    postprocessor_overrides: dict[str, Any] | None
+    dataset_stats: dict[str, dict[str, torch.Tensor]] | None
+
+
+def make_processor(
+    policy_cfg: PreTrainedConfig,
+    pretrained_path: str | None = None,
+    **kwargs: Unpack[ProcessorConfigKwargs],
+) -> tuple[RobotProcessor, RobotProcessor]:
+    """Make a processor instance for a given policy type.
+
+    This function creates the appropriate processor configuration based on the policy type.
+    Each policy type has its own processor with specific preprocessing steps.
+
+    Args:
+        policy_cfg: The config of the policy to create a processor for (e.g., "act", "diffusion", etc.)
+        pretrained_path: Optional path to load a pretrained processor from. If provided, loads
+            the processor from this path instead of creating a new one.
+        **kwargs: Additional keyword arguments passed to the processor creation.
+
+    Returns:
+        Tuple of (input_processor, output_processor) for the policy.
+
+    Raises:
+        NotImplementedError: If the policy type doesn't have a processor implemented.
+    """
+    if pretrained_path:
+        # Load a pretrained processor
+        # TODO(azouitine): Handle this case.
+        return (
+            RobotProcessor.from_pretrained(
+                pretrained_model_name_or_path=pretrained_path,
+                config_filename=kwargs.get("preprocessor_config_filename", "robot_preprocessor.json"),
+                overrides=kwargs.get("preprocessor_overrides", {}),
+            ),
+            RobotProcessor.from_pretrained(
+                pretrained_model_name_or_path=pretrained_path,
+                config_filename=kwargs.get("postprocessor_config_filename", "robot_postprocessor.json"),
+                overrides=kwargs.get("postprocessor_overrides", {}),
+            ),
+        )
+
+    # 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
+
+        processors = make_tdmpc_processor(
+            config=cast(TDMPCConfig, policy_cfg), dataset_stats=kwargs.get("dataset_stats")
+        )
+
+    elif policy_cfg.type == "diffusion":
+        from lerobot.policies.diffusion.processor_diffusion import make_diffusion_processor
+
+        processors = make_diffusion_processor(
+            cast(DiffusionConfig, policy_cfg), dataset_stats=kwargs.get("dataset_stats")
+        )
+
+    elif policy_cfg.type == "act":
+        from lerobot.policies.act.processor_act import make_act_processor
+
+        processors = make_act_processor(
+            config=cast(ACTConfig, policy_cfg), dataset_stats=kwargs.get("dataset_stats")
+        )
+
+    elif policy_cfg.type == "vqbet":
+        from lerobot.policies.vqbet.processor_vqbet import make_vqbet_processor
+
+        processors = make_vqbet_processor(
+            config=cast(VQBeTConfig, policy_cfg), dataset_stats=kwargs.get("dataset_stats")
+        )
+
+    elif policy_cfg.type == "pi0":
+        from lerobot.policies.pi0.processor_pi0 import make_pi0_processor
+
+        processors = make_pi0_processor(
+            config=cast(PI0Config, policy_cfg), dataset_stats=kwargs.get("dataset_stats")
+        )
+
+    elif policy_cfg.type == "pi0fast":
+        from lerobot.policies.pi0fast.processor_pi0fast import make_pi0fast_processor
+
+        processors = make_pi0fast_processor(
+            cast(PI0Config, policy_cfg), dataset_stats=kwargs.get("dataset_stats")
+        )
+
+    elif policy_cfg.type == "sac":
+        from lerobot.policies.sac.processor_sac import make_sac_processor
+
+        processors = make_sac_processor(
+            cast(SACConfig, policy_cfg), dataset_stats=kwargs.get("dataset_stats")
+        )
+
+    elif policy_cfg.type == "reward_classifier":
+        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")
+        )
+
+    elif policy_cfg.type == "smolvla":
+        from lerobot.policies.smolvla.processor_smolvla import make_smolvla_processor
+
+        processors = make_smolvla_processor(
+            cast(SmolVLAConfig, policy_cfg), dataset_stats=kwargs.get("dataset_stats")
+        )
+
+    elif policy_cfg.type == "rlearn":
+        from lerobot.policies.rlearn.processor_rlearn import make_rlearn_processor
+
+        processors = make_rlearn_processor(
+            cast(RLearNConfig, policy_cfg), dataset_stats=kwargs.get("dataset_stats")
+        )
+
+    else:
+        raise NotImplementedError(f"Processor for policy type '{policy_cfg.type}' is not implemented.")
+
+    return processors
+
+
 def make_policy(
     cfg: PreTrainedConfig,
     ds_meta: LeRobotDatasetMetadata | None = None,
     env_cfg: EnvConfig | None = None,
+    episode_data_index: dict | None = None,
 ) -> PreTrainedPolicy:
     """Make an instance of a policy class.
 
@@ -147,7 +287,6 @@ def make_policy(
     kwargs = {}
     if ds_meta is not None:
         features = dataset_to_policy_features(ds_meta.features)
-        kwargs["dataset_stats"] = ds_meta.stats
     else:
         if not cfg.pretrained_path:
             logging.warning(
@@ -155,12 +294,18 @@ def make_policy(
                 "rather than a dataset. Normalization modules inside the policy will have infinite values "
                 "by default without stats from a dataset."
             )
+        if env_cfg is None:
+            raise ValueError("env_cfg cannot be None when ds_meta is not provided")
         features = env_to_policy_features(env_cfg)
 
     cfg.output_features = {key: ft for key, ft in features.items() if ft.type is FeatureType.ACTION}
     cfg.input_features = {key: ft for key, ft in features.items() if key not in cfg.output_features}
     kwargs["config"] = cfg
 
+    # Pass episode_data_index for RLearN policy to calculate proper progress
+    if cfg.type == "rlearn" and episode_data_index is not None:
+        kwargs["episode_data_index"] = episode_data_index
+
     if cfg.pretrained_path:
         # Load a pretrained policy and override the config if needed (for example, if there are inference-time
         # hyperparameters that we want to vary).

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

src/lerobot/policies/pi0/processor_pi0.py

@@ -0,0 +1,120 @@
+#!/usr/bin/env python
+
+# Copyright 2025 Physical Intelligence and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Any
+
+import torch
+
+from lerobot.configs.types import PolicyFeature
+from lerobot.policies.pi0.configuration_pi0 import PI0Config
+from lerobot.processor import (
+    DeviceProcessor,
+    NormalizerProcessor,
+    RobotProcessor,
+    ToBatchProcessor,
+    TokenizerProcessor,
+    UnnormalizerProcessor,
+)
+from lerobot.processor.pipeline import (
+    EnvTransition,
+    ProcessorStep,
+    ProcessorStepRegistry,
+    TransitionKey,
+)
+from lerobot.processor.rename_processor import RenameProcessor
+
+
+@ProcessorStepRegistry.register(name="pi0_new_line_processor")
+class Pi0NewLineProcessor(ProcessorStep):
+    """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
+
+        task = complementary_data["task"]
+        if task is None:
+            return transition
+
+        # 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"
+        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]
+        # If task is neither string nor list of strings, leave unchanged
+
+        return transition
+
+    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 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(
+            features={**config.input_features, **config.output_features},
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+        ),
+        ToBatchProcessor(),
+        Pi0NewLineProcessor(),  # Add newlines before tokenization for PaliGemma
+        TokenizerProcessor(
+            tokenizer_name="google/paligemma-3b-pt-224",
+            max_length=config.tokenizer_max_length,
+            padding_side="right",
+            padding="max_length",
+        ),
+        DeviceProcessor(device=config.device),
+    ]
+
+    output_steps: list[ProcessorStep] = [
+        DeviceProcessor(device="cpu"),
+        UnnormalizerProcessor(
+            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"
+    )

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
-lerobot-train \
+python -m lerobot.scripts.train \
 --policy.path=lerobot/pi0fast_base \
 --dataset.repo_id=danaaubakirova/koch_test
 ```
 
 Example of training the pi0+FAST neural network with from scratch:
 ```bash
-lerobot-train \
+python -m lerobot.scripts.train \
 --policy.type=pi0fast \
 --dataset.repo_id=danaaubakirova/koch_test
 ```
@@ -58,7 +58,6 @@ from transformers.cache_utils import HybridCache, StaticCache
 from transformers.models.auto import CONFIG_MAPPING
 
 from lerobot.constants import ACTION, OBS_STATE
-from lerobot.policies.normalize import Normalize, Unnormalize
 from lerobot.policies.pi0fast.configuration_pi0fast import PI0FASTConfig
 from lerobot.policies.pretrained import PreTrainedPolicy
 
@@ -146,14 +145,6 @@ class PI0FASTPolicy(PreTrainedPolicy):
         config.validate_features()
         self.config = config
 
-        self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
-        self.normalize_targets = Normalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-        self.unnormalize_outputs = Unnormalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-
         self.language_tokenizer = AutoProcessor.from_pretrained("google/paligemma-3b-pt-224")
         self.model = PI0FAST(config)
 
@@ -221,8 +212,6 @@ class PI0FASTPolicy(PreTrainedPolicy):
         if self.config.adapt_to_pi_aloha:
             batch[OBS_STATE] = self._pi_aloha_decode_state(batch[OBS_STATE])
 
-        batch = self.normalize_inputs(batch)
-
         # Action queue logic for n_action_steps > 1. When the action_queue is depleted, populate it by
         # querying the policy.
         if len(self._action_queue) == 0:
@@ -235,8 +224,6 @@ class PI0FASTPolicy(PreTrainedPolicy):
             ]  # self.config.max_action_dim  # self.config.action_feature.shape[0]
             actions = actions[:, :, :original_action_dim]
 
-            actions = self.unnormalize_outputs({"action": actions})["action"]
-
             if self.config.adapt_to_pi_aloha:
                 actions = self._pi_aloha_encode_actions(actions)
 
@@ -249,8 +236,6 @@ class PI0FASTPolicy(PreTrainedPolicy):
         if self.config.adapt_to_pi_aloha:
             batch[OBS_STATE] = self._pi_aloha_decode_state(batch[OBS_STATE])
             batch[ACTION] = self._pi_aloha_encode_actions_inv(batch[ACTION])
-        batch = self.normalize_inputs(batch)
-        batch = self.normalize_targets(batch)
         loss_dict = self.model.forward(batch)
         return loss_dict["loss"], loss_dict
 

src/lerobot/policies/pi0fast/processor_pi0fast.py

@@ -0,0 +1,51 @@
+#!/usr/bin/env python
+
+# Copyright 2025 Physical Intelligence and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+
+from lerobot.policies.pi0.configuration_pi0 import PI0Config
+from lerobot.processor import (
+    DeviceProcessor,
+    NormalizerProcessor,
+    RenameProcessor,
+    RobotProcessor,
+    ToBatchProcessor,
+    UnnormalizerProcessor,
+)
+
+
+def make_pi0fast_processor(
+    config: PI0Config, dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None
+) -> tuple[RobotProcessor, RobotProcessor]:
+    input_steps = [
+        RenameProcessor(rename_map={}),  # To mimic the same processor as pretrained one
+        NormalizerProcessor(
+            features={**config.input_features, **config.output_features},
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+        ),
+        ToBatchProcessor(),
+        DeviceProcessor(device=config.device),
+    ]
+    output_steps = [
+        DeviceProcessor(device="cpu"),
+        UnnormalizerProcessor(
+            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"
+    )

src/lerobot/policies/rlearn/configuration_rlearn.py

@@ -0,0 +1,128 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass, field
+
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import NormalizationMode
+
+
+@PreTrainedConfig.register_subclass("rlearn")
+@dataclass
+class RLearNConfig(PreTrainedConfig):
+    """Configuration for a video-language conditioned reward model (RLearN).
+
+    Inputs:
+      - Visual frames (one or multiple cameras). Optionally a short sequence.
+      - A language instruction/goal string.
+
+    Output:
+      - Per-timestep reward logits or a single-step reward logit.
+
+    Notes:
+      - This follows the ReWiND paper architecture. It uses frozen vision/text encoders
+        (DINOv3 for vision, SigLIP2 for language) and trains a
+        lightweight temporal aggregator + head.
+    """
+
+    # Encoders - Use SigLIP2 for both vision and text (shared checkpoint)
+    vision_model_name: str = "google/siglip2-base-patch16-224"
+    text_model_name: str = "google/siglip2-base-patch16-224"
+    freeze_backbones: bool = True
+
+    # Sequence length, amount of past frames including current one to use in the temporal model
+    max_seq_len: int = 16
+    # Temporal sampling stride
+    temporal_sampling_stride: int = 3 # Open x mostly has fps 10, and rewind has seq len 16, ours is 30fps so 30/10 = 3 stride lenght to have same timeframe!
+
+    # Model dimensions and transformer
+    dim_model: int = 512
+    num_layers: int = 4
+    num_heads: int = 8
+    ff_mult: int = 4  # Feed-forward multiplier, hidden = dim_model * ff_mult
+    dropout: float = 0.05
+
+    # --- reward head options ---
+    use_categorical_rewards: bool = False      # classification over bins
+    num_reward_bins: int = 25
+    reward_min_value: float = 0.0              # for HL-Gauss range
+    reward_max_value: float = 1.0
+    use_hl_gauss_loss: bool = True             # if False -> plain regression
+    hl_gauss_num_bins: int = 25                # histogram resolution
+
+    # Inference-time subsampling and regularization
+    inference_stride: int = 1 # inference_stride is an extra, second downsampling applied in forward after window sampling/rewind. Keep it at 1 to disable extra skipping
+    frame_dropout_p: float = 0.10
+
+    # Training
+    learning_rate: float = 5e-4
+    weight_decay: float = 0.01
+    head_lr_multiplier: float = 5.0
+    logit_eps: float = 1e-4
+    regularizer_warmup_steps: int = 500
+    
+    # Performance optimizations
+    use_amp: bool = False
+    compile_model: bool = True
+
+    # ReWiND augmentation
+    rewind_prob: float = 0.3 #0.8
+    rewind_last3_prob: float = 0.0 #0.3
+    mismatch_prob: float = 0.0 #0.2
+    
+    # Normalization presets
+    normalization_mapping: dict[str, NormalizationMode] = field(
+        default_factory=lambda: {
+            "VISUAL": NormalizationMode.MEAN_STD,
+        }
+    )
+
+    # Required path to episodes.jsonl for episode boundaries
+    episodes_jsonl_path: str | None = "meta/episodes.jsonl"
+
+    def validate_features(self) -> None:
+        # Require at least one image feature. Language is recommended but optional (can be blank).
+        if not self.image_features:
+            raise ValueError(
+                "You must provide at least one image feature for RLearN (e.g. 'observation.image')."
+            )
+
+    @property
+    def observation_delta_indices(self) -> list | None:
+        # Request a long enough context so in-window stride sampling can be >1.
+        # We ask for (max_seq_len * temporal_sampling_stride) frames ending at t=0.
+        # Example: max_seq_len=16, temporal_sampling_stride=3 → 48 deltas → ~46 frames available.
+        total_needed = self.max_seq_len * max(1, int(self.temporal_sampling_stride))
+        return list(range(1 - total_needed, 1))
+
+    @property
+    def action_delta_indices(self) -> list | None:
+        # Not an action chunking policy.
+        return None
+
+    @property
+    def reward_delta_indices(self) -> list | None:
+        # ReWiND generates progress labels on-the-fly, doesn't need reward data
+        return None
+
+    def get_optimizer_preset(self):  # type: ignore[override]
+        from lerobot.optim.optimizers import AdamWConfig
+
+        return AdamWConfig(lr=self.learning_rate, weight_decay=self.weight_decay)
+
+    def get_scheduler_preset(self):  # type: ignore[override]
+        # No scheduler by default.
+        return None

src/lerobot/policies/rlearn/eval_script.py

@@ -0,0 +1,392 @@
+#!/usr/bin/env python
+
+"""
+Standalone evaluation script for RLearN models.
+
+This script evaluates RLearN reward models on episodes from a dataset,
+generating comparison plots between ground truth rewards and model predictions.
+
+Usage:
+    python src/lerobot/policies/rlearn/eval_script.py --model MODEL_NAME --dataset DATASET_REPO --episodes N
+
+Example:
+    python src/lerobot/policies/rlearn/eval_script.py --model pepijn223/rlearn_18 --dataset pepijn223/phone_pipeline_pickup1 --episodes 2
+"""
+
+import argparse
+import os
+import sys
+from pathlib import Path
+
+# Add src to path for imports
+sys.path.append(str(Path(__file__).parent.parent.parent.parent))
+
+import warnings
+
+import matplotlib.pyplot as plt
+import numpy as np
+import torch
+from scipy.stats import spearmanr
+from tqdm import tqdm
+
+warnings.filterwarnings("ignore")
+
+# LeRobot imports
+from lerobot.constants import OBS_IMAGE, OBS_IMAGES, OBS_LANGUAGE
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.policies.rlearn.modeling_rlearn import RLearNPolicy
+
+
+def _to_chw_float01(img):
+    """Ensure CHW float in [0,1]."""
+    if isinstance(img, np.ndarray):
+        img = torch.from_numpy(img)
+    # HWC -> CHW if needed
+    if len(img.shape) == 3 and img.shape[-1] in (1, 3, 4):
+        img = img.permute(2, 0, 1)
+    if img.dtype == torch.uint8:
+        img = img.float() / 255.0
+    else:
+        img = img.float()
+    return torch.clamp(img, 0.0, 1.0)
+
+
+def _get_language(frame_data):
+    lang = None
+    if OBS_LANGUAGE in frame_data:
+        lang = frame_data[OBS_LANGUAGE]
+        if isinstance(lang, list) and len(lang) > 0:
+            lang = lang[0]
+    elif "task" in frame_data:
+        lang = frame_data["task"]
+    return lang if isinstance(lang, str) else "No language provided"
+
+
+def _get_ground_truth_reward(frame_data):
+    """Try common keys for ground-truth reward. Return None if unavailable."""
+    for key in ("reward", "rewards", "gt_reward", "progress"):
+        if key in frame_data:
+            r = frame_data[key]
+            # unwrap single-element lists/arrays
+            if isinstance(r, (list, np.ndarray)) and np.array(r).size == 1:
+                r = float(np.array(r).reshape(-1)[0])
+            try:
+                return float(r)
+            except Exception:
+                pass
+    return None
+
+
+def extract_episode_frames_and_gt(dataset, episode_idx):
+    """Load a full episode: frames (T, C, H, W), language (str), gt_rewards (np.ndarray or None)."""
+    ep_start = dataset.episode_data_index["from"][episode_idx].item()
+    ep_end = dataset.episode_data_index["to"][episode_idx].item()
+    T = ep_end - ep_start
+
+    frames = []
+    gt_rewards = []
+    language = None
+
+    for t in range(T):
+        item = dataset[ep_start + t]
+
+        # image(s)
+        if OBS_IMAGES in item:
+            img = item[OBS_IMAGES]
+        elif OBS_IMAGE in item:
+            img = item[OBS_IMAGE]
+        else:
+            # try to find an image-like key
+            img_keys = [k for k in item.keys() if "image" in k.lower()]
+            if not img_keys:
+                continue
+            img = item[img_keys[0]]
+
+        frames.append(_to_chw_float01(img))
+
+        # language once
+        if language is None:
+            language = _get_language(item)
+
+        # ground-truth reward (optional)
+        r = _get_ground_truth_reward(item)
+        gt_rewards.append(r)
+
+    if not frames:
+        return None, None, None
+
+    frames = torch.stack(frames)  # (T, C, H, W)
+
+    # If all GT entries are None, treat as missing
+    if all(r is None for r in gt_rewards):
+        gt_rewards = None
+    else:
+        # Replace None by forward filling
+        arr = np.array([np.nan if r is None else float(r) for r in gt_rewards], dtype=float)
+        # forward/back fill
+        if np.isnan(arr[0]):
+            first_valid = np.flatnonzero(~np.isnan(arr))
+            if len(first_valid) > 0:
+                arr[0] = arr[first_valid[0]]
+            else:
+                arr[0] = 0.0
+        for i in range(1, len(arr)):
+            if np.isnan(arr[i]):
+                arr[i] = arr[i - 1]
+        gt_rewards = arr
+
+    return frames, language or "No language provided", gt_rewards
+
+
+@torch.no_grad()
+def predict_rewards_sliding(model, frames, language, max_seq_len=16, batch_size=64, device="cuda", temporal_stride: int | None = None):
+    """
+    Sliding-window prediction: for each frame i, create a window [max(0, i-L+1) .. i],
+    left-pad by repeating the first frame to length L (<= 16), and take the prediction 
+    corresponding to the current frame's position in the window.
+    Returns np.ndarray of shape (T,).
+    """
+    T = frames.shape[0]
+    cfg = getattr(model, "config", object())
+    L = int(getattr(cfg, "max_seq_len", max_seq_len))
+    L = min(L, max_seq_len)  # hard-cap at 16
+    # Use the same temporal stride as training (skip s-1 frames, take 1)
+    if temporal_stride is None:
+        temporal_stride = int(getattr(cfg, "temporal_sampling_stride", 1))
+    temporal_stride = max(1, int(temporal_stride))
+
+    # Preprocessed tensor on device
+    frames = frames.to(device)
+
+    windows = []
+    frame_positions = []  # Track which temporal position each frame should use
+    left_pad_counts = []  # Number of left-pad (OOB) frames per window
+    
+    for i in range(T):
+        # Build indices with stride s: [..., i-3, i] etc., left-padded by clamping to 0
+        idxs = [i - (L - 1 - j) * temporal_stride for j in range(L)]
+        pad_needed = sum(1 for k in idxs if k < 0)
+        clamped = [0 if k < 0 else (T - 1 if k >= T else k) for k in idxs]
+        window = frames[clamped]  # (L, C, H, W)
+
+        # Use the last temporal position (current frame) for reading model output
+        frame_pos = L - 1
+
+        windows.append(window)
+        frame_positions.append(frame_pos)
+        left_pad_counts.append(pad_needed)
+
+    preds = np.zeros(T, dtype=float)
+
+    for s in range(0, T, batch_size):
+        e = min(s + batch_size, T)
+        batch_windows = torch.stack(windows[s:e])  # (B, L, C, H, W)
+        batch_positions = frame_positions[s:e]
+
+        batch = {OBS_IMAGES: batch_windows, OBS_LANGUAGE: [language] * (e - s)}  # expects (B, L, C, H, W)
+
+        # Model returns (B, L) predictions for each temporal position
+        values = model.predict_rewards(batch)  # torch.Tensor (B, L)
+
+        # Apply eval-time padding rule: predictions for left-padded (OOB) frames are zero
+        if values.dim() == 2 and len(left_pad_counts) >= (e - s):
+            for b_idx in range(e - s):
+                pad_n = left_pad_counts[s + b_idx]
+                if pad_n > 0:
+                    values[b_idx, :pad_n] = 0.0
+
+        # Debug output removed - issue was identified and fixed
+
+        if values.dim() == 2:
+            # Extract the prediction corresponding to each frame's position in its window
+            batch_preds = []
+            for b_idx, pos in enumerate(batch_positions):
+                batch_preds.append(values[b_idx, pos].item())
+            preds[s:e] = np.array(batch_preds)
+        else:
+            # Fallback: if model returns (B,), use as is
+            preds[s:e] = values.detach().float().cpu().numpy()
+
+    return preds
+
+
+def plot_episode_eval(episode_idx, gt, pred, language, save_path=None, show=False, title_prefix="RLearN Eval"):
+    """Plot GT vs Predicted over time. Saves PNG if save_path is provided."""
+    T = len(pred)
+    x = np.arange(T)
+
+    plt.figure(figsize=(14, 8))
+    plt.plot(x, pred, linewidth=2.5, marker="o", markersize=3, label="Predicted Reward", color="blue")
+
+    if gt is not None:
+        plt.plot(x, gt, linestyle="--", linewidth=2.5, label="Ground-Truth Reward", color="orange")
+        # Correlation between GT and Pred
+        corr, p = spearmanr(gt, pred)
+        corr_str = f"ρ(GT, Pred) = {0.0 if np.isnan(corr) else corr:.3f} (p={0.0 if np.isnan(p) else p:.3f})"
+    else:
+        expected = np.linspace(0, 1, T)
+        plt.plot(x, expected, linestyle="--", linewidth=2.5, label="Expected Progress (0→1)", color="orange")
+        corr, p = spearmanr(x, pred)
+        corr_str = f"VOC-S ρ(t, Pred) = {0.0 if np.isnan(corr) else corr:.3f} (p={0.0 if np.isnan(p) else p:.3f})"
+
+    plt.title(f"{title_prefix} — Episode {episode_idx}\n{language}\n{corr_str}", fontsize=14)
+    plt.xlabel("Frame Index", fontsize=12)
+    plt.ylabel("Reward / Progress", fontsize=12)
+    plt.legend(fontsize=11)
+    plt.grid(True, alpha=0.3)
+    plt.tight_layout()
+
+    if save_path is not None:
+        plt.savefig(save_path, dpi=200, bbox_inches="tight")
+        print(f"Saved eval image to: {save_path}")
+
+    if show:
+        plt.show()
+    else:
+        plt.close()
+
+
+def eval_episode_sliding(
+    episode_idx, dataset, model, save_dir=".", device="cuda", max_seq_len=16, batch_size=64, title_prefix="RLearN Eval"
+):
+    """End-to-end: load episode, predict with sliding 16-frame windows, and save PNG."""
+    frames, language, gt = extract_episode_frames_and_gt(dataset, episode_idx)
+    if frames is None:
+        print(f"[Episode {episode_idx}] No frames found.")
+        return None
+
+    model.eval()
+
+    pred = predict_rewards_sliding(
+        model=model, frames=frames, language=language, max_seq_len=max_seq_len, batch_size=batch_size, device=device
+    )
+
+    # Basic stats
+    print(f"Episode {episode_idx}: T={len(pred)}, pred∈[{pred.min():.3f},{pred.max():.3f}]")
+    if gt is not None:
+        print(f"GT available: gt∈[{np.nanmin(gt):.3f},{np.nanmax(gt):.3f}]")
+
+    save_path = f"{save_dir}/episode_{episode_idx:04d}_eval.png"
+    plot_episode_eval(
+        episode_idx=episode_idx, gt=gt, pred=pred, language=language, save_path=save_path, show=False, title_prefix=title_prefix
+    )
+    return save_path
+
+
+def main():
+    """Main evaluation script for RLearN models."""
+    # Parse command line arguments
+    parser = argparse.ArgumentParser(description="Evaluate RLearN model on episodes with GT vs Predicted rewards")
+    parser.add_argument("--model", type=str, required=True, help="Model name/path (e.g., pepijn223/rlearn_mse5)")
+    parser.add_argument("--dataset", type=str, required=True, help="Dataset repo (e.g., pepijn223/phone_pipeline_pickup1)")
+    parser.add_argument("--episodes", type=int, default=5, help="Number of episodes to evaluate")
+    parser.add_argument("--output", type=str, default="./eval_results", help="Output directory for images")
+    parser.add_argument(
+        "--device",
+        type=str,
+        default="cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu",
+        help="Device to use",
+    )
+    parser.add_argument("--batch_size", type=int, default=32, help="Batch size for sliding window evaluation")
+
+    args = parser.parse_args()
+
+    # Create output directory
+    output_dir = Path(args.output)
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    print("🎯 RLearN Model Evaluation")
+    print("=" * 60)
+    print(f"Model: {args.model}")
+    print(f"Dataset: {args.dataset}")
+    print(f"Episodes: {args.episodes}")
+    print(f"Device: {args.device}")
+    print(f"Output: {output_dir}")
+    print("=" * 60)
+
+    try:
+        # Load dataset
+        print("📁 Loading dataset...")
+
+        dataset = LeRobotDataset(
+            repo_id=args.dataset,
+            episodes=list(range(min(args.episodes, 50))),  # Load enough episodes
+            download_videos=True,
+        )
+
+        print(f"✅ Dataset loaded: {dataset.num_episodes} episodes, {dataset.num_frames} frames")
+        print(f"   Features: {list(dataset.features.keys())}")
+        print(f"   FPS: {dataset.fps}")
+
+        # Load model
+        print("\n🤖 Loading model...")
+
+        model = RLearNPolicy.from_pretrained(args.model)
+        model = model.to(args.device)
+        model.eval()
+
+        print(f"✅ Model loaded on {args.device}")
+        print(f"   Parameters: {sum(p.numel() for p in model.parameters()):,}")
+        print(f"   Trainable: {sum(p.numel() for p in model.parameters() if p.requires_grad):,}")
+        print(f"   Max sequence length: {model.config.max_seq_len}")
+
+        # Select episodes to evaluate
+        total_available = min(dataset.num_episodes, args.episodes)
+        episode_indices = list(range(total_available))
+
+        print(f"\n📊 Evaluating {len(episode_indices)} episodes...")
+        print("=" * 60)
+
+        # Run sliding window evaluation on each episode
+        saved_paths = []
+        for i, ep_idx in enumerate(episode_indices):
+            print(f"\n[{i+1}/{len(episode_indices)}] Processing Episode {ep_idx}")
+            print("-" * 40)
+
+            try:
+                save_path = eval_episode_sliding(
+                    episode_idx=ep_idx,
+                    dataset=dataset,
+                    model=model,
+                    save_dir=str(output_dir),
+                    device=args.device,
+                    batch_size=args.batch_size,
+                    title_prefix="RLearN Ground Truth vs Predicted",
+                )
+
+                if save_path:
+                    saved_paths.append(save_path)
+
+            except Exception as e:
+                print(f"❌ Error processing episode {ep_idx}: {e}")
+                import traceback
+
+                traceback.print_exc()
+                continue
+
+        # Summary
+        print("\n" + "=" * 60)
+        print("✅ EVALUATION COMPLETE")
+        print(f"📈 Generated {len(saved_paths)} evaluation plots")
+        print(f"📁 Results saved to: {output_dir}")
+        print("\nGenerated files:")
+        for path in saved_paths:
+            print(f"  • {path}")
+
+        if saved_paths:
+            print(f"\n💡 View the plots to compare ground truth vs predicted rewards!")
+            print(f"   Each plot shows the model's sliding 16-frame window predictions")
+            print(f"   against available ground truth rewards over the episode timeline.")
+
+        return 0
+
+    except Exception as e:
+        print(f"❌ Error during evaluation: {e}")
+        import traceback
+
+        traceback.print_exc()
+        return 1
+
+
+if __name__ == "__main__":
+    exit(main())

src/lerobot/policies/rlearn/processor_rlearn.py

@@ -0,0 +1,128 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+from typing import Any
+
+from lerobot.configs.types import PolicyFeature
+from lerobot.constants import OBS_LANGUAGE
+from lerobot.policies.rlearn.configuration_rlearn import RLearNConfig
+from lerobot.processor import (
+    DeviceProcessor,
+    NormalizerProcessor,
+    RenameProcessor,
+    RobotProcessor,
+    ToBatchProcessor,
+    TokenizerProcessor,
+    UnnormalizerProcessor,
+)
+from lerobot.processor.pipeline import (
+    ComplementaryDataProcessor,
+    EnvTransition,
+    ProcessorStepRegistry,
+    TransitionKey,
+)
+
+
+def make_rlearn_processor(
+    config: RLearNConfig, dataset_stats: dict[str, dict[str, Any]] | None = None
+) -> tuple[RobotProcessor, RobotProcessor]:
+    """Build pre/post processors for RLearN.
+
+    Responsibilities moved out of the model:
+      - Normalize inputs (images) using dataset stats
+      - Ensure batching
+      - Map complementary_data.task to observation.language when available
+      - Tokenize language into observation.language.tokens / attention_mask
+      - Move to/from device
+    """
+
+    input_steps = [
+        # No renaming by default, but keep for future extensibility
+        RenameProcessor(rename_map={}),
+        # Move heavy normalization to GPU after transfer for better parallelism
+        ToBatchProcessor(),
+        RLearnLanguageFromTaskProcessor(),
+        # Use SigLIP2 for tokenizer to keep vocab aligned with text tower
+        TokenizerProcessor(
+            tokenizer_name=config.text_model_name,
+            max_length=64,
+            padding="max_length",
+            truncation=True,
+            padding_side="right",
+        ),
+        DeviceProcessor(device=config.device),
+        # Move normalization after GPU transfer to use GPU acceleration
+        NormalizerProcessor(
+            features={**config.input_features, **config.output_features},
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+        ),
+    ]
+
+    output_steps = [
+        DeviceProcessor(device="cpu"),
+        UnnormalizerProcessor(
+            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"
+    )
+
+
+@dataclass
+@ProcessorStepRegistry.register(name="rlearn_language_from_task")
+class RLearnLanguageFromTaskProcessor(ComplementaryDataProcessor):
+    """Copy complementary_data['task'] into observation['observation.language'] if present.
+
+    This ensures the model can consume a raw language string when tokenization is not used,
+    while TokenizerProcessor can still create tokenized fields.
+    """
+
+    task_key: str = "task"
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:  # type: ignore[override]
+        complementary_data = transition.get(TransitionKey.COMPLEMENTARY_DATA)
+        if not complementary_data or self.task_key not in complementary_data:
+            return transition
+
+        task = complementary_data.get(self.task_key)
+        if task is None:
+            return transition
+
+        # Normalize to list[str]
+        if isinstance(task, str):
+            task_list = [task]
+        elif isinstance(task, list) and all(isinstance(t, str) for t in task):
+            task_list = task
+        else:
+            return transition
+
+        observation = transition.get(TransitionKey.OBSERVATION) or {}
+        # Do not overwrite if user already provided observation.language
+        if OBS_LANGUAGE not in observation:
+            observation[OBS_LANGUAGE] = task_list
+            transition[TransitionKey.OBSERVATION] = observation
+        return transition
+
+    def transform_features(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:  # noqa: D401
+        # Adds nothing to features; only mirrors complementary_data.task into observation
+        return features
+
+    def get_config(self) -> dict[str, Any]:
+        return {"task_key": self.task_key}

src/lerobot/policies/rlearn/rlearn_plan.md

@@ -0,0 +1,101 @@
+## General Value/Reward Learning:
+
+I want to implement a general/universal vision and language value function or reward model for robotics/video tasks. Also called a video language conditioned reward model. Integrated with already existing LeRobot code if convenient, use the LeRobot Dataset for dataset and store the reward for a frame in the lerobot frame itself.
+
+Inspired by these papers:
+
+- ReWiND; https://arxiv.org/pdf/2505.10911 (Most applicable and main paper I want to implement ideas from) and code: https://github.com/lucidrains/rewind-reward-pytorch
+- LIV; https://arxiv.org/pdf/2306.00958 (Most applicable and 2nd main paper I want to implement ideas from) and code https://github.com/penn-pal-lab/LI
+- VLC: Video-Language Critic: Transferable Reward Functions for Language-Conditioned Robotics: https://arxiv.org/pdf/2405.19988 (Most applicable and 3rd paper I want to implement ideas from) and code: https://github.com/minttusofia/video_language_critic
+
+And these papers which are also relevant:
+
+- https://www.dyna.co/dyna-1/research (Main company I want to reproduce the eventual results from)
+- vip; https://arxiv.org/pdf/2210.00030
+- uvd; https://arxiv.org/pdf/2310.08581
+- vlm in context; https://arxiv.org/pdf/2411.04549
+- https://www.youtube.com/watch?v=JfZYtpEisoM
+
+Little less relevant but still similar papers:
+
+- Learning Generalizable Robotic Reward Functions from “In-The-Wild” Human Videos,
+- XIRL: Cross-embodiment Inverse Reinforcement Learning,
+- Video-Language Critic: Transferable Reward https://arxiv.org/pdf/2405.19988
+- Functions for Language-Conditioned Robotics,
+- LORel, Language-Driven Representation Learning for Robotics https://sites.google.com/view/robotlorel
+- RoboCLIP: One Demonstration is Enough to Learn Robot Policies https://arxiv.org/pdf/2310.07899
+- Points2Rewards: learn first key points and then uses the keypoints to learn general value function/policy https://semrob.github.io/docs/2025_rss_semrob.github.io_paper20.pdf
+- Language-Driven Representation Learning for Robotics: https://arxiv.org/pdf/2302.12766v1
+- R3M: A Universal Visual Representation for Robot Manipulation: https://arxiv.org/pdf/2203.12601v3
+
+Input should be the current image or whole video and the task goal specified in text/language. Output is current reward.
+Archiutecture:
+_ inputs: video o1:T (or current o1:t), language z;
+_ DINO v3 ViT-B/16 (86M params): https://huggingface.co/facebook/dinov3-vitb16-pretrain-lvd1689m for vision encoding
+\_ sentence-transformers/all-MiniLM-L12-v2: https://huggingface.co/sentence-transformers/all-MiniLM-L12-v2 for text encoding \* Temporal module: small causal transformer ("cross-modal sequential aggregator"), with first-frame positional embedding (to avoid position cheating), frame-dropout, and stride sampling; outputs per-timestep logits.
+
+Loss: See this chatgpt thread: https://chatgpt.com/s/t_68999a50a0b081919abc365cdd205e01
+
+Past images: (for example a reward method go to 3rd floor, has to know what floor it was on and what pas actions it did, can we attend or encorperate images of decision from history in one way?) Maybe via this paper: Learning Long-Context Diffusion Policies via Past-Token Prediction
+
+Amount of frames needed for test/generalization: 1M frames? or ~20% of IPEC-COMMUNITY/bc_z_lerobot
+
+Eval:
+Implement something like voc score , or ROC rank order correlation between reward leanredna and ev reward from sim, or use something else to do additional evaluation
+
+Ideas:
+
+- Incorporate training on multiple horizons: as in label same dataset for longer horizons: make a sandwich (long), put cheese on bread (medium) and even smaller horizons: go down or close gripper (small)
+- Incorporate navigation goals “walk towards the kitchen”, make sure we fix CLIP contrastive learning issue of positional text misunderstanding where model doesnnt learn difference between "horse right of cow" and "horse left of cow" “Move right” potentially train with more other data or even actionable world models such as Genie 3 (https://deepmind.google/discover/blog/genie-3-a-new-frontier-for-world-models/)
+
+How to use a general reward model (use cases): - Train rl policy on it - Success detection - Do exploraion - Do task via planning and search to optimize reward - Filter out bad episodes in large datasets from imitation learning
+
+Potential Datasets: (start with dataset that is most clean for this and works best with chosen way of doing evals)
+_ Epic-Kitchens-100
+_ Something-Something v. 2 Dataset https://www.qualcomm.com/developer/software/something-something-v-2-dataset
+_ Ego4D (3000 hours)
+_ Open X-Embodiment (OXE)
+\_ Agi bot world: https://huggingface.co/datasets/agibot-world/AgiBotWorld-Alpha
+
+- GalexiAI dataset: https://opengalaxea.github.io/G0/
+  _ GTEA+ Gaze: https://cbs.ic.gatech.edu/fpv/
+  _ YouCook2 dataset
+  \_ HOWTO100M: https://www.di.ens.fr/willow/research/howto100m/
+- Genie generated dataset?
+
+### TODOs:
+
+- Implement first architecture [x]
+- Implement processors [x]
+- Choose right loss metric(s) [x]
+- Make dataset with script that generated the dataset (IPEC-COMMUNITY/bc_z_lerobot) ready in lerobot format (and be able to visualize in dataset visualizer)
+  - Annotate with ReWiND-style 0→1 progress rewards [x]
+  - Visualize to check [x]
+- Implement eval score or metric that is robust and can deal with generalization/is a good metric to try different architectures. And use it in an eval jupyter notebook with visalization of the live reward next to the video for part of the dataset: VOC score and score with correct and incorrect language captions [x]
+- Do first training [x]
+- Implement on-the-fly progress label generation (no need for pre-annotated rewards) [x]
+- Try different losses
+  - Only rewind loss [x]
+  - Exactly similar to: https://github.com/lucidrains/rewind-reward-pytorch/blob/main/rewind_reward_pytorch/rewind_reward.py#L11 [x]
+  - Try DINO v2 as encoder Base 86 M: with https://huggingface.co/sentence-transformers/all-MiniLM-L12-v2 [x]
+  - Test rewind (evaluate) [x]
+- benchmark siglip 2 vs this implementation forward pass, debug speed [x]
+- use siglip 2 [x]
+- Fix evaluation bug !!! []
+- Fix sample episode padding bug !!! []
+- Overfit on one episode []
+- Cleanup code? [] + enable language loss
+- Convert python -m lerobot.datasets.v21.convert_dataset_v20_to_v21 --repo-id=IPEC-COMMUNITY/bc_z_lerobot and train on 1 percent
+- Then on 10 percent []
+- Ablation 16 sucessive frame vs 16 frame samples with stride 2 or 4 []
+- Add more artificial text to dataset generated by vlm (google gemini) []
+  - See google gemini vlm caption [] https://gemini.google.com/app/7e332ffaf32580f2
+  - Multiple captions per video, creat method to generate as much data as possible etc [] https://arxiv.org/abs/2508.13446, https://arxiv.org/pdf/2412.04453
+- Add other datasets from OXE metioned in rewind []
+- Extend evaluation []
+- Ablation for size vision encoder, language encoder, temporal head []
+- Ablation one mlp head per frame or single mlp head []
+- Add other datasets metnioned here []
+- How can we improve spatial aware learning? solve issue of Contrastive learning and position []
+
+

src/lerobot/policies/sac/modeling_sac.py

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

src/lerobot/policies/sac/processor_sac.py

@@ -0,0 +1,52 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team.
+# All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+
+from lerobot.policies.sac.configuration_sac import SACConfig
+from lerobot.processor import (
+    DeviceProcessor,
+    NormalizerProcessor,
+    RenameProcessor,
+    RobotProcessor,
+    ToBatchProcessor,
+    UnnormalizerProcessor,
+)
+
+
+def make_sac_processor(
+    config: SACConfig, dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None
+) -> tuple[RobotProcessor, RobotProcessor]:
+    input_steps = [
+        RenameProcessor(rename_map={}),
+        NormalizerProcessor(
+            features={**config.input_features, **config.output_features},
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+        ),
+        ToBatchProcessor(),
+        DeviceProcessor(device=config.device),
+    ]
+    output_steps = [
+        DeviceProcessor(device="cpu"),
+        UnnormalizerProcessor(
+            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"
+    )

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

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

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

@@ -0,0 +1,42 @@
+# !/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import torch
+
+from lerobot.policies.sac.reward_model.configuration_classifier import RewardClassifierConfig
+from lerobot.processor import (
+    DeviceProcessor,
+    IdentityProcessor,
+    NormalizerProcessor,
+    RobotProcessor,
+)
+
+
+def make_classifier_processor(
+    config: RewardClassifierConfig, dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None
+) -> tuple[RobotProcessor, RobotProcessor]:
+    input_steps = [
+        NormalizerProcessor(
+            features=config.input_features, norm_map=config.normalization_mapping, stats=dataset_stats
+        ),
+        NormalizerProcessor(
+            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
+        ),
+        DeviceProcessor(device=config.device),
+    ]
+    output_steps = [DeviceProcessor(device="cpu"), IdentityProcessor()]
+    return RobotProcessor(steps=input_steps, name="classifier_preprocessor"), RobotProcessor(
+        steps=output_steps, name="classifier_postprocessor"
+    )

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

src/lerobot/policies/smolvla/processor_smolvla.py

@@ -0,0 +1,109 @@
+#!/usr/bin/env python
+
+# Copyright 2025 HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Any
+
+import torch
+
+from lerobot.configs.types import PolicyFeature
+from lerobot.policies.smolvla.configuration_smolvla import SmolVLAConfig
+from lerobot.processor import (
+    DeviceProcessor,
+    NormalizerProcessor,
+    RenameProcessor,
+    RobotProcessor,
+    ToBatchProcessor,
+    TokenizerProcessor,
+    UnnormalizerProcessor,
+)
+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]:
+    input_steps = [
+        RenameProcessor(rename_map={}),  # To mimic the same processor as pretrained one
+        NormalizerProcessor(
+            features={**config.input_features, **config.output_features},
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+        ),
+        ToBatchProcessor(),
+        SmolVLANewLineProcessor(),
+        TokenizerProcessor(
+            tokenizer_name=config.vlm_model_name,
+            padding=config.pad_language_to,
+            padding_side="right",
+            max_length=config.tokenizer_max_length,
+        ),
+        DeviceProcessor(device=config.device),
+    ]
+    output_steps = [
+        DeviceProcessor(device="cpu"),
+        UnnormalizerProcessor(
+            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"
+    )
+
+
+@ProcessorStepRegistry.register(name="smolvla_new_line_processor")
+class SmolVLANewLineProcessor(ProcessorStep):
+    """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
+
+        task = complementary_data["task"]
+        if task is None:
+            return transition
+
+        # 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"
+        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]
+        # If task is neither string nor list of strings, leave unchanged
+
+        return transition
+
+    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/modeling_tdmpc.py

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

src/lerobot/policies/tdmpc/processor_tdmpc.py

@@ -0,0 +1,51 @@
+#!/usr/bin/env python
+
+# Copyright 2024 Nicklas Hansen, Xiaolong Wang, Hao Su,
+# and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import torch
+
+from lerobot.policies.tdmpc.configuration_tdmpc import TDMPCConfig
+from lerobot.processor import (
+    DeviceProcessor,
+    NormalizerProcessor,
+    RenameProcessor,
+    RobotProcessor,
+    ToBatchProcessor,
+    UnnormalizerProcessor,
+)
+
+
+def make_tdmpc_processor(
+    config: TDMPCConfig, dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None
+) -> tuple[RobotProcessor, RobotProcessor]:
+    input_steps = [
+        RenameProcessor(rename_map={}),
+        NormalizerProcessor(
+            features={**config.input_features, **config.output_features},
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+        ),
+        ToBatchProcessor(),
+        DeviceProcessor(device=config.device),
+    ]
+    output_steps = [
+        DeviceProcessor(device="cpu"),
+        UnnormalizerProcessor(
+            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"
+    )

src/lerobot/policies/vqbet/modeling_vqbet.py

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

src/lerobot/policies/vqbet/processor_vqbet.py

@@ -0,0 +1,52 @@
+#!/usr/bin/env python
+
+# Copyright 2024 Seungjae Lee and Yibin Wang and Haritheja Etukuru
+# and H. Jin Kim and Nur Muhammad Mahi Shafiullah and Lerrel Pinto
+# and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import torch
+
+from lerobot.policies.vqbet.configuration_vqbet import VQBeTConfig
+from lerobot.processor import (
+    DeviceProcessor,
+    NormalizerProcessor,
+    RenameProcessor,
+    RobotProcessor,
+    ToBatchProcessor,
+    UnnormalizerProcessor,
+)
+
+
+def make_vqbet_processor(
+    config: VQBeTConfig, dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None
+) -> tuple[RobotProcessor, RobotProcessor]:
+    input_steps = [
+        RenameProcessor(rename_map={}),  # Let the possibility to the user to rename the keys
+        NormalizerProcessor(
+            features={**config.input_features, **config.output_features},
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+        ),
+        ToBatchProcessor(),
+        DeviceProcessor(device=config.device),
+    ]
+    output_steps = [
+        DeviceProcessor(device="cpu"),
+        UnnormalizerProcessor(
+            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"
+    )

src/lerobot/processor/__init__.py

@@ -14,8 +14,9 @@
 # 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
+from .normalize_processor import NormalizerProcessor, UnnormalizerProcessor, hotswap_stats
 from .observation_processor import VanillaObservationProcessor
 from .pipeline import (
     ActionProcessor,
@@ -32,6 +33,7 @@ from .pipeline import (
     TruncatedProcessor,
 )
 from .rename_processor import RenameProcessor
+from .tokenizer_processor import TokenizerProcessor
 
 __all__ = [
     "ActionProcessor",
@@ -42,12 +44,15 @@ __all__ = [
     "InfoProcessor",
     "NormalizerProcessor",
     "UnnormalizerProcessor",
+    "hotswap_stats",
     "ObservationProcessor",
     "ProcessorStep",
     "ProcessorStepRegistry",
     "RenameProcessor",
     "RewardProcessor",
     "RobotProcessor",
+    "ToBatchProcessor",
+    "TokenizerProcessor",
     "TransitionKey",
     "TruncatedProcessor",
     "VanillaObservationProcessor",

src/lerobot/processor/batch_processor.py

@@ -0,0 +1,139 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from dataclasses import dataclass
+from typing import Any
+
+import torch
+from 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
+
+
+@dataclass
+@ProcessorStepRegistry.register(name="to_batch_processor")
+class ToBatchProcessor:
+    """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:
+
+    - For state observations (observation.state, observation.environment_state):
+      Adds batch dimension (unsqueeze at dim=0) if tensor is 1-dimensional
+
+    - For image observations (observation.image, observation.images.*):
+      Adds batch dimension (unsqueeze at dim=0) if tensor is 3-dimensional (H, W, C)
+
+    - For actions:
+      Adds batch dimension (unsqueeze at dim=0) if tensor is 1-dimensional
+
+    - For task field in complementary data:
+      Wraps string task in a list to add batch dimension
+      (task must be a string or list of strings)
+
+    This is useful when processing single transitions that need to be batched for
+    model inference or when converting from unbatched environment outputs to
+    batched model inputs.
+
+    The processor only modifies tensors that need batching and leaves already
+    batched tensors unchanged.
+
+    Example:
+        ```python
+        # State: (7,) -> (1, 7)
+        # Image: (224, 224, 3) -> (1, 224, 224, 3)
+        # Action: (4,) -> (1, 4)
+        # Task: "pick_cube" -> ["pick_cube"]
+        # Already batched: (1, 7) -> (1, 7) [unchanged]
+        ```
+    """
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        self._process_observation(transition)
+        self._process_action(transition)
+        self._process_complementary_data(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]:
+        return features