add: profiler for streaming performance

docs(dataset): add dataset v3 documentation (#1956 )
* add v3 doc * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix * update changes * iterate on review * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * add changes * create dataset section * Update docs/source/lerobot-dataset-v3.mdx Signed-off-by: Francesco Capuano <74058581+fracapuano@users.noreply.github.com> * Update docs/source/lerobot-dataset-v3.mdx Signed-off-by: Francesco Capuano <74058581+fracapuano@users.noreply.github.com> * Update docs/source/lerobot-dataset-v3.mdx Signed-off-by: Francesco Capuano <74058581+fracapuano@users.noreply.github.com> --------- Signed-off-by: Francesco Capuano <74058581+fracapuano@users.noreply.github.com> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> Co-authored-by: Michel Aractingi <michel.aractingi@huggingface.co> Co-authored-by: Francesco Capuano <74058581+fracapuano@users.noreply.github.com>
2026-05-11 22:59:50 +00:00 · 2025-09-18 17:28:14 +02:00 · 2025-09-16 17:45:38 +02:00 · 2025-09-15 18:52:30 +02:00 · 2025-09-15 14:08:01 +02:00 · 2025-09-15 09:53:30 +02:00
134 changed files with 14522 additions and 4295 deletions
@@ -30,7 +30,7 @@ pytest -sx tests/test_stuff.py::test_something
 ```

 ```bash
-python -m lerobot.scripts.train --some.option=true
+lerobot-train --some.option=true
 ```

 ## SECTION TO REMOVE BEFORE SUBMITTING YOUR PR
@@ -29,8 +29,8 @@ on:
 env:
  UV_VERSION: "0.8.0"
  PYTHON_VERSION: "3.10"
-  DOCKER_IMAGE_NAME_CPU: huggingface/lerobot-gpu:latest
-  DOCKER_IMAGE_NAME_GPU: huggingface/lerobot-cpu:latest
+  DOCKER_IMAGE_NAME_CPU: huggingface/lerobot-cpu:latest
+  DOCKER_IMAGE_NAME_GPU: huggingface/lerobot-gpu:latest

 # Ensures that only the latest commit is built, canceling older runs.
 concurrency:
@@ -44,7 +44,7 @@ test-end-to-end:
 	${MAKE} DEVICE=$(DEVICE) test-smolvla-ete-eval

 test-act-ete-train:
-	python -m lerobot.scripts.train \
+	lerobot-train \
 		--policy.type=act \
 		--policy.dim_model=64 \
 		--policy.n_action_steps=20 \
@@ -68,12 +68,12 @@ test-act-ete-train:
 		--output_dir=tests/outputs/act/

 test-act-ete-train-resume:
-	python -m lerobot.scripts.train \
+	lerobot-train \
 		--config_path=tests/outputs/act/checkpoints/000002/pretrained_model/train_config.json \
 		--resume=true

 test-act-ete-eval:
-	python -m lerobot.scripts.eval \
+	lerobot-eval \
 		--policy.path=tests/outputs/act/checkpoints/000004/pretrained_model \
 		--policy.device=$(DEVICE) \
 		--env.type=aloha \
@@ -82,7 +82,7 @@ test-act-ete-eval:
 		--eval.batch_size=1

 test-diffusion-ete-train:
-	python -m lerobot.scripts.train \
+	lerobot-train \
 		--policy.type=diffusion \
 		--policy.down_dims='[64,128,256]' \
 		--policy.diffusion_step_embed_dim=32 \
@@ -106,7 +106,7 @@ test-diffusion-ete-train:
 		--output_dir=tests/outputs/diffusion/

 test-diffusion-ete-eval:
-	python -m lerobot.scripts.eval \
+	lerobot-eval \
 		--policy.path=tests/outputs/diffusion/checkpoints/000002/pretrained_model \
 		--policy.device=$(DEVICE) \
 		--env.type=pusht \
@@ -115,7 +115,7 @@ test-diffusion-ete-eval:
 		--eval.batch_size=1

 test-tdmpc-ete-train:
-	python -m lerobot.scripts.train \
+	lerobot-train \
 		--policy.type=tdmpc \
 		--policy.device=$(DEVICE) \
 		--policy.push_to_hub=false \
@@ -137,7 +137,7 @@ test-tdmpc-ete-train:
 		--output_dir=tests/outputs/tdmpc/

 test-tdmpc-ete-eval:
-	python -m lerobot.scripts.eval \
+	lerobot-eval \
 		--policy.path=tests/outputs/tdmpc/checkpoints/000002/pretrained_model \
 		--policy.device=$(DEVICE) \
 		--env.type=xarm \
@@ -148,7 +148,7 @@ test-tdmpc-ete-eval:


 test-smolvla-ete-train:
-	python -m lerobot.scripts.train \
+	lerobot-train \
 		--policy.type=smolvla \
 		--policy.n_action_steps=20 \
 		--policy.chunk_size=20 \
@@ -171,7 +171,7 @@ test-smolvla-ete-train:
 		--output_dir=tests/outputs/smolvla/

 test-smolvla-ete-eval:
-	python -m lerobot.scripts.eval \
+	lerobot-eval \
 		--policy.path=tests/outputs/smolvla/checkpoints/000004/pretrained_model \
 		--policy.device=$(DEVICE) \
 		--env.type=aloha \
@@ -6,7 +6,7 @@

 <div align="center">

-[![Tests](https://github.com/huggingface/lerobot/actions/workflows/nightly.yml/badge.svg?branch=main)](https://github.com/huggingface/lerobot/actions/workflows/nighty.yml?query=branch%3Amain)
+[![Tests](https://github.com/huggingface/lerobot/actions/workflows/nightly.yml/badge.svg?branch=main)](https://github.com/huggingface/lerobot/actions/workflows/nightly.yml?query=branch%3Amain)
 [![Python versions](https://img.shields.io/pypi/pyversions/lerobot)](https://www.python.org/downloads/)
 [![License](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](https://github.com/huggingface/lerobot/blob/main/LICENSE)
 [![Status](https://img.shields.io/pypi/status/lerobot)](https://pypi.org/project/lerobot/)
@@ -101,6 +101,9 @@
 ## Installation

 LeRobot works with Python 3.10+ and PyTorch 2.2+.
+
+### Environment Setup
+
 Create a virtual environment with Python 3.10 and activate it, e.g. with [`miniconda`](https://docs.anaconda.com/free/miniconda/index.html):

 ```bash
@@ -124,10 +127,21 @@ conda install ffmpeg -c conda-forge
 >
 > - _[On Linux only]_ Install [ffmpeg build dependencies](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu#GettheDependencies) and [compile ffmpeg from source with libsvtav1](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu#libsvtav1), and make sure you use the corresponding ffmpeg binary to your install with `which ffmpeg`.

-Install 🤗 LeRobot:
+### Install LeRobot 🤗
+
+#### From Source
+
+First, clone the repository and navigate into the directory:

 ```bash
-pip install lerobot
+git clone https://github.com/huggingface/lerobot.git
+cd lerobot
+```
+
+Then, install the library in editable mode. This is useful if you plan to contribute to the code.
+
+```bash
+pip install -e .
 ```

 > **NOTE:** If you encounter build errors, you may need to install additional dependencies (`cmake`, `build-essential`, and `ffmpeg libs`). On Linux, run:
@@ -145,6 +159,34 @@ For instance, to install 🤗 LeRobot with aloha and pusht, use:
 pip install -e ".[aloha, pusht]"
 ```

+### Installation from PyPI
+
+**Core Library:**
+Install the base package with:
+
+```bash
+pip install lerobot
+```
+
+_This installs only the default dependencies._
+
+**Extra Features:**
+To install additional functionality, use one of the following:
+
+```bash
+pip install 'lerobot[all]'          # All available features
+pip install 'lerobot[aloha,pusht]'  # Specific features (Aloha & Pusht)
+pip install 'lerobot[feetech]'      # Feetech motor support
+```
+
+_Replace `[...]` with your desired features._
+
+**Available Tags:**
+For a full list of optional dependencies, see:
+https://pypi.org/project/lerobot/
+
+### Weights & Biases
+
 To use [Weights and Biases](https://docs.wandb.ai/quickstart) for experiment tracking, log in with

 ```bash
@@ -191,7 +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):
@@ -204,20 +246,30 @@ 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
-  ├ 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", ...]`)
-```
+  ├ 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')

 A `LeRobotDataset` is serialised using several widespread file formats for each of its parts, namely:

@@ -234,22 +286,22 @@ 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
-python -m lerobot.scripts.eval \
+lerobot-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
-python -m lerobot.scripts.eval --policy.path={OUTPUT_DIR}/checkpoints/last/pretrained_model
+lerobot-eval --policy.path={OUTPUT_DIR}/checkpoints/last/pretrained_model
 ```

-See `python -m lerobot.scripts.eval --help` for more instructions.
+See `lerobot-eval --help` for more instructions.

 ### Train your own policy

@@ -261,7 +313,7 @@ A link to the wandb logs for the run will also show up in yellow in your termina

 \<img src="https://raw.githubusercontent.com/huggingface/lerobot/main/media/wandb.png" alt="WandB logs example"\>

-Note: For efficiency, during training every checkpoint is evaluated on a low number of episodes. You may use `--eval.n_episodes=500` to evaluate on more episodes than the default. Or, after training, you may want to re-evaluate your best checkpoints on more episodes or change the evaluation settings. See `python -m lerobot.scripts.eval --help` for more instructions.
+Note: For efficiency, during training every checkpoint is evaluated on a low number of episodes. You may use `--eval.n_episodes=500` to evaluate on more episodes than the default. Or, after training, you may want to re-evaluate your best checkpoints on more episodes or change the evaluation settings. See `lerobot-eval --help` for more instructions.

 #### Reproduce state-of-the-art (SOTA)

@@ -269,7 +321,7 @@ We provide some pretrained policies on our [hub page](https://huggingface.co/ler
 You can reproduce their training by loading the config from their run. Simply running:

 ```bash
-python -m lerobot.scripts.train --config_path=lerobot/diffusion_pusht
+lerobot-train --config_path=lerobot/diffusion_pusht
 ```

 reproduces SOTA results for Diffusion Policy on the PushT task.
@@ -311,7 +363,7 @@ If you want, you can cite this work with:

 ```bibtex
@misc{cadene2024lerobot,
-    author = {Cadene, Remi and Alibert, Simon and Soare, Alexander and Gallouedec, Quentin and Zouitine, Adil and Palma, Steven and Kooijmans, Pepijn and Aractingi, Michel and Shukor, Mustafa and Aubakirova, Dana and Russi, Martino and Capuano, Francesco and Pascale, Caroline and Choghari, Jade and Moss, Jess and Wolf, Thomas},
+    author = {Cadene, Remi and Alibert, Simon and Soare, Alexander and Gallouedec, Quentin and Zouitine, Adil and Palma, Steven and Kooijmans, Pepijn and Aractingi, Michel and Shukor, Mustafa and Aubakirova, Dana and Russi, Martino and Capuano, Francesco and Pascal, Caroline and Choghari, Jade and Moss, Jess and Wolf, Thomas},
    title = {LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch},
    howpublished = "\url{https://github.com/huggingface/lerobot}",
    year = {2024}
@@ -108,7 +108,8 @@ def save_decoded_frames(


 def save_first_episode(imgs_dir: Path, dataset: LeRobotDataset) -> None:
-    ep_num_images = dataset.episode_data_index["to"][0].item()
+    episode_index = 0
+    ep_num_images = dataset.meta.episodes["length"][episode_index]
    if imgs_dir.exists() and len(list(imgs_dir.glob("frame_*.png"))) == ep_num_images:
        return

@@ -265,7 +266,8 @@ def benchmark_encoding_decoding(
            overwrite=True,
        )

-    ep_num_images = dataset.episode_data_index["to"][0].item()
+    episode_index = 0
+    ep_num_images = dataset.meta.episodes["length"][episode_index]
    width, height = tuple(dataset[0][dataset.meta.camera_keys[0]].shape[-2:])
    num_pixels = width * height
    video_size_bytes = video_path.stat().st_size
@@ -29,7 +29,7 @@ ENV DEBIAN_FRONTEND=noninteractive \

 # Install system dependencies and uv (as root)
 RUN apt-get update && apt-get install -y --no-install-recommends \
-    build-essential git curl libglib2.0-0 libegl1-mesa ffmpeg \
+    build-essential git curl libglib2.0-0 libegl1-mesa-dev ffmpeg \
    libusb-1.0-0-dev speech-dispatcher libgeos-dev portaudio19-dev \
    && curl -LsSf https://astral.sh/uv/install.sh | sh \
    && mv /root/.local/bin/uv /usr/local/bin/uv \
@@ -20,6 +20,12 @@
  - 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
@@ -35,10 +41,14 @@
    title: Koch v1.1
  - local: lekiwi
    title: LeKiwi
+  - local: reachy2
+    title: Reachy 2
  title: "Robots"
 - sections:
  - local: notebooks
    title: Notebooks
+  - local: feetech
+    title: Updating Feetech Firmware
  title: "Resources"
 - sections:
  - local: contributing
@@ -9,7 +9,7 @@ To instantiate a camera, you need a camera identifier. This identifier might cha
 To find the camera indices of the cameras plugged into your system, run the following script:

 ```bash
-python -m lerobot.find_cameras opencv # or realsense for Intel Realsense cameras
+lerobot-find-cameras opencv # or realsense for Intel Realsense cameras
 ```

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

 **Deploying and Testing the Model**
@@ -458,7 +458,7 @@ The reward classifier will automatically provide rewards based on the visual inp
 3. **Train the classifier**:

   ```bash
-   python -m lerobot.scripts.train --config_path src/lerobot/configs/reward_classifier_train_config.json
+   lerobot-train --config_path src/lerobot/configs/reward_classifier_train_config.json
   ```

 4. **Test the classifier**:
@@ -19,7 +19,7 @@ pip install -e ".[hopejr]"
 Before starting calibration and operation, you need to identify the USB ports for each HopeJR component. Run this script to find the USB ports for the arm, hand, glove, and exoskeleton:

 ```bash
-python -m lerobot.find_port
+lerobot-find-port
 ```

 This will display the available USB ports and their associated devices. Make note of the port paths (e.g., `/dev/tty.usbmodem58760433331`, `/dev/tty.usbmodem11301`) as you'll need to specify them in the `--robot.port` and `--teleop.port` parameters when recording data, replaying episodes, or running teleoperation scripts.
@@ -31,7 +31,7 @@ Before performing teleoperation, HopeJR's limbs need to be calibrated. Calibrati
 ### 1.1 Calibrate Robot Hand

 ```bash
-python -m lerobot.calibrate \
+lerobot-calibrate \
    --robot.type=hope_jr_hand \
    --robot.port=/dev/tty.usbmodem58760432281 \
    --robot.id=blue \
@@ -81,7 +81,7 @@ Once you have set the appropriate boundaries for all joints, click "Save" to sav
 ### 1.2 Calibrate Teleoperator Glove

 ```bash
-python -m lerobot.calibrate \
+lerobot-calibrate \
    --teleop.type=homunculus_glove \
    --teleop.port=/dev/tty.usbmodem11201 \
    --teleop.id=red \
@@ -120,7 +120,7 @@ Once calibration is complete, the system will save the calibration to `/Users/yo
 ### 1.3 Calibrate Robot Arm

 ```bash
-python -m lerobot.calibrate \
+lerobot-calibrate \
    --robot.type=hope_jr_arm \
    --robot.port=/dev/tty.usbserial-1110 \
    --robot.id=white
@@ -146,7 +146,7 @@ Use the calibration interface to set the range boundaries for each joint. Move e
 ### 1.4 Calibrate Teleoperator Exoskeleton

 ```bash
-python -m lerobot.calibrate \
+lerobot-calibrate \
    --teleop.type=homunculus_arm \
    --teleop.port=/dev/tty.usbmodem11201 \
    --teleop.id=black
@@ -178,7 +178,7 @@ Due to global variable conflicts in the Feetech middleware, teleoperation for ar
 ### Hand

 ```bash
-python -m lerobot.teleoperate \
+lerobot-teleoperate \
    --robot.type=hope_jr_hand \
    --robot.port=/dev/tty.usbmodem58760432281 \
    --robot.id=blue \
@@ -194,7 +194,7 @@ python -m lerobot.teleoperate \
 ### Arm

 ```bash
-python -m lerobot.teleoperate \
+lerobot-teleoperate \
    --robot.type=hope_jr_arm \
    --robot.port=/dev/tty.usbserial-1110 \
    --robot.id=white \
@@ -214,7 +214,7 @@ Record, Replay and Train with Hope-JR is still experimental.
 This step records the dataset, which can be seen as an example [here](https://huggingface.co/datasets/nepyope/hand_record_test_with_video_data/settings).

 ```bash
-python -m lerobot.record \
+lerobot-record \
    --robot.type=hope_jr_hand \
    --robot.port=/dev/tty.usbmodem58760432281 \
    --robot.id=right \
@@ -236,7 +236,7 @@ python -m lerobot.record \
 ### Replay

 ```bash
-python -m lerobot.replay \
+lerobot-replay \
    --robot.type=hope_jr_hand \
    --robot.port=/dev/tty.usbmodem58760432281 \
    --robot.id=right \
@@ -248,7 +248,7 @@ python -m lerobot.replay \
 ### Train

 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
  --dataset.repo_id=nepyope/hand_record_test_with_video_data \
  --policy.type=act \
  --output_dir=outputs/train/hopejr_hand \
@@ -263,7 +263,7 @@ python -m lerobot.scripts.train \
 This training run can be viewed as an example [here](https://wandb.ai/tino/lerobot/runs/rp0k8zvw?nw=nwusertino).

 ```bash
-python -m lerobot.record \
+lerobot-record \
  --robot.type=hope_jr_hand \
  --robot.port=/dev/tty.usbmodem58760432281 \
  --robot.id=right \
@@ -45,7 +45,7 @@ Note that the `id` associated with a robot is used to store the calibration file
 <hfoptions id="teleoperate_so101">
 <hfoption id="Command">
 ```bash
-python -m lerobot.teleoperate \
+lerobot-teleoperate \
    --robot.type=so101_follower \
    --robot.port=/dev/tty.usbmodem58760431541 \
    --robot.id=my_awesome_follower_arm \
@@ -101,7 +101,7 @@ With `rerun`, you can teleoperate again while simultaneously visualizing the cam
 <hfoptions id="teleoperate_koch_camera">
 <hfoption id="Command">
 ```bash
-python -m lerobot.teleoperate \
+lerobot-teleoperate \
    --robot.type=koch_follower \
    --robot.port=/dev/tty.usbmodem58760431541 \
    --robot.id=my_awesome_follower_arm \
@@ -174,7 +174,7 @@ Now you can record a dataset. To record 5 episodes and upload your dataset to th
 <hfoptions id="record">
 <hfoption id="Command">
 ```bash
-python -m lerobot.record \
+lerobot-record \
    --robot.type=so101_follower \
    --robot.port=/dev/tty.usbmodem585A0076841 \
    --robot.id=my_awesome_follower_arm \
@@ -294,7 +294,7 @@ dataset.push_to_hub()

 #### Dataset upload

-Locally, your dataset is stored in this folder: `~/.cache/huggingface/lerobot/{repo-id}`. At the end of data recording, your dataset will be uploaded on your Hugging Face page (e.g. https://huggingface.co/datasets/cadene/so101_test) that you can obtain by running:
+Locally, your dataset is stored in this folder: `~/.cache/huggingface/lerobot/{repo-id}`. At the end of data recording, your dataset will be uploaded on your Hugging Face page (e.g. `https://huggingface.co/datasets/${HF_USER}/so101_test`) that you can obtain by running:

 ```bash
 echo https://huggingface.co/datasets/${HF_USER}/so101_test
@@ -376,7 +376,7 @@ You can replay the first episode on your robot with either the command below or
 <hfoptions id="replay">
 <hfoption id="Command">
 ```bash
-python -m lerobot.replay \
+lerobot-replay \
    --robot.type=so101_follower \
    --robot.port=/dev/tty.usbmodem58760431541 \
    --robot.id=my_awesome_follower_arm \
@@ -428,10 +428,10 @@ Your robot should replicate movements similar to those you recorded. For example

 ## Train a policy

-To train a policy to control your robot, use the [`python -m lerobot.scripts.train`](../src/lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
+To train a policy to control your robot, use the [`lerobot-train`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:

 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
  --dataset.repo_id=${HF_USER}/so101_test \
  --policy.type=act \
  --output_dir=outputs/train/act_so101_test \
@@ -444,7 +444,7 @@ python -m lerobot.scripts.train \
 Let's explain the command:

 1. We provided the dataset as argument with `--dataset.repo_id=${HF_USER}/so101_test`.
-2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../src/lerobot/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor states, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
+2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor states, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
 3. We provided `policy.device=cuda` since we are training on a Nvidia GPU, but you could use `policy.device=mps` to train on Apple silicon.
 4. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.

@@ -453,7 +453,7 @@ Training should take several hours. You will find checkpoints in `outputs/train/
 To resume training from a checkpoint, below is an example command to resume from `last` checkpoint of the `act_so101_test` policy:

 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
  --config_path=outputs/train/act_so101_test/checkpoints/last/pretrained_model/train_config.json \
  --resume=true
 ```
@@ -490,7 +490,7 @@ You can use the `record` script from [`lerobot/record.py`](https://github.com/hu
 <hfoptions id="eval">
 <hfoption id="Command">
 ```bash
-python -m lerobot.record  \
+lerobot-record  \
  --robot.type=so100_follower \
  --robot.port=/dev/ttyACM1 \
  --robot.cameras="{ up: {type: opencv, index_or_path: /dev/video10, width: 640, height: 480, fps: 30}, side: {type: intelrealsense, serial_number_or_name: 233522074606, width: 640, height: 480, fps: 30}}" \
@@ -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 [`python -m lerobot.scripts.train`](../src/lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
+To train a policy to control your robot, use the [`lerobot-train`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:

 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
  --dataset.repo_id=${HF_USER}/il_gym \
  --policy.type=act \
  --output_dir=outputs/train/il_sim_test \
@@ -111,7 +111,7 @@ python -m lerobot.scripts.train \
 Let's explain the command:

 1. We provided the dataset as argument with `--dataset.repo_id=${HF_USER}/il_gym`.
-2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../src/lerobot/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor states, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
+2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor states, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
 3. We provided `policy.device=cuda` since we are training on a Nvidia GPU, but you could use `policy.device=mps` to train on Apple silicon.
 4. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.

@@ -1,15 +1,6 @@
 # Installation

-## Install LeRobot
-
-Currently only available from source.
-
-Download our source code:
-
-```bash
-git clone https://github.com/huggingface/lerobot.git
-cd lerobot
-```
+## Environment Setup

 Create a virtual environment with Python 3.10, using [`Miniconda`](https://docs.anaconda.com/miniconda/install/#quick-command-line-install)

@@ -40,12 +31,49 @@ conda install ffmpeg -c conda-forge
 >
 > - _[On Linux only]_ If you want to bring your own ffmpeg: Install [ffmpeg build dependencies](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu#GettheDependencies) and [compile ffmpeg from source with libsvtav1](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu#libsvtav1), and make sure you use the corresponding ffmpeg binary to your install with `which ffmpeg`.

-Install 🤗 LeRobot:
+## Install LeRobot 🤗
+
+### From Source
+
+First, clone the repository and navigate into the directory:
+
+```bash
+git clone https://github.com/huggingface/lerobot.git
+cd lerobot
+```
+
+Then, install the library in editable mode. This is useful if you plan to contribute to the code.

 ```bash
 pip install -e .
 ```

+### Installation from PyPI
+
+**Core Library:**
+Install the base package with:
+
+```bash
+pip install lerobot
+```
+
+_This installs only the default dependencies._
+
+**Extra Features:**
+To install additional functionality, use one of the following:
+
+```bash
+pip install 'lerobot[all]'          # All available features
+pip install 'lerobot[aloha,pusht]'  # Specific features (Aloha & Pusht)
+pip install 'lerobot[feetech]'      # Feetech motor support
+```
+
+_Replace `[...]` with your desired features._
+
+**Available Tags:**
+For a full list of optional dependencies, see:
+https://pypi.org/project/lerobot/
+
 ### Troubleshooting

 If you encounter build errors, you may need to install additional dependencies: `cmake`, `build-essential`, and `ffmpeg libs`.
@@ -31,7 +31,7 @@ pip install -e ".[dynamixel]"
 To find the port for each bus servo adapter, run this script:

 ```bash
-python -m lerobot.find_port
+lerobot-find-port
 ```

 <hfoptions id="example">
@@ -98,7 +98,7 @@ For a visual reference on how to set the motor ids please refer to [this video](
 <hfoption id="Command">

 ```bash
-python -m lerobot.setup_motors \
+lerobot-setup-motors \
    --robot.type=koch_follower \
    --robot.port=/dev/tty.usbmodem575E0031751 # <- paste here the port found at previous step
 ```
@@ -174,7 +174,7 @@ Do the same steps for the leader arm but modify the command or script accordingl
 <hfoption id="Command">

 ```bash
-python -m lerobot.setup_motors \
+lerobot-setup-motors \
    --teleop.type=koch_leader \
    --teleop.port=/dev/tty.usbmodem575E0031751 \  # <- paste here the port found at previous step
 ```
@@ -211,7 +211,7 @@ Run the following command or API example to calibrate the follower arm:
 <hfoption id="Command">

 ```bash
-python -m lerobot.calibrate \
+lerobot-calibrate \
    --robot.type=koch_follower \
    --robot.port=/dev/tty.usbmodem58760431551 \ # <- The port of your robot
    --robot.id=my_awesome_follower_arm  # <- Give the robot a unique name
@@ -249,7 +249,7 @@ Do the same steps to calibrate the leader arm, run the following command or API
 <hfoption id="Command">

 ```bash
-python -m lerobot.calibrate \
+lerobot-calibrate \
    --teleop.type=koch_leader \
    --teleop.port=/dev/tty.usbmodem58760431551 \ # <- The port of your robot
    --teleop.id=my_awesome_leader_arm  # <- Give the robot a unique name
@@ -60,7 +60,7 @@ First, we will assemble the two SO100/SO101 arms. One to attach to the mobile ba
 To find the port for each bus servo adapter, run this script:

 ```bash
-python -m lerobot.find_port
+lerobot-find-port
 ```

 <hfoptions id="example">
@@ -116,7 +116,7 @@ The instructions for configuring the motors can be found in the SO101 [docs](./s
 You can run this command to setup motors for LeKiwi. It will first setup the motors for arm (id 6..1) and then setup motors for wheels (9,8,7)

 ```bash
-python -m lerobot.setup_motors \
+lerobot-setup-motors \
    --robot.type=lekiwi \
    --robot.port=/dev/tty.usbmodem58760431551 # <- paste here the port found at previous step
 ```
@@ -174,7 +174,7 @@ The calibration process is very important because it allows a neural network tra
 Make sure the arm is connected to the Raspberry Pi and run this script or API example (on the Raspberry Pi via SSH) to launch calibration of the follower arm:

 ```bash
-python -m lerobot.calibrate \
+lerobot-calibrate \
    --robot.type=lekiwi \
    --robot.id=my_awesome_kiwi # <- Give the robot a unique name
 ```
@@ -193,7 +193,7 @@ Then, to calibrate the leader arm (which is attached to the laptop/pc). Run the
 <hfoption id="Command">

 ```bash
-python -m lerobot.calibrate \
+lerobot-calibrate \
    --teleop.type=so100_leader \
    --teleop.port=/dev/tty.usbmodem58760431551 \ # <- The port of your robot
    --teleop.id=my_awesome_leader_arm # <- Give the robot a unique name
@@ -0,0 +1,169 @@
+# 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.
@@ -0,0 +1,321 @@
+# 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
@@ -0,0 +1,288 @@
+# Reachy 2
+
+Reachy 2 is an open-source humanoid robot made by Pollen Robotics, specifically designed for the development of embodied AI and real-world applications.
+Check out [Pollen Robotics website](https://www.pollen-robotics.com/reachy/), or access [Reachy 2 documentation](https://docs.pollen-robotics.com/) for more information on the platform!
+
+## Teleoperate Reachy 2
+
+Currently, there are two ways to teleoperate Reachy 2:
+
+- Pollen Robotics’ VR teleoperation (not included in LeRobot).
+- Robot-to-robot teleoperation (use one Reachy 2 to control another).
+
+## Reachy 2 Simulation
+
+**(Linux only)** You can run Reachy 2 in simulation (Gazebo or MuJoCo) using the provided [Docker image](https://hub.docker.com/r/pollenrobotics/reachy2_core).
+
+1. Install [Docker Engine](https://docs.docker.com/engine/).
+2. Run (for MuJoCo):
+
+```
+docker run --rm -it \
+  --name reachy \
+  --privileged \
+  --network host \
+  --ipc host \
+  --device-cgroup-rule='c 189:* rwm' \
+  --group-add audio \
+  -e ROS_DOMAIN_ID="$ROS_DOMAIN_ID" \
+  -e DISPLAY="$DISPLAY" \
+  -e RCUTILS_CONSOLE_OUTPUT_FORMAT="[{severity}]: {message}" \
+  -e REACHY2_CORE_SERVICE_FAKE="${REACHY2_CORE_SERVICE_FAKE:-true}" \
+  -v /dev:/dev \
+  -v "$HOME/.reachy_config":/home/reachy/.reachy_config_override \
+  -v "$HOME/.reachy.log":/home/reachy/.ros/log \
+  -v /usr/lib/x86_64-linux-gnu:/opt/host-libs \
+  --entrypoint /package/launch.sh \
+  pollenrobotics/reachy2_core:1.7.5.9_deploy \
+  start_rviz:=true start_sdk_server:=true mujoco:=true
+```
+
+> If MuJoCo runs slowly (low simulation frequency), append `-e LD_LIBRARY_PATH="/opt/host-libs:$LD_LIBRARY_PATH" \` to the previous command to improve performance:
+>
+> ```
+> docker run --rm -it \
+>   --name reachy \
+>   --privileged \
+>   --network host \
+>   --ipc host \
+>   --device-cgroup-rule='c 189:* rwm' \
+>   --group-add audio \
+>   -e ROS_DOMAIN_ID="$ROS_DOMAIN_ID" \
+>   -e DISPLAY="$DISPLAY" \
+>   -e RCUTILS_CONSOLE_OUTPUT_FORMAT="[{severity}]: {message}" \
+>   -e REACHY2_CORE_SERVICE_FAKE="${REACHY2_CORE_SERVICE_FAKE:-true}" \
+>   -e LD_LIBRARY_PATH="/opt/host-libs:$LD_LIBRARY_PATH" \
+>   -v /dev:/dev \
+>   -v "$HOME/.reachy_config":/home/reachy/.reachy_config_override \
+>   -v "$HOME/.reachy.log":/home/reachy/.ros/log \
+>   -v /usr/lib/x86_64-linux-gnu:/opt/host-libs \
+>   --entrypoint /package/launch.sh \
+>   pollenrobotics/reachy2_core:1.7.5.9_deploy \
+>   start_rviz:=true start_sdk_server:=true mujoco:=true
+> ```
+
+## Setup
+
+### Prerequisites
+
+- On your robot, check the **service images** meet the minimum versions:
+  - **reachy2-core >= 1.7.5.2**
+  - **webrtc >= 2.0.1.1**
+
+Then, if you want to use VR teleoperation:
+
+- Install the [Reachy 2 teleoperation application](https://docs.pollen-robotics.com/teleoperation/teleoperation-introduction/discover-teleoperation/).
+  Use version **>=v1.2.0**
+
+We recommend using two computers: one for teleoperation (Windows required) and another for recording with LeRobot.
+
+### Install LeRobot
+
+Follow the [installation instructions](https://github.com/huggingface/lerobot#installation) to install LeRobot.
+
+Install LeRobot with Reachy 2 dependencies:
+
+```bash
+pip install -e ".[reachy2]"
+```
+
+### (Optional but recommended) Install pollen_data_acquisition_server
+
+How you manage Reachy 2 recording sessions is up to you, but the **easiest** way is to use this server so you can control sessions directly from the VR teleoperation app.
+
+> **Note:** Currently, only the VR teleoperation application works as a client for this server, so this step primarily targets teleoperation. You’re free to develop custom clients to manage sessions to your needs.
+
+In your LeRobot environment, install the server from source:
+
+```bash
+git clone https://github.com/pollen-robotics/pollen_data_acquisition_server.git
+cd pollen_data_acquisition_server
+pip install -e .
+```
+
+Find the [pollen_data_acquisition_server documentation here](https://github.com/pollen-robotics/pollen_data_acquisition_server).
+
+## Step 1: Recording
+
+### Get Reachy 2 IP address
+
+Before starting teleoperation and data recording, find the [robot's IP address](https://docs.pollen-robotics.com/getting-started/setup-reachy2/connect-reachy2/).
+We strongly recommend connecting all devices (PC and robot) via **Ethernet**.
+
+### Launch recording
+
+There are two ways to manage recording sessions when using the Reachy 2 VR teleoperation application:
+
+- **Using the data acquisition server (recommended for VR teleop)**: The VR app orchestrates sessions (via the server it tells LeRobot when to create datasets, start/stop episodes) while also controlling the robot’s motions.
+- **Using LeRobot’s record script**: LeRobot owns session control and decides when to start/stop episodes. If you also use the VR teleop app, it’s only for motion control.
+
+### Option 1: Using Pollen data acquisition server (recommended for VR teleop)
+
+Make sure you have installed pollen_data_acquisition_server, as explained in the Setup section.
+
+Launch the data acquisition server to be able to manage your session directly from the teleoperation application:
+
+```bash
+python -m pollen_data_acquisition_server.server
+```
+
+Then get into the teleoperation application and choose "Data acquisition session".
+You can finally setup your session by following the screens displayed.
+
+> Even without the VR app, you can use the `pollen_data_acquisition_server` with your own client implementation.
+
+### Option 2: Using lerobot.record
+
+Reachy 2 is fully supported by LeRobot’s recording features.
+If you choose this option but still want to use the VR teleoperation application, select "Standard session" in the app.
+
+**Example: start a recording without the mobile base:**
+First add reachy2 and reachy2_teleoperator to the imports of the record script. Then you can use the following command:
+
+```bash
+python -m lerobot.record \
+    --robot.type=reachy2 \
+    --robot.ip_address=192.168.0.200 \
+    --robot.id=r2-0000 \
+    --robot.use_external_commands=true \
+    --robot.with_mobile_base=false \
+    --teleop.type=reachy2_teleoperator \
+    --teleop.ip_address=192.168.0.200 \
+    --teleop.with_mobile_base=false \
+    --dataset.repo_id=pollen_robotics/record_test \
+    --dataset.single_task="Reachy 2 recording test" \
+    --dataset.num_episodes=1 \
+    --dataset.episode_time_s=5 \
+    --dataset.fps=15 \
+    --dataset.push_to_hub=true \
+    --dataset.private=true \
+    --display_data=true
+```
+
+#### Specific Options
+
+**Extended setup overview (all options included):**
+
+```bash
+python -m lerobot.record \
+    --robot.type=reachy2 \
+    --robot.ip_address=192.168.0.200 \
+    --robot.use_external_commands=true \
+    --robot.with_mobile_base=true \
+    --robot.with_l_arm=true \
+    --robot.with_r_arm=true \
+    --robot.with_neck=true \
+    --robot.with_antennas=true \
+    --robot.with_left_teleop_camera=true \
+    --robot.with_right_teleop_camera=true \
+    --robot.with_torso_camera=false \
+    --robot.disable_torque_on_disconnect=false \
+    --robot.max_relative_target=5.0 \
+    --teleop.type=reachy2_teleoperator \
+    --teleop.ip_address=192.168.0.200 \
+    --teleop.use_present_position=false \
+    --teleop.with_mobile_base=false \
+    --teleop.with_l_arm=true \
+    --teleop.with_r_arm=true \
+    --teleop.with_neck=true \
+    --teleop.with_antennas=true \
+    --dataset.repo_id=pollen_robotics/record_test \
+    --dataset.single_task="Reachy 2 recording test" \
+    --dataset.num_episodes=1 \
+    --dataset.episode_time_s=5 \
+    --dataset.fps=15 \
+    --dataset.push_to_hub=true \
+    --dataset.private=true \
+    --display_data=true
+```
+
+##### `--robot.use_external_commands`
+
+Determine whether LeRobot robot.send_action() sends commands to the robot.
+**Must** be set to false while using the VR teleoperation application, as the app already sends commands.
+
+##### `--teleop.use_present_position`
+
+Determine whether the teleoperator reads the goal or present position of the robot.
+Must be set to true if a compliant Reachy 2 is used to control another one.
+
+##### Use the relevant parts
+
+From our initial tests, recording **all** joints when only some are moving can reduce model quality with certain policies.
+To avoid this, you can exclude specific parts from recording and replay using:
+
+````
+--robot.with_<part>=false
+```,
+with `<part>` being one of : `mobile_base`, `l_arm`, `r_arm", `neck`, `antennas`.
+It determine whether the corresponding part is recorded in the observations. True if not set.
+
+By default, **all parts are recorded**.
+
+The same per-part mechanism is available in `reachy2_teleoperator` as well.
+
+````
+
+--teleop.with\_<part>
+
+```
+with `<part>` being one of : `mobile_base`, `l_arm`, `r_arm", `neck`, `antennas`.
+Determine whether the corresponding part is recorded in the actions. True if not set.
+
+> **Important:** In a given session, the **enabled parts must match** on both the robot and the teleoperator.
+For example, if the robot runs with `--robot.with_mobile_base=false`, the teleoperator must disable the same part `--teleoperator.with_mobile_base=false`.
+
+##### Use the relevant cameras
+
+You can do the same for **cameras**. By default, only the **teleoperation cameras** are recorded (both `left_teleop_camera` and `right_teleop_camera`). Enable or disable each camera with:
+
+```
+
+--robot.with_left_teleop_camera=<true|false>
+--robot.with_right_teleop_camera=<true|false>
+--robot.with_torso_camera=<true|false>
+
+````
+
+
+## Step 2: Replay
+
+Make sure the robot is configured with the same parts as the dataset:
+
+```bash
+python -m lerobot.replay \
+    --robot.type=reachy2 \
+    --robot.ip_address=192.168.0.200 \
+    --robot.use_external_commands=false \
+    --robot.with_mobile_base=false \
+    --dataset.repo_id=pollen_robotics/record_test \
+    --dataset.episode=0
+    --display_data=true
+````
+
+## Step 3: Train
+
+```bash
+python -m lerobot.scripts.train \
+  --dataset.repo_id=pollen_robotics/record_test \
+  --policy.type=act \
+  --output_dir=outputs/train/reachy2_test \
+  --job_name=reachy2 \
+  --policy.device=mps \
+  --wandb.enable=true \
+  --policy.repo_id=pollen_robotics/record_test_policy
+```
+
+## Step 4: Evaluate
+
+```bash
+python -m lerobot.record \
+  --robot.type=reachy2 \
+  --robot.ip_address=192.168.0.200 \
+  --display_data=false \
+  --dataset.repo_id=pollen_robotics/eval_record_test \
+  --dataset.single_task="Evaluate reachy2 policy" \
+  --dataset.num_episodes=10 \
+  --policy.path=outputs/train/reachy2_test/checkpoints/last/pretrained_model
+```
@@ -54,7 +54,7 @@ If you don't have a gpu device, you can train using our notebook on [![Google Co
 Pass your dataset to the training script using `--dataset.repo_id`. If you want to test your installation, run the following command where we use one of the datasets we collected for the [SmolVLA Paper](https://huggingface.co/papers/2506.01844).

 ```bash
-cd lerobot && python -m lerobot.scripts.train \
+cd lerobot && lerobot-train \
  --policy.path=lerobot/smolvla_base \
  --dataset.repo_id=${HF_USER}/mydataset \
  --batch_size=64 \
@@ -73,7 +73,7 @@ cd lerobot && python -m lerobot.scripts.train \
 Fine-tuning is an art. For a complete overview of the options for finetuning, run

 ```bash
-python -m lerobot.scripts.train --help
+lerobot-train --help
 ```

 <p align="center">
@@ -97,7 +97,7 @@ Similarly for when recording an episode, it is recommended that you are logged i
 Once you are logged in, you can run inference in your setup by doing:

 ```bash
-python -m lerobot.record \
+lerobot-record \
  --robot.type=so101_follower \
  --robot.port=/dev/ttyACM0 \ # <- Use your port
  --robot.id=my_blue_follower_arm \ # <- Use your robot id
@@ -26,7 +26,7 @@ Unlike the SO-101, the motor connectors are not easily accessible once the arm i
 To find the port for each bus servo adapter, run this script:

 ```bash
-python -m lerobot.find_port
+lerobot-find-port
 ```

 <hfoptions id="example">
@@ -93,7 +93,7 @@ For a visual reference on how to set the motor ids please refer to [this video](
 <hfoption id="Command">

 ```bash
-python -m lerobot.setup_motors \
+lerobot-setup-motors \
    --robot.type=so100_follower \
    --robot.port=/dev/tty.usbmodem585A0076841  # <- paste here the port found at previous step
 ```
@@ -168,7 +168,7 @@ Do the same steps for the leader arm.
 <hfoptions id="setup_motors">
 <hfoption id="Command">
 ```bash
-python -m lerobot.setup_motors \
+lerobot-setup-motors \
    --teleop.type=so100_leader \
    --teleop.port=/dev/tty.usbmodem575E0031751  # <- paste here the port found at previous step
 ```
@@ -568,7 +568,7 @@ Run the following command or API example to calibrate the follower arm:
 <hfoption id="Command">

 ```bash
-python -m lerobot.calibrate \
+lerobot-calibrate \
    --robot.type=so100_follower \
    --robot.port=/dev/tty.usbmodem58760431551 \ # <- The port of your robot
    --robot.id=my_awesome_follower_arm # <- Give the robot a unique name
@@ -606,7 +606,7 @@ Do the same steps to calibrate the leader arm, run the following command or API
 <hfoption id="Command">

 ```bash
-python -m lerobot.calibrate \
+lerobot-calibrate \
    --teleop.type=so100_leader \
    --teleop.port=/dev/tty.usbmodem58760431551 \ # <- The port of your robot
    --teleop.id=my_awesome_leader_arm # <- Give the robot a unique name
@@ -162,7 +162,7 @@ It is advisable to install one 3-pin cable in the motor after placing them befor
 To find the port for each bus servo adapter, connect MotorBus to your computer via USB and power. Run the following script and disconnect the MotorBus when prompted:

 ```bash
-python -m lerobot.find_port
+lerobot-find-port
 ```

 <hfoptions id="example">
@@ -240,7 +240,7 @@ Connect the usb cable from your computer and the power supply to the follower ar
 <hfoption id="Command">

 ```bash
-python -m lerobot.setup_motors \
+lerobot-setup-motors \
    --robot.type=so101_follower \
    --robot.port=/dev/tty.usbmodem585A0076841  # <- paste here the port found at previous step
 ```
@@ -316,7 +316,7 @@ Do the same steps for the leader arm.
 <hfoption id="Command">

 ```bash
-python -m lerobot.setup_motors \
+lerobot-setup-motors \
    --teleop.type=so101_leader \
    --teleop.port=/dev/tty.usbmodem575E0031751  # <- paste here the port found at previous step
 ```
@@ -353,7 +353,7 @@ Run the following command or API example to calibrate the follower arm:
 <hfoption id="Command">

 ```bash
-python -m lerobot.calibrate \
+lerobot-calibrate \
    --robot.type=so101_follower \
    --robot.port=/dev/tty.usbmodem58760431551 \ # <- The port of your robot
    --robot.id=my_awesome_follower_arm # <- Give the robot a unique name
@@ -402,7 +402,7 @@ Do the same steps to calibrate the leader arm, run the following command or API
 <hfoption id="Command">

 ```bash
-python -m lerobot.calibrate \
+lerobot-calibrate \
    --teleop.type=so101_leader \
    --teleop.port=/dev/tty.usbmodem58760431551 \ # <- The port of your robot
    --teleop.id=my_awesome_leader_arm # <- Give the robot a unique name
@@ -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 the episode_data_index. Here, we access
+# episodes, you can access the frame indices of any episode using dataset.meta.episodes. Here, we access
 # frame indices associated to the first episode:
 episode_index = 0
-from_idx = dataset.episode_data_index["from"][episode_index].item()
-to_idx = dataset.episode_data_index["to"][episode_index].item()
+from_idx = dataset.meta.episodes["dataset_from_index"][episode_index]
+to_idx = dataset.meta.episodes["dataset_to_index"][episode_index]

 # Then we grab all the image frames from the first camera:
 camera_key = dataset.meta.camera_keys[0]
@@ -62,7 +62,7 @@ By default, every field takes its default value specified in the dataclass. If a
 Let's say that we want to train [Diffusion Policy](../src/lerobot/policies/diffusion) on the [pusht](https://huggingface.co/datasets/lerobot/pusht) dataset, using the [gym_pusht](https://github.com/huggingface/gym-pusht) environment for evaluation. The command to do so would look like this:

 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
    --dataset.repo_id=lerobot/pusht \
    --policy.type=diffusion \
    --env.type=pusht
@@ -77,7 +77,7 @@ Let's break this down:
 Let's see another example. Let's say you've been training [ACT](../src/lerobot/policies/act) on [lerobot/aloha_sim_insertion_human](https://huggingface.co/datasets/lerobot/aloha_sim_insertion_human) using the [gym-aloha](https://github.com/huggingface/gym-aloha) environment for evaluation with:

 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
    --policy.type=act \
    --dataset.repo_id=lerobot/aloha_sim_insertion_human \
    --env.type=aloha \
@@ -90,7 +90,7 @@ We now want to train a different policy for aloha on another task. We'll change
 Looking at the [`AlohaEnv`](../src/lerobot/envs/configs.py) config, the task is `"AlohaInsertion-v0"` by default, which corresponds to the task we trained on in the command above. The [gym-aloha](https://github.com/huggingface/gym-aloha?tab=readme-ov-file#description) environment also has the `AlohaTransferCube-v0` task which corresponds to this other task we want to train on. Putting this together, we can train this new policy on this different task using:

 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
    --policy.type=act \
    --dataset.repo_id=lerobot/aloha_sim_transfer_cube_human \
    --env.type=aloha \
@@ -127,7 +127,7 @@ Now, let's assume that we want to reproduce the run just above. That run has pro
 We can then simply load the config values from this file using:

 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
    --config_path=outputs/train/act_aloha_transfer/checkpoints/last/pretrained_model/ \
    --output_dir=outputs/train/act_aloha_transfer_2
 ```
@@ -137,7 +137,7 @@ python -m lerobot.scripts.train \
 Similarly to Hydra, we can still override some parameters in the CLI if we want to, e.g.:

 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
    --config_path=outputs/train/act_aloha_transfer/checkpoints/last/pretrained_model/ \
    --output_dir=outputs/train/act_aloha_transfer_2
    --policy.n_action_steps=80
@@ -148,7 +148,7 @@ python -m lerobot.scripts.train \
 `--config_path` can also accept the repo_id of a repo on the hub that contains a `train_config.json` file, e.g. running:

 ```bash
-python -m lerobot.scripts.train --config_path=lerobot/diffusion_pusht
+lerobot-train --config_path=lerobot/diffusion_pusht
 ```

 will start a training run with the same configuration used for training [lerobot/diffusion_pusht](https://huggingface.co/lerobot/diffusion_pusht)
@@ -160,7 +160,7 @@ Being able to resume a training run is important in case it crashed or aborted f
 Let's reuse the command from the previous run and add a few more options:

 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
    --policy.type=act \
    --dataset.repo_id=lerobot/aloha_sim_transfer_cube_human \
    --env.type=aloha \
@@ -179,7 +179,7 @@ INFO 2025-01-24 16:10:56 ts/train.py:263 Checkpoint policy after step 100
 Now let's simulate a crash by killing the process (hit `ctrl`+`c`). We can then simply resume this run from the last checkpoint available with:

 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
    --config_path=outputs/train/run_resumption/checkpoints/last/pretrained_model/ \
    --resume=true
 ```
@@ -190,7 +190,7 @@ Another reason for which you might want to resume a run is simply to extend trai
 You could double the number of steps of the previous run with:

 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
    --config_path=outputs/train/run_resumption/checkpoints/last/pretrained_model/ \
    --resume=true \
    --steps=200000
@@ -224,7 +224,7 @@ In addition to the features currently in Draccus, we've added a special `.path`
 For example, we could fine-tune a [policy pre-trained on the aloha transfer task](https://huggingface.co/lerobot/act_aloha_sim_transfer_cube_human) on the aloha insertion task. We can achieve this with:

 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
    --policy.path=lerobot/act_aloha_sim_transfer_cube_human \
    --dataset.repo_id=lerobot/aloha_sim_insertion_human \
    --env.type=aloha \
@@ -270,7 +270,7 @@ We'll summarize here the main use cases to remember from this tutorial.
 #### Train a policy from scratch – CLI

 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
    --policy.type=act \  # <- select 'act' policy
    --env.type=pusht \  # <- select 'pusht' environment
    --dataset.repo_id=lerobot/pusht  # <- train on this dataset
@@ -279,7 +279,7 @@ python -m lerobot.scripts.train \
 #### Train a policy from scratch - config file + CLI

 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
    --config_path=path/to/pretrained_model \  # <- can also be a repo_id
    --policy.n_action_steps=80  # <- you may still override values
 ```
@@ -287,7 +287,7 @@ python -m lerobot.scripts.train \
 #### Resume/continue a training run

 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
    --config_path=checkpoint/pretrained_model/ \
    --resume=true \
    --steps=200000  # <- you can change some training parameters
@@ -296,7 +296,7 @@ python -m lerobot.scripts.train \
 #### Fine-tuning

 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
    --policy.path=lerobot/act_aloha_sim_transfer_cube_human \  # <- can also be a local path to a checkpoint
    --dataset.repo_id=lerobot/aloha_sim_insertion_human \
    --env.type=aloha \
@@ -0,0 +1,116 @@
+# 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()
@@ -18,7 +18,7 @@ Replays the actions of an episode from a dataset on a robot.
 Example:

 ```shell
-python -m lerobot.replay \
+lerobot-replay \
    --robot.type=so100_follower \
    --robot.port=/dev/tty.usbmodem58760431541 \
    --robot.id=black \
@@ -0,0 +1,85 @@
+#!/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()
@@ -0,0 +1,430 @@
+#!/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()
@@ -0,0 +1,148 @@
+#!/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()
@@ -0,0 +1,162 @@
+#!/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()
@@ -0,0 +1,281 @@
+#!/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()
@@ -25,7 +25,7 @@ discord = "https://discord.gg/s3KuuzsPFb"

 [project]
 name = "lerobot"
-version = "0.3.2"
+version = "0.3.4"
 description = "🤗 LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch"
 readme = "README.md"
 license = { text = "Apache-2.0" }
@@ -68,22 +68,22 @@ dependencies = [
    "einops>=0.8.0",
    "opencv-python-headless>=4.9.0",
    "av>=14.2.0",
-    "torch>=2.2.1",
-    "torchcodec>=0.2.1; 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')",
-    "torchvision>=0.21.0",
    "jsonlines>=4.0.0",
    "packaging>=24.2",
    "pynput>=1.7.7",
    "pyserial>=3.5",
    "wandb>=0.20.0",

+    "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
+    "torchvision>=0.21.0,<0.23.0", # TODO: Bumb dependency
+
    "draccus==0.10.0", # TODO: Remove ==
    "gymnasium>=0.29.1,<1.0.0", # TODO: Bumb dependency
    "rerun-sdk>=0.21.0,<0.23.0", # TODO: Bumb dependency

    # 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",
 ]
@@ -105,6 +105,7 @@ 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'",
@@ -140,6 +141,7 @@ all = [
    "lerobot[gamepad]",
    "lerobot[hopejr]",
    "lerobot[lekiwi]",
+    "lerobot[reachy2]",
    "lerobot[kinematics]",
    "lerobot[intelrealsense]",
    "lerobot[pi0]",
@@ -18,7 +18,7 @@ Helper to recalibrate your device (robot or teleoperator).
 Example:

 ```shell
-python -m lerobot.calibrate \
+lerobot-calibrate \
    --teleop.type=so100_leader \
    --teleop.port=/dev/tty.usbmodem58760431551 \
    --teleop.id=blue
@@ -60,7 +60,7 @@ class OpenCVCamera(Camera):
    or port changes, especially on Linux. Use the provided utility script to find
    available camera indices or paths:
    ```bash
-    python -m lerobot.find_cameras opencv
+    lerobot-find-cameras opencv
    ```

    The camera's default settings (FPS, resolution, color mode) are used unless
@@ -165,8 +165,7 @@ class OpenCVCamera(Camera):
            self.videocapture.release()
            self.videocapture = None
            raise ConnectionError(
-                f"Failed to open {self}."
-                f"Run `python -m lerobot.find_cameras opencv` to find available cameras."
+                f"Failed to open {self}.Run `lerobot-find-cameras opencv` to find available cameras."
            )

        self._configure_capture_settings()
@@ -0,0 +1,16 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .configuration_reachy2_camera import Reachy2CameraConfig
+from .reachy2_camera import Reachy2Camera
@@ -0,0 +1,78 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+
+from ..configs import CameraConfig, ColorMode
+
+
+@CameraConfig.register_subclass("reachy2_camera")
+@dataclass
+class Reachy2CameraConfig(CameraConfig):
+    """Configuration class for Reachy 2 camera devices.
+
+    This class provides configuration options for Reachy 2 cameras,
+    supporting both the teleop and depth cameras. It includes settings
+    for resolution, frame rate, color mode, and the selection of the cameras.
+
+    Example configurations:
+    ```python
+    # Basic configurations
+    Reachy2CameraConfig(
+        name="teleop",
+        image_type="left",
+        ip_address="192.168.0.200",  # IP address of the robot
+        fps=15,
+        width=640,
+        height=480,
+        color_mode=ColorMode.RGB,
+    )  # Left teleop camera, 640x480 @ 15FPS
+    ```
+
+    Attributes:
+        name: Name of the camera device. Can be "teleop" or "depth".
+        image_type: Type of image stream. For "teleop" camera, can be "left" or "right".
+                    For "depth" camera, can be "rgb" or "depth". (depth is not supported yet)
+        fps: Requested frames per second for the color stream.
+        width: Requested frame width in pixels for the color stream.
+        height: Requested frame height in pixels for the color stream.
+        color_mode: Color mode for image output (RGB or BGR). Defaults to RGB.
+        ip_address: IP address of the robot. Defaults to "localhost".
+        port: Port number for the camera server. Defaults to 50065.
+
+    Note:
+        - Only 3-channel color output (RGB/BGR) is currently supported.
+    """
+
+    name: str
+    image_type: str
+    color_mode: ColorMode = ColorMode.RGB
+    ip_address: str | None = "localhost"
+    port: int = 50065
+    # use_depth: bool = False
+
+    def __post_init__(self):
+        if self.name not in ["teleop", "depth"]:
+            raise ValueError(f"`name` is expected to be 'teleop' or 'depth', but {self.name} is provided.")
+        if (self.name == "teleop" and self.image_type not in ["left", "right"]) or (
+            self.name == "depth" and self.image_type not in ["rgb", "depth"]
+        ):
+            raise ValueError(
+                f"`image_type` is expected to be 'left' or 'right' for teleop camera, and 'rgb' or 'depth' for depth camera, but {self.image_type} is provided."
+            )
+
+        if self.color_mode not in ["rgb", "bgr"]:
+            raise ValueError(
+                f"`color_mode` is expected to be 'rgb' or 'bgr', but {self.color_mode} is provided."
+            )
@@ -0,0 +1,288 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Provides the Reachy2Camera class for capturing frames from Reachy 2 cameras using Reachy 2's CameraManager.
+"""
+
+import logging
+import os
+import platform
+import time
+from threading import Event, Lock, Thread
+from typing import Any
+
+# Fix MSMF hardware transform compatibility for Windows before importing cv2
+if platform.system() == "Windows" and "OPENCV_VIDEOIO_MSMF_ENABLE_HW_TRANSFORMS" not in os.environ:
+    os.environ["OPENCV_VIDEOIO_MSMF_ENABLE_HW_TRANSFORMS"] = "0"
+import cv2
+import numpy as np
+from reachy2_sdk.media.camera import CameraView
+from reachy2_sdk.media.camera_manager import CameraManager
+
+from lerobot.errors import DeviceNotConnectedError
+
+from ..camera import Camera
+from .configuration_reachy2_camera import ColorMode, Reachy2CameraConfig
+
+logger = logging.getLogger(__name__)
+
+
+class Reachy2Camera(Camera):
+    """
+    Manages Reachy 2 camera using Reachy 2 CameraManager.
+
+    This class provides a high-level interface to connect to, configure, and read
+    frames from Reachy 2 cameras. It supports both synchronous and asynchronous
+    frame reading.
+
+    An Reachy2Camera instance requires a camera name (e.g., "teleop") and an image
+    type (e.g., "left") to be specified in the configuration.
+
+    The camera's default settings (FPS, resolution, color mode) are used unless
+    overridden in the configuration.
+    """
+
+    def __init__(self, config: Reachy2CameraConfig):
+        """
+        Initializes the Reachy2Camera instance.
+
+        Args:
+            config: The configuration settings for the camera.
+        """
+        super().__init__(config)
+
+        self.config = config
+
+        self.fps = config.fps
+        self.color_mode = config.color_mode
+
+        self.cam_manager: CameraManager | None = None
+
+        self.thread: Thread | None = None
+        self.stop_event: Event | None = None
+        self.frame_lock: Lock = Lock()
+        self.latest_frame: np.ndarray | None = None
+        self.new_frame_event: Event = Event()
+
+    def __str__(self) -> str:
+        return f"{self.__class__.__name__}({self.config.name}, {self.config.image_type})"
+
+    @property
+    def is_connected(self) -> bool:
+        """Checks if the camera is currently connected and opened."""
+        if self.config.name == "teleop":
+            return self.cam_manager._grpc_connected and self.cam_manager.teleop if self.cam_manager else False
+        elif self.config.name == "depth":
+            return self.cam_manager._grpc_connected and self.cam_manager.depth if self.cam_manager else False
+        else:
+            raise ValueError(f"Invalid camera name '{self.config.name}'. Expected 'teleop' or 'depth'.")
+
+    def connect(self, warmup: bool = True):
+        """
+        Connects to the Reachy2 CameraManager as specified in the configuration.
+        """
+        self.cam_manager = CameraManager(host=self.config.ip_address, port=self.config.port)
+        self.cam_manager.initialize_cameras()
+
+        logger.info(f"{self} connected.")
+
+    @staticmethod
+    def find_cameras(ip_address: str = "localhost", port: int = 50065) -> list[dict[str, Any]]:
+        """
+        Detects available Reachy 2 cameras.
+
+        Returns:
+            List[Dict[str, Any]]: A list of dictionaries,
+            where each dictionary contains 'name', 'stereo',
+            and the default profile properties (width, height, fps).
+        """
+        initialized_cameras = []
+        camera_manager = CameraManager(host=ip_address, port=port)
+
+        for camera in [camera_manager.teleop, camera_manager.depth]:
+            if camera is None:
+                continue
+
+            height, width, _, _, _, _, _ = camera.get_parameters()
+
+            camera_info = {
+                "name": camera._cam_info.name,
+                "stereo": camera._cam_info.stereo,
+                "default_profile": {
+                    "width": width,
+                    "height": height,
+                    "fps": 30,
+                },
+            }
+            initialized_cameras.append(camera_info)
+
+        camera_manager.disconnect()
+        return initialized_cameras
+
+    def read(self, color_mode: ColorMode | None = None) -> np.ndarray:
+        """
+        Reads a single frame synchronously from the camera.
+
+        This is a blocking call.
+
+        Args:
+            color_mode (Optional[ColorMode]): If specified, overrides the default
+                color mode (`self.color_mode`) for this read operation (e.g.,
+                request RGB even if default is BGR).
+
+        Returns:
+            np.ndarray: The captured frame as a NumPy array in the format
+                       (height, width, channels), using the specified or default
+                       color mode and applying any configured rotation.
+        """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+
+        start_time = time.perf_counter()
+
+        frame = None
+
+        if self.cam_manager is None:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+        else:
+            if self.config.name == "teleop" and hasattr(self.cam_manager, "teleop"):
+                if self.config.image_type == "left":
+                    frame = self.cam_manager.teleop.get_frame(CameraView.LEFT, size=(640, 480))[0]
+                elif self.config.image_type == "right":
+                    frame = self.cam_manager.teleop.get_frame(CameraView.RIGHT, size=(640, 480))[0]
+            elif self.config.name == "depth" and hasattr(self.cam_manager, "depth"):
+                if self.config.image_type == "depth":
+                    frame = self.cam_manager.depth.get_depth_frame()[0]
+                elif self.config.image_type == "rgb":
+                    frame = self.cam_manager.depth.get_frame(size=(640, 480))[0]
+
+            if frame is None:
+                return np.empty((0, 0, 3), dtype=np.uint8)
+
+            if self.config.color_mode == "rgb":
+                frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+
+        read_duration_ms = (time.perf_counter() - start_time) * 1e3
+        logger.debug(f"{self} read took: {read_duration_ms:.1f}ms")
+
+        return frame
+
+    def _read_loop(self):
+        """
+        Internal loop run by the background thread for asynchronous reading.
+
+        On each iteration:
+        1. Reads a color frame
+        2. Stores result in latest_frame (thread-safe)
+        3. Sets new_frame_event to notify listeners
+
+        Stops on DeviceNotConnectedError, logs other errors and continues.
+        """
+        while not self.stop_event.is_set():
+            try:
+                color_image = self.read()
+
+                with self.frame_lock:
+                    self.latest_frame = color_image
+                self.new_frame_event.set()
+
+            except DeviceNotConnectedError:
+                break
+            except Exception as e:
+                logger.warning(f"Error reading frame in background thread for {self}: {e}")
+
+    def _start_read_thread(self) -> None:
+        """Starts or restarts the background read thread if it's not running."""
+        if self.thread is not None and self.thread.is_alive():
+            self.thread.join(timeout=0.1)
+        if self.stop_event is not None:
+            self.stop_event.set()
+
+        self.stop_event = Event()
+        self.thread = Thread(target=self._read_loop, args=(), name=f"{self}_read_loop")
+        self.thread.daemon = True
+        self.thread.start()
+
+    def _stop_read_thread(self) -> None:
+        """Signals the background read thread to stop and waits for it to join."""
+        if self.stop_event is not None:
+            self.stop_event.set()
+
+        if self.thread is not None and self.thread.is_alive():
+            self.thread.join(timeout=2.0)
+
+        self.thread = None
+        self.stop_event = None
+
+    def async_read(self, timeout_ms: float = 200) -> np.ndarray:
+        """
+        Reads the latest available frame asynchronously.
+
+        This method retrieves the most recent frame captured by the background
+        read thread. It does not block waiting for the camera hardware directly,
+        but may wait up to timeout_ms for the background thread to provide a frame.
+
+        Args:
+            timeout_ms (float): Maximum time in milliseconds to wait for a frame
+                to become available. Defaults to 200ms (0.2 seconds).
+
+        Returns:
+            np.ndarray: The latest captured frame as a NumPy array in the format
+                       (height, width, channels), processed according to configuration.
+
+        Raises:
+            DeviceNotConnectedError: If the camera is not connected.
+            TimeoutError: If no frame becomes available within the specified timeout.
+            RuntimeError: If an unexpected error occurs.
+        """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+
+        if self.thread is None or not self.thread.is_alive():
+            self._start_read_thread()
+
+        if not self.new_frame_event.wait(timeout=timeout_ms / 1000.0):
+            thread_alive = self.thread is not None and self.thread.is_alive()
+            raise TimeoutError(
+                f"Timed out waiting for frame from camera {self} after {timeout_ms} ms. "
+                f"Read thread alive: {thread_alive}."
+            )
+
+        with self.frame_lock:
+            frame = self.latest_frame
+            self.new_frame_event.clear()
+
+        if frame is None:
+            raise RuntimeError(f"Internal error: Event set but no frame available for {self}.")
+
+        return frame
+
+    def disconnect(self):
+        """
+        Stops the background read thread (if running).
+
+        Raises:
+            DeviceNotConnectedError: If the camera is already disconnected.
+        """
+        if not self.is_connected and self.thread is None:
+            raise DeviceNotConnectedError(f"{self} not connected.")
+
+        if self.thread is not None:
+            self._stop_read_thread()
+
+        if self.cam_manager is not None:
+            self.cam_manager.disconnect()
+
+        logger.info(f"{self} disconnected.")
@@ -51,7 +51,7 @@ class RealSenseCamera(Camera):

    Use the provided utility script to find available camera indices and default profiles:
    ```bash
-    python -m lerobot.find_cameras realsense
+    lerobot-find-cameras realsense
    ```

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

        self._configure_capture_settings()
@@ -37,8 +37,14 @@ def make_cameras_from_configs(camera_configs: dict[str, CameraConfig]) -> dict[s
            from .realsense.camera_realsense import RealSenseCamera

            cameras[key] = RealSenseCamera(cfg)
+
+        elif cfg.type == "reachy2_camera":
+            from .reachy2_camera.reachy2_camera import Reachy2Camera
+
+            cameras[key] = Reachy2Camera(cfg)
+
        else:
-            raise ValueError(f"The motor type '{cfg.type}' is not valid.")
+            raise ValueError(f"The camera type '{cfg.type}' is not valid.")

    return cameras

@@ -37,6 +37,7 @@ class DatasetConfig:
    revision: str | None = None
    use_imagenet_stats: bool = True
    video_backend: str = field(default_factory=get_safe_default_codec)
+    streaming: bool = False


@dataclass
@@ -27,6 +27,7 @@ from huggingface_hub.constants import CONFIG_NAME
 from huggingface_hub.errors import HfHubHTTPError

 from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
+from lerobot.constants import ACTION, OBS_STATE
 from lerobot.optim.optimizers import OptimizerConfig
 from lerobot.optim.schedulers import LRSchedulerConfig
 from lerobot.utils.hub import HubMixin
@@ -119,8 +120,8 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):

    @property
    def robot_state_feature(self) -> PolicyFeature | None:
-        for _, ft in self.input_features.items():
-            if ft.type is FeatureType.STATE:
+        for ft_name, ft in self.input_features.items():
+            if ft.type is FeatureType.STATE and ft_name == OBS_STATE:
                return ft
        return None

@@ -137,8 +138,8 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):

    @property
    def action_feature(self) -> PolicyFeature | None:
-        for _, ft in self.output_features.items():
-            if ft.type is FeatureType.ACTION:
+        for ft_name, ft in self.output_features.items():
+            if ft.type is FeatureType.ACTION and ft_name == ACTION:
                return ft
        return None

@@ -52,3 +52,8 @@ 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
@@ -0,0 +1,502 @@
+#!/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)
@@ -14,33 +14,13 @@

 import packaging.version

-V2_MESSAGE = """
+V30_MESSAGE = """
 The dataset you requested ({repo_id}) is in {version} format.

-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:
+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:
 ```
-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}
+python -m lerobot.datasets.v30.convert_dataset_v21_to_v30 --repo-id={repo_id}
 ```

 If you encounter a problem, contact LeRobot maintainers on [Discord](https://discord.com/invite/s3KuuzsPFb)
@@ -58,7 +38,12 @@ class CompatibilityError(Exception): ...

 class BackwardCompatibilityError(CompatibilityError):
    def __init__(self, repo_id: str, version: packaging.version.Version):
-        message = V2_MESSAGE.format(repo_id=repo_id, 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)."
+            )
        super().__init__(message)


@@ -25,6 +25,7 @@ from lerobot.datasets.lerobot_dataset import (
    LeRobotDatasetMetadata,
    MultiLeRobotDataset,
 )
+from lerobot.datasets.streaming_dataset import StreamingLeRobotDataset
 from lerobot.datasets.transforms import ImageTransforms

 IMAGENET_STATS = {
@@ -87,15 +88,26 @@ 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)
-        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,
-        )
+        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,
+            )
    else:
        raise NotImplementedError("The MultiLeRobotDataset isn't supported for now.")
        dataset = MultiLeRobotDataset(
@@ -337,13 +337,11 @@ def compute_sampler_weights(
    if len(offline_dataset) > 0:
        offline_data_mask_indices = []
        for start_index, end_index in zip(
-            offline_dataset.episode_data_index["from"],
-            offline_dataset.episode_data_index["to"],
+            offline_dataset.meta.episodes["dataset_from_index"],
+            offline_dataset.meta.episodes["dataset_to_index"],
            strict=True,
        ):
-            offline_data_mask_indices.extend(
-                range(start_index.item(), end_index.item() - offline_drop_n_last_frames)
-            )
+            offline_data_mask_indices.extend(range(start_index, end_index - 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(
@@ -21,7 +21,8 @@ import torch
 class EpisodeAwareSampler:
    def __init__(
        self,
-        episode_data_index: dict,
+        dataset_from_indices: list[int],
+        dataset_to_indices: list[int],
        episode_indices_to_use: list | None = None,
        drop_n_first_frames: int = 0,
        drop_n_last_frames: int = 0,
@@ -30,7 +31,8 @@ class EpisodeAwareSampler:
        """Sampler that optionally incorporates episode boundary information.

        Args:
-            episode_data_index: Dictionary with keys 'from' and 'to' containing the start and end indices of each episode.
+            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_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.
@@ -39,12 +41,10 @@ class EpisodeAwareSampler:
        """
        indices = []
        for episode_idx, (start_index, end_index) in enumerate(
-            zip(episode_data_index["from"], episode_data_index["to"], strict=True)
+            zip(dataset_from_indices, dataset_to_indices, strict=True)
        ):
            if episode_indices_to_use is None or episode_idx in episode_indices_to_use:
-                indices.extend(
-                    range(start_index.item() + drop_n_first_frames, end_index.item() - drop_n_last_frames)
-                )
+                indices.extend(range(start_index + drop_n_first_frames, end_index - drop_n_last_frames))

        self.indices = indices
        self.shuffle = shuffle
@@ -0,0 +1,535 @@
+#!/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)}"
+            )
@@ -17,43 +17,57 @@ import contextlib
 import importlib.resources
 import json
 import logging
-from collections.abc import Iterator
-from itertools import accumulate
+from collections import deque
+from collections.abc import Iterable, Iterator
 from pathlib import Path
 from pprint import pformat
-from types import SimpleNamespace
-from typing import Any
+from typing import Any, Deque, Generic, TypeVar

 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 DictLike, FeatureType, PolicyFeature
+from lerobot.configs.types import FeatureType, PolicyFeature
 from lerobot.datasets.backward_compatibility import (
-    V21_MESSAGE,
+    FUTURE_MESSAGE,
    BackwardCompatibilityError,
    ForwardCompatibilityError,
 )
 from lerobot.utils.utils import is_valid_numpy_dtype_string

-DEFAULT_CHUNK_SIZE = 1000  # Max number of episodes per chunk
+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

 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"

-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"
+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"

 DATASET_CARD_TEMPLATE = """
 ---
@@ -73,6 +87,67 @@ 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.
@@ -82,6 +157,7 @@ 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():
@@ -106,23 +182,13 @@ 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)):
@@ -152,24 +218,7 @@ def write_json(data: dict, fpath: Path) -> None:
        json.dump(data, f, indent=4, ensure_ascii=False)


-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):
+def write_info(info: dict, local_dir: Path) -> None:
    write_json(info, local_dir / INFO_PATH)


@@ -180,65 +229,68 @@ def load_info(local_dir: Path) -> dict:
    return info


-def write_stats(stats: dict, local_dir: Path):
+def write_stats(stats: dict, local_dir: Path) -> None:
    serialized_stats = serialize_dict(stats)
    write_json(serialized_stats, local_dir / STATS_PATH)


-def cast_stats_to_numpy(stats) -> dict[str, dict[str, np.ndarray]]:
+def cast_stats_to_numpy(stats: dict) -> 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]]:
+def load_stats(local_dir: Path) -> dict[str, dict[str, np.ndarray]] | None:
    if not (local_dir / STATS_PATH).exists():
        return None
    stats = load_json(local_dir / STATS_PATH)
    return cast_stats_to_numpy(stats)


-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 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 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 load_tasks(local_dir: Path) -> pandas.DataFrame:
+    tasks = pd.read_parquet(local_dir / DEFAULT_TASKS_PATH)
+    return tasks


-def write_episode(episode: dict, local_dir: Path):
-    append_jsonlines(episode, local_dir / EPISODES_PATH)
+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 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 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 backward_compatible_episodes_stats(
    stats: dict[str, dict[str, np.ndarray]], episodes: list[int]
-) -> dict[str, dict[str, np.ndarray]]:
+) -> dict[int, dict[str, dict[str, np.ndarray]]]:
    return dict.fromkeys(episodes, stats)


@@ -254,7 +306,7 @@ def load_image_as_numpy(
    return img_array


-def hf_transform_to_torch(items_dict: dict[torch.Tensor | None]):
+def hf_transform_to_torch(items_dict: dict[str, list[Any]]) -> dict[str, list[torch.Tensor | str]]:
    """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
@@ -299,7 +351,7 @@ 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(V21_MESSAGE.format(repo_id=repo_id, version=v_check))
+        logging.warning(FUTURE_MESSAGE.format(repo_id=repo_id, version=v_check))


 def get_repo_versions(repo_id: str) -> list[packaging.version.Version]:
@@ -476,6 +528,9 @@ def create_empty_dataset_info(
    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,
@@ -483,104 +538,17 @@ def create_empty_dataset_info(
        "total_episodes": 0,
        "total_frames": 0,
        "total_tasks": 0,
-        "total_videos": 0,
-        "total_chunks": 0,
-        "chunks_size": DEFAULT_CHUNK_SIZE,
+        "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,
        "fps": fps,
        "splits": {},
-        "data_path": DEFAULT_PARQUET_PATH,
+        "data_path": DEFAULT_DATA_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:
@@ -619,7 +587,7 @@ def get_delta_indices(delta_timestamps: dict[str, list[float]], fps: int) -> dic
    return delta_indices


-def cycle(iterable):
+def cycle(iterable: Any) -> Iterator[Any]:
    """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.
@@ -632,7 +600,7 @@ def cycle(iterable):
            iterator = iter(iterable)


-def create_branch(repo_id, *, branch: str, repo_type: str | None = None) -> None:
+def create_branch(repo_id: str, *, 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.
    """
@@ -685,76 +653,28 @@ def create_lerobot_dataset_card(
    )


-class IterableNamespace(SimpleNamespace):
-    """
-    A namespace object that supports both dictionary-like iteration and dot notation access.
-    Automatically converts nested dictionaries into IterableNamespaces.
-
-    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):
+def validate_frame(frame: dict, features: dict) -> None:
    expected_features = set(features) - set(DEFAULT_FEATURES)
    actual_features = set(frame)

-    error_message = validate_features_presence(actual_features, expected_features)
+    # 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")

-    common_features = actual_features & expected_features
-    for name in common_features - {"task"}:
+    # 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:
        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]):
+def validate_features_presence(actual_features: set[str], expected_features: set[str]) -> str:
    error_message = ""
    missing_features = expected_features - actual_features
    extra_features = actual_features - expected_features
@@ -769,7 +689,9 @@ 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):
+def validate_feature_dtype_and_shape(
+    name: str, feature: dict, value: np.ndarray | PILImage.Image | str
+) -> str:
    expected_dtype = feature["dtype"]
    expected_shape = feature["shape"]
    if is_valid_numpy_dtype_string(expected_dtype):
@@ -784,7 +706,7 @@ def validate_feature_dtype_and_shape(name: str, feature: dict, value: np.ndarray

 def validate_feature_numpy_array(
    name: str, expected_dtype: str, expected_shape: list[int], value: np.ndarray
-):
+) -> str:
    error_message = ""
    if isinstance(value, np.ndarray):
        actual_dtype = value.dtype
@@ -801,7 +723,9 @@ 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):
+def validate_feature_image_or_video(
+    name: str, expected_shape: list[str], value: np.ndarray | PILImage.Image
+) -> str:
    # 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):
@@ -817,13 +741,13 @@ def validate_feature_image_or_video(name: str, expected_shape: list[str], value:
    return error_message


-def validate_feature_string(name: str, value: str):
+def validate_feature_string(name: str, value: str) -> 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):
+def validate_episode_buffer(episode_buffer: dict, total_episodes: int, features: dict) -> None:
    if "size" not in episode_buffer:
        raise ValueError("size key not found in episode_buffer")

@@ -847,3 +771,238 @@ 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)
@@ -1,884 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-"""
-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()
@@ -1,687 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-"""
-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()
@@ -1,87 +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.
-
-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()
@@ -1,54 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-"""
-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()
@@ -1,114 +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.
-
-"""
-This script will help you convert any LeRobot dataset already pushed to the hub from codebase version 2.0 to
-2.1. 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 "v2.1".
-
-Usage:
-
-```bash
-python -m lerobot.datasets.v21.convert_dataset_v20_to_v21 \
-    --repo-id=aliberts/koch_tutorial
-```
-
-"""
-
-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(
-    repo_id: str,
-    branch: str | None = None,
-    num_workers: int = 4,
-):
-    with SuppressWarnings():
-        dataset = LeRobotDataset(repo_id, revision=V20, force_cache_sync=True)
-
-    if (dataset.root / EPISODES_STATS_PATH).is_file():
-        (dataset.root / EPISODES_STATS_PATH).unlink()
-
-    convert_stats(dataset, num_workers=num_workers)
-    ref_stats = load_stats(dataset.root)
-    check_aggregate_stats(dataset, ref_stats)
-
-    dataset.meta.info["codebase_version"] = CODEBASE_VERSION
-    write_info(dataset.meta.info, dataset.root)
-
-    dataset.push_to_hub(branch=branch, tag_version=False, allow_patterns="meta/")
-
-    # delete old stats.json file
-    if (dataset.root / STATS_PATH).is_file:
-        (dataset.root / STATS_PATH).unlink()
-
-    hub_api = HfApi()
-    if hub_api.file_exists(
-        repo_id=dataset.repo_id, filename=STATS_PATH, revision=branch, repo_type="dataset"
-    ):
-        hub_api.delete_file(
-            path_in_repo=STATS_PATH, repo_id=dataset.repo_id, revision=branch, repo_type="dataset"
-        )
-
-    hub_api.create_tag(repo_id, tag=CODEBASE_VERSION, revision=branch, repo_type="dataset")
-
-
-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(
-        "--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))
@@ -1,99 +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 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
-                )
@@ -0,0 +1,500 @@
+#!/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))
@@ -17,12 +17,15 @@ 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
@@ -168,15 +171,68 @@ 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,
@@ -185,27 +241,24 @@ 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

-    if importlib.util.find_spec("torchcodec"):
-        from torchcodec.decoders import VideoDecoder
-    else:
-        raise ImportError("torchcodec is required but not available.")
+    # Use cached decoder instead of creating new one each time
+    decoder = decoder_cache.get_decoder(str(video_path))

-    # initialize video decoder
-    decoder = VideoDecoder(video_path, device=device, seek_mode="approximate")
-    loaded_frames = []
    loaded_ts = []
+    loaded_frames = []
+
    # 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=False):
+    for frame, pts in zip(frames_batch.data, frames_batch.pts_seconds, strict=True):
        loaded_frames.append(frame)
        loaded_ts.append(pts.item())
        if log_loaded_timestamps:
@@ -236,10 +289,14 @@ def decode_video_frames_torchcodec(
    if log_loaded_timestamps:
        logging.info(f"{closest_ts=}")

-    # convert to float32 in [0,1] range (channel first)
-    closest_frames = closest_frames.type(torch.float32) / 255
+    # 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)}"
+        )

-    assert len(timestamps) == len(closest_frames)
    return closest_frames


@@ -263,7 +320,11 @@ def encode_video_frames(
    video_path = Path(video_path)
    imgs_dir = Path(imgs_dir)

-    video_path.parent.mkdir(parents=True, exist_ok=overwrite)
+    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)

    # Encoders/pixel formats incompatibility check
    if (vcodec == "libsvtav1" or vcodec == "hevc") and pix_fmt == "yuv444p":
@@ -273,9 +334,9 @@ 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)
@@ -300,7 +361,7 @@ 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)
@@ -331,6 +392,89 @@ 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
@@ -454,6 +598,28 @@ 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.
@@ -487,7 +653,7 @@ class VideoEncodingManager:
                f"Encoding remaining {self.dataset.episodes_since_last_encoding} episodes, "
                f"from episode {start_ep} to {end_ep - 1}"
            )
-            self.dataset.batch_encode_videos(start_ep, end_ep)
+            self.dataset._batch_save_episode_video(start_ep, end_ep)

        # Clean up episode images if recording was interrupted
        if exc_type is not None:
@@ -20,7 +20,7 @@ Helper to find the camera devices available in your system.
 Example:

 ```shell
-python -m lerobot.find_cameras
+lerobot-find-cameras
 ```
 """

@@ -18,7 +18,7 @@ Helper to find the USB port associated with your MotorsBus.
 Example:

 ```shell
-python -m lerobot.find_port
+lerobot-find-port
 ```
 """

@@ -107,6 +107,8 @@ X_SERIES_ENCODINGS_TABLE = {
    "Goal_PWM": X_SERIES_CONTROL_TABLE["Goal_PWM"][1],
    "Goal_Current": X_SERIES_CONTROL_TABLE["Goal_Current"][1],
    "Goal_Velocity": X_SERIES_CONTROL_TABLE["Goal_Velocity"][1],
+    "Goal_Position": X_SERIES_CONTROL_TABLE["Goal_Position"][1],
+    "Present_Position": X_SERIES_CONTROL_TABLE["Present_Position"][1],
    "Present_PWM": X_SERIES_CONTROL_TABLE["Present_PWM"][1],
    "Present_Current": X_SERIES_CONTROL_TABLE["Present_Current"][1],
    "Present_Velocity": X_SERIES_CONTROL_TABLE["Present_Velocity"][1],
@@ -222,7 +222,7 @@ class MotorsBus(abc.ABC):
    A MotorsBus subclass instance requires a port (e.g. `FeetechMotorsBus(port="/dev/tty.usbmodem575E0031751"`)).
    To find the port, you can run our utility script:
    ```bash
-    python -m lerobot.find_port.py
+    lerobot-find-port.py
    >>> Finding all available ports for the MotorsBus.
    >>> ["/dev/tty.usbmodem575E0032081", "/dev/tty.usbmodem575E0031751"]
    >>> Remove the usb cable from your MotorsBus and press Enter when done.
@@ -446,7 +446,7 @@ class MotorsBus(abc.ABC):
        except (FileNotFoundError, OSError, serial.SerialException) as e:
            raise ConnectionError(
                f"\nCould not connect on port '{self.port}'. Make sure you are using the correct port."
-                "\nTry running `python -m lerobot.find_port`\n"
+                "\nTry running `lerobot-find-port`\n"
            ) from e

    @abc.abstractmethod
@@ -30,7 +30,7 @@ pip install -e ".[pi0]"

 Example of finetuning the pi0 pretrained model (`pi0_base` in `openpi`):
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
 --policy.path=lerobot/pi0 \
 --dataset.repo_id=danaaubakirova/koch_test
 ```
@@ -38,7 +38,7 @@ python -m lerobot.scripts.train \
 Example of finetuning the pi0 neural network with PaliGemma and expert Gemma
 pretrained with VLM default parameters before pi0 finetuning:
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
 --policy.type=pi0 \
 --dataset.repo_id=danaaubakirova/koch_test
 ```
@@ -25,14 +25,14 @@ Disclaimer: It is not expected to perform as well as the original implementation

 Example of finetuning the pi0+FAST pretrained model (`pi0_fast_base` in `openpi`):
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
 --policy.path=lerobot/pi0fast_base \
 --dataset.repo_id=danaaubakirova/koch_test
 ```

 Example of training the pi0+FAST neural network with from scratch:
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
 --policy.type=pi0fast \
 --dataset.repo_id=danaaubakirova/koch_test
 ```
@@ -28,7 +28,7 @@ pip install -e ".[smolvla]"

 Example of finetuning the smolvla pretrained model (`smolvla_base`):
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
 --policy.path=lerobot/smolvla_base \
 --dataset.repo_id=danaaubakirova/svla_so100_task1_v3 \
 --batch_size=64 \
@@ -38,7 +38,7 @@ python -m lerobot.scripts.train \
 Example of finetuning a smolVLA. SmolVLA is composed of a pretrained VLM,
 and an action expert.
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
 --policy.type=smolvla \
 --dataset.repo_id=danaaubakirova/svla_so100_task1_v3 \
 --batch_size=64 \
@@ -0,0 +1,54 @@
+#!/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 .device_processor import DeviceProcessor
+from .normalize_processor import NormalizerProcessor, UnnormalizerProcessor
+from .observation_processor import VanillaObservationProcessor
+from .pipeline import (
+    ActionProcessor,
+    DoneProcessor,
+    EnvTransition,
+    IdentityProcessor,
+    InfoProcessor,
+    ObservationProcessor,
+    ProcessorStep,
+    ProcessorStepRegistry,
+    RewardProcessor,
+    RobotProcessor,
+    TransitionKey,
+    TruncatedProcessor,
+)
+from .rename_processor import RenameProcessor
+
+__all__ = [
+    "ActionProcessor",
+    "DeviceProcessor",
+    "DoneProcessor",
+    "EnvTransition",
+    "IdentityProcessor",
+    "InfoProcessor",
+    "NormalizerProcessor",
+    "UnnormalizerProcessor",
+    "ObservationProcessor",
+    "ProcessorStep",
+    "ProcessorStepRegistry",
+    "RenameProcessor",
+    "RewardProcessor",
+    "RobotProcessor",
+    "TransitionKey",
+    "TruncatedProcessor",
+    "VanillaObservationProcessor",
+]
@@ -0,0 +1,82 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from dataclasses import dataclass
+from typing import Any
+
+import torch
+
+from lerobot.configs.types import PolicyFeature
+from lerobot.processor.pipeline import EnvTransition, TransitionKey
+from lerobot.utils.utils import get_safe_torch_device
+
+
+@dataclass
+class DeviceProcessor:
+    """Processes transitions by moving tensors to the specified device.
+
+    This processor ensures that all tensors in the transition are moved to the
+    specified device (CPU or GPU) before they are returned.
+    """
+
+    device: torch.device = "cpu"
+
+    def __post_init__(self):
+        self.device = get_safe_torch_device(self.device)
+        self.non_blocking = "cuda" in str(self.device)
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        # Create a copy of the transition
+        new_transition = transition.copy()
+
+        # Process observation tensors
+        observation = transition.get(TransitionKey.OBSERVATION)
+        if observation is not None:
+            new_observation = {
+                k: v.to(self.device, non_blocking=self.non_blocking) if isinstance(v, torch.Tensor) else v
+                for k, v in observation.items()
+            }
+            new_transition[TransitionKey.OBSERVATION] = new_observation
+
+        # Process action tensor
+        action = transition.get(TransitionKey.ACTION)
+        if action is not None and isinstance(action, torch.Tensor):
+            new_transition[TransitionKey.ACTION] = action.to(self.device, non_blocking=self.non_blocking)
+
+        # Process reward tensor
+        reward = transition.get(TransitionKey.REWARD)
+        if reward is not None and isinstance(reward, torch.Tensor):
+            new_transition[TransitionKey.REWARD] = reward.to(self.device, non_blocking=self.non_blocking)
+
+        # Process done tensor
+        done = transition.get(TransitionKey.DONE)
+        if done is not None and isinstance(done, torch.Tensor):
+            new_transition[TransitionKey.DONE] = done.to(self.device, non_blocking=self.non_blocking)
+
+        # Process truncated tensor
+        truncated = transition.get(TransitionKey.TRUNCATED)
+        if truncated is not None and isinstance(truncated, torch.Tensor):
+            new_transition[TransitionKey.TRUNCATED] = truncated.to(
+                self.device, non_blocking=self.non_blocking
+            )
+
+        return new_transition
+
+    def get_config(self) -> dict[str, Any]:
+        """Return configuration for serialization."""
+        return {"device": self.device}
+
+    def feature_contract(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
+        return features
@@ -0,0 +1,331 @@
+from __future__ import annotations
+
+from collections.abc import Mapping
+from dataclasses import dataclass, field
+from typing import Any
+
+import numpy as np
+import torch
+from torch import Tensor
+
+from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.processor.pipeline import EnvTransition, ProcessorStepRegistry, TransitionKey
+
+
+def _convert_stats_to_tensors(stats: dict[str, dict[str, Any]]) -> dict[str, dict[str, Tensor]]:
+    """Convert numpy arrays and other types to torch tensors."""
+    tensor_stats: dict[str, dict[str, Tensor]] = {}
+    for key, sub in stats.items():
+        tensor_stats[key] = {}
+        for stat_name, value in sub.items():
+            if isinstance(value, np.ndarray):
+                tensor_val = torch.from_numpy(value.astype(np.float32))
+            elif isinstance(value, torch.Tensor):
+                tensor_val = value.to(dtype=torch.float32)
+            elif isinstance(value, (int, float, list, tuple)):
+                tensor_val = torch.tensor(value, dtype=torch.float32)
+            else:
+                raise TypeError(f"Unsupported type for stats['{key}']['{stat_name}']: {type(value)}")
+            tensor_stats[key][stat_name] = tensor_val
+    return tensor_stats
+
+
+@dataclass
+@ProcessorStepRegistry.register(name="normalizer_processor")
+class NormalizerProcessor:
+    """Normalizes observations and actions in a single processor step.
+
+    This processor handles normalization of both observation and action tensors
+    using either mean/std normalization or min/max scaling to a [-1, 1] range.
+
+    For each tensor key in the stats dictionary, the processor will:
+    - Use mean/std normalization if those statistics are provided: (x - mean) / std
+    - Use min/max scaling if those statistics are provided: 2 * (x - min) / (max - min) - 1
+
+    The processor can be configured to normalize only specific keys by setting
+    the normalize_keys parameter.
+    """
+
+    # Features and normalisation map are mandatory to match the design of normalize.py
+    features: dict[str, PolicyFeature]
+    norm_map: dict[FeatureType, NormalizationMode]
+
+    # Pre-computed statistics coming from dataset.meta.stats for instance.
+    stats: dict[str, dict[str, Any]] | None = None
+
+    # Explicit subset of keys to normalise. If ``None`` every key (except
+    # "action") found in ``stats`` will be normalised. Using a ``set`` makes
+    # membership checks O(1).
+    normalize_keys: set[str] | None = None
+
+    eps: float = 1e-8
+
+    _tensor_stats: dict[str, dict[str, Tensor]] = field(default_factory=dict, init=False, repr=False)
+
+    @classmethod
+    def from_lerobot_dataset(
+        cls,
+        dataset: LeRobotDataset,
+        features: dict[str, PolicyFeature],
+        norm_map: dict[FeatureType, NormalizationMode],
+        *,
+        normalize_keys: set[str] | None = None,
+        eps: float = 1e-8,
+    ) -> NormalizerProcessor:
+        """Factory helper that pulls statistics from a :class:`LeRobotDataset`.
+
+        The features and norm_map parameters are mandatory to match the design
+        pattern used in normalize.py.
+        """
+
+        return cls(
+            features=features,
+            norm_map=norm_map,
+            stats=dataset.meta.stats,
+            normalize_keys=normalize_keys,
+            eps=eps,
+        )
+
+    def __post_init__(self):
+        # Handle deserialization from JSON config
+        if self.features and isinstance(list(self.features.values())[0], dict):
+            # Features came from JSON - need to reconstruct PolicyFeature objects
+            reconstructed_features = {}
+            for key, ft_dict in self.features.items():
+                reconstructed_features[key] = PolicyFeature(
+                    type=FeatureType(ft_dict["type"]), shape=tuple(ft_dict["shape"])
+                )
+            self.features = reconstructed_features
+
+        if self.norm_map and isinstance(list(self.norm_map.keys())[0], str):
+            # norm_map came from JSON - need to reconstruct enum keys and values
+            reconstructed_norm_map = {}
+            for ft_type_str, norm_mode_str in self.norm_map.items():
+                reconstructed_norm_map[FeatureType(ft_type_str)] = NormalizationMode(norm_mode_str)
+            self.norm_map = reconstructed_norm_map
+
+        # Convert statistics once so we avoid repeated numpy→Tensor conversions
+        # during runtime.
+        self.stats = self.stats or {}
+        self._tensor_stats = _convert_stats_to_tensors(self.stats)
+
+        # Ensure *normalize_keys* is a set for fast look-ups and compare by
+        # value later when returning the configuration.
+        if self.normalize_keys is not None and not isinstance(self.normalize_keys, set):
+            self.normalize_keys = set(self.normalize_keys)
+
+    def _normalize_obs(self, observation):
+        if observation is None:
+            return None
+
+        # Decide which keys should be normalised for this call.
+        if self.normalize_keys is not None:
+            keys_to_norm = self.normalize_keys
+        else:
+            # Use feature map to skip action keys.
+            keys_to_norm = {k for k, ft in self.features.items() if ft.type is not FeatureType.ACTION}
+
+        processed = dict(observation)
+        for key in keys_to_norm:
+            if key not in processed or key not in self._tensor_stats:
+                continue
+
+            orig_val = processed[key]
+            tensor = (
+                orig_val.to(dtype=torch.float32)
+                if isinstance(orig_val, torch.Tensor)
+                else torch.as_tensor(orig_val, dtype=torch.float32)
+            )
+            stats = {k: v.to(tensor.device) for k, v in self._tensor_stats[key].items()}
+
+            if "mean" in stats and "std" in stats:
+                mean, std = stats["mean"], stats["std"]
+                processed[key] = (tensor - mean) / (std + self.eps)
+            elif "min" in stats and "max" in stats:
+                min_val, max_val = stats["min"], stats["max"]
+                processed[key] = 2 * (tensor - min_val) / (max_val - min_val + self.eps) - 1
+        return processed
+
+    def _normalize_action(self, action):
+        if action is None or "action" not in self._tensor_stats:
+            return action
+
+        tensor = (
+            action.to(dtype=torch.float32)
+            if isinstance(action, torch.Tensor)
+            else torch.as_tensor(action, dtype=torch.float32)
+        )
+        stats = {k: v.to(tensor.device) for k, v in self._tensor_stats["action"].items()}
+        if "mean" in stats and "std" in stats:
+            mean, std = stats["mean"], stats["std"]
+            return (tensor - mean) / (std + self.eps)
+        if "min" in stats and "max" in stats:
+            min_val, max_val = stats["min"], stats["max"]
+            return 2 * (tensor - min_val) / (max_val - min_val + self.eps) - 1
+        raise ValueError("Action stats must contain either ('mean','std') or ('min','max')")
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        observation = self._normalize_obs(transition.get(TransitionKey.OBSERVATION))
+        action = self._normalize_action(transition.get(TransitionKey.ACTION))
+
+        # Create a new transition with normalized values
+        new_transition = transition.copy()
+        new_transition[TransitionKey.OBSERVATION] = observation
+        new_transition[TransitionKey.ACTION] = action
+        return new_transition
+
+    def get_config(self) -> dict[str, Any]:
+        config = {
+            "eps": self.eps,
+            "features": {
+                key: {"type": ft.type.value, "shape": ft.shape} for key, ft in self.features.items()
+            },
+            "norm_map": {ft_type.value: norm_mode.value for ft_type, norm_mode in self.norm_map.items()},
+        }
+        if self.normalize_keys is not None:
+            # Serialise as a list for YAML / JSON friendliness
+            config["normalize_keys"] = sorted(self.normalize_keys)
+        return config
+
+    def state_dict(self) -> dict[str, Tensor]:
+        flat = {}
+        for key, sub in self._tensor_stats.items():
+            for stat_name, tensor in sub.items():
+                flat[f"{key}.{stat_name}"] = tensor
+        return flat
+
+    def load_state_dict(self, state: Mapping[str, Tensor]) -> None:
+        self._tensor_stats.clear()
+        for flat_key, tensor in state.items():
+            key, stat_name = flat_key.rsplit(".", 1)
+            self._tensor_stats.setdefault(key, {})[stat_name] = tensor
+
+    def reset(self):
+        pass
+
+    def feature_contract(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
+        return features
+
+
+@dataclass
+@ProcessorStepRegistry.register(name="unnormalizer_processor")
+class UnnormalizerProcessor:
+    """Inverse normalisation for observations and actions.
+
+    Exactly mirrors :class:`NormalizerProcessor` but applies the inverse
+    transform.
+    """
+
+    features: dict[str, PolicyFeature]
+    norm_map: dict[FeatureType, NormalizationMode]
+    stats: dict[str, dict[str, Any]] | None = None
+
+    _tensor_stats: dict[str, dict[str, Tensor]] = field(default_factory=dict, init=False, repr=False)
+
+    @classmethod
+    def from_lerobot_dataset(
+        cls,
+        dataset: LeRobotDataset,
+        features: dict[str, PolicyFeature],
+        norm_map: dict[FeatureType, NormalizationMode],
+    ) -> UnnormalizerProcessor:
+        return cls(features=features, norm_map=norm_map, stats=dataset.meta.stats)
+
+    def __post_init__(self):
+        # Handle deserialization from JSON config
+        if self.features and isinstance(list(self.features.values())[0], dict):
+            # Features came from JSON - need to reconstruct PolicyFeature objects
+            reconstructed_features = {}
+            for key, ft_dict in self.features.items():
+                reconstructed_features[key] = PolicyFeature(
+                    type=FeatureType(ft_dict["type"]), shape=tuple(ft_dict["shape"])
+                )
+            self.features = reconstructed_features
+
+        if self.norm_map and isinstance(list(self.norm_map.keys())[0], str):
+            # norm_map came from JSON - need to reconstruct enum keys and values
+            reconstructed_norm_map = {}
+            for ft_type_str, norm_mode_str in self.norm_map.items():
+                reconstructed_norm_map[FeatureType(ft_type_str)] = NormalizationMode(norm_mode_str)
+            self.norm_map = reconstructed_norm_map
+
+        self.stats = self.stats or {}
+        self._tensor_stats = _convert_stats_to_tensors(self.stats)
+
+    def _unnormalize_obs(self, observation):
+        if observation is None:
+            return None
+        keys = [k for k, ft in self.features.items() if ft.type is not FeatureType.ACTION]
+        processed = dict(observation)
+        for key in keys:
+            if key not in processed or key not in self._tensor_stats:
+                continue
+            orig_val = processed[key]
+            tensor = (
+                orig_val.to(dtype=torch.float32)
+                if isinstance(orig_val, torch.Tensor)
+                else torch.as_tensor(orig_val, dtype=torch.float32)
+            )
+            stats = {k: v.to(tensor.device) for k, v in self._tensor_stats[key].items()}
+            if "mean" in stats and "std" in stats:
+                mean, std = stats["mean"], stats["std"]
+                processed[key] = tensor * std + mean
+            elif "min" in stats and "max" in stats:
+                min_val, max_val = stats["min"], stats["max"]
+                processed[key] = (tensor + 1) / 2 * (max_val - min_val) + min_val
+        return processed
+
+    def _unnormalize_action(self, action):
+        if action is None or "action" not in self._tensor_stats:
+            return action
+        tensor = (
+            action.to(dtype=torch.float32)
+            if isinstance(action, torch.Tensor)
+            else torch.as_tensor(action, dtype=torch.float32)
+        )
+        stats = {k: v.to(tensor.device) for k, v in self._tensor_stats["action"].items()}
+        if "mean" in stats and "std" in stats:
+            mean, std = stats["mean"], stats["std"]
+            return tensor * std + mean
+        if "min" in stats and "max" in stats:
+            min_val, max_val = stats["min"], stats["max"]
+            return (tensor + 1) / 2 * (max_val - min_val) + min_val
+        raise ValueError("Action stats must contain either ('mean','std') or ('min','max')")
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        observation = self._unnormalize_obs(transition.get(TransitionKey.OBSERVATION))
+        action = self._unnormalize_action(transition.get(TransitionKey.ACTION))
+
+        # Create a new transition with unnormalized values
+        new_transition = transition.copy()
+        new_transition[TransitionKey.OBSERVATION] = observation
+        new_transition[TransitionKey.ACTION] = action
+        return new_transition
+
+    def get_config(self) -> dict[str, Any]:
+        return {
+            "features": {
+                key: {"type": ft.type.value, "shape": ft.shape} for key, ft in self.features.items()
+            },
+            "norm_map": {ft_type.value: norm_mode.value for ft_type, norm_mode in self.norm_map.items()},
+        }
+
+    def state_dict(self) -> dict[str, Tensor]:
+        flat = {}
+        for key, sub in self._tensor_stats.items():
+            for stat_name, tensor in sub.items():
+                flat[f"{key}.{stat_name}"] = tensor
+        return flat
+
+    def load_state_dict(self, state: Mapping[str, Tensor]) -> None:
+        self._tensor_stats.clear()
+        for flat_key, tensor in state.items():
+            key, stat_name = flat_key.rsplit(".", 1)
+            self._tensor_stats.setdefault(key, {})[stat_name] = tensor
+
+    def reset(self):
+        pass
+
+    def feature_contract(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
+        return features
@@ -0,0 +1,157 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from dataclasses import dataclass
+
+import einops
+import numpy as np
+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 ObservationProcessor, ProcessorStepRegistry
+
+
+@dataclass
+@ProcessorStepRegistry.register(name="observation_processor")
+class VanillaObservationProcessor(ObservationProcessor):
+    """
+    Processes environment observations into the LeRobot format by handling both images and states.
+
+    Image processing:
+        - Converts channel-last (H, W, C) images to channel-first (C, H, W)
+        - Normalizes uint8 images ([0, 255]) to float32 ([0, 1])
+        - Adds a batch dimension if missing
+        - Supports single images and image dictionaries
+
+    State processing:
+        - Maps 'environment_state' to observation.environment_state
+        - Maps 'agent_pos' to observation.state
+        - Converts numpy arrays to tensors
+        - Adds a batch dimension if missing
+    """
+
+    def _process_single_image(self, img: np.ndarray) -> Tensor:
+        """Process a single image array."""
+        # Convert to tensor
+        img_tensor = torch.from_numpy(img)
+
+        # Add batch dimension if needed
+        if img_tensor.ndim == 3:
+            img_tensor = img_tensor.unsqueeze(0)
+
+        # Validate image format
+        _, h, w, c = img_tensor.shape
+        if not (c < h and c < w):
+            raise ValueError(f"Expected channel-last images, but got shape {img_tensor.shape}")
+
+        if img_tensor.dtype != torch.uint8:
+            raise ValueError(f"Expected torch.uint8 images, but got {img_tensor.dtype}")
+
+        # Convert to channel-first format
+        img_tensor = einops.rearrange(img_tensor, "b h w c -> b c h w").contiguous()
+
+        # Convert to float32 and normalize to [0, 1]
+        img_tensor = img_tensor.type(torch.float32) / 255.0
+
+        return img_tensor
+
+    def _process_observation(self, observation):
+        """
+        Processes both image and state observations.
+        """
+
+        processed_obs = observation.copy()
+
+        if "pixels" in processed_obs:
+            pixels = processed_obs.pop("pixels")
+
+            if isinstance(pixels, dict):
+                imgs = {f"{OBS_IMAGES}.{key}": img for key, img in pixels.items()}
+            else:
+                imgs = {OBS_IMAGE: pixels}
+
+            for imgkey, img in imgs.items():
+                processed_obs[imgkey] = self._process_single_image(img)
+
+        if "environment_state" in processed_obs:
+            env_state_np = processed_obs.pop("environment_state")
+            env_state = torch.from_numpy(env_state_np).float()
+            if env_state.dim() == 1:
+                env_state = env_state.unsqueeze(0)
+            processed_obs[OBS_ENV_STATE] = env_state
+
+        if "agent_pos" in processed_obs:
+            agent_pos_np = processed_obs.pop("agent_pos")
+            agent_pos = torch.from_numpy(agent_pos_np).float()
+            if agent_pos.dim() == 1:
+                agent_pos = agent_pos.unsqueeze(0)
+            processed_obs[OBS_STATE] = agent_pos
+
+        return processed_obs
+
+    def observation(self, observation):
+        return self._process_observation(observation)
+
+    def feature_contract(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
+        """Transforms feature keys to a standardized contract.
+
+        This method handles several renaming patterns:
+        - Exact matches (e.g., 'pixels' -> 'OBS_IMAGE').
+        - Prefixed exact matches (e.g., 'observation.pixels' -> 'OBS_IMAGE').
+        - Prefix matches (e.g., 'pixels.cam1' -> 'OBS_IMAGES.cam1').
+        - Prefixed prefix matches (e.g., 'observation.pixels.cam1' -> 'OBS_IMAGES.cam1').
+        - environment_state -> OBS_ENV_STATE,
+        - agent_pos -> OBS_STATE,
+        - observation.environment_state -> OBS_ENV_STATE,
+        - observation.agent_pos -> OBS_STATE
+        """
+        exact_pairs = {
+            "pixels": OBS_IMAGE,
+            "environment_state": OBS_ENV_STATE,
+            "agent_pos": OBS_STATE,
+        }
+
+        prefix_pairs = {
+            "pixels.": f"{OBS_IMAGES}.",
+        }
+
+        for key in list(features.keys()):
+            matched_prefix = False
+            for old_prefix, new_prefix in prefix_pairs.items():
+                prefixed_old = f"observation.{old_prefix}"
+                if key.startswith(prefixed_old):
+                    suffix = key[len(prefixed_old) :]
+                    features[f"{new_prefix}{suffix}"] = features.pop(key)
+                    matched_prefix = True
+                    break
+
+                if key.startswith(old_prefix):
+                    suffix = key[len(old_prefix) :]
+                    features[f"{new_prefix}{suffix}"] = features.pop(key)
+                    matched_prefix = True
+                    break
+
+            if matched_prefix:
+                continue
+
+            for old, new in exact_pairs.items():
+                if key == old or key == f"observation.{old}":
+                    if key in features:
+                        features[new] = features.pop(key)
+                        break
+
+        return features
@@ -0,0 +1,51 @@
+#!/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 typing import Any
+
+from lerobot.configs.types import PolicyFeature
+from lerobot.processor.pipeline import (
+    ObservationProcessor,
+    ProcessorStepRegistry,
+)
+
+
+@dataclass
+@ProcessorStepRegistry.register(name="rename_processor")
+class RenameProcessor(ObservationProcessor):
+    """Rename processor that renames keys in the observation."""
+
+    rename_map: dict[str, str] = field(default_factory=dict)
+
+    def observation(self, observation):
+        processed_obs = {}
+        for key, value in observation.items():
+            if key in self.rename_map:
+                processed_obs[self.rename_map[key]] = value
+            else:
+                processed_obs[key] = value
+
+        return processed_obs
+
+    def get_config(self) -> dict[str, Any]:
+        return {"rename_map": self.rename_map}
+
+    def feature_contract(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
+        """Transforms:
+        - Each key in the observation that appears in `rename_map` is renamed to its value.
+        - Keys not in `rename_map` remain unchanged.
+        """
+        return {self.rename_map.get(k, k): v for k, v in features.items()}
@@ -18,7 +18,7 @@ Records a dataset. Actions for the robot can be either generated by teleoperatio
 Example:

 ```shell
-python -m lerobot.record \
+lerobot-record \
    --robot.type=so100_follower \
    --robot.port=/dev/tty.usbmodem58760431541 \
    --robot.cameras="{laptop: {type: opencv, camera_index: 0, width: 640, height: 480}}" \
@@ -36,7 +36,7 @@ python -m lerobot.record \

 Example recording with bimanual so100:
 ```shell
-python -m lerobot.record \
+lerobot-record \
  --robot.type=bi_so100_follower \
  --robot.left_arm_port=/dev/tty.usbmodem5A460851411 \
  --robot.right_arm_port=/dev/tty.usbmodem5A460812391 \
@@ -209,7 +209,14 @@ def record_loop(
            (
                t
                for t in teleop
-                if isinstance(t, (so100_leader.SO100Leader, so101_leader.SO101Leader, koch_leader.KochLeader))
+                if isinstance(
+                    t,
+                    (
+                        so100_leader.SO100Leader,
+                        so101_leader.SO101Leader,
+                        koch_leader.KochLeader,
+                    ),
+                )
            ),
            None,
        )
@@ -272,8 +279,8 @@ def record_loop(

        if dataset is not None:
            action_frame = build_dataset_frame(dataset.features, sent_action, prefix="action")
-            frame = {**observation_frame, **action_frame}
-            dataset.add_frame(frame, task=single_task)
+            frame = {**observation_frame, **action_frame, "task": single_task}
+            dataset.add_frame(frame)

        if display_data:
            log_rerun_data(observation, action)
@@ -18,7 +18,7 @@ Replays the actions of an episode from a dataset on a robot.
 Examples:

 ```shell
-python -m lerobot.replay \
+lerobot-replay \
    --robot.type=so100_follower \
    --robot.port=/dev/tty.usbmodem58760431541 \
    --robot.id=black \
@@ -28,7 +28,7 @@ python -m lerobot.replay \

 Example replay with bimanual so100:
 ```shell
-python -m lerobot.replay \
+lerobot-replay \
  --robot.type=bi_so100_follower \
  --robot.left_arm_port=/dev/tty.usbmodem5A460851411 \
  --robot.right_arm_port=/dev/tty.usbmodem5A460812391 \
@@ -55,6 +55,7 @@ from lerobot.robots import (  # noqa: F401
    hope_jr,
    koch_follower,
    make_robot_from_config,
+    reachy2,
    so100_follower,
    so101_follower,
 )
@@ -92,11 +93,15 @@ def replay(cfg: ReplayConfig):

    robot = make_robot_from_config(cfg.robot)
    dataset = LeRobotDataset(cfg.dataset.repo_id, root=cfg.dataset.root, episodes=[cfg.dataset.episode])
-    actions = dataset.hf_dataset.select_columns("action")
+
+    # Filter dataset to only include frames from the specified episode since episodes are chunked in dataset V3.0
+    episode_frames = dataset.hf_dataset.filter(lambda x: x["episode_index"] == cfg.dataset.episode)
+    actions = episode_frames.select_columns("action")
+
    robot.connect()

    log_say("Replaying episode", cfg.play_sounds, blocking=True)
-    for idx in range(dataset.num_frames):
+    for idx in range(len(episode_frames)):
        start_episode_t = time.perf_counter()

        action_array = actions[idx]["action"]
@@ -29,10 +29,10 @@ class BiSO100FollowerConfig(RobotConfig):

    # Optional
    left_arm_disable_torque_on_disconnect: bool = True
-    left_arm_max_relative_target: int | None = None
+    left_arm_max_relative_target: float | dict[str, float] | None = None
    left_arm_use_degrees: bool = False
    right_arm_disable_torque_on_disconnect: bool = True
-    right_arm_max_relative_target: int | None = None
+    right_arm_max_relative_target: float | dict[str, float] | None = None
    right_arm_use_degrees: bool = False

    # cameras (shared between both arms)
@@ -44,8 +44,8 @@ class HopeJrArmConfig(RobotConfig):
    disable_torque_on_disconnect: bool = True

    # `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
-    # Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
-    # the number of motors in your follower arms.
-    max_relative_target: int | None = None
+    # Set this to a positive scalar to have the same value for all motors, or a dictionary that maps motor
+    # names to the max_relative_target value for that motor.
+    max_relative_target: float | dict[str, float] | None = None

    cameras: dict[str, CameraConfig] = field(default_factory=dict)
@@ -28,9 +28,9 @@ class KochFollowerConfig(RobotConfig):
    disable_torque_on_disconnect: bool = True

    # `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
-    # Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
-    # the number of motors in your follower arms.
-    max_relative_target: int | None = None
+    # Set this to a positive scalar to have the same value for all motors, or a dictionary that maps motor
+    # names to the max_relative_target value for that motor.
+    max_relative_target: float | dict[str, float] | None = None

    # cameras
    cameras: dict[str, CameraConfig] = field(default_factory=dict)
@@ -110,6 +110,7 @@ class KochFollower(Robot):
        return self.bus.is_calibrated

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

@@ -39,9 +39,9 @@ class LeKiwiConfig(RobotConfig):
    disable_torque_on_disconnect: bool = True

    # `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
-    # Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
-    # the number of motors in your follower arms.
-    max_relative_target: int | None = None
+    # Set this to a positive scalar to have the same value for all motors, or a dictionary that maps motor
+    # names to the max_relative_target value for that motor.
+    max_relative_target: float | dict[str, float] | None = None

    cameras: dict[str, CameraConfig] = field(default_factory=lekiwi_cameras_config)

@@ -0,0 +1,25 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .configuration_reachy2 import Reachy2RobotConfig
+from .robot_reachy2 import (
+    REACHY2_ANTENNAS_JOINTS,
+    REACHY2_L_ARM_JOINTS,
+    REACHY2_NECK_JOINTS,
+    REACHY2_R_ARM_JOINTS,
+    REACHY2_VEL,
+    Reachy2Robot,
+)
@@ -0,0 +1,107 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass, field
+
+from lerobot.cameras import CameraConfig
+from lerobot.cameras.configs import ColorMode
+from lerobot.cameras.reachy2_camera import Reachy2CameraConfig
+
+from ..config import RobotConfig
+
+
+@RobotConfig.register_subclass("reachy2")
+@dataclass
+class Reachy2RobotConfig(RobotConfig):
+    # `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
+    # Set this to a positive scalar to have the same value for all motors.
+    max_relative_target: float | None = None
+
+    # IP address of the Reachy 2 robot
+    ip_address: str | None = "localhost"
+
+    # If True, turn_off_smoothly() will be sent to the robot before disconnecting.
+    disable_torque_on_disconnect: bool = False
+
+    # Tag for external commands control
+    # Set to True if you use an external commands system to control the robot,
+    # such as the official teleoperation application: https://github.com/pollen-robotics/Reachy2Teleoperation
+    # If True, robot.send_action() will not send commands to the robot.
+    use_external_commands: bool = False
+
+    # Robot parts
+    # Set to False to not add the corresponding joints part to the robot list of joints.
+    # By default, all parts are set to True.
+    with_mobile_base: bool = True
+    with_l_arm: bool = True
+    with_r_arm: bool = True
+    with_neck: bool = True
+    with_antennas: bool = True
+
+    # Robot cameras
+    # Set to True if you want to use the corresponding cameras in the observations.
+    # By default, only the teleop cameras are used.
+    with_left_teleop_camera: bool = True
+    with_right_teleop_camera: bool = True
+    with_torso_camera: bool = False
+
+    cameras: dict[str, CameraConfig] = field(default_factory=dict)
+
+    def __post_init__(self) -> None:
+        # Add cameras with same ip_address as the robot
+        if self.with_left_teleop_camera:
+            self.cameras["teleop_left"] = Reachy2CameraConfig(
+                name="teleop",
+                image_type="left",
+                ip_address=self.ip_address,
+                fps=15,
+                width=640,
+                height=480,
+                color_mode=ColorMode.RGB,
+            )
+        if self.with_right_teleop_camera:
+            self.cameras["teleop_right"] = Reachy2CameraConfig(
+                name="teleop",
+                image_type="right",
+                ip_address=self.ip_address,
+                fps=15,
+                width=640,
+                height=480,
+                color_mode=ColorMode.RGB,
+            )
+        if self.with_torso_camera:
+            self.cameras["torso_rgb"] = Reachy2CameraConfig(
+                name="depth",
+                image_type="rgb",
+                ip_address=self.ip_address,
+                fps=15,
+                width=640,
+                height=480,
+                color_mode=ColorMode.RGB,
+            )
+
+        super().__post_init__()
+
+        if not (
+            self.with_mobile_base
+            or self.with_l_arm
+            or self.with_r_arm
+            or self.with_neck
+            or self.with_antennas
+        ):
+            raise ValueError(
+                "No Reachy2Robot part used.\n"
+                "At least one part of the robot must be set to True "
+                "(with_mobile_base, with_l_arm, with_r_arm, with_neck, with_antennas)"
+            )
@@ -0,0 +1,230 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import time
+from typing import Any
+
+import numpy as np
+from reachy2_sdk import ReachySDK
+
+from lerobot.cameras.utils import make_cameras_from_configs
+
+from ..robot import Robot
+from ..utils import ensure_safe_goal_position
+from .configuration_reachy2 import Reachy2RobotConfig
+
+# {lerobot_keys: reachy2_sdk_keys}
+REACHY2_NECK_JOINTS = {
+    "neck_yaw.pos": "head.neck.yaw",
+    "neck_pitch.pos": "head.neck.pitch",
+    "neck_roll.pos": "head.neck.roll",
+}
+
+REACHY2_ANTENNAS_JOINTS = {
+    "l_antenna.pos": "head.l_antenna",
+    "r_antenna.pos": "head.r_antenna",
+}
+
+REACHY2_R_ARM_JOINTS = {
+    "r_shoulder_pitch.pos": "r_arm.shoulder.pitch",
+    "r_shoulder_roll.pos": "r_arm.shoulder.roll",
+    "r_elbow_yaw.pos": "r_arm.elbow.yaw",
+    "r_elbow_pitch.pos": "r_arm.elbow.pitch",
+    "r_wrist_roll.pos": "r_arm.wrist.roll",
+    "r_wrist_pitch.pos": "r_arm.wrist.pitch",
+    "r_wrist_yaw.pos": "r_arm.wrist.yaw",
+    "r_gripper.pos": "r_arm.gripper",
+}
+
+REACHY2_L_ARM_JOINTS = {
+    "l_shoulder_pitch.pos": "l_arm.shoulder.pitch",
+    "l_shoulder_roll.pos": "l_arm.shoulder.roll",
+    "l_elbow_yaw.pos": "l_arm.elbow.yaw",
+    "l_elbow_pitch.pos": "l_arm.elbow.pitch",
+    "l_wrist_roll.pos": "l_arm.wrist.roll",
+    "l_wrist_pitch.pos": "l_arm.wrist.pitch",
+    "l_wrist_yaw.pos": "l_arm.wrist.yaw",
+    "l_gripper.pos": "l_arm.gripper",
+}
+
+REACHY2_VEL = {
+    "mobile_base.vx": "vx",
+    "mobile_base.vy": "vy",
+    "mobile_base.vtheta": "vtheta",
+}
+
+
+class Reachy2Robot(Robot):
+    """
+    [Reachy 2](https://www.pollen-robotics.com/reachy/), by Pollen Robotics.
+    """
+
+    config_class = Reachy2RobotConfig
+    name = "reachy2"
+
+    def __init__(self, config: Reachy2RobotConfig):
+        super().__init__(config)
+
+        self.config = config
+        self.robot_type = self.config.type
+        self.use_external_commands = self.config.use_external_commands
+
+        self.reachy: None | ReachySDK = None
+        self.cameras = make_cameras_from_configs(config.cameras)
+
+        self.logs: dict[str, float] = {}
+
+        self.joints_dict: dict[str, str] = self._generate_joints_dict()
+
+    @property
+    def observation_features(self) -> dict[str, Any]:
+        return {**self.motors_features, **self.camera_features}
+
+    @property
+    def action_features(self) -> dict[str, type]:
+        return self.motors_features
+
+    @property
+    def camera_features(self) -> dict[str, tuple[int | None, int | None, int]]:
+        return {cam: (self.cameras[cam].height, self.cameras[cam].width, 3) for cam in self.cameras}
+
+    @property
+    def motors_features(self) -> dict[str, type]:
+        if self.config.with_mobile_base:
+            return {
+                **dict.fromkeys(
+                    self.joints_dict.keys(),
+                    float,
+                ),
+                **dict.fromkeys(
+                    REACHY2_VEL.keys(),
+                    float,
+                ),
+            }
+        else:
+            return dict.fromkeys(self.joints_dict.keys(), float)
+
+    @property
+    def is_connected(self) -> bool:
+        return self.reachy.is_connected() if self.reachy is not None else False
+
+    def connect(self, calibrate: bool = False) -> None:
+        self.reachy = ReachySDK(self.config.ip_address)
+        if not self.is_connected:
+            raise ConnectionError()
+
+        for cam in self.cameras.values():
+            cam.connect()
+
+        self.configure()
+
+    def configure(self) -> None:
+        if self.reachy is not None:
+            self.reachy.turn_on()
+            self.reachy.reset_default_limits()
+
+    @property
+    def is_calibrated(self) -> bool:
+        return True
+
+    def calibrate(self) -> None:
+        pass
+
+    def _generate_joints_dict(self) -> dict[str, str]:
+        joints = {}
+        if self.config.with_neck:
+            joints.update(REACHY2_NECK_JOINTS)
+        if self.config.with_l_arm:
+            joints.update(REACHY2_L_ARM_JOINTS)
+        if self.config.with_r_arm:
+            joints.update(REACHY2_R_ARM_JOINTS)
+        if self.config.with_antennas:
+            joints.update(REACHY2_ANTENNAS_JOINTS)
+        return joints
+
+    def _get_state(self) -> dict[str, float]:
+        if self.reachy is not None:
+            pos_dict = {k: self.reachy.joints[v].present_position for k, v in self.joints_dict.items()}
+            if not self.config.with_mobile_base:
+                return pos_dict
+            vel_dict = {k: self.reachy.mobile_base.odometry[v] for k, v in REACHY2_VEL.items()}
+            return {**pos_dict, **vel_dict}
+        else:
+            return {}
+
+    def get_observation(self) -> dict[str, np.ndarray]:
+        obs_dict: dict[str, Any] = {}
+
+        # Read Reachy 2 state
+        before_read_t = time.perf_counter()
+        obs_dict.update(self._get_state())
+        self.logs["read_pos_dt_s"] = time.perf_counter() - before_read_t
+
+        # Capture images from cameras
+        for cam_key, cam in self.cameras.items():
+            obs_dict[cam_key] = cam.async_read()
+
+        return obs_dict
+
+    def send_action(self, action: dict[str, Any]) -> dict[str, Any]:
+        if self.reachy is not None:
+            if not self.is_connected:
+                raise ConnectionError()
+
+            before_write_t = time.perf_counter()
+
+            vel = {}
+            goal_pos = {}
+            for key, val in action.items():
+                if key not in self.joints_dict:
+                    if key not in REACHY2_VEL:
+                        raise KeyError(f"Key '{key}' is not a valid motor key in Reachy 2.")
+                    else:
+                        vel[REACHY2_VEL[key]] = float(val)
+                else:
+                    if not self.use_external_commands and self.config.max_relative_target is not None:
+                        goal_pos[key] = float(val)
+                        goal_present_pos = {
+                            key: (
+                                goal_pos[key],
+                                self.reachy.joints[self.joints_dict[key]].present_position,
+                            )
+                        }
+                        safe_goal_pos = ensure_safe_goal_position(
+                            goal_present_pos, float(self.config.max_relative_target)
+                        )
+                        val = safe_goal_pos[key]
+                    self.reachy.joints[self.joints_dict[key]].goal_position = float(val)
+
+            if self.config.with_mobile_base:
+                self.reachy.mobile_base.set_goal_speed(vel["vx"], vel["vy"], vel["vtheta"])
+
+            # We don't send the goal positions if we control Reachy 2 externally
+            if not self.use_external_commands:
+                self.reachy.send_goal_positions()
+                if self.config.with_mobile_base:
+                    self.reachy.mobile_base.send_speed_command()
+
+            self.logs["write_pos_dt_s"] = time.perf_counter() - before_write_t
+        return action
+
+    def disconnect(self) -> None:
+        if self.reachy is not None:
+            for cam in self.cameras.values():
+                cam.disconnect()
+            if self.config.disable_torque_on_disconnect:
+                self.reachy.turn_off_smoothly()
+            self.reachy.disconnect()
@@ -30,9 +30,9 @@ class SO100FollowerConfig(RobotConfig):
    disable_torque_on_disconnect: bool = True

    # `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
-    # Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
-    # the number of motors in your follower arms.
-    max_relative_target: int | None = None
+    # Set this to a positive scalar to have the same value for all motors, or a dictionary that maps motor
+    # names to the max_relative_target value for that motor.
+    max_relative_target: float | dict[str, float] | None = None

    # cameras
    cameras: dict[str, CameraConfig] = field(default_factory=dict)
@@ -161,6 +161,11 @@ class SO100Follower(Robot):
                self.bus.write("I_Coefficient", motor, 0)
                self.bus.write("D_Coefficient", motor, 32)

+                if motor == "gripper":
+                    self.bus.write("Max_Torque_Limit", motor, 500)  # 50% of max torque to avoid burnout
+                    self.bus.write("Protection_Current", motor, 250)  # 50% of max current to avoid burnout
+                    self.bus.write("Overload_Torque", motor, 25)  # 25% torque when overloaded
+
    def setup_motors(self) -> None:
        for motor in reversed(self.bus.motors):
            input(f"Connect the controller board to the '{motor}' motor only and press enter.")
@@ -30,9 +30,9 @@ class SO101FollowerConfig(RobotConfig):
    disable_torque_on_disconnect: bool = True

    # `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
-    # Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
-    # the number of motors in your follower arms.
-    max_relative_target: int | None = None
+    # Set this to a positive scalar to have the same value for all motors, or a dictionary that maps motor
+    # names to the max_relative_target value for that motor.
+    max_relative_target: float | dict[str, float] | None = None

    # cameras
    cameras: dict[str, CameraConfig] = field(default_factory=dict)
@@ -157,6 +157,13 @@ class SO101Follower(Robot):
                self.bus.write("I_Coefficient", motor, 0)
                self.bus.write("D_Coefficient", motor, 32)

+                if motor == "gripper":
+                    self.bus.write(
+                        "Max_Torque_Limit", motor, 500
+                    )  # 50% of the max torque limit to avoid burnout
+                    self.bus.write("Protection_Current", motor, 250)  # 50% of max current to avoid burnout
+                    self.bus.write("Overload_Torque", motor, 25)  # 25% torque when overloaded
+
    def setup_motors(self) -> None:
        for motor in reversed(self.bus.motors):
            input(f"Connect the controller board to the '{motor}' motor only and press enter.")
@@ -24,11 +24,6 @@ from ..config import RobotConfig
@RobotConfig.register_subclass("stretch3")
@dataclass
 class Stretch3RobotConfig(RobotConfig):
-    # `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
-    # Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
-    # the number of motors in your follower arms.
-    max_relative_target: int | None = None
-
    # cameras
    cameras: dict[str, CameraConfig] = field(
        default_factory=lambda: {
@@ -61,6 +61,10 @@ def make_robot_from_config(config: RobotConfig) -> Robot:
        from .bi_so100_follower import BiSO100Follower

        return BiSO100Follower(config)
+    elif config.type == "reachy2":
+        from .reachy2 import Reachy2Robot
+
+        return Reachy2Robot(config)
    elif config.type == "mock_robot":
        from tests.mocks.mock_robot import MockRobot

@@ -70,7 +74,7 @@ def make_robot_from_config(config: RobotConfig) -> Robot:


 def ensure_safe_goal_position(
-    goal_present_pos: dict[str, tuple[float, float]], max_relative_target: float | dict[float]
+    goal_present_pos: dict[str, tuple[float, float]], max_relative_target: float | dict[str, float]
 ) -> dict[str, float]:
    """Caps relative action target magnitude for safety."""

@@ -115,11 +115,11 @@ If you uploaded your dataset to the hub with `--control.push_to_hub=true`, you c
 echo ${HF_USER}/aloha_test
 ```

-If you didn't upload with `--control.push_to_hub=false`, you can also visualize it locally with:
+If you didn't upload with `--control.push_to_hub=false`, you can also visualize it locally with [Rerun](https://github.com/rerun-io/rerun):

 ```bash
-python -m lerobot.scripts.visualize_dataset_html \
-  --repo-id ${HF_USER}/aloha_test
+python -m lerobot.scripts.visualize_dataset \
+  --repo-id ${HF_USER}/aloha_test --episode 0
 ```

 ## Replay an episode
@@ -141,10 +141,10 @@ python lerobot/scripts/control_robot.py \

 ## Train a policy

-To train a policy to control your robot, use the [`python -m lerobot.scripts.train`](../src/lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
+To train a policy to control your robot, use the [`lerobot-train`](../src/lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:

 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
  --dataset.repo_id=${HF_USER}/aloha_test \
  --policy.type=act \
  --output_dir=outputs/train/act_aloha_test \
@@ -28,15 +28,15 @@ class ViperXConfig(RobotConfig):

    # /!\ FOR SAFETY, READ THIS /!\
    # `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
-    # Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
-    # the number of motors in your follower arms.
+    # Set this to a positive scalar to have the same value for all motors, or a dictionary that maps motor
+    # names to the max_relative_target value for that motor.
    # For Aloha, for every goal position request, motor rotations are capped at 5 degrees by default.
    # When you feel more confident with teleoperation or running the policy, you can extend
    # this safety limit and even removing it by setting it to `null`.
    # Also, everything is expected to work safely out-of-the-box, but we highly advise to
    # first try to teleoperate the grippers only (by commenting out the rest of the motors in this yaml),
    # then to gradually add more motors (by uncommenting), until you can teleoperate both arms fully
-    max_relative_target: int | None = 5
+    max_relative_target: float | dict[str, float] = 5.0

    # cameras
    cameras: dict[str, CameraConfig] = field(default_factory=dict)
@@ -21,7 +21,7 @@ You want to evaluate a model from the hub (eg: https://huggingface.co/lerobot/di
 for 10 episodes.

 ```
-python -m lerobot.scripts.eval \
+lerobot-eval \
    --policy.path=lerobot/diffusion_pusht \
    --env.type=pusht \
    --eval.batch_size=10 \
@@ -32,7 +32,7 @@ python -m lerobot.scripts.eval \

 OR, you want to evaluate a model checkpoint from the LeRobot training script for 10 episodes.
 ```
-python -m lerobot.scripts.eval \
+lerobot-eval \
    --policy.path=outputs/train/diffusion_pusht/checkpoints/005000/pretrained_model \
    --env.type=pusht \
    --eval.batch_size=10 \
@@ -0,0 +1,234 @@
+import argparse
+import datetime
+import os
+import time
+from collections.abc import Callable
+
+import numpy as np
+import pandas as pd
+from tqdm import tqdm
+
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.streaming_dataset import StreamingLeRobotDataset
+
+
+def profile_throughput_indexed(
+    dataset: LeRobotDataset, num_samples: int, warmup_iters: int = 3
+) -> np.ndarray:
+    """Measure per-item access time on an indexable LeRobotDataset.
+
+    Accesses dataset[i % len(dataset)] for ``num_samples`` iterations, with an initial warmup.
+    """
+    next_times = np.zeros(num_samples)
+    total = len(dataset)
+
+    # warmup
+    for k in range(warmup_iters):
+        _ = dataset[k % total]
+
+    for j in tqdm(range(num_samples), desc="Profiling dataset throughput", unit="item"):
+        start_time = time.perf_counter()
+        _ = dataset[j % total]
+        end_time = time.perf_counter()
+        next_times[j] = end_time - start_time
+
+    return next_times
+
+
+def profile_throughput(
+    dataset: StreamingLeRobotDataset, num_samples: int, warmup_iters: int = 3
+) -> np.ndarray:
+    """Measure ``.next()`` call latency on a streaming dataset.
+
+    Performs a configurable warmup. This does not numerically "normalize" times; it simply
+    avoids including initialization overhead in the timing window.
+    """
+    next_times = np.zeros(num_samples)
+    iter_dataset = iter(dataset)
+
+    # warmup
+    for _ in range(warmup_iters):
+        _ = next(iter_dataset)
+
+    for j in tqdm(range(num_samples), desc="Profiling throughput", unit="call"):
+        start_time = time.perf_counter()
+        _sample = next(iter_dataset)
+        end_time = time.perf_counter()
+        next_times[j] = end_time - start_time
+
+    return next_times
+
+
+def profile_init(dataset_factory: Callable[[], StreamingLeRobotDataset], num_runs: int) -> np.ndarray:
+    """Measure time-to-first-sample by re-instantiating the dataset ``num_runs`` times.
+
+    Using a factory avoids unsafe ``deepcopy`` of objects that may own threads or file handles.
+    """
+    init_times = np.zeros(num_runs)
+    for i in tqdm(range(num_runs), desc="Profiling init", unit="run"):
+        fresh_dataset = dataset_factory()
+        iter_dataset = iter(fresh_dataset)
+        start_time = time.perf_counter()
+        _ = next(iter_dataset)
+        end_time = time.perf_counter()
+        init_times[i] = end_time - start_time
+
+    return init_times
+
+
+def profile_randomness(dataset: StreamingLeRobotDataset, num_samples: int) -> float:
+    """Measure how random the sample order is via correlation.
+
+    Returns a Pearson correlation between retrieved frame index and iteration index.
+    - ~0: random order
+    - ~+1: strictly increasing (in-order)
+    - ~-1: strictly decreasing (reverse order)
+    """
+    frame_indices = np.zeros(num_samples, dtype=float)
+    iter_indices = np.arange(num_samples, dtype=float)
+
+    iter_dataset = iter(dataset)
+
+    for i in tqdm(range(num_samples), desc="Profiling randomness", unit="sample"):
+        sample = next(iter_dataset)
+        if "index" in sample:
+            frame_idx_value = sample["index"]
+        elif "frame_index" in sample:
+            frame_idx_value = sample["frame_index"]
+        else:
+            raise KeyError("Sample is missing 'index' (or 'frame_index') required to compute randomness.")
+        frame_indices[i] = float(frame_idx_value)
+
+    # Guard against degenerate cases
+    if num_samples < 2 or np.std(frame_indices) == 0 or np.std(iter_indices) == 0:
+        return np.nan, None
+
+    correlation = float(np.corrcoef(frame_indices, iter_indices)[0, 1])
+    return correlation
+
+
+def profile_streaming_dataset(
+    repo_id: str,
+    delta_timestamps: dict[str, list[float]] | None = None,
+    num_samples: int = 100,
+    warmup_iters: int = 10,
+    buffer_size: int = 1000,
+) -> tuple[np.ndarray, np.ndarray, float]:
+    """Run init, throughput, and randomness profiles on a StreamingLeRobotDataset."""
+
+    def dataset_factory() -> StreamingLeRobotDataset:
+        return StreamingLeRobotDataset(repo_id, delta_timestamps=delta_timestamps, buffer_size=buffer_size)
+
+    # Measure init by repeated instantiation
+    init_times = profile_init(dataset_factory, num_runs=warmup_iters)
+
+    # Throughput and randomness on a single fresh dataset instance
+    dataset = dataset_factory()
+    next_times = profile_throughput(dataset, num_samples=num_samples, warmup_iters=warmup_iters)
+    correlation = profile_randomness(dataset, num_samples=num_samples)
+
+    return init_times, next_times, correlation
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Profile StreamingLeRobotDataset performance metrics.")
+    parser.add_argument(
+        "--repo-id",
+        type=str,
+        default="lerobot/svla_so101_pickplace",
+        help="Dataset repo_id to profile.",
+    )
+    parser.add_argument(
+        "--num-samples",
+        type=int,
+        default=1000,
+        help="Number of samples to measure for throughput/randomness.",
+    )
+    parser.add_argument(
+        "--warmup-iters",
+        type=int,
+        default=10,
+        help="Number of iterations for init and throughput warmup.",
+    )
+    parser.add_argument(
+        "--buffer-size",
+        type=int,
+        default=1000,
+        help="Buffer size for the streaming dataset.",
+    )
+    parser.add_argument(
+        "--with-delta-timestamps",
+        action="store_true",
+        help="Profile with delta timestamps.",
+    )
+    parser.add_argument(
+        "--compare-with-local",
+        action="store_true",
+        help="Also profile local LeRobotDataset throughput for comparison.",
+    )
+    parser.add_argument(
+        "--outdir",
+        type=str,
+        default=os.path.join("outputs", "benchmarks"),
+        help="Directory to write CSVs/PNGs to.",
+    )
+
+    args = parser.parse_args()
+
+    delta_timestamps = (
+        None
+        if not args.with_delta_timestamps
+        else {
+            "observation.state": [-2.0, -1.0, -0.5, 0.0, 0.5, 1.0],
+            "action": [
+                -0.1,
+                0.0,
+                0.1,
+                0.2,
+                0.3,
+                0.4,
+                0.5,
+                0.6,
+                0.7,
+                0.8,
+                0.9,
+                1.0,
+            ],
+        }
+    )
+
+    init_times, next_times, correlation = profile_streaming_dataset(
+        repo_id=args.repo_id,
+        delta_timestamps=delta_timestamps,
+        num_samples=args.num_samples,
+        warmup_iters=args.warmup_iters,
+        buffer_size=args.buffer_size,
+    )
+
+    os.makedirs(args.outdir, exist_ok=True)
+
+    repo_id_str = args.repo_id.replace("/", "-")
+    date_str = datetime.datetime.now().strftime("%Y-%m-%d")
+    name_suffix = f"{repo_id_str}_buf{args.buffer_size}_{date_str}"
+
+    # Visualization disabled by default; figures are not created or saved.
+
+    init_df = pd.DataFrame({"init_times": init_times})
+    next_df = pd.DataFrame({"next_times": next_times})
+    correlation_df = pd.DataFrame({"correlation": [correlation]})
+
+    init_df.to_csv(os.path.join(args.outdir, f"init_times_{name_suffix}.csv"), index=False)
+    next_df.to_csv(os.path.join(args.outdir, f"next_times_{name_suffix}.csv"), index=False)
+    correlation_df.to_csv(os.path.join(args.outdir, f"correlation_{name_suffix}.csv"), index=False)
+
+    if args.compare_with_local:
+        # Profile local non-streaming dataset throughput for comparison
+        local_ds = LeRobotDataset(args.repo_id, delta_timestamps=delta_timestamps)
+        local_next_times = profile_throughput_indexed(
+            local_ds, num_samples=args.num_samples, warmup_iters=args.warmup_iters
+        )
+        local_df = pd.DataFrame({"next_times": local_next_times})
+        local_df.to_csv(
+            os.path.join(args.outdir, f"next_times_local_{repo_id_str}_{date_str}.csv"),
+            index=False,
+        )
@@ -226,7 +226,8 @@ def convert_lerobot_dataset_to_cropper_lerobot_dataset(
                value = value.unsqueeze(0)
            new_frame[key] = value

-        new_dataset.add_frame(new_frame, task=task)
+        new_frame["task"] = task
+        new_dataset.add_frame(new_frame)

        if frame["episode_index"].item() != prev_episode_index:
            # Save the episode
@@ -2129,7 +2129,8 @@ def record_dataset(env, policy, cfg):
            frame["complementary_info.discrete_penalty"] = torch.tensor(
                [info.get("discrete_penalty", 0.0)], dtype=torch.float32
            )
-            dataset.add_frame(frame, task=cfg.task)
+            frame["task"] = cfg.task
+            dataset.add_frame(frame)

            # Maintain consistent timing
            if cfg.fps:
@@ -302,11 +302,6 @@ class RobotClient:

                    self.logger.debug(f"Current latest action: {latest_action}")

-                    # Get queue state before changes
-                    old_size, old_timesteps = self._inspect_action_queue()
-                    if not old_timesteps:
-                        old_timesteps = [latest_action]  # queue was empty
-
                    # Get queue state before changes
                    old_size, old_timesteps = self._inspect_action_queue()
                    if not old_timesteps:
@@ -166,7 +166,8 @@ def train(cfg: TrainPipelineConfig):
    if hasattr(cfg.policy, "drop_n_last_frames"):
        shuffle = False
        sampler = EpisodeAwareSampler(
-            dataset.episode_data_index,
+            dataset.meta.episodes["dataset_from_index"],
+            dataset.meta.episodes["dataset_to_index"],
            drop_n_last_frames=cfg.policy.drop_n_last_frames,
            shuffle=True,
        )
@@ -178,10 +179,11 @@ def train(cfg: TrainPipelineConfig):
        dataset,
        num_workers=cfg.num_workers,
        batch_size=cfg.batch_size,
-        shuffle=shuffle,
+        shuffle=shuffle and not cfg.dataset.streaming,
        sampler=sampler,
        pin_memory=device.type == "cuda",
        drop_last=False,
+        prefetch_factor=2,
    )
    dl_iter = cycle(dataloader)

@@ -207,6 +209,9 @@ def train(cfg: TrainPipelineConfig):

        for key in batch:
            if isinstance(batch[key], torch.Tensor):
+                if batch[key].dtype != torch.bool:
+                    batch[key] = batch[key].type(torch.float32) if device.type == "mps" else batch[key]
+
                batch[key] = batch[key].to(device, non_blocking=device.type == "cuda")

        train_tracker, output_dict = update_policy(
@@ -79,8 +79,8 @@ from lerobot.datasets.lerobot_dataset import LeRobotDataset

 class EpisodeSampler(torch.utils.data.Sampler):
    def __init__(self, dataset: LeRobotDataset, episode_index: int):
-        from_idx = dataset.episode_data_index["from"][episode_index].item()
-        to_idx = dataset.episode_data_index["to"][episode_index].item()
+        from_idx = dataset.meta.episodes["dataset_from_index"][episode_index]
+        to_idx = dataset.meta.episodes["dataset_to_index"][episode_index]
        self.frame_ids = range(from_idx, to_idx)

    def __iter__(self) -> Iterator:
@@ -283,7 +283,7 @@ def main():
    tolerance_s = kwargs.pop("tolerance_s")

    logging.info("Loading dataset")
-    dataset = LeRobotDataset(repo_id, root=root, tolerance_s=tolerance_s)
+    dataset = LeRobotDataset(repo_id, episodes=[args.episode_index], root=root, tolerance_s=tolerance_s)

    visualize_dataset(dataset, **vars(args))

@@ -1,482 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-""" Visualize data of **all** frames of any episode of a dataset of type LeRobotDataset.
-
-Note: The last frame of the episode doesnt always correspond to a final state.
-That's because our datasets are composed of transition from state to state up to
-the antepenultimate state associated to the ultimate action to arrive in the final state.
-However, there might not be a transition from a final state to another state.
-
-Note: This script aims to visualize the data used to train the neural networks.
-~What you see is what you get~. When visualizing image modality, it is often expected to observe
-lossly compression artifacts since these images have been decoded from compressed mp4 videos to
-save disk space. The compression factor applied has been tuned to not affect success rate.
-
-Example of usage:
-
- Visualize data stored on a local machine:
-```bash
-local$ python -m lerobot.scripts.visualize_dataset_html \
-    --repo-id lerobot/pusht
-
-local$ open http://localhost:9090
-```
-
- Visualize data stored on a distant machine with a local viewer:
-```bash
-distant$ python -m lerobot.scripts.visualize_dataset_html \
-    --repo-id lerobot/pusht
-
-local$ ssh -L 9090:localhost:9090 distant  # create a ssh tunnel
-local$ open http://localhost:9090
-```
-
- Select episodes to visualize:
-```bash
-python -m lerobot.scripts.visualize_dataset_html \
-    --repo-id lerobot/pusht \
-    --episodes 7 3 5 1 4
-```
-"""
-
-import argparse
-import csv
-import json
-import logging
-import re
-import shutil
-import tempfile
-from io import StringIO
-from pathlib import Path
-
-import numpy as np
-import pandas as pd
-import requests
-from flask import Flask, redirect, render_template, request, url_for
-
-from lerobot import available_datasets
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.datasets.utils import IterableNamespace
-from lerobot.utils.utils import init_logging
-
-
-def run_server(
-    dataset: LeRobotDataset | IterableNamespace | None,
-    episodes: list[int] | None,
-    host: str,
-    port: str,
-    static_folder: Path,
-    template_folder: Path,
-):
-    app = Flask(__name__, static_folder=static_folder.resolve(), template_folder=template_folder.resolve())
-    app.config["SEND_FILE_MAX_AGE_DEFAULT"] = 0  # specifying not to cache
-
-    @app.route("/")
-    def hommepage(dataset=dataset):
-        if dataset:
-            dataset_namespace, dataset_name = dataset.repo_id.split("/")
-            return redirect(
-                url_for(
-                    "show_episode",
-                    dataset_namespace=dataset_namespace,
-                    dataset_name=dataset_name,
-                    episode_id=0,
-                )
-            )
-
-        dataset_param, episode_param = None, None
-        all_params = request.args
-        if "dataset" in all_params:
-            dataset_param = all_params["dataset"]
-        if "episode" in all_params:
-            episode_param = int(all_params["episode"])
-
-        if dataset_param:
-            dataset_namespace, dataset_name = dataset_param.split("/")
-            return redirect(
-                url_for(
-                    "show_episode",
-                    dataset_namespace=dataset_namespace,
-                    dataset_name=dataset_name,
-                    episode_id=episode_param if episode_param is not None else 0,
-                )
-            )
-
-        featured_datasets = [
-            "lerobot/aloha_static_cups_open",
-            "lerobot/columbia_cairlab_pusht_real",
-            "lerobot/taco_play",
-        ]
-        return render_template(
-            "visualize_dataset_homepage.html",
-            featured_datasets=featured_datasets,
-            lerobot_datasets=available_datasets,
-        )
-
-    @app.route("/<string:dataset_namespace>/<string:dataset_name>")
-    def show_first_episode(dataset_namespace, dataset_name):
-        first_episode_id = 0
-        return redirect(
-            url_for(
-                "show_episode",
-                dataset_namespace=dataset_namespace,
-                dataset_name=dataset_name,
-                episode_id=first_episode_id,
-            )
-        )
-
-    @app.route("/<string:dataset_namespace>/<string:dataset_name>/episode_<int:episode_id>")
-    def show_episode(dataset_namespace, dataset_name, episode_id, dataset=dataset, episodes=episodes):
-        repo_id = f"{dataset_namespace}/{dataset_name}"
-        try:
-            if dataset is None:
-                dataset = get_dataset_info(repo_id)
-        except FileNotFoundError:
-            return (
-                "Make sure to convert your LeRobotDataset to v2 & above. See how to convert your dataset at https://github.com/huggingface/lerobot/pull/461",
-                400,
-            )
-        dataset_version = (
-            str(dataset.meta._version) if isinstance(dataset, LeRobotDataset) else dataset.codebase_version
-        )
-        match = re.search(r"v(\d+)\.", dataset_version)
-        if match:
-            major_version = int(match.group(1))
-            if major_version < 2:
-                return "Make sure to convert your LeRobotDataset to v2 & above."
-
-        episode_data_csv_str, columns, ignored_columns = get_episode_data(dataset, episode_id)
-        dataset_info = {
-            "repo_id": f"{dataset_namespace}/{dataset_name}",
-            "num_samples": dataset.num_frames
-            if isinstance(dataset, LeRobotDataset)
-            else dataset.total_frames,
-            "num_episodes": dataset.num_episodes
-            if isinstance(dataset, LeRobotDataset)
-            else dataset.total_episodes,
-            "fps": dataset.fps,
-        }
-        if isinstance(dataset, LeRobotDataset):
-            video_paths = [
-                dataset.meta.get_video_file_path(episode_id, key) for key in dataset.meta.video_keys
-            ]
-            videos_info = [
-                {
-                    "url": url_for("static", filename=str(video_path).replace("\\", "/")),
-                    "filename": video_path.parent.name,
-                }
-                for video_path in video_paths
-            ]
-            tasks = dataset.meta.episodes[episode_id]["tasks"]
-        else:
-            video_keys = [key for key, ft in dataset.features.items() if ft["dtype"] == "video"]
-            videos_info = [
-                {
-                    "url": f"https://huggingface.co/datasets/{repo_id}/resolve/main/"
-                    + dataset.video_path.format(
-                        episode_chunk=int(episode_id) // dataset.chunks_size,
-                        video_key=video_key,
-                        episode_index=episode_id,
-                    ),
-                    "filename": video_key,
-                }
-                for video_key in video_keys
-            ]
-
-            response = requests.get(
-                f"https://huggingface.co/datasets/{repo_id}/resolve/main/meta/episodes.jsonl", timeout=5
-            )
-            response.raise_for_status()
-            # Split into lines and parse each line as JSON
-            tasks_jsonl = [json.loads(line) for line in response.text.splitlines() if line.strip()]
-
-            filtered_tasks_jsonl = [row for row in tasks_jsonl if row["episode_index"] == episode_id]
-            tasks = filtered_tasks_jsonl[0]["tasks"]
-
-        videos_info[0]["language_instruction"] = tasks
-
-        if episodes is None:
-            episodes = list(
-                range(dataset.num_episodes if isinstance(dataset, LeRobotDataset) else dataset.total_episodes)
-            )
-
-        return render_template(
-            "visualize_dataset_template.html",
-            episode_id=episode_id,
-            episodes=episodes,
-            dataset_info=dataset_info,
-            videos_info=videos_info,
-            episode_data_csv_str=episode_data_csv_str,
-            columns=columns,
-            ignored_columns=ignored_columns,
-        )
-
-    app.run(host=host, port=port)
-
-
-def get_ep_csv_fname(episode_id: int):
-    ep_csv_fname = f"episode_{episode_id}.csv"
-    return ep_csv_fname
-
-
-def get_episode_data(dataset: LeRobotDataset | IterableNamespace, episode_index):
-    """Get a csv str containing timeseries data of an episode (e.g. state and action).
-    This file will be loaded by Dygraph javascript to plot data in real time."""
-    columns = []
-
-    selected_columns = [col for col, ft in dataset.features.items() if ft["dtype"] in ["float32", "int32"]]
-    selected_columns.remove("timestamp")
-
-    ignored_columns = []
-    for column_name in selected_columns:
-        shape = dataset.features[column_name]["shape"]
-        shape_dim = len(shape)
-        if shape_dim > 1:
-            selected_columns.remove(column_name)
-            ignored_columns.append(column_name)
-
-    # init header of csv with state and action names
-    header = ["timestamp"]
-
-    for column_name in selected_columns:
-        dim_state = (
-            dataset.meta.shapes[column_name][0]
-            if isinstance(dataset, LeRobotDataset)
-            else dataset.features[column_name].shape[0]
-        )
-
-        if "names" in dataset.features[column_name] and dataset.features[column_name]["names"]:
-            column_names = dataset.features[column_name]["names"]
-            while not isinstance(column_names, list):
-                column_names = list(column_names.values())[0]
-        else:
-            column_names = [f"{column_name}_{i}" for i in range(dim_state)]
-        columns.append({"key": column_name, "value": column_names})
-
-        header += column_names
-
-    selected_columns.insert(0, "timestamp")
-
-    if isinstance(dataset, LeRobotDataset):
-        from_idx = dataset.episode_data_index["from"][episode_index]
-        to_idx = dataset.episode_data_index["to"][episode_index]
-        data = (
-            dataset.hf_dataset.select(range(from_idx, to_idx))
-            .select_columns(selected_columns)
-            .with_format("pandas")
-        )
-    else:
-        repo_id = dataset.repo_id
-
-        url = f"https://huggingface.co/datasets/{repo_id}/resolve/main/" + dataset.data_path.format(
-            episode_chunk=int(episode_index) // dataset.chunks_size, episode_index=episode_index
-        )
-        df = pd.read_parquet(url)
-        data = df[selected_columns]  # Select specific columns
-
-    rows = np.hstack(
-        (
-            np.expand_dims(data["timestamp"], axis=1),
-            *[np.vstack(data[col]) for col in selected_columns[1:]],
-        )
-    ).tolist()
-
-    # Convert data to CSV string
-    csv_buffer = StringIO()
-    csv_writer = csv.writer(csv_buffer)
-    # Write header
-    csv_writer.writerow(header)
-    # Write data rows
-    csv_writer.writerows(rows)
-    csv_string = csv_buffer.getvalue()
-
-    return csv_string, columns, ignored_columns
-
-
-def get_episode_video_paths(dataset: LeRobotDataset, ep_index: int) -> list[str]:
-    # get first frame of episode (hack to get video_path of the episode)
-    first_frame_idx = dataset.episode_data_index["from"][ep_index].item()
-    return [
-        dataset.hf_dataset.select_columns(key)[first_frame_idx][key]["path"]
-        for key in dataset.meta.video_keys
-    ]
-
-
-def get_episode_language_instruction(dataset: LeRobotDataset, ep_index: int) -> list[str]:
-    # check if the dataset has language instructions
-    if "language_instruction" not in dataset.features:
-        return None
-
-    # get first frame index
-    first_frame_idx = dataset.episode_data_index["from"][ep_index].item()
-
-    language_instruction = dataset.hf_dataset[first_frame_idx]["language_instruction"]
-    # TODO (michel-aractingi) hack to get the sentence, some strings in openx are badly stored
-    # with the tf.tensor appearing in the string
-    return language_instruction.removeprefix("tf.Tensor(b'").removesuffix("', shape=(), dtype=string)")
-
-
-def get_dataset_info(repo_id: str) -> IterableNamespace:
-    response = requests.get(
-        f"https://huggingface.co/datasets/{repo_id}/resolve/main/meta/info.json", timeout=5
-    )
-    response.raise_for_status()  # Raises an HTTPError for bad responses
-    dataset_info = response.json()
-    dataset_info["repo_id"] = repo_id
-    return IterableNamespace(dataset_info)
-
-
-def visualize_dataset_html(
-    dataset: LeRobotDataset | None,
-    episodes: list[int] | None = None,
-    output_dir: Path | None = None,
-    serve: bool = True,
-    host: str = "127.0.0.1",
-    port: int = 9090,
-    force_override: bool = False,
-) -> Path | None:
-    init_logging()
-
-    template_dir = Path(__file__).resolve().parent.parent / "templates"
-
-    if output_dir is None:
-        # Create a temporary directory that will be automatically cleaned up
-        output_dir = tempfile.mkdtemp(prefix="lerobot_visualize_dataset_")
-
-    output_dir = Path(output_dir)
-    if output_dir.exists():
-        if force_override:
-            shutil.rmtree(output_dir)
-        else:
-            logging.info(f"Output directory already exists. Loading from it: '{output_dir}'")
-
-    output_dir.mkdir(parents=True, exist_ok=True)
-
-    static_dir = output_dir / "static"
-    static_dir.mkdir(parents=True, exist_ok=True)
-
-    if dataset is None:
-        if serve:
-            run_server(
-                dataset=None,
-                episodes=None,
-                host=host,
-                port=port,
-                static_folder=static_dir,
-                template_folder=template_dir,
-            )
-    else:
-        # Create a simlink from the dataset video folder containing mp4 files to the output directory
-        # so that the http server can get access to the mp4 files.
-        if isinstance(dataset, LeRobotDataset):
-            ln_videos_dir = static_dir / "videos"
-            if not ln_videos_dir.exists():
-                ln_videos_dir.symlink_to((dataset.root / "videos").resolve().as_posix())
-
-        if serve:
-            run_server(dataset, episodes, host, port, static_dir, template_dir)
-
-
-def main():
-    parser = argparse.ArgumentParser()
-
-    parser.add_argument(
-        "--repo-id",
-        type=str,
-        default=None,
-        help="Name of hugging face repositery containing a LeRobotDataset dataset (e.g. `lerobot/pusht` for https://huggingface.co/datasets/lerobot/pusht).",
-    )
-    parser.add_argument(
-        "--root",
-        type=Path,
-        default=None,
-        help="Root directory for a dataset stored locally (e.g. `--root data`). By default, the dataset will be loaded from hugging face cache folder, or downloaded from the hub if available.",
-    )
-    parser.add_argument(
-        "--load-from-hf-hub",
-        type=int,
-        default=0,
-        help="Load videos and parquet files from HF Hub rather than local system.",
-    )
-    parser.add_argument(
-        "--episodes",
-        type=int,
-        nargs="*",
-        default=None,
-        help="Episode indices to visualize (e.g. `0 1 5 6` to load episodes of index 0, 1, 5 and 6). By default loads all episodes.",
-    )
-    parser.add_argument(
-        "--output-dir",
-        type=Path,
-        default=None,
-        help="Directory path to write html files and kickoff a web server. By default write them to 'outputs/visualize_dataset/REPO_ID'.",
-    )
-    parser.add_argument(
-        "--serve",
-        type=int,
-        default=1,
-        help="Launch web server.",
-    )
-    parser.add_argument(
-        "--host",
-        type=str,
-        default="127.0.0.1",
-        help="Web host used by the http server.",
-    )
-    parser.add_argument(
-        "--port",
-        type=int,
-        default=9090,
-        help="Web port used by the http server.",
-    )
-    parser.add_argument(
-        "--force-override",
-        type=int,
-        default=0,
-        help="Delete the output directory if it exists already.",
-    )
-
-    parser.add_argument(
-        "--tolerance-s",
-        type=float,
-        default=1e-4,
-        help=(
-            "Tolerance in seconds used to ensure data timestamps respect the dataset fps value"
-            "This is argument passed to the constructor of LeRobotDataset and maps to its tolerance_s constructor argument"
-            "If not given, defaults to 1e-4."
-        ),
-    )
-
-    args = parser.parse_args()
-    kwargs = vars(args)
-    repo_id = kwargs.pop("repo_id")
-    load_from_hf_hub = kwargs.pop("load_from_hf_hub")
-    root = kwargs.pop("root")
-    tolerance_s = kwargs.pop("tolerance_s")
-
-    dataset = None
-    if repo_id:
-        dataset = (
-            LeRobotDataset(repo_id, root=root, tolerance_s=tolerance_s)
-            if not load_from_hf_hub
-            else get_dataset_info(repo_id)
-        )
-
-    visualize_dataset_html(dataset, **vars(args))
-
-
-if __name__ == "__main__":
-    main()
@@ -18,7 +18,7 @@ Helper to set motor ids and baudrate.
 Example:

 ```shell
-python -m lerobot.setup_motors \
+lerobot-setup-motors \
    --teleop.type=so100_leader \
    --teleop.port=/dev/tty.usbmodem575E0031751
 ```
@@ -18,7 +18,7 @@ Simple script to control a robot from teleoperation.
 Example:

 ```shell
-python -m lerobot.teleoperate \
+lerobot-teleoperate \
    --robot.type=so101_follower \
    --robot.port=/dev/tty.usbmodem58760431541 \
    --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 1920, height: 1080, fps: 30}}" \
@@ -32,7 +32,7 @@ python -m lerobot.teleoperate \
 Example teleoperation with bimanual so100:

 ```shell
-python -m lerobot.teleoperate \
+lerobot-teleoperate \
  --robot.type=bi_so100_follower \
  --robot.left_arm_port=/dev/tty.usbmodem5A460851411 \
  --robot.right_arm_port=/dev/tty.usbmodem5A460812391 \
@@ -88,6 +88,7 @@ class KochLeader(Teleoperator):
        return self.bus.is_calibrated

    def calibrate(self) -> None:
+        self.bus.disable_torque()
        if self.calibration:
            # Calibration file exists, ask user whether to use it or run new calibration
            user_input = input(
@@ -98,7 +99,6 @@ class KochLeader(Teleoperator):
                self.bus.write_calibration(self.calibration)
                return
        logger.info(f"\nRunning calibration of {self}")
-        self.bus.disable_torque()
        for motor in self.bus.motors:
            self.bus.write("Operating_Mode", motor, OperatingMode.EXTENDED_POSITION.value)

--- a/Show More
+++ b/Show More