chore(dummy): add comments + better visualization

Co-authored-by: Pepijn <pepijn@huggingface.co>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

.github/workflows/test-docker-build.yml

@@ -41,7 +41,7 @@ jobs:
 
       - name: Get changed files
         id: changed-files
-        uses: tj-actions/changed-files@3f54ebb830831fc121d3263c1857cfbdc310cdb9 #v42
+        uses: tj-actions/changed-files@v44
         with:
           files: docker/**
           json: "true"

.github/workflows/test.yml

@@ -126,7 +126,7 @@ jobs:
       # portaudio19-dev is needed to install pyaudio
         run: |
           sudo apt-get update && \
-          sudo apt-get install -y libegl1-mesa-dev ffmpeg portaudio19-dev
+          sudo apt-get install -y libegl1-mesa-dev portaudio19-dev
 
       - name: Install uv and python
         uses: astral-sh/setup-uv@v5

.pre-commit-config.yaml

@@ -36,8 +36,8 @@ repos:
       - id: end-of-file-fixer
       - id: trailing-whitespace
 
-  - repo: https://github.com/adhtruong/mirrors-typos
-    rev: v1.31.1
+  - repo: https://github.com/crate-ci/typos
+    rev: v1.30.2
     hooks:
       - id: typos
         args: [--force-exclude]
@@ -48,7 +48,7 @@ repos:
     -   id: pyupgrade
 
   - repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.11.5
+    rev: v0.9.10
     hooks:
       - id: ruff
         args: [--fix]
@@ -57,12 +57,12 @@ repos:
 
   ##### Security #####
   - repo: https://github.com/gitleaks/gitleaks
-    rev: v8.24.3
+    rev: v8.24.0
     hooks:
       - id: gitleaks
 
   - repo: https://github.com/woodruffw/zizmor-pre-commit
-    rev: v1.5.2
+    rev: v1.4.1
     hooks:
       - id: zizmor
 

README.md

@@ -98,25 +98,14 @@ conda create -y -n lerobot python=3.10
 conda activate lerobot
 ```
 
-When using `miniconda`, install `ffmpeg` in your environment:
-```bash
-conda install ffmpeg -c conda-forge
-```
-
-> **NOTE:** This usually installs `ffmpeg 7.X` for your platform compiled with the `libsvtav1` encoder. If `libsvtav1` is not supported (check supported encoders with `ffmpeg -encoders`), you can:
->  - _[On any platform]_ Explicitly install `ffmpeg 7.X` using:
->  ```bash
->  conda install ffmpeg=7.1.1 -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:
 ```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:
-`sudo apt-get install cmake build-essential python3-dev pkg-config libavformat-dev libavcodec-dev libavdevice-dev libavutil-dev libswscale-dev libswresample-dev libavfilter-dev pkg-config`. For other systems, see: [Compiling PyAV](https://pyav.org/docs/develop/overview/installation.html#bring-your-own-ffmpeg)
+> **NOTE:** Depending on your platform, If you encounter any build errors during this step
+you may need to install `cmake` and `build-essential` for building some of our dependencies.
+On linux: `sudo apt-get install cmake build-essential`
 
 For simulations, 🤗 LeRobot comes with gymnasium environments that can be installed as extras:
 - [aloha](https://github.com/huggingface/gym-aloha)
@@ -198,7 +187,6 @@ 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:
 
 ```
-TODO: IMPROVE
 dataset attributes:
   ├ hf_dataset: a Hugging Face dataset (backed by Arrow/parquet). Typical features example:
   │  ├ observation.images.cam_high (VideoFrame):
@@ -211,7 +199,7 @@ dataset attributes:
   │  ├ timestamp (float32): timestamp in the episode
   │  ├ next.done (bool): indicates the end of en episode ; True for the last frame in each episode
   │  └ index (int64): general index in the whole dataset
-  ├ meta: contains 2 tensors with the start and end indices of each episode
+  ├ 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

benchmarks/video/README.md

@@ -51,7 +51,7 @@ For a comprehensive list and documentation of these parameters, see the ffmpeg d
 ### Decoding parameters
 **Decoder**
 We tested two video decoding backends from torchvision:
-- `pyav`
+- `pyav` (default)
 - `video_reader` (requires to build torchvision from source)
 
 **Requested timestamps**

benchmarks/video/capture_camera_feed.py

@@ -17,21 +17,12 @@
 
 import argparse
 import datetime as dt
-import os
-import time
 from pathlib import Path
 
 import cv2
-import rerun as rr
-
-# see https://rerun.io/docs/howto/visualization/limit-ram
-RERUN_MEMORY_LIMIT = os.getenv("LEROBOT_RERUN_MEMORY_LIMIT", "5%")
 
 
-def display_and_save_video_stream(output_dir: Path, fps: int, width: int, height: int, duration: int):
-    rr.init("lerobot_capture_camera_feed")
-    rr.spawn(memory_limit=RERUN_MEMORY_LIMIT)
-
+def display_and_save_video_stream(output_dir: Path, fps: int, width: int, height: int):
     now = dt.datetime.now()
     capture_dir = output_dir / f"{now:%Y-%m-%d}" / f"{now:%H-%M-%S}"
     if not capture_dir.exists():
@@ -48,21 +39,24 @@ def display_and_save_video_stream(output_dir: Path, fps: int, width: int, height
     cap.set(cv2.CAP_PROP_FRAME_HEIGHT, height)
 
     frame_index = 0
-    start_time = time.time()
-    while time.time() - start_time < duration:
+    while True:
         ret, frame = cap.read()
 
         if not ret:
             print("Error: Could not read frame.")
             break
-        rr.log("video/stream", rr.Image(frame.numpy()), static=True)
+
+        cv2.imshow("Video Stream", frame)
         cv2.imwrite(str(capture_dir / f"frame_{frame_index:06d}.png"), frame)
         frame_index += 1
 
-    # Release the capture
-    cap.release()
+        # Break the loop on 'q' key press
+        if cv2.waitKey(1) & 0xFF == ord("q"):
+            break
 
-    # TODO(Steven): Add a graceful shutdown via a close() method for the Viewer context, though not currently supported in the Rerun API.
+    # Release the capture and destroy all windows
+    cap.release()
+    cv2.destroyAllWindows()
 
 
 if __name__ == "__main__":
@@ -92,11 +86,5 @@ if __name__ == "__main__":
         default=720,
         help="Height of the captured images.",
     )
-    parser.add_argument(
-        "--duration",
-        type=int,
-        default=20,
-        help="Duration in seconds for which the video stream should be captured.",
-    )
     args = parser.parse_args()
     display_and_save_video_stream(**vars(args))

benchmarks/video/run_video_benchmark.py

@@ -67,7 +67,7 @@ def parse_int_or_none(value) -> int | None:
 def check_datasets_formats(repo_ids: list) -> None:
     for repo_id in repo_ids:
         dataset = LeRobotDataset(repo_id)
-        if len(dataset.meta.video_keys) > 0:
+        if dataset.video:
             raise ValueError(
                 f"Use only image dataset for running this benchmark. Video dataset provided: {repo_id}"
             )
@@ -108,8 +108,7 @@ def save_decoded_frames(
 
 
 def save_first_episode(imgs_dir: Path, dataset: LeRobotDataset) -> None:
-    episode_index = 0
-    ep_num_images = dataset.meta.episodes["length"][episode_index]
+    ep_num_images = dataset.episode_data_index["to"][0].item()
     if imgs_dir.exists() and len(list(imgs_dir.glob("frame_*.png"))) == ep_num_images:
         return
 
@@ -266,8 +265,7 @@ def benchmark_encoding_decoding(
             overwrite=True,
         )
 
-    episode_index = 0
-    ep_num_images = dataset.meta.episodes["length"][episode_index]
+    ep_num_images = dataset.episode_data_index["to"][0].item()
     width, height = tuple(dataset[0][dataset.meta.camera_keys[0]].shape[-2:])
     num_pixels = width * height
     video_size_bytes = video_path.stat().st_size

docker/lerobot-gpu-dev/Dockerfile

@@ -14,7 +14,7 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
     tcpdump sysstat screen tmux \
     libglib2.0-0 libgl1-mesa-glx libegl1-mesa \
     speech-dispatcher portaudio19-dev libgeos-dev \
-    python${PYTHON_VERSION} python${PYTHON_VERSION}-venv python${PYTHON_VERSION}-dev \
+    python${PYTHON_VERSION} python${PYTHON_VERSION}-venv \
     && apt-get clean && rm -rf /var/lib/apt/lists/*
 
 # Install ffmpeg build dependencies. See:

examples/1_load_lerobot_dataset.py

@@ -92,11 +92,11 @@ print(dataset.hf_dataset)
 # LeRobot datasets also subclasses PyTorch datasets so you can do everything you know and love from working
 # with the latter, like iterating through the dataset.
 # The __getitem__ iterates over the frames of the dataset. Since our datasets are also structured by
-# episodes, you can access the frame indices of any episode using dataset.meta.episodes. Here, we access
+# episodes, you can access the frame indices of any episode using the episode_data_index. Here, we access
 # frame indices associated to the first episode:
 episode_index = 0
-from_idx = dataset.meta.episodes["dataset_from_index"][episode_index]
-to_idx = dataset.meta.episodes["dataset_to_index"][episode_index]
+from_idx = dataset.episode_data_index["from"][episode_index].item()
+to_idx = dataset.episode_data_index["to"][episode_index].item()
 
 # Then we grab all the image frames from the first camera:
 camera_key = dataset.meta.camera_keys[0]
@@ -119,7 +119,7 @@ print(dataset.features[camera_key]["shape"])
 delta_timestamps = {
     # loads 4 images: 1 second before current frame, 500 ms before, 200 ms before, and current frame
     camera_key: [-1, -0.5, -0.20, 0],
-    # loads 6 state vectors: 1.5 seconds before, 1 second before, ... 200 ms, 100 ms, and current frame
+    # loads 8 state vectors: 1.5 seconds before, 1 second before, ... 200 ms, 100 ms, and current frame
     "observation.state": [-1.5, -1, -0.5, -0.20, -0.10, 0],
     # loads 64 action vectors: current frame, 1 frame in the future, 2 frames, ... 63 frames in the future
     "action": [t / dataset.fps for t in range(64)],
@@ -143,6 +143,6 @@ dataloader = torch.utils.data.DataLoader(
 
 for batch in dataloader:
     print(f"{batch[camera_key].shape=}")  # (32, 4, c, h, w)
-    print(f"{batch['observation.state'].shape=}")  # (32, 6, c)
+    print(f"{batch['observation.state'].shape=}")  # (32, 5, c)
     print(f"{batch['action'].shape=}")  # (32, 64, c)
     break

examples/2_evaluate_pretrained_policy.py

@@ -18,7 +18,7 @@ training outputs directory. In the latter case, you might want to run examples/3
 
 It requires the installation of the 'gym_pusht' simulation environment. Install it by running:
 ```bash
-pip install -e ".[pusht]"
+pip install -e ".[pusht]"`
 ```
 """
 

examples/4_train_policy_with_script.md

@@ -4,7 +4,7 @@ This tutorial will explain the training script, how to use it, and particularly
 
 ## The training script
 
-LeRobot offers a training script at [`lerobot/scripts/train.py`](../lerobot/scripts/train.py). At a high level it does the following:
+LeRobot offers a training script at [`lerobot/scripts/train.py`](../../lerobot/scripts/train.py). At a high level it does the following:
 
 - Initialize/load a configuration for the following steps using.
 - Instantiates a dataset.
@@ -21,7 +21,7 @@ In the training script, the main function `train` expects a `TrainPipelineConfig
 def train(cfg: TrainPipelineConfig):
 ```
 
-You can inspect the `TrainPipelineConfig` defined in [`lerobot/configs/train.py`](../lerobot/configs/train.py) (which is heavily commented and meant to be a reference to understand any option)
+You can inspect the `TrainPipelineConfig` defined in [`lerobot/configs/train.py`](../../lerobot/configs/train.py) (which is heavily commented and meant to be a reference to understand any option)
 
 When running the script, inputs for the command line are parsed thanks to the `@parser.wrap()` decorator and an instance of this class is automatically generated. Under the hood, this is done with [Draccus](https://github.com/dlwh/draccus) which is a tool dedicated for this purpose. If you're familiar with Hydra, Draccus can similarly load configurations from config files (.json, .yaml) and also override their values through command line inputs. Unlike Hydra, these configurations are pre-defined in the code through dataclasses rather than being defined entirely in config files. This allows for more rigorous serialization/deserialization, typing, and to manipulate configuration as objects directly in the code and not as dictionaries or namespaces (which enables nice features in an IDE such as autocomplete, jump-to-def, etc.)
 
@@ -50,7 +50,7 @@ By default, every field takes its default value specified in the dataclass. If a
 
 ## Specifying values from the CLI
 
-Let's say that we want to train [Diffusion Policy](../lerobot/common/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:
+Let's say that we want to train [Diffusion Policy](../../lerobot/common/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 lerobot/scripts/train.py \
     --dataset.repo_id=lerobot/pusht \
@@ -60,10 +60,10 @@ python lerobot/scripts/train.py \
 
 Let's break this down:
 - To specify the dataset, we just need to specify its `repo_id` on the hub which is the only required argument in the `DatasetConfig`. The rest of the fields have default values and in this case we are fine with those so we can just add the option `--dataset.repo_id=lerobot/pusht`.
-- To specify the policy, we can just select diffusion policy using `--policy` appended with `.type`. Here, `.type` is a special argument which allows us to select config classes inheriting from `draccus.ChoiceRegistry` and that have been decorated with the `register_subclass()` method. To have a better explanation of this feature, have a look at this [Draccus demo](https://github.com/dlwh/draccus?tab=readme-ov-file#more-flexible-configuration-with-choice-types). In our code, we use this mechanism mainly to select policies, environments, robots, and some other components like optimizers. The policies available to select are located in [lerobot/common/policies](../lerobot/common/policies)
-- Similarly, we select the environment with `--env.type=pusht`. The different environment configs are available in [`lerobot/common/envs/configs.py`](../lerobot/common/envs/configs.py)
+- To specify the policy, we can just select diffusion policy using `--policy` appended with `.type`. Here, `.type` is a special argument which allows us to select config classes inheriting from `draccus.ChoiceRegistry` and that have been decorated with the `register_subclass()` method. To have a better explanation of this feature, have a look at this [Draccus demo](https://github.com/dlwh/draccus?tab=readme-ov-file#more-flexible-configuration-with-choice-types). In our code, we use this mechanism mainly to select policies, environments, robots, and some other components like optimizers. The policies available to select are located in [lerobot/common/policies](../../lerobot/common/policies)
+- Similarly, we select the environment with `--env.type=pusht`. The different environment configs are available in [`lerobot/common/envs/configs.py`](../../lerobot/common/envs/configs.py)
 
-Let's see another example. Let's say you've been training [ACT](../lerobot/common/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:
+Let's see another example. Let's say you've been training [ACT](../../lerobot/common/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 lerobot/scripts/train.py \
     --policy.type=act \
@@ -74,7 +74,7 @@ python lerobot/scripts/train.py \
 > Notice we added `--output_dir` to explicitly tell where to write outputs from this run (checkpoints, training state, configs etc.). This is not mandatory and if you don't specify it, a default directory will be created from the current date and time, env.type and policy.type. This will typically look like `outputs/train/2025-01-24/16-10-05_aloha_act`.
 
 We now want to train a different policy for aloha on another task. We'll change the dataset and use [lerobot/aloha_sim_transfer_cube_human](https://huggingface.co/datasets/lerobot/aloha_sim_transfer_cube_human) instead. Of course, we also need to change the task of the environment as well to match this other task.
-Looking at the [`AlohaEnv`](../lerobot/common/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:
+Looking at the [`AlohaEnv`](../../lerobot/common/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 lerobot/scripts/train.py \
     --policy.type=act \

examples/7_get_started_with_real_robot.md

@@ -46,6 +46,13 @@ Using `uv`:
 uv sync --extra "dynamixel"
 ```
 
+/!\ For Linux only, ffmpeg and opencv requires conda install for now. Run this exact sequence of commands:
+```bash
+conda install -c conda-forge ffmpeg
+pip uninstall opencv-python
+conda install -c conda-forge "opencv>=4.10.0"
+```
+
 You are now ready to plug the 5V power supply to the motor bus of the leader arm (the smaller one) since all its motors only require 5V.
 
 Then plug the 12V power supply to the motor bus of the follower arm. It has two motors that need 12V, and the rest will be powered with 5V through the voltage convertor.
@@ -55,9 +62,6 @@ Finally, connect both arms to your computer via USB. Note that the USB doesn't p
 Now you are ready to configure your motors for the first time, as detailed in the sections below. In the upcoming sections, you'll learn about our classes and functions by running some python code in an interactive session, or by copy-pasting it in a python file.
 
 If you have already configured your motors the first time, you can streamline the process by directly running the teleoperate script (which is detailed further in the tutorial):
-
-> **NOTE:** To visualize the data, enable `--control.display_data=true`. This streams the data using `rerun`.
-
 ```bash
 python lerobot/scripts/control_robot.py \
   --robot.type=koch \
@@ -288,11 +292,6 @@ Steps:
    - Scan for devices. All 12 motors should appear.
    - Select the motors one by one and move the arm. Check that the graphical indicator near the top right shows the movement.
 
-** There is a common issue with the Dynamixel XL430-W250 motors where the motors become undiscoverable after upgrading their firmware from Mac and Windows Dynamixel Wizard2 applications.  When this occurs, it is required to do a firmware recovery (Select `DYNAMIXEL Firmware Recovery` and follow the prompts).   There are two known workarounds to conduct this firmware reset:
-  1) Install the Dynamixel Wizard on a linux machine and complete the firmware recovery
-  2) Use the Dynamixel U2D2 in order to perform the reset with Windows or Mac.  This U2D2 can be purchased [here](https://www.robotis.us/u2d2/).
-  For either solution, open DYNAMIXEL Wizard 2.0 and select the appropriate port. You will likely be unable to see the motor in the GUI at this time. Select `Firmware Recovery`, carefully choose the correct model, and wait for the process to complete. Finally, re-scan to confirm the firmware recovery was successful.
-
 **Read and Write with DynamixelMotorsBus**
 
 To get familiar with how `DynamixelMotorsBus` communicates with the motors, you can start by reading data from them. Copy past this code in the same interactive python session:
@@ -830,6 +829,16 @@ It contains:
 - `dtRphone:33.84 (29.5hz)` which is the delta time of capturing an image from the phone camera in the thread running asynchronously.
 
 Troubleshooting:
+- On Linux, if you encounter a hanging issue when using cameras, uninstall opencv and re-install it with conda:
+```bash
+pip uninstall opencv-python
+conda install -c conda-forge opencv=4.10.0
+```
+- On Linux, if you encounter any issue during video encoding with `ffmpeg: unknown encoder libsvtav1`, you can:
+  - install with conda-forge by running `conda install -c conda-forge ffmpeg` (it should be compiled with `libsvtav1`),
+  - or, install [Homebrew](https://brew.sh) and run `brew install ffmpeg` (it should be compiled with `libsvtav1`),
+  - or, 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`.
 - On Linux, if the left and right arrow keys and escape key don't have any effect during data recording, make sure you've set the `$DISPLAY` environment variable. See [pynput limitations](https://pynput.readthedocs.io/en/latest/limitations.html#linux).
 
 At the end of data recording, your dataset will be uploaded on your Hugging Face page (e.g. https://huggingface.co/datasets/cadene/koch_test) that you can obtain by running:

examples/advanced/1_add_image_transforms.py

@@ -31,7 +31,7 @@ dataset = LeRobotDataset(dataset_repo_id, episodes=[0])
 # This is equivalent to `dataset = LeRobotDataset(dataset_repo_id, image_transforms=None)`
 
 # Get the index of the first observation in the first episode
-first_idx = dataset.meta.episodes["dataset_from_index"][0]
+first_idx = dataset.episode_data_index["from"][0].item()
 
 # Get the frame corresponding to the first camera
 frame = dataset[first_idx][dataset.meta.camera_keys[0]]

examples/port_datasets/droid_rlds/README.md

@@ -1,144 +0,0 @@
-# Port DROID 1.0.1 dataset to LeRobotDataset
-
-## Download
-
-TODO
-
-It will take 2 TB in your local disk.
-
-## Port on a single computer
-
-First, install tensorflow dataset utilities to read from raw files:
-```bash
-pip install tensorflow
-pip install tensorflow_datasets
-```
-
-Then run this script to start porting the dataset:
-```bash
-python examples/port_datasets/droid_rlds/port_droid.py \
-    --raw-dir /your/data/droid/1.0.1 \
-    --repo-id your_id/droid_1.0.1 \
-    --push-to-hub
-```
-
-It will take 400GB in your local disk.
-
-As usual, your LeRobotDataset will be stored in your huggingface/lerobot cache folder.
-
-WARNING: it will take 7 days for porting the dataset locally and 3 days to upload, so we will need to parallelize over multiple nodes on a slurm cluster.
-
-NOTE: For development, run this script to start porting a shard:
-```bash
-python examples/port_datasets/droid_rlds/port.py \
-    --raw-dir /your/data/droid/1.0.1 \
-    --repo-id your_id/droid_1.0.1 \
-    --num-shards 2048 \
-    --shard-index 0
-```
-
-## Port over SLURM
-
-Install slurm utilities from Hugging Face:
-```bash
-pip install datatrove
-```
-
-
-### 1. Port one shard per job
-
-Run this script to start porting shards of the dataset:
-```bash
-python examples/port_datasets/droid_rlds/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
-```
-
-**Note on how to set your command line arguments**
-
-Regarding `--partition`, find yours by running:
-```bash
-info --format="%R"`
-```
-and select the CPU partition if you have one. No GPU needed.
-
-Regarding `--workers`, it is the number of slurm jobs you will launch in parallel. 2048 is the maximum number, since there is 2048 shards in Droid. This big number will certainly max-out your cluster.
-
-Regarding `--cpus-per-task` and `--mem-per-cpu`, by default it will use ~16GB of RAM (8*1950M) which is recommended to load the raw frames and 8 CPUs which can be useful to parallelize the encoding of the frames.
-
-Find the number of CPUs and Memory of the nodes of your partition by running:
-```bash
-sinfo -N -p your_partition -h -o "%N cpus=%c mem=%m"
-```
-
-**Useful commands to check progress and debug**
-
-Check if your jobs are running:
-```bash
-squeue -u $USER`
-```
-
-You should see a list with job indices like `15125385_155` where `15125385` is the index of the run and `155` is the worker index. The output/print of this worker is written in real time in `/your/logs/job_name/slurm_jobs/15125385_155.out`. For instance, you can inspect the content of this file by running `less /your/logs/job_name/slurm_jobs/15125385_155.out`.
-
-Check the progression of your jobs by running:
-```bash
-jobs_status /your/logs
-```
-
-If it's not 100% and no more slurm job is running, it means that some of them failed. Inspect the logs by running:
-```bash
-failed_logs /your/logs/job_name
-```
-
-If there is an issue in the code, you can fix it in debug mode with `--slurm 0` which allows to set breakpoint:
-```bash
-python examples/port_datasets/droid_rlds/slurm_port_shards.py --slurm 0 ...
-```
-
-And you can relaunch the same command, which will skip the completed jobs:
-```bash
-python examples/port_datasets/droid_rlds/slurm_port_shards.py --slurm 1 ...
-```
-
-Once all jobs are completed, you will have one dataset per shard (e.g. `droid_1.0.1_world_2048_rank_1594`) saved on disk in your `/lerobot/home/dir/your_id` directory. You can find your `/lerobot/home/dir` by running:
-```bash
-python -c "from lerobot.common.constants import HF_LEROBOT_HOME;print(HF_LEROBOT_HOME)"
-```
-
-
-### 2. Aggregate all shards
-
-Run this script to start aggregation:
-```bash
-python examples/port_datasets/droid_rlds/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
-```
-
-Once all jobs are completed, you will have one dataset your `/lerobot/home/dir/your_id/droid_1.0.1` directory.
-
-
-### 3. Upload dataset
-
-Run this script to start uploading:
-```bash
-python examples/port_datasets/droid_rlds/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
-```

examples/port_datasets/droid_rlds/port_droid.py

@@ -1,430 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-import argparse
-import logging
-import time
-from pathlib import Path
-
-import numpy as np
-import tensorflow_datasets as tfds
-
-from lerobot.common.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
-from lerobot.common.utils.utils import get_elapsed_time_in_days_hours_minutes_seconds
-
-DROID_SHARDS = 2048
-DROID_FPS = 15
-DROID_ROBOT_TYPE = "Franka"
-
-# Dataset schema slightly adapted from: https://droid-dataset.github.io/droid/the-droid-dataset.html#-dataset-schema
-DROID_FEATURES = {
-    # true on first step of the episode
-    "is_first": {
-        "dtype": "bool",
-        "shape": (1,),
-        "names": None,
-    },
-    # true on last step of the episode
-    "is_last": {
-        "dtype": "bool",
-        "shape": (1,),
-        "names": None,
-    },
-    # true on last step of the episode if it is a terminal step, True for demos
-    "is_terminal": {
-        "dtype": "bool",
-        "shape": (1,),
-        "names": None,
-    },
-    # language_instruction is also stored as "task" to follow LeRobot standard
-    "language_instruction": {
-        "dtype": "string",
-        "shape": (1,),
-        "names": None,
-    },
-    "language_instruction_2": {
-        "dtype": "string",
-        "shape": (1,),
-        "names": None,
-    },
-    "language_instruction_3": {
-        "dtype": "string",
-        "shape": (1,),
-        "names": None,
-    },
-    "observation.state.gripper_position": {
-        "dtype": "float32",
-        "shape": (1,),
-        "names": {
-            "axes": ["gripper"],
-        },
-    },
-    "observation.state.cartesian_position": {
-        "dtype": "float32",
-        "shape": (6,),
-        "names": {
-            "axes": ["x", "y", "z", "roll", "pitch", "yaw"],
-        },
-    },
-    "observation.state.joint_position": {
-        "dtype": "float32",
-        "shape": (7,),
-        "names": {
-            "axes": ["joint_0", "joint_1", "joint_2", "joint_3", "joint_4", "joint_5", "joint_6"],
-        },
-    },
-    # Add this new feature to follow LeRobot standard of using joint position + gripper
-    "observation.state": {
-        "dtype": "float32",
-        "shape": (8,),
-        "names": {
-            "axes": ["joint_0", "joint_1", "joint_2", "joint_3", "joint_4", "joint_5", "joint_6", "gripper"],
-        },
-    },
-    # Initially called wrist_image_left
-    "observation.images.wrist_left": {
-        "dtype": "video",
-        "shape": (180, 320, 3),
-        "names": [
-            "height",
-            "width",
-            "channels",
-        ],
-    },
-    # Initially called exterior_image_1_left
-    "observation.images.exterior_1_left": {
-        "dtype": "video",
-        "shape": (180, 320, 3),
-        "names": [
-            "height",
-            "width",
-            "channels",
-        ],
-    },
-    # Initially called exterior_image_2_left
-    "observation.images.exterior_2_left": {
-        "dtype": "video",
-        "shape": (180, 320, 3),
-        "names": [
-            "height",
-            "width",
-            "channels",
-        ],
-    },
-    "action.gripper_position": {
-        "dtype": "float32",
-        "shape": (1,),
-        "names": {
-            "axes": ["gripper"],
-        },
-    },
-    "action.gripper_velocity": {
-        "dtype": "float32",
-        "shape": (1,),
-        "names": {
-            "axes": ["gripper"],
-        },
-    },
-    "action.cartesian_position": {
-        "dtype": "float32",
-        "shape": (6,),
-        "names": {
-            "axes": ["x", "y", "z", "roll", "pitch", "yaw"],
-        },
-    },
-    "action.cartesian_velocity": {
-        "dtype": "float32",
-        "shape": (6,),
-        "names": {
-            "axes": ["x", "y", "z", "roll", "pitch", "yaw"],
-        },
-    },
-    "action.joint_position": {
-        "dtype": "float32",
-        "shape": (7,),
-        "names": {
-            "axes": ["joint_0", "joint_1", "joint_2", "joint_3", "joint_4", "joint_5", "joint_6"],
-        },
-    },
-    "action.joint_velocity": {
-        "dtype": "float32",
-        "shape": (7,),
-        "names": {
-            "axes": ["joint_0", "joint_1", "joint_2", "joint_3", "joint_4", "joint_5", "joint_6"],
-        },
-    },
-    # This feature was called "action" in RLDS dataset and consists of [6x joint velocities, 1x gripper position]
-    "action.original": {
-        "dtype": "float32",
-        "shape": (7,),
-        "names": {
-            "axes": ["x", "y", "z", "roll", "pitch", "yaw", "gripper"],
-        },
-    },
-    # Add this new feature to follow LeRobot standard of using joint position + gripper
-    "action": {
-        "dtype": "float32",
-        "shape": (8,),
-        "names": {
-            "axes": ["joint_0", "joint_1", "joint_2", "joint_3", "joint_4", "joint_5", "joint_6", "gripper"],
-        },
-    },
-    "discount": {
-        "dtype": "float32",
-        "shape": (1,),
-        "names": None,
-    },
-    "reward": {
-        "dtype": "float32",
-        "shape": (1,),
-        "names": None,
-    },
-    # Meta data that are the same for all frames in the episode
-    "task_category": {
-        "dtype": "string",
-        "shape": (1,),
-        "names": None,
-    },
-    "building": {
-        "dtype": "string",
-        "shape": (1,),
-        "names": None,
-    },
-    "collector_id": {
-        "dtype": "string",
-        "shape": (1,),
-        "names": None,
-    },
-    "date": {
-        "dtype": "string",
-        "shape": (1,),
-        "names": None,
-    },
-    "camera_extrinsics.wrist_left": {
-        "dtype": "float32",
-        "shape": (6,),
-        "names": {
-            "axes": ["x", "y", "z", "roll", "pitch", "yaw"],
-        },
-    },
-    "camera_extrinsics.exterior_1_left": {
-        "dtype": "float32",
-        "shape": (6,),
-        "names": {
-            "axes": ["x", "y", "z", "roll", "pitch", "yaw"],
-        },
-    },
-    "camera_extrinsics.exterior_2_left": {
-        "dtype": "float32",
-        "shape": (6,),
-        "names": {
-            "axes": ["x", "y", "z", "roll", "pitch", "yaw"],
-        },
-    },
-    "is_episode_successful": {
-        "dtype": "bool",
-        "shape": (1,),
-        "names": None,
-    },
-}
-
-
-def is_episode_successful(tf_episode_metadata):
-    # Adapted from: https://github.com/droid-dataset/droid_policy_learning/blob/dd1020eb20d981f90b5ff07dc80d80d5c0cb108b/robomimic/utils/rlds_utils.py#L8
-    return "/success/" in tf_episode_metadata["file_path"].numpy().decode()
-
-
-def generate_lerobot_frames(tf_episode):
-    m = tf_episode["episode_metadata"]
-    frame_meta = {
-        "task_category": m["building"].numpy().decode(),
-        "building": m["building"].numpy().decode(),
-        "collector_id": m["collector_id"].numpy().decode(),
-        "date": m["date"].numpy().decode(),
-        "camera_extrinsics.wrist_left": m["extrinsics_wrist_cam"].numpy(),
-        "camera_extrinsics.exterior_1_left": m["extrinsics_exterior_cam_1"].numpy(),
-        "camera_extrinsics.exterior_2_left": m["extrinsics_exterior_cam_2"].numpy(),
-        "is_episode_successful": np.array([is_episode_successful(m)]),
-    }
-    for f in tf_episode["steps"]:
-        # Dataset schema slightly adapted from: https://droid-dataset.github.io/droid/the-droid-dataset.html#-dataset-schema
-        frame = {
-            "is_first": np.array([f["is_first"].numpy()]),
-            "is_last": np.array([f["is_last"].numpy()]),
-            "is_terminal": np.array([f["is_terminal"].numpy()]),
-            "language_instruction": f["language_instruction"].numpy().decode(),
-            "language_instruction_2": f["language_instruction_2"].numpy().decode(),
-            "language_instruction_3": f["language_instruction_3"].numpy().decode(),
-            "observation.state.gripper_position": f["observation"]["gripper_position"].numpy(),
-            "observation.state.cartesian_position": f["observation"]["cartesian_position"].numpy(),
-            "observation.state.joint_position": f["observation"]["joint_position"].numpy(),
-            "observation.images.wrist_left": f["observation"]["wrist_image_left"].numpy(),
-            "observation.images.exterior_1_left": f["observation"]["exterior_image_1_left"].numpy(),
-            "observation.images.exterior_2_left": f["observation"]["exterior_image_2_left"].numpy(),
-            "action.gripper_position": f["action_dict"]["gripper_position"].numpy(),
-            "action.gripper_velocity": f["action_dict"]["gripper_velocity"].numpy(),
-            "action.cartesian_position": f["action_dict"]["cartesian_position"].numpy(),
-            "action.cartesian_velocity": f["action_dict"]["cartesian_velocity"].numpy(),
-            "action.joint_position": f["action_dict"]["joint_position"].numpy(),
-            "action.joint_velocity": f["action_dict"]["joint_velocity"].numpy(),
-            "discount": np.array([f["discount"].numpy()]),
-            "reward": np.array([f["reward"].numpy()]),
-            "action.original": f["action"].numpy(),
-        }
-
-        # language_instruction is also stored as "task" to follow LeRobot standard
-        frame["task"] = frame["language_instruction"]
-
-        # Add this new feature to follow LeRobot standard of using joint position + gripper
-        frame["observation.state"] = np.concatenate(
-            [frame["observation.state.joint_position"], frame["observation.state.gripper_position"]]
-        )
-        frame["action"] = np.concatenate([frame["action.joint_position"], frame["action.gripper_position"]])
-
-        # Meta data that are the same for all frames in the episode
-        frame.update(frame_meta)
-
-        # Cast fp64 to fp32
-        for key in frame:
-            if isinstance(frame[key], np.ndarray) and frame[key].dtype == np.float64:
-                frame[key] = frame[key].astype(np.float32)
-
-        yield frame
-
-
-def port_droid(
-    raw_dir: Path,
-    repo_id: str,
-    push_to_hub: bool = False,
-    num_shards: int | None = None,
-    shard_index: int | None = None,
-):
-    dataset_name = raw_dir.parent.name
-    version = raw_dir.name
-    data_dir = raw_dir.parent.parent
-
-    builder = tfds.builder(f"{dataset_name}/{version}", data_dir=data_dir, version="")
-
-    if num_shards is not None:
-        tfds_num_shards = builder.info.splits["train"].num_shards
-        if tfds_num_shards != DROID_SHARDS:
-            raise ValueError(
-                f"Number of shards of Droid dataset is expected to be {DROID_SHARDS} but is {tfds_num_shards}."
-            )
-        if num_shards != tfds_num_shards:
-            raise ValueError(
-                f"We only shard over the fixed number of shards provided by tensorflow dataset ({tfds_num_shards}), but {num_shards} shards provided instead."
-            )
-        if shard_index >= tfds_num_shards:
-            raise ValueError(
-                f"Shard index is greater than the num of shards ({shard_index} >= {num_shards})."
-            )
-
-        raw_dataset = builder.as_dataset(split=f"train[{shard_index}shard]")
-    else:
-        raw_dataset = builder.as_dataset(split="train")
-
-    lerobot_dataset = LeRobotDataset.create(
-        repo_id=repo_id,
-        robot_type=DROID_ROBOT_TYPE,
-        fps=DROID_FPS,
-        features=DROID_FEATURES,
-    )
-
-    start_time = time.time()
-    num_episodes = raw_dataset.cardinality().numpy().item()
-    logging.info(f"Number of episodes {num_episodes}")
-
-    for episode_index, episode in enumerate(raw_dataset):
-        elapsed_time = time.time() - start_time
-        d, h, m, s = get_elapsed_time_in_days_hours_minutes_seconds(elapsed_time)
-
-        logging.info(
-            f"{episode_index} / {num_episodes} episodes processed (after {d} days, {h} hours, {m} minutes, {s:.3f} seconds)"
-        )
-
-        for frame in generate_lerobot_frames(episode):
-            lerobot_dataset.add_frame(frame)
-
-        lerobot_dataset.save_episode()
-        logging.info("Save_episode")
-
-    if push_to_hub:
-        lerobot_dataset.push_to_hub(
-            # Add openx tag, since it belongs to the openx collection of datasets
-            tags=["openx"],
-            private=False,
-        )
-
-
-def validate_dataset(repo_id):
-    """Sanity check that ensure meta data can be loaded and all files are present."""
-    meta = LeRobotDatasetMetadata(repo_id)
-
-    if meta.total_episodes == 0:
-        raise ValueError("Number of episodes is 0.")
-
-    for ep_idx in range(meta.total_episodes):
-        data_path = meta.root / meta.get_data_file_path(ep_idx)
-
-        if not data_path.exists():
-            raise ValueError(f"Parquet file is missing in: {data_path}")
-
-        for vid_key in meta.video_keys:
-            vid_path = meta.root / meta.get_video_file_path(ep_idx, vid_key)
-            if not vid_path.exists():
-                raise ValueError(f"Video file is missing in: {vid_path}")
-
-
-def main():
-    parser = argparse.ArgumentParser()
-
-    parser.add_argument(
-        "--raw-dir",
-        type=Path,
-        required=True,
-        help="Directory containing input raw datasets (e.g. `path/to/dataset` or `path/to/dataset/version).",
-    )
-    parser.add_argument(
-        "--repo-id",
-        type=str,
-        help="Repositery identifier on Hugging Face: a community or a user name `/` the name of the dataset, required when push-to-hub is True",
-    )
-    parser.add_argument(
-        "--push-to-hub",
-        action="store_true",
-        help="Upload to hub.",
-    )
-    parser.add_argument(
-        "--num-shards",
-        type=int,
-        default=None,
-        help="Number of shards. Can be either None to load the full dataset, or 2048 to load one of the 2048 tensorflow dataset files.",
-    )
-    parser.add_argument(
-        "--shard-index",
-        type=int,
-        default=None,
-        help="Index of the shard. Can be either None to load the full dataset, or in [0,2047] to load one of the 2048 tensorflow dataset files.",
-    )
-
-    args = parser.parse_args()
-
-    port_droid(**vars(args))
-
-
-if __name__ == "__main__":
-    main()

examples/port_datasets/droid_rlds/slurm_aggregate_shards.py

@@ -1,293 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-import argparse
-import logging
-from pathlib import Path
-
-import tqdm
-from datatrove.executor import LocalPipelineExecutor
-from datatrove.executor.slurm import SlurmPipelineExecutor
-from datatrove.pipeline.base import PipelineStep
-
-from examples.port_datasets.droid_rlds.port_droid import DROID_SHARDS
-from lerobot.common.datasets.aggregate import validate_all_metadata
-from lerobot.common.datasets.lerobot_dataset import LeRobotDatasetMetadata
-from lerobot.common.datasets.utils import (
-    legacy_write_episode_stats,
-    legacy_write_task,
-    write_episode,
-    write_info,
-)
-from lerobot.common.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
-
-        self.create_aggr_dataset()
-
-    def create_aggr_dataset(self):
-        init_logging()
-
-        logging.info("Start aggregate_datasets")
-
-        all_metadata = [LeRobotDatasetMetadata(repo_id) for repo_id in self.repo_ids]
-
-        fps, robot_type, features = validate_all_metadata(all_metadata)
-
-        # Create resulting dataset folder
-        aggr_meta = LeRobotDatasetMetadata.create(
-            repo_id=self.aggr_repo_id,
-            fps=fps,
-            robot_type=robot_type,
-            features=features,
-        )
-
-        logging.info("Find all tasks")
-        # find all tasks, deduplicate them, create new task indices for each dataset
-        # indexed by dataset index
-        datasets_task_index_to_aggr_task_index = {}
-        aggr_task_index = 0
-        for dataset_index, meta in enumerate(tqdm.tqdm(all_metadata, desc="Find all tasks")):
-            task_index_to_aggr_task_index = {}
-
-            for task_index, task in meta.tasks.items():
-                if task not in aggr_meta.task_to_task_index:
-                    # add the task to aggr tasks mappings
-                    aggr_meta.tasks[aggr_task_index] = task
-                    aggr_meta.task_to_task_index[task] = aggr_task_index
-                    aggr_task_index += 1
-
-                # add task_index anyway
-                task_index_to_aggr_task_index[task_index] = aggr_meta.task_to_task_index[task]
-
-            datasets_task_index_to_aggr_task_index[dataset_index] = task_index_to_aggr_task_index
-
-        logging.info("Prepare copy data and videos")
-        datasets_ep_idx_to_aggr_ep_idx = {}
-        datasets_aggr_episode_index_shift = {}
-        aggr_episode_index_shift = 0
-        for dataset_index, meta in enumerate(tqdm.tqdm(all_metadata, desc="Prepare copy data and videos")):
-            ep_idx_to_aggr_ep_idx = {}
-
-            for episode_index in range(meta.total_episodes):
-                aggr_episode_index = episode_index + aggr_episode_index_shift
-                ep_idx_to_aggr_ep_idx[episode_index] = aggr_episode_index
-
-            datasets_ep_idx_to_aggr_ep_idx[dataset_index] = ep_idx_to_aggr_ep_idx
-            datasets_aggr_episode_index_shift[dataset_index] = aggr_episode_index_shift
-
-            # populate episodes
-            for episode_index, episode_dict in meta.episodes.items():
-                aggr_episode_index = episode_index + aggr_episode_index_shift
-                episode_dict["episode_index"] = aggr_episode_index
-                aggr_meta.episodes[aggr_episode_index] = episode_dict
-
-            # populate episodes_stats
-            for episode_index, episode_stats in meta.episodes_stats.items():
-                aggr_episode_index = episode_index + aggr_episode_index_shift
-                aggr_meta.episodes_stats[aggr_episode_index] = episode_stats
-
-            # populate info
-            aggr_meta.info["total_episodes"] += meta.total_episodes
-            aggr_meta.info["total_frames"] += meta.total_frames
-            aggr_meta.info["total_videos"] += len(aggr_meta.video_keys) * meta.total_episodes
-
-            aggr_episode_index_shift += meta.total_episodes
-
-        logging.info("Write meta data")
-        aggr_meta.info["total_tasks"] = len(aggr_meta.tasks)
-        aggr_meta.info["total_chunks"] = aggr_meta.get_episode_chunk(aggr_episode_index_shift - 1)
-        aggr_meta.info["splits"] = {"train": f"0:{aggr_meta.info['total_episodes']}"}
-
-        # create a new episodes jsonl with updated episode_index using write_episode
-        for episode_dict in tqdm.tqdm(aggr_meta.episodes.values(), desc="Write episodes"):
-            write_episode(episode_dict, aggr_meta.root)
-
-        # create a new episode_stats jsonl with updated episode_index using write_episode_stats
-        for episode_index, episode_stats in tqdm.tqdm(
-            aggr_meta.episodes_stats.items(), desc="Write episodes stats"
-        ):
-            legacy_write_episode_stats(episode_index, episode_stats, aggr_meta.root)
-
-        # create a new task jsonl with updated episode_index using write_task
-        for task_index, task in tqdm.tqdm(aggr_meta.tasks.items(), desc="Write tasks"):
-            legacy_write_task(task_index, task, aggr_meta.root)
-
-        write_info(aggr_meta.info, aggr_meta.root)
-
-        self.datasets_task_index_to_aggr_task_index = datasets_task_index_to_aggr_task_index
-        self.datasets_ep_idx_to_aggr_ep_idx = datasets_ep_idx_to_aggr_ep_idx
-        self.datasets_aggr_episode_index_shift = datasets_aggr_episode_index_shift
-
-        logging.info("Meta data done writing!")
-
-    def run(self, data=None, rank: int = 0, world_size: int = 1):
-        import logging
-        import shutil
-
-        import pandas as pd
-
-        from lerobot.common.datasets.aggregate import get_update_episode_and_task_func
-        from lerobot.common.datasets.lerobot_dataset import LeRobotDatasetMetadata
-        from lerobot.common.utils.utils import init_logging
-
-        init_logging()
-
-        aggr_meta = LeRobotDatasetMetadata(self.aggr_repo_id)
-        all_metadata = [LeRobotDatasetMetadata(repo_id) for repo_id in self.repo_ids]
-
-        if world_size != len(all_metadata):
-            raise ValueError()
-
-        dataset_index = rank
-        meta = all_metadata[dataset_index]
-        aggr_episode_index_shift = self.datasets_aggr_episode_index_shift[dataset_index]
-
-        logging.info("Copy data")
-        for episode_index in range(meta.total_episodes):
-            aggr_episode_index = self.datasets_ep_idx_to_aggr_ep_idx[dataset_index][episode_index]
-            data_path = meta.root / meta.get_data_file_path(episode_index)
-            aggr_data_path = aggr_meta.root / aggr_meta.get_data_file_path(aggr_episode_index)
-
-            # update episode_index and task_index
-            df = pd.read_parquet(data_path)
-            update_row_func = get_update_episode_and_task_func(
-                aggr_episode_index_shift, self.datasets_task_index_to_aggr_task_index[dataset_index]
-            )
-            df = df.apply(update_row_func, axis=1)
-
-            aggr_data_path.parent.mkdir(parents=True, exist_ok=True)
-            df.to_parquet(aggr_data_path)
-
-        logging.info("Copy videos")
-        for episode_index in range(meta.total_episodes):
-            aggr_episode_index = episode_index + aggr_episode_index_shift
-            for vid_key in meta.video_keys:
-                video_path = meta.root / meta.get_video_file_path(episode_index, vid_key)
-                aggr_video_path = aggr_meta.root / aggr_meta.get_video_file_path(aggr_episode_index, vid_key)
-                aggr_video_path.parent.mkdir(parents=True, exist_ok=True)
-                shutil.copy(video_path, aggr_video_path)
-
-                # copy_command = f"cp {video_path} {aggr_video_path} &"
-                # subprocess.Popen(copy_command, shell=True)
-
-        logging.info("Done!")
-
-
-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:
-        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="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=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
-
-    repo_ids = [f"{args.repo_id}_world_{DROID_SHARDS}_rank_{rank}" for rank in range(DROID_SHARDS)]
-    aggregate_executor = make_aggregate_executor(repo_ids, **kwargs)
-    aggregate_executor.run()
-
-
-if __name__ == "__main__":
-    main()

examples/port_datasets/droid_rlds/slurm_port_shards.py

@@ -1,147 +0,0 @@
-import argparse
-from pathlib import Path
-
-from datatrove.executor import LocalPipelineExecutor
-from datatrove.executor.slurm import SlurmPipelineExecutor
-from datatrove.pipeline.base import PipelineStep
-
-from examples.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 examples.port_datasets.droid_rlds.port_droid import port_droid, validate_dataset
-        from lerobot.common.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:
-            pass
-
-        port_droid(
-            self.raw_dir,
-            shard_repo_id,
-            push_to_hub=False,
-            num_shards=world_size,
-            shard_index=rank,
-        )
-
-        validate_dataset(shard_repo_id)
-
-
-def make_port_executor(
-    raw_dir, repo_id, job_name, logs_dir, workers, partition, cpus_per_task, mem_per_cpu, slurm=True
-):
-    kwargs = {
-        "pipeline": [
-            PortDroidShards(raw_dir, repo_id),
-        ],
-        "logging_dir": str(logs_dir / job_name),
-    }
-
-    if slurm:
-        kwargs.update(
-            {
-                "job_name": job_name,
-                "tasks": DROID_SHARDS,
-                "workers": workers,
-                "time": "08:00:00",
-                "partition": partition,
-                "cpus_per_task": cpus_per_task,
-                "sbatch_args": {"mem-per-cpu": mem_per_cpu},
-            }
-        )
-        executor = SlurmPipelineExecutor(**kwargs)
-    else:
-        kwargs.update(
-            {
-                "tasks": 1,
-                "workers": 1,
-            }
-        )
-        executor = LocalPipelineExecutor(**kwargs)
-
-    return executor
-
-
-def main():
-    parser = argparse.ArgumentParser()
-
-    parser.add_argument(
-        "--raw-dir",
-        type=Path,
-        required=True,
-        help="Directory containing input raw datasets (e.g. `path/to/dataset` or `path/to/dataset/version).",
-    )
-    parser.add_argument(
-        "--repo-id",
-        type=str,
-        help="Repositery identifier on Hugging Face: a community or a user name `/` the name of the dataset, required when push-to-hub is True.",
-    )
-    parser.add_argument(
-        "--logs-dir",
-        type=Path,
-        help="Path to logs directory for `datatrove`.",
-    )
-    parser.add_argument(
-        "--job-name",
-        type=str,
-        default="port_droid",
-        help="Job name used in slurm, and name of the directory created inside the provided logs directory.",
-    )
-    parser.add_argument(
-        "--slurm",
-        type=int,
-        default=1,
-        help="Launch over slurm. Use `--slurm 0` to launch sequentially (useful to debug).",
-    )
-    parser.add_argument(
-        "--workers",
-        type=int,
-        default=2048,
-        help="Number of slurm workers. It should be less than the maximum number of shards.",
-    )
-    parser.add_argument(
-        "--partition",
-        type=str,
-        help="Slurm partition. Ideally a CPU partition. No need for GPU partition.",
-    )
-    parser.add_argument(
-        "--cpus-per-task",
-        type=int,
-        default=8,
-        help="Number of cpus that each slurm worker will use.",
-    )
-    parser.add_argument(
-        "--mem-per-cpu",
-        type=str,
-        default="1950M",
-        help="Memory per cpu that each worker will use.",
-    )
-
-    args = parser.parse_args()
-    kwargs = vars(args)
-    kwargs["slurm"] = kwargs.pop("slurm") == 1
-    port_executor = make_port_executor(**kwargs)
-    port_executor.run()
-
-
-if __name__ == "__main__":
-    main()

examples/port_datasets/droid_rlds/slurm_upload.py

@@ -1,263 +0,0 @@
-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 examples.port_datasets.droid_rlds.port_droid import DROID_SHARDS
-from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDatasetMetadata
-from lerobot.common.datasets.utils import create_lerobot_dataset_card
-from lerobot.common.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)
-
-        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.common.datasets.lerobot_dataset import LeRobotDatasetMetadata
-        from lerobot.common.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()

examples/port_datasets/pusht_zarr.py

@@ -0,0 +1,243 @@
+# 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 shutil
+from pathlib import Path
+
+import numpy as np
+from huggingface_hub import HfApi
+
+from lerobot.common.constants import HF_LEROBOT_HOME
+from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDataset
+from lerobot.common.datasets.push_dataset_to_hub._download_raw import download_raw
+
+PUSHT_TASK = "Push the T-shaped blue block onto the T-shaped green target surface."
+PUSHT_FEATURES = {
+    "observation.state": {
+        "dtype": "float32",
+        "shape": (2,),
+        "names": {
+            "axes": ["x", "y"],
+        },
+    },
+    "action": {
+        "dtype": "float32",
+        "shape": (2,),
+        "names": {
+            "axes": ["x", "y"],
+        },
+    },
+    "next.reward": {
+        "dtype": "float32",
+        "shape": (1,),
+        "names": None,
+    },
+    "next.success": {
+        "dtype": "bool",
+        "shape": (1,),
+        "names": None,
+    },
+    "observation.environment_state": {
+        "dtype": "float32",
+        "shape": (16,),
+        "names": [
+            "keypoints",
+        ],
+    },
+    "observation.image": {
+        "dtype": None,
+        "shape": (3, 96, 96),
+        "names": [
+            "channels",
+            "height",
+            "width",
+        ],
+    },
+}
+
+
+def build_features(mode: str) -> dict:
+    features = PUSHT_FEATURES
+    if mode == "keypoints":
+        features.pop("observation.image")
+    else:
+        features.pop("observation.environment_state")
+        features["observation.image"]["dtype"] = mode
+
+    return features
+
+
+def load_raw_dataset(zarr_path: Path):
+    try:
+        from lerobot.common.datasets.push_dataset_to_hub._diffusion_policy_replay_buffer import (
+            ReplayBuffer as DiffusionPolicyReplayBuffer,
+        )
+    except ModuleNotFoundError as e:
+        print("`gym_pusht` is not installed. Please install it with `pip install 'lerobot[gym_pusht]'`")
+        raise e
+
+    zarr_data = DiffusionPolicyReplayBuffer.copy_from_path(zarr_path)
+    return zarr_data
+
+
+def calculate_coverage(zarr_data):
+    try:
+        import pymunk
+        from gym_pusht.envs.pusht import PushTEnv, pymunk_to_shapely
+    except ModuleNotFoundError as e:
+        print("`gym_pusht` is not installed. Please install it with `pip install 'lerobot[gym_pusht]'`")
+        raise e
+
+    block_pos = zarr_data["state"][:, 2:4]
+    block_angle = zarr_data["state"][:, 4]
+
+    num_frames = len(block_pos)
+
+    coverage = np.zeros((num_frames,), dtype=np.float32)
+    # 8 keypoints with 2 coords each
+    keypoints = np.zeros((num_frames, 16), dtype=np.float32)
+
+    # Set x, y, theta (in radians)
+    goal_pos_angle = np.array([256, 256, np.pi / 4])
+    goal_body = PushTEnv.get_goal_pose_body(goal_pos_angle)
+
+    for i in range(num_frames):
+        space = pymunk.Space()
+        space.gravity = 0, 0
+        space.damping = 0
+
+        # Add walls.
+        walls = [
+            PushTEnv.add_segment(space, (5, 506), (5, 5), 2),
+            PushTEnv.add_segment(space, (5, 5), (506, 5), 2),
+            PushTEnv.add_segment(space, (506, 5), (506, 506), 2),
+            PushTEnv.add_segment(space, (5, 506), (506, 506), 2),
+        ]
+        space.add(*walls)
+
+        block_body, block_shapes = PushTEnv.add_tee(space, block_pos[i].tolist(), block_angle[i].item())
+        goal_geom = pymunk_to_shapely(goal_body, block_body.shapes)
+        block_geom = pymunk_to_shapely(block_body, block_body.shapes)
+        intersection_area = goal_geom.intersection(block_geom).area
+        goal_area = goal_geom.area
+        coverage[i] = intersection_area / goal_area
+        keypoints[i] = PushTEnv.get_keypoints(block_shapes).flatten()
+
+    return coverage, keypoints
+
+
+def calculate_success(coverage: float, success_threshold: float):
+    return coverage > success_threshold
+
+
+def calculate_reward(coverage: float, success_threshold: float):
+    return np.clip(coverage / success_threshold, 0, 1)
+
+
+def main(raw_dir: Path, repo_id: str, mode: str = "video", push_to_hub: bool = True):
+    if mode not in ["video", "image", "keypoints"]:
+        raise ValueError(mode)
+
+    if (HF_LEROBOT_HOME / repo_id).exists():
+        shutil.rmtree(HF_LEROBOT_HOME / repo_id)
+
+    if not raw_dir.exists():
+        download_raw(raw_dir, repo_id="lerobot-raw/pusht_raw")
+
+    zarr_data = load_raw_dataset(zarr_path=raw_dir / "pusht_cchi_v7_replay.zarr")
+
+    env_state = zarr_data["state"][:]
+    agent_pos = env_state[:, :2]
+
+    action = zarr_data["action"][:]
+    image = zarr_data["img"]  # (b, h, w, c)
+
+    if image.dtype == np.float32 and image.max() == np.float32(255):
+        # HACK: images are loaded as float32 but they actually encode uint8 data
+        image = image.astype(np.uint8)
+
+    episode_data_index = {
+        "from": np.concatenate(([0], zarr_data.meta["episode_ends"][:-1])),
+        "to": zarr_data.meta["episode_ends"],
+    }
+
+    # Calculate success and reward based on the overlapping area
+    # of the T-object and the T-area.
+    coverage, keypoints = calculate_coverage(zarr_data)
+    success = calculate_success(coverage, success_threshold=0.95)
+    reward = calculate_reward(coverage, success_threshold=0.95)
+
+    features = build_features(mode)
+    dataset = LeRobotDataset.create(
+        repo_id=repo_id,
+        fps=10,
+        robot_type="2d pointer",
+        features=features,
+        image_writer_threads=4,
+    )
+    episodes = range(len(episode_data_index["from"]))
+    for ep_idx in episodes:
+        from_idx = episode_data_index["from"][ep_idx]
+        to_idx = episode_data_index["to"][ep_idx]
+        num_frames = to_idx - from_idx
+
+        for frame_idx in range(num_frames):
+            i = from_idx + frame_idx
+            idx = i + (frame_idx < num_frames - 1)
+            frame = {
+                "action": action[i],
+                # Shift reward and success by +1 until the last item of the episode
+                "next.reward": reward[idx : idx + 1],
+                "next.success": success[idx : idx + 1],
+                "task": PUSHT_TASK,
+            }
+
+            frame["observation.state"] = agent_pos[i]
+
+            if mode == "keypoints":
+                frame["observation.environment_state"] = keypoints[i]
+            else:
+                frame["observation.image"] = image[i]
+
+            dataset.add_frame(frame)
+
+        dataset.save_episode()
+
+    if push_to_hub:
+        dataset.push_to_hub()
+        hub_api = HfApi()
+        hub_api.create_tag(repo_id, tag=CODEBASE_VERSION, repo_type="dataset")
+
+
+if __name__ == "__main__":
+    # To try this script, modify the repo id with your own HuggingFace user (e.g cadene/pusht)
+    repo_id = "lerobot/pusht"
+
+    modes = ["video", "image", "keypoints"]
+    # Uncomment if you want to try with a specific mode
+    # modes = ["video"]
+    # modes = ["image"]
+    # modes = ["keypoints"]
+
+    raw_dir = Path("data/lerobot-raw/pusht_raw")
+    for mode in modes:
+        if mode in ["image", "keypoints"]:
+            repo_id += f"_{mode}"
+
+        # download and load raw dataset, create LeRobotDataset, populate it, push to hub
+        main(raw_dir, repo_id=repo_id, mode=mode)
+
+        # Uncomment if you want to load the local dataset and explore it
+        # dataset = LeRobotDataset(repo_id=repo_id)
+        # breakpoint()

lerobot/common/cameras/__init__.py

@@ -0,0 +1,4 @@
+from .camera import Camera
+from .configs import CameraConfig
+
+__all__ = ["Camera", "CameraConfig"]

lerobot/common/cameras/camera.py

@@ -0,0 +1,25 @@
+import abc
+
+import numpy as np
+
+
+class Camera(abc.ABC):
+    @abc.abstractmethod
+    def connect(self):
+        pass
+
+    @abc.abstractmethod
+    def read(self, temporary_color: str | None = None) -> np.ndarray:
+        pass
+
+    @abc.abstractmethod
+    def async_read(self) -> np.ndarray:
+        pass
+
+    @abc.abstractmethod
+    def disconnect(self):
+        pass
+
+    def __del__(self):
+        if getattr(self, "is_connected", False):
+            self.disconnect()

lerobot/common/cameras/configs.py

@@ -0,0 +1,11 @@
+import abc
+from dataclasses import dataclass
+
+import draccus
+
+
+@dataclass
+class CameraConfig(draccus.ChoiceRegistry, abc.ABC):
+    @property
+    def type(self) -> str:
+        return self.get_choice_name(self.__class__)

lerobot/common/cameras/intel/__init__.py

@@ -0,0 +1,4 @@
+from .camera_realsense import RealSenseCamera
+from .configuration_realsense import RealSenseCameraConfig
+
+__all__ = ["RealSenseCamera", "RealSenseCameraConfig"]

lerobot/common/cameras/intel/camera_realsense.py

@@ -31,14 +31,15 @@ from threading import Thread
 import numpy as np
 from PIL import Image
 
-from lerobot.common.robot_devices.cameras.configs import IntelRealSenseCameraConfig
-from lerobot.common.robot_devices.utils import (
-    RobotDeviceAlreadyConnectedError,
-    RobotDeviceNotConnectedError,
+from lerobot.common.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+from lerobot.common.utils.robot_utils import (
     busy_wait,
 )
 from lerobot.common.utils.utils import capture_timestamp_utc
 
+from ..camera import Camera
+from .configuration_realsense import RealSenseCameraConfig
+
 SERIAL_NUMBER_INDEX = 1
 
 
@@ -108,13 +109,11 @@ def save_images_from_cameras(
     cameras = []
     for cam_sn in serial_numbers:
         print(f"{cam_sn=}")
-        config = IntelRealSenseCameraConfig(
-            serial_number=cam_sn, fps=fps, width=width, height=height, mock=mock
-        )
-        camera = IntelRealSenseCamera(config)
+        config = RealSenseCameraConfig(serial_number=cam_sn, fps=fps, width=width, height=height, mock=mock)
+        camera = RealSenseCamera(config)
         camera.connect()
         print(
-            f"IntelRealSenseCamera({camera.serial_number}, fps={camera.fps}, width={camera.capture_width}, height={camera.capture_height}, color_mode={camera.color_mode})"
+            f"RealSenseCamera({camera.serial_number}, fps={camera.fps}, width={camera.capture_width}, height={camera.capture_height}, color_mode={camera.color_mode})"
         )
         cameras.append(camera)
 
@@ -166,11 +165,11 @@ def save_images_from_cameras(
             camera.disconnect()
 
 
-class IntelRealSenseCamera:
+class RealSenseCamera(Camera):
     """
-    The IntelRealSenseCamera class is similar to OpenCVCamera class but adds additional features for Intel Real Sense cameras:
+    The RealSenseCamera class is similar to OpenCVCamera class but adds additional features for Intel Real Sense cameras:
     - is instantiated with the serial number of the camera - won't randomly change as it can be the case of OpenCVCamera for Linux,
-    - can also be instantiated with the camera's name — if it's unique — using IntelRealSenseCamera.init_from_name(),
+    - can also be instantiated with the camera's name — if it's unique — using RealSenseCamera.init_from_name(),
     - depth map can be returned.
 
     To find the camera indices of your cameras, you can run our utility script that will save a few frames for each camera:
@@ -178,15 +177,15 @@ class IntelRealSenseCamera:
     python lerobot/common/robot_devices/cameras/intelrealsense.py --images-dir outputs/images_from_intelrealsense_cameras
     ```
 
-    When an IntelRealSenseCamera is instantiated, if no specific config is provided, the default fps, width, height and color_mode
+    When an RealSenseCamera is instantiated, if no specific config is provided, the default fps, width, height and color_mode
     of the given camera will be used.
 
     Example of instantiating with a serial number:
     ```python
-    from lerobot.common.robot_devices.cameras.configs import IntelRealSenseCameraConfig
+    from lerobot.common.robot_devices.cameras.configs import RealSenseCameraConfig
 
-    config = IntelRealSenseCameraConfig(serial_number=128422271347)
-    camera = IntelRealSenseCamera(config)
+    config = RealSenseCameraConfig(serial_number=128422271347)
+    camera = RealSenseCamera(config)
     camera.connect()
     color_image = camera.read()
     # when done using the camera, consider disconnecting
@@ -195,21 +194,21 @@ class IntelRealSenseCamera:
 
     Example of instantiating with a name if it's unique:
     ```
-    config = IntelRealSenseCameraConfig(name="Intel RealSense D405")
+    config = RealSenseCameraConfig(name="Intel RealSense D405")
     ```
 
     Example of changing default fps, width, height and color_mode:
     ```python
-    config = IntelRealSenseCameraConfig(serial_number=128422271347, fps=30, width=1280, height=720)
-    config = IntelRealSenseCameraConfig(serial_number=128422271347, fps=90, width=640, height=480)
-    config = IntelRealSenseCameraConfig(serial_number=128422271347, fps=90, width=640, height=480, color_mode="bgr")
+    config = RealSenseCameraConfig(serial_number=128422271347, fps=30, width=1280, height=720)
+    config = RealSenseCameraConfig(serial_number=128422271347, fps=90, width=640, height=480)
+    config = RealSenseCameraConfig(serial_number=128422271347, fps=90, width=640, height=480, color_mode="bgr")
     # Note: might error out upon `camera.connect()` if these settings are not compatible with the camera
     ```
 
     Example of returning depth:
     ```python
-    config = IntelRealSenseCameraConfig(serial_number=128422271347, use_depth=True)
-    camera = IntelRealSenseCamera(config)
+    config = RealSenseCameraConfig(serial_number=128422271347, use_depth=True)
+    camera = RealSenseCamera(config)
     camera.connect()
     color_image, depth_map = camera.read()
     ```
@@ -217,7 +216,7 @@ class IntelRealSenseCamera:
 
     def __init__(
         self,
-        config: IntelRealSenseCameraConfig,
+        config: RealSenseCameraConfig,
     ):
         self.config = config
         if config.name is not None:
@@ -282,9 +281,7 @@ class IntelRealSenseCamera:
 
     def connect(self):
         if self.is_connected:
-            raise RobotDeviceAlreadyConnectedError(
-                f"IntelRealSenseCamera({self.serial_number}) is already connected."
-            )
+            raise DeviceAlreadyConnectedError(f"RealSenseCamera({self.serial_number}) is already connected.")
 
         if self.mock:
             import tests.cameras.mock_pyrealsense2 as rs
@@ -330,7 +327,7 @@ class IntelRealSenseCamera:
                     "To find the serial number you should use, run `python lerobot/common/robot_devices/cameras/intelrealsense.py`."
                 )
 
-            raise OSError(f"Can't access IntelRealSenseCamera({self.serial_number}).")
+            raise OSError(f"Can't access RealSenseCamera({self.serial_number}).")
 
         color_stream = profile.get_stream(rs.stream.color)
         color_profile = color_stream.as_video_stream_profile()
@@ -342,15 +339,15 @@ class IntelRealSenseCamera:
         if self.fps is not None and not math.isclose(self.fps, actual_fps, rel_tol=1e-3):
             # Using `OSError` since it's a broad that encompasses issues related to device communication
             raise OSError(
-                f"Can't set {self.fps=} for IntelRealSenseCamera({self.serial_number}). Actual value is {actual_fps}."
+                f"Can't set {self.fps=} for RealSenseCamera({self.serial_number}). Actual value is {actual_fps}."
             )
         if self.capture_width is not None and self.capture_width != actual_width:
             raise OSError(
-                f"Can't set {self.capture_width=} for IntelRealSenseCamera({self.serial_number}). Actual value is {actual_width}."
+                f"Can't set {self.capture_width=} for RealSenseCamera({self.serial_number}). Actual value is {actual_width}."
             )
         if self.capture_height is not None and self.capture_height != actual_height:
             raise OSError(
-                f"Can't set {self.capture_height=} for IntelRealSenseCamera({self.serial_number}). Actual value is {actual_height}."
+                f"Can't set {self.capture_height=} for RealSenseCamera({self.serial_number}). Actual value is {actual_height}."
             )
 
         self.fps = round(actual_fps)
@@ -370,8 +367,8 @@ class IntelRealSenseCamera:
         If you are reading data from other sensors, we advise to use `camera.async_read()` which is non blocking version of `camera.read()`.
         """
         if not self.is_connected:
-            raise RobotDeviceNotConnectedError(
-                f"IntelRealSenseCamera({self.serial_number}) is not connected. Try running `camera.connect()` first."
+            raise DeviceNotConnectedError(
+                f"RealSenseCamera({self.serial_number}) is not connected. Try running `camera.connect()` first."
             )
 
         if self.mock:
@@ -386,7 +383,7 @@ class IntelRealSenseCamera:
         color_frame = frame.get_color_frame()
 
         if not color_frame:
-            raise OSError(f"Can't capture color image from IntelRealSenseCamera({self.serial_number}).")
+            raise OSError(f"Can't capture color image from RealSenseCamera({self.serial_number}).")
 
         color_image = np.asanyarray(color_frame.get_data())
 
@@ -418,7 +415,7 @@ class IntelRealSenseCamera:
         if self.use_depth:
             depth_frame = frame.get_depth_frame()
             if not depth_frame:
-                raise OSError(f"Can't capture depth image from IntelRealSenseCamera({self.serial_number}).")
+                raise OSError(f"Can't capture depth image from RealSenseCamera({self.serial_number}).")
 
             depth_map = np.asanyarray(depth_frame.get_data())
 
@@ -445,8 +442,8 @@ class IntelRealSenseCamera:
     def async_read(self):
         """Access the latest color image"""
         if not self.is_connected:
-            raise RobotDeviceNotConnectedError(
-                f"IntelRealSenseCamera({self.serial_number}) is not connected. Try running `camera.connect()` first."
+            raise DeviceNotConnectedError(
+                f"RealSenseCamera({self.serial_number}) is not connected. Try running `camera.connect()` first."
             )
 
         if self.thread is None:
@@ -472,8 +469,8 @@ class IntelRealSenseCamera:
 
     def disconnect(self):
         if not self.is_connected:
-            raise RobotDeviceNotConnectedError(
-                f"IntelRealSenseCamera({self.serial_number}) is not connected. Try running `camera.connect()` first."
+            raise DeviceNotConnectedError(
+                f"RealSenseCamera({self.serial_number}) is not connected. Try running `camera.connect()` first."
             )
 
         if self.thread is not None and self.thread.is_alive():
@@ -495,14 +492,14 @@ class IntelRealSenseCamera:
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser(
-        description="Save a few frames using `IntelRealSenseCamera` for all cameras connected to the computer, or a selected subset."
+        description="Save a few frames using `RealSenseCamera` for all cameras connected to the computer, or a selected subset."
     )
     parser.add_argument(
         "--serial-numbers",
         type=int,
         nargs="*",
         default=None,
-        help="List of serial numbers used to instantiate the `IntelRealSenseCamera`. If not provided, find and use all available camera indices.",
+        help="List of serial numbers used to instantiate the `RealSenseCamera`. If not provided, find and use all available camera indices.",
     )
     parser.add_argument(
         "--fps",
@@ -512,13 +509,13 @@ if __name__ == "__main__":
     )
     parser.add_argument(
         "--width",
-        type=int,
+        type=str,
         default=640,
         help="Set the width for all cameras. If not provided, use the default width of each camera.",
     )
     parser.add_argument(
         "--height",
-        type=int,
+        type=str,
         default=480,
         help="Set the height for all cameras. If not provided, use the default height of each camera.",
     )

lerobot/common/cameras/intel/configuration_realsense.py

@@ -12,67 +12,24 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import abc
 from dataclasses import dataclass
 
-import draccus
-
-
-@dataclass
-class CameraConfig(draccus.ChoiceRegistry, abc.ABC):
-    @property
-    def type(self) -> str:
-        return self.get_choice_name(self.__class__)
-
-
-@CameraConfig.register_subclass("opencv")
-@dataclass
-class OpenCVCameraConfig(CameraConfig):
-    """
-    Example of tested options for Intel Real Sense D405:
-
-    ```python
-    OpenCVCameraConfig(0, 30, 640, 480)
-    OpenCVCameraConfig(0, 60, 640, 480)
-    OpenCVCameraConfig(0, 90, 640, 480)
-    OpenCVCameraConfig(0, 30, 1280, 720)
-    ```
-    """
-
-    camera_index: int
-    fps: int | None = None
-    width: int | None = None
-    height: int | None = None
-    color_mode: str = "rgb"
-    channels: int | None = None
-    rotation: int | None = None
-    mock: bool = False
-
-    def __post_init__(self):
-        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."
-            )
-
-        self.channels = 3
-
-        if self.rotation not in [-90, None, 90, 180]:
-            raise ValueError(f"`rotation` must be in [-90, None, 90, 180] (got {self.rotation})")
+from ..configs import CameraConfig
 
 
 @CameraConfig.register_subclass("intelrealsense")
 @dataclass
-class IntelRealSenseCameraConfig(CameraConfig):
+class RealSenseCameraConfig(CameraConfig):
     """
     Example of tested options for Intel Real Sense D405:
 
     ```python
-    IntelRealSenseCameraConfig(128422271347, 30, 640, 480)
-    IntelRealSenseCameraConfig(128422271347, 60, 640, 480)
-    IntelRealSenseCameraConfig(128422271347, 90, 640, 480)
-    IntelRealSenseCameraConfig(128422271347, 30, 1280, 720)
-    IntelRealSenseCameraConfig(128422271347, 30, 640, 480, use_depth=True)
-    IntelRealSenseCameraConfig(128422271347, 30, 640, 480, rotation=90)
+    RealSenseCameraConfig(128422271347, 30, 640, 480)
+    RealSenseCameraConfig(128422271347, 60, 640, 480)
+    RealSenseCameraConfig(128422271347, 90, 640, 480)
+    RealSenseCameraConfig(128422271347, 30, 1280, 720)
+    RealSenseCameraConfig(128422271347, 30, 640, 480, use_depth=True)
+    RealSenseCameraConfig(128422271347, 30, 640, 480, rotation=90)
     ```
     """
 

lerobot/common/cameras/opencv/__init__.py

@@ -0,0 +1,4 @@
+from .camera_opencv import OpenCVCamera
+from .configuration_opencv import OpenCVCameraConfig
+
+__all__ = ["OpenCVCamera", "OpenCVCameraConfig"]

lerobot/common/cameras/opencv/camera_opencv.py

@@ -29,14 +29,15 @@ from threading import Thread
 import numpy as np
 from PIL import Image
 
-from lerobot.common.robot_devices.cameras.configs import OpenCVCameraConfig
-from lerobot.common.robot_devices.utils import (
-    RobotDeviceAlreadyConnectedError,
-    RobotDeviceNotConnectedError,
+from lerobot.common.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+from lerobot.common.utils.robot_utils import (
     busy_wait,
 )
 from lerobot.common.utils.utils import capture_timestamp_utc
 
+from ..camera import Camera
+from .configuration_opencv import OpenCVCameraConfig
+
 # The maximum opencv device index depends on your operating system. For instance,
 # if you have 3 cameras, they should be associated to index 0, 1, and 2. This is the case
 # on MacOS. However, on Ubuntu, the indices are different like 6, 16, 23.
@@ -190,7 +191,7 @@ def save_images_from_cameras(
     print(f"Images have been saved to {images_dir}")
 
 
-class OpenCVCamera:
+class OpenCVCamera(Camera):
     """
     The OpenCVCamera class allows to efficiently record images from cameras. It relies on opencv2 to communicate
     with the cameras. Most cameras are compatible. For more info, see the [Video I/O with OpenCV Overview](https://docs.opencv.org/4.x/d0/da7/videoio_overview.html).
@@ -283,7 +284,7 @@ class OpenCVCamera:
 
     def connect(self):
         if self.is_connected:
-            raise RobotDeviceAlreadyConnectedError(f"OpenCVCamera({self.camera_index}) is already connected.")
+            raise DeviceAlreadyConnectedError(f"OpenCVCamera({self.camera_index}) is already connected.")
 
         if self.mock:
             import tests.cameras.mock_cv2 as cv2
@@ -375,7 +376,7 @@ class OpenCVCamera:
         If you are reading data from other sensors, we advise to use `camera.async_read()` which is non blocking version of `camera.read()`.
         """
         if not self.is_connected:
-            raise RobotDeviceNotConnectedError(
+            raise DeviceNotConnectedError(
                 f"OpenCVCamera({self.camera_index}) is not connected. Try running `camera.connect()` first."
             )
 
@@ -432,7 +433,7 @@ class OpenCVCamera:
 
     def async_read(self):
         if not self.is_connected:
-            raise RobotDeviceNotConnectedError(
+            raise DeviceNotConnectedError(
                 f"OpenCVCamera({self.camera_index}) is not connected. Try running `camera.connect()` first."
             )
 
@@ -454,7 +455,7 @@ class OpenCVCamera:
 
     def disconnect(self):
         if not self.is_connected:
-            raise RobotDeviceNotConnectedError(
+            raise DeviceNotConnectedError(
                 f"OpenCVCamera({self.camera_index}) is not connected. Try running `camera.connect()` first."
             )
 
@@ -492,13 +493,13 @@ if __name__ == "__main__":
     )
     parser.add_argument(
         "--width",
-        type=int,
+        type=str,
         default=None,
         help="Set the width for all cameras. If not provided, use the default width of each camera.",
     )
     parser.add_argument(
         "--height",
-        type=int,
+        type=str,
         default=None,
         help="Set the height for all cameras. If not provided, use the default height of each camera.",
     )

lerobot/common/cameras/opencv/configuration_opencv.py

@@ -0,0 +1,38 @@
+from dataclasses import dataclass
+
+from ..configs import CameraConfig
+
+
+@CameraConfig.register_subclass("opencv")
+@dataclass
+class OpenCVCameraConfig(CameraConfig):
+    """
+    Example of tested options for Intel Real Sense D405:
+
+    ```python
+    OpenCVCameraConfig(0, 30, 640, 480)
+    OpenCVCameraConfig(0, 60, 640, 480)
+    OpenCVCameraConfig(0, 90, 640, 480)
+    OpenCVCameraConfig(0, 30, 1280, 720)
+    ```
+    """
+
+    camera_index: int
+    fps: int | None = None
+    width: int | None = None
+    height: int | None = None
+    color_mode: str = "rgb"
+    channels: int | None = None
+    rotation: int | None = None
+    mock: bool = False
+
+    def __post_init__(self):
+        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."
+            )
+
+        self.channels = 3
+
+        if self.rotation not in [-90, None, 90, 180]:
+            raise ValueError(f"`rotation` must be in [-90, None, 90, 180] (got {self.rotation})")

lerobot/common/cameras/utils.py

@@ -0,0 +1,21 @@
+from .camera import Camera
+from .configs import CameraConfig
+
+
+def make_cameras_from_configs(camera_configs: dict[str, CameraConfig]) -> dict[str, Camera]:
+    cameras = {}
+
+    for key, cfg in camera_configs.items():
+        if cfg.type == "opencv":
+            from .opencv import OpenCVCamera
+
+            cameras[key] = OpenCVCamera(cfg)
+
+        elif cfg.type == "intelrealsense":
+            from .intel.camera_realsense import RealSenseCamera
+
+            cameras[key] = RealSenseCamera(cfg)
+        else:
+            raise ValueError(f"The motor type '{cfg.type}' is not valid.")
+
+    return cameras

lerobot/common/constants.py

@@ -17,12 +17,15 @@ from pathlib import Path
 
 from huggingface_hub.constants import HF_HOME
 
-OBS_ENV = "observation.environment_state"
-OBS_ROBOT = "observation.state"
+OBS_ENV_STATE = "observation.environment_state"
+OBS_STATE = "observation.state"
 OBS_IMAGE = "observation.image"
 OBS_IMAGES = "observation.images"
 ACTION = "action"
 
+ROBOTS = "robots"
+TELEOPERATORS = "teleoperators"
+
 # files & directories
 CHECKPOINTS_DIR = "checkpoints"
 LAST_CHECKPOINT_LINK = "last"
@@ -34,12 +37,16 @@ OPTIMIZER_STATE = "optimizer_state.safetensors"
 OPTIMIZER_PARAM_GROUPS = "optimizer_param_groups.json"
 SCHEDULER_STATE = "scheduler_state.json"
 
-# cache dir
-default_cache_path = Path(HF_HOME) / "lerobot"
-HF_LEROBOT_HOME = Path(os.getenv("HF_LEROBOT_HOME", default_cache_path)).expanduser()
-
 if "LEROBOT_HOME" in os.environ:
     raise ValueError(
         f"You have a 'LEROBOT_HOME' environment variable set to '{os.getenv('LEROBOT_HOME')}'.\n"
         "'LEROBOT_HOME' is deprecated, please use 'HF_LEROBOT_HOME' instead."
     )
+
+# cache dir
+default_cache_path = Path(HF_HOME) / "lerobot"
+HF_LEROBOT_HOME = Path(os.getenv("HF_LEROBOT_HOME", default_cache_path)).expanduser()
+
+# calibration dir
+default_calibration_path = HF_LEROBOT_HOME / ".calibration"
+HF_LEROBOT_CALIBRATION = Path(os.getenv("HF_LEROBOT_CALIBRATION", default_calibration_path)).expanduser()

lerobot/common/datasets/aggregate.py

@@ -1,416 +0,0 @@
-import logging
-import shutil
-from pathlib import Path
-
-import pandas as pd
-import tqdm
-
-from lerobot.common.constants import HF_LEROBOT_HOME
-from lerobot.common.datasets.compute_stats import aggregate_stats
-from lerobot.common.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
-from lerobot.common.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,
-    concat_video_files,
-    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.common.utils.utils import init_logging
-
-
-def validate_all_metadata(all_metadata: list[LeRobotDatasetMetadata]):
-    # validate same fps, robot_type, features
-
-    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):
-    def _update(row):
-        row["episode_index"] = row["episode_index"] + dst_meta["total_episodes"]
-        row["index"] = row["index"] + dst_meta["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,
-):
-    def _update(row):
-        row["meta/episodes/chunk_index"] = row["meta/episodes/chunk_index"] + meta_idx["chunk_index"]
-        row["meta/episodes/file_index"] = row["meta/episodes/file_index"] + meta_idx["file_index"]
-        row["data/chunk_index"] = row["data/chunk_index"] + data_idx["chunk_index"]
-        row["data/file_index"] = row["data/file_index"] + data_idx["file_index"]
-        for key, video_idx in videos_idx.items():
-            row[f"videos/{key}/chunk_index"] = row[f"videos/{key}/chunk_index"] + video_idx["chunk_index"]
-            row[f"videos/{key}/file_index"] = row[f"videos/{key}/file_index"] + video_idx["file_index"]
-            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, aggr_root=None):
-    logging.info("Start aggregate_datasets")
-
-    # Load metadata
-    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"]
-    image_keys = [key for key in features if features[key]["dtype"] == "image"]
-
-    # Initialize output dataset metadata
-    dst_meta = LeRobotDatasetMetadata.create(
-        repo_id=aggr_repo_id,
-        fps=fps,
-        robot_type=robot_type,
-        features=features,
-        root=aggr_root,
-    )
-
-    # Aggregate task info
-    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)
-
-    # Track counters and indices
-    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 = {}
-
-    # Process each dataset
-    for src_meta in tqdm.tqdm(all_metadata, desc="Copy data and videos"):
-        videos_idx = aggregate_videos(src_meta, dst_meta, videos_idx)
-        data_idx = aggregate_data(src_meta, dst_meta, data_idx)
-    
-        meta_idx = aggregate_metadata(src_meta, dst_meta, meta_idx, data_idx, videos_idx, video_keys, image_keys)
-
-        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.")
-
-
-# -------------------------------
-# Helper Functions
-# -------------------------------
-
-
-def aggregate_videos(src_meta, dst_meta, videos_idx):
-    """
-    Aggregates video chunks from a dataset into the aggregated dataset folder.
-    """
-    for key, video_idx in videos_idx.items():
-        # Get unique (chunk, file) combinations
-        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,
-            )
-        }
-
-        # Current target chunk/file index
-        chunk_idx = video_idx["chunk_idx"]
-        file_idx = video_idx["file_idx"]
-
-        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 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
-
-            # 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 >= DEFAULT_VIDEO_FILE_SIZE_IN_MB:
-                # Rotate to a new chunk/file
-                chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, DEFAULT_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:
-                # Append to existing video file
-                concat_video_files(
-                    [dst_path, src_path],
-                    dst_meta.root,
-                    key,
-                    chunk_idx,
-                    file_idx,
-                )
-
-                if src_size + dst_size >= DEFAULT_DATA_FILE_SIZE_IN_MB:
-                    # Size limit is reached, prepare new parquet file
-                    aggr_data_chunk_idx, aggr_data_file_idx = update_chunk_file_indices(
-                        aggr_data_chunk_idx, aggr_data_file_idx, DEFAULT_CHUNK_SIZE
-                    )
-                    aggr_path = aggr_root / DEFAULT_DATA_PATH.format(
-                        chunk_index=aggr_data_chunk_idx, file_index=aggr_data_file_idx
-                    )
-                    aggr_path.parent.mkdir(parents=True, exist_ok=True)
-                    df.to_parquet(aggr_path)
-                else:
-                    # Update the existing parquet file with new rows
-                    aggr_df = pd.read_parquet(aggr_path)
-                    df = pd.concat([aggr_df, df], ignore_index=True)
-                    to_parquet_with_hf_images(df, aggr_path, dst_meta.image_keys)
-
-        return videos_idx
-
-
-def aggregate_data(src_meta, dst_meta, data_idx):
-    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
-        )
-    }
-    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,
-            DEFAULT_DATA_FILE_SIZE_IN_MB,
-            DEFAULT_CHUNK_SIZE,
-            DEFAULT_DATA_PATH,
-            contains_images=len(dst_meta.image_keys) > 0
-        )
-
-    return data_idx
-
-
-def aggregate_metadata(src_meta, dst_meta, meta_idx, data_idx, videos_idx):
-    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,
-        )
-    }
-
-    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 video_keys:
-        #     video_idx[k]["latest_duration"] += video_idx[k]["episode_duration"]
-
-        append_or_create_parquet_file(
-            df,
-            src_path,
-            meta_idx,
-            DEFAULT_DATA_FILE_SIZE_IN_MB,
-            DEFAULT_CHUNK_SIZE,
-            DEFAULT_EPISODES_PATH,
-        )
-
-    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,
-):
-    """
-    Safely appends or creates a Parquet file at dst_path based on size constraints.
-
-    Parameters:
-        df (pd.DataFrame): Data to write.
-        src_path (Path): Path to source file (used to get size).
-        idx (dict): Dictionary containing 'chunk' and 'file' indices.
-        max_mb (float): Maximum allowed file size in MB.
-        chunk_size (int): Maximum number of files per chunk.
-        default_path (str): Format string for generating a new file path.
-
-    Returns:
-        dict: Updated index dictionary.
-    """
-    # Initial destination path
-    dst_path = aggr_root / DEFAULT_DATA_PATH.format(
-        chunk_index=idx["chunk"], file_index=idx["file"]
-    )
-
-    # If destination file doesn't exist, just write the new one
-    if not dst_path.exists():
-        dst_path.parent.mkdir(parents=True, exist_ok=True)
-        df.to_parquet(dst_path)
-        return idx
-
-    # Otherwise, check if we exceed the size limit
-    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:
-        # File is too large, move to a new one
-        idx["chunk"], idx["file"] = update_chunk_file_indices(idx["chunk"], idx["file"], chunk_size)
-        new_path = dst_path.parent / default_path.format(chunk_index=idx["chunk"], file_index=idx["file"])
-        new_path.parent.mkdir(parents=True, exist_ok=True)
-        final_df = df
-    else:
-        # Append to existing file
-        existing_df = pd.read_parquet(dst_path)
-        final_df = pd.concat([existing_df, df], ignore_index=True)
-
-    if contains_images:
-        to_parquet_with_hf_images(final_df, new_path)
-    else:
-        final_df.to_parquet(new_path)
-
-    return idx
-
-
-def finalize_aggregation(aggr_meta, all_metadata):
-    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)
-
-
-if __name__ == "__main__":
-    init_logging()
-
-    num_shards = 2048
-    repo_id = "cadene/droid_1.0.1_v30"
-    aggr_repo_id = f"{repo_id}_compact_6"
-    tags = ["openx"]
-
-    # num_shards = 210
-    # repo_id = "cadene/agibot_alpha_v30"
-    # aggr_repo_id = f"{repo_id}"
-    # tags = None
-
-    # aggr_root = Path(f"/tmp/{aggr_repo_id}")
-    aggr_root = HF_LEROBOT_HOME / aggr_repo_id
-    if aggr_root.exists():
-        shutil.rmtree(aggr_root)
-
-    repo_ids = []
-    roots = []
-    for rank in range(num_shards):
-        shard_repo_id = f"{repo_id}_world_{num_shards}_rank_{rank}"
-        shard_root = HF_LEROBOT_HOME / shard_repo_id
-        try:
-            meta = LeRobotDatasetMetadata(shard_repo_id, root=shard_root)
-            if len(meta.video_keys) == 0:
-                continue
-            repo_ids.append(shard_repo_id)
-            roots.append(shard_root)
-        except:
-            pass
-
-        if rank == 1:
-            break
-
-    aggregate_datasets(
-        repo_ids,
-        aggr_repo_id,
-        roots=roots,
-        aggr_root=aggr_root,
-    )
-
-    aggr_dataset = LeRobotDataset(repo_id=aggr_repo_id, root=aggr_root)
-    # for i in tqdm.tqdm(range(len(aggr_dataset))):
-    #     aggr_dataset[i]
-    #     pass
-    aggr_dataset.push_to_hub(tags=tags, upload_large_folder=True)

lerobot/common/datasets/backward_compatibility.py

@@ -47,18 +47,6 @@ If you encounter a problem, contact LeRobot maintainers on [Discord](https://dis
 or open an [issue on GitHub](https://github.com/huggingface/lerobot/issues/new/choose).
 """
 
-V30_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 lerobot/common/datasets/v30/convert_dataset_v21_to_v30.py --repo-id={repo_id}
-```
-
-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).
-"""
-
 FUTURE_MESSAGE = """
 The dataset you requested ({repo_id}) is only available in {version} format.
 As we cannot ensure forward compatibility with it, please update your current version of lerobot.
@@ -70,14 +58,7 @@ class CompatibilityError(Exception): ...
 
 class BackwardCompatibilityError(CompatibilityError):
     def __init__(self, repo_id: str, version: packaging.version.Version):
-        if version.major == 3:
-            message = V30_MESSAGE.format(repo_id=repo_id, version=version)
-        elif version.major == 2:
-            message = V2_MESSAGE.format(repo_id=repo_id, version=version)
-        else:
-            raise NotImplementedError(
-                "Contact the maintainer on [Discord](https://discord.com/invite/s3KuuzsPFb)."
-            )
+        message = V2_MESSAGE.format(repo_id=repo_id, version=version)
         super().__init__(message)
 
 

lerobot/common/datasets/lerobot_dataset.py

@@ -16,18 +16,16 @@
 import contextlib
 import logging
 import shutil
-import tempfile
 from pathlib import Path
 from typing import Callable
 
 import datasets
 import numpy as np
 import packaging.version
-import pandas as pd
 import PIL.Image
 import torch
 import torch.utils
-from datasets import Dataset
+from datasets import concatenate_datasets, load_dataset
 from huggingface_hub import HfApi, snapshot_download
 from huggingface_hub.constants import REPOCARD_NAME
 from huggingface_hub.errors import RevisionNotFoundError
@@ -36,52 +34,46 @@ from lerobot.common.constants import HF_LEROBOT_HOME
 from lerobot.common.datasets.compute_stats import aggregate_stats, compute_episode_stats
 from lerobot.common.datasets.image_writer import AsyncImageWriter, write_image
 from lerobot.common.datasets.utils import (
-    DEFAULT_EPISODES_PATH,
     DEFAULT_FEATURES,
     DEFAULT_IMAGE_PATH,
     INFO_PATH,
+    TASKS_PATH,
+    append_jsonlines,
+    backward_compatible_episodes_stats,
     check_delta_timestamps,
+    check_timestamps_sync,
     check_version_compatibility,
-    concat_video_files,
     create_empty_dataset_info,
     create_lerobot_dataset_card,
     embed_images,
-    flatten_dict,
     get_delta_indices,
+    get_episode_data_index,
     get_features_from_robot,
-    get_hf_dataset_size_in_mb,
     get_hf_features_from_features,
-    get_parquet_file_size_in_mb,
-    get_parquet_num_frames,
     get_safe_version,
-    get_video_duration_in_s,
-    get_video_size_in_mb,
     hf_transform_to_torch,
     is_valid_version,
     load_episodes,
+    load_episodes_stats,
     load_info,
-    load_nested_dataset,
     load_stats,
     load_tasks,
-    to_parquet_with_hf_images,
-    update_chunk_file_indices,
     validate_episode_buffer,
     validate_frame,
+    write_episode,
+    write_episode_stats,
     write_info,
     write_json,
-    write_stats,
-    write_tasks,
 )
 from lerobot.common.datasets.video_utils import (
     VideoFrame,
-    decode_video_frames,
+    decode_video_frames_torchvision,
     encode_video_frames,
-    get_safe_default_codec,
     get_video_info,
 )
-from lerobot.common.robot_devices.robots.utils import Robot
+from lerobot.common.robots.utils import Robot
 
-CODEBASE_VERSION = "v3.0"
+CODEBASE_VERSION = "v2.1"
 
 
 class LeRobotDatasetMetadata:
@@ -105,18 +97,20 @@ class LeRobotDatasetMetadata:
                 self.revision = get_safe_version(self.repo_id, self.revision)
 
             (self.root / "meta").mkdir(exist_ok=True, parents=True)
-            # TODO(rcadene): instead of downloading all episodes metadata files,
-            # download only the ones associated to the requested episodes. This would
-            # require adding `episodes: list[int]` as argument.
             self.pull_from_repo(allow_patterns="meta/")
             self.load_metadata()
 
     def load_metadata(self):
         self.info = load_info(self.root)
         check_version_compatibility(self.repo_id, self._version, CODEBASE_VERSION)
-        self.tasks = load_tasks(self.root)
+        self.tasks, self.task_to_task_index = load_tasks(self.root)
         self.episodes = load_episodes(self.root)
-        self.stats = load_stats(self.root)
+        if self._version < packaging.version.parse("v2.1"):
+            self.stats = load_stats(self.root)
+            self.episodes_stats = backward_compatible_episodes_stats(self.stats, self.episodes)
+        else:
+            self.episodes_stats = load_episodes_stats(self.root)
+            self.stats = aggregate_stats(list(self.episodes_stats.values()))
 
     def pull_from_repo(
         self,
@@ -138,19 +132,18 @@ class LeRobotDatasetMetadata:
         return packaging.version.parse(self.info["codebase_version"])
 
     def get_data_file_path(self, ep_index: int) -> Path:
-        ep = self.episodes[ep_index]
-        chunk_idx = ep["data/chunk_index"]
-        file_idx = ep["data/file_index"]
-        fpath = self.data_path.format(chunk_index=chunk_idx, file_index=file_idx)
+        ep_chunk = self.get_episode_chunk(ep_index)
+        fpath = self.data_path.format(episode_chunk=ep_chunk, episode_index=ep_index)
         return Path(fpath)
 
     def get_video_file_path(self, ep_index: int, vid_key: str) -> Path:
-        ep = self.episodes[ep_index]
-        chunk_idx = ep[f"videos/{vid_key}/chunk_index"]
-        file_idx = ep[f"videos/{vid_key}/file_index"]
-        fpath = self.video_path.format(video_key=vid_key, chunk_index=chunk_idx, file_index=file_idx)
+        ep_chunk = self.get_episode_chunk(ep_index)
+        fpath = self.video_path.format(episode_chunk=ep_chunk, video_key=vid_key, episode_index=ep_index)
         return Path(fpath)
 
+    def get_episode_chunk(self, ep_index: int) -> int:
+        return ep_index // self.chunks_size
+
     @property
     def data_path(self) -> str:
         """Formattable string for the parquet files."""
@@ -216,109 +209,40 @@ class LeRobotDatasetMetadata:
         """Total number of different tasks performed in this dataset."""
         return self.info["total_tasks"]
 
+    @property
+    def total_chunks(self) -> int:
+        """Total number of chunks (groups of episodes)."""
+        return self.info["total_chunks"]
+
     @property
     def chunks_size(self) -> int:
-        """Max number of files per chunk."""
+        """Max number of episodes per chunk."""
         return self.info["chunks_size"]
 
-    @property
-    def data_files_size_in_mb(self) -> int:
-        """Max size of data file in mega bytes."""
-        return self.info["data_files_size_in_mb"]
-
-    @property
-    def video_files_size_in_mb(self) -> int:
-        """Max size of video file in mega bytes."""
-        return self.info["video_files_size_in_mb"]
-
     def get_task_index(self, task: str) -> int | None:
         """
         Given a task in natural language, returns its task_index if the task already exists in the dataset,
         otherwise return None.
         """
-        if task in self.tasks.index:
-            return int(self.tasks.loc[task].task_index)
-        else:
-            return None
+        return self.task_to_task_index.get(task, None)
 
-    def save_episode_tasks(self, tasks: list[str]):
-        if len(set(tasks)) != len(tasks):
-            raise ValueError(f"Tasks are not unique: {tasks}")
-
-        if self.tasks is None:
-            new_tasks = tasks
-            task_indices = range(len(tasks))
-            self.tasks = pd.DataFrame({"task_index": task_indices}, index=tasks)
-        else:
-            new_tasks = [task for task in tasks if task not in self.tasks.index]
-            new_task_indices = range(len(self.tasks), len(self.tasks) + len(new_tasks))
-            for task_idx, task in zip(new_task_indices, new_tasks, strict=False):
-                self.tasks.loc[task] = task_idx
-
-        if len(new_tasks) > 0:
-            # Update on disk
-            write_tasks(self.tasks, self.root)
-
-    def _save_episode_metadata(self, episode_dict: dict) -> None:
-        """Save episode metadata to a parquet file and update the Hugging Face dataset of episodes metadata.
-
-        This function processes episodes metadata from a dictionary, converts it into a Hugging Face dataset,
-        and saves it as a parquet file. It handles both the creation of new parquet files and the
-        updating of existing ones based on size constraints. After saving the metadata, it reloads
-        the Hugging Face dataset to ensure it is up-to-date.
-
-        Notes: We both need to update parquet files and HF dataset:
-        - `pandas` loads parquet file in RAM
-        - `datasets` relies on a memory mapping from pyarrow (no RAM). It either converts parquet files to a pyarrow cache on disk,
-          or loads directly from pyarrow cache.
+    def add_task(self, task: str):
         """
-        # Convert buffer into HF Dataset
-        episode_dict = {key: [value] for key, value in episode_dict.items()}
-        ep_dataset = Dataset.from_dict(episode_dict)
-        ep_size_in_mb = get_hf_dataset_size_in_mb(ep_dataset)
-        df = pd.DataFrame(ep_dataset)
-        num_frames = episode_dict["length"][0]
+        Given a task in natural language, add it to the dictionary of tasks.
+        """
+        if task in self.task_to_task_index:
+            raise ValueError(f"The task '{task}' already exists and can't be added twice.")
 
-        if self.episodes is None:
-            # Initialize indices and frame count for a new dataset made of the first episode data
-            chunk_idx, file_idx = 0, 0
-            df["meta/episodes/chunk_index"] = [chunk_idx]
-            df["meta/episodes/file_index"] = [file_idx]
-            df["dataset_from_index"] = [0]
-            df["dataset_to_index"] = [num_frames]
-        else:
-            # Retrieve information from the latest parquet file
-            latest_ep = self.episodes[-1]
-            chunk_idx = latest_ep["meta/episodes/chunk_index"]
-            file_idx = latest_ep["meta/episodes/file_index"]
+        task_index = self.info["total_tasks"]
+        self.task_to_task_index[task] = task_index
+        self.tasks[task_index] = task
+        self.info["total_tasks"] += 1
 
-            latest_path = self.root / DEFAULT_EPISODES_PATH.format(chunk_index=chunk_idx, file_index=file_idx)
-            latest_size_in_mb = get_parquet_file_size_in_mb(latest_path)
-
-            if latest_size_in_mb + ep_size_in_mb >= self.data_files_size_in_mb:
-                # Size limit is reached, prepare new parquet file
-                chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, self.chunks_size)
-
-            # Update the existing pandas dataframe with new row
-            df["meta/episodes/chunk_index"] = [chunk_idx]
-            df["meta/episodes/file_index"] = [file_idx]
-            df["dataset_from_index"] = [latest_ep["dataset_to_index"]]
-            df["dataset_to_index"] = [latest_ep["dataset_to_index"] + num_frames]
-
-            if latest_size_in_mb + ep_size_in_mb < self.data_files_size_in_mb:
-                # Size limit wasnt reached, concatenate latest dataframe with new one
-                latest_df = pd.read_parquet(latest_path)
-                df = pd.concat([latest_df, df], ignore_index=True)
-
-        # Write the resulting dataframe from RAM to disk
-        path = self.root / DEFAULT_EPISODES_PATH.format(chunk_index=chunk_idx, file_index=file_idx)
-        path.parent.mkdir(parents=True, exist_ok=True)
-        df.to_parquet(path, index=False)
-
-        # Update the Hugging Face dataset by reloading it.
-        # This process should be fast because only the latest Parquet file has been modified.
-        # Therefore, only this file needs to be converted to PyArrow; the rest is loaded from the PyArrow memory-mapped cache.
-        self.episodes = load_episodes(self.root)
+        task_dict = {
+            "task_index": task_index,
+            "task": task,
+        }
+        append_jsonlines(task_dict, self.root / TASKS_PATH)
 
     def save_episode(
         self,
@@ -326,28 +250,32 @@ class LeRobotDatasetMetadata:
         episode_length: int,
         episode_tasks: list[str],
         episode_stats: dict[str, dict],
-        episode_metadata: dict,
     ) -> None:
+        self.info["total_episodes"] += 1
+        self.info["total_frames"] += episode_length
+
+        chunk = self.get_episode_chunk(episode_index)
+        if chunk >= self.total_chunks:
+            self.info["total_chunks"] += 1
+
+        self.info["splits"] = {"train": f"0:{self.info['total_episodes']}"}
+        self.info["total_videos"] += len(self.video_keys)
+        if len(self.video_keys) > 0:
+            self.update_video_info()
+
+        write_info(self.info, self.root)
+
         episode_dict = {
             "episode_index": episode_index,
             "tasks": episode_tasks,
             "length": episode_length,
         }
-        episode_dict.update(episode_metadata)
-        episode_dict.update(flatten_dict({"stats": episode_stats}))
-        self._save_episode_metadata(episode_dict)
+        self.episodes[episode_index] = episode_dict
+        write_episode(episode_dict, self.root)
 
-        # Update info
-        self.info["total_episodes"] += 1
-        self.info["total_frames"] += episode_length
-        self.info["total_tasks"] = len(self.tasks)
-        self.info["splits"] = {"train": f"0:{self.info['total_episodes']}"}
-        if len(self.video_keys) > 0:
-            self.update_video_info()
-        write_info(self.info, self.root)
-
-        self.stats = aggregate_stats([self.stats, episode_stats]) if self.stats is not None else episode_stats
-        write_stats(self.stats, self.root)
+        self.episodes_stats[episode_index] = episode_stats
+        self.stats = aggregate_stats([self.stats, episode_stats]) if self.stats else episode_stats
+        write_episode_stats(episode_index, episode_stats, self.root)
 
     def update_video_info(self) -> None:
         """
@@ -412,9 +340,8 @@ class LeRobotDatasetMetadata:
 
             features = {**features, **DEFAULT_FEATURES}
 
-        obj.tasks = None
-        obj.episodes = None
-        obj.stats = None
+        obj.tasks, obj.task_to_task_index = {}, {}
+        obj.episodes_stats, obj.stats, obj.episodes = {}, {}, {}
         obj.info = create_empty_dataset_info(CODEBASE_VERSION, fps, robot_type, features, use_videos)
         if len(obj.video_keys) > 0 and not use_videos:
             raise ValueError()
@@ -535,8 +462,8 @@ class LeRobotDataset(torch.utils.data.Dataset):
             download_videos (bool, optional): Flag to download the videos. Note that when set to True but the
                 video files are already present on local disk, they won't be downloaded again. Defaults to
                 True.
-            video_backend (str | None, optional): Video backend to use for decoding videos. Defaults to torchcodec when available int the platform; otherwise, defaults to 'pyav'.
-                You can also use the 'pyav' decoder used by Torchvision, which used to be the default option, or 'video_reader' which is another decoder of Torchvision.
+            video_backend (str | None, optional): Video backend to use for decoding videos. There is currently
+                a single option which is the pyav decoder used by Torchvision. Defaults to pyav.
         """
         super().__init__()
         self.repo_id = repo_id
@@ -546,7 +473,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         self.episodes = episodes
         self.tolerance_s = tolerance_s
         self.revision = revision if revision else CODEBASE_VERSION
-        self.video_backend = video_backend if video_backend else get_safe_default_codec()
+        self.video_backend = video_backend if video_backend else "pyav"
         self.delta_indices = None
 
         # Unused attributes
@@ -559,17 +486,29 @@ class LeRobotDataset(torch.utils.data.Dataset):
         self.meta = LeRobotDatasetMetadata(
             self.repo_id, self.root, self.revision, force_cache_sync=force_cache_sync
         )
+        if self.episodes is not None and self.meta._version >= packaging.version.parse("v2.1"):
+            episodes_stats = [self.meta.episodes_stats[ep_idx] for ep_idx in self.episodes]
+            self.stats = aggregate_stats(episodes_stats)
 
         # Load actual data
         try:
             if force_cache_sync:
                 raise FileNotFoundError
+            assert all((self.root / fpath).is_file() for fpath in self.get_episodes_file_paths())
             self.hf_dataset = self.load_hf_dataset()
         except (AssertionError, FileNotFoundError, NotADirectoryError):
             self.revision = get_safe_version(self.repo_id, self.revision)
-            self.download(download_videos)
+            self.download_episodes(download_videos)
             self.hf_dataset = self.load_hf_dataset()
 
+        self.episode_data_index = get_episode_data_index(self.meta.episodes, self.episodes)
+
+        # Check timestamps
+        timestamps = torch.stack(self.hf_dataset["timestamp"]).numpy()
+        episode_indices = torch.stack(self.hf_dataset["episode_index"]).numpy()
+        ep_data_index_np = {k: t.numpy() for k, t in self.episode_data_index.items()}
+        check_timestamps_sync(timestamps, episode_indices, ep_data_index_np, self.fps, self.tolerance_s)
+
         # Setup delta_indices
         if self.delta_timestamps is not None:
             check_delta_timestamps(self.delta_timestamps, self.fps, self.tolerance_s)
@@ -645,7 +584,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
             ignore_patterns=ignore_patterns,
         )
 
-    def download(self, download_videos: bool = True) -> None:
+    def download_episodes(self, download_videos: bool = True) -> None:
         """Downloads the dataset from the given 'repo_id' at the provided version. If 'episodes' is given, this
         will only download those episodes (selected by their episode_index). If 'episodes' is None, the whole
         dataset will be downloaded. Thanks to the behavior of snapshot_download, if the files are already present
@@ -653,10 +592,11 @@ class LeRobotDataset(torch.utils.data.Dataset):
         """
         # TODO(rcadene, aliberts): implement faster transfer
         # https://huggingface.co/docs/huggingface_hub/en/guides/download#faster-downloads
-        ignore_patterns = None if download_videos else "videos/"
         files = None
+        ignore_patterns = None if download_videos else "videos/"
         if self.episodes is not None:
             files = self.get_episodes_file_paths()
+
         self.pull_from_repo(allow_patterns=files, ignore_patterns=ignore_patterns)
 
     def get_episodes_file_paths(self) -> list[Path]:
@@ -669,13 +609,19 @@ class LeRobotDataset(torch.utils.data.Dataset):
                 for ep_idx in episodes
             ]
             fpaths += video_files
-        # episodes are stored in the same files, so we return unique paths only
-        fpaths = list(set(fpaths))
+
         return fpaths
 
     def load_hf_dataset(self) -> datasets.Dataset:
         """hf_dataset contains all the observations, states, actions, rewards, etc."""
-        hf_dataset = load_nested_dataset(self.root / "data")
+        if self.episodes is None:
+            path = str(self.root / "data")
+            hf_dataset = load_dataset("parquet", data_dir=path, split="train")
+        else:
+            files = [str(self.root / self.meta.get_data_file_path(ep_idx)) for ep_idx in self.episodes]
+            hf_dataset = load_dataset("parquet", data_files=files, split="train")
+
+        # TODO(aliberts): hf_dataset.set_format("torch")
         hf_dataset.set_transform(hf_transform_to_torch)
         return hf_dataset
 
@@ -683,6 +629,8 @@ class LeRobotDataset(torch.utils.data.Dataset):
         features = get_hf_features_from_features(self.features)
         ft_dict = {col: [] for col in features}
         hf_dataset = datasets.Dataset.from_dict(ft_dict, features=features, split="train")
+
+        # TODO(aliberts): hf_dataset.set_format("torch")
         hf_dataset.set_transform(hf_transform_to_torch)
         return hf_dataset
 
@@ -714,16 +662,15 @@ class LeRobotDataset(torch.utils.data.Dataset):
             return get_hf_features_from_features(self.features)
 
     def _get_query_indices(self, idx: int, ep_idx: int) -> tuple[dict[str, list[int | bool]]]:
-        ep = self.meta.episodes[ep_idx]
-        ep_start = ep["dataset_from_index"]
-        ep_end = ep["dataset_to_index"]
+        ep_start = self.episode_data_index["from"][ep_idx]
+        ep_end = self.episode_data_index["to"][ep_idx]
         query_indices = {
-            key: [max(ep_start, min(ep_end - 1, idx + delta)) for delta in delta_idx]
+            key: [max(ep_start.item(), min(ep_end.item() - 1, idx + delta)) for delta in delta_idx]
             for key, delta_idx in self.delta_indices.items()
         }
         padding = {  # Pad values outside of current episode range
             f"{key}_is_pad": torch.BoolTensor(
-                [(idx + delta < ep_start) | (idx + delta >= ep_end) for delta in delta_idx]
+                [(idx + delta < ep_start.item()) | (idx + delta >= ep_end.item()) for delta in delta_idx]
             )
             for key, delta_idx in self.delta_indices.items()
         }
@@ -737,7 +684,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         query_timestamps = {}
         for key in self.meta.video_keys:
             if query_indices is not None and key in query_indices:
-                timestamps = self.hf_dataset[query_indices[key]]["timestamp"]
+                timestamps = self.hf_dataset.select(query_indices[key])["timestamp"]
                 query_timestamps[key] = torch.stack(timestamps).tolist()
             else:
                 query_timestamps[key] = [current_ts]
@@ -746,7 +693,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
 
     def _query_hf_dataset(self, query_indices: dict[str, list[int]]) -> dict:
         return {
-            key: torch.stack(self.hf_dataset[q_idx][key])
+            key: torch.stack(self.hf_dataset.select(q_idx)[key])
             for key, q_idx in query_indices.items()
             if key not in self.meta.video_keys
         }
@@ -757,17 +704,12 @@ class LeRobotDataset(torch.utils.data.Dataset):
         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.
         """
-        ep = self.meta.episodes[ep_idx]
         item = {}
         for vid_key, query_ts in query_timestamps.items():
-            # Episodes are stored sequentially on a single mp4 to reduce the number of files.
-            # Thus we load the start timestamp of the episode on this mp4 and
-            # shift the query timestamp accordingly.
-            from_timestamp = ep[f"videos/{vid_key}/from_timestamp"]
-            shifted_query_ts = [from_timestamp + ts for ts in query_ts]
-
             video_path = self.root / self.meta.get_video_file_path(ep_idx, vid_key)
-            frames = decode_video_frames(video_path, shifted_query_ts, self.tolerance_s, self.video_backend)
+            frames = decode_video_frames_torchvision(
+                video_path, query_ts, self.tolerance_s, self.video_backend
+            )
             item[vid_key] = frames.squeeze(0)
 
         return item
@@ -805,7 +747,8 @@ class LeRobotDataset(torch.utils.data.Dataset):
 
         # Add task as a string
         task_idx = item["task_index"].item()
-        item["task"] = self.meta.tasks.iloc[task_idx].name
+        item["task"] = self.meta.tasks[task_idx]
+
         return item
 
     def __repr__(self):
@@ -835,9 +778,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         )
         return self.root / fpath
 
-    def _get_image_file_dir(self, episode_index: int, image_key: str) -> Path:
-        return self._get_image_file_path(episode_index, image_key, frame_index=0).parent
-
     def _save_image(self, image: torch.Tensor | np.ndarray | PIL.Image.Image, fpath: Path) -> None:
         if self.image_writer is None:
             if isinstance(image, torch.Tensor):
@@ -916,8 +856,11 @@ class LeRobotDataset(torch.utils.data.Dataset):
         episode_buffer["index"] = np.arange(self.meta.total_frames, self.meta.total_frames + episode_length)
         episode_buffer["episode_index"] = np.full((episode_length,), episode_index)
 
-        # Update tasks and task indices with new tasks if any
-        self.meta.save_episode_tasks(episode_tasks)
+        # Add new tasks to the tasks dictionary
+        for task in episode_tasks:
+            task_index = self.meta.get_task_index(task)
+            if task_index is None:
+                self.meta.add_task(task)
 
         # Given tasks in natural language, find their corresponding task indices
         episode_buffer["task_index"] = np.array([self.meta.get_task_index(task) for task in tasks])
@@ -929,154 +872,51 @@ class LeRobotDataset(torch.utils.data.Dataset):
                 continue
             episode_buffer[key] = np.stack(episode_buffer[key])
 
-        # Wait for image writer to end, so that episode stats over images can be computed
         self._wait_image_writer()
+        self._save_episode_table(episode_buffer, episode_index)
         ep_stats = compute_episode_stats(episode_buffer, self.features)
 
-        ep_metadata = self._save_episode_data(episode_buffer)
-        for video_key in self.meta.video_keys:
-            ep_metadata.update(self._save_episode_video(video_key, episode_index))
+        if len(self.meta.video_keys) > 0:
+            video_paths = self.encode_episode_videos(episode_index)
+            for key in self.meta.video_keys:
+                episode_buffer[key] = video_paths[key]
 
-        # `meta.save_episode` need to be executed after encoding the videos
-        self.meta.save_episode(episode_index, episode_length, episode_tasks, ep_stats, ep_metadata)
+        # `meta.save_episode` be executed after encoding the videos
+        self.meta.save_episode(episode_index, episode_length, episode_tasks, ep_stats)
 
-        # TODO(rcadene): remove? there is only one episode in the episode buffer, no need for ep_data_index
-        # ep_data_index = get_episode_data_index(self.meta.episodes, [episode_index])
-        # ep_data_index_np = {k: t.numpy() for k, t in ep_data_index.items()}
-        # check_timestamps_sync(
-        #     episode_buffer["timestamp"],
-        #     episode_buffer["episode_index"],
-        #     ep_data_index_np,
-        #     self.fps,
-        #     self.tolerance_s,
-        # )
+        ep_data_index = get_episode_data_index(self.meta.episodes, [episode_index])
+        ep_data_index_np = {k: t.numpy() for k, t in ep_data_index.items()}
+        check_timestamps_sync(
+            episode_buffer["timestamp"],
+            episode_buffer["episode_index"],
+            ep_data_index_np,
+            self.fps,
+            self.tolerance_s,
+        )
+
+        video_files = list(self.root.rglob("*.mp4"))
+        assert len(video_files) == self.num_episodes * len(self.meta.video_keys)
+
+        parquet_files = list(self.root.rglob("*.parquet"))
+        assert len(parquet_files) == self.num_episodes
 
-        # TODO(rcadene): images are also deleted in clear_episode_buffer
         # delete images
         img_dir = self.root / "images"
         if img_dir.is_dir():
             shutil.rmtree(self.root / "images")
 
-        if not episode_data:
-            # Reset episode buffer
+        if not episode_data:  # Reset the buffer
             self.episode_buffer = self.create_episode_buffer()
 
-    def _save_episode_data(self, episode_buffer: dict) -> dict:
-        """Save episode data to a parquet file and update the Hugging Face dataset of frames data.
-
-        This function processes episodes data from a buffer, converts it into a Hugging Face dataset,
-        and saves it as a parquet file. It handles both the creation of new parquet files and the
-        updating of existing ones based on size constraints. After saving the data, it reloads
-        the Hugging Face dataset to ensure it is up-to-date.
-
-        Notes: We both need to update parquet files and HF dataset:
-        - `pandas` loads parquet file in RAM
-        - `datasets` relies on a memory mapping from pyarrow (no RAM). It either converts parquet files to a pyarrow cache on disk,
-          or loads directly from pyarrow cache.
-        """
-        # Convert buffer into HF Dataset
-        ep_dict = {key: episode_buffer[key] for key in self.hf_features}
-        ep_dataset = datasets.Dataset.from_dict(ep_dict, features=self.hf_features, split="train")
+    def _save_episode_table(self, episode_buffer: dict, episode_index: int) -> None:
+        episode_dict = {key: episode_buffer[key] for key in self.hf_features}
+        ep_dataset = datasets.Dataset.from_dict(episode_dict, features=self.hf_features, split="train")
         ep_dataset = embed_images(ep_dataset)
-        ep_size_in_mb = get_hf_dataset_size_in_mb(ep_dataset)
-        ep_num_frames = len(ep_dataset)
-        df = pd.DataFrame(ep_dataset)
-
-        if self.meta.episodes is None:
-            # Initialize indices and frame count for a new dataset made of the first episode data
-            chunk_idx, file_idx = 0, 0
-            latest_num_frames = 0
-        else:
-            # Retrieve information from the latest parquet file
-            latest_ep = self.meta.episodes[-1]
-            chunk_idx = latest_ep["data/chunk_index"]
-            file_idx = latest_ep["data/file_index"]
-
-            latest_path = self.root / self.meta.data_path.format(chunk_index=chunk_idx, file_index=file_idx)
-            latest_size_in_mb = get_parquet_file_size_in_mb(latest_path)
-            latest_num_frames = get_parquet_num_frames(latest_path)
-
-            # Determine if a new parquet file is needed
-            if latest_size_in_mb + ep_size_in_mb >= self.meta.data_files_size_in_mb:
-                # Size limit is reached, prepare new parquet file
-                chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, self.meta.chunks_size)
-                latest_num_frames = 0
-            else:
-                # Update the existing parquet file with new rows
-                latest_df = pd.read_parquet(latest_path)
-                df = pd.concat([latest_df, df], ignore_index=True)
-
-        # Write the resulting dataframe from RAM to disk
-        path = self.root / self.meta.data_path.format(chunk_index=chunk_idx, file_index=file_idx)
-        path.parent.mkdir(parents=True, exist_ok=True)
-        if len(self.meta.image_keys) > 0:
-            to_parquet_with_hf_images(df, path)
-        else:
-            df.to_parquet(path)
-
-        # Update the Hugging Face dataset by reloading it.
-        # This process should be fast because only the latest Parquet file has been modified.
-        # Therefore, only this file needs to be converted to PyArrow; the rest is loaded from the PyArrow memory-mapped cache.
-        self.hf_dataset = self.load_hf_dataset()
-
-        metadata = {
-            "data/chunk_index": chunk_idx,
-            "data/file_index": file_idx,
-            "dataset_from_index": latest_num_frames,
-            "dataset_to_index": latest_num_frames + ep_num_frames,
-        }
-        return metadata
-
-    def _save_episode_video(self, video_key: str, episode_index: int):
-        # Encode episode frames into a temporary video
-        ep_path = self._encode_temporary_episode_video(video_key, episode_index)
-        ep_size_in_mb = get_video_size_in_mb(ep_path)
-        ep_duration_in_s = get_video_duration_in_s(ep_path)
-
-        if self.meta.episodes is None:
-            # Initialize indices for a new dataset made of the first episode data
-            chunk_idx, file_idx = 0, 0
-            latest_duration_in_s = 0
-            new_path = self.root / self.meta.video_path.format(
-                video_key=video_key, chunk_index=chunk_idx, file_index=file_idx
-            )
-            new_path.parent.mkdir(parents=True, exist_ok=True)
-            shutil.move(str(ep_path), str(new_path))
-        else:
-            # Retrieve information from the latest video file
-            latest_ep = self.meta.episodes[-1]
-            chunk_idx = latest_ep[f"videos/{video_key}/chunk_index"]
-            file_idx = latest_ep[f"videos/{video_key}/file_index"]
-
-            latest_path = self.root / self.meta.video_path.format(
-                video_key=video_key, chunk_index=chunk_idx, file_index=file_idx
-            )
-            latest_size_in_mb = get_video_size_in_mb(latest_path)
-            latest_duration_in_s = get_video_duration_in_s(latest_path)
-
-            if latest_size_in_mb + ep_size_in_mb >= self.meta.video_files_size_in_mb:
-                # Move temporary episode video to a new video file in the dataset
-                chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, self.meta.chunks_size)
-                new_path = self.root / self.meta.video_path.format(
-                    video_key=video_key, chunk_index=chunk_idx, file_index=file_idx
-                )
-                new_path.parent.mkdir(parents=True, exist_ok=True)
-                shutil.move(str(ep_path), str(new_path))
-            else:
-                # Update latest video file
-                concat_video_files([latest_path, ep_path], self.root, video_key, chunk_idx, file_idx)
-
-        # Remove temporary directory
-        shutil.rmtree(str(ep_path.parent))
-
-        metadata = {
-            "episode_index": episode_index,
-            f"videos/{video_key}/chunk_index": chunk_idx,
-            f"videos/{video_key}/file_index": file_idx,
-            f"videos/{video_key}/from_timestamp": latest_duration_in_s,
-            f"videos/{video_key}/to_timestamp": latest_duration_in_s + ep_duration_in_s,
-        }
-        return metadata
+        self.hf_dataset = concatenate_datasets([self.hf_dataset, ep_dataset])
+        self.hf_dataset.set_transform(hf_transform_to_torch)
+        ep_data_path = self.root / self.meta.get_data_file_path(ep_index=episode_index)
+        ep_data_path.parent.mkdir(parents=True, exist_ok=True)
+        ep_dataset.to_parquet(ep_data_path)
 
     def clear_episode_buffer(self) -> None:
         episode_index = self.episode_buffer["episode_index"]
@@ -1116,16 +956,34 @@ class LeRobotDataset(torch.utils.data.Dataset):
         if self.image_writer is not None:
             self.image_writer.wait_until_done()
 
-    def _encode_temporary_episode_video(self, video_key: str, episode_index: int) -> dict:
+    def encode_videos(self) -> None:
         """
         Use ffmpeg to convert frames stored as png into mp4 videos.
         Note: `encode_video_frames` is a blocking call. Making it asynchronous shouldn't speedup encoding,
         since video encoding with ffmpeg is already using multithreading.
         """
-        temp_path = Path(tempfile.mkdtemp(dir=self.root)) / f"{video_key}_{episode_index:03d}.mp4"
-        img_dir = self._get_image_file_dir(episode_index, video_key)
-        encode_video_frames(img_dir, temp_path, self.fps, overwrite=True)
-        return temp_path
+        for ep_idx in range(self.meta.total_episodes):
+            self.encode_episode_videos(ep_idx)
+
+    def encode_episode_videos(self, episode_index: int) -> dict:
+        """
+        Use ffmpeg to convert frames stored as png into mp4 videos.
+        Note: `encode_video_frames` is a blocking call. Making it asynchronous shouldn't speedup encoding,
+        since video encoding with ffmpeg is already using multithreading.
+        """
+        video_paths = {}
+        for key in self.meta.video_keys:
+            video_path = self.root / self.meta.get_video_file_path(episode_index, key)
+            video_paths[key] = str(video_path)
+            if video_path.is_file():
+                # Skip if video is already encoded. Could be the case when resuming data recording.
+                continue
+            img_dir = self._get_image_file_path(
+                episode_index=episode_index, image_key=key, frame_index=0
+            ).parent
+            encode_video_frames(img_dir, video_path, self.fps, overwrite=True)
+
+        return video_paths
 
     @classmethod
     def create(
@@ -1170,7 +1028,8 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj.image_transforms = None
         obj.delta_timestamps = None
         obj.delta_indices = None
-        obj.video_backend = video_backend if video_backend is not None else get_safe_default_codec()
+        obj.episode_data_index = None
+        obj.video_backend = video_backend if video_backend is not None else "pyav"
         return obj
 
 
@@ -1195,7 +1054,7 @@ class MultiLeRobotDataset(torch.utils.data.Dataset):
         super().__init__()
         self.repo_ids = repo_ids
         self.root = Path(root) if root else HF_LEROBOT_HOME
-        self.tolerances_s = tolerances_s if tolerances_s else dict.fromkeys(repo_ids, 0.0001)
+        self.tolerances_s = tolerances_s if tolerances_s else {repo_id: 1e-4 for repo_id in repo_ids}
         # Construct the underlying datasets passing everything but `transform` and `delta_timestamps` which
         # are handled by this class.
         self._datasets = [

lerobot/common/datasets/online_buffer.py

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

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/mobile_cabinet.txt

@@ -0,0 +1,85 @@
+https://drive.google.com/file/d/1_SOJkgfP5yZyVjMhTt3nwhvyUjcnlI51/view?usp=drive_link
+https://drive.google.com/file/d/1rmgN8UUzph1qwJnzG1d-uOafodn-gLvb/view?usp=drive_link
+https://drive.google.com/file/d/1NYQ-XxsBVinB6dUoZmVWweT83367P3i2/view?usp=drive_link
+https://drive.google.com/file/d/1oAv_j74zxxCJieMG7r5Vl2BeHK1__3s3/view?usp=drive_link
+https://drive.google.com/file/d/1wFUJQROsrTJt64YRuIeExhFjr2wnK5uu/view?usp=drive_link
+https://drive.google.com/file/d/1KzL3Tt0Le7jVl58XVRUcmigmXjyiuhbK/view?usp=drive_link
+https://drive.google.com/file/d/1qy_YBladeHtianSSGtgAPSHtMin7msvf/view?usp=drive_link
+https://drive.google.com/file/d/1rA_F0V_qL_nyuC_0aBKCisF4-0TIkF2Y/view?usp=drive_link
+https://drive.google.com/file/d/1hw-8qMpz9VgSt62XoASqNRuPECpCwJQP/view?usp=drive_link
+https://drive.google.com/file/d/1BpHOl9rKMzdvNGka6js7C0s40hH6vnDA/view?usp=drive_link
+https://drive.google.com/file/d/1PazhkhiDnJ-OUMyDVDFxEZNKQQqHiNWS/view?usp=drive_link
+https://drive.google.com/file/d/1lZ665R6ATl57dypxH4dGJ2NSt6XYnbuz/view?usp=drive_link
+https://drive.google.com/file/d/1V9HzLaf-tlG15wUzT7KrTDCS_z1vi5NV/view?usp=drive_link
+https://drive.google.com/file/d/1aKauWiXoKqbNwn_2xs4MrmLlaNYlVNmO/view?usp=drive_link
+https://drive.google.com/file/d/1WVD5DFhriO1YmmOgiVHhacR6HWoTPxav/view?usp=drive_link
+https://drive.google.com/file/d/1_X43WgeBAsfkhH9EmpyPki8U9joMeAGC/view?usp=drive_link
+https://drive.google.com/file/d/1t8x0GqWoNKWtnBsB7_D40Z34nL9ak4kf/view?usp=drive_link
+https://drive.google.com/file/d/15V_f26WaKOXjKnq2T3HRWAmtQUi4lbu2/view?usp=drive_link
+https://drive.google.com/file/d/11VFIAsiSDsMOBANgrOcZBpKB9AFWnLy7/view?usp=drive_link
+https://drive.google.com/file/d/1M0NS7vVaxJv3FHnuRYtdwTFYF7We4LxP/view?usp=drive_link
+https://drive.google.com/file/d/1mR0OItTNqFnVLoczcyKYlm6drAy778lO/view?usp=drive_link
+https://drive.google.com/file/d/1NbVFWDQAh-z4JJ4D-Zw6Lps9kdvpqh2j/view?usp=drive_link
+https://drive.google.com/file/d/1JQoZGBzl4W3QG26-n39tefcGN0fDRMbB/view?usp=drive_link
+https://drive.google.com/file/d/1VBjHl-TvZpncopvasIP5G9gecbB2a5f6/view?usp=drive_link
+https://drive.google.com/file/d/1VzSf6zaB21nahm7MsPwroXbJ84NIwq0b/view?usp=drive_link
+https://drive.google.com/file/d/1OtNnfMEydNtZOcivs4k6E_uJSpf8PkGy/view?usp=drive_link
+https://drive.google.com/file/d/14nVvpvsrFr_03Pa_N7MKzwnRwibOUYM6/view?usp=drive_link
+https://drive.google.com/file/d/1M8li6duiO2r3lv_9HhF_XJn0oZUIEK5F/view?usp=drive_link
+https://drive.google.com/file/d/1Cpzea6fO14lxAaNfSBifqoa4ekhCiLD1/view?usp=drive_link
+https://drive.google.com/file/d/1mbxRTm5vlbsY9UJ0jfjM6j9D7kPJjBpG/view?usp=drive_link
+https://drive.google.com/file/d/1RXD1i6IfWsHRlCxVmG04h2h5Ycm_WwZN/view?usp=drive_link
+https://drive.google.com/file/d/1QFqFSwDGOk1BkgGmqgCcc2BRWnJ6R3MA/view?usp=drive_link
+https://drive.google.com/file/d/1bFqWR8DQM0ZUxxtS2bl-RANQvukeFLzp/view?usp=drive_link
+https://drive.google.com/file/d/1pR-rH3yNGoyPdD4hJ6-3lXQ-PstBx9du/view?usp=drive_link
+https://drive.google.com/file/d/107OAwLY-hva9HeQLIK7VCh-ytdDabVjr/view?usp=drive_link
+https://drive.google.com/file/d/1Tpl08QOaSZ37GTO4awFWSdD8wBR9xdlT/view?usp=drive_link
+https://drive.google.com/file/d/1MR164AOM-0S1T6RX8xKTV2IHyaCvpqAW/view?usp=drive_link
+https://drive.google.com/file/d/1_wknJfVnStIhJ82lU_QtcrwahsqYIsr8/view?usp=drive_link
+https://drive.google.com/file/d/1ZuEktWrbYkTx0l5pj3WiZ2CJrfbDOHNo/view?usp=drive_link
+https://drive.google.com/file/d/15G_10hkkkq6yxvyI5NGZirlF-RzduR2F/view?usp=drive_link
+https://drive.google.com/file/d/1DBKxg3ONqh7dhLuX6oh1Yyo2x383V1Hp/view?usp=drive_link
+https://drive.google.com/file/d/1B5iDBkTUr5vopDddV_fHud18SqAHhauS/view?usp=drive_link
+https://drive.google.com/file/d/1acwFV0eenRkki1QcjSKH5xqOtys-P3Pr/view?usp=drive_link
+https://drive.google.com/file/d/1S47BI83xyrh-FKXsvAQqer98Biu_p8XK/view?usp=drive_link
+https://drive.google.com/file/d/1JL6DmBZl3uyq9dyLfgSqtGF06e7E9JwM/view?usp=drive_link
+https://drive.google.com/file/d/16WvRS4Kjog8Pxgr0E3sGGnI01YwL9Uql/view?usp=drive_link
+https://drive.google.com/file/d/12ttGqL33IPWg0-s1SD44rr22M6LiSQBr/view?usp=drive_link
+https://drive.google.com/file/d/1OyZqqnldTU_DliRbr6x0C4a_iWPwIN7j/view?usp=drive_link
+https://drive.google.com/file/d/1oYk00IpLnR9fesLfD15Ebe7nVBffEbcS/view?usp=drive_link
+https://drive.google.com/file/d/1eyE2-MQduCEqCd-5_kl5zsoOEERAzpZD/view?usp=drive_link
+https://drive.google.com/file/d/1ir1Ya-vO0d97pfvbePlUeuKTTRc0qIMU/view?usp=drive_link
+https://drive.google.com/file/d/1hOi-JnqlMt47gVnLZHMTqeojyYVErohl/view?usp=drive_link
+https://drive.google.com/file/d/1NFFw5_PqigQ7xGqsL-MNq2B1r5yAscCf/view?usp=drive_link
+https://drive.google.com/file/d/1uftq1-Zlh8d2sNLWrlVcKYQUwZTD7o24/view?usp=drive_link
+https://drive.google.com/file/d/1-ax19dSLPacVgk000T-m3l4flPcg07pM/view?usp=drive_link
+https://drive.google.com/file/d/126y-lgn86-ZmCz8hooF1THKJGGObw3OB/view?usp=drive_link
+https://drive.google.com/file/d/1JiDniK0VmDIkk92AbBILb8J2Ba59PWML/view?usp=drive_link
+https://drive.google.com/file/d/1kr8nPIRljiU0R4J9SMgj80o1FPQxzu9z/view?usp=drive_link
+https://drive.google.com/file/d/1bbThWRij1pKBh_kFgV8FwK0sXtTHBoLX/view?usp=drive_link
+https://drive.google.com/file/d/1WenzDW6lxk1xkOFm-OiGFfc0ROskAuKU/view?usp=drive_link
+https://drive.google.com/file/d/1MiKRzuzUn1yN-k_6kPJJzIGy7dT-nnsD/view?usp=drive_link
+https://drive.google.com/file/d/17rRg2tcmB-gNhQ0KoZJQmNfyFeoij1jH/view?usp=drive_link
+https://drive.google.com/file/d/11mokBpvrY3ld6sY5WztREtJ1jgqfQV70/view?usp=drive_link
+https://drive.google.com/file/d/1Il_6IOx9NDp1bX_KHizJfBwzTufTmn86/view?usp=drive_link
+https://drive.google.com/file/d/1KswtJGsxJ7eeBDAmNA_aeLjOxcH6MIxa/view?usp=drive_link
+https://drive.google.com/file/d/1gzMhi5uWu4C3Y6WbQ3L-08V96GxTZrRR/view?usp=drive_link
+https://drive.google.com/file/d/1nRQFtaBxfUCYc2W90Qibh0kHCt6YQCfc/view?usp=drive_link
+https://drive.google.com/file/d/1vs-gyW-KheqHbUATwAhA2mmR9GOGw7f_/view?usp=drive_link
+https://drive.google.com/file/d/1MuxzGOA2fgLaHryq82KkQumtuRJGcUOC/view?usp=drive_link
+https://drive.google.com/file/d/1IIwxZnGlqrXLUXqG6yMO0r7uhCvhpk9e/view?usp=drive_link
+https://drive.google.com/file/d/1vE7XPyaFcXP4DtTY5Y9WKIt7zWgmX-Cr/view?usp=drive_link
+https://drive.google.com/file/d/1j-bIV09gr21RC3-x1N_pK4RPLV3fmWKz/view?usp=drive_link
+https://drive.google.com/file/d/1t3nW1rD3S-EL0Oymb5U7ZAj5UMkydkln/view?usp=drive_link
+https://drive.google.com/file/d/14hbfHCdMKtJZ41F9CQReMec2jeRFTOqR/view?usp=drive_link
+https://drive.google.com/file/d/1x-hUyOSne5BW0AzQ3W6_Pf4g5yXQWi9M/view?usp=drive_link
+https://drive.google.com/file/d/1sw9JqRg6E-3P84I3ZhzTrJMu0vuiaMmP/view?usp=drive_link
+https://drive.google.com/file/d/1LuqhQlL4MGZhB_6THmkovRxrlP26BbdC/view?usp=drive_link
+https://drive.google.com/file/d/15C5K6v_lkjnMSmUvVyqHQKwh2N166e7K/view?usp=drive_link
+https://drive.google.com/file/d/1ns_9eSsQeeoZ10nlbkLy8tu0GmJFSnkt/view?usp=drive_link
+https://drive.google.com/file/d/1NpzWJeK6CqjxzjIMYe6aYdX8xGsQwD4o/view?usp=drive_link
+https://drive.google.com/file/d/1NMLezwufKJ9_8xTc9KQThSzVVD71B9Ui/view?usp=drive_link
+https://drive.google.com/file/d/1aa71DCUqs6oXlIxX35jgsmsgm-NlDxPV/view?usp=drive_link
+https://drive.google.com/file/d/1UJzkIZzAL0j-D5YQBnoq7mHvttASy12O/view?usp=drive_link
+https://drive.google.com/file/d/1nPgx36HIJFb7oI94VbRzWjpPP2GANxzG/view?usp=drive_link
+https://drive.google.com/file/d/1NovAP-KVJjqcuvWy3d6G4ptGGAIDqcCx/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/mobile_chair.txt

@@ -0,0 +1,55 @@
+https://drive.google.com/file/d/11M3Ye0r5agMaaicPbVGD0q2Hb3rGklbb/view?usp=drive_link
+https://drive.google.com/file/d/1-tx7SvYYgSvXCvnf_EI2OVdwK-CkFY6S/view?usp=drive_link
+https://drive.google.com/file/d/1EWJunmOpMHaU1hE106wwpbkGYcjQXYAF/view?usp=drive_link
+https://drive.google.com/file/d/1IDn95Z7FSiCckrSENtGV4u3RyFHNQSDY/view?usp=drive_link
+https://drive.google.com/file/d/1CwzvWj1i7QOtqrZvsCZ6BdZaKNDfpN32/view?usp=drive_link
+https://drive.google.com/file/d/1HvAvlhm77nAD3Td24QPSeq8lw-Rl_aOh/view?usp=drive_link
+https://drive.google.com/file/d/1t-suKYOPhXH666RpAYNRp2QU_DOy3AeM/view?usp=drive_link
+https://drive.google.com/file/d/18xpKgWh7RWyjMN5PkLTOo-AxsAadAuRw/view?usp=drive_link
+https://drive.google.com/file/d/1oci5Eto-ztv-AQNz8EnwZveBIhxvk-xJ/view?usp=drive_link
+https://drive.google.com/file/d/1Y-t_4vxdE6NpHO0DLJR8f3mD0Q-Wj5-c/view?usp=drive_link
+https://drive.google.com/file/d/1lylRqbbbB8bgtpsBWMPACmHJreuKmllv/view?usp=drive_link
+https://drive.google.com/file/d/1yliSyMig_NXShWfQx6qyW7Ijf2Y5lFK6/view?usp=drive_link
+https://drive.google.com/file/d/1XXhwJsJbeb7KXAooGvJapnm9bjnGUmxS/view?usp=drive_link
+https://drive.google.com/file/d/1_xs1f3hW2JArKyvfF7UWubWjyROGTLs6/view?usp=drive_link
+https://drive.google.com/file/d/1WVEHpr6EqKCZbkHapQSTXJq4xE4SWFT-/view?usp=drive_link
+https://drive.google.com/file/d/1RqOHv9pEQGvW8NUA7ynffFmG999TL_Az/view?usp=drive_link
+https://drive.google.com/file/d/1cu5AgD2gh-uA3PFJmzxxzNaF3qOSlYY1/view?usp=drive_link
+https://drive.google.com/file/d/1SsrXqiPclNrnYToPZ9Uq-k3y0C4qdHT1/view?usp=drive_link
+https://drive.google.com/file/d/1-J7EXf0vjkLIfSqT8ICEsP6CTjzSLBop/view?usp=drive_link
+https://drive.google.com/file/d/11O7ewUmoZXfyyKjy_6B5RW4DpjICxqBT/view?usp=drive_link
+https://drive.google.com/file/d/1iic44kZoCsjNsfAz2cMstZ9-WQvAhblF/view?usp=drive_link
+https://drive.google.com/file/d/1yLV1lVX-2WnWQldGlnQZ0x7QBuDiVkL3/view?usp=drive_link
+https://drive.google.com/file/d/1Tybp9ru98TTbGn4eyROpUQwDFuALWXmk/view?usp=drive_link
+https://drive.google.com/file/d/13E9OTMiipVJByDs5-J19oWwAz7l94LTN/view?usp=drive_link
+https://drive.google.com/file/d/1EeTpJQdMSliw4JzSMtJ6CyTvVdexjM4M/view?usp=drive_link
+https://drive.google.com/file/d/1NHyNwoFqzeAu-1_PSpq5JfxaiD_xbpn9/view?usp=drive_link
+https://drive.google.com/file/d/1fJcS0phDp4xm_FyGaJ5wr9Pe4KqtHaxD/view?usp=drive_link
+https://drive.google.com/file/d/12AqrLUaewDPEcFRqPZeZFb_TQ0Lfi3At/view?usp=drive_link
+https://drive.google.com/file/d/1x_hd4Qsq1oJS-aj2t3qM7WbbV7KZj05b/view?usp=drive_link
+https://drive.google.com/file/d/14OUSUArmsB068hs6BuEIXQhI1Cyz8Sf0/view?usp=drive_link
+https://drive.google.com/file/d/16zlzh1T5zeUJQnFf382NXkFEKEnDub4O/view?usp=drive_link
+https://drive.google.com/file/d/1IbDltmN-NEFCNtr1TO4ILxEgQ94rtjWv/view?usp=drive_link
+https://drive.google.com/file/d/15gmlf8Gx9455pZ1AlqcCSwh3nDPxMzSr/view?usp=drive_link
+https://drive.google.com/file/d/1qHpRL1oZfIMo_vxnm8qfwQ-7l0BZIVva/view?usp=drive_link
+https://drive.google.com/file/d/1H1xskIgiFZivkYn23rMzH3xePGOh3VTC/view?usp=drive_link
+https://drive.google.com/file/d/1avls6Pv0kYiCMNVknbc1zQsgy64MUDMM/view?usp=drive_link
+https://drive.google.com/file/d/1MmWVgCj5khc8KMIifmt3EzF1o-CtPyyn/view?usp=drive_link
+https://drive.google.com/file/d/1U0kCc_xqW0WNppf4sbnK14euWKdPZtzB/view?usp=drive_link
+https://drive.google.com/file/d/16CaEyQscOuhLj23PEGDTL9DeyNkohkMn/view?usp=drive_link
+https://drive.google.com/file/d/1Iu8uM6UUJ0zW8tvN-9UiOe_4oSNzEutg/view?usp=drive_link
+https://drive.google.com/file/d/1UImqiBaIxCR-1DNJaZhHqeHhaySOtVIr/view?usp=drive_link
+https://drive.google.com/file/d/1VpU2V_leIoRIyv_lAvE7eLHBG8DxCTnp/view?usp=drive_link
+https://drive.google.com/file/d/1_Q8J27OT3Xby7QY6yHvIJauFRWEMxkRm/view?usp=drive_link
+https://drive.google.com/file/d/1bantmVo1L9Xz4tbiNw_a1UC2Z_HPO1wT/view?usp=drive_link
+https://drive.google.com/file/d/1IRIXMJMCBDkBjbaHvAlEiBogSvZ1jK_3/view?usp=drive_link
+https://drive.google.com/file/d/1mAHXKjiFbjwydypW2t5Lv8_H5x6nHegl/view?usp=drive_link
+https://drive.google.com/file/d/1SfyY796fLrBCMY39OcyuxZafqSCRZPZk/view?usp=drive_link
+https://drive.google.com/file/d/1X-44sZ8CcfzIskc0dvSx882o1yFhHaZB/view?usp=drive_link
+https://drive.google.com/file/d/1BOIWCCCk6DLD4Bmvc75ZbbLi9AQm-1ao/view?usp=drive_link
+https://drive.google.com/file/d/1RuyDtRE1kk76sw-wP8vx5SgLoPF3PA_H/view?usp=drive_link
+https://drive.google.com/file/d/1c4eoQiBbGuy3CTAQDUSkd84Ponh1roAQ/view?usp=drive_link
+https://drive.google.com/file/d/19PXB9z4Ljq6dsbf9TqcOrrP5SRbw2Tc_/view?usp=drive_link
+https://drive.google.com/file/d/1nn1VVZVoIXWdYDozR7XHXE4mPLQG80PQ/view?usp=drive_link
+https://drive.google.com/file/d/1MBdFGOKPV8GUhwoSsJ_Ky3qAMLM2Bv3K/view?usp=drive_link
+https://drive.google.com/file/d/1of3k_M-7Nh3I1TndcWedxK4ca9dn8Sc5/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/mobile_elevator.txt

@@ -0,0 +1,20 @@
+https://drive.google.com/file/d/12ctkOAdkCNGN1JLbZb5ww3XTBn2LFpGI/view?usp=drive_link
+https://drive.google.com/file/d/1G_Vd46_4fq6O64gHHjUbJX5Ld44ZZx0y/view?usp=drive_link
+https://drive.google.com/file/d/1uKgUy73B3xBogQAOUhfZjO0X5qZGsi2c/view?usp=drive_link
+https://drive.google.com/file/d/1fu9cIrfI-fE2LhdGUxbx7-8Ci_PF8Ypm/view?usp=drive_link
+https://drive.google.com/file/d/1Ygk9ZPJzx8xw2A9JF3NHbJ44TqnvSTQR/view?usp=drive_link
+https://drive.google.com/file/d/18m5xPuccNsEB20WPshm3zhxmXc6k63ED/view?usp=drive_link
+https://drive.google.com/file/d/1DiqqxC44rriviRQpqogcv0-EB-Y6nr9g/view?usp=drive_link
+https://drive.google.com/file/d/1qPdaoTVDizJXkfXLioWU7iJ8hqCXSyOQ/view?usp=drive_link
+https://drive.google.com/file/d/1Fj9kIA_mG7f67WFfACJEaZ7izcHG7vUm/view?usp=drive_link
+https://drive.google.com/file/d/1WpYehZnI2P7dUdJPfkE-ij1rqCnjZEbB/view?usp=drive_link
+https://drive.google.com/file/d/1_zwWkT4jPyzB38STWb6whlzsPzXmfA9r/view?usp=drive_link
+https://drive.google.com/file/d/1U6-J4I_fPlSFFGfhZPxS5_YzKXwXIZYp/view?usp=drive_link
+https://drive.google.com/file/d/1pRhxxcTfZp5tQo_EScvJUwfc3amiS6Vk/view?usp=drive_link
+https://drive.google.com/file/d/1lWLntqra83RlYU_gN7Vostnfydf6gutd/view?usp=drive_link
+https://drive.google.com/file/d/1vIBKo0x-NYEHV1FvRpco1lQMpRdAWAIL/view?usp=drive_link
+https://drive.google.com/file/d/1pdrLV3JTQou_XH0Aap61Ssf60iVKm1jJ/view?usp=drive_link
+https://drive.google.com/file/d/1QTsLoQ7SwmKdQHjBGVDaR2uTwfFwtrOf/view?usp=drive_link
+https://drive.google.com/file/d/1Gytai8M_12J36GY6L_TulEcOC-035jwS/view?usp=drive_link
+https://drive.google.com/file/d/14LJudNc629NT-i8xreXtzl27ce_DxOFJ/view?usp=drive_link
+https://drive.google.com/file/d/1sBvPCODbzxGAI0S3lgN5cSG9Go3lRi00/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/mobile_shrimp.txt

@@ -0,0 +1,18 @@
+https://drive.google.com/file/d/1MJn9GbC8p9lN4gC9KDMLEkTkP_gGpXj0/view?usp=drive_link
+https://drive.google.com/file/d/1-4LXgjl7ZCOgp-8GCJmFRD8OeqN5Jf7-/view?usp=drive_link
+https://drive.google.com/file/d/1Ho06Ce0SPbqU3juaMxNUwAt3zCRLGC8W/view?usp=drive_link
+https://drive.google.com/file/d/1ivHoj7_7olBSxH-Y8kqXEW7ttITK-45j/view?usp=drive_link
+https://drive.google.com/file/d/1qjY4hM_IvZ8cq2II_n9MeJbvyeuN4oBP/view?usp=drive_link
+https://drive.google.com/file/d/1rKVhO_f92-7sw13T8hTVrza3B9oAVgoy/view?usp=drive_link
+https://drive.google.com/file/d/1pcLPHO8fBkc1-CRa88tyQtEueE4xiXNi/view?usp=drive_link
+https://drive.google.com/file/d/1Vev_chCsIeEdvQ8poEYNsOJFGy_QU8kZ/view?usp=drive_link
+https://drive.google.com/file/d/1l5G4zpRkxSLCQjvGPYSN4zfCvVRQuzMz/view?usp=drive_link
+https://drive.google.com/file/d/14vgthE1eoakXkr2-DRw50E6lAqYOiUuE/view?usp=drive_link
+https://drive.google.com/file/d/17nPSmKKmgQ2B7zkzWrZYiLM3RBuFod82/view?usp=drive_link
+https://drive.google.com/file/d/1QcDsxplVvb_ID9BVrihl5FvlC-j7waXi/view?usp=drive_link
+https://drive.google.com/file/d/18pEejBpI-eEVaWAAjBCyC0vgbX3T1Esj/view?usp=drive_link
+https://drive.google.com/file/d/1H8eH6_IRODtEFT6WoM77ltR5OoOrqXmI/view?usp=drive_link
+https://drive.google.com/file/d/1IWlpFRZhoxyG4nS13CWK4leZVk5wbNx4/view?usp=drive_link
+https://drive.google.com/file/d/1PbZA8_OCGmMLxNP9xbkLRSChniL4uGxl/view?usp=drive_link
+https://drive.google.com/file/d/1p9XAdmG2f_WeflNO4DIJ_tr1rK6M9B4B/view?usp=drive_link
+https://drive.google.com/file/d/1nS59Et1cNAvKo3Y4SeSGRuZD5TvBbCF3/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/mobile_wash_pan.txt

@@ -0,0 +1 @@
+https://drive.google.com/drive/folders/1S8eFg98IaGAIKVZ8QFWG1bx4mHa-O204

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/mobile_wipe_wine.txt

@@ -0,0 +1,4 @@
+https://drive.google.com/drive/folders/1tC_g1AJ8lglBLY-fjsQrG6DMBa3Ucp-0
+https://drive.google.com/file/d/1fG_Yi2MJrFjiUVN3XoiWXLtTxHlwwaDv/view?usp=drive_link
+https://drive.google.com/file/d/1WX32VWfzzX3Blmd06DRxLwFbMJfVe7P4/view?usp=drive_link
+https://drive.google.com/file/d/18onsX3vXg3xkFwP5bVUCjdV4n9TRn0C9/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/sim_insertion_human.txt

@@ -0,0 +1,3 @@
+https://drive.google.com/drive/folders/1RgyD0JgTX30H4IM5XZn8I3zSV_mr8pyF
+https://drive.google.com/file/d/18Cudl6nikDtgRolea7je8iF_gGKzynOP/view?usp=drive_link
+https://drive.google.com/file/d/1C1kZYyROzs-PrLc0SkDgUgMi4-L3lauE/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/sim_insertion_scripted.txt

@@ -0,0 +1,3 @@
+https://drive.google.com/drive/folders/1TsojQQSXtHEoGnqgJ3gmpPQR2DPLtS2N
+https://drive.google.com/file/d/1wfMSZ24oOh5KR_0aaP3Cnu_c4ZCveduB/view?usp=drive_link
+https://drive.google.com/file/d/17EuCUWS6uCCr6yyNzpXdcdE-_TTNCKtf/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/sim_transfer_cube_human.txt

@@ -0,0 +1,3 @@
+https://drive.google.com/drive/folders/1sc-E4QYW7A0o23m1u2VWNGVq5smAsfCo
+https://drive.google.com/file/d/18smMymtr8tIxaNUQ61gW6dG50pt3MvGq/view?usp=drive_link
+https://drive.google.com/file/d/1Nk7l53d9sJoGDBKAOnNrExX5nLacATc6/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/sim_transfer_cube_scripted.txt

@@ -0,0 +1,3 @@
+https://drive.google.com/drive/folders/1aRyoOhQwxhyt1J8XgEig4s6kzaw__LXj
+https://drive.google.com/file/d/1pnGIOd-E4-rhz2P3VxpknMKRZCoKt6eI/view?usp=drive_link
+https://drive.google.com/file/d/1GKReZHrXU73NMiC5zKCq_UtqPVtYq8eo/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_battery.txt

@@ -0,0 +1,2 @@
+https://drive.google.com/drive/folders/19qS_n7vKgDcPeTMnvDHQ5-n73xEbJz5D
+https://drive.google.com/file/d/1oC31By0A2bsBeHyUwBdQw1z4ng6yi9Za/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_candy.txt

@@ -0,0 +1,2 @@
+https://drive.google.com/drive/folders/1m5rQ6UVH8Q9RQp_6c0CxkQ88-L-ScO7q
+https://drive.google.com/file/d/1wHz2qcmwcVG0C0CZ9MjQDQcmj4OY9_a3/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_coffee.txt

@@ -0,0 +1,2 @@
+https://drive.google.com/drive/folders/1seQGay470nGQ-knBI5TjsTr8iL9Qws5q
+https://drive.google.com/file/d/1T89hSX5U99wLGvGTE7yUBaQPOpyj6Sai/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_coffee_new.txt

@@ -0,0 +1,2 @@
+https://drive.google.com/drive/folders/1t3eDc5Rg0DveyRe8oTm6Dia_FYU5mXyf
+https://drive.google.com/file/d/1TXFaduTakvS0ZWJqKCX-HIvYglum_5CY/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_cups_open.txt

@@ -0,0 +1,2 @@
+https://drive.google.com/drive/folders/1Z9X3DNzd6LS0FFjQemNUMoMA5yk5VQOh
+https://drive.google.com/file/d/1Wlyc0vTkjXuWB6zbaVOWhEfD7BmPgUV_/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_fork_pick_up.txt

@@ -0,0 +1,53 @@
+https://drive.google.com/drive/folders/1DYgB4ifX4uIid9m9jnC0Zdz8Nf7ZC0fc
+https://drive.google.com/file/d/1Eb-NRNk_FmVleCbU_Ng5Y4dfcjTKN7Rv/view?usp=drive_link
+https://drive.google.com/file/d/1dkhjEADakT-44l9jf-nK4x89kr4yG_qb/view?usp=drive_link
+https://drive.google.com/file/d/14hDhgcZkVqNExGb4tIXpSjMshhqZETch/view?usp=drive_link
+https://drive.google.com/file/d/1zVMEHpHbuNyP5A_lYU7RPSLB-4V0yfZw/view?usp=drive_link
+https://drive.google.com/file/d/1JtgDjBvy7FnRpFzrx_foC3quorYQFAR-/view?usp=drive_link
+https://drive.google.com/file/d/1EHdneB6F-PP0dQlX8qPaXbxmKoBy_YwO/view?usp=drive_link
+https://drive.google.com/file/d/17Z0jjVBy1OPKREPu77_n_rQzorDiapji/view?usp=drive_link
+https://drive.google.com/file/d/1F4i23qPJ_qTf5jWjfLo4ARGJChznYWt3/view?usp=drive_link
+https://drive.google.com/file/d/1kZtXWM3uS0-rLblydBfJ0mMcVnMMXw9w/view?usp=drive_link
+https://drive.google.com/file/d/1mNODox87xFfY5Z_o5mcLsr8SHb39jDik/view?usp=drive_link
+https://drive.google.com/file/d/1Ob44VdmEUA93FKDECiRb5Ogz2xQg5IWp/view?usp=drive_link
+https://drive.google.com/file/d/1fdQLdjj3Cwv33R1wZhfrLz9Del8mqgHb/view?usp=drive_link
+https://drive.google.com/file/d/1Yu3L3ft21zP__XL8pCfhb788ZleuW1n5/view?usp=drive_link
+https://drive.google.com/file/d/1ozBBWXVZ9hXDh9ooHUNroHdYm8UDqnhJ/view?usp=drive_link
+https://drive.google.com/file/d/1o0TGqvfWw_Lunxb5ubKDS21Lr_WC0h75/view?usp=drive_link
+https://drive.google.com/file/d/1jZnd5eP5L6BH5l98BPN6OnoQx3fu8e9n/view?usp=drive_link
+https://drive.google.com/file/d/1S5sYbz8wcLYp0V67v13i4PRcBxodn4Hg/view?usp=drive_link
+https://drive.google.com/file/d/1rFeg_x6ftJYwPtBv34D3h2L2cpDLeR4G/view?usp=drive_link
+https://drive.google.com/file/d/1GvS3lcm4o6nm_scUk0XxKeVFNmzjucDZ/view?usp=drive_link
+https://drive.google.com/file/d/1-9i0riphC7NhhDahcQfD1QoBXP5gF90A/view?usp=drive_link
+https://drive.google.com/file/d/15p_IqGsMbKuvzMS872THAZr-3SBtb1Fr/view?usp=drive_link
+https://drive.google.com/file/d/1ToyYcBfJL8gbQn0q_59zPLsFmm7dmMJo/view?usp=drive_link
+https://drive.google.com/file/d/1e_7PNH7CYafE4pAebP7ZdI7XFbmEcy_i/view?usp=drive_link
+https://drive.google.com/file/d/1JoabvGVsIQdug2xOhUIhetEIyDM91y_Y/view?usp=drive_link
+https://drive.google.com/file/d/1kOMw1y0lmnVaCjwZICfzCsx6e0Z8MNGR/view?usp=drive_link
+https://drive.google.com/file/d/16it_wd1JOevUQTK2_CvF_pBACTgpIPgM/view?usp=drive_link
+https://drive.google.com/file/d/1IRcCj9HnJSfbyMgr5XEERGlEnWeZQwOc/view?usp=drive_link
+https://drive.google.com/file/d/1Z2dIJfq_S3liGmPN9Rphvkmucnmw7tlb/view?usp=drive_link
+https://drive.google.com/file/d/1J3NoAjzndGx9yNyaBOJHdNny1epzUoBt/view?usp=drive_link
+https://drive.google.com/file/d/18nOvxV1k8FSmBrhT4TPo2sKKSZXougyx/view?usp=drive_link
+https://drive.google.com/file/d/1CT8FxclafFMjSd7gCWVw3VSeryeiF04i/view?usp=drive_link
+https://drive.google.com/file/d/16M9KVqQMFfSsXfypK0bocFft8Nz3j2Rt/view?usp=drive_link
+https://drive.google.com/file/d/18QPVkw6bj6HW8LTPrQLWrrUX4R6RcF42/view?usp=drive_link
+https://drive.google.com/file/d/1hQTVtA5hBTE_StXpJafTZJ3tgt2VQQ_t/view?usp=drive_link
+https://drive.google.com/file/d/1Dn-d5g69H6EgAWgsFdrcbJKtz7ySsCQ8/view?usp=drive_link
+https://drive.google.com/file/d/13hMr16483P7ALYv73yMRUN37fJdVQM62/view?usp=drive_link
+https://drive.google.com/file/d/1848yN3XMN5zJMEgApt6KzrWgfRPfimtv/view?usp=drive_link
+https://drive.google.com/file/d/1oAD9kSnS0fTgj-CjD4u9VdZ5X67IOIMa/view?usp=drive_link
+https://drive.google.com/file/d/1ilzIWLCCG5b_KgF5s0wdN2I5-lFNpwC1/view?usp=drive_link
+https://drive.google.com/file/d/1rjsT2YBjnidxod1s9s-myAYz8boHr-WB/view?usp=drive_link
+https://drive.google.com/file/d/18Gg48HTub15bd8qzbhiCUufbVy0fbN5G/view?usp=drive_link
+https://drive.google.com/file/d/1WsSnQSqmMTVSRwrhT1Y-v782My2zcjLm/view?usp=drive_link
+https://drive.google.com/file/d/1ea9ZCvoyc-xqiFXgeDcA_mOWsw7VUuoi/view?usp=drive_link
+https://drive.google.com/file/d/1wv1v3-XhPgbNzp62BXbJTDzMPu2tlDUc/view?usp=drive_link
+https://drive.google.com/file/d/18-ikzt8LoZ83Gi3goKCELs4U4z8hrRoF/view?usp=drive_link
+https://drive.google.com/file/d/16Bjhp7JNCXkGuLvyNcZowAx3W-Y-15DV/view?usp=drive_link
+https://drive.google.com/file/d/1Gc-KRI-xwcp1fMR55ugbrLg_5y3SPde-/view?usp=drive_link
+https://drive.google.com/file/d/1oP72Q386Z4Sy5MMm-t5yNogIe5Van_9k/view?usp=drive_link
+https://drive.google.com/file/d/112T90eDUDVH-SyOV7UnZl5bscAH2hcfq/view?usp=drive_link
+https://drive.google.com/file/d/1y-uKOesRRhjgDtFbG_j65f4SGg0v8XDg/view?usp=drive_link
+https://drive.google.com/file/d/1LOP05OagoI3km-ZKQBrS204A85UVk7Ok/view?usp=drive_link
+https://drive.google.com/file/d/1QkHQKgasVzWsmdPvkXgGhWyQ84d93_Az/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_pingpong_test.txt

@@ -0,0 +1 @@
+https://drive.google.com/drive/folders/1Ut2cv6o6Pkfgg46DgwVUM7Z5PkNG8eJ-

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_pro_pencil.txt

@@ -0,0 +1 @@
+https://drive.google.com/drive/folders/1FqxPV0PgvgIu8XFjtvZSPSExuNcxVVAY

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_screw_driver.txt

@@ -0,0 +1,2 @@
+https://drive.google.com/drive/folders/1SKtG0ct9q0nVdYssJNMWSOjikcXliT58
+https://drive.google.com/file/d/1nchD21O30B3i3LDoqramo1zgW5YvpJIN/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_tape.txt

@@ -0,0 +1,2 @@
+https://drive.google.com/drive/folders/1_4DHf2cma0xsChLQFghwigX6Ukti5-zQ
+https://drive.google.com/file/d/1_8vS4hDNDgUQY-SmekrNaa7dF67QJYU-/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_thread_velcro.txt

@@ -0,0 +1,2 @@
+https://drive.google.com/drive/folders/1_4DHf2cma0xsChLQFghwigX6Ukti5-zQ
+https://drive.google.com/file/d/1_8vS4hDNDgUQY-SmekrNaa7dF67QJYU-/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_towel.txt

@@ -0,0 +1,2 @@
+https://drive.google.com/drive/folders/1fAD7vkyTGTFB_nGXIKofCU1U05oE3MFv
+https://drive.google.com/file/d/1XzyQ2B6LLvcurIonOpEu4nij2qwNWshH/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_vinh_cup.txt

@@ -0,0 +1,53 @@
+https://drive.google.com/drive/folders/13EQsVsnxT86K20QAoyE_YpsFbQ7fZQdu
+https://drive.google.com/file/d/1-W_JHghZG65FNTVhw1SXhtQrazdLL3Ue/view?usp=drive_link
+https://drive.google.com/file/d/1VwRJgdWUo-2nQaNM7Bs77-fsm8iwUxEo/view?usp=drive_link
+https://drive.google.com/file/d/1wFzGRo5iYA13WLi6IV1ry64RyahQBFio/view?usp=drive_link
+https://drive.google.com/file/d/1IKtQzQ-n-UTv64hYpReu2R4cqUvmNQqD/view?usp=drive_link
+https://drive.google.com/file/d/1GicVci9OiuuZZH79i5Mg7AtWod94MzwT/view?usp=drive_link
+https://drive.google.com/file/d/1JVnIoR7EIQp70T4eAf9RX65JcTrzsjQc/view?usp=drive_link
+https://drive.google.com/file/d/1W2xr4h23ucjPrc-mBEeqnACsfaImpc0p/view?usp=drive_link
+https://drive.google.com/file/d/10xj_0V7A07o3uCa7v5omUrTC0YlPW8H3/view?usp=drive_link
+https://drive.google.com/file/d/1FOc3EMaCy8Mb0_a7PuXLAwKwvxkbKmwU/view?usp=drive_link
+https://drive.google.com/file/d/143PgDXBcf2GQ0Q07ZPMVMfBgZDd5sLJG/view?usp=drive_link
+https://drive.google.com/file/d/1pE5Tyj0LlGbGWvUzuhixp86Ibu55Ez3I/view?usp=drive_link
+https://drive.google.com/file/d/141668b1VzX80ncrVJPzhkoAeIFB4MEK9/view?usp=drive_link
+https://drive.google.com/file/d/1bw12lo37p1ZvRvErHsll7cEYi2OxscvZ/view?usp=drive_link
+https://drive.google.com/file/d/1zfnMFvbgBjl6SzYhksbaOzfbwLrCN6tb/view?usp=drive_link
+https://drive.google.com/file/d/1-GIszA6mUJMaNB-tdh9r9skc77SWA0VX/view?usp=drive_link
+https://drive.google.com/file/d/1fTB0zWFYU6zh4IIUFT2zX_OkwYqmElwY/view?usp=drive_link
+https://drive.google.com/file/d/1gPIPNKGmrO9c7gKF7SP0SuUYbIBBq8z1/view?usp=drive_link
+https://drive.google.com/file/d/12JeJ-dQd5lYyn6PlDOGdE-ChVeiZ-Uv0/view?usp=drive_link
+https://drive.google.com/file/d/100_20cgCqerU6qoh3TfTbwLy9mlDAFEG/view?usp=drive_link
+https://drive.google.com/file/d/111oAGJ76ku_pYgbBoIdZAC1_XEQcPI__/view?usp=drive_link
+https://drive.google.com/file/d/1UhC8L-354ZQ2gblPFGI35EMsVwfpuKa0/view?usp=drive_link
+https://drive.google.com/file/d/1sIXQSgUR_xdrNtGrL6QGBnkLMKErsIp1/view?usp=drive_link
+https://drive.google.com/file/d/16Ax77bDSIXnsn4GFL8XYKKT1P6bPpfMd/view?usp=drive_link
+https://drive.google.com/file/d/1pgRVYwwVIsWq_qsWqZpe1UBzZfF5Fa9D/view?usp=drive_link
+https://drive.google.com/file/d/1jtimaZkWsY1P5gC2bbS64H_WCUU7HXN2/view?usp=drive_link
+https://drive.google.com/file/d/1N6Bh02P-RiTEgtx1YH1Db_X3TGpP-X_r/view?usp=drive_link
+https://drive.google.com/file/d/14Fy8EwJ8d9Vh97Yt1VOvUChSCrfIjBij/view?usp=drive_link
+https://drive.google.com/file/d/1IRuv42dvIMPuKhcMZmuXaBjJ-lPFOmQd/view?usp=drive_link
+https://drive.google.com/file/d/16XWzNY2D8ucVVn5geBgsVdhm3ppO4que/view?usp=drive_link
+https://drive.google.com/file/d/1xsVOoQgthK_L_SDrmq_JvQgUpAvPEAY8/view?usp=drive_link
+https://drive.google.com/file/d/1bZbw66DyEMvnJnzkdUUNbKjvNKg8KFYM/view?usp=drive_link
+https://drive.google.com/file/d/1CyTVkdrNGGpouCXr4CfhKbMzE6Ah3oo3/view?usp=drive_link
+https://drive.google.com/file/d/1hDRyeM-XEDpHXpptbT8LvNnlQUR3PWOh/view?usp=drive_link
+https://drive.google.com/file/d/1XhHWxbra8Iy5irQZ83IvxwaJqHq9x4s1/view?usp=drive_link
+https://drive.google.com/file/d/1haZcn6aM1o4JlmP9tJj3x2enrxiPaDSD/view?usp=drive_link
+https://drive.google.com/file/d/1ypDyuUTbljaBZ34f-t7lj3O_0bRmyX2n/view?usp=drive_link
+https://drive.google.com/file/d/1ILEEZo_tA9_ChIAprr2mPaNVKZi5vXsO/view?usp=drive_link
+https://drive.google.com/file/d/1U7nVYFaGE8vVTfLCW33D74xOjDcqfgyJ/view?usp=drive_link
+https://drive.google.com/file/d/1rZ93_rmCov5SMDxPkfM3qthcRELZrQX6/view?usp=drive_link
+https://drive.google.com/file/d/1mYO1b_csddtyE3qT6cwLiw-m2w2_1Lxh/view?usp=drive_link
+https://drive.google.com/file/d/1xz7Q5x2jikY8wJQjMRQpRws6AnfWlHm5/view?usp=drive_link
+https://drive.google.com/file/d/1OO8GaO-0FrSZRd1kxMYwBmubyiLOWnbl/view?usp=drive_link
+https://drive.google.com/file/d/1EXn4NVDmf-4_HCy34mYwT-vwK2CFI9ev/view?usp=drive_link
+https://drive.google.com/file/d/10hH70XhXRL9C5SnAG4toHtfHqfJUJo4H/view?usp=drive_link
+https://drive.google.com/file/d/18tiBcxea0guUai4lwsXQvt0q2LZ8ZnnJ/view?usp=drive_link
+https://drive.google.com/file/d/1Q8R8qv37vk5PQ5kQ2ibx6BFLOySD0VpX/view?usp=drive_link
+https://drive.google.com/file/d/17aNriHzjhdibCyuUjQoMFZqjybJZtggG/view?usp=drive_link
+https://drive.google.com/file/d/1LVjEYHSdeKm6CotU1QguIeNEPaIaFl_1/view?usp=drive_link
+https://drive.google.com/file/d/1ufAhE_EkgJ85slg2EW8aW_grOzE_Lmxd/view?usp=drive_link
+https://drive.google.com/file/d/1wtzLtXrkw9eXRGESTPIOlpl1tInu-b2m/view?usp=drive_link
+https://drive.google.com/file/d/1Mk5qvVtD_QHwGOUApRq76TUw2T5THu6f/view?usp=drive_link
+https://drive.google.com/file/d/1y1WQ3hboWVJ68KEYQQ3OhreGuaUpSgwc/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_vinh_cup_left.txt

@@ -0,0 +1,52 @@
+https://drive.google.com/drive/folders/1dxWh6YFZUDt6qXIoxgD9bla3CiFjZ11C
+https://drive.google.com/file/d/1hNBJN00SCAlOl0ZEgm7RRGbAGDjyBs0p/view?usp=drive_link
+https://drive.google.com/file/d/17He0CVwXGeoMmXg4SHKo-osNn7YPKVL7/view?usp=drive_link
+https://drive.google.com/file/d/1laNKUVID1x2CV6a2O2WQjwFewKu4lidL/view?usp=drive_link
+https://drive.google.com/file/d/1pNf36xbZJGRArYLmNAvRj5y6CoqdC6kB/view?usp=drive_link
+https://drive.google.com/file/d/1_4E1-y3JXk5I0ebycLYM70YDPK9g52gZ/view?usp=drive_link
+https://drive.google.com/file/d/1PHfzhGPdbolKyOpS3FnR2w7Q8zUlJXSk/view?usp=drive_link
+https://drive.google.com/file/d/17ls2PPN-Pi3tEuK059cwV2_iDT8aGhOO/view?usp=drive_link
+https://drive.google.com/file/d/1LWsg6PmCT00Kv_N_slrmcwKmQPGoBT3k/view?usp=drive_link
+https://drive.google.com/file/d/12LckrchoHTUVH7rxi8J7zD9dA19GXvoW/view?usp=drive_link
+https://drive.google.com/file/d/1VqrJKjAIkj5gtFXL69grdSeu9CyaqnSw/view?usp=drive_link
+https://drive.google.com/file/d/1g5rQYDBZvW-kUtYPeyF3qmd53v6k7kXu/view?usp=drive_link
+https://drive.google.com/file/d/10kUgaSJ0TS7teaG83G3Rf_DG4XGrBt6A/view?usp=drive_link
+https://drive.google.com/file/d/1je9XmneZQZvTma5adMJICUPDovW3ppei/view?usp=drive_link
+https://drive.google.com/file/d/1v28r6bedwZGbUPVVTVImXhK-42XdtGfj/view?usp=drive_link
+https://drive.google.com/file/d/1-TEEx9sGVvzMMaNXYfQMtY2JJ6cvl0dT/view?usp=drive_link
+https://drive.google.com/file/d/1YdBKdJFP9rJWBUX7qrOYL_gfUA8o6J9M/view?usp=drive_link
+https://drive.google.com/file/d/1X9vffwQHNUSKLXr2RlYNtbWDIFCIDfdF/view?usp=drive_link
+https://drive.google.com/file/d/11hqesqa5kvEe5FABUnZRcvmOhR373cYM/view?usp=drive_link
+https://drive.google.com/file/d/1ltTTECjEcbQPgS3UPRgMzaE2x9n6H7dC/view?usp=drive_link
+https://drive.google.com/file/d/1Zxqfa29JdwT-bfMpivi6IG2vz34d21dD/view?usp=drive_link
+https://drive.google.com/file/d/11LQlVxS5hz494dYUJ_PNRPx2NHIJbQns/view?usp=drive_link
+https://drive.google.com/file/d/1i1JhNtnZpO_E8rAv8gxBP3ZTZRvcvsZi/view?usp=drive_link
+https://drive.google.com/file/d/11jOXAr2EULUO4Qkm748634lg4UUFho5U/view?usp=drive_link
+https://drive.google.com/file/d/1rj67wur8DdB_Pipwx24bY43xu4X1eQ5e/view?usp=drive_link
+https://drive.google.com/file/d/15ZTm6lO6f_JQy_4SNfrOu3iPYn1Ro8mh/view?usp=drive_link
+https://drive.google.com/file/d/1q4gBtqWPJtCwXEvknGgN0WHGp7Vfn1b9/view?usp=drive_link
+https://drive.google.com/file/d/1t17keyre47AYqm8GgXiQ7EcvcUkeSiDQ/view?usp=drive_link
+https://drive.google.com/file/d/1OYUPGxtZgOF86Ng_BEOTXm_XOYpuQPsO/view?usp=drive_link
+https://drive.google.com/file/d/1cBjbGHi3dwWHtx6r9EQJi0JT_CE3LuHt/view?usp=drive_link
+https://drive.google.com/file/d/14qaMyF0mcbCB-fCYKNyo5_2NahSC6D5u/view?usp=drive_link
+https://drive.google.com/file/d/12FgX86eA7Y5co9ULBVK80XMsiKQSs-Ri/view?usp=drive_link
+https://drive.google.com/file/d/1yvoHWidf-jdBVw6qCCXOFfkVwKj_2hPk/view?usp=drive_link
+https://drive.google.com/file/d/1a2SugsSDlC8UtUrFzp-_KAwyZckQOvdQ/view?usp=drive_link
+https://drive.google.com/file/d/1l8pILBFSAosypWJMza2K09Vm7rug9axm/view?usp=drive_link
+https://drive.google.com/file/d/1hfPQ8dBCk97PnOhq6_MIISm3IEzcOxJG/view?usp=drive_link
+https://drive.google.com/file/d/1PPAUwlJCFKpms8cqF_k1v2_fCgDBOc3S/view?usp=drive_link
+https://drive.google.com/file/d/1lVKQZeqFfK3amEmLuFhYLUFQ2eyE8rOW/view?usp=drive_link
+https://drive.google.com/file/d/1K9iPMLfDowcIFoyzpvgn88dQ6x6kVwNG/view?usp=drive_link
+https://drive.google.com/file/d/1PNvMqG9tL7QxeLaYBGHiWYR6SYb5iIct/view?usp=drive_link
+https://drive.google.com/file/d/1xkRtzbvIkUsylx9hrFLGQsJn0h1EYu-5/view?usp=drive_link
+https://drive.google.com/file/d/1nxMRrJlSayjDIfr5CmHO1NzAw3COhsLi/view?usp=drive_link
+https://drive.google.com/file/d/1Qs3WEyMGrmagiHIkkFEueWNnJhkUeR1s/view?usp=drive_link
+https://drive.google.com/file/d/1D-G2_Q0SS3M8zyJbg_XzkF2ANPw1HTuX/view?usp=drive_link
+https://drive.google.com/file/d/1mdmJsDGO-YtJAOF_yPKl6lq4PJOIbQhT/view?usp=drive_link
+https://drive.google.com/file/d/11m9bwfop_sPmnQr_8amB6EEsrbAeG_z5/view?usp=drive_link
+https://drive.google.com/file/d/19tyYt5FMn5kru0g9o2nMJhKPnsDqkIZv/view?usp=drive_link
+https://drive.google.com/file/d/1XvTpUdsVTZ-vydvdYYmynbma--HfUGSl/view?usp=drive_link
+https://drive.google.com/file/d/1MO3hFu68J6NohTzr9aB_fY02VA6QSOqj/view?usp=drive_link
+https://drive.google.com/file/d/1Lh-UjwAk__04YOTWINF_QGVU8SjetVaY/view?usp=drive_link
+https://drive.google.com/file/d/1jkSOUwZV5GJ7rZlVeErjcu0DBQs8Np0d/view?usp=drive_link
+https://drive.google.com/file/d/1VIN1eLI-93WrVQwCjsv6XQr353DqqBYA/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_ziploc_slide.txt

@@ -0,0 +1,8 @@
+https://drive.google.com/drive/folders/1EgKar7rWBmTIRmeJYZciSwjZx3uP2mHO
+https://drive.google.com/file/d/12eYWQO15atK2hBjXhynPJd9MKAj_42pz/view?usp=drive_link
+https://drive.google.com/file/d/1Ul4oEeICJDjgfYTl4H1uaisTzVYIM6wd/view?usp=drive_link
+https://drive.google.com/file/d/1WSF-OG8lKSe2wVYCv5D1aJNipxpgddk-/view?usp=drive_link
+https://drive.google.com/file/d/1_ppD5j5sFh26aWW0JmhLzJMeNB-lCArk/view?usp=drive_link
+https://drive.google.com/file/d/1WUp846dgWXYhu4oJfhHxiU6YL_7N6s4W/view?usp=drive_link
+https://drive.google.com/file/d/1HRZNAIoAQw_uYiPwnBvtBioQoqiqoXdA/view?usp=drive_link
+https://drive.google.com/file/d/1hedGq-QDMnIn8GlXXBC3GiEJ_Y-LTxyt/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_diffusion_policy_replay_buffer.py

@@ -0,0 +1,634 @@
+#!/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.
+"""Helper code for loading PushT dataset from Diffusion Policy (https://diffusion-policy.cs.columbia.edu/)
+
+Copied from the original Diffusion Policy repository and used in our `download_and_upload_dataset.py` script.
+"""
+
+from __future__ import annotations
+
+import math
+import numbers
+import os
+from functools import cached_property
+
+import numcodecs
+import numpy as np
+import zarr
+
+
+def check_chunks_compatible(chunks: tuple, shape: tuple):
+    assert len(shape) == len(chunks)
+    for c in chunks:
+        assert isinstance(c, numbers.Integral)
+        assert c > 0
+
+
+def rechunk_recompress_array(group, name, chunks=None, chunk_length=None, compressor=None, tmp_key="_temp"):
+    old_arr = group[name]
+    if chunks is None:
+        chunks = (chunk_length,) + old_arr.chunks[1:] if chunk_length is not None else old_arr.chunks
+    check_chunks_compatible(chunks, old_arr.shape)
+
+    if compressor is None:
+        compressor = old_arr.compressor
+
+    if (chunks == old_arr.chunks) and (compressor == old_arr.compressor):
+        # no change
+        return old_arr
+
+    # rechunk recompress
+    group.move(name, tmp_key)
+    old_arr = group[tmp_key]
+    n_copied, n_skipped, n_bytes_copied = zarr.copy(
+        source=old_arr,
+        dest=group,
+        name=name,
+        chunks=chunks,
+        compressor=compressor,
+    )
+    del group[tmp_key]
+    arr = group[name]
+    return arr
+
+
+def get_optimal_chunks(shape, dtype, target_chunk_bytes=2e6, max_chunk_length=None):
+    """
+    Common shapes
+    T,D
+    T,N,D
+    T,H,W,C
+    T,N,H,W,C
+    """
+    itemsize = np.dtype(dtype).itemsize
+    # reversed
+    rshape = list(shape[::-1])
+    if max_chunk_length is not None:
+        rshape[-1] = int(max_chunk_length)
+    split_idx = len(shape) - 1
+    for i in range(len(shape) - 1):
+        this_chunk_bytes = itemsize * np.prod(rshape[:i])
+        next_chunk_bytes = itemsize * np.prod(rshape[: i + 1])
+        if this_chunk_bytes <= target_chunk_bytes and next_chunk_bytes > target_chunk_bytes:
+            split_idx = i
+
+    rchunks = rshape[:split_idx]
+    item_chunk_bytes = itemsize * np.prod(rshape[:split_idx])
+    this_max_chunk_length = rshape[split_idx]
+    next_chunk_length = min(this_max_chunk_length, math.ceil(target_chunk_bytes / item_chunk_bytes))
+    rchunks.append(next_chunk_length)
+    len_diff = len(shape) - len(rchunks)
+    rchunks.extend([1] * len_diff)
+    chunks = tuple(rchunks[::-1])
+    # print(np.prod(chunks) * itemsize / target_chunk_bytes)
+    return chunks
+
+
+class ReplayBuffer:
+    """
+    Zarr-based temporal datastructure.
+    Assumes first dimension to be time. Only chunk in time dimension.
+    """
+
+    def __init__(self, root: zarr.Group | dict[str, dict]):
+        """
+        Dummy constructor. Use copy_from* and create_from* class methods instead.
+        """
+        assert "data" in root
+        assert "meta" in root
+        assert "episode_ends" in root["meta"]
+        for value in root["data"].values():
+            assert value.shape[0] == root["meta"]["episode_ends"][-1]
+        self.root = root
+
+    # ============= create constructors ===============
+    @classmethod
+    def create_empty_zarr(cls, storage=None, root=None):
+        if root is None:
+            if storage is None:
+                storage = zarr.MemoryStore()
+            root = zarr.group(store=storage)
+        root.require_group("data", overwrite=False)
+        meta = root.require_group("meta", overwrite=False)
+        if "episode_ends" not in meta:
+            meta.zeros("episode_ends", shape=(0,), dtype=np.int64, compressor=None, overwrite=False)
+        return cls(root=root)
+
+    @classmethod
+    def create_empty_numpy(cls):
+        root = {"data": {}, "meta": {"episode_ends": np.zeros((0,), dtype=np.int64)}}
+        return cls(root=root)
+
+    @classmethod
+    def create_from_group(cls, group, **kwargs):
+        if "data" not in group:
+            # create from stratch
+            buffer = cls.create_empty_zarr(root=group, **kwargs)
+        else:
+            # already exist
+            buffer = cls(root=group, **kwargs)
+        return buffer
+
+    @classmethod
+    def create_from_path(cls, zarr_path, mode="r", **kwargs):
+        """
+        Open a on-disk zarr directly (for dataset larger than memory).
+        Slower.
+        """
+        group = zarr.open(os.path.expanduser(zarr_path), mode)
+        return cls.create_from_group(group, **kwargs)
+
+    # ============= copy constructors ===============
+    @classmethod
+    def copy_from_store(
+        cls,
+        src_store,
+        store=None,
+        keys=None,
+        chunks: dict[str, tuple] | None = None,
+        compressors: dict | str | numcodecs.abc.Codec | None = None,
+        if_exists="replace",
+        **kwargs,
+    ):
+        """
+        Load to memory.
+        """
+        src_root = zarr.group(src_store)
+        if chunks is None:
+            chunks = {}
+        if compressors is None:
+            compressors = {}
+        root = None
+        if store is None:
+            # numpy backend
+            meta = {}
+            for key, value in src_root["meta"].items():
+                if len(value.shape) == 0:
+                    meta[key] = np.array(value)
+                else:
+                    meta[key] = value[:]
+
+            if keys is None:
+                keys = src_root["data"].keys()
+            data = {}
+            for key in keys:
+                arr = src_root["data"][key]
+                data[key] = arr[:]
+
+            root = {"meta": meta, "data": data}
+        else:
+            root = zarr.group(store=store)
+            # copy without recompression
+            n_copied, n_skipped, n_bytes_copied = zarr.copy_store(
+                source=src_store, dest=store, source_path="/meta", dest_path="/meta", if_exists=if_exists
+            )
+            data_group = root.create_group("data", overwrite=True)
+            if keys is None:
+                keys = src_root["data"].keys()
+            for key in keys:
+                value = src_root["data"][key]
+                cks = cls._resolve_array_chunks(chunks=chunks, key=key, array=value)
+                cpr = cls._resolve_array_compressor(compressors=compressors, key=key, array=value)
+                if cks == value.chunks and cpr == value.compressor:
+                    # copy without recompression
+                    this_path = "/data/" + key
+                    n_copied, n_skipped, n_bytes_copied = zarr.copy_store(
+                        source=src_store,
+                        dest=store,
+                        source_path=this_path,
+                        dest_path=this_path,
+                        if_exists=if_exists,
+                    )
+                else:
+                    # copy with recompression
+                    n_copied, n_skipped, n_bytes_copied = zarr.copy(
+                        source=value,
+                        dest=data_group,
+                        name=key,
+                        chunks=cks,
+                        compressor=cpr,
+                        if_exists=if_exists,
+                    )
+        buffer = cls(root=root)
+        return buffer
+
+    @classmethod
+    def copy_from_path(
+        cls,
+        zarr_path,
+        backend=None,
+        store=None,
+        keys=None,
+        chunks: dict[str, tuple] | None = None,
+        compressors: dict | str | numcodecs.abc.Codec | None = None,
+        if_exists="replace",
+        **kwargs,
+    ):
+        """
+        Copy a on-disk zarr to in-memory compressed.
+        Recommended
+        """
+        if chunks is None:
+            chunks = {}
+        if compressors is None:
+            compressors = {}
+        if backend == "numpy":
+            print("backend argument is deprecated!")
+            store = None
+        group = zarr.open(os.path.expanduser(zarr_path), "r")
+        return cls.copy_from_store(
+            src_store=group.store,
+            store=store,
+            keys=keys,
+            chunks=chunks,
+            compressors=compressors,
+            if_exists=if_exists,
+            **kwargs,
+        )
+
+    # ============= save methods ===============
+    def save_to_store(
+        self,
+        store,
+        chunks: dict[str, tuple] | None = None,
+        compressors: str | numcodecs.abc.Codec | dict | None = None,
+        if_exists="replace",
+        **kwargs,
+    ):
+        root = zarr.group(store)
+        if chunks is None:
+            chunks = {}
+        if compressors is None:
+            compressors = {}
+        if self.backend == "zarr":
+            # recompression free copy
+            n_copied, n_skipped, n_bytes_copied = zarr.copy_store(
+                source=self.root.store,
+                dest=store,
+                source_path="/meta",
+                dest_path="/meta",
+                if_exists=if_exists,
+            )
+        else:
+            meta_group = root.create_group("meta", overwrite=True)
+            # save meta, no chunking
+            for key, value in self.root["meta"].items():
+                _ = meta_group.array(name=key, data=value, shape=value.shape, chunks=value.shape)
+
+        # save data, chunk
+        data_group = root.create_group("data", overwrite=True)
+        for key, value in self.root["data"].items():
+            cks = self._resolve_array_chunks(chunks=chunks, key=key, array=value)
+            cpr = self._resolve_array_compressor(compressors=compressors, key=key, array=value)
+            if isinstance(value, zarr.Array):
+                if cks == value.chunks and cpr == value.compressor:
+                    # copy without recompression
+                    this_path = "/data/" + key
+                    n_copied, n_skipped, n_bytes_copied = zarr.copy_store(
+                        source=self.root.store,
+                        dest=store,
+                        source_path=this_path,
+                        dest_path=this_path,
+                        if_exists=if_exists,
+                    )
+                else:
+                    # copy with recompression
+                    n_copied, n_skipped, n_bytes_copied = zarr.copy(
+                        source=value,
+                        dest=data_group,
+                        name=key,
+                        chunks=cks,
+                        compressor=cpr,
+                        if_exists=if_exists,
+                    )
+            else:
+                # numpy
+                _ = data_group.array(name=key, data=value, chunks=cks, compressor=cpr)
+        return store
+
+    def save_to_path(
+        self,
+        zarr_path,
+        chunks: dict[str, tuple] | None = None,
+        compressors: str | numcodecs.abc.Codec | dict | None = None,
+        if_exists="replace",
+        **kwargs,
+    ):
+        if chunks is None:
+            chunks = {}
+        if compressors is None:
+            compressors = {}
+        store = zarr.DirectoryStore(os.path.expanduser(zarr_path))
+        return self.save_to_store(
+            store, chunks=chunks, compressors=compressors, if_exists=if_exists, **kwargs
+        )
+
+    @staticmethod
+    def resolve_compressor(compressor="default"):
+        if compressor == "default":
+            compressor = numcodecs.Blosc(cname="lz4", clevel=5, shuffle=numcodecs.Blosc.NOSHUFFLE)
+        elif compressor == "disk":
+            compressor = numcodecs.Blosc("zstd", clevel=5, shuffle=numcodecs.Blosc.BITSHUFFLE)
+        return compressor
+
+    @classmethod
+    def _resolve_array_compressor(cls, compressors: dict | str | numcodecs.abc.Codec, key, array):
+        # allows compressor to be explicitly set to None
+        cpr = "nil"
+        if isinstance(compressors, dict):
+            if key in compressors:
+                cpr = cls.resolve_compressor(compressors[key])
+            elif isinstance(array, zarr.Array):
+                cpr = array.compressor
+        else:
+            cpr = cls.resolve_compressor(compressors)
+        # backup default
+        if cpr == "nil":
+            cpr = cls.resolve_compressor("default")
+        return cpr
+
+    @classmethod
+    def _resolve_array_chunks(cls, chunks: dict | tuple, key, array):
+        cks = None
+        if isinstance(chunks, dict):
+            if key in chunks:
+                cks = chunks[key]
+            elif isinstance(array, zarr.Array):
+                cks = array.chunks
+        elif isinstance(chunks, tuple):
+            cks = chunks
+        else:
+            raise TypeError(f"Unsupported chunks type {type(chunks)}")
+        # backup default
+        if cks is None:
+            cks = get_optimal_chunks(shape=array.shape, dtype=array.dtype)
+        # check
+        check_chunks_compatible(chunks=cks, shape=array.shape)
+        return cks
+
+    # ============= properties =================
+    @cached_property
+    def data(self):
+        return self.root["data"]
+
+    @cached_property
+    def meta(self):
+        return self.root["meta"]
+
+    def update_meta(self, data):
+        # sanitize data
+        np_data = {}
+        for key, value in data.items():
+            if isinstance(value, np.ndarray):
+                np_data[key] = value
+            else:
+                arr = np.array(value)
+                if arr.dtype == object:
+                    raise TypeError(f"Invalid value type {type(value)}")
+                np_data[key] = arr
+
+        meta_group = self.meta
+        if self.backend == "zarr":
+            for key, value in np_data.items():
+                _ = meta_group.array(
+                    name=key, data=value, shape=value.shape, chunks=value.shape, overwrite=True
+                )
+        else:
+            meta_group.update(np_data)
+
+        return meta_group
+
+    @property
+    def episode_ends(self):
+        return self.meta["episode_ends"]
+
+    def get_episode_idxs(self):
+        import numba
+
+        numba.jit(nopython=True)
+
+        def _get_episode_idxs(episode_ends):
+            result = np.zeros((episode_ends[-1],), dtype=np.int64)
+            for i in range(len(episode_ends)):
+                start = 0
+                if i > 0:
+                    start = episode_ends[i - 1]
+                end = episode_ends[i]
+                for idx in range(start, end):
+                    result[idx] = i
+            return result
+
+        return _get_episode_idxs(self.episode_ends)
+
+    @property
+    def backend(self):
+        backend = "numpy"
+        if isinstance(self.root, zarr.Group):
+            backend = "zarr"
+        return backend
+
+    # =========== dict-like API ==============
+    def __repr__(self) -> str:
+        if self.backend == "zarr":
+            return str(self.root.tree())
+        else:
+            return super().__repr__()
+
+    def keys(self):
+        return self.data.keys()
+
+    def values(self):
+        return self.data.values()
+
+    def items(self):
+        return self.data.items()
+
+    def __getitem__(self, key):
+        return self.data[key]
+
+    def __contains__(self, key):
+        return key in self.data
+
+    # =========== our API ==============
+    @property
+    def n_steps(self):
+        if len(self.episode_ends) == 0:
+            return 0
+        return self.episode_ends[-1]
+
+    @property
+    def n_episodes(self):
+        return len(self.episode_ends)
+
+    @property
+    def chunk_size(self):
+        if self.backend == "zarr":
+            return next(iter(self.data.arrays()))[-1].chunks[0]
+        return None
+
+    @property
+    def episode_lengths(self):
+        ends = self.episode_ends[:]
+        ends = np.insert(ends, 0, 0)
+        lengths = np.diff(ends)
+        return lengths
+
+    def add_episode(
+        self,
+        data: dict[str, np.ndarray],
+        chunks: dict[str, tuple] | None = None,
+        compressors: str | numcodecs.abc.Codec | dict | None = None,
+    ):
+        if chunks is None:
+            chunks = {}
+        if compressors is None:
+            compressors = {}
+        assert len(data) > 0
+        is_zarr = self.backend == "zarr"
+
+        curr_len = self.n_steps
+        episode_length = None
+        for value in data.values():
+            assert len(value.shape) >= 1
+            if episode_length is None:
+                episode_length = len(value)
+            else:
+                assert episode_length == len(value)
+        new_len = curr_len + episode_length
+
+        for key, value in data.items():
+            new_shape = (new_len,) + value.shape[1:]
+            # create array
+            if key not in self.data:
+                if is_zarr:
+                    cks = self._resolve_array_chunks(chunks=chunks, key=key, array=value)
+                    cpr = self._resolve_array_compressor(compressors=compressors, key=key, array=value)
+                    arr = self.data.zeros(
+                        name=key, shape=new_shape, chunks=cks, dtype=value.dtype, compressor=cpr
+                    )
+                else:
+                    # copy data to prevent modify
+                    arr = np.zeros(shape=new_shape, dtype=value.dtype)
+                    self.data[key] = arr
+            else:
+                arr = self.data[key]
+                assert value.shape[1:] == arr.shape[1:]
+                # same method for both zarr and numpy
+                if is_zarr:
+                    arr.resize(new_shape)
+                else:
+                    arr.resize(new_shape, refcheck=False)
+            # copy data
+            arr[-value.shape[0] :] = value
+
+        # append to episode ends
+        episode_ends = self.episode_ends
+        if is_zarr:
+            episode_ends.resize(episode_ends.shape[0] + 1)
+        else:
+            episode_ends.resize(episode_ends.shape[0] + 1, refcheck=False)
+        episode_ends[-1] = new_len
+
+        # rechunk
+        if is_zarr and episode_ends.chunks[0] < episode_ends.shape[0]:
+            rechunk_recompress_array(self.meta, "episode_ends", chunk_length=int(episode_ends.shape[0] * 1.5))
+
+    def drop_episode(self):
+        is_zarr = self.backend == "zarr"
+        episode_ends = self.episode_ends[:].copy()
+        assert len(episode_ends) > 0
+        start_idx = 0
+        if len(episode_ends) > 1:
+            start_idx = episode_ends[-2]
+        for value in self.data.values():
+            new_shape = (start_idx,) + value.shape[1:]
+            if is_zarr:
+                value.resize(new_shape)
+            else:
+                value.resize(new_shape, refcheck=False)
+        if is_zarr:
+            self.episode_ends.resize(len(episode_ends) - 1)
+        else:
+            self.episode_ends.resize(len(episode_ends) - 1, refcheck=False)
+
+    def pop_episode(self):
+        assert self.n_episodes > 0
+        episode = self.get_episode(self.n_episodes - 1, copy=True)
+        self.drop_episode()
+        return episode
+
+    def extend(self, data):
+        self.add_episode(data)
+
+    def get_episode(self, idx, copy=False):
+        idx = list(range(len(self.episode_ends)))[idx]
+        start_idx = 0
+        if idx > 0:
+            start_idx = self.episode_ends[idx - 1]
+        end_idx = self.episode_ends[idx]
+        result = self.get_steps_slice(start_idx, end_idx, copy=copy)
+        return result
+
+    def get_episode_slice(self, idx):
+        start_idx = 0
+        if idx > 0:
+            start_idx = self.episode_ends[idx - 1]
+        end_idx = self.episode_ends[idx]
+        return slice(start_idx, end_idx)
+
+    def get_steps_slice(self, start, stop, step=None, copy=False):
+        _slice = slice(start, stop, step)
+
+        result = {}
+        for key, value in self.data.items():
+            x = value[_slice]
+            if copy and isinstance(value, np.ndarray):
+                x = x.copy()
+            result[key] = x
+        return result
+
+    # =========== chunking =============
+    def get_chunks(self) -> dict:
+        assert self.backend == "zarr"
+        chunks = {}
+        for key, value in self.data.items():
+            chunks[key] = value.chunks
+        return chunks
+
+    def set_chunks(self, chunks: dict):
+        assert self.backend == "zarr"
+        for key, value in chunks.items():
+            if key in self.data:
+                arr = self.data[key]
+                if value != arr.chunks:
+                    check_chunks_compatible(chunks=value, shape=arr.shape)
+                    rechunk_recompress_array(self.data, key, chunks=value)
+
+    def get_compressors(self) -> dict:
+        assert self.backend == "zarr"
+        compressors = {}
+        for key, value in self.data.items():
+            compressors[key] = value.compressor
+        return compressors
+
+    def set_compressors(self, compressors: dict):
+        assert self.backend == "zarr"
+        for key, value in compressors.items():
+            if key in self.data:
+                arr = self.data[key]
+                compressor = self.resolve_compressor(value)
+                if compressor != arr.compressor:
+                    rechunk_recompress_array(self.data, key, compressor=compressor)

lerobot/common/datasets/push_dataset_to_hub/_download_raw.py

@@ -0,0 +1,202 @@
+#!/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 file contains download scripts for raw datasets.
+
+Example of usage:
+```
+python lerobot/common/datasets/push_dataset_to_hub/_download_raw.py \
+--raw-dir data/lerobot-raw/pusht_raw \
+--repo-id lerobot-raw/pusht_raw
+```
+"""
+
+import argparse
+import logging
+import warnings
+from pathlib import Path
+
+from huggingface_hub import snapshot_download
+
+from lerobot.common.datasets.push_dataset_to_hub.utils import check_repo_id
+
+# {raw_repo_id: raw_format}
+AVAILABLE_RAW_REPO_IDS = {
+    "lerobot-raw/aloha_mobile_cabinet_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_mobile_chair_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_mobile_elevator_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_mobile_shrimp_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_mobile_wash_pan_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_mobile_wipe_wine_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_sim_insertion_human_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_sim_insertion_scripted_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_sim_transfer_cube_human_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_sim_transfer_cube_scripted_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_battery_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_candy_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_coffee_new_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_coffee_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_cups_open_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_fork_pick_up_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_pingpong_test_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_pro_pencil_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_screw_driver_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_tape_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_thread_velcro_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_towel_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_vinh_cup_left_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_vinh_cup_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_ziploc_slide_raw": "aloha_hdf5",
+    "lerobot-raw/umi_cup_in_the_wild_raw": "umi_zarr",
+    "lerobot-raw/pusht_raw": "pusht_zarr",
+    "lerobot-raw/unitreeh1_fold_clothes_raw": "aloha_hdf5",
+    "lerobot-raw/unitreeh1_rearrange_objects_raw": "aloha_hdf5",
+    "lerobot-raw/unitreeh1_two_robot_greeting_raw": "aloha_hdf5",
+    "lerobot-raw/unitreeh1_warehouse_raw": "aloha_hdf5",
+    "lerobot-raw/xarm_lift_medium_raw": "xarm_pkl",
+    "lerobot-raw/xarm_lift_medium_replay_raw": "xarm_pkl",
+    "lerobot-raw/xarm_push_medium_raw": "xarm_pkl",
+    "lerobot-raw/xarm_push_medium_replay_raw": "xarm_pkl",
+    "lerobot-raw/fractal20220817_data_raw": "openx_rlds.fractal20220817_data",
+    "lerobot-raw/kuka_raw": "openx_rlds.kuka",
+    "lerobot-raw/bridge_openx_raw": "openx_rlds.bridge_openx",
+    "lerobot-raw/taco_play_raw": "openx_rlds.taco_play",
+    "lerobot-raw/jaco_play_raw": "openx_rlds.jaco_play",
+    "lerobot-raw/berkeley_cable_routing_raw": "openx_rlds.berkeley_cable_routing",
+    "lerobot-raw/roboturk_raw": "openx_rlds.roboturk",
+    "lerobot-raw/nyu_door_opening_surprising_effectiveness_raw": "openx_rlds.nyu_door_opening_surprising_effectiveness",
+    "lerobot-raw/viola_raw": "openx_rlds.viola",
+    "lerobot-raw/berkeley_autolab_ur5_raw": "openx_rlds.berkeley_autolab_ur5",
+    "lerobot-raw/toto_raw": "openx_rlds.toto",
+    "lerobot-raw/language_table_raw": "openx_rlds.language_table",
+    "lerobot-raw/columbia_cairlab_pusht_real_raw": "openx_rlds.columbia_cairlab_pusht_real",
+    "lerobot-raw/stanford_kuka_multimodal_dataset_raw": "openx_rlds.stanford_kuka_multimodal_dataset",
+    "lerobot-raw/nyu_rot_dataset_raw": "openx_rlds.nyu_rot_dataset",
+    "lerobot-raw/io_ai_tech_raw": "openx_rlds.io_ai_tech",
+    "lerobot-raw/stanford_hydra_dataset_raw": "openx_rlds.stanford_hydra_dataset",
+    "lerobot-raw/austin_buds_dataset_raw": "openx_rlds.austin_buds_dataset",
+    "lerobot-raw/nyu_franka_play_dataset_raw": "openx_rlds.nyu_franka_play_dataset",
+    "lerobot-raw/maniskill_dataset_raw": "openx_rlds.maniskill_dataset",
+    "lerobot-raw/furniture_bench_dataset_raw": "openx_rlds.furniture_bench_dataset",
+    "lerobot-raw/cmu_franka_exploration_dataset_raw": "openx_rlds.cmu_franka_exploration_dataset",
+    "lerobot-raw/ucsd_kitchen_dataset_raw": "openx_rlds.ucsd_kitchen_dataset",
+    "lerobot-raw/ucsd_pick_and_place_dataset_raw": "openx_rlds.ucsd_pick_and_place_dataset",
+    "lerobot-raw/spoc_raw": "openx_rlds.spoc",
+    "lerobot-raw/austin_sailor_dataset_raw": "openx_rlds.austin_sailor_dataset",
+    "lerobot-raw/austin_sirius_dataset_raw": "openx_rlds.austin_sirius_dataset",
+    "lerobot-raw/bc_z_raw": "openx_rlds.bc_z",
+    "lerobot-raw/utokyo_pr2_opening_fridge_raw": "openx_rlds.utokyo_pr2_opening_fridge",
+    "lerobot-raw/utokyo_pr2_tabletop_manipulation_raw": "openx_rlds.utokyo_pr2_tabletop_manipulation",
+    "lerobot-raw/utokyo_xarm_pick_and_place_raw": "openx_rlds.utokyo_xarm_pick_and_place",
+    "lerobot-raw/utokyo_xarm_bimanual_raw": "openx_rlds.utokyo_xarm_bimanual",
+    "lerobot-raw/utokyo_saytap_raw": "openx_rlds.utokyo_saytap",
+    "lerobot-raw/robo_net_raw": "openx_rlds.robo_net",
+    "lerobot-raw/robo_set_raw": "openx_rlds.robo_set",
+    "lerobot-raw/berkeley_mvp_raw": "openx_rlds.berkeley_mvp",
+    "lerobot-raw/berkeley_rpt_raw": "openx_rlds.berkeley_rpt",
+    "lerobot-raw/kaist_nonprehensile_raw": "openx_rlds.kaist_nonprehensile",
+    "lerobot-raw/stanford_mask_vit_raw": "openx_rlds.stanford_mask_vit",
+    "lerobot-raw/tokyo_u_lsmo_raw": "openx_rlds.tokyo_u_lsmo",
+    "lerobot-raw/dlr_sara_pour_raw": "openx_rlds.dlr_sara_pour",
+    "lerobot-raw/dlr_sara_grid_clamp_raw": "openx_rlds.dlr_sara_grid_clamp",
+    "lerobot-raw/dlr_edan_shared_control_raw": "openx_rlds.dlr_edan_shared_control",
+    "lerobot-raw/asu_table_top_raw": "openx_rlds.asu_table_top",
+    "lerobot-raw/stanford_robocook_raw": "openx_rlds.stanford_robocook",
+    "lerobot-raw/imperialcollege_sawyer_wrist_cam_raw": "openx_rlds.imperialcollege_sawyer_wrist_cam",
+    "lerobot-raw/iamlab_cmu_pickup_insert_raw": "openx_rlds.iamlab_cmu_pickup_insert",
+    "lerobot-raw/uiuc_d3field_raw": "openx_rlds.uiuc_d3field",
+    "lerobot-raw/utaustin_mutex_raw": "openx_rlds.utaustin_mutex",
+    "lerobot-raw/berkeley_fanuc_manipulation_raw": "openx_rlds.berkeley_fanuc_manipulation",
+    "lerobot-raw/cmu_playing_with_food_raw": "openx_rlds.cmu_playing_with_food",
+    "lerobot-raw/cmu_play_fusion_raw": "openx_rlds.cmu_play_fusion",
+    "lerobot-raw/cmu_stretch_raw": "openx_rlds.cmu_stretch",
+    "lerobot-raw/berkeley_gnm_recon_raw": "openx_rlds.berkeley_gnm_recon",
+    "lerobot-raw/berkeley_gnm_cory_hall_raw": "openx_rlds.berkeley_gnm_cory_hall",
+    "lerobot-raw/berkeley_gnm_sac_son_raw": "openx_rlds.berkeley_gnm_sac_son",
+    "lerobot-raw/droid_raw": "openx_rlds.droid",
+    "lerobot-raw/droid_100_raw": "openx_rlds.droid100",
+    "lerobot-raw/fmb_raw": "openx_rlds.fmb",
+    "lerobot-raw/dobbe_raw": "openx_rlds.dobbe",
+    "lerobot-raw/usc_cloth_sim_raw": "openx_rlds.usc_cloth_sim",
+    "lerobot-raw/plex_robosuite_raw": "openx_rlds.plex_robosuite",
+    "lerobot-raw/conq_hose_manipulation_raw": "openx_rlds.conq_hose_manipulation",
+    "lerobot-raw/vima_raw": "openx_rlds.vima",
+    "lerobot-raw/robot_vqa_raw": "openx_rlds.robot_vqa",
+    "lerobot-raw/mimic_play_raw": "openx_rlds.mimic_play",
+    "lerobot-raw/tidybot_raw": "openx_rlds.tidybot",
+    "lerobot-raw/eth_agent_affordances_raw": "openx_rlds.eth_agent_affordances",
+}
+
+
+def download_raw(raw_dir: Path, repo_id: str):
+    check_repo_id(repo_id)
+    user_id, dataset_id = repo_id.split("/")
+
+    if not dataset_id.endswith("_raw"):
+        warnings.warn(
+            f"""`dataset_id` ({dataset_id}) doesn't end with '_raw' (e.g. 'lerobot/pusht_raw'). Following this
+             naming convention by renaming your repository is advised, but not mandatory.""",
+            stacklevel=1,
+        )
+
+    # Send warning if raw_dir isn't well formatted
+    if raw_dir.parts[-2] != user_id or raw_dir.parts[-1] != dataset_id:
+        warnings.warn(
+            f"""`raw_dir` ({raw_dir}) doesn't contain a community or user id `/` the name of the dataset that
+             match the `repo_id` (e.g. 'data/lerobot/pusht_raw'). Following this naming convention is advised,
+             but not mandatory.""",
+            stacklevel=1,
+        )
+    raw_dir.mkdir(parents=True, exist_ok=True)
+
+    logging.info(f"Start downloading from huggingface.co/{user_id} for {dataset_id}")
+    snapshot_download(repo_id, repo_type="dataset", local_dir=raw_dir)
+    logging.info(f"Finish downloading from huggingface.co/{user_id} for {dataset_id}")
+
+
+def download_all_raw_datasets(data_dir: Path | None = None):
+    if data_dir is None:
+        data_dir = Path("data")
+    for repo_id in AVAILABLE_RAW_REPO_IDS:
+        raw_dir = data_dir / repo_id
+        download_raw(raw_dir, repo_id)
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description=f"""A script to download raw datasets from Hugging Face hub to a local directory. Here is a
+            non exhaustive list of available repositories to use in `--repo-id`: {list(AVAILABLE_RAW_REPO_IDS.keys())}""",
+    )
+
+    parser.add_argument(
+        "--raw-dir",
+        type=Path,
+        required=True,
+        help="Directory containing input raw datasets (e.g. `data/aloha_mobile_chair_raw` or `data/pusht_raw).",
+    )
+    parser.add_argument(
+        "--repo-id",
+        type=str,
+        required=True,
+        help="""Repositery identifier on Hugging Face: a community or a user name `/` the name of
+        the dataset (e.g. `lerobot/pusht_raw`, `cadene/aloha_sim_insertion_human_raw`).""",
+    )
+    args = parser.parse_args()
+    download_raw(**vars(args))
+
+
+if __name__ == "__main__":
+    main()

lerobot/common/datasets/push_dataset_to_hub/_encode_datasets.py

@@ -0,0 +1,184 @@
+#!/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.
+"""
+Use this script to batch encode lerobot dataset from their raw format to LeRobotDataset and push their updated
+version to the hub. Under the hood, this script reuses 'push_dataset_to_hub.py'. It assumes that you already
+downloaded raw datasets, which you can do with the related '_download_raw.py' script.
+
+For instance, for codebase_version = 'v1.6', the following command was run, assuming raw datasets from
+lerobot-raw were downloaded in 'raw/datasets/directory':
+```bash
+python lerobot/common/datasets/push_dataset_to_hub/_encode_datasets.py \
+  --raw-dir raw/datasets/directory \
+  --raw-repo-ids lerobot-raw \
+  --local-dir push/datasets/directory \
+  --tests-data-dir tests/data \
+  --push-repo lerobot \
+  --vcodec libsvtav1 \
+  --pix-fmt yuv420p \
+  --g 2 \
+  --crf 30
+```
+"""
+
+import argparse
+from pathlib import Path
+
+from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
+from lerobot.common.datasets.push_dataset_to_hub._download_raw import AVAILABLE_RAW_REPO_IDS
+from lerobot.common.datasets.push_dataset_to_hub.utils import check_repo_id
+from lerobot.scripts.push_dataset_to_hub import push_dataset_to_hub
+
+
+def get_push_repo_id_from_raw(raw_repo_id: str, push_repo: str) -> str:
+    dataset_id_raw = raw_repo_id.split("/")[1]
+    dataset_id = dataset_id_raw.removesuffix("_raw")
+    return f"{push_repo}/{dataset_id}"
+
+
+def encode_datasets(
+    raw_dir: Path,
+    raw_repo_ids: list[str],
+    push_repo: str,
+    vcodec: str,
+    pix_fmt: str,
+    g: int,
+    crf: int,
+    local_dir: Path | None = None,
+    tests_data_dir: Path | None = None,
+    raw_format: str | None = None,
+    dry_run: bool = False,
+) -> None:
+    if len(raw_repo_ids) == 1 and raw_repo_ids[0].lower() == "lerobot-raw":
+        raw_repo_ids_format = AVAILABLE_RAW_REPO_IDS
+    else:
+        if raw_format is None:
+            raise ValueError(raw_format)
+        raw_repo_ids_format = {id_: raw_format for id_ in raw_repo_ids}
+
+    for raw_repo_id, repo_raw_format in raw_repo_ids_format.items():
+        check_repo_id(raw_repo_id)
+        dataset_repo_id_push = get_push_repo_id_from_raw(raw_repo_id, push_repo)
+        dataset_raw_dir = raw_dir / raw_repo_id
+        dataset_dir = local_dir / dataset_repo_id_push if local_dir is not None else None
+        encoding = {
+            "vcodec": vcodec,
+            "pix_fmt": pix_fmt,
+            "g": g,
+            "crf": crf,
+        }
+
+        if not (dataset_raw_dir).is_dir():
+            raise NotADirectoryError(dataset_raw_dir)
+
+        if not dry_run:
+            push_dataset_to_hub(
+                dataset_raw_dir,
+                raw_format=repo_raw_format,
+                repo_id=dataset_repo_id_push,
+                local_dir=dataset_dir,
+                resume=True,
+                encoding=encoding,
+                tests_data_dir=tests_data_dir,
+            )
+        else:
+            print(
+                f"DRY RUN: {dataset_raw_dir}  -->  {dataset_dir}  -->  {dataset_repo_id_push}@{CODEBASE_VERSION}"
+            )
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--raw-dir",
+        type=Path,
+        default=Path("data"),
+        help="Directory where raw datasets are located.",
+    )
+    parser.add_argument(
+        "--raw-repo-ids",
+        type=str,
+        nargs="*",
+        default=["lerobot-raw"],
+        help="""Raw dataset repo ids. if 'lerobot-raw', the keys from `AVAILABLE_RAW_REPO_IDS` will be
+            used and raw datasets will be fetched from the 'lerobot-raw/' repo and pushed with their
+            associated format. It is assumed that each dataset is located at `raw_dir / raw_repo_id` """,
+    )
+    parser.add_argument(
+        "--raw-format",
+        type=str,
+        default=None,
+        help="""Raw format to use for the raw repo-ids. Must be specified if --raw-repo-ids is not
+            'lerobot-raw'""",
+    )
+    parser.add_argument(
+        "--local-dir",
+        type=Path,
+        default=None,
+        help="""When provided, writes the dataset converted to LeRobotDataset format in this directory
+        (e.g. `data/lerobot/aloha_mobile_chair`).""",
+    )
+    parser.add_argument(
+        "--push-repo",
+        type=str,
+        default="lerobot",
+        help="Repo to upload datasets to",
+    )
+    parser.add_argument(
+        "--vcodec",
+        type=str,
+        default="libsvtav1",
+        help="Codec to use for encoding videos",
+    )
+    parser.add_argument(
+        "--pix-fmt",
+        type=str,
+        default="yuv420p",
+        help="Pixel formats (chroma subsampling) to be used for encoding",
+    )
+    parser.add_argument(
+        "--g",
+        type=int,
+        default=2,
+        help="Group of pictures sizes to be used for encoding.",
+    )
+    parser.add_argument(
+        "--crf",
+        type=int,
+        default=30,
+        help="Constant rate factors to be used for encoding.",
+    )
+    parser.add_argument(
+        "--tests-data-dir",
+        type=Path,
+        default=None,
+        help=(
+            "When provided, save tests artifacts into the given directory "
+            "(e.g. `--tests-data-dir tests/data` will save to tests/data/{--repo-id})."
+        ),
+    )
+    parser.add_argument(
+        "--dry-run",
+        type=int,
+        default=0,
+        help="If not set to 0, this script won't download or upload anything.",
+    )
+    args = parser.parse_args()
+    encode_datasets(**vars(args))
+
+
+if __name__ == "__main__":
+    main()

lerobot/common/datasets/push_dataset_to_hub/_umi_imagecodecs_numcodecs.py

@@ -0,0 +1,326 @@
+#!/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.
+# imagecodecs/numcodecs.py
+
+# Copyright (c) 2021-2022, Christoph Gohlke
+# All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are met:
+#
+# 1. Redistributions of source code must retain the above copyright notice,
+#    this list of conditions and the following disclaimer.
+#
+# 2. Redistributions in binary form must reproduce the above copyright notice,
+#    this list of conditions and the following disclaimer in the documentation
+#    and/or other materials provided with the distribution.
+#
+# 3. Neither the name of the copyright holder nor the names of its
+#    contributors may be used to endorse or promote products derived from
+#    this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+# ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
+# LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+# CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+# SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+# INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+# CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+# ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+# POSSIBILITY OF SUCH DAMAGE.
+
+# Copied from: https://github.com/real-stanford/universal_manipulation_interface/blob/298776ce251f33b6b3185a98d6e7d1f9ad49168b/diffusion_policy/codecs/imagecodecs_numcodecs.py#L1
+"""Additional numcodecs implemented using imagecodecs."""
+
+__version__ = "2022.9.26"
+
+__all__ = ("register_codecs",)
+
+import imagecodecs
+import numpy
+from numcodecs.abc import Codec
+from numcodecs.registry import get_codec, register_codec
+
+# TODO (azouitine): Remove useless codecs
+
+
+def protective_squeeze(x: numpy.ndarray):
+    """
+    Squeeze dim only if it's not the last dim.
+    Image dim expected to be *, H, W, C
+    """
+    img_shape = x.shape[-3:]
+    if len(x.shape) > 3:
+        n_imgs = numpy.prod(x.shape[:-3])
+        if n_imgs > 1:
+            img_shape = (-1,) + img_shape
+    return x.reshape(img_shape)
+
+
+def get_default_image_compressor(**kwargs):
+    if imagecodecs.JPEGXL:
+        # has JPEGXL
+        this_kwargs = {
+            "effort": 3,
+            "distance": 0.3,
+            # bug in libjxl, invalid codestream for non-lossless
+            # when decoding speed > 1
+            "decodingspeed": 1,
+        }
+        this_kwargs.update(kwargs)
+        return JpegXl(**this_kwargs)
+    else:
+        this_kwargs = {"level": 50}
+        this_kwargs.update(kwargs)
+        return Jpeg2k(**this_kwargs)
+
+
+class Jpeg2k(Codec):
+    """JPEG 2000 codec for numcodecs."""
+
+    codec_id = "imagecodecs_jpeg2k"
+
+    def __init__(
+        self,
+        level=None,
+        codecformat=None,
+        colorspace=None,
+        tile=None,
+        reversible=None,
+        bitspersample=None,
+        resolutions=None,
+        numthreads=None,
+        verbose=0,
+    ):
+        self.level = level
+        self.codecformat = codecformat
+        self.colorspace = colorspace
+        self.tile = None if tile is None else tuple(tile)
+        self.reversible = reversible
+        self.bitspersample = bitspersample
+        self.resolutions = resolutions
+        self.numthreads = numthreads
+        self.verbose = verbose
+
+    def encode(self, buf):
+        buf = protective_squeeze(numpy.asarray(buf))
+        return imagecodecs.jpeg2k_encode(
+            buf,
+            level=self.level,
+            codecformat=self.codecformat,
+            colorspace=self.colorspace,
+            tile=self.tile,
+            reversible=self.reversible,
+            bitspersample=self.bitspersample,
+            resolutions=self.resolutions,
+            numthreads=self.numthreads,
+            verbose=self.verbose,
+        )
+
+    def decode(self, buf, out=None):
+        return imagecodecs.jpeg2k_decode(buf, verbose=self.verbose, numthreads=self.numthreads, out=out)
+
+
+class JpegXl(Codec):
+    """JPEG XL codec for numcodecs."""
+
+    codec_id = "imagecodecs_jpegxl"
+
+    def __init__(
+        self,
+        # encode
+        level=None,
+        effort=None,
+        distance=None,
+        lossless=None,
+        decodingspeed=None,
+        photometric=None,
+        planar=None,
+        usecontainer=None,
+        # decode
+        index=None,
+        keeporientation=None,
+        # both
+        numthreads=None,
+    ):
+        """
+        Return JPEG XL image from numpy array.
+        Float must be in nominal range 0..1.
+
+        Currently L, LA, RGB, RGBA images are supported in contig mode.
+        Extra channels are only supported for grayscale images in planar mode.
+
+        Parameters
+        ----------
+        level : Default to None, i.e. not overwriting lossess and decodingspeed options.
+            When < 0: Use lossless compression
+            When in [0,1,2,3,4]: Sets the decoding speed tier for the provided options.
+                Minimum is 0 (slowest to decode, best quality/density), and maximum
+                is 4 (fastest to decode, at the cost of some quality/density).
+        effort : Default to 3.
+            Sets encoder effort/speed level without affecting decoding speed.
+            Valid values are, from faster to slower speed: 1:lightning 2:thunder
+                3:falcon 4:cheetah 5:hare 6:wombat 7:squirrel 8:kitten 9:tortoise.
+            Speed: lightning, thunder, falcon, cheetah, hare, wombat, squirrel, kitten, tortoise
+            control the encoder effort in ascending order.
+            This also affects memory usage: using lower effort will typically reduce memory
+            consumption during encoding.
+            lightning and thunder are fast modes useful for lossless mode (modular).
+            falcon disables all of the following tools.
+            cheetah enables coefficient reordering, context clustering, and heuristics for selecting DCT sizes and quantization steps.
+            hare enables Gaborish filtering, chroma from luma, and an initial estimate of quantization steps.
+            wombat enables error diffusion quantization and full DCT size selection heuristics.
+            squirrel (default) enables dots, patches, and spline detection, and full context clustering.
+            kitten optimizes the adaptive quantization for a psychovisual metric.
+            tortoise enables a more thorough adaptive quantization search.
+        distance : Default to 1.0
+            Sets the distance level for lossy compression: target max butteraugli distance,
+            lower = higher quality. Range: 0 .. 15. 0.0 = mathematically lossless
+            (however, use JxlEncoderSetFrameLossless instead to use true lossless,
+            as setting distance to 0 alone is not the only requirement).
+            1.0 = visually lossless. Recommended range: 0.5 .. 3.0.
+        lossess : Default to False.
+            Use lossess encoding.
+        decodingspeed : Default to 0.
+            Duplicate to level. [0,4]
+        photometric : Return JxlColorSpace value.
+            Default logic is quite complicated but works most of the time.
+            Accepted value:
+                int: [-1,3]
+                str: ['RGB',
+                    'WHITEISZERO', 'MINISWHITE',
+                    'BLACKISZERO', 'MINISBLACK', 'GRAY',
+                    'XYB', 'KNOWN']
+        planar : Enable multi-channel mode.
+            Default to false.
+        usecontainer :
+            Forces the encoder to use the box-based container format (BMFF)
+            even when not necessary.
+            When using JxlEncoderUseBoxes, JxlEncoderStoreJPEGMetadata or
+            JxlEncoderSetCodestreamLevel with level 10, the encoder will
+            automatically also use the container format, it is not necessary
+            to use JxlEncoderUseContainer for those use cases.
+            By default this setting is disabled.
+        index : Selectively decode frames for animation.
+            Default to 0, decode all frames.
+            When set to > 0, decode that frame index only.
+        keeporientation :
+            Enables or disables preserving of as-in-bitstream pixeldata orientation.
+            Some images are encoded with an Orientation tag indicating that the
+            decoder must perform a rotation and/or mirroring to the encoded image data.
+
+            If skip_reorientation is JXL_FALSE (the default): the decoder will apply
+            the transformation from the orientation setting, hence rendering the image
+            according to its specified intent. When producing a JxlBasicInfo, the decoder
+            will always set the orientation field to JXL_ORIENT_IDENTITY (matching the
+            returned pixel data) and also align xsize and ysize so that they correspond
+            to the width and the height of the returned pixel data.
+
+            If skip_reorientation is JXL_TRUE: the decoder will skip applying the
+            transformation from the orientation setting, returning the image in
+            the as-in-bitstream pixeldata orientation. This may be faster to decode
+            since the decoder doesnt have to apply the transformation, but can
+            cause wrong display of the image if the orientation tag is not correctly
+            taken into account by the user.
+
+            By default, this option is disabled, and the returned pixel data is
+            re-oriented according to the images Orientation setting.
+        threads : Default to 1.
+            If <= 0, use all cores.
+            If > 32, clipped to 32.
+        """
+
+        self.level = level
+        self.effort = effort
+        self.distance = distance
+        self.lossless = bool(lossless)
+        self.decodingspeed = decodingspeed
+        self.photometric = photometric
+        self.planar = planar
+        self.usecontainer = usecontainer
+        self.index = index
+        self.keeporientation = keeporientation
+        self.numthreads = numthreads
+
+    def encode(self, buf):
+        # TODO: only squeeze all but last dim
+        buf = protective_squeeze(numpy.asarray(buf))
+        return imagecodecs.jpegxl_encode(
+            buf,
+            level=self.level,
+            effort=self.effort,
+            distance=self.distance,
+            lossless=self.lossless,
+            decodingspeed=self.decodingspeed,
+            photometric=self.photometric,
+            planar=self.planar,
+            usecontainer=self.usecontainer,
+            numthreads=self.numthreads,
+        )
+
+    def decode(self, buf, out=None):
+        return imagecodecs.jpegxl_decode(
+            buf,
+            index=self.index,
+            keeporientation=self.keeporientation,
+            numthreads=self.numthreads,
+            out=out,
+        )
+
+
+def _flat(out):
+    """Return numpy array as contiguous view of bytes if possible."""
+    if out is None:
+        return None
+    view = memoryview(out)
+    if view.readonly or not view.contiguous:
+        return None
+    return view.cast("B")
+
+
+def register_codecs(codecs=None, force=False, verbose=True):
+    """Register codecs in this module with numcodecs."""
+    for name, cls in globals().items():
+        if not hasattr(cls, "codec_id") or name == "Codec":
+            continue
+        if codecs is not None and cls.codec_id not in codecs:
+            continue
+        try:
+            try:  # noqa: SIM105
+                get_codec({"id": cls.codec_id})
+            except TypeError:
+                # registered, but failed
+                pass
+        except ValueError:
+            # not registered yet
+            pass
+        else:
+            if not force:
+                if verbose:
+                    log_warning(f"numcodec {cls.codec_id!r} already registered")
+                continue
+            if verbose:
+                log_warning(f"replacing registered numcodec {cls.codec_id!r}")
+        register_codec(cls)
+
+
+def log_warning(msg, *args, **kwargs):
+    """Log message with level WARNING."""
+    import logging
+
+    logging.getLogger(__name__).warning(msg, *args, **kwargs)

lerobot/common/datasets/push_dataset_to_hub/aloha_hdf5_format.py

@@ -0,0 +1,233 @@
+#!/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.
+"""
+Contains utilities to process raw data format of HDF5 files like in: https://github.com/tonyzhaozh/act
+"""
+
+import gc
+import shutil
+from pathlib import Path
+
+import h5py
+import numpy as np
+import torch
+import tqdm
+from datasets import Dataset, Features, Image, Sequence, Value
+from PIL import Image as PILImage
+
+from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
+from lerobot.common.datasets.push_dataset_to_hub.utils import (
+    calculate_episode_data_index,
+    concatenate_episodes,
+    get_default_encoding,
+    save_images_concurrently,
+)
+from lerobot.common.datasets.utils import (
+    hf_transform_to_torch,
+)
+from lerobot.common.datasets.video_utils import VideoFrame, encode_video_frames
+
+
+def get_cameras(hdf5_data):
+    # ignore depth channel, not currently handled
+    # TODO(rcadene): add depth
+    rgb_cameras = [key for key in hdf5_data["/observations/images"].keys() if "depth" not in key]  # noqa: SIM118
+    return rgb_cameras
+
+
+def check_format(raw_dir) -> bool:
+    # only frames from simulation are uncompressed
+    compressed_images = "sim" not in raw_dir.name
+
+    hdf5_paths = list(raw_dir.glob("episode_*.hdf5"))
+    assert len(hdf5_paths) != 0
+    for hdf5_path in hdf5_paths:
+        with h5py.File(hdf5_path, "r") as data:
+            assert "/action" in data
+            assert "/observations/qpos" in data
+
+            assert data["/action"].ndim == 2
+            assert data["/observations/qpos"].ndim == 2
+
+            num_frames = data["/action"].shape[0]
+            assert num_frames == data["/observations/qpos"].shape[0]
+
+            for camera in get_cameras(data):
+                assert num_frames == data[f"/observations/images/{camera}"].shape[0]
+
+                if compressed_images:
+                    assert data[f"/observations/images/{camera}"].ndim == 2
+                else:
+                    assert data[f"/observations/images/{camera}"].ndim == 4
+                    b, h, w, c = data[f"/observations/images/{camera}"].shape
+                    assert c < h and c < w, f"Expect (h,w,c) image format but ({h=},{w=},{c=}) provided."
+
+
+def load_from_raw(
+    raw_dir: Path,
+    videos_dir: Path,
+    fps: int,
+    video: bool,
+    episodes: list[int] | None = None,
+    encoding: dict | None = None,
+):
+    # only frames from simulation are uncompressed
+    compressed_images = "sim" not in raw_dir.name
+
+    hdf5_files = sorted(raw_dir.glob("episode_*.hdf5"))
+    num_episodes = len(hdf5_files)
+
+    ep_dicts = []
+    ep_ids = episodes if episodes else range(num_episodes)
+    for ep_idx in tqdm.tqdm(ep_ids):
+        ep_path = hdf5_files[ep_idx]
+        with h5py.File(ep_path, "r") as ep:
+            num_frames = ep["/action"].shape[0]
+
+            # last step of demonstration is considered done
+            done = torch.zeros(num_frames, dtype=torch.bool)
+            done[-1] = True
+
+            state = torch.from_numpy(ep["/observations/qpos"][:])
+            action = torch.from_numpy(ep["/action"][:])
+            if "/observations/qvel" in ep:
+                velocity = torch.from_numpy(ep["/observations/qvel"][:])
+            if "/observations/effort" in ep:
+                effort = torch.from_numpy(ep["/observations/effort"][:])
+
+            ep_dict = {}
+
+            for camera in get_cameras(ep):
+                img_key = f"observation.images.{camera}"
+
+                if compressed_images:
+                    import cv2
+
+                    # load one compressed image after the other in RAM and uncompress
+                    imgs_array = []
+                    for data in ep[f"/observations/images/{camera}"]:
+                        imgs_array.append(cv2.imdecode(data, 1))
+                    imgs_array = np.array(imgs_array)
+
+                else:
+                    # load all images in RAM
+                    imgs_array = ep[f"/observations/images/{camera}"][:]
+
+                if video:
+                    # save png images in temporary directory
+                    tmp_imgs_dir = videos_dir / "tmp_images"
+                    save_images_concurrently(imgs_array, tmp_imgs_dir)
+
+                    # encode images to a mp4 video
+                    fname = f"{img_key}_episode_{ep_idx:06d}.mp4"
+                    video_path = videos_dir / fname
+                    encode_video_frames(tmp_imgs_dir, video_path, fps, **(encoding or {}))
+
+                    # clean temporary images directory
+                    shutil.rmtree(tmp_imgs_dir)
+
+                    # store the reference to the video frame
+                    ep_dict[img_key] = [
+                        {"path": f"videos/{fname}", "timestamp": i / fps} for i in range(num_frames)
+                    ]
+                else:
+                    ep_dict[img_key] = [PILImage.fromarray(x) for x in imgs_array]
+
+            ep_dict["observation.state"] = state
+            if "/observations/velocity" in ep:
+                ep_dict["observation.velocity"] = velocity
+            if "/observations/effort" in ep:
+                ep_dict["observation.effort"] = effort
+            ep_dict["action"] = action
+            ep_dict["episode_index"] = torch.tensor([ep_idx] * num_frames)
+            ep_dict["frame_index"] = torch.arange(0, num_frames, 1)
+            ep_dict["timestamp"] = torch.arange(0, num_frames, 1) / fps
+            ep_dict["next.done"] = done
+            # TODO(rcadene): add reward and success by computing them in sim
+
+            assert isinstance(ep_idx, int)
+            ep_dicts.append(ep_dict)
+
+        gc.collect()
+
+    data_dict = concatenate_episodes(ep_dicts)
+
+    total_frames = data_dict["frame_index"].shape[0]
+    data_dict["index"] = torch.arange(0, total_frames, 1)
+    return data_dict
+
+
+def to_hf_dataset(data_dict, video) -> Dataset:
+    features = {}
+
+    keys = [key for key in data_dict if "observation.images." in key]
+    for key in keys:
+        if video:
+            features[key] = VideoFrame()
+        else:
+            features[key] = Image()
+
+    features["observation.state"] = Sequence(
+        length=data_dict["observation.state"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    if "observation.velocity" in data_dict:
+        features["observation.velocity"] = Sequence(
+            length=data_dict["observation.velocity"].shape[1], feature=Value(dtype="float32", id=None)
+        )
+    if "observation.effort" in data_dict:
+        features["observation.effort"] = Sequence(
+            length=data_dict["observation.effort"].shape[1], feature=Value(dtype="float32", id=None)
+        )
+    features["action"] = Sequence(
+        length=data_dict["action"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    features["episode_index"] = Value(dtype="int64", id=None)
+    features["frame_index"] = Value(dtype="int64", id=None)
+    features["timestamp"] = Value(dtype="float32", id=None)
+    features["next.done"] = Value(dtype="bool", id=None)
+    features["index"] = Value(dtype="int64", id=None)
+
+    hf_dataset = Dataset.from_dict(data_dict, features=Features(features))
+    hf_dataset.set_transform(hf_transform_to_torch)
+    return hf_dataset
+
+
+def from_raw_to_lerobot_format(
+    raw_dir: Path,
+    videos_dir: Path,
+    fps: int | None = None,
+    video: bool = True,
+    episodes: list[int] | None = None,
+    encoding: dict | None = None,
+):
+    # sanity check
+    check_format(raw_dir)
+
+    if fps is None:
+        fps = 50
+
+    data_dict = load_from_raw(raw_dir, videos_dir, fps, video, episodes, encoding)
+    hf_dataset = to_hf_dataset(data_dict, video)
+    episode_data_index = calculate_episode_data_index(hf_dataset)
+    info = {
+        "codebase_version": CODEBASE_VERSION,
+        "fps": fps,
+        "video": video,
+    }
+    if video:
+        info["encoding"] = get_default_encoding()
+
+    return hf_dataset, episode_data_index, info

lerobot/common/datasets/push_dataset_to_hub/cam_png_format.py

@@ -0,0 +1,107 @@
+#!/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.
+"""
+Contains utilities to process raw data format of png images files recorded with capture_camera_feed.py
+"""
+
+from pathlib import Path
+
+import torch
+from datasets import Dataset, Features, Image, Value
+from PIL import Image as PILImage
+
+from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
+from lerobot.common.datasets.push_dataset_to_hub.utils import (
+    calculate_episode_data_index,
+    concatenate_episodes,
+)
+from lerobot.common.datasets.utils import hf_transform_to_torch
+from lerobot.common.datasets.video_utils import VideoFrame
+
+
+def check_format(raw_dir: Path) -> bool:
+    image_paths = list(raw_dir.glob("frame_*.png"))
+    if len(image_paths) == 0:
+        raise ValueError
+
+
+def load_from_raw(raw_dir: Path, fps: int, episodes: list[int] | None = None):
+    if episodes is not None:
+        # TODO(aliberts): add support for multi-episodes.
+        raise NotImplementedError()
+
+    ep_dict = {}
+    ep_idx = 0
+
+    image_paths = sorted(raw_dir.glob("frame_*.png"))
+    num_frames = len(image_paths)
+
+    ep_dict["observation.image"] = [PILImage.open(x) for x in image_paths]
+    ep_dict["episode_index"] = torch.tensor([ep_idx] * num_frames)
+    ep_dict["frame_index"] = torch.arange(0, num_frames, 1)
+    ep_dict["timestamp"] = torch.arange(0, num_frames, 1) / fps
+
+    ep_dicts = [ep_dict]
+    data_dict = concatenate_episodes(ep_dicts)
+    total_frames = data_dict["frame_index"].shape[0]
+    data_dict["index"] = torch.arange(0, total_frames, 1)
+    return data_dict
+
+
+def to_hf_dataset(data_dict, video) -> Dataset:
+    features = {}
+    if video:
+        features["observation.image"] = VideoFrame()
+    else:
+        features["observation.image"] = Image()
+
+    features["episode_index"] = Value(dtype="int64", id=None)
+    features["frame_index"] = Value(dtype="int64", id=None)
+    features["timestamp"] = Value(dtype="float32", id=None)
+    features["index"] = Value(dtype="int64", id=None)
+
+    hf_dataset = Dataset.from_dict(data_dict, features=Features(features))
+    hf_dataset.set_transform(hf_transform_to_torch)
+    return hf_dataset
+
+
+def from_raw_to_lerobot_format(
+    raw_dir: Path,
+    videos_dir: Path,
+    fps: int | None = None,
+    video: bool = True,
+    episodes: list[int] | None = None,
+    encoding: dict | None = None,
+):
+    if video or episodes or encoding is not None:
+        # TODO(aliberts): support this
+        raise NotImplementedError
+
+    # sanity check
+    check_format(raw_dir)
+
+    if fps is None:
+        fps = 30
+
+    data_dict = load_from_raw(raw_dir, videos_dir, fps, video, episodes)
+    hf_dataset = to_hf_dataset(data_dict, video)
+    episode_data_index = calculate_episode_data_index(hf_dataset)
+    info = {
+        "codebase_version": CODEBASE_VERSION,
+        "fps": fps,
+        "video": video,
+    }
+    return hf_dataset, episode_data_index, info

lerobot/common/datasets/push_dataset_to_hub/dora_parquet_format.py

@@ -0,0 +1,233 @@
+#!/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.
+"""
+Contains utilities to process raw data format from dora-record
+"""
+
+import re
+import warnings
+from pathlib import Path
+
+import pandas as pd
+import torch
+from datasets import Dataset, Features, Image, Sequence, Value
+
+from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
+from lerobot.common.datasets.push_dataset_to_hub.utils import calculate_episode_data_index
+from lerobot.common.datasets.utils import (
+    hf_transform_to_torch,
+)
+from lerobot.common.datasets.video_utils import VideoFrame
+
+
+def check_format(raw_dir) -> bool:
+    assert raw_dir.exists()
+
+    leader_file = list(raw_dir.glob("*.parquet"))
+    if len(leader_file) == 0:
+        raise ValueError(f"Missing parquet files in '{raw_dir}'")
+    return True
+
+
+def load_from_raw(raw_dir: Path, videos_dir: Path, fps: int, video: bool, episodes: list[int] | None = None):
+    # Load data stream that will be used as reference for the timestamps synchronization
+    reference_files = list(raw_dir.glob("observation.images.cam_*.parquet"))
+    if len(reference_files) == 0:
+        raise ValueError(f"Missing reference files for camera, starting with  in '{raw_dir}'")
+    # select first camera in alphanumeric order
+    reference_key = sorted(reference_files)[0].stem
+    reference_df = pd.read_parquet(raw_dir / f"{reference_key}.parquet")
+    reference_df = reference_df[["timestamp_utc", reference_key]]
+
+    # Merge all data stream using nearest backward strategy
+    df = reference_df
+    for path in raw_dir.glob("*.parquet"):
+        key = path.stem  # action or observation.state or ...
+        if key == reference_key:
+            continue
+        if "failed_episode_index" in key:
+            # TODO(rcadene): add support for removing episodes that are tagged as "failed"
+            continue
+        modality_df = pd.read_parquet(path)
+        modality_df = modality_df[["timestamp_utc", key]]
+        df = pd.merge_asof(
+            df,
+            modality_df,
+            on="timestamp_utc",
+            # "nearest" is the best option over "backward", since the latter can desynchronizes camera timestamps by
+            # matching timestamps that are too far apart, in order to fit the backward constraints. It's not the case for "nearest".
+            # However, note that "nearest" might synchronize the reference camera with other cameras on slightly future timestamps.
+            # are too far apart.
+            direction="nearest",
+            tolerance=pd.Timedelta(f"{1 / fps} seconds"),
+        )
+    # Remove rows with episode_index -1 which indicates data that correspond to in-between episodes
+    df = df[df["episode_index"] != -1]
+
+    image_keys = [key for key in df if "observation.images." in key]
+
+    def get_episode_index(row):
+        episode_index_per_cam = {}
+        for key in image_keys:
+            path = row[key][0]["path"]
+            match = re.search(r"_(\d{6}).mp4", path)
+            if not match:
+                raise ValueError(path)
+            episode_index = int(match.group(1))
+            episode_index_per_cam[key] = episode_index
+        if len(set(episode_index_per_cam.values())) != 1:
+            raise ValueError(
+                f"All cameras are expected to belong to the same episode, but getting {episode_index_per_cam}"
+            )
+        return episode_index
+
+    df["episode_index"] = df.apply(get_episode_index, axis=1)
+
+    # dora only use arrays, so single values are encapsulated into a list
+    df["frame_index"] = df.groupby("episode_index").cumcount()
+    df = df.reset_index()
+    df["index"] = df.index
+
+    # set 'next.done' to True for the last frame of each episode
+    df["next.done"] = False
+    df.loc[df.groupby("episode_index").tail(1).index, "next.done"] = True
+
+    df["timestamp"] = df["timestamp_utc"].map(lambda x: x.timestamp())
+    # each episode starts with timestamp 0 to match the ones from the video
+    df["timestamp"] = df.groupby("episode_index")["timestamp"].transform(lambda x: x - x.iloc[0])
+
+    del df["timestamp_utc"]
+
+    # sanity check
+    has_nan = df.isna().any().any()
+    if has_nan:
+        raise ValueError("Dataset contains Nan values.")
+
+    # sanity check episode indices go from 0 to n-1
+    ep_ids = [ep_idx for ep_idx, _ in df.groupby("episode_index")]
+    expected_ep_ids = list(range(df["episode_index"].max() + 1))
+    if ep_ids != expected_ep_ids:
+        raise ValueError(f"Episodes indices go from {ep_ids} instead of {expected_ep_ids}")
+
+    # Create symlink to raw videos directory (that needs to be absolute not relative)
+    videos_dir.parent.mkdir(parents=True, exist_ok=True)
+    videos_dir.symlink_to((raw_dir / "videos").absolute())
+
+    # sanity check the video paths are well formatted
+    for key in df:
+        if "observation.images." not in key:
+            continue
+        for ep_idx in ep_ids:
+            video_path = videos_dir / f"{key}_episode_{ep_idx:06d}.mp4"
+            if not video_path.exists():
+                raise ValueError(f"Video file not found in {video_path}")
+
+    data_dict = {}
+    for key in df:
+        # is video frame
+        if "observation.images." in key:
+            # we need `[0] because dora only use arrays, so single values are encapsulated into a list.
+            # it is the case for video_frame dictionary = [{"path": ..., "timestamp": ...}]
+            data_dict[key] = [video_frame[0] for video_frame in df[key].values]
+
+            # sanity check the video path is well formatted
+            video_path = videos_dir.parent / data_dict[key][0]["path"]
+            if not video_path.exists():
+                raise ValueError(f"Video file not found in {video_path}")
+        # is number
+        elif df[key].iloc[0].ndim == 0 or df[key].iloc[0].shape[0] == 1:
+            data_dict[key] = torch.from_numpy(df[key].values)
+        # is vector
+        elif df[key].iloc[0].shape[0] > 1:
+            data_dict[key] = torch.stack([torch.from_numpy(x.copy()) for x in df[key].values])
+        else:
+            raise ValueError(key)
+
+    return data_dict
+
+
+def to_hf_dataset(data_dict, video) -> Dataset:
+    features = {}
+
+    keys = [key for key in data_dict if "observation.images." in key]
+    for key in keys:
+        if video:
+            features[key] = VideoFrame()
+        else:
+            features[key] = Image()
+
+    features["observation.state"] = Sequence(
+        length=data_dict["observation.state"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    if "observation.velocity" in data_dict:
+        features["observation.velocity"] = Sequence(
+            length=data_dict["observation.velocity"].shape[1], feature=Value(dtype="float32", id=None)
+        )
+    if "observation.effort" in data_dict:
+        features["observation.effort"] = Sequence(
+            length=data_dict["observation.effort"].shape[1], feature=Value(dtype="float32", id=None)
+        )
+    features["action"] = Sequence(
+        length=data_dict["action"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    features["episode_index"] = Value(dtype="int64", id=None)
+    features["frame_index"] = Value(dtype="int64", id=None)
+    features["timestamp"] = Value(dtype="float32", id=None)
+    features["next.done"] = Value(dtype="bool", id=None)
+    features["index"] = Value(dtype="int64", id=None)
+
+    hf_dataset = Dataset.from_dict(data_dict, features=Features(features))
+    hf_dataset.set_transform(hf_transform_to_torch)
+    return hf_dataset
+
+
+def from_raw_to_lerobot_format(
+    raw_dir: Path,
+    videos_dir: Path,
+    fps: int | None = None,
+    video: bool = True,
+    episodes: list[int] | None = None,
+    encoding: dict | None = None,
+):
+    # sanity check
+    check_format(raw_dir)
+
+    if fps is None:
+        fps = 30
+    else:
+        raise NotImplementedError()
+
+    if not video:
+        raise NotImplementedError()
+
+    if encoding is not None:
+        warnings.warn(
+            "Video encoding is currently done outside of LeRobot for the dora_parquet format.",
+            stacklevel=1,
+        )
+
+    data_df = load_from_raw(raw_dir, videos_dir, fps, episodes)
+    hf_dataset = to_hf_dataset(data_df, video)
+    episode_data_index = calculate_episode_data_index(hf_dataset)
+    info = {
+        "codebase_version": CODEBASE_VERSION,
+        "fps": fps,
+        "video": video,
+    }
+    if video:
+        info["encoding"] = "unknown"
+
+    return hf_dataset, episode_data_index, info

lerobot/common/datasets/push_dataset_to_hub/openx_rlds_format.py

@@ -0,0 +1,312 @@
+#!/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.
+"""
+For all datasets in the RLDS format.
+For https://github.com/google-deepmind/open_x_embodiment (OPENX) datasets.
+
+NOTE: You need to install tensorflow and tensorflow_datasets before running this script.
+
+Example:
+    python lerobot/scripts/push_dataset_to_hub.py \
+        --raw-dir /path/to/data/bridge_dataset/1.0.0/ \
+        --repo-id your_hub/sampled_bridge_data_v2 \
+        --raw-format rlds \
+        --episodes 3 4 5 8 9
+
+Exact dataset fps defined in openx/config.py, obtained from:
+    https://docs.google.com/spreadsheets/d/1rPBD77tk60AEIGZrGSODwyyzs5FgCU9Uz3h-3_t2A9g/edit?gid=0#gid=0&range=R:R
+"""
+
+import shutil
+from pathlib import Path
+
+import numpy as np
+import tensorflow as tf
+import tensorflow_datasets as tfds
+import torch
+import tqdm
+from datasets import Dataset, Features, Image, Sequence, Value
+from PIL import Image as PILImage
+
+from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
+from lerobot.common.datasets.push_dataset_to_hub.utils import (
+    calculate_episode_data_index,
+    concatenate_episodes,
+    get_default_encoding,
+    save_images_concurrently,
+)
+from lerobot.common.datasets.utils import (
+    hf_transform_to_torch,
+)
+from lerobot.common.datasets.video_utils import VideoFrame, encode_video_frames
+
+np.set_printoptions(precision=2)
+
+
+def tf_to_torch(data):
+    return torch.from_numpy(data.numpy())
+
+
+def tf_img_convert(img):
+    if img.dtype == tf.string:
+        img = tf.io.decode_image(img, expand_animations=False, dtype=tf.uint8)
+    elif img.dtype != tf.uint8:
+        raise ValueError(f"Unsupported image dtype: found with dtype {img.dtype}")
+    return img.numpy()
+
+
+def _broadcast_metadata_rlds(i: tf.Tensor, traj: dict) -> dict:
+    """
+    In the RLDS format, each trajectory has some top-level metadata that is explicitly separated out, and a "steps"
+    entry. This function moves the "steps" entry to the top level, broadcasting any metadata to the length of the
+    trajectory. This function also adds the extra metadata fields `_len`, `_traj_index`, and `_frame_index`.
+
+    NOTE: adapted from DLimp library https://github.com/kvablack/dlimp/
+    """
+    steps = traj.pop("steps")
+
+    traj_len = tf.shape(tf.nest.flatten(steps)[0])[0]
+
+    # broadcast metadata to the length of the trajectory
+    metadata = tf.nest.map_structure(lambda x: tf.repeat(x, traj_len), traj)
+
+    # put steps back in
+    assert "traj_metadata" not in steps
+    traj = {**steps, "traj_metadata": metadata}
+
+    assert "_len" not in traj
+    assert "_traj_index" not in traj
+    assert "_frame_index" not in traj
+    traj["_len"] = tf.repeat(traj_len, traj_len)
+    traj["_traj_index"] = tf.repeat(i, traj_len)
+    traj["_frame_index"] = tf.range(traj_len)
+
+    return traj
+
+
+def load_from_raw(
+    raw_dir: Path,
+    videos_dir: Path,
+    fps: int,
+    video: bool,
+    episodes: list[int] | None = None,
+    encoding: dict | None = None,
+):
+    """
+    Args:
+        raw_dir (Path): _description_
+        videos_dir (Path): _description_
+        fps (int): _description_
+        video (bool): _description_
+        episodes (list[int] | None, optional): _description_. Defaults to None.
+    """
+    ds_builder = tfds.builder_from_directory(str(raw_dir))
+    dataset = ds_builder.as_dataset(
+        split="all",
+        decoders={"steps": tfds.decode.SkipDecoding()},
+    )
+
+    dataset_info = ds_builder.info
+    print("dataset_info: ", dataset_info)
+
+    ds_length = len(dataset)
+    dataset = dataset.take(ds_length)
+    # "flatten" the dataset as such we can apply trajectory level map() easily
+    # each [obs][key] has a shape of (frame_size, ...)
+    dataset = dataset.enumerate().map(_broadcast_metadata_rlds)
+
+    # we will apply the standardization transform if the dataset_name is provided
+    # if the dataset name is not provided and the goal is to convert any rlds formatted dataset
+    # search for 'image' keys in the observations
+    image_keys = []
+    state_keys = []
+    observation_info = dataset_info.features["steps"]["observation"]
+    for key in observation_info:
+        # check whether the key is for an image or a vector observation
+        if len(observation_info[key].shape) == 3:
+            # only adding uint8 images discards depth images
+            if observation_info[key].dtype == tf.uint8:
+                image_keys.append(key)
+        else:
+            state_keys.append(key)
+
+    lang_key = "language_instruction" if "language_instruction" in dataset.element_spec else None
+
+    print(" - image_keys: ", image_keys)
+    print(" - lang_key: ", lang_key)
+
+    it = iter(dataset)
+
+    ep_dicts = []
+    # Init temp path to save ep_dicts in case of crash
+    tmp_ep_dicts_dir = videos_dir.parent.joinpath("ep_dicts")
+    tmp_ep_dicts_dir.mkdir(parents=True, exist_ok=True)
+
+    # check if ep_dicts have already been saved in /tmp
+    starting_ep_idx = 0
+    saved_ep_dicts = [ep.__str__() for ep in tmp_ep_dicts_dir.iterdir()]
+    if len(saved_ep_dicts) > 0:
+        saved_ep_dicts.sort()
+        # get last ep_idx number
+        starting_ep_idx = int(saved_ep_dicts[-1][-13:-3]) + 1
+        for i in range(starting_ep_idx):
+            episode = next(it)
+            ep_dicts.append(torch.load(saved_ep_dicts[i]))
+
+    # if we user specified episodes, skip the ones not in the list
+    if episodes is not None:
+        if ds_length == 0:
+            raise ValueError("No episodes found.")
+        # convert episodes index to sorted list
+        episodes = sorted(episodes)
+
+    for ep_idx in tqdm.tqdm(range(starting_ep_idx, ds_length)):
+        episode = next(it)
+
+        # if user specified episodes, skip the ones not in the list
+        if episodes is not None:
+            if len(episodes) == 0:
+                break
+            if ep_idx == episodes[0]:
+                # process this episode
+                print(" selecting episode idx: ", ep_idx)
+                episodes.pop(0)
+            else:
+                continue  # skip
+
+        num_frames = episode["action"].shape[0]
+
+        ep_dict = {}
+        for key in state_keys:
+            ep_dict[f"observation.{key}"] = tf_to_torch(episode["observation"][key])
+
+        ep_dict["action"] = tf_to_torch(episode["action"])
+        ep_dict["next.reward"] = tf_to_torch(episode["reward"]).float()
+        ep_dict["next.done"] = tf_to_torch(episode["is_last"])
+        ep_dict["is_terminal"] = tf_to_torch(episode["is_terminal"])
+        ep_dict["is_first"] = tf_to_torch(episode["is_first"])
+        ep_dict["discount"] = tf_to_torch(episode["discount"])
+
+        # If lang_key is present, convert the entire tensor at once
+        if lang_key is not None:
+            ep_dict["language_instruction"] = [x.numpy().decode("utf-8") for x in episode[lang_key]]
+
+        ep_dict["timestamp"] = torch.arange(0, num_frames, 1) / fps
+        ep_dict["episode_index"] = torch.tensor([ep_idx] * num_frames)
+        ep_dict["frame_index"] = torch.arange(0, num_frames, 1)
+
+        image_array_dict = {key: [] for key in image_keys}
+
+        for im_key in image_keys:
+            imgs = episode["observation"][im_key]
+            image_array_dict[im_key] = [tf_img_convert(img) for img in imgs]
+
+        # loop through all cameras
+        for im_key in image_keys:
+            img_key = f"observation.images.{im_key}"
+            imgs_array = image_array_dict[im_key]
+            imgs_array = np.array(imgs_array)
+            if video:
+                # save png images in temporary directory
+                tmp_imgs_dir = videos_dir / "tmp_images"
+                save_images_concurrently(imgs_array, tmp_imgs_dir)
+
+                # encode images to a mp4 video
+                fname = f"{img_key}_episode_{ep_idx:06d}.mp4"
+                video_path = videos_dir / fname
+                encode_video_frames(tmp_imgs_dir, video_path, fps, **(encoding or {}))
+
+                # clean temporary images directory
+                shutil.rmtree(tmp_imgs_dir)
+
+                # store the reference to the video frame
+                ep_dict[img_key] = [
+                    {"path": f"videos/{fname}", "timestamp": i / fps} for i in range(num_frames)
+                ]
+            else:
+                ep_dict[img_key] = [PILImage.fromarray(x) for x in imgs_array]
+
+        path_ep_dict = tmp_ep_dicts_dir.joinpath(
+            "ep_dict_" + "0" * (10 - len(str(ep_idx))) + str(ep_idx) + ".pt"
+        )
+        torch.save(ep_dict, path_ep_dict)
+
+        ep_dicts.append(ep_dict)
+
+    data_dict = concatenate_episodes(ep_dicts)
+
+    total_frames = data_dict["frame_index"].shape[0]
+    data_dict["index"] = torch.arange(0, total_frames, 1)
+    return data_dict
+
+
+def to_hf_dataset(data_dict, video) -> Dataset:
+    features = {}
+
+    for key in data_dict:
+        # check if vector state obs
+        if key.startswith("observation.") and "observation.images." not in key:
+            features[key] = Sequence(length=data_dict[key].shape[1], feature=Value(dtype="float32", id=None))
+        # check if image obs
+        elif "observation.images." in key:
+            if video:
+                features[key] = VideoFrame()
+            else:
+                features[key] = Image()
+
+    if "language_instruction" in data_dict:
+        features["language_instruction"] = Value(dtype="string", id=None)
+
+    features["action"] = Sequence(
+        length=data_dict["action"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+
+    features["is_terminal"] = Value(dtype="bool", id=None)
+    features["is_first"] = Value(dtype="bool", id=None)
+    features["discount"] = Value(dtype="float32", id=None)
+
+    features["episode_index"] = Value(dtype="int64", id=None)
+    features["frame_index"] = Value(dtype="int64", id=None)
+    features["timestamp"] = Value(dtype="float32", id=None)
+    features["next.reward"] = Value(dtype="float32", id=None)
+    features["next.done"] = Value(dtype="bool", id=None)
+    features["index"] = Value(dtype="int64", id=None)
+
+    hf_dataset = Dataset.from_dict(data_dict, features=Features(features))
+    hf_dataset.set_transform(hf_transform_to_torch)
+    return hf_dataset
+
+
+def from_raw_to_lerobot_format(
+    raw_dir: Path,
+    videos_dir: Path,
+    fps: int | None = None,
+    video: bool = True,
+    episodes: list[int] | None = None,
+    encoding: dict | None = None,
+):
+    data_dict = load_from_raw(raw_dir, videos_dir, fps, video, episodes, encoding)
+    hf_dataset = to_hf_dataset(data_dict, video)
+    episode_data_index = calculate_episode_data_index(hf_dataset)
+    info = {
+        "codebase_version": CODEBASE_VERSION,
+        "fps": fps,
+        "video": video,
+    }
+    if video:
+        info["encoding"] = get_default_encoding()
+
+    return hf_dataset, episode_data_index, info

lerobot/common/datasets/push_dataset_to_hub/pusht_zarr_format.py

@@ -0,0 +1,275 @@
+#!/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.
+"""Process zarr files formatted like in: https://github.com/real-stanford/diffusion_policy"""
+
+import shutil
+from pathlib import Path
+
+import numpy as np
+import torch
+import tqdm
+import zarr
+from datasets import Dataset, Features, Image, Sequence, Value
+from PIL import Image as PILImage
+
+from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
+from lerobot.common.datasets.push_dataset_to_hub.utils import (
+    calculate_episode_data_index,
+    concatenate_episodes,
+    get_default_encoding,
+    save_images_concurrently,
+)
+from lerobot.common.datasets.utils import (
+    hf_transform_to_torch,
+)
+from lerobot.common.datasets.video_utils import VideoFrame, encode_video_frames
+
+
+def check_format(raw_dir):
+    zarr_path = raw_dir / "pusht_cchi_v7_replay.zarr"
+    zarr_data = zarr.open(zarr_path, mode="r")
+
+    required_datasets = {
+        "data/action",
+        "data/img",
+        "data/keypoint",
+        "data/n_contacts",
+        "data/state",
+        "meta/episode_ends",
+    }
+    for dataset in required_datasets:
+        assert dataset in zarr_data
+    nb_frames = zarr_data["data/img"].shape[0]
+
+    required_datasets.remove("meta/episode_ends")
+
+    assert all(nb_frames == zarr_data[dataset].shape[0] for dataset in required_datasets)
+
+
+def load_from_raw(
+    raw_dir: Path,
+    videos_dir: Path,
+    fps: int,
+    video: bool,
+    episodes: list[int] | None = None,
+    keypoints_instead_of_image: bool = False,
+    encoding: dict | None = None,
+):
+    try:
+        import pymunk
+        from gym_pusht.envs.pusht import PushTEnv, pymunk_to_shapely
+
+        from lerobot.common.datasets.push_dataset_to_hub._diffusion_policy_replay_buffer import (
+            ReplayBuffer as DiffusionPolicyReplayBuffer,
+        )
+    except ModuleNotFoundError as e:
+        print("`gym_pusht` is not installed. Please install it with `pip install 'lerobot[gym_pusht]'`")
+        raise e
+    # as define in gmy-pusht env: https://github.com/huggingface/gym-pusht/blob/e0684ff988d223808c0a9dcfaba9dc4991791370/gym_pusht/envs/pusht.py#L174
+    success_threshold = 0.95  # 95% coverage,
+
+    zarr_path = raw_dir / "pusht_cchi_v7_replay.zarr"
+    zarr_data = DiffusionPolicyReplayBuffer.copy_from_path(zarr_path)
+
+    episode_ids = torch.from_numpy(zarr_data.get_episode_idxs())
+    assert len(
+        {zarr_data[key].shape[0] for key in zarr_data.keys()}  # noqa: SIM118
+    ), "Some data type dont have the same number of total frames."
+
+    # TODO(rcadene): verify that goal pose is expected to be fixed
+    goal_pos_angle = np.array([256, 256, np.pi / 4])  # x, y, theta (in radians)
+    goal_body = PushTEnv.get_goal_pose_body(goal_pos_angle)
+
+    imgs = torch.from_numpy(zarr_data["img"])  # b h w c
+    states = torch.from_numpy(zarr_data["state"])
+    actions = torch.from_numpy(zarr_data["action"])
+
+    # load data indices from which each episode starts and ends
+    from_ids, to_ids = [], []
+    from_idx = 0
+    for to_idx in zarr_data.meta["episode_ends"]:
+        from_ids.append(from_idx)
+        to_ids.append(to_idx)
+        from_idx = to_idx
+
+    num_episodes = len(from_ids)
+
+    ep_dicts = []
+    ep_ids = episodes if episodes else range(num_episodes)
+    for ep_idx, selected_ep_idx in tqdm.tqdm(enumerate(ep_ids)):
+        from_idx = from_ids[selected_ep_idx]
+        to_idx = to_ids[selected_ep_idx]
+        num_frames = to_idx - from_idx
+
+        # sanity check
+        assert (episode_ids[from_idx:to_idx] == ep_idx).all()
+
+        # get image
+        if not keypoints_instead_of_image:
+            image = imgs[from_idx:to_idx]
+            assert image.min() >= 0.0
+            assert image.max() <= 255.0
+            image = image.type(torch.uint8)
+
+        # get state
+        state = states[from_idx:to_idx]
+        agent_pos = state[:, :2]
+        block_pos = state[:, 2:4]
+        block_angle = state[:, 4]
+
+        # get reward, success, done, and (maybe) keypoints
+        reward = torch.zeros(num_frames)
+        success = torch.zeros(num_frames, dtype=torch.bool)
+        if keypoints_instead_of_image:
+            keypoints = torch.zeros(num_frames, 16)  # 8 keypoints each with 2 coords
+        done = torch.zeros(num_frames, dtype=torch.bool)
+        for i in range(num_frames):
+            space = pymunk.Space()
+            space.gravity = 0, 0
+            space.damping = 0
+
+            # Add walls.
+            walls = [
+                PushTEnv.add_segment(space, (5, 506), (5, 5), 2),
+                PushTEnv.add_segment(space, (5, 5), (506, 5), 2),
+                PushTEnv.add_segment(space, (506, 5), (506, 506), 2),
+                PushTEnv.add_segment(space, (5, 506), (506, 506), 2),
+            ]
+            space.add(*walls)
+
+            block_body, block_shapes = PushTEnv.add_tee(space, block_pos[i].tolist(), block_angle[i].item())
+            goal_geom = pymunk_to_shapely(goal_body, block_body.shapes)
+            block_geom = pymunk_to_shapely(block_body, block_body.shapes)
+            intersection_area = goal_geom.intersection(block_geom).area
+            goal_area = goal_geom.area
+            coverage = intersection_area / goal_area
+            reward[i] = np.clip(coverage / success_threshold, 0, 1)
+            success[i] = coverage > success_threshold
+            if keypoints_instead_of_image:
+                keypoints[i] = torch.from_numpy(PushTEnv.get_keypoints(block_shapes).flatten())
+
+        # last step of demonstration is considered done
+        done[-1] = True
+
+        ep_dict = {}
+
+        if not keypoints_instead_of_image:
+            imgs_array = [x.numpy() for x in image]
+            img_key = "observation.image"
+            if video:
+                # save png images in temporary directory
+                tmp_imgs_dir = videos_dir / "tmp_images"
+                save_images_concurrently(imgs_array, tmp_imgs_dir)
+
+                # encode images to a mp4 video
+                fname = f"{img_key}_episode_{ep_idx:06d}.mp4"
+                video_path = videos_dir / fname
+                encode_video_frames(tmp_imgs_dir, video_path, fps, **(encoding or {}))
+
+                # clean temporary images directory
+                shutil.rmtree(tmp_imgs_dir)
+
+                # store the reference to the video frame
+                ep_dict[img_key] = [
+                    {"path": f"videos/{fname}", "timestamp": i / fps} for i in range(num_frames)
+                ]
+            else:
+                ep_dict[img_key] = [PILImage.fromarray(x) for x in imgs_array]
+
+        ep_dict["observation.state"] = agent_pos
+        if keypoints_instead_of_image:
+            ep_dict["observation.environment_state"] = keypoints
+        ep_dict["action"] = actions[from_idx:to_idx]
+        ep_dict["episode_index"] = torch.tensor([ep_idx] * num_frames, dtype=torch.int64)
+        ep_dict["frame_index"] = torch.arange(0, num_frames, 1)
+        ep_dict["timestamp"] = torch.arange(0, num_frames, 1) / fps
+        # ep_dict["next.observation.image"] = image[1:],
+        # ep_dict["next.observation.state"] = agent_pos[1:],
+        # TODO(rcadene)] = verify that reward and done are aligned with image and agent_pos
+        ep_dict["next.reward"] = torch.cat([reward[1:], reward[[-1]]])
+        ep_dict["next.done"] = torch.cat([done[1:], done[[-1]]])
+        ep_dict["next.success"] = torch.cat([success[1:], success[[-1]]])
+        ep_dicts.append(ep_dict)
+    data_dict = concatenate_episodes(ep_dicts)
+
+    total_frames = data_dict["frame_index"].shape[0]
+    data_dict["index"] = torch.arange(0, total_frames, 1)
+    return data_dict
+
+
+def to_hf_dataset(data_dict, video, keypoints_instead_of_image: bool = False):
+    features = {}
+
+    if not keypoints_instead_of_image:
+        if video:
+            features["observation.image"] = VideoFrame()
+        else:
+            features["observation.image"] = Image()
+
+    features["observation.state"] = Sequence(
+        length=data_dict["observation.state"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    if keypoints_instead_of_image:
+        features["observation.environment_state"] = Sequence(
+            length=data_dict["observation.environment_state"].shape[1],
+            feature=Value(dtype="float32", id=None),
+        )
+    features["action"] = Sequence(
+        length=data_dict["action"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    features["episode_index"] = Value(dtype="int64", id=None)
+    features["frame_index"] = Value(dtype="int64", id=None)
+    features["timestamp"] = Value(dtype="float32", id=None)
+    features["next.reward"] = Value(dtype="float32", id=None)
+    features["next.done"] = Value(dtype="bool", id=None)
+    features["next.success"] = Value(dtype="bool", id=None)
+    features["index"] = Value(dtype="int64", id=None)
+
+    hf_dataset = Dataset.from_dict(data_dict, features=Features(features))
+    hf_dataset.set_transform(hf_transform_to_torch)
+    return hf_dataset
+
+
+def from_raw_to_lerobot_format(
+    raw_dir: Path,
+    videos_dir: Path,
+    fps: int | None = None,
+    video: bool = True,
+    episodes: list[int] | None = None,
+    encoding: dict | None = None,
+):
+    # Manually change this to True to use keypoints of the T instead of an image observation (but don't merge
+    # with True). Also make sure to use video = 0 in the `push_dataset_to_hub.py` script.
+    keypoints_instead_of_image = False
+
+    # sanity check
+    check_format(raw_dir)
+
+    if fps is None:
+        fps = 10
+
+    data_dict = load_from_raw(raw_dir, videos_dir, fps, video, episodes, keypoints_instead_of_image, encoding)
+    hf_dataset = to_hf_dataset(data_dict, video, keypoints_instead_of_image)
+    episode_data_index = calculate_episode_data_index(hf_dataset)
+    info = {
+        "codebase_version": CODEBASE_VERSION,
+        "fps": fps,
+        "video": video if not keypoints_instead_of_image else 0,
+    }
+    if video:
+        info["encoding"] = get_default_encoding()
+
+    return hf_dataset, episode_data_index, info

lerobot/common/datasets/push_dataset_to_hub/umi_zarr_format.py

@@ -0,0 +1,234 @@
+#!/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.
+"""Process UMI (Universal Manipulation Interface) data stored in Zarr format like in: https://github.com/real-stanford/universal_manipulation_interface"""
+
+import logging
+import shutil
+from pathlib import Path
+
+import torch
+import tqdm
+import zarr
+from datasets import Dataset, Features, Image, Sequence, Value
+from PIL import Image as PILImage
+
+from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
+from lerobot.common.datasets.push_dataset_to_hub._umi_imagecodecs_numcodecs import register_codecs
+from lerobot.common.datasets.push_dataset_to_hub.utils import (
+    calculate_episode_data_index,
+    concatenate_episodes,
+    get_default_encoding,
+    save_images_concurrently,
+)
+from lerobot.common.datasets.utils import (
+    hf_transform_to_torch,
+)
+from lerobot.common.datasets.video_utils import VideoFrame, encode_video_frames
+
+
+def check_format(raw_dir) -> bool:
+    zarr_path = raw_dir / "cup_in_the_wild.zarr"
+    zarr_data = zarr.open(zarr_path, mode="r")
+
+    required_datasets = {
+        "data/robot0_demo_end_pose",
+        "data/robot0_demo_start_pose",
+        "data/robot0_eef_pos",
+        "data/robot0_eef_rot_axis_angle",
+        "data/robot0_gripper_width",
+        "meta/episode_ends",
+        "data/camera0_rgb",
+    }
+    for dataset in required_datasets:
+        if dataset not in zarr_data:
+            return False
+
+    # mandatory to access zarr_data
+    register_codecs()
+    nb_frames = zarr_data["data/camera0_rgb"].shape[0]
+
+    required_datasets.remove("meta/episode_ends")
+    assert all(nb_frames == zarr_data[dataset].shape[0] for dataset in required_datasets)
+
+
+def load_from_raw(
+    raw_dir: Path,
+    videos_dir: Path,
+    fps: int,
+    video: bool,
+    episodes: list[int] | None = None,
+    encoding: dict | None = None,
+):
+    zarr_path = raw_dir / "cup_in_the_wild.zarr"
+    zarr_data = zarr.open(zarr_path, mode="r")
+
+    # We process the image data separately because it is too large to fit in memory
+    end_pose = torch.from_numpy(zarr_data["data/robot0_demo_end_pose"][:])
+    start_pos = torch.from_numpy(zarr_data["data/robot0_demo_start_pose"][:])
+    eff_pos = torch.from_numpy(zarr_data["data/robot0_eef_pos"][:])
+    eff_rot_axis_angle = torch.from_numpy(zarr_data["data/robot0_eef_rot_axis_angle"][:])
+    gripper_width = torch.from_numpy(zarr_data["data/robot0_gripper_width"][:])
+
+    states_pos = torch.cat([eff_pos, eff_rot_axis_angle], dim=1)
+    states = torch.cat([states_pos, gripper_width], dim=1)
+
+    episode_ends = zarr_data["meta/episode_ends"][:]
+    num_episodes = episode_ends.shape[0]
+
+    # We convert it in torch tensor later because the jit function does not support torch tensors
+    episode_ends = torch.from_numpy(episode_ends)
+
+    # load data indices from which each episode starts and ends
+    from_ids, to_ids = [], []
+    from_idx = 0
+    for to_idx in episode_ends:
+        from_ids.append(from_idx)
+        to_ids.append(to_idx)
+        from_idx = to_idx
+
+    ep_dicts_dir = videos_dir / "ep_dicts"
+    ep_dicts_dir.mkdir(exist_ok=True, parents=True)
+    ep_dicts = []
+
+    ep_ids = episodes if episodes else range(num_episodes)
+    for ep_idx, selected_ep_idx in tqdm.tqdm(enumerate(ep_ids)):
+        ep_dict_path = ep_dicts_dir / f"{ep_idx}"
+        if not ep_dict_path.is_file():
+            from_idx = from_ids[selected_ep_idx]
+            to_idx = to_ids[selected_ep_idx]
+            num_frames = to_idx - from_idx
+
+            # TODO(rcadene): save temporary images of the episode?
+
+            state = states[from_idx:to_idx]
+
+            ep_dict = {}
+
+            # load 57MB of images in RAM (400x224x224x3 uint8)
+            imgs_array = zarr_data["data/camera0_rgb"][from_idx:to_idx]
+            img_key = "observation.image"
+            if video:
+                fname = f"{img_key}_episode_{ep_idx:06d}.mp4"
+                video_path = videos_dir / fname
+                if not video_path.is_file():
+                    # save png images in temporary directory
+                    tmp_imgs_dir = videos_dir / "tmp_images"
+                    save_images_concurrently(imgs_array, tmp_imgs_dir)
+
+                    # encode images to a mp4 video
+                    encode_video_frames(tmp_imgs_dir, video_path, fps, **(encoding or {}))
+
+                    # clean temporary images directory
+                    shutil.rmtree(tmp_imgs_dir)
+
+                # store the reference to the video frame
+                ep_dict[img_key] = [
+                    {"path": f"videos/{fname}", "timestamp": i / fps} for i in range(num_frames)
+                ]
+            else:
+                ep_dict[img_key] = [PILImage.fromarray(x) for x in imgs_array]
+
+            ep_dict["observation.state"] = state
+            ep_dict["episode_index"] = torch.tensor([ep_idx] * num_frames, dtype=torch.int64)
+            ep_dict["frame_index"] = torch.arange(0, num_frames, 1)
+            ep_dict["timestamp"] = torch.arange(0, num_frames, 1) / fps
+            ep_dict["episode_data_index_from"] = torch.tensor([from_idx] * num_frames)
+            ep_dict["episode_data_index_to"] = torch.tensor([from_idx + num_frames] * num_frames)
+            ep_dict["end_pose"] = end_pose[from_idx:to_idx]
+            ep_dict["start_pos"] = start_pos[from_idx:to_idx]
+            ep_dict["gripper_width"] = gripper_width[from_idx:to_idx]
+            torch.save(ep_dict, ep_dict_path)
+        else:
+            ep_dict = torch.load(ep_dict_path)
+
+        ep_dicts.append(ep_dict)
+
+    data_dict = concatenate_episodes(ep_dicts)
+
+    total_frames = data_dict["frame_index"].shape[0]
+    data_dict["index"] = torch.arange(0, total_frames, 1)
+    return data_dict
+
+
+def to_hf_dataset(data_dict, video):
+    features = {}
+
+    if video:
+        features["observation.image"] = VideoFrame()
+    else:
+        features["observation.image"] = Image()
+
+    features["observation.state"] = Sequence(
+        length=data_dict["observation.state"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    features["episode_index"] = Value(dtype="int64", id=None)
+    features["frame_index"] = Value(dtype="int64", id=None)
+    features["timestamp"] = Value(dtype="float32", id=None)
+    features["index"] = Value(dtype="int64", id=None)
+    features["episode_data_index_from"] = Value(dtype="int64", id=None)
+    features["episode_data_index_to"] = Value(dtype="int64", id=None)
+    # `start_pos` and `end_pos` respectively represent the positions of the end-effector
+    # at the beginning and the end of the episode.
+    # `gripper_width` indicates the distance between the grippers, and this value is included
+    # in the state vector, which comprises the concatenation of the end-effector position
+    # and gripper width.
+    features["end_pose"] = Sequence(
+        length=data_dict["end_pose"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    features["start_pos"] = Sequence(
+        length=data_dict["start_pos"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    features["gripper_width"] = Sequence(
+        length=data_dict["gripper_width"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+
+    hf_dataset = Dataset.from_dict(data_dict, features=Features(features))
+    hf_dataset.set_transform(hf_transform_to_torch)
+    return hf_dataset
+
+
+def from_raw_to_lerobot_format(
+    raw_dir: Path,
+    videos_dir: Path,
+    fps: int | None = None,
+    video: bool = True,
+    episodes: list[int] | None = None,
+    encoding: dict | None = None,
+):
+    # sanity check
+    check_format(raw_dir)
+
+    if fps is None:
+        # For umi cup in the wild: https://arxiv.org/pdf/2402.10329#table.caption.16
+        fps = 10
+
+    if not video:
+        logging.warning(
+            "Generating UMI dataset without `video=True` creates ~150GB on disk and requires ~80GB in RAM."
+        )
+
+    data_dict = load_from_raw(raw_dir, videos_dir, fps, video, episodes, encoding)
+    hf_dataset = to_hf_dataset(data_dict, video)
+    episode_data_index = calculate_episode_data_index(hf_dataset)
+    info = {
+        "codebase_version": CODEBASE_VERSION,
+        "fps": fps,
+        "video": video,
+    }
+    if video:
+        info["encoding"] = get_default_encoding()
+
+    return hf_dataset, episode_data_index, info

lerobot/common/datasets/push_dataset_to_hub/xarm_pkl_format.py

@@ -0,0 +1,200 @@
+#!/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.
+"""Process pickle files formatted like in: https://github.com/fyhMer/fowm"""
+
+import pickle
+import shutil
+from pathlib import Path
+
+import einops
+import torch
+import tqdm
+from datasets import Dataset, Features, Image, Sequence, Value
+from PIL import Image as PILImage
+
+from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
+from lerobot.common.datasets.push_dataset_to_hub.utils import (
+    calculate_episode_data_index,
+    concatenate_episodes,
+    get_default_encoding,
+    save_images_concurrently,
+)
+from lerobot.common.datasets.utils import (
+    hf_transform_to_torch,
+)
+from lerobot.common.datasets.video_utils import VideoFrame, encode_video_frames
+
+
+def check_format(raw_dir):
+    keys = {"actions", "rewards", "dones"}
+    nested_keys = {"observations": {"rgb", "state"}, "next_observations": {"rgb", "state"}}
+
+    xarm_files = list(raw_dir.glob("*.pkl"))
+    assert len(xarm_files) > 0
+
+    with open(xarm_files[0], "rb") as f:
+        dataset_dict = pickle.load(f)
+
+    assert isinstance(dataset_dict, dict)
+    assert all(k in dataset_dict for k in keys)
+
+    # Check for consistent lengths in nested keys
+    expected_len = len(dataset_dict["actions"])
+    assert all(len(dataset_dict[key]) == expected_len for key in keys if key in dataset_dict)
+
+    for key, subkeys in nested_keys.items():
+        nested_dict = dataset_dict.get(key, {})
+        assert all(len(nested_dict[subkey]) == expected_len for subkey in subkeys if subkey in nested_dict)
+
+
+def load_from_raw(
+    raw_dir: Path,
+    videos_dir: Path,
+    fps: int,
+    video: bool,
+    episodes: list[int] | None = None,
+    encoding: dict | None = None,
+):
+    pkl_path = raw_dir / "buffer.pkl"
+
+    with open(pkl_path, "rb") as f:
+        pkl_data = pickle.load(f)
+
+    # load data indices from which each episode starts and ends
+    from_ids, to_ids = [], []
+    from_idx, to_idx = 0, 0
+    for done in pkl_data["dones"]:
+        to_idx += 1
+        if not done:
+            continue
+        from_ids.append(from_idx)
+        to_ids.append(to_idx)
+        from_idx = to_idx
+
+    num_episodes = len(from_ids)
+
+    ep_dicts = []
+    ep_ids = episodes if episodes else range(num_episodes)
+    for ep_idx, selected_ep_idx in tqdm.tqdm(enumerate(ep_ids)):
+        from_idx = from_ids[selected_ep_idx]
+        to_idx = to_ids[selected_ep_idx]
+        num_frames = to_idx - from_idx
+
+        image = torch.tensor(pkl_data["observations"]["rgb"][from_idx:to_idx])
+        image = einops.rearrange(image, "b c h w -> b h w c")
+        state = torch.tensor(pkl_data["observations"]["state"][from_idx:to_idx])
+        action = torch.tensor(pkl_data["actions"][from_idx:to_idx])
+        # TODO(rcadene): we have a missing last frame which is the observation when the env is done
+        # it is critical to have this frame for tdmpc to predict a "done observation/state"
+        # next_image = torch.tensor(pkl_data["next_observations"]["rgb"][from_idx:to_idx])
+        # next_state = torch.tensor(pkl_data["next_observations"]["state"][from_idx:to_idx])
+        next_reward = torch.tensor(pkl_data["rewards"][from_idx:to_idx])
+        next_done = torch.tensor(pkl_data["dones"][from_idx:to_idx])
+
+        ep_dict = {}
+
+        imgs_array = [x.numpy() for x in image]
+        img_key = "observation.image"
+        if video:
+            # save png images in temporary directory
+            tmp_imgs_dir = videos_dir / "tmp_images"
+            save_images_concurrently(imgs_array, tmp_imgs_dir)
+
+            # encode images to a mp4 video
+            fname = f"{img_key}_episode_{ep_idx:06d}.mp4"
+            video_path = videos_dir / fname
+            encode_video_frames(tmp_imgs_dir, video_path, fps, **(encoding or {}))
+
+            # clean temporary images directory
+            shutil.rmtree(tmp_imgs_dir)
+
+            # store the reference to the video frame
+            ep_dict[img_key] = [{"path": f"videos/{fname}", "timestamp": i / fps} for i in range(num_frames)]
+        else:
+            ep_dict[img_key] = [PILImage.fromarray(x) for x in imgs_array]
+
+        ep_dict["observation.state"] = state
+        ep_dict["action"] = action
+        ep_dict["episode_index"] = torch.tensor([ep_idx] * num_frames, dtype=torch.int64)
+        ep_dict["frame_index"] = torch.arange(0, num_frames, 1)
+        ep_dict["timestamp"] = torch.arange(0, num_frames, 1) / fps
+        # ep_dict["next.observation.image"] = next_image
+        # ep_dict["next.observation.state"] = next_state
+        ep_dict["next.reward"] = next_reward
+        ep_dict["next.done"] = next_done
+        ep_dicts.append(ep_dict)
+
+    data_dict = concatenate_episodes(ep_dicts)
+
+    total_frames = data_dict["frame_index"].shape[0]
+    data_dict["index"] = torch.arange(0, total_frames, 1)
+    return data_dict
+
+
+def to_hf_dataset(data_dict, video):
+    features = {}
+
+    if video:
+        features["observation.image"] = VideoFrame()
+    else:
+        features["observation.image"] = Image()
+
+    features["observation.state"] = Sequence(
+        length=data_dict["observation.state"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    features["action"] = Sequence(
+        length=data_dict["action"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    features["episode_index"] = Value(dtype="int64", id=None)
+    features["frame_index"] = Value(dtype="int64", id=None)
+    features["timestamp"] = Value(dtype="float32", id=None)
+    features["next.reward"] = Value(dtype="float32", id=None)
+    features["next.done"] = Value(dtype="bool", id=None)
+    features["index"] = Value(dtype="int64", id=None)
+    # TODO(rcadene): add success
+    # features["next.success"] = Value(dtype='bool', id=None)
+
+    hf_dataset = Dataset.from_dict(data_dict, features=Features(features))
+    hf_dataset.set_transform(hf_transform_to_torch)
+    return hf_dataset
+
+
+def from_raw_to_lerobot_format(
+    raw_dir: Path,
+    videos_dir: Path,
+    fps: int | None = None,
+    video: bool = True,
+    episodes: list[int] | None = None,
+    encoding: dict | None = None,
+):
+    # sanity check
+    check_format(raw_dir)
+
+    if fps is None:
+        fps = 15
+
+    data_dict = load_from_raw(raw_dir, videos_dir, fps, video, episodes, encoding)
+    hf_dataset = to_hf_dataset(data_dict, video)
+    episode_data_index = calculate_episode_data_index(hf_dataset)
+    info = {
+        "codebase_version": CODEBASE_VERSION,
+        "fps": fps,
+        "video": video,
+    }
+    if video:
+        info["encoding"] = get_default_encoding()
+
+    return hf_dataset, episode_data_index, info

lerobot/common/datasets/sampler.py

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

lerobot/common/datasets/utils.py

@@ -17,23 +17,18 @@ import contextlib
 import importlib.resources
 import json
 import logging
-import shutil
-import subprocess
-import tempfile
 from collections.abc import Iterator
+from itertools import accumulate
 from pathlib import Path
 from pprint import pformat
 from types import SimpleNamespace
 from typing import Any
 
 import datasets
+import jsonlines
 import numpy as np
 import packaging.version
-import pandas
-import pandas as pd
-import pyarrow.parquet as pq
 import torch
-from datasets import Dataset, concatenate_datasets
 from datasets.table import embed_table_storage
 from huggingface_hub import DatasetCard, DatasetCardData, HfApi
 from huggingface_hub.errors import RevisionNotFoundError
@@ -45,27 +40,21 @@ from lerobot.common.datasets.backward_compatibility import (
     BackwardCompatibilityError,
     ForwardCompatibilityError,
 )
-from lerobot.common.robot_devices.robots.utils import Robot
+from lerobot.common.robots.utils import Robot
 from lerobot.common.utils.utils import is_valid_numpy_dtype_string
-from lerobot.configs.types import FeatureType, PolicyFeature
+from lerobot.configs.types import DictLike, FeatureType, PolicyFeature
 
-DEFAULT_CHUNK_SIZE = 1000  # Max number of files per chunk
-DEFAULT_DATA_FILE_SIZE_IN_MB = 100  # Max size per file
-DEFAULT_VIDEO_FILE_SIZE_IN_MB = 500  # Max size per file
+DEFAULT_CHUNK_SIZE = 1000  # Max number of episodes per chunk
 
 INFO_PATH = "meta/info.json"
+EPISODES_PATH = "meta/episodes.jsonl"
 STATS_PATH = "meta/stats.json"
+EPISODES_STATS_PATH = "meta/episodes_stats.jsonl"
+TASKS_PATH = "meta/tasks.jsonl"
 
-EPISODES_DIR = "meta/episodes"
-DATA_DIR = "data"
-VIDEO_DIR = "videos"
-
-CHUNK_FILE_PATTERN = "chunk-{chunk_index:03d}/file-{file_index:03d}"
-DEFAULT_TASKS_PATH = "meta/tasks.parquet"
-DEFAULT_EPISODES_PATH = EPISODES_DIR + "/" + CHUNK_FILE_PATTERN + ".parquet"
-DEFAULT_DATA_PATH = DATA_DIR + "/" + CHUNK_FILE_PATTERN + ".parquet"
-DEFAULT_VIDEO_PATH = VIDEO_DIR + "/{video_key}/" + CHUNK_FILE_PATTERN + ".mp4"
-DEFAULT_IMAGE_PATH = "images/{image_key}/episode-{episode_index:06d}/frame-{frame_index:06d}.png"
+DEFAULT_VIDEO_PATH = "videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4"
+DEFAULT_PARQUET_PATH = "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet"
+DEFAULT_IMAGE_PATH = "images/{image_key}/episode_{episode_index:06d}/frame_{frame_index:06d}.png"
 
 DATASET_CARD_TEMPLATE = """
 ---
@@ -86,115 +75,6 @@ DEFAULT_FEATURES = {
 }
 
 
-def get_parquet_file_size_in_mb(parquet_path):
-    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_pd_dataframe_size_in_mb(df: pandas.DataFrame) -> int:
-    # TODO(rcadene): unused?
-    memory_usage_bytes = df.memory_usage(deep=True).sum()
-    return memory_usage_bytes / (1024**2)
-
-
-def update_chunk_file_indices(chunk_idx: int, file_idx: int, chunks_size: 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) -> 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
-    """
-    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)) for path in paths]
-    return concatenate_datasets(datasets)
-
-
-def get_parquet_num_frames(parquet_path):
-    metadata = pq.read_metadata(parquet_path)
-    return metadata.num_rows
-
-
-def get_video_size_in_mb(mp4_path: Path):
-    file_size_bytes = mp4_path.stat().st_size
-    file_size_mb = file_size_bytes / (1024**2)
-    return file_size_mb
-
-
-def concat_video_files(paths_to_cat: list[Path], root: Path, video_key: str, chunk_idx: int, file_idx: int):
-    # TODO(rcadene): move to video_utils.py
-    # TODO(rcadene): add docstring
-    tmp_dir = Path(tempfile.mkdtemp(dir=root))
-    # Create a text file with the list of files to concatenate
-    path_concat_video_files = tmp_dir / "concat_video_files.txt"
-    with open(path_concat_video_files, "w") as f:
-        for ep_path in paths_to_cat:
-            f.write(f"file '{str(ep_path)}'\n")
-
-    path_tmp_output = tmp_dir / "tmp_output.mp4"
-    command = [
-        "ffmpeg",
-        "-y",
-        "-f",
-        "concat",
-        "-safe",
-        "0",
-        "-i",
-        str(path_concat_video_files),
-        "-c",
-        "copy",
-        str(path_tmp_output),
-    ]
-    subprocess.run(command, check=True)
-
-    output_path = root / DEFAULT_VIDEO_PATH.format(
-        video_key=video_key, chunk_index=chunk_idx, file_index=file_idx
-    )
-    output_path.parent.mkdir(parents=True, exist_ok=True)
-    shutil.move(str(path_tmp_output), str(output_path))
-    shutil.rmtree(str(tmp_dir))
-
-
-def get_video_duration_in_s(mp4_file: Path):
-    # TODO(rcadene): move to video_utils.py
-    command = [
-        "ffprobe",
-        "-v",
-        "error",
-        "-show_entries",
-        "format=duration",
-        "-of",
-        "default=noprint_wrappers=1:nokey=1",
-        str(mp4_file),
-    ]
-    result = subprocess.run(
-        command,
-        stdout=subprocess.PIPE,
-        stderr=subprocess.STDOUT,
-    )
-    return float(result.stdout)
-
-
 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.
 
@@ -227,13 +107,23 @@ def unflatten_dict(d: dict, sep: str = "/") -> dict:
     return outdict
 
 
+def get_nested_item(obj: DictLike, flattened_key: str, sep: str = "/") -> Any:
+    split_keys = flattened_key.split(sep)
+    getter = obj[split_keys[0]]
+    if len(split_keys) == 1:
+        return getter
+
+    for key in split_keys[1:]:
+        getter = getter[key]
+
+    return getter
+
+
 def serialize_dict(stats: dict[str, torch.Tensor | np.ndarray | dict]) -> dict:
     serialized_dict = {}
     for key, value in flatten_dict(stats).items():
         if isinstance(value, (torch.Tensor, np.ndarray)):
             serialized_dict[key] = value.tolist()
-        elif isinstance(value, list) and isinstance(value[0], (int, float, list)):
-            serialized_dict[key] = value
         elif isinstance(value, np.generic):
             serialized_dict[key] = value.item()
         elif isinstance(value, (int, float)):
@@ -263,6 +153,23 @@ 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):
     write_json(info, local_dir / INFO_PATH)
 
@@ -291,48 +198,49 @@ def load_stats(local_dir: Path) -> dict[str, dict[str, np.ndarray]]:
     return cast_stats_to_numpy(stats)
 
 
-def write_hf_dataset(hf_dataset: Dataset, local_dir: Path):
-    if get_hf_dataset_size_in_mb(hf_dataset) > DEFAULT_DATA_FILE_SIZE_IN_MB:
-        raise NotImplementedError("Contact a maintainer.")
-
-    path = local_dir / DEFAULT_DATA_PATH.format(chunk_index=0, file_index=0)
-    path.parent.mkdir(parents=True, exist_ok=True)
-    hf_dataset.to_parquet(path)
+def write_task(task_index: int, task: dict, local_dir: Path):
+    task_dict = {
+        "task_index": task_index,
+        "task": task,
+    }
+    append_jsonlines(task_dict, local_dir / TASKS_PATH)
 
 
-def write_tasks(tasks: pandas.DataFrame, local_dir: Path):
-    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):
-    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):
-    if get_hf_dataset_size_in_mb(episodes) > DEFAULT_DATA_FILE_SIZE_IN_MB:
-        raise NotImplementedError("Contact a maintainer.")
-
-    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 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 write_episode_stats(episode_index: int, episode_stats: dict, local_dir: Path):
+    # We wrap episode_stats in a dictionary since `episode_stats["episode_index"]`
+    # is a dictionary of stats and not an integer.
+    episode_stats = {"episode_index": episode_index, "stats": serialize_dict(episode_stats)}
+    append_jsonlines(episode_stats, local_dir / EPISODES_STATS_PATH)
+
+
+def load_episodes_stats(local_dir: Path) -> dict:
+    episodes_stats = load_jsonlines(local_dir / EPISODES_STATS_PATH)
+    return {
+        item["episode_index"]: cast_stats_to_numpy(item["stats"])
+        for item in sorted(episodes_stats, key=lambda x: x["episode_index"])
+    }
 
 
 def backward_compatible_episodes_stats(
     stats: dict[str, dict[str, np.ndarray]], episodes: list[int]
 ) -> dict[str, dict[str, np.ndarray]]:
-    return dict.fromkeys(episodes, stats)
+    return {ep_idx: stats for ep_idx in episodes}
 
 
 def load_image_as_numpy(
@@ -480,7 +388,6 @@ def get_hf_features_from_features(features: dict) -> datasets.Features:
 
 
 def get_features_from_robot(robot: Robot, use_videos: bool = True) -> dict:
-    # TODO(rcadene): add fps for each feature
     camera_ft = {}
     if robot.cameras:
         camera_ft = {
@@ -534,17 +441,31 @@ 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,
-        "data_files_size_in_mb": DEFAULT_DATA_FILE_SIZE_IN_MB,
-        "video_files_size_in_mb": DEFAULT_VIDEO_FILE_SIZE_IN_MB,
         "fps": fps,
         "splits": {},
-        "data_path": DEFAULT_DATA_PATH,
+        "data_path": DEFAULT_PARQUET_PATH,
         "video_path": DEFAULT_VIDEO_PATH if use_videos else None,
         "features": features,
     }
 
 
+def get_episode_data_index(
+    episode_dicts: dict[dict], episodes: list[int] | None = None
+) -> dict[str, torch.Tensor]:
+    episode_lengths = {ep_idx: ep_dict["length"] for ep_idx, ep_dict in episode_dicts.items()}
+    if episodes is not None:
+        episode_lengths = {ep_idx: episode_lengths[ep_idx] for ep_idx in episodes}
+
+    cumulative_lengths = list(accumulate(episode_lengths.values()))
+    return {
+        "from": torch.LongTensor([0] + cumulative_lengths[:-1]),
+        "to": torch.LongTensor(cumulative_lengths),
+    }
+
+
 def check_timestamps_sync(
     timestamps: np.ndarray,
     episode_indices: np.ndarray,
@@ -890,11 +811,3 @@ def validate_episode_buffer(episode_buffer: dict, total_episodes: int, features:
             f"In episode_buffer not in features: {buffer_keys - set(features)}"
             f"In features not in episode_buffer: {set(features) - buffer_keys}"
         )
-
-
-def to_parquet_with_hf_images(df: pandas.DataFrame, path: Path):
-    """ 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 formated 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)

lerobot/common/datasets/v2/batch_convert_dataset_v1_to_v2.py

@@ -27,7 +27,7 @@ from textwrap import dedent
 
 from lerobot import available_datasets
 from lerobot.common.datasets.v2.convert_dataset_v1_to_v2 import convert_dataset
-from lerobot.common.robot_devices.robots.configs import AlohaRobotConfig
+from lerobot.common.robots.aloha.configuration_aloha import AlohaRobotConfig
 
 LOCAL_DIR = Path("data/")
 

lerobot/common/datasets/v2/convert_dataset_v1_to_v2.py

@@ -121,12 +121,12 @@ from safetensors.torch import load_file
 
 from lerobot.common.datasets.utils import (
     DEFAULT_CHUNK_SIZE,
-    DEFAULT_DATA_PATH,
+    DEFAULT_PARQUET_PATH,
     DEFAULT_VIDEO_PATH,
+    EPISODES_PATH,
     INFO_PATH,
-    LEGACY_EPISODES_PATH,
-    LEGACY_TASKS_PATH,
     STATS_PATH,
+    TASKS_PATH,
     create_branch,
     create_lerobot_dataset_card,
     flatten_dict,
@@ -141,8 +141,8 @@ from lerobot.common.datasets.video_utils import (
     get_image_pixel_channels,
     get_video_info,
 )
-from lerobot.common.robot_devices.robots.configs import RobotConfig
-from lerobot.common.robot_devices.robots.utils import make_robot_config
+from lerobot.common.robots import RobotConfig
+from lerobot.common.robots.utils import make_robot_config
 
 V16 = "v1.6"
 V20 = "v2.0"
@@ -291,12 +291,14 @@ def split_parquet_by_episodes(
     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_DATA_PATH.split("/")[:-1]).format(episode_chunk=ep_chunk)
+        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_DATA_PATH.format(episode_chunk=ep_chunk, episode_index=ep_idx)
+            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
@@ -479,7 +481,7 @@ def convert_dataset(
 
     # Tasks
     if single_task:
-        tasks_by_episodes = dict.fromkeys(episode_indices, single_task)
+        tasks_by_episodes = {ep_idx: single_task for ep_idx in episode_indices}
         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:
@@ -494,7 +496,7 @@ def convert_dataset(
 
     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 / LEGACY_TASKS_PATH)
+    write_jsonlines(tasks, v20_dir / TASKS_PATH)
     features["task_index"] = {
         "dtype": "int64",
         "shape": (1,),
@@ -544,7 +546,7 @@ def convert_dataset(
         {"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 / LEGACY_EPISODES_PATH)
+    write_jsonlines(episodes, v20_dir / EPISODES_PATH)
 
     # Assemble metadata v2.0
     metadata_v2_0 = {
@@ -558,7 +560,7 @@ def convert_dataset(
         "chunks_size": DEFAULT_CHUNK_SIZE,
         "fps": metadata_v1["fps"],
         "splits": {"train": f"0:{total_episodes}"},
-        "data_path": DEFAULT_DATA_PATH,
+        "data_path": DEFAULT_PARQUET_PATH,
         "video_path": DEFAULT_VIDEO_PATH if video_keys else None,
         "features": features,
     }

lerobot/common/datasets/v21/convert_dataset_v20_to_v21.py

@@ -37,7 +37,7 @@ import logging
 from huggingface_hub import HfApi
 
 from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDataset
-from lerobot.common.datasets.utils import LEGACY_EPISODES_STATS_PATH, STATS_PATH, load_stats, write_info
+from lerobot.common.datasets.utils import EPISODES_STATS_PATH, STATS_PATH, load_stats, write_info
 from lerobot.common.datasets.v21.convert_stats import check_aggregate_stats, convert_stats
 
 V20 = "v2.0"
@@ -61,8 +61,8 @@ def convert_dataset(
     with SuppressWarnings():
         dataset = LeRobotDataset(repo_id, revision=V20, force_cache_sync=True)
 
-    if (dataset.root / LEGACY_EPISODES_STATS_PATH).is_file():
-        (dataset.root / LEGACY_EPISODES_STATS_PATH).unlink()
+    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)

lerobot/common/datasets/v21/convert_stats.py

@@ -19,7 +19,7 @@ from tqdm import tqdm
 
 from lerobot.common.datasets.compute_stats import aggregate_stats, get_feature_stats, sample_indices
 from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.common.datasets.utils import legacy_write_episode_stats
+from lerobot.common.datasets.utils import write_episode_stats
 
 
 def sample_episode_video_frames(dataset: LeRobotDataset, episode_index: int, ft_key: str) -> np.ndarray:
@@ -72,7 +72,7 @@ def convert_stats(dataset: LeRobotDataset, num_workers: int = 0):
             convert_episode_stats(dataset, ep_idx)
 
     for ep_idx in tqdm(range(total_episodes)):
-        legacy_write_episode_stats(ep_idx, dataset.meta.episodes_stats[ep_idx], dataset.root)
+        write_episode_stats(ep_idx, dataset.meta.episodes_stats[ep_idx], dataset.root)
 
 
 def check_aggregate_stats(

lerobot/common/datasets/v30/convert_dataset_v21_to_v30.py

@@ -1,452 +0,0 @@
-"""
-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 "v2.1".
-
-Usage:
-
-```bash
-python lerobot/common/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.common.constants import HF_LEROBOT_HOME
-from lerobot.common.datasets.compute_stats import aggregate_stats
-from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDataset
-from lerobot.common.datasets.utils import (
-    DEFAULT_CHUNK_SIZE,
-    DEFAULT_DATA_FILE_SIZE_IN_MB,
-    DEFAULT_DATA_PATH,
-    DEFAULT_VIDEO_FILE_SIZE_IN_MB,
-    DEFAULT_VIDEO_PATH,
-    cast_stats_to_numpy,
-    concat_video_files,
-    flatten_dict,
-    get_parquet_file_size_in_mb,
-    get_parquet_num_frames,
-    get_video_duration_in_s,
-    get_video_size_in_mb,
-    load_info,
-    update_chunk_file_indices,
-    write_episodes,
-    write_info,
-    write_stats,
-    write_tasks,
-)
-
-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"
-
-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, new_root):
-    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 < DEFAULT_DATA_FILE_SIZE_IN_MB:
-            paths_to_cat.append(ep_path)
-            continue
-
-        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_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)
-        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):
-    # 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)
-        ep_metadata = {
-            "episode_index": ep_idx,
-            f"videos/{video_key}/chunk_index": chunk_idx,
-            f"videos/{video_key}/file_index": file_idx,
-            f"videos/{video_key}/from_timestamp": duration_in_s,
-            f"videos/{video_key}/to_timestamp": duration_in_s + ep_duration_in_s,
-        }
-        size_in_mb += ep_size_in_mb
-        duration_in_s += ep_duration_in_s
-        episodes_metadata.append(ep_metadata)
-        ep_idx += 1
-
-        if size_in_mb < DEFAULT_VIDEO_FILE_SIZE_IN_MB:
-            paths_to_cat.append(ep_path)
-            continue
-
-        concat_video_files(paths_to_cat, new_root, video_key, chunk_idx, file_idx)
-
-        # Reset for the next file
-        size_in_mb = ep_size_in_mb
-        duration_in_s = ep_duration_in_s
-        paths_to_cat = [ep_path]
-
-        chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, DEFAULT_CHUNK_SIZE)
-
-    # Write remaining videos if any
-    if paths_to_cat:
-        concat_video_files(paths_to_cat, new_root, video_key, chunk_idx, 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):
-    info = load_info(root)
-    info["codebase_version"] = "v3.0"
-    del info["total_chunks"]
-    del info["total_videos"]
-    info["data_files_size_in_mb"] = DEFAULT_DATA_FILE_SIZE_IN_MB
-    info["video_files_size_in_mb"] = DEFAULT_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,
-    num_workers: int = 4,
-):
-    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 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)
-    convert_tasks(root, new_root)
-    episodes_metadata = convert_data(root, new_root)
-    episodes_videos_metadata = convert_videos(root, new_root)
-    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(
-        "--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))

lerobot/common/datasets/video_utils.py

@@ -13,7 +13,6 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-import importlib
 import json
 import logging
 import subprocess
@@ -30,46 +29,6 @@ from datasets.features.features import register_feature
 from PIL import Image
 
 
-def get_safe_default_codec():
-    if importlib.util.find_spec("torchcodec"):
-        return "torchcodec"
-    else:
-        logging.warning(
-            "'torchcodec' is not available in your platform, falling back to 'pyav' as a default decoder"
-        )
-        return "pyav"
-
-
-def decode_video_frames(
-    video_path: Path | str,
-    timestamps: list[float],
-    tolerance_s: float,
-    backend: str | None = None,
-) -> torch.Tensor:
-    """
-    Decodes video frames using the specified backend.
-
-    Args:
-        video_path (Path): Path to the video file.
-        timestamps (list[float]): List of timestamps to extract frames.
-        tolerance_s (float): Allowed deviation in seconds for frame retrieval.
-        backend (str, optional): Backend to use for decoding. Defaults to "torchcodec" when available in the platform; otherwise, defaults to "pyav"..
-
-    Returns:
-        torch.Tensor: Decoded frames.
-
-    Currently supports torchcodec on cpu and pyav.
-    """
-    if backend is None:
-        backend = get_safe_default_codec()
-    if backend == "torchcodec":
-        return decode_video_frames_torchcodec(video_path, timestamps, tolerance_s)
-    elif backend in ["pyav", "video_reader"]:
-        return decode_video_frames_torchvision(video_path, timestamps, tolerance_s, backend)
-    else:
-        raise ValueError(f"Unsupported video backend: {backend}")
-
-
 def decode_video_frames_torchvision(
     video_path: Path | str,
     timestamps: list[float],
@@ -155,7 +114,6 @@ def decode_video_frames_torchvision(
     )
 
     # get closest frames to the query timestamps
-    # TODO(rcadene): remove torch.stack
     closest_frames = torch.stack([loaded_frames[idx] for idx in argmin_])
     closest_ts = loaded_ts[argmin_]
 
@@ -169,81 +127,6 @@ def decode_video_frames_torchvision(
     return closest_frames
 
 
-def decode_video_frames_torchcodec(
-    video_path: Path | str,
-    timestamps: list[float],
-    tolerance_s: float,
-    device: str = "cpu",
-    log_loaded_timestamps: bool = False,
-) -> torch.Tensor:
-    """Loads frames associated with the requested timestamps of a video using torchcodec.
-
-    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,
-    the encoder stores a reference frame (or a key frame) and subsequent frames as differences relative to
-    that key frame. As a consequence, to access a requested frame, we need to load the preceding key frame,
-    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 importlib.util.find_spec("torchcodec"):
-        from torchcodec.decoders import VideoDecoder
-    else:
-        raise ImportError("torchcodec is required but not available.")
-
-    # initialize video decoder
-    decoder = VideoDecoder(video_path, device=device, seek_mode="approximate")
-    loaded_frames = []
-    loaded_ts = []
-    # get metadata for frame information
-    metadata = decoder.metadata
-    average_fps = metadata.average_fps
-
-    # convert timestamps to frame indices
-    frame_indices = [round(ts * average_fps) for ts in timestamps]
-
-    # retrieve frames based on indices
-    frames_batch = decoder.get_frames_at(indices=frame_indices)
-
-    for frame, pts in zip(frames_batch.data, frames_batch.pts_seconds, strict=False):
-        loaded_frames.append(frame)
-        loaded_ts.append(pts.item())
-        if log_loaded_timestamps:
-            logging.info(f"Frame loaded at timestamp={pts:.4f}")
-
-    query_ts = torch.tensor(timestamps)
-    loaded_ts = torch.tensor(loaded_ts)
-
-    # compute distances between each query timestamp and loaded timestamps
-    dist = torch.cdist(query_ts[:, None], loaded_ts[:, None], p=1)
-    min_, argmin_ = dist.min(1)
-
-    is_within_tol = min_ < tolerance_s
-    assert is_within_tol.all(), (
-        f"One or several query timestamps unexpectedly violate the tolerance ({min_[~is_within_tol]} > {tolerance_s=})."
-        "It means that the closest frame that can be loaded from the video is too far away in time."
-        "This might be due to synchronization issues with timestamps during data collection."
-        "To be safe, we advise to ignore this item during training."
-        f"\nqueried timestamps: {query_ts}"
-        f"\nloaded timestamps: {loaded_ts}"
-        f"\nvideo: {video_path}"
-    )
-
-    # get closest frames to the query timestamps
-    closest_frames = torch.stack([loaded_frames[idx] for idx in argmin_])
-    closest_ts = loaded_ts[argmin_]
-
-    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
-
-    assert len(timestamps) == len(closest_frames)
-    return closest_frames
-
-
 def encode_video_frames(
     imgs_dir: Path | str,
     video_path: Path | str,
@@ -253,19 +136,18 @@ def encode_video_frames(
     g: int | None = 2,
     crf: int | None = 30,
     fast_decode: int = 0,
-    log_level: str | None = "quiet",
+    log_level: str | None = "error",
     overwrite: bool = False,
 ) -> None:
     """More info on ffmpeg arguments tuning on `benchmark/video/README.md`"""
     video_path = Path(video_path)
-    imgs_dir = Path(imgs_dir)
     video_path.parent.mkdir(parents=True, exist_ok=True)
 
     ffmpeg_args = OrderedDict(
         [
             ("-f", "image2"),
             ("-r", str(fps)),
-            ("-i", str(imgs_dir / "frame-%06d.png")),
+            ("-i", str(imgs_dir / "frame_%06d.png")),
             ("-vcodec", vcodec),
             ("-pix_fmt", pix_fmt),
         ]

lerobot/common/envs/configs.py

@@ -17,7 +17,7 @@ from dataclasses import dataclass, field
 
 import draccus
 
-from lerobot.common.constants import ACTION, OBS_ENV, OBS_IMAGE, OBS_IMAGES, OBS_ROBOT
+from lerobot.common.constants import ACTION, OBS_ENV_STATE, OBS_IMAGE, OBS_IMAGES, OBS_STATE
 from lerobot.configs.types import FeatureType, PolicyFeature
 
 
@@ -53,7 +53,7 @@ class AlohaEnv(EnvConfig):
     features_map: dict[str, str] = field(
         default_factory=lambda: {
             "action": ACTION,
-            "agent_pos": OBS_ROBOT,
+            "agent_pos": OBS_STATE,
             "top": f"{OBS_IMAGE}.top",
             "pixels/top": f"{OBS_IMAGES}.top",
         }
@@ -94,8 +94,8 @@ class PushtEnv(EnvConfig):
     features_map: dict[str, str] = field(
         default_factory=lambda: {
             "action": ACTION,
-            "agent_pos": OBS_ROBOT,
-            "environment_state": OBS_ENV,
+            "agent_pos": OBS_STATE,
+            "environment_state": OBS_ENV_STATE,
             "pixels": OBS_IMAGE,
         }
     )
@@ -136,7 +136,7 @@ class XarmEnv(EnvConfig):
     features_map: dict[str, str] = field(
         default_factory=lambda: {
             "action": ACTION,
-            "agent_pos": OBS_ROBOT,
+            "agent_pos": OBS_STATE,
             "pixels": OBS_IMAGE,
         }
     )

lerobot/common/envs/utils.py

@@ -13,11 +13,7 @@
 # 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 warnings
-from typing import Any
-
 import einops
-import gymnasium as gym
 import numpy as np
 import torch
 from torch import Tensor
@@ -90,38 +86,3 @@ def env_to_policy_features(env_cfg: EnvConfig) -> dict[str, PolicyFeature]:
         policy_features[policy_key] = feature
 
     return policy_features
-
-
-def are_all_envs_same_type(env: gym.vector.VectorEnv) -> bool:
-    first_type = type(env.envs[0])  # Get type of first env
-    return all(type(e) is first_type for e in env.envs)  # Fast type check
-
-
-def check_env_attributes_and_types(env: gym.vector.VectorEnv) -> None:
-    with warnings.catch_warnings():
-        warnings.simplefilter("once", UserWarning)  # Apply filter only in this function
-
-        if not (hasattr(env.envs[0], "task_description") and hasattr(env.envs[0], "task")):
-            warnings.warn(
-                "The environment does not have 'task_description' and 'task'. Some policies require these features.",
-                UserWarning,
-                stacklevel=2,
-            )
-        if not are_all_envs_same_type(env):
-            warnings.warn(
-                "The environments have different types. Make sure you infer the right task from each environment. Empty task will be passed instead.",
-                UserWarning,
-                stacklevel=2,
-            )
-
-
-def add_envs_task(env: gym.vector.VectorEnv, observation: dict[str, Any]) -> dict[str, Any]:
-    """Adds task feature to the observation dict with respect to the first environment attribute."""
-    if hasattr(env.envs[0], "task_description"):
-        observation["task"] = env.call("task_description")
-    elif hasattr(env.envs[0], "task"):
-        observation["task"] = env.call("task")
-    else:  #  For envs without language instructions, e.g. aloha transfer cube and etc.
-        num_envs = observation[list(observation.keys())[0]].shape[0]
-        observation["task"] = ["" for _ in range(num_envs)]
-    return observation

lerobot/common/errors.py

@@ -0,0 +1,17 @@
+class DeviceNotConnectedError(ConnectionError):
+    """Exception raised when the device is not connected."""
+
+    def __init__(self, message="This device is not connected. Try calling `connect()` first."):
+        self.message = message
+        super().__init__(self.message)
+
+
+class DeviceAlreadyConnectedError(ConnectionError):
+    """Exception raised when the device is already connected."""
+
+    def __init__(
+        self,
+        message="This device is already connected. Try not calling `connect()` twice.",
+    ):
+        self.message = message
+        super().__init__(self.message)

lerobot/common/motors/__init__.py

@@ -0,0 +1,3 @@
+from .motors_bus import MotorsBus
+
+__all__ = ["MotorsBus"]

lerobot/common/motors/dynamixel/__init__.py

@@ -0,0 +1,4 @@
+from .dynamixel import DynamixelMotorsBus, TorqueMode, set_operating_mode
+from .dynamixel_calibration import run_arm_calibration
+
+__all__ = ["DynamixelMotorsBus", "TorqueMode", "set_operating_mode", "run_arm_calibration"]

lerobot/common/motors/dynamixel/dynamixel.py

@@ -22,8 +22,7 @@ from copy import deepcopy
 import numpy as np
 import tqdm
 
-from lerobot.common.robot_devices.motors.configs import DynamixelMotorsBusConfig
-from lerobot.common.robot_devices.utils import RobotDeviceAlreadyConnectedError, RobotDeviceNotConnectedError
+from lerobot.common.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
 from lerobot.common.utils.utils import capture_timestamp_utc
 
 PROTOCOL_VERSION = 2.0
@@ -288,11 +287,10 @@ class DynamixelMotorsBus:
     motor_index = 6
     motor_model = "xl330-m288"
 
-    config = DynamixelMotorsBusConfig(
+    motors_bus = DynamixelMotorsBus(
         port="/dev/tty.usbmodem575E0031751",
         motors={motor_name: (motor_index, motor_model)},
     )
-    motors_bus = DynamixelMotorsBus(config)
     motors_bus.connect()
 
     position = motors_bus.read("Present_Position")
@@ -308,11 +306,13 @@ class DynamixelMotorsBus:
 
     def __init__(
         self,
-        config: DynamixelMotorsBusConfig,
+        port: str,
+        motors: dict[str, tuple[int, str]],
+        mock: bool = False,
     ):
-        self.port = config.port
-        self.motors = config.motors
-        self.mock = config.mock
+        self.port = port
+        self.motors = motors
+        self.mock = mock
 
         self.model_ctrl_table = deepcopy(MODEL_CONTROL_TABLE)
         self.model_resolution = deepcopy(MODEL_RESOLUTION)
@@ -327,7 +327,7 @@ class DynamixelMotorsBus:
 
     def connect(self):
         if self.is_connected:
-            raise RobotDeviceAlreadyConnectedError(
+            raise DeviceAlreadyConnectedError(
                 f"DynamixelMotorsBus({self.port}) is already connected. Do not call `motors_bus.connect()` twice."
             )
 
@@ -684,7 +684,7 @@ class DynamixelMotorsBus:
 
     def read(self, data_name, motor_names: str | list[str] | None = None):
         if not self.is_connected:
-            raise RobotDeviceNotConnectedError(
+            raise DeviceNotConnectedError(
                 f"DynamixelMotorsBus({self.port}) is not connected. You need to run `motors_bus.connect()`."
             )
 
@@ -786,7 +786,7 @@ class DynamixelMotorsBus:
 
     def write(self, data_name, values: int | float | np.ndarray, motor_names: str | list[str] | None = None):
         if not self.is_connected:
-            raise RobotDeviceNotConnectedError(
+            raise DeviceNotConnectedError(
                 f"DynamixelMotorsBus({self.port}) is not connected. You need to run `motors_bus.connect()`."
             )
 
@@ -855,7 +855,7 @@ class DynamixelMotorsBus:
 
     def disconnect(self):
         if not self.is_connected:
-            raise RobotDeviceNotConnectedError(
+            raise DeviceNotConnectedError(
                 f"DynamixelMotorsBus({self.port}) is not connected. Try running `motors_bus.connect()` first."
             )
 
@@ -871,3 +871,25 @@ class DynamixelMotorsBus:
     def __del__(self):
         if getattr(self, "is_connected", False):
             self.disconnect()
+
+
+def set_operating_mode(arm: DynamixelMotorsBus):
+    if (arm.read("Torque_Enable") != TorqueMode.DISABLED.value).any():
+        raise ValueError("To run set robot preset, the torque must be disabled on all motors.")
+
+    # Use 'extended position mode' for all motors except gripper, because in joint mode the servos can't
+    # rotate more than 360 degrees (from 0 to 4095) And some mistake can happen while assembling the arm,
+    # you could end up with a servo with a position 0 or 4095 at a crucial point See [
+    # https://emanual.robotis.com/docs/en/dxl/x/x_series/#operating-mode11]
+    all_motors_except_gripper = [name for name in arm.motor_names if name != "gripper"]
+    if len(all_motors_except_gripper) > 0:
+        # 4 corresponds to Extended Position on Koch motors
+        arm.write("Operating_Mode", 4, all_motors_except_gripper)
+
+    # Use 'position control current based' for gripper to be limited by the limit of the current.
+    # For the follower gripper, it means it can grasp an object without forcing too much even tho,
+    # it's goal position is a complete grasp (both gripper fingers are ordered to join and reach a touch).
+    # For the leader gripper, it means we can use it as a physical trigger, since we can force with our finger
+    # to make it move, and it will move back to its original target position when we release the force.
+    # 5 corresponds to Current Controlled Position on Koch gripper motors "xl330-m077, xl330-m288"
+    arm.write("Operating_Mode", 5, "gripper")

lerobot/common/motors/dynamixel/dynamixel_calibration.py

@@ -17,12 +17,12 @@
 
 import numpy as np
 
-from lerobot.common.robot_devices.motors.dynamixel import (
+from ..motors_bus import MotorsBus
+from .dynamixel import (
     CalibrationMode,
     TorqueMode,
     convert_degrees_to_steps,
 )
-from lerobot.common.robot_devices.motors.utils import MotorsBus
 
 URL_TEMPLATE = (
     "https://raw.githubusercontent.com/huggingface/lerobot/main/media/{robot}/{arm}_{position}.webp"

lerobot/common/motors/feetech/__init__.py

@@ -0,0 +1,9 @@
+from .feetech import FeetechMotorsBus, TorqueMode
+from .feetech_calibration import apply_feetech_offsets_from_calibration, run_full_arm_calibration
+
+__all__ = [
+    "FeetechMotorsBus",
+    "TorqueMode",
+    "apply_feetech_offsets_from_calibration",
+    "run_full_arm_calibration",
+]

lerobot/common/motors/feetech/feetech.py

@@ -13,8 +13,6 @@
 # limitations under the License.
 
 import enum
-import logging
-import math
 import time
 import traceback
 from copy import deepcopy
@@ -22,8 +20,7 @@ from copy import deepcopy
 import numpy as np
 import tqdm
 
-from lerobot.common.robot_devices.motors.configs import FeetechMotorsBusConfig
-from lerobot.common.robot_devices.utils import RobotDeviceAlreadyConnectedError, RobotDeviceNotConnectedError
+from lerobot.common.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
 from lerobot.common.utils.utils import capture_timestamp_utc
 
 PROTOCOL_VERSION = 0
@@ -32,13 +29,6 @@ TIMEOUT_MS = 1000
 
 MAX_ID_RANGE = 252
 
-# The following bounds define the lower and upper joints range (after calibration).
-# For joints in degree (i.e. revolute joints), their nominal range is [-180, 180] degrees
-# which corresponds to a half rotation on the left and half rotation on the right.
-# Some joints might require higher range, so we allow up to [-270, 270] degrees until
-# an error is raised.
-LOWER_BOUND_DEGREE = -270
-UPPER_BOUND_DEGREE = 270
 # For joints in percentage (i.e. joints that move linearly like the prismatic joint of a gripper),
 # their nominal range is [0, 100] %. For instance, for Aloha gripper, 0% is fully
 # closed, and 100% is fully open. To account for slight calibration issue, we allow up to
@@ -48,7 +38,6 @@ UPPER_BOUND_LINEAR = 110
 
 HALF_TURN_DEGREE = 180
 
-
 # See this link for STS3215 Memory Table:
 # https://docs.google.com/spreadsheets/d/1GVs7W1VS1PqdhA1nW-abeyAHhTUxKUdR/edit?usp=sharing&ouid=116566590112741600240&rtpof=true&sd=true
 # data_name: (address, size_byte)
@@ -114,8 +103,6 @@ SCS_SERIES_BAUDRATE_TABLE = {
 }
 
 CALIBRATION_REQUIRED = ["Goal_Position", "Present_Position"]
-CONVERT_UINT32_TO_INT32_REQUIRED = ["Goal_Position", "Present_Position"]
-
 
 MODEL_CONTROL_TABLE = {
     "scs_series": SCS_SERIES_CONTROL_TABLE,
@@ -137,15 +124,63 @@ NUM_READ_RETRY = 20
 NUM_WRITE_RETRY = 20
 
 
-def convert_degrees_to_steps(degrees: float | np.ndarray, models: str | list[str]) -> np.ndarray:
-    """This function converts the degree range to the step range for indicating motors rotation.
-    It assumes a motor achieves a full rotation by going from -180 degree position to +180.
-    The motor resolution (e.g. 4096) corresponds to the number of steps needed to achieve a full rotation.
+def convert_ticks_to_degrees(ticks, model):
+    resolutions = MODEL_RESOLUTION[model]
+    # Convert the ticks to degrees
+    return ticks * (360.0 / resolutions)
+
+
+def convert_degrees_to_ticks(degrees, model):
+    resolutions = MODEL_RESOLUTION[model]
+    # Convert degrees to motor ticks
+    return int(degrees * (resolutions / 360.0))
+
+
+def adjusted_to_homing_ticks(raw_motor_ticks: int, model: str, motorbus, motor_id: int) -> int:
     """
-    resolutions = [MODEL_RESOLUTION[model] for model in models]
-    steps = degrees / 180 * np.array(resolutions) / 2
-    steps = steps.astype(int)
-    return steps
+    Takes a raw reading [0..(res-1)] (e.g. 0..4095) and shifts it so that '2048'
+    becomes 0 in the homed coordinate system ([-2048..+2047] for 4096 resolution).
+    """
+    resolutions = MODEL_RESOLUTION[model]
+
+    # Shift raw ticks by half-resolution so 2048 -> 0, then wrap [0..res-1].
+    ticks = (raw_motor_ticks - (resolutions // 2)) % resolutions
+
+    # If above halfway, fold it into negative territory => [-2048..+2047].
+    if ticks > (resolutions // 2):
+        ticks -= resolutions
+
+    # Flip sign if drive_mode is set.
+    drive_mode = 0
+    if motorbus.calibration is not None:
+        drive_mode = motorbus.calibration["drive_mode"][motor_id - 1]
+
+    if drive_mode:
+        ticks *= -1
+
+    return ticks
+
+
+def adjusted_to_motor_ticks(adjusted_pos: int, model: str, motorbus, motor_id: int) -> int:
+    """
+    Inverse of adjusted_to_homing_ticks(). Takes a 'homed' position in [-2048..+2047]
+    and recovers the raw [0..(res-1)] ticks with 2048 as midpoint.
+    """
+    # Flip sign if drive_mode was set.
+    drive_mode = 0
+    if motorbus.calibration is not None:
+        drive_mode = motorbus.calibration["drive_mode"][motor_id - 1]
+
+    if drive_mode:
+        adjusted_pos *= -1
+
+    resolutions = MODEL_RESOLUTION[model]
+
+    # Shift by +half-resolution and wrap into [0..res-1].
+    # This undoes the earlier shift by -half-resolution.
+    ticks = (adjusted_pos + (resolutions // 2)) % resolutions
+
+    return ticks
 
 
 def convert_to_bytes(value, bytes, mock=False):
@@ -267,11 +302,10 @@ class FeetechMotorsBus:
     motor_index = 6
     motor_model = "sts3215"
 
-    config = FeetechMotorsBusConfig(
+    motors_bus = FeetechMotorsBus(
         port="/dev/tty.usbmodem575E0031751",
         motors={motor_name: (motor_index, motor_model)},
     )
-    motors_bus = FeetechMotorsBus(config)
     motors_bus.connect()
 
     position = motors_bus.read("Present_Position")
@@ -287,11 +321,13 @@ class FeetechMotorsBus:
 
     def __init__(
         self,
-        config: FeetechMotorsBusConfig,
+        port: str,
+        motors: dict[str, tuple[int, str]],
+        mock: bool = False,
     ):
-        self.port = config.port
-        self.motors = config.motors
-        self.mock = config.mock
+        self.port = port
+        self.motors = motors
+        self.mock = mock
 
         self.model_ctrl_table = deepcopy(MODEL_CONTROL_TABLE)
         self.model_resolution = deepcopy(MODEL_RESOLUTION)
@@ -304,11 +340,9 @@ class FeetechMotorsBus:
         self.group_writers = {}
         self.logs = {}
 
-        self.track_positions = {}
-
     def connect(self):
         if self.is_connected:
-            raise RobotDeviceAlreadyConnectedError(
+            raise DeviceAlreadyConnectedError(
                 f"FeetechMotorsBus({self.port}) is already connected. Do not call `motors_bus.connect()` twice."
             )
 
@@ -402,33 +436,7 @@ class FeetechMotorsBus:
     def set_calibration(self, calibration: dict[str, list]):
         self.calibration = calibration
 
-    def apply_calibration_autocorrect(self, values: np.ndarray | list, motor_names: list[str] | None):
-        """This function apply the calibration, automatically detects out of range errors for motors values and attempt to correct.
-
-        For more info, see docstring of `apply_calibration` and `autocorrect_calibration`.
-        """
-        try:
-            values = self.apply_calibration(values, motor_names)
-        except JointOutOfRangeError as e:
-            print(e)
-            self.autocorrect_calibration(values, motor_names)
-            values = self.apply_calibration(values, motor_names)
-        return values
-
     def apply_calibration(self, values: np.ndarray | list, motor_names: list[str] | None):
-        """Convert from unsigned int32 joint position range [0, 2**32[ to the universal float32 nominal degree range ]-180.0, 180.0[ with
-        a "zero position" at 0 degree.
-
-        Note: We say "nominal degree range" since the motors can take values outside this range. For instance, 190 degrees, if the motor
-        rotate more than a half a turn from the zero position. However, most motors can't rotate more than 180 degrees and will stay in this range.
-
-        Joints values are original in [0, 2**32[ (unsigned int32). Each motor are expected to complete a full rotation
-        when given a goal position that is + or - their resolution. For instance, feetech xl330-m077 have a resolution of 4096, and
-        at any position in their original range, let's say the position 56734, they complete a full rotation clockwise by moving to 60830,
-        or anticlockwise by moving to 52638. The position in the original range is arbitrary and might change a lot between each motor.
-        To harmonize between motors of the same model, different robots, or even models of different brands, we propose to work
-        in the centered nominal degree range ]-180, 180[.
-        """
         if motor_names is None:
             motor_names = self.motor_names
 
@@ -440,34 +448,11 @@ class FeetechMotorsBus:
             calib_mode = self.calibration["calib_mode"][calib_idx]
 
             if CalibrationMode[calib_mode] == CalibrationMode.DEGREE:
-                drive_mode = self.calibration["drive_mode"][calib_idx]
-                homing_offset = self.calibration["homing_offset"][calib_idx]
-                _, model = self.motors[name]
-                resolution = self.model_resolution[model]
+                motor_idx, model = self.motors[name]
 
-                # Update direction of rotation of the motor to match between leader and follower.
-                # In fact, the motor of the leader for a given joint can be assembled in an
-                # opposite direction in term of rotation than the motor of the follower on the same joint.
-                if drive_mode:
-                    values[i] *= -1
-
-                # Convert from range [-2**31, 2**31[ to
-                # nominal range ]-resolution, resolution[ (e.g. ]-2048, 2048[)
-                values[i] += homing_offset
-
-                # Convert from range ]-resolution, resolution[ to
-                # universal float32 centered degree range ]-180, 180[
-                values[i] = values[i] / (resolution // 2) * HALF_TURN_DEGREE
-
-                if (values[i] < LOWER_BOUND_DEGREE) or (values[i] > UPPER_BOUND_DEGREE):
-                    raise JointOutOfRangeError(
-                        f"Wrong motor position range detected for {name}. "
-                        f"Expected to be in nominal range of [-{HALF_TURN_DEGREE}, {HALF_TURN_DEGREE}] degrees (a full rotation), "
-                        f"with a maximum range of [{LOWER_BOUND_DEGREE}, {UPPER_BOUND_DEGREE}] degrees to account for joints that can rotate a bit more, "
-                        f"but present value is {values[i]} degree. "
-                        "This might be due to a cable connection issue creating an artificial 360 degrees jump in motor values. "
-                        "You need to recalibrate by running: `python lerobot/scripts/control_robot.py calibrate`"
-                    )
+                # Convert raw motor ticks to homed ticks, then convert the homed ticks to degrees
+                values[i] = adjusted_to_homing_ticks(values[i], model, self, motor_idx)
+                values[i] = convert_ticks_to_degrees(values[i], model)
 
             elif CalibrationMode[calib_mode] == CalibrationMode.LINEAR:
                 start_pos = self.calibration["start_pos"][calib_idx]
@@ -489,103 +474,6 @@ class FeetechMotorsBus:
 
         return values
 
-    def autocorrect_calibration(self, values: np.ndarray | list, motor_names: list[str] | None):
-        """This function automatically detects issues with values of motors after calibration, and correct for these issues.
-
-        Some motors might have values outside of expected maximum bounds after calibration.
-        For instance, for a joint in degree, its value can be outside [-270, 270] degrees, which is totally unexpected given
-        a nominal range of [-180, 180] degrees, which represents half a turn to the left or right starting from zero position.
-
-        Known issues:
-        #1: Motor value randomly shifts of a full turn, caused by hardware/connection errors.
-        #2: Motor internal homing offset is shifted of a full turn, caused by using default calibration (e.g Aloha).
-        #3: motor internal homing offset is shifted of less or more than a full turn, caused by using default calibration
-            or by human error during manual calibration.
-
-        Issues #1 and #2 can be solved by shifting the calibration homing offset by a full turn.
-        Issue #3 will be visually detected by user and potentially captured by the safety feature `max_relative_target`,
-        that will slow down the motor, raise an error asking to recalibrate. Manual recalibrating will solve the issue.
-
-        Note: A full turn corresponds to 360 degrees but also to 4096 steps for a motor resolution of 4096.
-        """
-        if motor_names is None:
-            motor_names = self.motor_names
-
-        # Convert from unsigned int32 original range [0, 2**32] to signed float32 range
-        values = values.astype(np.float32)
-
-        for i, name in enumerate(motor_names):
-            calib_idx = self.calibration["motor_names"].index(name)
-            calib_mode = self.calibration["calib_mode"][calib_idx]
-
-            if CalibrationMode[calib_mode] == CalibrationMode.DEGREE:
-                drive_mode = self.calibration["drive_mode"][calib_idx]
-                homing_offset = self.calibration["homing_offset"][calib_idx]
-                _, model = self.motors[name]
-                resolution = self.model_resolution[model]
-
-                if drive_mode:
-                    values[i] *= -1
-
-                # Convert from initial range to range [-180, 180] degrees
-                calib_val = (values[i] + homing_offset) / (resolution // 2) * HALF_TURN_DEGREE
-                in_range = (calib_val > LOWER_BOUND_DEGREE) and (calib_val < UPPER_BOUND_DEGREE)
-
-                # Solve this inequality to find the factor to shift the range into [-180, 180] degrees
-                # values[i] = (values[i] + homing_offset + resolution * factor) / (resolution // 2) * HALF_TURN_DEGREE
-                # - HALF_TURN_DEGREE <= (values[i] + homing_offset + resolution * factor) / (resolution // 2) * HALF_TURN_DEGREE <= HALF_TURN_DEGREE
-                # (- HALF_TURN_DEGREE / HALF_TURN_DEGREE * (resolution // 2) - values[i] - homing_offset) / resolution <= factor <= (HALF_TURN_DEGREE / 180 * (resolution // 2) - values[i] - homing_offset) / resolution
-                low_factor = (
-                    -HALF_TURN_DEGREE / HALF_TURN_DEGREE * (resolution // 2) - values[i] - homing_offset
-                ) / resolution
-                upp_factor = (
-                    HALF_TURN_DEGREE / HALF_TURN_DEGREE * (resolution // 2) - values[i] - homing_offset
-                ) / resolution
-
-            elif CalibrationMode[calib_mode] == CalibrationMode.LINEAR:
-                start_pos = self.calibration["start_pos"][calib_idx]
-                end_pos = self.calibration["end_pos"][calib_idx]
-
-                # Convert from initial range to range [0, 100] in %
-                calib_val = (values[i] - start_pos) / (end_pos - start_pos) * 100
-                in_range = (calib_val > LOWER_BOUND_LINEAR) and (calib_val < UPPER_BOUND_LINEAR)
-
-                # Solve this inequality to find the factor to shift the range into [0, 100] %
-                # values[i] = (values[i] - start_pos + resolution * factor) / (end_pos + resolution * factor - start_pos - resolution * factor) * 100
-                # values[i] = (values[i] - start_pos + resolution * factor) / (end_pos - start_pos) * 100
-                # 0 <= (values[i] - start_pos + resolution * factor) / (end_pos - start_pos) * 100 <= 100
-                # (start_pos - values[i]) / resolution <= factor <= (end_pos - values[i]) / resolution
-                low_factor = (start_pos - values[i]) / resolution
-                upp_factor = (end_pos - values[i]) / resolution
-
-            if not in_range:
-                # Get first integer between the two bounds
-                if low_factor < upp_factor:
-                    factor = math.ceil(low_factor)
-
-                    if factor > upp_factor:
-                        raise ValueError(f"No integer found between bounds [{low_factor=}, {upp_factor=}]")
-                else:
-                    factor = math.ceil(upp_factor)
-
-                    if factor > low_factor:
-                        raise ValueError(f"No integer found between bounds [{low_factor=}, {upp_factor=}]")
-
-                if CalibrationMode[calib_mode] == CalibrationMode.DEGREE:
-                    out_of_range_str = f"{LOWER_BOUND_DEGREE} < {calib_val} < {UPPER_BOUND_DEGREE} degrees"
-                    in_range_str = f"{LOWER_BOUND_DEGREE} < {calib_val} < {UPPER_BOUND_DEGREE} degrees"
-                elif CalibrationMode[calib_mode] == CalibrationMode.LINEAR:
-                    out_of_range_str = f"{LOWER_BOUND_LINEAR} < {calib_val} < {UPPER_BOUND_LINEAR} %"
-                    in_range_str = f"{LOWER_BOUND_LINEAR} < {calib_val} < {UPPER_BOUND_LINEAR} %"
-
-                logging.warning(
-                    f"Auto-correct calibration of motor '{name}' by shifting value by {abs(factor)} full turns, "
-                    f"from '{out_of_range_str}' to '{in_range_str}'."
-                )
-
-                # A full turn corresponds to 360 degrees but also to 4096 steps for a motor resolution of 4096.
-                self.calibration["homing_offset"][calib_idx] += resolution * factor
-
     def revert_calibration(self, values: np.ndarray | list, motor_names: list[str] | None):
         """Inverse of `apply_calibration`."""
         if motor_names is None:
@@ -596,23 +484,11 @@ class FeetechMotorsBus:
             calib_mode = self.calibration["calib_mode"][calib_idx]
 
             if CalibrationMode[calib_mode] == CalibrationMode.DEGREE:
-                drive_mode = self.calibration["drive_mode"][calib_idx]
-                homing_offset = self.calibration["homing_offset"][calib_idx]
-                _, model = self.motors[name]
-                resolution = self.model_resolution[model]
+                motor_idx, model = self.motors[name]
 
-                # Convert from nominal 0-centered degree range [-180, 180] to
-                # 0-centered resolution range (e.g. [-2048, 2048] for resolution=4096)
-                values[i] = values[i] / HALF_TURN_DEGREE * (resolution // 2)
-
-                # Subtract the homing offsets to come back to actual motor range of values
-                # which can be arbitrary.
-                values[i] -= homing_offset
-
-                # Remove drive mode, which is the rotation direction of the motor, to come back to
-                # actual motor rotation direction which can be arbitrary.
-                if drive_mode:
-                    values[i] *= -1
+                # Convert degrees to homed ticks, then convert the homed ticks to raw ticks
+                values[i] = convert_degrees_to_ticks(values[i], model)
+                values[i] = adjusted_to_motor_ticks(values[i], model, self, motor_idx)
 
             elif CalibrationMode[calib_mode] == CalibrationMode.LINEAR:
                 start_pos = self.calibration["start_pos"][calib_idx]
@@ -625,43 +501,6 @@ class FeetechMotorsBus:
         values = np.round(values).astype(np.int32)
         return values
 
-    def avoid_rotation_reset(self, values, motor_names, data_name):
-        if data_name not in self.track_positions:
-            self.track_positions[data_name] = {
-                "prev": [None] * len(self.motor_names),
-                # Assume False at initialization
-                "below_zero": [False] * len(self.motor_names),
-                "above_max": [False] * len(self.motor_names),
-            }
-
-        track = self.track_positions[data_name]
-
-        if motor_names is None:
-            motor_names = self.motor_names
-
-        for i, name in enumerate(motor_names):
-            idx = self.motor_names.index(name)
-
-            if track["prev"][idx] is None:
-                track["prev"][idx] = values[i]
-                continue
-
-            # Detect a full rotation occurred
-            if abs(track["prev"][idx] - values[i]) > 2048:
-                # Position went below 0 and got reset to 4095
-                if track["prev"][idx] < values[i]:
-                    # So we set negative value by adding a full rotation
-                    values[i] -= 4096
-
-                # Position went above 4095 and got reset to 0
-                elif track["prev"][idx] > values[i]:
-                    # So we add a full rotation
-                    values[i] += 4096
-
-            track["prev"][idx] = values[i]
-
-        return values
-
     def read_with_motor_ids(self, motor_models, motor_ids, data_name, num_retry=NUM_READ_RETRY):
         if self.mock:
             import tests.motors.mock_scservo_sdk as scs
@@ -707,7 +546,7 @@ class FeetechMotorsBus:
             import scservo_sdk as scs
 
         if not self.is_connected:
-            raise RobotDeviceNotConnectedError(
+            raise DeviceNotConnectedError(
                 f"FeetechMotorsBus({self.port}) is not connected. You need to run `motors_bus.connect()`."
             )
 
@@ -735,7 +574,7 @@ class FeetechMotorsBus:
             self.port_handler.ser.reset_output_buffer()
             self.port_handler.ser.reset_input_buffer()
 
-            # create new group reader
+            # Create new group reader
             self.group_readers[group_key] = scs.GroupSyncRead(
                 self.port_handler, self.packet_handler, addr, bytes
             )
@@ -760,15 +599,8 @@ class FeetechMotorsBus:
 
         values = np.array(values)
 
-        # Convert to signed int to use range [-2048, 2048] for our motor positions.
-        if data_name in CONVERT_UINT32_TO_INT32_REQUIRED:
-            values = values.astype(np.int32)
-
-        if data_name in CALIBRATION_REQUIRED:
-            values = self.avoid_rotation_reset(values, motor_names, data_name)
-
         if data_name in CALIBRATION_REQUIRED and self.calibration is not None:
-            values = self.apply_calibration_autocorrect(values, motor_names)
+            values = self.apply_calibration(values, motor_names)
 
         # log the number of seconds it took to read the data from the motors
         delta_ts_name = get_log_name("delta_timestamp_s", "read", data_name, motor_names)
@@ -811,7 +643,7 @@ class FeetechMotorsBus:
 
     def write(self, data_name, values: int | float | np.ndarray, motor_names: str | list[str] | None = None):
         if not self.is_connected:
-            raise RobotDeviceNotConnectedError(
+            raise DeviceNotConnectedError(
                 f"FeetechMotorsBus({self.port}) is not connected. You need to run `motors_bus.connect()`."
             )
 
@@ -880,7 +712,7 @@ class FeetechMotorsBus:
 
     def disconnect(self):
         if not self.is_connected:
-            raise RobotDeviceNotConnectedError(
+            raise DeviceNotConnectedError(
                 f"FeetechMotorsBus({self.port}) is not connected. Try running `motors_bus.connect()` first."
             )
 

lerobot/common/motors/feetech/feetech_calibration.py

@@ -0,0 +1,254 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import numpy as np
+
+from ..motors_bus import MotorsBus
+from .feetech import (
+    CalibrationMode,
+    FeetechMotorsBus,
+    TorqueMode,
+)
+
+URL_TEMPLATE = (
+    "https://raw.githubusercontent.com/huggingface/lerobot/main/media/{robot}/{arm}_{position}.webp"
+)
+
+
+def disable_torque(arm: MotorsBus):
+    if (arm.read("Torque_Enable") != TorqueMode.DISABLED.value).any():
+        raise ValueError("To run calibration, the torque must be disabled on all motors.")
+
+
+def get_calibration_modes(arm: MotorsBus):
+    """Returns calibration modes for each motor (DEGREE for rotational, LINEAR for gripper)."""
+    return [
+        CalibrationMode.LINEAR.name if name == "gripper" else CalibrationMode.DEGREE.name
+        for name in arm.motor_names
+    ]
+
+
+def reset_offset(motor_id, motor_bus):
+    # Open the write lock, changes to EEPROM do NOT persist yet
+    motor_bus.write("Lock", 1)
+
+    # Set offset to 0
+    motor_name = motor_bus.motor_names[motor_id - 1]
+    motor_bus.write("Offset", 0, motor_names=[motor_name])
+
+    # Close the write lock, changes to EEPROM do persist
+    motor_bus.write("Lock", 0)
+
+    # Confirm that the offset is zero by reading it back
+    confirmed_offset = motor_bus.read("Offset")[motor_id - 1]
+    print(f"Offset for motor {motor_id} reset to: {confirmed_offset}")
+    return confirmed_offset
+
+
+def calibrate_homing_motor(motor_id, motor_bus):
+    reset_offset(motor_id, motor_bus)
+
+    home_ticks = motor_bus.read("Present_Position")[motor_id - 1]  # Read index starts at 0
+    print(f"Encoder offset (present position in homing position): {home_ticks}")
+
+    return home_ticks
+
+
+def calibrate_linear_motor(motor_id, motor_bus):
+    motor_names = motor_bus.motor_names
+    motor_name = motor_names[motor_id - 1]
+
+    reset_offset(motor_id, motor_bus)
+
+    input(f"Close the {motor_name}, then press Enter...")
+    start_pos = motor_bus.read("Present_Position")[motor_id - 1]  # Read index starts ar 0
+    print(f"  [Motor {motor_id}] start position recorded: {start_pos}")
+
+    input(f"Open the {motor_name} fully, then press Enter...")
+    end_pos = motor_bus.read("Present_Position")[motor_id - 1]  # Read index starts ar 0
+    print(f"  [Motor {motor_id}] end position recorded: {end_pos}")
+
+    return start_pos, end_pos
+
+
+def single_motor_calibration(arm: MotorsBus, motor_id: int):
+    """Calibrates a single motor and returns its calibration data for updating the calibration file."""
+
+    disable_torque(arm)
+    print(f"\n--- Calibrating Motor {motor_id} ---")
+
+    start_pos = 0
+    end_pos = 0
+    encoder_offset = 0
+
+    if motor_id == 6:
+        start_pos, end_pos = calibrate_linear_motor(motor_id, arm)
+    else:
+        input("Move the motor to (zero) position, then press Enter...")
+        encoder_offset = calibrate_homing_motor(motor_id, arm)
+
+    print(f"Calibration for motor ID:{motor_id} done.")
+
+    # Create a calibration dictionary for the single motor
+    calib_dict = {
+        "homing_offset": int(encoder_offset),
+        "drive_mode": 0,
+        "start_pos": int(start_pos),
+        "end_pos": int(end_pos),
+        "calib_mode": get_calibration_modes(arm)[motor_id - 1],
+        "motor_name": arm.motor_names[motor_id - 1],
+    }
+
+    return calib_dict
+
+
+def run_full_arm_calibration(arm: MotorsBus, robot_type: str, arm_name: str, arm_type: str):
+    """
+    Runs a full calibration process for all motors in a robotic arm.
+
+    This function calibrates each motor in the arm, determining encoder offsets and
+    start/end positions for linear and rotational motors. The calibration data is then
+    stored in a dictionary for later use.
+
+    **Calibration Process:**
+    - The user is prompted to move the arm to its homing position before starting.
+    - Motors with rotational motion are calibrated using a homing method.
+    - Linear actuators (e.g., grippers) are calibrated separately.
+    - Encoder offsets, start positions, and end positions are recorded.
+
+    **Example Usage:**
+    ```python
+    run_full_arm_calibration(arm, "so100", "left", "follower")
+    ```
+    """
+    disable_torque(arm)
+
+    print(f"\nRunning calibration of {robot_type} {arm_name} {arm_type}...")
+
+    print("\nMove arm to homing position (middle)")
+    print(
+        "See: " + URL_TEMPLATE.format(robot=robot_type, arm=arm_type, position="zero")
+    )  # TODO(pepijn): replace with new instruction homing pos (all motors in middle) in tutorial
+    input("Press Enter to continue...")
+
+    start_positions = np.zeros(len(arm.motor_indices))
+    end_positions = np.zeros(len(arm.motor_indices))
+    encoder_offsets = np.zeros(len(arm.motor_indices))
+
+    modes = get_calibration_modes(arm)
+
+    for i, motor_id in enumerate(arm.motor_indices):
+        if modes[i] == CalibrationMode.DEGREE.name:
+            encoder_offsets[i] = calibrate_homing_motor(motor_id, arm)
+            start_positions[i] = 0
+            end_positions[i] = 0
+
+    for i, motor_id in enumerate(arm.motor_indices):
+        if modes[i] == CalibrationMode.LINEAR.name:
+            start_positions[i], end_positions[i] = calibrate_linear_motor(motor_id, arm)
+            encoder_offsets[i] = 0
+
+    print("\nMove arm to rest position")
+    input("Press Enter to continue...")
+
+    print(f"\n calibration of {robot_type} {arm_name} {arm_type} done!")
+
+    # Force drive_mode values (can be static)
+    drive_modes = [0, 1, 0, 0, 1, 0]
+
+    calib_dict = {
+        "homing_offset": encoder_offsets.astype(int).tolist(),
+        "drive_mode": drive_modes,
+        "start_pos": start_positions.astype(int).tolist(),
+        "end_pos": end_positions.astype(int).tolist(),
+        "calib_mode": get_calibration_modes(arm),
+        "motor_names": arm.motor_names,
+    }
+    return calib_dict
+
+
+def run_full_auto_arm_calibration(arm: MotorsBus, robot_type: str, arm_name: str, arm_type: str):
+    """TODO(pepijn): Add this method later as extra
+    Example of usage:
+    ```python
+    run_full_auto_arm_calibration(arm, "so100", "left", "follower")
+    ```
+    """
+    print(f"\nRunning calibration of {robot_type} {arm_name} {arm_type}...")
+
+
+def apply_feetech_offsets_from_calibration(motorsbus: FeetechMotorsBus, calibration_dict: dict):
+    """
+    Reads 'calibration_dict' containing 'homing_offset' and 'motor_names',
+    then writes each motor's offset to the servo's internal Offset (0x1F) in EPROM.
+
+    This version is modified so each homed position (originally 0) will now read
+    2047, i.e. 180° away from 0 in the 4096-count circle. Offsets are permanently
+    stored in EEPROM, so the servo's Present_Position is hardware-shifted even
+    after power cycling.
+
+    Steps:
+      1) Subtract 2047 from the old offset (so 0 -> 2047).
+      2) Clamp to [-2047..+2047].
+      3) Encode sign bit and magnitude into a 12-bit number.
+    """
+
+    homing_offsets = calibration_dict["homing_offset"]
+    motor_names = calibration_dict["motor_names"]
+    start_pos = calibration_dict["start_pos"]
+
+    # Open the write lock, changes to EEPROM do NOT persist yet
+    motorsbus.write("Lock", 1)
+
+    # For each motor, set the 'Offset' parameter
+    for m_name, old_offset in zip(motor_names, homing_offsets, strict=False):
+        # If bus doesn’t have a motor named m_name, skip
+        if m_name not in motorsbus.motors:
+            print(f"Warning: '{m_name}' not found in motorsbus.motors; skipping offset.")
+            continue
+
+        if m_name == "gripper":
+            old_offset = start_pos  # If gripper set the offset to the start position of the gripper
+            continue
+
+        # Shift the offset so the homed position reads 2047
+        new_offset = old_offset - 2047
+
+        # Clamp to [-2047..+2047]
+        if new_offset > 2047:
+            new_offset = 2047
+            print(
+                f"Warning: '{new_offset}' is getting clamped because its larger then 2047; This should not happen!"
+            )
+        elif new_offset < -2047:
+            new_offset = -2047
+            print(
+                f"Warning: '{new_offset}' is getting clamped because its smaller then -2047; This should not happen!"
+            )
+
+        # Determine the direction (sign) bit and magnitude
+        direction_bit = 1 if new_offset < 0 else 0
+        magnitude = abs(new_offset)
+
+        # Combine sign bit (bit 11) with the magnitude (bits 0..10)
+        servo_offset = (direction_bit << 11) | magnitude
+
+        # Write offset to servo
+        motorsbus.write("Offset", servo_offset, motor_names=m_name)
+        print(
+            f"Set offset for {m_name}: "
+            f"old_offset={old_offset}, new_offset={new_offset}, servo_encoded={magnitude} + direction={direction_bit}"
+        )
+
+    motorsbus.write("Lock", 0)
+    print("Offsets have been saved to EEPROM successfully.")

lerobot/common/motors/motors_bus.py

@@ -0,0 +1,46 @@
+import abc
+
+
+class MotorsBus(abc.ABC):
+    """The main LeRobot class for implementing motors buses."""
+
+    def __init__(
+        self,
+        motors: dict[str, tuple[int, str]],
+    ):
+        self.motors = motors
+
+    def __len__(self):
+        return len(self.motors)
+
+    @abc.abstractmethod
+    def connect(self):
+        pass
+
+    @abc.abstractmethod
+    def reconnect(self):
+        pass
+
+    @abc.abstractmethod
+    def set_calibration(self, calibration: dict[str, list]):
+        pass
+
+    @abc.abstractmethod
+    def apply_calibration(self):
+        pass
+
+    @abc.abstractmethod
+    def revert_calibration(self):
+        pass
+
+    @abc.abstractmethod
+    def read(self):
+        pass
+
+    @abc.abstractmethod
+    def write(self):
+        pass
+
+    @abc.abstractmethod
+    def disconnect(self):
+        pass

lerobot/common/motors/utils.py

@@ -12,22 +12,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import Protocol
-
-from lerobot.common.robot_devices.motors.configs import (
-    DynamixelMotorsBusConfig,
-    FeetechMotorsBusConfig,
-    MotorsBusConfig,
-)
-
-
-class MotorsBus(Protocol):
-    def motor_names(self): ...
-    def set_calibration(self): ...
-    def apply_calibration(self): ...
-    def revert_calibration(self): ...
-    def read(self): ...
-    def write(self): ...
+from .configs import MotorsBusConfig
+from .motors_bus import MotorsBus
 
 
 def make_motors_buses_from_configs(motors_bus_configs: dict[str, MotorsBusConfig]) -> list[MotorsBus]:
@@ -35,12 +21,12 @@ def make_motors_buses_from_configs(motors_bus_configs: dict[str, MotorsBusConfig
 
     for key, cfg in motors_bus_configs.items():
         if cfg.type == "dynamixel":
-            from lerobot.common.robot_devices.motors.dynamixel import DynamixelMotorsBus
+            from .dynamixel import DynamixelMotorsBus
 
             motors_buses[key] = DynamixelMotorsBus(cfg)
 
         elif cfg.type == "feetech":
-            from lerobot.common.robot_devices.motors.feetech import FeetechMotorsBus
+            from lerobot.common.motors.feetech.feetech import FeetechMotorsBus
 
             motors_buses[key] = FeetechMotorsBus(cfg)
 
@@ -52,13 +38,16 @@ def make_motors_buses_from_configs(motors_bus_configs: dict[str, MotorsBusConfig
 
 def make_motors_bus(motor_type: str, **kwargs) -> MotorsBus:
     if motor_type == "dynamixel":
-        from lerobot.common.robot_devices.motors.dynamixel import DynamixelMotorsBus
+        from .configs import DynamixelMotorsBusConfig
+        from .dynamixel import DynamixelMotorsBus
 
         config = DynamixelMotorsBusConfig(**kwargs)
         return DynamixelMotorsBus(config)
 
     elif motor_type == "feetech":
-        from lerobot.common.robot_devices.motors.feetech import FeetechMotorsBus
+        from feetech import FeetechMotorsBus
+
+        from .configs import FeetechMotorsBusConfig
 
         config = FeetechMotorsBusConfig(**kwargs)
         return FeetechMotorsBus(config)

lerobot/common/policies/act/modeling_act.py

@@ -119,7 +119,9 @@ class ACTPolicy(PreTrainedPolicy):
         batch = self.normalize_inputs(batch)
         if self.config.image_features:
             batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
-            batch["observation.images"] = [batch[key] for key in self.config.image_features]
+            batch["observation.images"] = torch.stack(
+                [batch[key] for key in self.config.image_features], dim=-4
+            )
 
         # If we are doing temporal ensembling, do online updates where we keep track of the number of actions
         # we are ensembling over.
@@ -147,8 +149,9 @@ class ACTPolicy(PreTrainedPolicy):
         batch = self.normalize_inputs(batch)
         if self.config.image_features:
             batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
-            batch["observation.images"] = [batch[key] for key in self.config.image_features]
-
+            batch["observation.images"] = torch.stack(
+                [batch[key] for key in self.config.image_features], dim=-4
+            )
         batch = self.normalize_targets(batch)
         actions_hat, (mu_hat, log_sigma_x2_hat) = self.model(batch)
 
@@ -410,10 +413,11 @@ class ACT(nn.Module):
                 "actions must be provided when using the variational objective in training mode."
             )
 
-        if "observation.images" in batch:
-            batch_size = batch["observation.images"][0].shape[0]
-        else:
-            batch_size = batch["observation.environment_state"].shape[0]
+        batch_size = (
+            batch["observation.images"]
+            if "observation.images" in batch
+            else batch["observation.environment_state"]
+        ).shape[0]
 
         # Prepare the latent for input to the transformer encoder.
         if self.config.use_vae and "action" in batch:
@@ -486,21 +490,20 @@ class ACT(nn.Module):
             all_cam_features = []
             all_cam_pos_embeds = []
 
-            # For a list of images, the H and W may vary but H*W is constant.
-            for img in batch["observation.images"]:
-                cam_features = self.backbone(img)["feature_map"]
+            for cam_index in range(batch["observation.images"].shape[-4]):
+                cam_features = self.backbone(batch["observation.images"][:, cam_index])["feature_map"]
+                # TODO(rcadene, alexander-soare): remove call to `.to` to speedup forward ; precompute and use
+                # buffer
                 cam_pos_embed = self.encoder_cam_feat_pos_embed(cam_features).to(dtype=cam_features.dtype)
-                cam_features = self.encoder_img_feat_input_proj(cam_features)
-
-                # Rearrange features to (sequence, batch, dim).
-                cam_features = einops.rearrange(cam_features, "b c h w -> (h w) b c")
-                cam_pos_embed = einops.rearrange(cam_pos_embed, "b c h w -> (h w) b c")
-
+                cam_features = self.encoder_img_feat_input_proj(cam_features)  # (B, C, h, w)
                 all_cam_features.append(cam_features)
                 all_cam_pos_embeds.append(cam_pos_embed)
-
-            encoder_in_tokens.extend(torch.cat(all_cam_features, axis=0))
-            encoder_in_pos_embed.extend(torch.cat(all_cam_pos_embeds, axis=0))
+            # Concatenate camera observation feature maps and positional embeddings along the width dimension,
+            # and move to (sequence, batch, dim).
+            all_cam_features = torch.cat(all_cam_features, axis=-1)
+            encoder_in_tokens.extend(einops.rearrange(all_cam_features, "b c h w -> (h w) b c"))
+            all_cam_pos_embeds = torch.cat(all_cam_pos_embeds, axis=-1)
+            encoder_in_pos_embed.extend(einops.rearrange(all_cam_pos_embeds, "b c h w -> (h w) b c"))
 
         # Stack all tokens along the sequence dimension.
         encoder_in_tokens = torch.stack(encoder_in_tokens, axis=0)

lerobot/common/policies/diffusion/modeling_diffusion.py

@@ -33,7 +33,7 @@ from diffusers.schedulers.scheduling_ddim import DDIMScheduler
 from diffusers.schedulers.scheduling_ddpm import DDPMScheduler
 from torch import Tensor, nn
 
-from lerobot.common.constants import OBS_ENV, OBS_ROBOT
+from lerobot.common.constants import OBS_ENV_STATE, OBS_STATE
 from lerobot.common.policies.diffusion.configuration_diffusion import DiffusionConfig
 from lerobot.common.policies.normalize import Normalize, Unnormalize
 from lerobot.common.policies.pretrained import PreTrainedPolicy
@@ -238,8 +238,8 @@ class DiffusionModel(nn.Module):
 
     def _prepare_global_conditioning(self, batch: dict[str, Tensor]) -> Tensor:
         """Encode image features and concatenate them all together along with the state vector."""
-        batch_size, n_obs_steps = batch[OBS_ROBOT].shape[:2]
-        global_cond_feats = [batch[OBS_ROBOT]]
+        batch_size, n_obs_steps = batch[OBS_STATE].shape[:2]
+        global_cond_feats = [batch[OBS_STATE]]
         # Extract image features.
         if self.config.image_features:
             if self.config.use_separate_rgb_encoder_per_camera:
@@ -269,7 +269,7 @@ class DiffusionModel(nn.Module):
             global_cond_feats.append(img_features)
 
         if self.config.env_state_feature:
-            global_cond_feats.append(batch[OBS_ENV])
+            global_cond_feats.append(batch[OBS_ENV_STATE])
 
         # Concatenate features then flatten to (B, global_cond_dim).
         return torch.cat(global_cond_feats, dim=-1).flatten(start_dim=1)

lerobot/common/policies/factory.py

@@ -25,7 +25,6 @@ from lerobot.common.envs.utils import env_to_policy_features
 from lerobot.common.policies.act.configuration_act import ACTConfig
 from lerobot.common.policies.diffusion.configuration_diffusion import DiffusionConfig
 from lerobot.common.policies.pi0.configuration_pi0 import PI0Config
-from lerobot.common.policies.pi0fast.configuration_pi0fast import PI0FASTConfig
 from lerobot.common.policies.pretrained import PreTrainedPolicy
 from lerobot.common.policies.tdmpc.configuration_tdmpc import TDMPCConfig
 from lerobot.common.policies.vqbet.configuration_vqbet import VQBeTConfig
@@ -55,10 +54,6 @@ def get_policy_class(name: str) -> PreTrainedPolicy:
         from lerobot.common.policies.pi0.modeling_pi0 import PI0Policy
 
         return PI0Policy
-    elif name == "pi0fast":
-        from lerobot.common.policies.pi0fast.modeling_pi0fast import PI0FASTPolicy
-
-        return PI0FASTPolicy
     else:
         raise NotImplementedError(f"Policy with name {name} is not implemented.")
 
@@ -74,8 +69,6 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
         return VQBeTConfig(**kwargs)
     elif policy_type == "pi0":
         return PI0Config(**kwargs)
-    elif policy_type == "pi0fast":
-        return PI0FASTConfig(**kwargs)
     else:
         raise ValueError(f"Policy type '{policy_type}' is not available.")
 

lerobot/common/policies/pi0/modeling_pi0.py

@@ -57,7 +57,7 @@ import torch.nn.functional as F  # noqa: N812
 from torch import Tensor, nn
 from transformers import AutoTokenizer
 
-from lerobot.common.constants import ACTION, OBS_ROBOT
+from lerobot.common.constants import ACTION, OBS_STATE
 from lerobot.common.policies.normalize import Normalize, Unnormalize
 from lerobot.common.policies.pi0.configuration_pi0 import PI0Config
 from lerobot.common.policies.pi0.paligemma_with_expert import (
@@ -271,7 +271,7 @@ class PI0Policy(PreTrainedPolicy):
         self.eval()
 
         if self.config.adapt_to_pi_aloha:
-            batch[OBS_ROBOT] = self._pi_aloha_decode_state(batch[OBS_ROBOT])
+            batch[OBS_STATE] = self._pi_aloha_decode_state(batch[OBS_STATE])
 
         batch = self.normalize_inputs(batch)
 
@@ -303,7 +303,7 @@ class PI0Policy(PreTrainedPolicy):
     def forward(self, batch: dict[str, Tensor], noise=None, time=None) -> tuple[Tensor, dict[str, Tensor]]:
         """Do a full training forward pass to compute the loss"""
         if self.config.adapt_to_pi_aloha:
-            batch[OBS_ROBOT] = self._pi_aloha_decode_state(batch[OBS_ROBOT])
+            batch[OBS_STATE] = self._pi_aloha_decode_state(batch[OBS_STATE])
             batch[ACTION] = self._pi_aloha_encode_actions_inv(batch[ACTION])
 
         batch = self.normalize_inputs(batch)
@@ -313,7 +313,7 @@ class PI0Policy(PreTrainedPolicy):
         state = self.prepare_state(batch)
         lang_tokens, lang_masks = self.prepare_language(batch)
         actions = self.prepare_action(batch)
-        actions_is_pad = batch.get("action_is_pad")
+        actions_is_pad = batch.get("actions_is_pad")
 
         loss_dict = {}
         losses = self.model.forward(images, img_masks, lang_tokens, lang_masks, state, actions, noise, time)
@@ -380,7 +380,7 @@ class PI0Policy(PreTrainedPolicy):
 
     def prepare_language(self, batch) -> tuple[Tensor, Tensor]:
         """Tokenize the text input"""
-        device = batch[OBS_ROBOT].device
+        device = batch[OBS_STATE].device
         tasks = batch["task"]
 
         # PaliGemma prompt has to end with a new line
@@ -427,7 +427,7 @@ class PI0Policy(PreTrainedPolicy):
 
     def prepare_state(self, batch):
         """Pad state"""
-        state = pad_vector(batch[OBS_ROBOT], self.config.max_state_dim)
+        state = pad_vector(batch[OBS_STATE], self.config.max_state_dim)
         return state
 
     def prepare_action(self, batch):

lerobot/common/policies/pi0fast/configuration_pi0fast.py

@@ -1,136 +0,0 @@
-from dataclasses import dataclass, field
-
-from lerobot.common.optim.optimizers import AdamWConfig
-from lerobot.common.optim.schedulers import (
-    CosineDecayWithWarmupSchedulerConfig,
-)
-from lerobot.configs.policies import PreTrainedConfig
-from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
-
-
-@PreTrainedConfig.register_subclass("pi0fast")
-@dataclass
-class PI0FASTConfig(PreTrainedConfig):
-    # Input / output structure.
-    n_obs_steps: int = 1
-    chunk_size: int = 10
-    n_action_steps: int = 5
-
-    normalization_mapping: dict[str, NormalizationMode] = field(
-        default_factory=lambda: {
-            "VISUAL": NormalizationMode.IDENTITY,
-            "STATE": NormalizationMode.MEAN_STD,
-            "ACTION": NormalizationMode.MEAN_STD,
-        }
-    )
-
-    # Shorter state and action vectors will be padded
-    max_state_dim: int = 32  # 32
-    max_action_dim: int = 32  # 32
-
-    # Image preprocessing
-    resize_imgs_with_padding: tuple[int, int] = (224, 224)
-    interpolate_like_pi: bool = False
-
-    # Add empty images. Used by pi0_aloha_sim which adds the empty
-    # left and right wrist cameras in addition to the top camera.
-    empty_cameras: int = 0
-
-    # Converts the joint and gripper values from the standard Aloha space to
-    # the space used by the pi internal runtime which was used to train the base model.
-    adapt_to_pi_aloha: bool = False
-
-    # Converts joint dimensions to deltas with respect to the current state before passing to the model.
-    # Gripper dimensions will remain in absolute values.
-    use_delta_joint_actions_aloha: bool = False
-
-    # Tokenizer
-    tokenizer_max_length: int = 48
-
-    # Projector
-    proj_width: int = 1024
-
-    # Decoding
-    max_decoding_steps: int = 256
-    fast_skip_tokens: int = 128  # Skip last 128 tokens in PaliGemma vocab since they are special tokens
-    max_input_seq_len: int = 256  # 512
-
-    # Utils
-    use_cache: bool = True
-
-    # Frozen parameters
-    freeze_vision_encoder: bool = True
-    freeze_lm_head: bool = True
-
-    # Training presets
-    optimizer_lr: float = 1e-4
-    optimizer_betas: tuple[float, float] = (0.9, 0.95)
-    optimizer_eps: float = 1e-8
-    optimizer_weight_decay: float = 1e-5
-
-    scheduler_warmup_steps: int = 1_000
-    scheduler_decay_steps: int = 30_000
-    scheduler_decay_lr: float = 2.5e-6
-
-    checkpoint_path: str = None
-
-    padding_side: str = "right"
-
-    precision: str = "bfloat16"
-    grad_clip_norm: float = 1
-
-    # Allows padding/truncation of generated action tokens during detokenization to ensure decoding.
-    # In the original version, tensors of 0s were generated if shapes didn't match for stable decoding.
-    relaxed_action_decoding: bool = True
-
-    def __post_init__(self):
-        super().__post_init__()
-
-        """Input validation (not exhaustive)."""
-        if self.n_action_steps > self.chunk_size:
-            raise ValueError(
-                f"The chunk size is the upper bound for the number of action steps per model invocation. Got "
-                f"{self.n_action_steps} for `n_action_steps` and {self.chunk_size} for `chunk_size`."
-            )
-        if self.n_obs_steps != 1:
-            raise ValueError(
-                f"Multiple observation steps not handled yet. Got `nobs_steps={self.n_obs_steps}`"
-            )
-
-    def validate_features(self) -> None:
-        for i in range(self.empty_cameras):
-            key = f"observation.images.empty_camera_{i}"
-            empty_camera = PolicyFeature(
-                type=FeatureType.VISUAL,
-                shape=(3, 480, 640),
-            )
-            self.input_features[key] = empty_camera
-
-    def get_optimizer_preset(self) -> AdamWConfig:
-        return AdamWConfig(
-            lr=self.optimizer_lr,
-            betas=self.optimizer_betas,
-            eps=self.optimizer_eps,
-            weight_decay=self.optimizer_weight_decay,
-            grad_clip_norm=self.grad_clip_norm,
-        )
-
-    def get_scheduler_preset(self):
-        return CosineDecayWithWarmupSchedulerConfig(
-            peak_lr=self.optimizer_lr,
-            decay_lr=self.scheduler_decay_lr,
-            num_warmup_steps=self.scheduler_warmup_steps,
-            num_decay_steps=self.scheduler_decay_steps,
-        )
-
-    @property
-    def observation_delta_indices(self) -> None:
-        return None
-
-    @property
-    def action_delta_indices(self) -> list:
-        return list(range(self.chunk_size))
-
-    @property
-    def reward_delta_indices(self) -> None:
-        return None

lerobot/common/policies/pi0fast/modeling_pi0fast.py

@@ -1,973 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2025 Physical Intelligence and The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-"""
-π0+FAST: Efficient Action Tokenization for Vision-Language-Action Models
-
-[Paper](https://arxiv.org/abs/2501.09747)
-[Jax code](https://github.com/Physical-Intelligence/openpi)
-
-Designed by Physical Intelligence. Ported from Jax by Hugging Face.
-
-Example of finetuning the pi0+FAST pretrained model (`pi0_fast_base` in `openpi`):
-```bash
-python lerobot/scripts/train.py \
---policy.path=lerobot/pi0fast_base \
---dataset.repo_id=danaaubakirova/koch_test
-```
-
-Example of training the pi0+FAST neural network with from scratch:
-```bash
-python lerobot/scripts/train.py \
---policy.type=pi0fast \
---dataset.repo_id=danaaubakirova/koch_test
-```
-
-Example of using the pi0 pretrained model outside LeRobot training framework:
-```python
-policy = PI0FASTPolicy.from_pretrained("lerobot/pi0fast_base")
-```
-
-"""
-
-from collections import deque
-from functools import partial
-
-import numpy as np
-import torch
-import torch.nn.functional as F  # noqa: N812
-from PIL import Image
-from scipy.fft import idct
-from torch import Tensor, nn
-from transformers import AutoProcessor, AutoTokenizer, PaliGemmaForConditionalGeneration
-from transformers.cache_utils import HybridCache, StaticCache
-from transformers.models.auto import CONFIG_MAPPING
-
-from lerobot.common.constants import ACTION, OBS_ROBOT
-from lerobot.common.policies.normalize import Normalize, Unnormalize
-from lerobot.common.policies.pi0fast.configuration_pi0fast import PI0FASTConfig
-from lerobot.common.policies.pretrained import PreTrainedPolicy
-
-PRECISION = {
-    "float16": torch.float16,
-    "float32": torch.float32,
-    "bfloat16": torch.bfloat16,
-}
-
-
-def normalize(x, min_val, max_val):
-    return (x - min_val) / (max_val - min_val)
-
-
-def unnormalize(x, min_val, max_val):
-    return x * (max_val - min_val) + min_val
-
-
-def safe_arcsin(value):
-    # This ensures that the input stays within
-    # [−1,1] to avoid invalid values for arcsin
-    return torch.arcsin(torch.clamp(value, -1.0, 1.0))
-
-
-def aloha_gripper_to_angular(value):
-    # Aloha transforms the gripper positions into a linear space. The following code
-    # reverses this transformation to be consistent with pi0 which is pretrained in
-    # angular space.
-    #
-    # These values are coming from the Aloha code:
-    # PUPPET_GRIPPER_POSITION_OPEN, PUPPET_GRIPPER_POSITION_CLOSED
-    value = unnormalize(value, min_val=0.01844, max_val=0.05800)
-
-    # This is the inverse of the angular to linear transformation inside the Interbotix code.
-    def linear_to_radian(linear_position, arm_length, horn_radius):
-        value = (horn_radius**2 + linear_position**2 - arm_length**2) / (2 * horn_radius * linear_position)
-        return safe_arcsin(value)
-
-    # The constants are taken from the Interbotix code.
-    value = linear_to_radian(value, arm_length=0.036, horn_radius=0.022)
-
-    # Normalize to [0, 1].
-    # The values 0.4 and 1.5 were measured on an actual Trossen robot.
-    return normalize(value, min_val=0.4, max_val=1.5)
-
-
-def aloha_gripper_from_angular(value):
-    # Convert from the gripper position used by pi0 to the gripper position that is used by Aloha.
-    # Note that the units are still angular but the range is different.
-
-    # The values 0.4 and 1.5 were measured on an actual Trossen robot.
-    value = unnormalize(value, min_val=0.4, max_val=1.5)
-
-    # These values are coming from the Aloha code:
-    # PUPPET_GRIPPER_JOINT_OPEN, PUPPET_GRIPPER_JOINT_CLOSE
-    return normalize(value, min_val=-0.6213, max_val=1.4910)
-
-
-def aloha_gripper_from_angular_inv(value):
-    # Directly inverts the gripper_from_angular function.
-    value = unnormalize(value, min_val=-0.6213, max_val=1.4910)
-    return normalize(value, min_val=0.4, max_val=1.5)
-
-
-class PI0FASTPolicy(PreTrainedPolicy):
-    """Wrapper class around PI0FAST tokenizer and model to train and run inference within LeRobot."""
-
-    config_class = PI0FASTConfig
-    name = "pi0fast"
-
-    def __init__(
-        self,
-        config: PI0FASTConfig,
-        dataset_stats: dict[str, dict[str, Tensor]] | None = None,
-    ):
-        """
-        Args:
-            config: Policy configuration class instance or None, in which case the default instantiation of
-                    the configuration class is used.
-            dataset_stats: Dataset statistics to be used for normalization. If not passed here, it is expected
-                that they will be passed with a call to `load_state_dict` before the policy is used.
-        """
-
-        super().__init__(config)
-        config.validate_features()
-        self.config = config
-
-        self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
-        self.normalize_targets = Normalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-        self.unnormalize_outputs = Unnormalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-
-        self.language_tokenizer = AutoProcessor.from_pretrained("google/paligemma-3b-pt-224")
-        self.model = PI0FAST(config)
-
-        self.reset()
-
-    def reset(self):
-        """This should be called whenever the environment is reset."""
-        self._action_queue = deque([], maxlen=self.config.n_action_steps)
-
-    def get_optim_params(self) -> dict:
-        return self.parameters()
-
-    def _pi_aloha_decode_state(self, state):
-        # Flip the joints.
-        for motor_idx in [1, 2, 8, 9]:
-            state[:, motor_idx] *= -1
-        # Reverse the gripper transformation that is being applied by the Aloha runtime.
-        for motor_idx in [6, 13]:
-            state[:, motor_idx] = aloha_gripper_to_angular(state[:, motor_idx])
-        return state
-
-    def _pi_aloha_encode_actions(self, actions):
-        # Flip the joints.
-        for motor_idx in [1, 2, 8, 9]:
-            actions[:, :, motor_idx] *= -1
-        # Reverse the gripper transformation that is being applied by the Aloha runtime.
-        for motor_idx in [6, 13]:
-            actions[:, :, motor_idx] = aloha_gripper_from_angular(actions[:, :, motor_idx])
-        return actions
-
-    def _pi_aloha_encode_actions_inv(self, actions):
-        # Flip the joints again.
-        for motor_idx in [1, 2, 8, 9]:
-            actions[:, :, motor_idx] *= -1
-        # Reverse the gripper transformation that is being applied by the Aloha runtime.
-        for motor_idx in [6, 13]:
-            actions[:, :, motor_idx] = aloha_gripper_from_angular_inv(actions[:, :, motor_idx])
-        return actions
-
-    @torch.no_grad
-    def select_action(self, batch: dict[str, Tensor]) -> Tensor:
-        """Select a single action given environment observations.
-
-        This method wraps `select_actions` in order to return one action at a time for execution in the
-        environment. It works by managing the actions in a queue and only calling `select_actions` when the
-        queue is empty.
-        """
-        self.eval()
-
-        if self.config.adapt_to_pi_aloha:
-            batch[OBS_ROBOT] = self._pi_aloha_decode_state(batch[OBS_ROBOT])
-
-        batch = self.normalize_inputs(batch)
-
-        # Action queue logic for n_action_steps > 1. When the action_queue is depleted, populate it by
-        # querying the policy.
-        if len(self._action_queue) == 0:
-            actions = self.model.generate_actions(batch)
-
-            actions = actions[:, : self.config.n_action_steps]
-
-            original_action_dim = self.config.action_feature.shape[
-                0
-            ]  # self.config.max_action_dim  # self.config.action_feature.shape[0]
-            actions = actions[:, :, :original_action_dim]
-
-            actions = self.unnormalize_outputs({"action": actions})["action"]
-
-            if self.config.adapt_to_pi_aloha:
-                actions = self._pi_aloha_encode_actions(actions)
-
-            # `self.model.forward` returns a (batch_size, n_action_steps, action_dim) tensor, but the queue
-            # effectively has shape (n_action_steps, batch_size, *), hence the transpose.
-            self._action_queue.extend(actions.transpose(0, 1))
-        return self._action_queue.popleft()
-
-    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
-        if self.config.adapt_to_pi_aloha:
-            batch[OBS_ROBOT] = self._pi_aloha_decode_state(batch[OBS_ROBOT])
-            batch[ACTION] = self._pi_aloha_encode_actions_inv(batch[ACTION])
-        batch = self.normalize_inputs(batch)
-        batch = self.normalize_targets(batch)
-        loss_dict = self.model.forward(batch)
-        return loss_dict["loss"], loss_dict
-
-
-def block_causal_update_causal_mask(
-    attention_mask,
-    token_type_ids=None,
-    past_key_values=None,
-    cache_position=None,
-    input_tensor=None,
-    attn_implementation: str = "eager",
-    dtype: torch.dtype = "float32",
-):
-    """
-    Update the causal mask during training and generation. It can be customized to different attention masks.
-    """
-    if attn_implementation == "flash_attention_2":
-        if attention_mask is not None and 0.0 in attention_mask:
-            return attention_mask
-        return None
-    using_static_cache = isinstance(past_key_values, StaticCache)
-    min_dtype = torch.finfo(dtype).min
-
-    if input_tensor is None:
-        input_tensor = attention_mask
-
-    inputs_lead_dim, sequence_length = input_tensor.shape[:2]
-
-    if using_static_cache or isinstance(past_key_values, HybridCache):
-        target_length = past_key_values.get_max_cache_shape()
-    else:
-        target_length = (
-            attention_mask.shape[-1]
-            if isinstance(attention_mask, torch.Tensor)
-            else cache_position[0] + sequence_length + 1
-        )
-
-    # Handle precomputed attention masks
-    if attention_mask is not None and attention_mask.dim() == 4:
-        return attention_mask
-
-    # Causal mask initialization
-    causal_mask = torch.full(
-        (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=cache_position.device
-    )
-
-    # Standard causal masking (triu ensures tokens can only attend to past)
-    if sequence_length != 1:
-        causal_mask = torch.triu(causal_mask, diagonal=1)
-
-        # Apply block causal mask
-        if token_type_ids is not None:
-            token_type_ids = token_type_ids.to(causal_mask.device).bool()
-            cumsum = torch.cumsum(token_type_ids, dim=1)
-            block_causal_mask = cumsum[:, None, :] <= cumsum[:, :, None]
-
-            # Combine causal_mask with block-wise attention mask
-            causal_mask = torch.where(block_causal_mask, 0.0, causal_mask)
-            causal_mask = causal_mask[:, None, :, :]
-        else:
-            # Apply past cache position constraint
-            causal_mask *= torch.arange(target_length, device=cache_position.device) > cache_position.reshape(
-                -1, 1
-            )
-            causal_mask = causal_mask[None, None, :, :].expand(inputs_lead_dim, 1, -1, -1)
-    else:
-        # Apply past cache position constraint
-        causal_mask *= torch.arange(target_length, device=cache_position.device) > cache_position.reshape(
-            -1, 1
-        )
-        causal_mask = causal_mask[None, None, :, :].expand(inputs_lead_dim, 1, -1, -1)
-
-    if attention_mask is not None:
-        causal_mask = causal_mask.clone()  # Copy to contiguous memory for in-place edits
-        mask_length = attention_mask.shape[-1]
-
-        # Apply padding mask
-        padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :].to(
-            causal_mask.device
-        )
-        padding_mask = padding_mask == 0
-        causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
-            padding_mask, min_dtype
-        )
-
-    return causal_mask
-
-
-def prepare_inputs_for_generation(
-    # self,
-    input_ids,
-    past_key_values=None,
-    inputs_embeds=None,
-    cache_position=None,
-    position_ids=None,
-    pixel_values=None,
-    attention_mask=None,
-    token_type_ids=None,
-    use_cache=True,
-    num_logits_to_keep=None,
-    labels=None,
-    self=None,
-    **kwargs,
-):
-    # create block causal attention
-    if cache_position[0] > 0 and input_ids.shape[1] > 0:
-        input_tensor = input_ids[:, -1:]
-        new_positions = (
-            torch.ones(
-                (position_ids.shape[0], input_ids.shape[1]),
-                dtype=position_ids.dtype,
-                device=position_ids.device,
-            ).cumsum(-1)
-            + position_ids[:, -1:]
-        )
-        position_ids = torch.cat([position_ids, new_positions], dim=-1)
-    else:
-        input_tensor = inputs_embeds
-    attention_mask = block_causal_update_causal_mask(
-        attention_mask=attention_mask,
-        past_key_values=past_key_values,
-        cache_position=cache_position,
-        input_tensor=input_tensor,
-        token_type_ids=token_type_ids,
-        dtype=self.dtype,
-        attn_implementation=self.config.text_config._attn_implementation,
-    )
-    # Overwritten -- custom `position_ids` and `pixel_values` handling
-    model_inputs = self.language_model.prepare_inputs_for_generation(
-        input_ids,
-        past_key_values=past_key_values,
-        inputs_embeds=inputs_embeds,
-        attention_mask=attention_mask,
-        position_ids=position_ids,
-        cache_position=cache_position,
-        use_cache=use_cache,
-        num_logits_to_keep=num_logits_to_keep,
-        token_type_ids=token_type_ids,
-        **kwargs,
-    )
-
-    # Position_ids in Paligemma are 1-indexed
-    if model_inputs.get("position_ids") is not None:
-        model_inputs["position_ids"] += 1
-    # If we're in cached decoding stage, pixel values should be None because input ids do not contain special image token anymore
-    # Otherwise we need pixel values to be passed to model. NOTE: use_cache=False needs pixel_values always
-    if cache_position[0] == 0:
-        model_inputs["pixel_values"] = pixel_values
-    is_training = token_type_ids is not None and labels is not None
-    if cache_position[0] == 0 and isinstance(past_key_values, HybridCache):
-        input_tensor = inputs_embeds if inputs_embeds is not None else input_ids
-        causal_mask = self._update_causal_mask(
-            attention_mask, token_type_ids, past_key_values, cache_position, input_tensor, is_training
-        )
-        model_inputs["attention_mask"] = causal_mask
-
-    return model_inputs
-
-
-class PI0FAST(nn.Module):
-    def __init__(self, config: PI0FASTConfig):
-        super().__init__()
-        self.config = config
-
-        # TODO: move tokenizers in Policy
-        fast_tokenizer_path = "physical-intelligence/fast"
-        pi0_paligemma_path = "google/paligemma-3b-pt-224"
-        self.paligemma_tokenizer = AutoTokenizer.from_pretrained(pi0_paligemma_path)
-        self.processor = AutoProcessor.from_pretrained(pi0_paligemma_path)
-        self.fast_tokenizer = AutoProcessor.from_pretrained(fast_tokenizer_path, trust_remote_code=True)
-        self.fast_skip_tokens = self.config.fast_skip_tokens
-        self.max_input_seq_len = self.config.max_input_seq_len
-        self.action_horizon = self.config.chunk_size
-        self.action_dim = self.config.action_feature.shape[
-            0
-        ]  # self.config.max_action_dim  # self.config.action_feature.shape[0]
-        precision = config.precision
-        torch_precision = PRECISION.get(precision, torch.float32)
-        self.pad_token_id = (
-            self.paligemma_tokenizer.pad_token_id
-            if hasattr(self.paligemma_tokenizer, "pad_token_id")
-            else self.paligemma_tokenizer.eos_token_id
-        )
-
-        paligemma_config = CONFIG_MAPPING["paligemma"](
-            transformers_version="4.48.1",
-            _vocab_size=257152,
-            bos_token_id=2,
-            eos_token_id=1,
-            hidden_size=2048,
-            image_token_index=257152,
-            model_type="paligemma",
-            pad_token_id=0,
-            projection_dim=2048,
-            text_config={
-                "hidden_activation": "gelu_pytorch_tanh",
-                "hidden_size": 2048,
-                "intermediate_size": 16384,
-                "model_type": "gemma",
-                "num_attention_heads": 8,
-                "num_hidden_layers": 18,
-                "num_image_tokens": 256,
-                "num_key_value_heads": 1,
-                "torch_dtype": precision,
-                "vocab_size": 257152,
-                "_attn_implementation": "eager",
-            },
-            vision_config={
-                "hidden_size": 1152,
-                "intermediate_size": 4304,
-                "model_type": "siglip_vision_model",
-                "num_attention_heads": 16,
-                "num_hidden_layers": 27,
-                "num_image_tokens": 256,
-                "patch_size": 14,
-                "projection_dim": 2048,
-                "projector_hidden_act": "gelu_pytorch_tanh",
-                "torch_dtype": precision,
-                "vision_use_head": False,
-            },
-        )
-        self.pi0_paligemma = PaliGemmaForConditionalGeneration(config=paligemma_config)
-
-        self.pi0_paligemma.prepare_inputs_for_generation = partial(
-            prepare_inputs_for_generation, self=self.pi0_paligemma
-        )
-        # change important stuff in bf16
-        params_to_change_dtype = [
-            "language_model",
-            "vision_tower",
-            "multi_modal",
-        ]
-        for name, param in self.pi0_paligemma.named_parameters():
-            if any(selector in name for selector in params_to_change_dtype):
-                param.data = param.data.to(dtype=torch_precision)
-        self.set_requires_grad()
-        self.image_keys = self.config.image_features.keys()
-        self.ignore_index = self.pi0_paligemma.config.ignore_index
-        self.padding_side = self.config.padding_side
-
-    def set_requires_grad(self):
-        if self.config.freeze_vision_encoder:
-            self.pi0_paligemma.vision_tower.eval()
-            for params in self.pi0_paligemma.vision_tower.parameters():
-                params.requires_grad = False
-        # To avoid unused params issue with distributed training
-        if self.config.freeze_lm_head:
-            for name, params in self.pi0_paligemma.named_parameters():
-                if "embed_tokens" in name:  # lm heads and embedding layer are tied
-                    params.requires_grad = False
-
-    def embed_tokens(self, tokens: torch.Tensor):
-        return self.pi0_paligemma.language_model.model.embed_tokens(tokens)
-
-    def prepare_inputs_for_generation(self, *args, **kwargs):
-        return self.pi0_paligemma.prepare_inputs_for_generation(*args, **kwargs)
-
-    def prepare_images(self, batch):
-        """Preprocess LeRobot batch into Pi0 inputs"""
-        images = []
-        img_masks = []
-        present_img_keys = [key for key in self.image_keys if key in batch]
-        if len(present_img_keys) == 0:
-            raise ValueError(
-                f"All image features are missing from the batch. At least one expected. (batch: {batch.keys()}) (image_features:{self.config.image_features})"
-            )
-
-        # Preprocess image features present in the batch
-        num_empty_cameras = 0
-        for key in self.image_keys:
-            if key in present_img_keys:
-                img = batch[key]
-
-                if self.config.resize_imgs_with_padding is not None:
-                    img = resize_with_pad(
-                        img,
-                        *self.config.resize_imgs_with_padding,
-                        pad_value=0,
-                        interpolate_like_pi=self.config.interpolate_like_pi,
-                    )
-
-                # Normalize from range [0,1] to [-1,1] as expacted by siglip
-                img = img * 2.0 - 1.0
-
-                bsize = img.shape[0]
-                device = img.device
-                mask = torch.ones(bsize, dtype=torch.bool, device=device)
-            else:
-                if num_empty_cameras >= self.config.empty_cameras:
-                    continue
-                img = torch.ones_like(img) * -1
-                bsize = img.shape[0]
-                device = img.device
-                mask = torch.ones(bsize, dtype=torch.bool, device=device)
-                num_empty_cameras += 1
-
-            images.append(img)
-            img_masks.append(mask)
-        return images, img_masks
-
-    def normalize_actions(self, actions: torch.Tensor) -> torch.Tensor:
-        mins = actions.amin(dim=(1, 2), keepdim=True)  # [0]
-        maxs = actions.amax(dim=(1, 2), keepdim=True)  # [0]
-        return 2 * (actions - mins) / (maxs - mins + 1e-8) - 1
-
-    def _act_tokens_to_paligemma_tokens(self, tokens: torch.Tensor) -> torch.Tensor:
-        out = self.paligemma_tokenizer.vocab_size - 1 - self.fast_skip_tokens - tokens
-        return out
-
-    def fast_tokenizer_wrapper(self, actions_norm):
-        """
-        A wrapper for self.fast_tokenizer that ensures batch processing,
-        conversion to PyTorch tensors, and returns a dictionary without padding.
-        """
-        batch_tokens = self.fast_tokenizer(actions_norm)
-        fast_out = self.processor.tokenizer.pad({"input_ids": batch_tokens}, return_tensors="pt")
-
-        return fast_out
-
-    def create_token_type_ids(self, padded_mask: torch.Tensor, prefix_len: int) -> torch.Tensor:
-        token_type_ids = torch.zeros_like(padded_mask, dtype=torch.bool)
-        # Compute cumulative sum mask
-        cumsum_mask = (padded_mask != 0).cumsum(dim=1)
-        # Suffix block (everything after prefix_len)
-        suffix_mask = cumsum_mask > prefix_len
-        token_type_ids = suffix_mask
-        return token_type_ids
-
-    def create_input_tokens(self, state, lang_text, actions=None):
-        bsize = state.shape[0]
-        device = state.device
-        bins = torch.linspace(-1, 1, 256 + 1, device=device)[:-1]
-        discretized = torch.bucketize(state, bins) - 1
-        discretized = discretized[:, :32]
-
-        prefix_texts = []
-        state_text = []
-        for txt, disc in zip(lang_text, discretized, strict=False):
-            cleaned = txt.lower().strip().replace("_", " ")
-            state_str = " ".join(str(val.item()) for val in disc)
-            prefix_texts.append(f"Task: {cleaned}, State: {state_str};\n")
-            state_text.append(f"State: {state_str};\n")
-
-        prefix_out = self.paligemma_tokenizer(
-            prefix_texts, add_special_tokens=True, return_tensors="pt", padding="longest", truncation=False
-        )
-        prefix_ids = prefix_out["input_ids"].to(device)
-        prefix_mask = prefix_out["attention_mask"].to(device)
-        prefix_lens = prefix_mask.sum(dim=1)[:, None].cpu()
-
-        if actions is not None:
-            actions_norm = self.normalize_actions(actions)
-            actions_pad = F.pad(
-                actions_norm, (0, max(0, self.config.max_action_dim - actions_norm.shape[2])), value=0
-            )[:, :, : self.config.max_action_dim]
-            fast_out = self.fast_tokenizer_wrapper(
-                actions_pad.cpu(),
-            )
-            act_ids = fast_out["input_ids"]
-            act_mask = fast_out["attention_mask"].to(device)
-
-            act_ids = self._act_tokens_to_paligemma_tokens(act_ids).to(device)
-            # Replace action with 0 to pad tokens
-            act_ids = torch.where(
-                act_ids == self.paligemma_tokenizer.vocab_size - 1 - self.fast_skip_tokens,
-                self.pad_token_id,
-                act_ids,
-            )
-
-            eos_token = torch.tensor(
-                [self.paligemma_tokenizer.eos_token_id], dtype=torch.long, device=device
-            ).expand(bsize, -1)
-            eos_mask = torch.tensor([1], dtype=torch.long, device=device).expand(bsize, -1)
-            bos = self.paligemma_tokenizer("Action: ", add_special_tokens=False, return_tensors="pt")
-            bos_token = bos["input_ids"].expand(act_ids.shape[0], -1).to(device)
-            bos_mask = bos["attention_mask"].expand(act_ids.shape[0], -1).to(device)
-            act_ids = torch.cat([bos_token, act_ids, eos_token], dim=1)
-            act_mask = torch.cat([bos_mask, act_mask, eos_mask], dim=1)
-            act_mask = act_mask.to(device)
-        else:
-            act_ids = torch.empty(bsize, self.pad_token_id, dtype=torch.long, device=device)
-            act_mask = torch.empty(bsize, 0, dtype=torch.long, device=device)
-        final_ids = torch.cat([prefix_ids, act_ids], dim=1)
-
-        final_mask = torch.cat([prefix_mask, act_mask], dim=1)
-        batch_inputs = {"input_ids": final_ids.tolist(), "attention_mask": final_mask.tolist()}
-
-        # Use tokenizer pad function
-        padded_output = self.paligemma_tokenizer.pad(
-            batch_inputs, padding="longest", max_length=180, return_tensors="pt"
-        )
-        padded_mask = padded_output["attention_mask"]
-
-        # define tensor of padding lengths
-        att_mask = (padded_mask != 0).cumsum(dim=1) > prefix_lens
-
-        token_type_ids = self.create_token_type_ids(padded_mask=padded_mask, prefix_len=prefix_lens)
-
-        padded_output["padded_mask"] = padded_output.pop("attention_mask")
-        padded_output["attention_mask"] = att_mask
-        # loss is computed not on prefix, and not on padding
-        padded_output["loss_mask"] = att_mask & padded_output["padded_mask"]
-        padded_output["token_type_ids"] = token_type_ids
-        return padded_output
-
-    def shift_padding_side(
-        self,
-        tokens: torch.Tensor,
-        ar_mask: torch.Tensor,
-        padding_mask: torch.Tensor,
-        loss_mask: torch.Tensor,
-        targets: torch.Tensor,
-        token_type_ids: torch.Tensor,
-        padding_side: str = "right",
-    ) -> tuple[torch.Tensor]:
-        if padding_side not in ["right", "left"]:
-            return tokens, ar_mask, padding_mask, loss_mask, targets, token_type_ids
-
-        new_tokens = torch.empty_like(tokens)
-        new_ar_masks = torch.empty_like(ar_mask)
-        new_padding_mask = torch.empty_like(padding_mask)
-        new_loss_mask = torch.empty_like(loss_mask)
-        new_targets = torch.empty_like(targets)
-        new_token_type_ids = torch.empty_like(token_type_ids)
-        batch_size = tokens.shape[0]
-        for i in range(batch_size):
-            padding_indices = torch.where(padding_mask[i] == 0)[0]
-            non_padding_indices = torch.where(padding_mask[i] == 1)[0]
-            if padding_side == "left":
-                new_indices = torch.cat((padding_indices, non_padding_indices), dim=0)
-            else:
-                new_indices = torch.cat((non_padding_indices, padding_indices), dim=0)
-            new_tokens[i] = tokens[i].index_select(0, new_indices)
-            new_ar_masks[i] = ar_mask[i].index_select(0, new_indices)
-            new_padding_mask[i] = padding_mask[i].index_select(0, new_indices)
-            new_loss_mask[i] = loss_mask[i].index_select(0, new_indices)
-            new_targets[i] = targets[i].index_select(0, new_indices)
-            new_token_type_ids[i] = token_type_ids[i].index_select(0, new_indices)
-
-        return new_tokens, new_ar_masks, new_padding_mask, new_loss_mask, new_targets, new_token_type_ids
-
-    def forward(self, batch: dict[str, Tensor]):
-        device = batch[OBS_ROBOT].device
-        # TODO: keep like this or move to the policy .forward
-        images, img_masks = self.prepare_images(batch)
-
-        padded_outs = self.create_input_tokens(
-            state=batch[OBS_ROBOT],
-            lang_text=batch["task"],
-            actions=batch[ACTION],
-        )
-
-        embs, pad_masks, _, targets, loss_mask, token_type_ids = self.embed_inputs(
-            images,
-            img_masks,
-            padded_outs["input_ids"],
-            padded_outs["padded_mask"],
-            padded_outs["attention_mask"],
-            padded_outs["loss_mask"],
-            padded_outs["token_type_ids"],
-            padding_side=self.padding_side,
-        )
-        position_ids = torch.cumsum(pad_masks, dim=1) - 1
-        token_type_ids = token_type_ids.to(dtype=torch.int64)
-        past_seen_tokens = 0
-        cache_position = torch.arange(past_seen_tokens, past_seen_tokens + embs.shape[1], device=embs.device)
-        pad_masks = block_causal_update_causal_mask(
-            attention_mask=pad_masks,
-            past_key_values=None,
-            cache_position=cache_position,
-            input_tensor=embs,
-            token_type_ids=token_type_ids,
-            dtype=self.pi0_paligemma.dtype,
-            attn_implementation=self.pi0_paligemma.config.text_config._attn_implementation,
-        )
-        outputs = self.pi0_paligemma.forward(
-            input_ids=None,
-            token_type_ids=None,
-            attention_mask=pad_masks,
-            position_ids=position_ids,
-            past_key_values=None,
-            inputs_embeds=embs,
-            use_cache=False,
-            labels=None,
-        )
-
-        logits = outputs.logits
-
-        loss_fct = nn.CrossEntropyLoss(reduction="none")
-
-        # Shift left for next-step prediction
-        logits = logits[:, :-1, :]
-        targets = targets[:, 1:].to(device)  # Shift targets
-        loss_mask = loss_mask[:, 1:].to(device)  # Ensure correct shape
-
-        # Compute per-token loss
-        token_loss = loss_fct(logits.reshape(-1, logits.shape[-1]), targets.reshape(-1))
-
-        # Apply loss mask
-        token_loss = token_loss * loss_mask.reshape(-1)
-
-        # Compute final loss
-        loss = token_loss.sum() / torch.clamp(loss_mask.sum(), min=1)
-
-        # Return loss dictionary
-        loss_dict = {"ce_loss": loss.item(), "loss": loss}
-        return loss_dict
-
-    def decode_actions_with_fast(
-        self,
-        tokens: list[list[int]],
-        *,
-        time_horizon: int | None = None,
-        action_dim: int | None = None,
-        relaxed_decoding: bool = True,
-    ) -> np.array:
-        """
-        Adapt original decoding in FAST to always return actions instead of zeros.
-        """
-        self.time_horizon = (
-            time_horizon or self.fast_tokenizer.time_horizon or self.fast_tokenizer.called_time_horizon
-        )
-        self.action_dim = (
-            action_dim or self.fast_tokenizer.action_dim or self.fast_tokenizer.called_action_dim
-        )
-
-        # Cache the time horizon and action dimension for the next call
-        self.called_time_horizon = self.time_horizon
-        self.called_action_dim = self.action_dim
-
-        assert self.time_horizon is not None and self.action_dim is not None, (
-            "Tokenizer not initialized, call encode() once or pass in time_horizon and action_dim."
-        )
-
-        decoded_actions = []
-        for token in tokens:
-            try:
-                decoded_tokens = self.fast_tokenizer.bpe_tokenizer.decode(token)
-                decoded_dct_coeff = np.array(list(map(ord, decoded_tokens))) + self.fast_tokenizer.min_token
-                if relaxed_decoding:
-                    # Expected sequence length
-                    expected_seq_len = self.time_horizon * self.action_dim
-                    diff = expected_seq_len - decoded_dct_coeff.shape[0]
-                    # Apply truncation if too long
-                    if diff < 0:
-                        decoded_dct_coeff = decoded_dct_coeff[:expected_seq_len]  # Truncate on the right
-                    # Apply padding if too short
-                    elif diff > 0:
-                        decoded_dct_coeff = np.pad(
-                            decoded_dct_coeff, (0, diff), mode="constant", constant_values=0
-                        )
-
-                decoded_dct_coeff = decoded_dct_coeff.reshape(-1, self.action_dim)
-                assert decoded_dct_coeff.shape == (
-                    self.time_horizon,
-                    self.action_dim,
-                ), (
-                    f"Decoded DCT coefficients have shape {decoded_dct_coeff.shape}, expected ({self.time_horizon}, {self.action_dim})"
-                )
-            except Exception as e:
-                print(f"Error decoding tokens: {e}")
-                print(f"Tokens: {token}")
-                decoded_dct_coeff = np.zeros((self.time_horizon, self.action_dim))
-            decoded_actions.append(idct(decoded_dct_coeff / self.fast_tokenizer.scale, axis=0, norm="ortho"))
-        return np.stack(decoded_actions)
-
-    def extract_actions(self, tokens: torch.Tensor, action_horizon: int, action_dim: int) -> torch.Tensor:
-        """
-        Extracts actions from predicted output tokens using the FAST model.
-
-        Args:
-            tokens (torch.Tensor): The input tensor of tokenized outputs.
-            action_horizon (int): The number of timesteps for actions.
-            action_dim (int): The dimensionality of each action.
-
-        Returns:
-            torch.Tensor: The extracted actions as a tensor of shape (action_horizon, action_dim).
-        """
-        # Decode predicted output tokens
-        decoded_tokens = self.paligemma_tokenizer.batch_decode(tokens, skip_special_tokens=True)
-        cleaned_tokens = [
-            tokens_sequence.replace("Action:", "").replace(":", "").strip().split("|")[0].strip()
-            for tokens_sequence in decoded_tokens
-        ]
-        raw_action_tokens = [
-            self.processor.tokenizer.encode(sample_tokens, return_tensors="pt", padding=False)
-            for sample_tokens in cleaned_tokens
-        ]  # something like this should be robust #looks good
-        action_tokens = [
-            self._act_tokens_to_paligemma_tokens(raw_action_token) for raw_action_token in raw_action_tokens
-        ]
-        # returns the tensor of decoded actions per sample in a list
-        decoded_actions = [
-            torch.tensor(
-                self.decode_actions_with_fast(
-                    tok.tolist(),
-                    time_horizon=action_horizon,
-                    action_dim=action_dim,
-                    relaxed_decoding=self.config.relaxed_action_decoding,
-                ),
-                device=tokens.device,
-            ).squeeze(0)
-            for tok in action_tokens
-        ]
-
-        return torch.stack(
-            decoded_actions,
-            dim=0,
-        )
-
-    def generate_actions(self, batch: dict[str, Tensor]):
-        # TODO: keep like this or move to the policy .forward
-        images, img_masks = self.prepare_images(batch)
-
-        padded_outs = self.create_input_tokens(state=batch[OBS_ROBOT], lang_text=batch["task"], actions=None)
-        embs, pad_masks, att_masks2, targets, loss_mask, token_type_ids = self.embed_inputs(
-            images,
-            img_masks,
-            padded_outs["input_ids"],
-            padded_outs["padded_mask"],
-            padded_outs["attention_mask"],
-            padded_outs["loss_mask"],
-            padded_outs["token_type_ids"],
-            padding_side="left",
-        )
-        token_type_ids = token_type_ids.to(dtype=torch.int64)
-        prefix_position_ids = torch.cumsum(pad_masks, dim=1) - 1
-        output_tokens = self.pi0_paligemma.generate(
-            input_ids=None,
-            attention_mask=pad_masks,
-            position_ids=prefix_position_ids,
-            past_key_values=None,
-            inputs_embeds=embs,
-            use_cache=self.config.use_cache,
-            max_new_tokens=self.config.max_decoding_steps,
-            do_sample=False,
-            num_beams=1,
-            token_type_ids=token_type_ids,
-        )
-        actions = self.extract_actions(output_tokens, self.action_horizon, self.action_dim)
-        return actions
-
-    def embed_image(self, image: torch.Tensor):
-        return self.pi0_paligemma.get_image_features(image)
-
-    def embed_inputs(
-        self,
-        images,
-        img_masks,
-        tokens,
-        pad_mask,
-        ar_mask,
-        loss_mask,
-        token_type_ids,
-        padding_side: str = "right",
-    ):
-        # TODO: avoid list in python and torch.cat ; prefer pre-allocation with torch.empty
-        # images are a list of same size
-        # vectorizing everything!
-        device = images[0].device
-        image_embedding_dim = images[0].shape[-1]  # TODO should be from self.config
-        all_images = torch.stack(images, dim=1).to(device)
-        b, n, c, h, w = all_images.shape
-        all_images = all_images.view(b * n, c, h, w)
-        embedded = self.embed_image(all_images).to(device)
-        b_n, p, image_embedding_dim = embedded.shape  # Extract current dimensions
-        m = b_n // b  # Compute the number of images per sample dynamically
-
-        # Reshape dynamically
-        embedded = embedded.view(b, m, p, image_embedding_dim)
-        tokens_embs = self.embed_tokens(tokens.to(device))
-
-        img_masks = torch.stack(img_masks, dim=1).unsqueeze(-1).to(device)
-        num_img_emb = embedded.shape[2]
-        img_pad_masks = img_masks.repeat(1, 1, num_img_emb).view(b, -1)
-        img_att_masks = torch.zeros((b, n, num_img_emb), dtype=torch.long, device=device).reshape(b, -1)
-
-        image_target_tokens = (
-            torch.ones((b, n, num_img_emb), dtype=torch.long, device=device) * self.pad_token_id
-        ).reshape(b, -1)
-        image_loss_mask = torch.zeros((b, n, num_img_emb), dtype=torch.long, device=device).reshape(b, -1)
-
-        embedded = embedded.reshape(b, n * num_img_emb, image_embedding_dim)  # Shape: (B, N*P, D)
-
-        embs = torch.cat([embedded, tokens_embs], dim=1).to(device)
-        pad_masks = torch.cat([img_pad_masks, pad_mask.to(device)], dim=1)
-        att_masks = torch.cat([img_att_masks, ar_mask.to(device)], dim=1)
-        loss_masks = torch.cat([image_loss_mask, loss_mask.to(device)], dim=1)
-        targets = torch.cat([image_target_tokens, tokens.to(device)], dim=1)
-        token_type_ids = torch.cat([img_att_masks, token_type_ids.to(device)], dim=1)
-
-        # Shift pad tokens to the left (.generate()) or right (.train())
-        embs, att_masks, pad_masks, loss_masks, targets, token_type_ids = self.shift_padding_side(
-            embs, att_masks, pad_masks, loss_masks, targets, token_type_ids, padding_side=padding_side
-        )
-
-        targets = torch.where(targets == self.pad_token_id, self.ignore_index, targets)
-        return embs, pad_masks, att_masks, targets, loss_masks, token_type_ids
-
-
-def resize_with_pad(img, width, height, pad_value=0, interpolate_like_pi=True):
-    # assume no-op when width height fits already
-    if img.ndim != 4:
-        raise ValueError(f"(b,c,h,w) expected, but {img.shape}")
-
-    cur_height, cur_width = img.shape[2:]
-
-    ratio = max(cur_width / width, cur_height / height)
-    resized_height = int(cur_height / ratio)
-    resized_width = int(cur_width / ratio)
-
-    if interpolate_like_pi:
-        img = (img * 255.0).to(dtype=torch.uint8)
-        img = img.permute(0, 2, 3, 1)
-        original_device = img.device
-        img = img.to(device="cpu").numpy()
-        imgs = []
-        for sub_img in img:
-            sub_img = Image.fromarray(sub_img)
-            resized_img = sub_img.resize((resized_width, resized_height), resample=2)
-            resized_img = torch.from_numpy(np.array(resized_img))
-            imgs.append(resized_img)
-        img = torch.stack(imgs, dim=0)
-        img = img.permute(0, 3, 1, 2)
-        resized_img = img.to(device=original_device, dtype=torch.float32) / 255.0
-    else:
-        resized_img = F.interpolate(
-            img, size=(resized_height, resized_width), mode="bilinear", align_corners=False
-        )
-
-    pad_height = max(0, int(height - resized_height))
-    pad_width = max(0, int(width - resized_width))
-
-    # pad on left and top of image
-    padded_img = F.pad(resized_img, (pad_width, 0, pad_height, 0), value=pad_value)
-    return padded_img

lerobot/common/policies/tdmpc/modeling_tdmpc.py

@@ -35,7 +35,7 @@ import torch.nn as nn
 import torch.nn.functional as F  # noqa: N812
 from torch import Tensor
 
-from lerobot.common.constants import OBS_ENV, OBS_ROBOT
+from lerobot.common.constants import OBS_ENV_STATE, OBS_STATE
 from lerobot.common.policies.normalize import Normalize, Unnormalize
 from lerobot.common.policies.pretrained import PreTrainedPolicy
 from lerobot.common.policies.tdmpc.configuration_tdmpc import TDMPCConfig
@@ -122,7 +122,7 @@ class TDMPCPolicy(PreTrainedPolicy):
 
         # When the action queue is depleted, populate it again by querying the policy.
         if len(self._queues["action"]) == 0:
-            batch = {key: torch.stack(list(self._queues[key]), dim=1) for key in batch if key in self._queues}
+            batch = {key: torch.stack(list(self._queues[key]), dim=1) for key in batch}
 
             # Remove the time dimensions as it is not handled yet.
             for key in batch:
@@ -753,9 +753,9 @@ class TDMPCObservationEncoder(nn.Module):
                 )
             )
         if self.config.env_state_feature:
-            feat.append(self.env_state_enc_layers(obs_dict[OBS_ENV]))
+            feat.append(self.env_state_enc_layers(obs_dict[OBS_ENV_STATE]))
         if self.config.robot_state_feature:
-            feat.append(self.state_enc_layers(obs_dict[OBS_ROBOT]))
+            feat.append(self.state_enc_layers(obs_dict[OBS_STATE]))
         return torch.stack(feat, dim=0).mean(0)
 
 

lerobot/common/robot_devices/cameras/utils.py

@@ -1,67 +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 typing import Protocol
-
-import numpy as np
-
-from lerobot.common.robot_devices.cameras.configs import (
-    CameraConfig,
-    IntelRealSenseCameraConfig,
-    OpenCVCameraConfig,
-)
-
-
-# Defines a camera type
-class Camera(Protocol):
-    def connect(self): ...
-    def read(self, temporary_color: str | None = None) -> np.ndarray: ...
-    def async_read(self) -> np.ndarray: ...
-    def disconnect(self): ...
-
-
-def make_cameras_from_configs(camera_configs: dict[str, CameraConfig]) -> list[Camera]:
-    cameras = {}
-
-    for key, cfg in camera_configs.items():
-        if cfg.type == "opencv":
-            from lerobot.common.robot_devices.cameras.opencv import OpenCVCamera
-
-            cameras[key] = OpenCVCamera(cfg)
-
-        elif cfg.type == "intelrealsense":
-            from lerobot.common.robot_devices.cameras.intelrealsense import IntelRealSenseCamera
-
-            cameras[key] = IntelRealSenseCamera(cfg)
-        else:
-            raise ValueError(f"The camera type '{cfg.type}' is not valid.")
-
-    return cameras
-
-
-def make_camera(camera_type, **kwargs) -> Camera:
-    if camera_type == "opencv":
-        from lerobot.common.robot_devices.cameras.opencv import OpenCVCamera
-
-        config = OpenCVCameraConfig(**kwargs)
-        return OpenCVCamera(config)
-
-    elif camera_type == "intelrealsense":
-        from lerobot.common.robot_devices.cameras.intelrealsense import IntelRealSenseCamera
-
-        config = IntelRealSenseCameraConfig(**kwargs)
-        return IntelRealSenseCamera(config)
-
-    else:
-        raise ValueError(f"The camera type '{camera_type}' is not valid.")

lerobot/common/robot_devices/robots/configs.py

@@ -1,613 +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 abc
-from dataclasses import dataclass, field
-from typing import Sequence
-
-import draccus
-
-from lerobot.common.robot_devices.cameras.configs import (
-    CameraConfig,
-    IntelRealSenseCameraConfig,
-    OpenCVCameraConfig,
-)
-from lerobot.common.robot_devices.motors.configs import (
-    DynamixelMotorsBusConfig,
-    FeetechMotorsBusConfig,
-    MotorsBusConfig,
-)
-
-
-@dataclass
-class RobotConfig(draccus.ChoiceRegistry, abc.ABC):
-    @property
-    def type(self) -> str:
-        return self.get_choice_name(self.__class__)
-
-
-# TODO(rcadene, aliberts): remove ManipulatorRobotConfig abstraction
-@dataclass
-class ManipulatorRobotConfig(RobotConfig):
-    leader_arms: dict[str, MotorsBusConfig] = field(default_factory=lambda: {})
-    follower_arms: dict[str, MotorsBusConfig] = field(default_factory=lambda: {})
-    cameras: dict[str, CameraConfig] = field(default_factory=lambda: {})
-
-    # Optionally limit 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 (assumes all follower arms have the same number of
-    # motors).
-    max_relative_target: list[float] | float | None = None
-
-    # Optionally set the leader arm in torque mode with the gripper motor set to this angle. This makes it
-    # possible to squeeze the gripper and have it spring back to an open position on its own. If None, the
-    # gripper is not put in torque mode.
-    gripper_open_degree: float | None = None
-
-    mock: bool = False
-
-    def __post_init__(self):
-        if self.mock:
-            for arm in self.leader_arms.values():
-                if not arm.mock:
-                    arm.mock = True
-            for arm in self.follower_arms.values():
-                if not arm.mock:
-                    arm.mock = True
-            for cam in self.cameras.values():
-                if not cam.mock:
-                    cam.mock = True
-
-        if self.max_relative_target is not None and isinstance(self.max_relative_target, Sequence):
-            for name in self.follower_arms:
-                if len(self.follower_arms[name].motors) != len(self.max_relative_target):
-                    raise ValueError(
-                        f"len(max_relative_target)={len(self.max_relative_target)} but the follower arm with name {name} has "
-                        f"{len(self.follower_arms[name].motors)} motors. Please make sure that the "
-                        f"`max_relative_target` list has as many parameters as there are motors per arm. "
-                        "Note: This feature does not yet work with robots where different follower arms have "
-                        "different numbers of motors."
-                    )
-
-
-@RobotConfig.register_subclass("aloha")
-@dataclass
-class AlohaRobotConfig(ManipulatorRobotConfig):
-    # Specific to Aloha, LeRobot comes with default calibration files. Assuming the motors have been
-    # properly assembled, no manual calibration step is expected. If you need to run manual calibration,
-    # simply update this path to ".cache/calibration/aloha"
-    calibration_dir: str = ".cache/calibration/aloha_default"
-
-    # /!\ 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.
-    # 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
-
-    leader_arms: dict[str, MotorsBusConfig] = field(
-        default_factory=lambda: {
-            "left": DynamixelMotorsBusConfig(
-                # window_x
-                port="/dev/ttyDXL_leader_left",
-                motors={
-                    # name: (index, model)
-                    "waist": [1, "xm430-w350"],
-                    "shoulder": [2, "xm430-w350"],
-                    "shoulder_shadow": [3, "xm430-w350"],
-                    "elbow": [4, "xm430-w350"],
-                    "elbow_shadow": [5, "xm430-w350"],
-                    "forearm_roll": [6, "xm430-w350"],
-                    "wrist_angle": [7, "xm430-w350"],
-                    "wrist_rotate": [8, "xl430-w250"],
-                    "gripper": [9, "xc430-w150"],
-                },
-            ),
-            "right": DynamixelMotorsBusConfig(
-                # window_x
-                port="/dev/ttyDXL_leader_right",
-                motors={
-                    # name: (index, model)
-                    "waist": [1, "xm430-w350"],
-                    "shoulder": [2, "xm430-w350"],
-                    "shoulder_shadow": [3, "xm430-w350"],
-                    "elbow": [4, "xm430-w350"],
-                    "elbow_shadow": [5, "xm430-w350"],
-                    "forearm_roll": [6, "xm430-w350"],
-                    "wrist_angle": [7, "xm430-w350"],
-                    "wrist_rotate": [8, "xl430-w250"],
-                    "gripper": [9, "xc430-w150"],
-                },
-            ),
-        }
-    )
-
-    follower_arms: dict[str, MotorsBusConfig] = field(
-        default_factory=lambda: {
-            "left": DynamixelMotorsBusConfig(
-                port="/dev/ttyDXL_follower_left",
-                motors={
-                    # name: (index, model)
-                    "waist": [1, "xm540-w270"],
-                    "shoulder": [2, "xm540-w270"],
-                    "shoulder_shadow": [3, "xm540-w270"],
-                    "elbow": [4, "xm540-w270"],
-                    "elbow_shadow": [5, "xm540-w270"],
-                    "forearm_roll": [6, "xm540-w270"],
-                    "wrist_angle": [7, "xm540-w270"],
-                    "wrist_rotate": [8, "xm430-w350"],
-                    "gripper": [9, "xm430-w350"],
-                },
-            ),
-            "right": DynamixelMotorsBusConfig(
-                port="/dev/ttyDXL_follower_right",
-                motors={
-                    # name: (index, model)
-                    "waist": [1, "xm540-w270"],
-                    "shoulder": [2, "xm540-w270"],
-                    "shoulder_shadow": [3, "xm540-w270"],
-                    "elbow": [4, "xm540-w270"],
-                    "elbow_shadow": [5, "xm540-w270"],
-                    "forearm_roll": [6, "xm540-w270"],
-                    "wrist_angle": [7, "xm540-w270"],
-                    "wrist_rotate": [8, "xm430-w350"],
-                    "gripper": [9, "xm430-w350"],
-                },
-            ),
-        }
-    )
-
-    # Troubleshooting: If one of your IntelRealSense cameras freeze during
-    # data recording due to bandwidth limit, you might need to plug the camera
-    # on another USB hub or PCIe card.
-    cameras: dict[str, CameraConfig] = field(
-        default_factory=lambda: {
-            "cam_high": IntelRealSenseCameraConfig(
-                serial_number=128422271347,
-                fps=30,
-                width=640,
-                height=480,
-            ),
-            "cam_low": IntelRealSenseCameraConfig(
-                serial_number=130322270656,
-                fps=30,
-                width=640,
-                height=480,
-            ),
-            "cam_left_wrist": IntelRealSenseCameraConfig(
-                serial_number=218622272670,
-                fps=30,
-                width=640,
-                height=480,
-            ),
-            "cam_right_wrist": IntelRealSenseCameraConfig(
-                serial_number=130322272300,
-                fps=30,
-                width=640,
-                height=480,
-            ),
-        }
-    )
-
-    mock: bool = False
-
-
-@RobotConfig.register_subclass("koch")
-@dataclass
-class KochRobotConfig(ManipulatorRobotConfig):
-    calibration_dir: str = ".cache/calibration/koch"
-    # `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
-
-    leader_arms: dict[str, MotorsBusConfig] = field(
-        default_factory=lambda: {
-            "main": DynamixelMotorsBusConfig(
-                port="/dev/tty.usbmodem585A0085511",
-                motors={
-                    # name: (index, model)
-                    "shoulder_pan": [1, "xl330-m077"],
-                    "shoulder_lift": [2, "xl330-m077"],
-                    "elbow_flex": [3, "xl330-m077"],
-                    "wrist_flex": [4, "xl330-m077"],
-                    "wrist_roll": [5, "xl330-m077"],
-                    "gripper": [6, "xl330-m077"],
-                },
-            ),
-        }
-    )
-
-    follower_arms: dict[str, MotorsBusConfig] = field(
-        default_factory=lambda: {
-            "main": DynamixelMotorsBusConfig(
-                port="/dev/tty.usbmodem585A0076891",
-                motors={
-                    # name: (index, model)
-                    "shoulder_pan": [1, "xl430-w250"],
-                    "shoulder_lift": [2, "xl430-w250"],
-                    "elbow_flex": [3, "xl330-m288"],
-                    "wrist_flex": [4, "xl330-m288"],
-                    "wrist_roll": [5, "xl330-m288"],
-                    "gripper": [6, "xl330-m288"],
-                },
-            ),
-        }
-    )
-
-    cameras: dict[str, CameraConfig] = field(
-        default_factory=lambda: {
-            "laptop": OpenCVCameraConfig(
-                camera_index=0,
-                fps=30,
-                width=640,
-                height=480,
-            ),
-            "phone": OpenCVCameraConfig(
-                camera_index=1,
-                fps=30,
-                width=640,
-                height=480,
-            ),
-        }
-    )
-
-    # ~ Koch specific settings ~
-    # Sets the leader arm in torque mode with the gripper motor set to this angle. This makes it possible
-    # to squeeze the gripper and have it spring back to an open position on its own.
-    gripper_open_degree: float = 35.156
-
-    mock: bool = False
-
-
-@RobotConfig.register_subclass("koch_bimanual")
-@dataclass
-class KochBimanualRobotConfig(ManipulatorRobotConfig):
-    calibration_dir: str = ".cache/calibration/koch_bimanual"
-    # `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
-
-    leader_arms: dict[str, MotorsBusConfig] = field(
-        default_factory=lambda: {
-            "left": DynamixelMotorsBusConfig(
-                port="/dev/tty.usbmodem585A0085511",
-                motors={
-                    # name: (index, model)
-                    "shoulder_pan": [1, "xl330-m077"],
-                    "shoulder_lift": [2, "xl330-m077"],
-                    "elbow_flex": [3, "xl330-m077"],
-                    "wrist_flex": [4, "xl330-m077"],
-                    "wrist_roll": [5, "xl330-m077"],
-                    "gripper": [6, "xl330-m077"],
-                },
-            ),
-            "right": DynamixelMotorsBusConfig(
-                port="/dev/tty.usbmodem575E0031751",
-                motors={
-                    # name: (index, model)
-                    "shoulder_pan": [1, "xl330-m077"],
-                    "shoulder_lift": [2, "xl330-m077"],
-                    "elbow_flex": [3, "xl330-m077"],
-                    "wrist_flex": [4, "xl330-m077"],
-                    "wrist_roll": [5, "xl330-m077"],
-                    "gripper": [6, "xl330-m077"],
-                },
-            ),
-        }
-    )
-
-    follower_arms: dict[str, MotorsBusConfig] = field(
-        default_factory=lambda: {
-            "left": DynamixelMotorsBusConfig(
-                port="/dev/tty.usbmodem585A0076891",
-                motors={
-                    # name: (index, model)
-                    "shoulder_pan": [1, "xl430-w250"],
-                    "shoulder_lift": [2, "xl430-w250"],
-                    "elbow_flex": [3, "xl330-m288"],
-                    "wrist_flex": [4, "xl330-m288"],
-                    "wrist_roll": [5, "xl330-m288"],
-                    "gripper": [6, "xl330-m288"],
-                },
-            ),
-            "right": DynamixelMotorsBusConfig(
-                port="/dev/tty.usbmodem575E0032081",
-                motors={
-                    # name: (index, model)
-                    "shoulder_pan": [1, "xl430-w250"],
-                    "shoulder_lift": [2, "xl430-w250"],
-                    "elbow_flex": [3, "xl330-m288"],
-                    "wrist_flex": [4, "xl330-m288"],
-                    "wrist_roll": [5, "xl330-m288"],
-                    "gripper": [6, "xl330-m288"],
-                },
-            ),
-        }
-    )
-
-    cameras: dict[str, CameraConfig] = field(
-        default_factory=lambda: {
-            "laptop": OpenCVCameraConfig(
-                camera_index=0,
-                fps=30,
-                width=640,
-                height=480,
-            ),
-            "phone": OpenCVCameraConfig(
-                camera_index=1,
-                fps=30,
-                width=640,
-                height=480,
-            ),
-        }
-    )
-
-    # ~ Koch specific settings ~
-    # Sets the leader arm in torque mode with the gripper motor set to this angle. This makes it possible
-    # to squeeze the gripper and have it spring back to an open position on its own.
-    gripper_open_degree: float = 35.156
-
-    mock: bool = False
-
-
-@RobotConfig.register_subclass("moss")
-@dataclass
-class MossRobotConfig(ManipulatorRobotConfig):
-    calibration_dir: str = ".cache/calibration/moss"
-    # `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
-
-    leader_arms: dict[str, MotorsBusConfig] = field(
-        default_factory=lambda: {
-            "main": FeetechMotorsBusConfig(
-                port="/dev/tty.usbmodem58760431091",
-                motors={
-                    # name: (index, model)
-                    "shoulder_pan": [1, "sts3215"],
-                    "shoulder_lift": [2, "sts3215"],
-                    "elbow_flex": [3, "sts3215"],
-                    "wrist_flex": [4, "sts3215"],
-                    "wrist_roll": [5, "sts3215"],
-                    "gripper": [6, "sts3215"],
-                },
-            ),
-        }
-    )
-
-    follower_arms: dict[str, MotorsBusConfig] = field(
-        default_factory=lambda: {
-            "main": FeetechMotorsBusConfig(
-                port="/dev/tty.usbmodem585A0076891",
-                motors={
-                    # name: (index, model)
-                    "shoulder_pan": [1, "sts3215"],
-                    "shoulder_lift": [2, "sts3215"],
-                    "elbow_flex": [3, "sts3215"],
-                    "wrist_flex": [4, "sts3215"],
-                    "wrist_roll": [5, "sts3215"],
-                    "gripper": [6, "sts3215"],
-                },
-            ),
-        }
-    )
-
-    cameras: dict[str, CameraConfig] = field(
-        default_factory=lambda: {
-            "laptop": OpenCVCameraConfig(
-                camera_index=0,
-                fps=30,
-                width=640,
-                height=480,
-            ),
-            "phone": OpenCVCameraConfig(
-                camera_index=1,
-                fps=30,
-                width=640,
-                height=480,
-            ),
-        }
-    )
-
-    mock: bool = False
-
-
-@RobotConfig.register_subclass("so100")
-@dataclass
-class So100RobotConfig(ManipulatorRobotConfig):
-    calibration_dir: str = ".cache/calibration/so100"
-    # `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
-
-    leader_arms: dict[str, MotorsBusConfig] = field(
-        default_factory=lambda: {
-            "main": FeetechMotorsBusConfig(
-                port="/dev/tty.usbmodem58760431091",
-                motors={
-                    # name: (index, model)
-                    "shoulder_pan": [1, "sts3215"],
-                    "shoulder_lift": [2, "sts3215"],
-                    "elbow_flex": [3, "sts3215"],
-                    "wrist_flex": [4, "sts3215"],
-                    "wrist_roll": [5, "sts3215"],
-                    "gripper": [6, "sts3215"],
-                },
-            ),
-        }
-    )
-
-    follower_arms: dict[str, MotorsBusConfig] = field(
-        default_factory=lambda: {
-            "main": FeetechMotorsBusConfig(
-                port="/dev/tty.usbmodem585A0076891",
-                motors={
-                    # name: (index, model)
-                    "shoulder_pan": [1, "sts3215"],
-                    "shoulder_lift": [2, "sts3215"],
-                    "elbow_flex": [3, "sts3215"],
-                    "wrist_flex": [4, "sts3215"],
-                    "wrist_roll": [5, "sts3215"],
-                    "gripper": [6, "sts3215"],
-                },
-            ),
-        }
-    )
-
-    cameras: dict[str, CameraConfig] = field(
-        default_factory=lambda: {
-            "laptop": OpenCVCameraConfig(
-                camera_index=0,
-                fps=30,
-                width=640,
-                height=480,
-            ),
-            "phone": OpenCVCameraConfig(
-                camera_index=1,
-                fps=30,
-                width=640,
-                height=480,
-            ),
-        }
-    )
-
-    mock: bool = False
-
-
-@RobotConfig.register_subclass("stretch")
-@dataclass
-class StretchRobotConfig(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: dict[str, CameraConfig] = field(
-        default_factory=lambda: {
-            "navigation": OpenCVCameraConfig(
-                camera_index="/dev/hello-nav-head-camera",
-                fps=10,
-                width=1280,
-                height=720,
-                rotation=-90,
-            ),
-            "head": IntelRealSenseCameraConfig(
-                name="Intel RealSense D435I",
-                fps=30,
-                width=640,
-                height=480,
-                rotation=90,
-            ),
-            "wrist": IntelRealSenseCameraConfig(
-                name="Intel RealSense D405",
-                fps=30,
-                width=640,
-                height=480,
-            ),
-        }
-    )
-
-    mock: bool = False
-
-
-@RobotConfig.register_subclass("lekiwi")
-@dataclass
-class LeKiwiRobotConfig(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
-
-    # Network Configuration
-    ip: str = "192.168.0.193"
-    port: int = 5555
-    video_port: int = 5556
-
-    cameras: dict[str, CameraConfig] = field(
-        default_factory=lambda: {
-            "front": OpenCVCameraConfig(
-                camera_index="/dev/video0", fps=30, width=640, height=480, rotation=90
-            ),
-            "wrist": OpenCVCameraConfig(
-                camera_index="/dev/video2", fps=30, width=640, height=480, rotation=180
-            ),
-        }
-    )
-
-    calibration_dir: str = ".cache/calibration/lekiwi"
-
-    leader_arms: dict[str, MotorsBusConfig] = field(
-        default_factory=lambda: {
-            "main": FeetechMotorsBusConfig(
-                port="/dev/tty.usbmodem585A0077581",
-                motors={
-                    # name: (index, model)
-                    "shoulder_pan": [1, "sts3215"],
-                    "shoulder_lift": [2, "sts3215"],
-                    "elbow_flex": [3, "sts3215"],
-                    "wrist_flex": [4, "sts3215"],
-                    "wrist_roll": [5, "sts3215"],
-                    "gripper": [6, "sts3215"],
-                },
-            ),
-        }
-    )
-
-    follower_arms: dict[str, MotorsBusConfig] = field(
-        default_factory=lambda: {
-            "main": FeetechMotorsBusConfig(
-                port="/dev/ttyACM0",
-                motors={
-                    # name: (index, model)
-                    "shoulder_pan": [1, "sts3215"],
-                    "shoulder_lift": [2, "sts3215"],
-                    "elbow_flex": [3, "sts3215"],
-                    "wrist_flex": [4, "sts3215"],
-                    "wrist_roll": [5, "sts3215"],
-                    "gripper": [6, "sts3215"],
-                    "left_wheel": (7, "sts3215"),
-                    "back_wheel": (8, "sts3215"),
-                    "right_wheel": (9, "sts3215"),
-                },
-            ),
-        }
-    )
-
-    teleop_keys: dict[str, str] = field(
-        default_factory=lambda: {
-            # Movement
-            "forward": "w",
-            "backward": "s",
-            "left": "a",
-            "right": "d",
-            "rotate_left": "z",
-            "rotate_right": "x",
-            # Speed control
-            "speed_up": "r",
-            "speed_down": "f",
-            # quit teleop
-            "quit": "q",
-        }
-    )
-
-    mock: bool = False