nit

- Updated standard deviation parameterization in SACConfig to 'softplus' with defined min and max values for improved stability.
- Modified action sampling in SACPolicy to use reparameterized sampling, ensuring better gradient flow and log probability calculations.
- Cleaned up log probability calculations in TanhMultivariateNormalDiag for clarity and efficiency.
- Increased evaluation frequency in YAML configuration to 50000 for more efficient training cycles.

These changes aim to enhance the robustness and performance of the SAC implementation during training and inference.

.github/PULL_REQUEST_TEMPLATE.md

@@ -21,7 +21,7 @@ Provide a simple way for the reviewer to try out your changes.
 
 Examples:
 ```bash
-DATA_DIR=tests/data pytest -sx tests/test_stuff.py::test_something
+pytest -sx tests/test_stuff.py::test_something
 ```
 ```bash
 python lerobot/scripts/train.py --some.option=true

.github/workflows/nightly-tests.yml

@@ -7,10 +7,8 @@ on:
   schedule:
     - cron: "0 2 * * *"
 
-env:
-  DATA_DIR: tests/data
+# env:
   # SLACK_API_TOKEN: ${{ secrets.SLACK_API_TOKEN }}
-
 jobs:
   run_all_tests_cpu:
     name: CPU
@@ -30,13 +28,9 @@ jobs:
         working-directory: /lerobot
     steps:
       - name: Tests
-        env:
-          DATA_DIR: tests/data
         run: pytest -v --cov=./lerobot --disable-warnings tests
 
       - name: Tests end-to-end
-        env:
-          DATA_DIR: tests/data
         run: make test-end-to-end
 
 

.github/workflows/quality.yml

@@ -50,7 +50,7 @@ jobs:
         uses: actions/checkout@v3
 
       - name: Install poetry
-        run: pipx install poetry
+        run: pipx install "poetry<2.0.0"
 
       - name: Poetry check
         run: poetry check
@@ -64,7 +64,7 @@ jobs:
         uses: actions/checkout@v3
 
       - name: Install poetry
-        run: pipx install poetry
+        run: pipx install "poetry<2.0.0"
 
       - name: Install poetry-relax
         run: poetry self add poetry-relax

.github/workflows/test.yml

@@ -29,7 +29,6 @@ jobs:
     name: Pytest
     runs-on: ubuntu-latest
     env:
-      DATA_DIR: tests/data
       MUJOCO_GL: egl
     steps:
       - uses: actions/checkout@v4
@@ -70,7 +69,6 @@ jobs:
     name: Pytest (minimal install)
     runs-on: ubuntu-latest
     env:
-      DATA_DIR: tests/data
       MUJOCO_GL: egl
     steps:
       - uses: actions/checkout@v4
@@ -103,40 +101,39 @@ jobs:
             -W ignore::UserWarning:gymnasium.utils.env_checker:247 \
             && rm -rf tests/outputs outputs
 
+  # TODO(aliberts, rcadene): redesign after v2 migration / removing hydra
+  # end-to-end:
+  #   name: End-to-end
+  #   runs-on: ubuntu-latest
+  #   env:
+  #     MUJOCO_GL: egl
+  #   steps:
+  #     - uses: actions/checkout@v4
+  #       with:
+  #         lfs: true  # Ensure LFS files are pulled
 
-  end-to-end:
-    name: End-to-end
-    runs-on: ubuntu-latest
-    env:
-      DATA_DIR: tests/data
-      MUJOCO_GL: egl
-    steps:
-      - uses: actions/checkout@v4
-        with:
-          lfs: true  # Ensure LFS files are pulled
+  #     - name: Install apt dependencies
+  #     # portaudio19-dev is needed to install pyaudio
+  #       run: |
+  #         sudo apt-get update && \
+  #         sudo apt-get install -y libegl1-mesa-dev portaudio19-dev
 
-      - name: Install apt dependencies
-      # portaudio19-dev is needed to install pyaudio
-        run: |
-          sudo apt-get update && \
-          sudo apt-get install -y libegl1-mesa-dev portaudio19-dev
+  #     - name: Install poetry
+  #       run: |
+  #         pipx install poetry && poetry config virtualenvs.in-project true
+  #         echo "${{ github.workspace }}/.venv/bin" >> $GITHUB_PATH
 
-      - name: Install poetry
-        run: |
-          pipx install poetry && poetry config virtualenvs.in-project true
-          echo "${{ github.workspace }}/.venv/bin" >> $GITHUB_PATH
+  #     - name: Set up Python 3.10
+  #       uses: actions/setup-python@v5
+  #       with:
+  #         python-version: "3.10"
+  #         cache: "poetry"
 
-      - name: Set up Python 3.10
-        uses: actions/setup-python@v5
-        with:
-          python-version: "3.10"
-          cache: "poetry"
+  #     - name: Install poetry dependencies
+  #       run: |
+  #         poetry install --all-extras
 
-      - name: Install poetry dependencies
-        run: |
-          poetry install --all-extras
-
-      - name: Test end-to-end
-        run: |
-          make test-end-to-end \
-            && rm -rf outputs
+  #     - name: Test end-to-end
+  #       run: |
+  #         make test-end-to-end \
+  #           && rm -rf outputs

.pre-commit-config.yaml

@@ -3,7 +3,7 @@ default_language_version:
     python: python3.10
 repos:
   - repo: https://github.com/pre-commit/pre-commit-hooks
-    rev: v4.6.0
+    rev: v5.0.0
     hooks:
       - id: check-added-large-files
       - id: debug-statements
@@ -14,11 +14,11 @@ repos:
       - id: end-of-file-fixer
       - id: trailing-whitespace
   - repo: https://github.com/asottile/pyupgrade
-    rev: v3.16.0
+    rev: v3.19.0
     hooks:
     -   id: pyupgrade
   - repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.5.2
+    rev: v0.8.2
     hooks:
       - id: ruff
         args: [--fix]
@@ -32,6 +32,6 @@ repos:
           - "--check"
           - "--no-update"
   - repo: https://github.com/gitleaks/gitleaks
-    rev: v8.18.4
+    rev: v8.21.2
     hooks:
       - id: gitleaks

CONTRIBUTING.md

@@ -267,7 +267,7 @@ We use `pytest` in order to run the tests. From the root of the
 repository, here's how to run tests with `pytest` for the library:
 
 ```bash
-DATA_DIR="tests/data" python -m pytest -sv ./tests
+python -m pytest -sv ./tests
 ```
 
 

README.md

@@ -68,7 +68,7 @@
 
 ### Acknowledgment
 
-- Thanks to Tony Zaho, Zipeng Fu and colleagues for open sourcing ACT policy, ALOHA environments and datasets. Ours are adapted from [ALOHA](https://tonyzhaozh.github.io/aloha) and [Mobile ALOHA](https://mobile-aloha.github.io).
+- Thanks to Tony Zhao, Zipeng Fu and colleagues for open sourcing ACT policy, ALOHA environments and datasets. Ours are adapted from [ALOHA](https://tonyzhaozh.github.io/aloha) and [Mobile ALOHA](https://mobile-aloha.github.io).
 - Thanks to Cheng Chi, Zhenjia Xu and colleagues for open sourcing Diffusion policy, Pusht environment and datasets, as well as UMI datasets. Ours are adapted from [Diffusion Policy](https://diffusion-policy.cs.columbia.edu) and [UMI Gripper](https://umi-gripper.github.io).
 - Thanks to Nicklas Hansen, Yunhai Feng and colleagues for open sourcing TDMPC policy, Simxarm environments and datasets. Ours are adapted from [TDMPC](https://github.com/nicklashansen/tdmpc) and [FOWM](https://www.yunhaifeng.com/FOWM).
 - Thanks to Antonio Loquercio and Ashish Kumar for their early support.
@@ -153,10 +153,12 @@ python lerobot/scripts/visualize_dataset.py \
     --episode-index 0
 ```
 
-or from a dataset in a local folder with the root `DATA_DIR` environment variable (in the following case the dataset will be searched for in `./my_local_data_dir/lerobot/pusht`)
+or from a dataset in a local folder with the `root` option and the `--local-files-only` (in the following case the dataset will be searched for in `./my_local_data_dir/lerobot/pusht`)
 ```bash
-DATA_DIR='./my_local_data_dir' python lerobot/scripts/visualize_dataset.py \
+python lerobot/scripts/visualize_dataset.py \
     --repo-id lerobot/pusht \
+    --root ./my_local_data_dir \
+    --local-files-only 1 \
     --episode-index 0
 ```
 
@@ -208,12 +210,10 @@ dataset attributes:
 
 A `LeRobotDataset` is serialised using several widespread file formats for each of its parts, namely:
 - hf_dataset stored using Hugging Face datasets library serialization to parquet
-- videos are stored in mp4 format to save space or png files
-- episode_data_index saved using `safetensor` tensor serialization format
-- stats saved using `safetensor` tensor serialization format
-- info are saved using JSON
+- videos are stored in mp4 format to save space
+- metadata are stored in plain json/jsonl files
 
-Dataset can be uploaded/downloaded from the HuggingFace hub seamlessly. To work on a local dataset, you can set the `DATA_DIR` environment variable to your root dataset folder as illustrated in the above section on dataset visualization.
+Dataset can be uploaded/downloaded from the HuggingFace hub seamlessly. To work on a local dataset, you can use the `local_files_only` argument and specify its location with the `root` argument if it's not in the default `~/.cache/huggingface/lerobot` location.
 
 ### Evaluate a pretrained policy
 

benchmarks/video/README.md

@@ -21,7 +21,7 @@ How to decode videos?
 
 ## Variables
 **Image content & size**
-We don't expect the same optimal settings for a dataset of images from a simulation, or from real-world in an appartment, or in a factory, or outdoor, or with lots of moving objects in the scene, etc. Similarly, loading times might not vary linearly with the image size (resolution).
+We don't expect the same optimal settings for a dataset of images from a simulation, or from real-world in an apartment, or in a factory, or outdoor, or with lots of moving objects in the scene, etc. Similarly, loading times might not vary linearly with the image size (resolution).
 For these reasons, we run this benchmark on four representative datasets:
 - `lerobot/pusht_image`: (96 x 96 pixels) simulation with simple geometric shapes, fixed camera.
 - `aliberts/aloha_mobile_shrimp_image`: (480 x 640 pixels) real-world indoor, moving camera.
@@ -63,7 +63,7 @@ This of course is affected by the `-g` parameter during encoding, which specifie
 
 Note that this differs significantly from a typical use case like watching a movie, in which every frame is loaded sequentially from the beginning to the end and it's acceptable to have big values for `-g`.
 
-Additionally, because some policies might request single timestamps that are a few frames appart, we also have the following scenario:
+Additionally, because some policies might request single timestamps that are a few frames apart, we also have the following scenario:
 - `2_frames_4_space`: 2 frames with 4 consecutive frames of spacing in between (e.g `[t, t + 5 / fps]`),
 
 However, due to how video decoding is implemented with `pyav`, we don't have access to an accurate seek so in practice this scenario is essentially the same as `6_frames` since all 6 frames between `t` and `t + 5 / fps` will be decoded.
@@ -85,8 +85,8 @@ However, due to how video decoding is implemented with `pyav`, we don't have acc
 **Average Structural Similarity Index Measure (higher is better)**
 `avg_ssim` evaluates the perceived quality of images by comparing luminance, contrast, and structure. SSIM values range from -1 to 1, where 1 indicates perfect similarity.
 
-One aspect that can't be measured here with those metrics is the compatibility of the encoding accross platforms, in particular on web browser, for visualization purposes.
-h264, h265 and AV1 are all commonly used codecs and should not be pose an issue. However, the chroma subsampling (`pix_fmt`) format might affect compatibility:
+One aspect that can't be measured here with those metrics is the compatibility of the encoding across platforms, in particular on web browser, for visualization purposes.
+h264, h265 and AV1 are all commonly used codecs and should not pose an issue. However, the chroma subsampling (`pix_fmt`) format might affect compatibility:
 - `yuv420p` is more widely supported across various platforms, including web browsers.
 - `yuv444p` offers higher color fidelity but might not be supported as broadly.
 
@@ -116,7 +116,7 @@ Additional encoding parameters exist that are not included in this benchmark. In
 - `-preset` which allows for selecting encoding presets. This represents a collection of options that will provide a certain encoding speed to compression ratio. By leaving this parameter unspecified, it is considered to be `medium` for libx264 and libx265 and `8` for libsvtav1.
 - `-tune` which allows to optimize the encoding for certains aspects (e.g. film quality, fast decoding, etc.).
 
-See the documentation mentioned above for more detailled info on these settings and for a more comprehensive list of other parameters.
+See the documentation mentioned above for more detailed info on these settings and for a more comprehensive list of other parameters.
 
 Similarly on the decoding side, other decoders exist but are not implemented in our current benchmark. To name a few:
 - `torchaudio`

benchmarks/video/run_video_benchmark.py

@@ -266,7 +266,7 @@ def benchmark_encoding_decoding(
         )
 
     ep_num_images = dataset.episode_data_index["to"][0].item()
-    width, height = tuple(dataset[0][dataset.camera_keys[0]].shape[-2:])
+    width, height = tuple(dataset[0][dataset.meta.camera_keys[0]].shape[-2:])
     num_pixels = width * height
     video_size_bytes = video_path.stat().st_size
     images_size_bytes = get_directory_size(imgs_dir)

examples/10_use_so100.md

@@ -1,25 +1,31 @@
-This tutorial explains how to use [SO-100](https://github.com/TheRobotStudio/SO-ARM100) with LeRobot.
+# Using the [SO-100](https://github.com/TheRobotStudio/SO-ARM100) with LeRobot
 
-## Source the parts
+
+## A. Source the parts
 
 Follow this [README](https://github.com/TheRobotStudio/SO-ARM100). It contains the bill of materials, with link to source the parts, as well as the instructions to 3D print the parts, and advices if it's your first time printing or if you don't own a 3D printer already.
 
 **Important**: Before assembling, you will first need to configure your motors. To this end, we provide a nice script, so let's first install LeRobot. After configuration, we will also guide you through assembly.
 
-## Install LeRobot
+## B. Install LeRobot
 
 On your computer:
 
 1. [Install Miniconda](https://docs.anaconda.com/miniconda/#quick-command-line-install):
 ```bash
 mkdir -p ~/miniconda3
+# Linux:
 wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O ~/miniconda3/miniconda.sh
+# Mac M-series: 
+# curl https://repo.anaconda.com/miniconda/Miniconda3-latest-MacOSX-arm64.sh -o ~/miniconda3/miniconda.sh
+# Mac Intel: 
+# curl https://repo.anaconda.com/miniconda/Miniconda3-latest-MacOSX-x86_64.sh -o ~/miniconda3/miniconda.sh
 bash ~/miniconda3/miniconda.sh -b -u -p ~/miniconda3
 rm ~/miniconda3/miniconda.sh
 ~/miniconda3/bin/conda init bash
 ```
 
-2. Restart shell or `source ~/.bashrc`
+2. Restart shell or `source ~/.bashrc` (*Mac*: `source ~/.bash_profile`) or `source ~/.zshrc` if you're using zshell
 
 3. Create and activate a fresh conda environment for lerobot
 ```bash
@@ -36,23 +42,30 @@ git clone https://github.com/huggingface/lerobot.git ~/lerobot
 cd ~/lerobot && pip install -e ".[feetech]"
 ```
 
-For Linux only (not Mac), install extra dependencies for recording datasets:
+*For Linux only (not Mac)*: install extra dependencies for recording datasets:
 ```bash
 conda install -y -c conda-forge ffmpeg
 pip uninstall -y opencv-python
 conda install -y -c conda-forge "opencv>=4.10.0"
 ```
 
-## Configure the motors
+## C. Configure the motors
 
-Follow steps 1 of the [assembly video](https://www.youtube.com/watch?v=FioA2oeFZ5I) which illustrates the use of our scripts below.
+### 1. Find the USB ports associated to each arm
 
-**Find USB ports associated to your arms**
-To find the correct ports for each arm, run the utility script twice:
+Designate one bus servo adapter and 6 motors for your leader arm, and similarly the other bus servo adapter and 6 motors for the follower arm.
+
+#### a. Run the script to find ports
+
+Follow Step 1 of the [assembly video](https://www.youtube.com/watch?v=FioA2oeFZ5I), which illustrates the use of our scripts below.
+
+To find the port for each bus servo adapter, run the utility script:
 ```bash
 python lerobot/scripts/find_motors_bus_port.py
 ```
 
+#### b. Example outputs
+
 Example output when identifying the leader arm's port (e.g., `/dev/tty.usbmodem575E0031751` on Mac, or possibly `/dev/ttyACM0` on Linux):
 ```
 Finding all available ports for the MotorBus.
@@ -64,7 +77,6 @@ Remove the usb cable from your DynamixelMotorsBus and press Enter when done.
 The port of this DynamixelMotorsBus is /dev/tty.usbmodem575E0031751
 Reconnect the usb cable.
 ```
-
 Example output when identifying the follower arm's port (e.g., `/dev/tty.usbmodem575E0032081`, or possibly `/dev/ttyACM1` on Linux):
 ```
 Finding all available ports for the MotorBus.
@@ -77,13 +89,20 @@ The port of this DynamixelMotorsBus is /dev/tty.usbmodem575E0032081
 Reconnect the usb cable.
 ```
 
-Troubleshooting: On Linux, you might need to give access to the USB ports by running:
+#### c. Troubleshooting
+On Linux, you might need to give access to the USB ports by running:
 ```bash
 sudo chmod 666 /dev/ttyACM0
 sudo chmod 666 /dev/ttyACM1
 ```
 
-**Configure your motors**
+#### d. Update YAML file
+
+Now that you have the ports, modify the *port* sections in `so100.yaml` 
+
+### 2. Configure the motors
+
+#### a. Set IDs for all 12 motors
 Plug your first motor and run this script to set its ID to 1. It will also set its present position to 2048, so expect your motor to rotate:
 ```bash
 python lerobot/scripts/configure_motor.py \
@@ -94,7 +113,7 @@ python lerobot/scripts/configure_motor.py \
   --ID 1
 ```
 
-Note: These motors are currently limitated. They can take values between 0 and 4096 only, which corresponds to a full turn. They can't turn more than that. 2048 is at the middle of this range, so we can take -2048 steps (180 degrees anticlockwise) and reach the maximum range, or take +2048 steps (180 degrees clockwise) and reach the maximum range. The configuration step also sets the homing offset to 0, so that if you misassembled the arm, you can always update the homing offset to account for a shift up to ± 2048 steps (± 180 degrees).
+*Note: These motors are currently limitated. They can take values between 0 and 4096 only, which corresponds to a full turn. They can't turn more than that. 2048 is at the middle of this range, so we can take -2048 steps (180 degrees anticlockwise) and reach the maximum range, or take +2048 steps (180 degrees clockwise) and reach the maximum range. The configuration step also sets the homing offset to 0, so that if you misassembled the arm, you can always update the homing offset to account for a shift up to ± 2048 steps (± 180 degrees).*
 
 Then unplug your motor and plug the second motor and set its ID to 2.
 ```bash
@@ -108,23 +127,25 @@ python lerobot/scripts/configure_motor.py \
 
 Redo the process for all your motors until ID 6. Do the same for the 6 motors of the leader arm.
 
-**Remove the gears of the 6 leader motors**
-Follow step 2 of the [assembly video](https://www.youtube.com/watch?v=FioA2oeFZ5I). You need to remove the gear for the motors of the leader arm. As a result, you will only use the position encoding of the motor and reduce friction to more easily operate the leader arm.
 
-**Add motor horn to the motors**
-Follow step 3 of the [assembly video](https://www.youtube.com/watch?v=FioA2oeFZ5I). For SO-100, you need to align the holes on the motor horn to the motor spline to be approximately 1:30, 4:30, 7:30 and 10:30.
+#### b. Remove the gears of the 6 leader motors
+
+Follow step 2 of the [assembly video](https://youtu.be/FioA2oeFZ5I?t=248). You need to remove the gear for the motors of the leader arm. As a result, you will only use the position encoding of the motor and reduce friction to more easily operate the leader arm.
+
+#### c. Add motor horn to all 12 motors
+Follow step 3 of the [assembly video](https://youtu.be/FioA2oeFZ5I?t=569). For SO-100, you need to align the holes on the motor horn to the motor spline to be approximately 1:30, 4:30, 7:30 and 10:30.
 Try to avoid rotating the motor while doing so to keep position 2048 set during configuration. It is especially tricky for the leader motors as it is more sensible without the gears, but it's ok if it's a bit rotated.
 
-## Assemble the arms
+## D. Assemble the arms
 
-Follow step 4 of the [assembly video](https://www.youtube.com/watch?v=FioA2oeFZ5I). The first arm should take a bit more than 1 hour to assemble, but once you get use to it, you can do it under 1 hour for the second arm.
+Follow step 4 of the [assembly video](https://youtu.be/FioA2oeFZ5I?t=610). The first arm should take a bit more than 1 hour to assemble, but once you get use to it, you can do it under 1 hour for the second arm.
 
-## Calibrate
+## E. Calibrate
 
 Next, you'll need to calibrate your SO-100 robot to ensure that the leader and follower arms have the same position values when they are in the same physical position. This calibration is essential because it allows a neural network trained on one SO-100 robot to work on another.
 
-**Manual calibration of follower arm**
-/!\ Contrarily to step 6 of the [assembly video](https://www.youtube.com/watch?v=FioA2oeFZ5I) which illustrates the auto calibration, we will actually do manual calibration of follower for now.
+#### a. Manual calibration of follower arm
+/!\ Contrarily to step 6 of the [assembly video](https://youtu.be/FioA2oeFZ5I?t=724) which illustrates the auto calibration, we will actually do manual calibration of follower for now.
 
 You will need to move the follower arm to these positions sequentially:
 
@@ -139,8 +160,8 @@ python lerobot/scripts/control_robot.py calibrate \
     --robot-overrides '~cameras' --arms main_follower
 ```
 
-**Manual calibration of leader arm**
-Follow step 6 of the [assembly video](https://www.youtube.com/watch?v=FioA2oeFZ5I) which illustrates the manual calibration. You will need to move the leader arm to these positions sequentially:
+#### b. Manual calibration of leader arm
+Follow step 6 of the [assembly video](https://youtu.be/FioA2oeFZ5I?t=724) which illustrates the manual calibration. You will need to move the leader arm to these positions sequentially:
 
 | 1. Zero position | 2. Rotated position | 3. Rest position |
 |---|---|---|
@@ -153,7 +174,7 @@ python lerobot/scripts/control_robot.py calibrate \
     --robot-overrides '~cameras' --arms main_leader
 ```
 
-## Teleoperate
+## F. Teleoperate
 
 **Simple teleop**
 Then you are ready to teleoperate your robot! Run this simple script (it won't connect and display the cameras):
@@ -165,14 +186,14 @@ python lerobot/scripts/control_robot.py teleoperate \
 ```
 
 
-**Teleop with displaying cameras**
+#### a. Teleop with displaying cameras
 Follow [this guide to setup your cameras](https://github.com/huggingface/lerobot/blob/main/examples/7_get_started_with_real_robot.md#c-add-your-cameras-with-opencvcamera). Then you will be able to display the cameras on your computer while you are teleoperating by running the following code. This is useful to prepare your setup before recording your first dataset.
 ```bash
 python lerobot/scripts/control_robot.py teleoperate \
     --robot-path lerobot/configs/robot/so100.yaml
 ```
 
-## Record a dataset
+## G. Record a dataset
 
 Once you're familiar with teleoperation, you can record your first dataset with SO-100.
 
@@ -192,7 +213,6 @@ Record 2 episodes and upload your dataset to the hub:
 python lerobot/scripts/control_robot.py record \
     --robot-path lerobot/configs/robot/so100.yaml \
     --fps 30 \
-    --root data \
     --repo-id ${HF_USER}/so100_test \
     --tags so100 tutorial \
     --warmup-time-s 5 \
@@ -202,7 +222,7 @@ python lerobot/scripts/control_robot.py record \
     --push-to-hub 1
 ```
 
-## Visualize a dataset
+## H. Visualize a dataset
 
 If you uploaded your dataset to the hub with `--push-to-hub 1`, you can [visualize your dataset online](https://huggingface.co/spaces/lerobot/visualize_dataset) by copy pasting your repo id given by:
 ```bash
@@ -212,27 +232,25 @@ echo ${HF_USER}/so100_test
 If you didn't upload with `--push-to-hub 0`, you can also visualize it locally with:
 ```bash
 python lerobot/scripts/visualize_dataset_html.py \
-  --root data \
   --repo-id ${HF_USER}/so100_test
 ```
 
-## Replay an episode
+## I. Replay an episode
 
 Now try to replay the first episode on your robot:
 ```bash
-DATA_DIR=data python lerobot/scripts/control_robot.py replay \
+python lerobot/scripts/control_robot.py replay \
     --robot-path lerobot/configs/robot/so100.yaml \
     --fps 30 \
-    --root data \
     --repo-id ${HF_USER}/so100_test \
     --episode 0
 ```
 
-## Train a policy
+## J. Train a policy
 
 To train a policy to control your robot, use the [`python lerobot/scripts/train.py`](../lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
 ```bash
-DATA_DIR=data python lerobot/scripts/train.py \
+python lerobot/scripts/train.py \
   dataset_repo_id=${HF_USER}/so100_test \
   policy=act_so100_real \
   env=so100_real \
@@ -248,18 +266,16 @@ Let's explain it:
 3. We provided an environment as argument with `env=so100_real`. This loads configurations from [`lerobot/configs/env/so100_real.yaml`](../lerobot/configs/env/so100_real.yaml).
 4. We provided `device=cuda` since we are training on a Nvidia GPU, but you can also use `device=mps` if you are using a Mac with Apple silicon, or `device=cpu` otherwise.
 5. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
-6. We added `DATA_DIR=data` to access your dataset stored in your local `data` directory. If you dont provide `DATA_DIR`, your dataset will be downloaded from Hugging Face hub to your cache folder `$HOME/.cache/hugginface`. In future versions of `lerobot`, both directories will be in sync.
 
 Training should take several hours. You will find checkpoints in `outputs/train/act_so100_test/checkpoints`.
 
-## Evaluate your policy
+## K. Evaluate your policy
 
 You can use the `record` function from [`lerobot/scripts/control_robot.py`](../lerobot/scripts/control_robot.py) but with a policy checkpoint as input. For instance, run this command to record 10 evaluation episodes:
 ```bash
 python lerobot/scripts/control_robot.py record \
   --robot-path lerobot/configs/robot/so100.yaml \
   --fps 30 \
-  --root data \
   --repo-id ${HF_USER}/eval_act_so100_test \
   --tags so100 tutorial eval \
   --warmup-time-s 5 \
@@ -273,7 +289,7 @@ As you can see, it's almost the same command as previously used to record your t
 1. There is an additional `-p` argument which indicates the path to your policy checkpoint with  (e.g. `-p outputs/train/eval_so100_test/checkpoints/last/pretrained_model`). You can also use the model repository if you uploaded a model checkpoint to the hub (e.g. `-p ${HF_USER}/act_so100_test`).
 2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `--repo-id ${HF_USER}/eval_act_so100_test`).
 
-## More
+## L. More Information
 
 Follow this [previous tutorial](https://github.com/huggingface/lerobot/blob/main/examples/7_get_started_with_real_robot.md#4-train-a-policy-on-your-data) for a more in-depth tutorial on controlling real robots with LeRobot.
 

examples/11_use_moss.md

@@ -192,7 +192,6 @@ Record 2 episodes and upload your dataset to the hub:
 python lerobot/scripts/control_robot.py record \
     --robot-path lerobot/configs/robot/moss.yaml \
     --fps 30 \
-    --root data \
     --repo-id ${HF_USER}/moss_test \
     --tags moss tutorial \
     --warmup-time-s 5 \
@@ -212,7 +211,6 @@ echo ${HF_USER}/moss_test
 If you didn't upload with `--push-to-hub 0`, you can also visualize it locally with:
 ```bash
 python lerobot/scripts/visualize_dataset_html.py \
-  --root data \
   --repo-id ${HF_USER}/moss_test
 ```
 
@@ -220,10 +218,9 @@ python lerobot/scripts/visualize_dataset_html.py \
 
 Now try to replay the first episode on your robot:
 ```bash
-DATA_DIR=data python lerobot/scripts/control_robot.py replay \
+python lerobot/scripts/control_robot.py replay \
     --robot-path lerobot/configs/robot/moss.yaml \
     --fps 30 \
-    --root data \
     --repo-id ${HF_USER}/moss_test \
     --episode 0
 ```
@@ -232,7 +229,7 @@ DATA_DIR=data python lerobot/scripts/control_robot.py replay \
 
 To train a policy to control your robot, use the [`python lerobot/scripts/train.py`](../lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
 ```bash
-DATA_DIR=data python lerobot/scripts/train.py \
+python lerobot/scripts/train.py \
   dataset_repo_id=${HF_USER}/moss_test \
   policy=act_moss_real \
   env=moss_real \
@@ -248,7 +245,6 @@ Let's explain it:
 3. We provided an environment as argument with `env=moss_real`. This loads configurations from [`lerobot/configs/env/moss_real.yaml`](../lerobot/configs/env/moss_real.yaml).
 4. We provided `device=cuda` since we are training on a Nvidia GPU, but you can also use `device=mps` if you are using a Mac with Apple silicon, or `device=cpu` otherwise.
 5. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
-6. We added `DATA_DIR=data` to access your dataset stored in your local `data` directory. If you dont provide `DATA_DIR`, your dataset will be downloaded from Hugging Face hub to your cache folder `$HOME/.cache/hugginface`. In future versions of `lerobot`, both directories will be in sync.
 
 Training should take several hours. You will find checkpoints in `outputs/train/act_moss_test/checkpoints`.
 
@@ -259,7 +255,6 @@ You can use the `record` function from [`lerobot/scripts/control_robot.py`](../l
 python lerobot/scripts/control_robot.py record \
   --robot-path lerobot/configs/robot/moss.yaml \
   --fps 30 \
-  --root data \
   --repo-id ${HF_USER}/eval_act_moss_test \
   --tags moss tutorial eval \
   --warmup-time-s 5 \

examples/12_train_hilserl_classifier.md

@@ -0,0 +1,83 @@
+# Training a HIL-SERL Reward Classifier with LeRobot
+
+This tutorial provides step-by-step instructions for training a reward classifier using LeRobot.
+
+---
+
+## Training Script Overview
+
+LeRobot includes a ready-to-use training script located at [`lerobot/scripts/train_hilserl_classifier.py`](../../lerobot/scripts/train_hilserl_classifier.py). Here's an outline of its workflow:
+
+1. **Configuration Loading**
+   The script uses Hydra to load a configuration file for subsequent steps. (Details on Hydra follow below.)
+
+2. **Dataset Initialization**
+   It loads a `LeRobotDataset` containing images and rewards. To optimize performance, a weighted random sampler is used to balance class sampling.
+
+3. **Classifier Initialization**
+   A lightweight classification head is built on top of a frozen, pretrained image encoder from HuggingFace. The classifier outputs either:
+   - A single probability (binary classification), or
+   - Logits (multi-class classification).
+
+4. **Training Loop Execution**
+   The script performs:
+   - Forward and backward passes,
+   - Optimization steps,
+   - Periodic logging, evaluation, and checkpoint saving.
+
+---
+
+## Configuring with Hydra
+
+For detailed information about Hydra usage, refer to [`examples/4_train_policy_with_script.md`](../examples/4_train_policy_with_script.md). However, note that training the reward classifier differs slightly and requires a separate configuration file.
+
+### Config File Setup
+
+The default `default.yaml` cannot launch the reward classifier training directly. Instead, you need a configuration file like [`lerobot/configs/policy/hilserl_classifier.yaml`](../../lerobot/configs/policy/hilserl_classifier.yaml), with the following adjustment:
+
+Replace the `dataset_repo_id` field with the identifier for your dataset, which contains images and sparse rewards:
+
+```yaml
+# Example: lerobot/configs/policy/reward_classifier.yaml
+dataset_repo_id: "my_dataset_repo_id"
+## Typical logs and metrics
+```
+When you start the training process, you will first see your full configuration being printed in the terminal. You can check it to make sure that you config it correctly and your config is not overrided by other files. The final configuration will also be saved with the checkpoint.
+
+After that, you will see training log like this one:
+
+```
+[2024-11-29 18:26:36,999][root][INFO] -
+Epoch 5/5
+Training:  82%|██████████████████████████████████████████████████████████████████████████████▋                 | 91/111 [00:50<00:09,  2.04it/s, loss=0.2999, acc=69.99%]
+```
+
+or evaluation log like:
+
+```
+Validation: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:20<00:00,  1.37it/s]
+```
+
+### Metrics Tracking with Weights & Biases (WandB)
+
+If `wandb.enable` is set to `true`, the training and evaluation logs will also be saved in WandB. This allows you to track key metrics in real-time, including:
+
+- **Training Metrics**:
+  - `train/accuracy`
+  - `train/loss`
+  - `train/dataloading_s`
+- **Evaluation Metrics**:
+  - `eval/accuracy`
+  - `eval/loss`
+  - `eval/eval_s`
+
+#### Additional Features
+
+You can also log sample predictions during evaluation. Each logged sample will include:
+
+- The **input image**.
+- The **predicted label**.
+- The **true label**.
+- The **classifier's "confidence" (logits/probability)**.
+
+These logs can be useful for diagnosing and debugging performance issues.

examples/1_load_lerobot_dataset.py

@@ -3,78 +3,120 @@ This script demonstrates the use of `LeRobotDataset` class for handling and proc
 It illustrates how to load datasets, manipulate them, and apply transformations suitable for machine learning tasks in PyTorch.
 
 Features included in this script:
-- Loading a dataset and accessing its properties.
-- Filtering data by episode number.
-- Converting tensor data for visualization.
-- Saving video files from dataset frames.
+- Viewing a dataset's metadata and exploring its properties.
+- Loading an existing dataset from the hub or a subset of it.
+- Accessing frames by episode number.
 - Using advanced dataset features like timestamp-based frame selection.
 - Demonstrating compatibility with PyTorch DataLoader for batch processing.
 
 The script ends with examples of how to batch process data using PyTorch's DataLoader.
 """
 
-from pathlib import Path
 from pprint import pprint
 
-import imageio
 import torch
+from huggingface_hub import HfApi
 
 import lerobot
-from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
 
+# We ported a number of existing datasets ourselves, use this to see the list:
 print("List of available datasets:")
 pprint(lerobot.available_datasets)
 
-# Let's take one for this example
-repo_id = "lerobot/pusht"
+# You can also browse through the datasets created/ported by the community on the hub using the hub api:
+hub_api = HfApi()
+repo_ids = [info.id for info in hub_api.list_datasets(task_categories="robotics", tags=["LeRobot"])]
+pprint(repo_ids)
 
-# You can easily load a dataset from a Hugging Face repository
+# Or simply explore them in your web browser directly at:
+# https://huggingface.co/datasets?other=LeRobot
+
+# Let's take this one for this example
+repo_id = "lerobot/aloha_mobile_cabinet"
+# We can have a look and fetch its metadata to know more about it:
+ds_meta = LeRobotDatasetMetadata(repo_id)
+
+# By instantiating just this class, you can quickly access useful information about the content and the
+# structure of the dataset without downloading the actual data yet (only metadata files — which are
+# lightweight).
+print(f"Total number of episodes: {ds_meta.total_episodes}")
+print(f"Average number of frames per episode: {ds_meta.total_frames / ds_meta.total_episodes:.3f}")
+print(f"Frames per second used during data collection: {ds_meta.fps}")
+print(f"Robot type: {ds_meta.robot_type}")
+print(f"keys to access images from cameras: {ds_meta.camera_keys=}\n")
+
+print("Tasks:")
+print(ds_meta.tasks)
+print("Features:")
+pprint(ds_meta.features)
+
+# You can also get a short summary by simply printing the object:
+print(ds_meta)
+
+# You can then load the actual dataset from the hub.
+# Either load any subset of episodes:
+dataset = LeRobotDataset(repo_id, episodes=[0, 10, 11, 23])
+
+# And see how many frames you have:
+print(f"Selected episodes: {dataset.episodes}")
+print(f"Number of episodes selected: {dataset.num_episodes}")
+print(f"Number of frames selected: {dataset.num_frames}")
+
+# Or simply load the entire dataset:
 dataset = LeRobotDataset(repo_id)
+print(f"Number of episodes selected: {dataset.num_episodes}")
+print(f"Number of frames selected: {dataset.num_frames}")
 
-# LeRobotDataset is actually a thin wrapper around an underlying Hugging Face dataset
-# (see https://huggingface.co/docs/datasets/index for more information).
-print(dataset)
+# The previous metadata class is contained in the 'meta' attribute of the dataset:
+print(dataset.meta)
+
+# LeRobotDataset actually wraps an underlying Hugging Face dataset
+# (see https://huggingface.co/docs/datasets for more information).
 print(dataset.hf_dataset)
 
-# And provides additional utilities for robotics and compatibility with Pytorch
-print(f"\naverage number of frames per episode: {dataset.num_samples / dataset.num_episodes:.3f}")
-print(f"frames per second used during data collection: {dataset.fps=}")
-print(f"keys to access images from cameras: {dataset.camera_keys=}\n")
-
-# Access frame indexes associated to first episode
+# LeRobot datasets also subclasses PyTorch datasets so you can do everything you know and love from working
+# with the latter, like iterating through the dataset.
+# The __getitem__ iterates over the frames of the dataset. Since our datasets are also structured by
+# episodes, you can access the frame indices of any episode using the episode_data_index. Here, we access
+# frame indices associated to the first episode:
 episode_index = 0
 from_idx = dataset.episode_data_index["from"][episode_index].item()
 to_idx = dataset.episode_data_index["to"][episode_index].item()
 
-# LeRobot datasets actually subclass PyTorch datasets so you can do everything you know and love from working
-# with the latter, like iterating through the dataset. Here we grab all the image frames.
-frames = [dataset[idx]["observation.image"] for idx in range(from_idx, to_idx)]
+# Then we grab all the image frames from the first camera:
+camera_key = dataset.meta.camera_keys[0]
+frames = [dataset[idx][camera_key] for idx in range(from_idx, to_idx)]
 
-# Video frames are now float32 in range [0,1] channel first (c,h,w) to follow pytorch convention. To visualize
-# them, we convert to uint8 in range [0,255]
-frames = [(frame * 255).type(torch.uint8) for frame in frames]
-# and to channel last (h,w,c).
-frames = [frame.permute((1, 2, 0)).numpy() for frame in frames]
+# The objects returned by the dataset are all torch.Tensors
+print(type(frames[0]))
+print(frames[0].shape)
 
-# Finally, we save the frames to a mp4 video for visualization.
-Path("outputs/examples/1_load_lerobot_dataset").mkdir(parents=True, exist_ok=True)
-imageio.mimsave("outputs/examples/1_load_lerobot_dataset/episode_0.mp4", frames, fps=dataset.fps)
+# Since we're using pytorch, the shape is in pytorch, channel-first convention (c, h, w).
+# We can compare this shape with the information available for that feature
+pprint(dataset.features[camera_key])
+# In particular:
+print(dataset.features[camera_key]["shape"])
+# The shape is in (h, w, c) which is a more universal format.
 
 # For many machine learning applications we need to load the history of past observations or trajectories of
 # future actions. Our datasets can load previous and future frames for each key/modality, using timestamps
 # differences with the current loaded frame. For instance:
 delta_timestamps = {
     # loads 4 images: 1 second before current frame, 500 ms before, 200 ms before, and current frame
-    "observation.image": [-1, -0.5, -0.20, 0],
-    # loads 8 state vectors: 1.5 seconds before, 1 second before, ... 20 ms, 10 ms, and current frame
-    "observation.state": [-1.5, -1, -0.5, -0.20, -0.10, -0.02, -0.01, 0],
+    camera_key: [-1, -0.5, -0.20, 0],
+    # 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)],
 }
+# Note that in any case, these delta_timestamps values need to be multiples of (1/fps) so that added to any
+# timestamp, you still get a valid timestamp.
+
 dataset = LeRobotDataset(repo_id, delta_timestamps=delta_timestamps)
-print(f"\n{dataset[0]['observation.image'].shape=}")  # (4,c,h,w)
-print(f"{dataset[0]['observation.state'].shape=}")  # (8,c)
-print(f"{dataset[0]['action'].shape=}\n")  # (64,c)
+print(f"\n{dataset[0][camera_key].shape=}")  # (4, c, h, w)
+print(f"{dataset[0]['observation.state'].shape=}")  # (6, c)
+print(f"{dataset[0]['action'].shape=}\n")  # (64, c)
 
 # Finally, our datasets are fully compatible with PyTorch dataloaders and samplers because they are just
 # PyTorch datasets.
@@ -84,8 +126,9 @@ dataloader = torch.utils.data.DataLoader(
     batch_size=32,
     shuffle=True,
 )
+
 for batch in dataloader:
-    print(f"{batch['observation.image'].shape=}")  # (32,4,c,h,w)
-    print(f"{batch['observation.state'].shape=}")  # (32,8,c)
-    print(f"{batch['action'].shape=}")  # (32,64,c)
+    print(f"{batch[camera_key].shape=}")  # (32, 4, c, h, w)
+    print(f"{batch['observation.state'].shape=}")  # (32, 5, c)
+    print(f"{batch['action'].shape=}")  # (32, 64, c)
     break

examples/3_train_policy.py

@@ -40,7 +40,7 @@ dataset = LeRobotDataset("lerobot/pusht", delta_timestamps=delta_timestamps)
 # For this example, no arguments need to be passed because the defaults are set up for PushT.
 # If you're doing something different, you will likely need to change at least some of the defaults.
 cfg = DiffusionConfig()
-policy = DiffusionPolicy(cfg, dataset_stats=dataset.stats)
+policy = DiffusionPolicy(cfg, dataset_stats=dataset.meta.stats)
 policy.train()
 policy.to(device)
 

examples/6_add_image_transforms.py

@@ -1,7 +1,7 @@
 """
 This script demonstrates how to use torchvision's image transformation with LeRobotDataset for data
 augmentation purposes. The transformations are passed to the dataset as an argument upon creation, and
-transforms are applied to the observation images before they are returned in the dataset's __get_item__.
+transforms are applied to the observation images before they are returned in the dataset's __getitem__.
 """
 
 from pathlib import Path
@@ -10,17 +10,17 @@ from torchvision.transforms import ToPILImage, v2
 
 from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
 
-dataset_repo_id = "lerobot/aloha_static_tape"
+dataset_repo_id = "lerobot/aloha_static_screw_driver"
 
 # Create a LeRobotDataset with no transformations
-dataset = LeRobotDataset(dataset_repo_id)
+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.episode_data_index["from"][0].item()
 
 # Get the frame corresponding to the first camera
-frame = dataset[first_idx][dataset.camera_keys[0]]
+frame = dataset[first_idx][dataset.meta.camera_keys[0]]
 
 
 # Define the transformations
@@ -28,15 +28,16 @@ transforms = v2.Compose(
     [
         v2.ColorJitter(brightness=(0.5, 1.5)),
         v2.ColorJitter(contrast=(0.5, 1.5)),
+        v2.ColorJitter(hue=(-0.1, 0.1)),
         v2.RandomAdjustSharpness(sharpness_factor=2, p=1),
     ]
 )
 
 # Create another LeRobotDataset with the defined transformations
-transformed_dataset = LeRobotDataset(dataset_repo_id, image_transforms=transforms)
+transformed_dataset = LeRobotDataset(dataset_repo_id, episodes=[0], image_transforms=transforms)
 
 # Get a frame from the transformed dataset
-transformed_frame = transformed_dataset[first_idx][transformed_dataset.camera_keys[0]]
+transformed_frame = transformed_dataset[first_idx][transformed_dataset.meta.camera_keys[0]]
 
 # Create a directory to store output images
 output_dir = Path("outputs/image_transforms")

examples/7_get_started_with_real_robot.md

@@ -29,7 +29,7 @@ For a visual walkthrough of the assembly process, you can refer to [this video t
 
 ## 2. Configure motors, calibrate arms, teleoperate your Koch v1.1
 
-First, install the additional dependencies required for robots built with dynamixel motors like Koch v1.1 by running one of the following commands.
+First, install the additional dependencies required for robots built with dynamixel motors like Koch v1.1 by running one of the following commands (make sure gcc is installed).
 
 Using `pip`:
 ```bash
@@ -778,7 +778,6 @@ Now run this to record 2 episodes:
 python lerobot/scripts/control_robot.py record \
   --robot-path lerobot/configs/robot/koch.yaml \
   --fps 30 \
-  --root data \
   --repo-id ${HF_USER}/koch_test \
   --tags tutorial \
   --warmup-time-s 5 \
@@ -787,7 +786,7 @@ python lerobot/scripts/control_robot.py record \
   --num-episodes 2
 ```
 
-This will write your dataset locally to `{root}/{repo-id}` (e.g. `data/cadene/koch_test`) and push it on the hub at `https://huggingface.co/datasets/{HF_USER}/{repo-id}`. Your dataset will be automatically tagged with `LeRobot` for the community to find it easily, and you can also add custom tags (in this case `tutorial` for example).
+This will write your dataset locally to `~/.cache/huggingface/lerobot/{repo-id}` (e.g. `data/cadene/koch_test`) and push it on the hub at `https://huggingface.co/datasets/{HF_USER}/{repo-id}`. Your dataset will be automatically tagged with `LeRobot` for the community to find it easily, and you can also add custom tags (in this case `tutorial` for example).
 
 You can look for other LeRobot datasets on the hub by searching for `LeRobot` tags: https://huggingface.co/datasets?other=LeRobot
 
@@ -840,7 +839,6 @@ In the coming months, we plan to release a foundational model for robotics. We a
 You can visualize your dataset by running:
 ```bash
 python lerobot/scripts/visualize_dataset_html.py \
-  --root data \
   --repo-id ${HF_USER}/koch_test
 ```
 
@@ -858,7 +856,6 @@ To replay the first episode of the dataset you just recorded, run the following
 python lerobot/scripts/control_robot.py replay \
   --robot-path lerobot/configs/robot/koch.yaml \
   --fps 30 \
-  --root data \
   --repo-id ${HF_USER}/koch_test \
   --episode 0
 ```
@@ -871,7 +868,7 @@ Your robot should replicate movements similar to those you recorded. For example
 
 To train a policy to control your robot, use the [`python lerobot/scripts/train.py`](../lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
 ```bash
-DATA_DIR=data python lerobot/scripts/train.py \
+python lerobot/scripts/train.py \
   dataset_repo_id=${HF_USER}/koch_test \
   policy=act_koch_real \
   env=koch_real \
@@ -918,7 +915,6 @@ env:
 It should match your dataset (e.g. `fps: 30`) and your robot (e.g. `state_dim: 6` and `action_dim: 6`). We are still working on simplifying this in future versions of `lerobot`.
 4. We provided `device=cuda` since we are training on a Nvidia GPU, but you could use `device=mps` to train on Apple silicon.
 5. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
-6. We added `DATA_DIR=data` to access your dataset stored in your local `data` directory. If you dont provide `DATA_DIR`, your dataset will be downloaded from Hugging Face hub to your cache folder `$HOME/.cache/hugginface`. In future versions of `lerobot`, both directories will be in sync.
 
 For more information on the `train` script see the previous tutorial: [`examples/4_train_policy_with_script.md`](../examples/4_train_policy_with_script.md)
 
@@ -991,7 +987,6 @@ To this end, you can use the `record` function from [`lerobot/scripts/control_ro
 python lerobot/scripts/control_robot.py record \
   --robot-path lerobot/configs/robot/koch.yaml \
   --fps 30 \
-  --root data \
   --repo-id ${HF_USER}/eval_koch_test \
   --tags tutorial eval \
   --warmup-time-s 5 \
@@ -1010,7 +1005,6 @@ As you can see, it's almost the same command as previously used to record your t
 You can then visualize your evaluation dataset by running the same command as before but with the new inference dataset as argument:
 ```bash
 python lerobot/scripts/visualize_dataset.py \
-  --root data \
   --repo-id ${HF_USER}/eval_koch_test
 ```
 

examples/8_use_stretch.md

@@ -128,7 +128,6 @@ Record one episode:
 python lerobot/scripts/control_robot.py record \
     --robot-path lerobot/configs/robot/stretch.yaml \
     --fps 20 \
-    --root data \
     --repo-id ${HF_USER}/stretch_test \
     --tags stretch tutorial \
     --warmup-time-s 3 \
@@ -146,7 +145,6 @@ Now try to replay this episode (make sure the robot's initial position is the sa
 python lerobot/scripts/control_robot.py replay \
     --robot-path lerobot/configs/robot/stretch.yaml \
     --fps 20 \
-    --root data \
     --repo-id ${HF_USER}/stretch_test \
     --episode 0
 ```

examples/9_use_aloha.md

@@ -56,7 +56,7 @@ python lerobot/scripts/control_robot.py teleoperate \
     --robot-overrides max_relative_target=5
 ```
 
-By adding `--robot-overrides max_relative_target=5`, we override the default value for `max_relative_target` defined in `lerobot/configs/robot/aloha.yaml`. It is expected to be `5` to limit the magnitude of the movement for more safety, but the teloperation won't be smooth. When you feel confident, you can disable this limit by adding `--robot-overrides max_relative_target=null` to the command line:
+By adding `--robot-overrides max_relative_target=5`, we override the default value for `max_relative_target` defined in `lerobot/configs/robot/aloha.yaml`. It is expected to be `5` to limit the magnitude of the movement for more safety, but the teleoperation won't be smooth. When you feel confident, you can disable this limit by adding `--robot-overrides max_relative_target=null` to the command line:
 ```bash
 python lerobot/scripts/control_robot.py teleoperate \
     --robot-path lerobot/configs/robot/aloha.yaml \
@@ -84,7 +84,6 @@ python lerobot/scripts/control_robot.py record \
     --robot-path lerobot/configs/robot/aloha.yaml \
     --robot-overrides max_relative_target=null \
     --fps 30 \
-    --root data \
     --repo-id ${HF_USER}/aloha_test \
     --tags aloha tutorial \
     --warmup-time-s 5 \
@@ -104,7 +103,6 @@ echo ${HF_USER}/aloha_test
 If you didn't upload with `--push-to-hub 0`, you can also visualize it locally with:
 ```bash
 python lerobot/scripts/visualize_dataset_html.py \
-  --root data \
   --repo-id ${HF_USER}/aloha_test
 ```
 
@@ -119,7 +117,6 @@ python lerobot/scripts/control_robot.py replay \
     --robot-path lerobot/configs/robot/aloha.yaml \
     --robot-overrides max_relative_target=null \
     --fps 30 \
-    --root data \
     --repo-id ${HF_USER}/aloha_test \
     --episode 0
 ```
@@ -128,7 +125,7 @@ python lerobot/scripts/control_robot.py replay \
 
 To train a policy to control your robot, use the [`python lerobot/scripts/train.py`](../lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
 ```bash
-DATA_DIR=data python lerobot/scripts/train.py \
+python lerobot/scripts/train.py \
   dataset_repo_id=${HF_USER}/aloha_test \
   policy=act_aloha_real \
   env=aloha_real \
@@ -144,7 +141,6 @@ Let's explain it:
 3. We provided an environment as argument with `env=aloha_real`. This loads configurations from [`lerobot/configs/env/aloha_real.yaml`](../lerobot/configs/env/aloha_real.yaml). Note: this yaml defines 18 dimensions for the `state_dim` and `action_dim`, corresponding to 18 motors, not 14 motors as used in previous Aloha work. This is because, we include the `shoulder_shadow` and `elbow_shadow` motors for simplicity.
 4. We provided `device=cuda` since we are training on a Nvidia GPU.
 5. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
-6. We added `DATA_DIR=data` to access your dataset stored in your local `data` directory. If you dont provide `DATA_DIR`, your dataset will be downloaded from Hugging Face hub to your cache folder `$HOME/.cache/hugginface`. In future versions of `lerobot`, both directories will be in sync.
 
 Training should take several hours. You will find checkpoints in `outputs/train/act_aloha_test/checkpoints`.
 
@@ -156,7 +152,6 @@ python lerobot/scripts/control_robot.py record \
   --robot-path lerobot/configs/robot/aloha.yaml \
   --robot-overrides max_relative_target=null \
   --fps 30 \
-  --root data \
   --repo-id ${HF_USER}/eval_act_aloha_test \
   --tags aloha tutorial eval \
   --warmup-time-s 5 \

examples/advanced/2_calculate_validation_loss.py

@@ -14,7 +14,7 @@ from pathlib import Path
 import torch
 from huggingface_hub import snapshot_download
 
-from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
 from lerobot.common.policies.diffusion.modeling_diffusion import DiffusionPolicy
 
 device = torch.device("cuda")
@@ -41,26 +41,20 @@ delta_timestamps = {
 }
 
 # Load the last 10% of episodes of the dataset as a validation set.
-# - Load full dataset
-full_dataset = LeRobotDataset("lerobot/pusht", split="train")
-# - Calculate train and val subsets
-num_train_episodes = math.floor(full_dataset.num_episodes * 90 / 100)
-num_val_episodes = full_dataset.num_episodes - num_train_episodes
-print(f"Number of episodes in full dataset: {full_dataset.num_episodes}")
-print(f"Number of episodes in training dataset (90% subset): {num_train_episodes}")
-print(f"Number of episodes in validation dataset (10% subset): {num_val_episodes}")
-# - Get first frame index of the validation set
-first_val_frame_index = full_dataset.episode_data_index["from"][num_train_episodes].item()
-# - Load frames subset belonging to validation set using the `split` argument.
-#   It utilizes the `datasets` library's syntax for slicing datasets.
-#   For more information on the Slice API, please see:
-#   https://huggingface.co/docs/datasets/v2.19.0/loading#slice-splits
-train_dataset = LeRobotDataset(
-    "lerobot/pusht", split=f"train[:{first_val_frame_index}]", delta_timestamps=delta_timestamps
-)
-val_dataset = LeRobotDataset(
-    "lerobot/pusht", split=f"train[{first_val_frame_index}:]", delta_timestamps=delta_timestamps
-)
+# - Load dataset metadata
+dataset_metadata = LeRobotDatasetMetadata("lerobot/pusht")
+# - Calculate train and val episodes
+total_episodes = dataset_metadata.total_episodes
+episodes = list(range(dataset_metadata.total_episodes))
+num_train_episodes = math.floor(total_episodes * 90 / 100)
+train_episodes = episodes[:num_train_episodes]
+val_episodes = episodes[num_train_episodes:]
+print(f"Number of episodes in full dataset: {total_episodes}")
+print(f"Number of episodes in training dataset (90% subset): {len(train_episodes)}")
+print(f"Number of episodes in validation dataset (10% subset): {len(val_episodes)}")
+# - Load train an val datasets
+train_dataset = LeRobotDataset("lerobot/pusht", episodes=train_episodes, delta_timestamps=delta_timestamps)
+val_dataset = LeRobotDataset("lerobot/pusht", episodes=val_episodes, delta_timestamps=delta_timestamps)
 print(f"Number of frames in training dataset (90% subset): {len(train_dataset)}")
 print(f"Number of frames in validation dataset (10% subset): {len(val_dataset)}")
 

examples/port_datasets/pusht_zarr.py

@@ -0,0 +1,222 @@
+import shutil
+from pathlib import Path
+
+import numpy as np
+import torch
+
+from lerobot.common.datasets.lerobot_dataset import LEROBOT_HOME, 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": [
+            "channel",
+            "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,))
+    # 8 keypoints with 2 coords each
+    keypoints = np.zeros((num_frames, 16))
+
+    # 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] = torch.from_numpy(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 (LEROBOT_HOME / repo_id).exists():
+        shutil.rmtree(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)
+
+    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
+            frame = {
+                "action": torch.from_numpy(action[i]),
+                # Shift reward and success by +1 until the last item of the episode
+                "next.reward": reward[i + (frame_idx < num_frames - 1)],
+                "next.success": success[i + (frame_idx < num_frames - 1)],
+            }
+
+            frame["observation.state"] = torch.from_numpy(agent_pos[i])
+
+            if mode == "keypoints":
+                frame["observation.environment_state"] = torch.from_numpy(keypoints[i])
+            else:
+                frame["observation.image"] = torch.from_numpy(image[i])
+
+            dataset.add_frame(frame)
+
+        dataset.save_episode(task=PUSHT_TASK)
+
+    dataset.consolidate()
+
+    if push_to_hub:
+        dataset.push_to_hub()
+
+
+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, local_files_only=True)
+        # breakpoint()

lerobot/__init__.py

@@ -181,8 +181,8 @@ available_real_world_datasets = [
     "lerobot/usc_cloth_sim",
 ]
 
-available_datasets = list(
-    itertools.chain(*available_datasets_per_env.values(), available_real_world_datasets)
+available_datasets = sorted(
+    set(itertools.chain(*available_datasets_per_env.values(), available_real_world_datasets))
 )
 
 # lists all available policies from `lerobot/common/policies`

lerobot/common/datasets/card_template.md

@@ -0,0 +1,27 @@
+---
+# For reference on dataset card metadata, see the spec: https://github.com/huggingface/hub-docs/blob/main/datasetcard.md?plain=1
+# Doc / guide: https://huggingface.co/docs/hub/datasets-cards
+{{ card_data }}
+---
+
+This dataset was created using [LeRobot](https://github.com/huggingface/lerobot).
+
+## Dataset Description
+
+{{ dataset_description | default("", true) }}
+
+- **Homepage:** {{ url | default("[More Information Needed]", true)}}
+- **Paper:** {{ paper | default("[More Information Needed]", true)}}
+- **License:** {{ license | default("[More Information Needed]", true)}}
+
+## Dataset Structure
+
+{{ dataset_structure | default("[More Information Needed]", true)}}
+
+## Citation
+
+**BibTeX:**
+
+```bibtex
+{{ citation_bibtex | default("[More Information Needed]", true)}}
+```

lerobot/common/datasets/compute_stats.py

@@ -19,9 +19,6 @@ from math import ceil
 import einops
 import torch
 import tqdm
-from datasets import Image
-
-from lerobot.common.datasets.video_utils import VideoFrame
 
 
 def get_stats_einops_patterns(dataset, num_workers=0):
@@ -39,15 +36,13 @@ def get_stats_einops_patterns(dataset, num_workers=0):
     batch = next(iter(dataloader))
 
     stats_patterns = {}
-    for key, feats_type in dataset.features.items():
-        # NOTE: skip language_instruction embedding in stats computation
-        if key == "language_instruction":
-            continue
 
+    for key in dataset.features:
         # sanity check that tensors are not float64
         assert batch[key].dtype != torch.float64
 
-        if isinstance(feats_type, (VideoFrame, Image)):
+        # if isinstance(feats_type, (VideoFrame, Image)):
+        if key in dataset.meta.camera_keys:
             # sanity check that images are channel first
             _, c, h, w = batch[key].shape
             assert c < h and c < w, f"expect channel first images, but instead {batch[key].shape}"
@@ -63,7 +58,7 @@ def get_stats_einops_patterns(dataset, num_workers=0):
         elif batch[key].ndim == 1:
             stats_patterns[key] = "b -> 1"
         else:
-            raise ValueError(f"{key}, {feats_type}, {batch[key].shape}")
+            raise ValueError(f"{key}, {batch[key].shape}")
 
     return stats_patterns
 
@@ -175,39 +170,45 @@ def aggregate_stats(ls_datasets) -> dict[str, torch.Tensor]:
     """
     data_keys = set()
     for dataset in ls_datasets:
-        data_keys.update(dataset.stats.keys())
+        data_keys.update(dataset.meta.stats.keys())
     stats = {k: {} for k in data_keys}
     for data_key in data_keys:
         for stat_key in ["min", "max"]:
             # compute `max(dataset_0["max"], dataset_1["max"], ...)`
             stats[data_key][stat_key] = einops.reduce(
-                torch.stack([d.stats[data_key][stat_key] for d in ls_datasets if data_key in d.stats], dim=0),
+                torch.stack(
+                    [ds.meta.stats[data_key][stat_key] for ds in ls_datasets if data_key in ds.meta.stats],
+                    dim=0,
+                ),
                 "n ... -> ...",
                 stat_key,
             )
-        total_samples = sum(d.num_samples for d in ls_datasets if data_key in d.stats)
+        total_samples = sum(d.num_frames for d in ls_datasets if data_key in d.meta.stats)
         # Compute the "sum" statistic by multiplying each mean by the number of samples in the respective
         # dataset, then divide by total_samples to get the overall "mean".
-        # NOTE: the brackets around (d.num_samples / total_samples) are needed tor minimize the risk of
+        # NOTE: the brackets around (d.num_frames / total_samples) are needed tor minimize the risk of
         # numerical overflow!
         stats[data_key]["mean"] = sum(
-            d.stats[data_key]["mean"] * (d.num_samples / total_samples)
+            d.meta.stats[data_key]["mean"] * (d.num_frames / total_samples)
             for d in ls_datasets
-            if data_key in d.stats
+            if data_key in d.meta.stats
         )
         # The derivation for standard deviation is a little more involved but is much in the same spirit as
         # the computation of the mean.
         # Given two sets of data where the statistics are known:
         # σ_combined = sqrt[ (n1 * (σ1^2 + d1^2) + n2 * (σ2^2 + d2^2)) / (n1 + n2) ]
         # where d1 = μ1 - μ_combined, d2 = μ2 - μ_combined
-        # NOTE: the brackets around (d.num_samples / total_samples) are needed tor minimize the risk of
+        # NOTE: the brackets around (d.num_frames / total_samples) are needed tor minimize the risk of
         # numerical overflow!
         stats[data_key]["std"] = torch.sqrt(
             sum(
-                (d.stats[data_key]["std"] ** 2 + (d.stats[data_key]["mean"] - stats[data_key]["mean"]) ** 2)
-                * (d.num_samples / total_samples)
+                (
+                    d.meta.stats[data_key]["std"] ** 2
+                    + (d.meta.stats[data_key]["mean"] - stats[data_key]["mean"]) ** 2
+                )
+                * (d.num_frames / total_samples)
                 for d in ls_datasets
-                if data_key in d.stats
+                if data_key in d.meta.stats
             )
         )
     return stats

lerobot/common/datasets/factory.py

@@ -91,9 +91,9 @@ def make_dataset(cfg, split: str = "train") -> LeRobotDataset | MultiLeRobotData
         )
 
     if isinstance(cfg.dataset_repo_id, str):
+        # TODO (aliberts): add 'episodes' arg from config after removing hydra
         dataset = LeRobotDataset(
             cfg.dataset_repo_id,
-            split=split,
             delta_timestamps=cfg.training.get("delta_timestamps"),
             image_transforms=image_transforms,
             video_backend=cfg.video_backend,
@@ -101,7 +101,6 @@ def make_dataset(cfg, split: str = "train") -> LeRobotDataset | MultiLeRobotData
     else:
         dataset = MultiLeRobotDataset(
             cfg.dataset_repo_id,
-            split=split,
             delta_timestamps=cfg.training.get("delta_timestamps"),
             image_transforms=image_transforms,
             video_backend=cfg.video_backend,
@@ -112,6 +111,6 @@ def make_dataset(cfg, split: str = "train") -> LeRobotDataset | MultiLeRobotData
             for stats_type, listconfig in stats_dict.items():
                 # example of stats_type: min, max, mean, std
                 stats = OmegaConf.to_container(listconfig, resolve=True)
-                dataset.stats[key][stats_type] = torch.tensor(stats, dtype=torch.float32)
+                dataset.meta.stats[key][stats_type] = torch.tensor(stats, dtype=torch.float32)
 
     return dataset

lerobot/common/datasets/image_writer.py

@@ -0,0 +1,160 @@
+#!/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 multiprocessing
+import queue
+import threading
+from pathlib import Path
+
+import numpy as np
+import PIL.Image
+import torch
+
+
+def safe_stop_image_writer(func):
+    def wrapper(*args, **kwargs):
+        try:
+            return func(*args, **kwargs)
+        except Exception as e:
+            dataset = kwargs.get("dataset")
+            image_writer = getattr(dataset, "image_writer", None) if dataset else None
+            if image_writer is not None:
+                print("Waiting for image writer to terminate...")
+                image_writer.stop()
+            raise e
+
+    return wrapper
+
+
+def image_array_to_image(image_array: np.ndarray) -> PIL.Image.Image:
+    # TODO(aliberts): handle 1 channel and 4 for depth images
+    if image_array.ndim == 3 and image_array.shape[0] in [1, 3]:
+        # Transpose from pytorch convention (C, H, W) to (H, W, C)
+        image_array = image_array.transpose(1, 2, 0)
+    if image_array.dtype != np.uint8:
+        # Assume the image is in [0, 1] range for floating-point data
+        image_array = np.clip(image_array, 0, 1)
+        image_array = (image_array * 255).astype(np.uint8)
+    return PIL.Image.fromarray(image_array)
+
+
+def write_image(image: np.ndarray | PIL.Image.Image, fpath: Path):
+    try:
+        if isinstance(image, np.ndarray):
+            img = image_array_to_image(image)
+        elif isinstance(image, PIL.Image.Image):
+            img = image
+        else:
+            raise TypeError(f"Unsupported image type: {type(image)}")
+        img.save(fpath)
+    except Exception as e:
+        print(f"Error writing image {fpath}: {e}")
+
+
+def worker_thread_loop(queue: queue.Queue):
+    while True:
+        item = queue.get()
+        if item is None:
+            queue.task_done()
+            break
+        image_array, fpath = item
+        write_image(image_array, fpath)
+        queue.task_done()
+
+
+def worker_process(queue: queue.Queue, num_threads: int):
+    threads = []
+    for _ in range(num_threads):
+        t = threading.Thread(target=worker_thread_loop, args=(queue,))
+        t.daemon = True
+        t.start()
+        threads.append(t)
+    for t in threads:
+        t.join()
+
+
+class AsyncImageWriter:
+    """
+    This class abstract away the initialisation of processes or/and threads to
+    save images on disk asynchrounously, which is critical to control a robot and record data
+    at a high frame rate.
+
+    When `num_processes=0`, it creates a threads pool of size `num_threads`.
+    When `num_processes>0`, it creates processes pool of size `num_processes`, where each subprocess starts
+    their own threads pool of size `num_threads`.
+
+    The optimal number of processes and threads depends on your computer capabilities.
+    We advise to use 4 threads per camera with 0 processes. If the fps is not stable, try to increase or lower
+    the number of threads. If it is still not stable, try to use 1 subprocess, or more.
+    """
+
+    def __init__(self, num_processes: int = 0, num_threads: int = 1):
+        self.num_processes = num_processes
+        self.num_threads = num_threads
+        self.queue = None
+        self.threads = []
+        self.processes = []
+        self._stopped = False
+
+        if num_threads <= 0 and num_processes <= 0:
+            raise ValueError("Number of threads and processes must be greater than zero.")
+
+        if self.num_processes == 0:
+            # Use threading
+            self.queue = queue.Queue()
+            for _ in range(self.num_threads):
+                t = threading.Thread(target=worker_thread_loop, args=(self.queue,))
+                t.daemon = True
+                t.start()
+                self.threads.append(t)
+        else:
+            # Use multiprocessing
+            self.queue = multiprocessing.JoinableQueue()
+            for _ in range(self.num_processes):
+                p = multiprocessing.Process(target=worker_process, args=(self.queue, self.num_threads))
+                p.daemon = True
+                p.start()
+                self.processes.append(p)
+
+    def save_image(self, image: torch.Tensor | np.ndarray | PIL.Image.Image, fpath: Path):
+        if isinstance(image, torch.Tensor):
+            # Convert tensor to numpy array to minimize main process time
+            image = image.cpu().numpy()
+        self.queue.put((image, fpath))
+
+    def wait_until_done(self):
+        self.queue.join()
+
+    def stop(self):
+        if self._stopped:
+            return
+
+        if self.num_processes == 0:
+            for _ in self.threads:
+                self.queue.put(None)
+            for t in self.threads:
+                t.join()
+        else:
+            num_nones = self.num_processes * self.num_threads
+            for _ in range(num_nones):
+                self.queue.put(None)
+            for p in self.processes:
+                p.join()
+                if p.is_alive():
+                    p.terminate()
+            self.queue.close()
+            self.queue.join_thread()
+
+        self._stopped = True

lerobot/common/datasets/online_buffer.py

@@ -187,7 +187,7 @@ class OnlineBuffer(torch.utils.data.Dataset):
         assert data[OnlineBuffer.INDEX_KEY][0].item() == 0
 
         # Shift the incoming indices if necessary.
-        if self.num_samples > 0:
+        if self.num_frames > 0:
             last_episode_index = self._data[OnlineBuffer.EPISODE_INDEX_KEY][next_index - 1]
             last_data_index = self._data[OnlineBuffer.INDEX_KEY][next_index - 1]
             data[OnlineBuffer.EPISODE_INDEX_KEY] += last_episode_index + 1
@@ -227,11 +227,11 @@ class OnlineBuffer(torch.utils.data.Dataset):
         )
 
     @property
-    def num_samples(self) -> int:
+    def num_frames(self) -> int:
         return np.count_nonzero(self._data[OnlineBuffer.OCCUPANCY_MASK_KEY])
 
     def __len__(self):
-        return self.num_samples
+        return self.num_frames
 
     def _item_to_tensors(self, item: dict) -> dict:
         item_ = {}

lerobot/common/datasets/populate_dataset.py

@@ -1,468 +0,0 @@
-"""Functions to create an empty dataset, and populate it with frames."""
-# TODO(rcadene, aliberts): to adapt as class methods of next version of LeRobotDataset
-
-import concurrent
-import json
-import logging
-import multiprocessing
-import shutil
-from pathlib import Path
-
-import torch
-import tqdm
-from PIL import Image
-
-from lerobot.common.datasets.compute_stats import compute_stats
-from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDataset
-from lerobot.common.datasets.push_dataset_to_hub.aloha_hdf5_format import to_hf_dataset
-from lerobot.common.datasets.push_dataset_to_hub.utils import concatenate_episodes, get_default_encoding
-from lerobot.common.datasets.utils import calculate_episode_data_index, create_branch
-from lerobot.common.datasets.video_utils import encode_video_frames
-from lerobot.common.utils.utils import log_say
-from lerobot.scripts.push_dataset_to_hub import (
-    push_dataset_card_to_hub,
-    push_meta_data_to_hub,
-    push_videos_to_hub,
-    save_meta_data,
-)
-
-########################################################################################
-# Asynchrounous saving of images on disk
-########################################################################################
-
-
-def safe_stop_image_writer(func):
-    # TODO(aliberts): Allow to pass custom exceptions
-    # (e.g. ThreadServiceExit, KeyboardInterrupt, SystemExit, UnpluggedError, DynamixelCommError)
-    def wrapper(*args, **kwargs):
-        try:
-            return func(*args, **kwargs)
-        except Exception as e:
-            image_writer = kwargs.get("dataset", {}).get("image_writer")
-            if image_writer is not None:
-                print("Waiting for image writer to terminate...")
-                stop_image_writer(image_writer, timeout=20)
-            raise e
-
-    return wrapper
-
-
-def save_image(img_tensor, key, frame_index, episode_index, videos_dir: str):
-    img = Image.fromarray(img_tensor.numpy())
-    path = Path(videos_dir) / f"{key}_episode_{episode_index:06d}" / f"frame_{frame_index:06d}.png"
-    path.parent.mkdir(parents=True, exist_ok=True)
-    img.save(str(path), quality=100)
-
-
-def loop_to_save_images_in_threads(image_queue, num_threads):
-    if num_threads < 1:
-        raise NotImplementedError(f"Only `num_threads>=1` is supported for now, but {num_threads=} given.")
-
-    with concurrent.futures.ThreadPoolExecutor(max_workers=num_threads) as executor:
-        futures = []
-        while True:
-            # Blocks until a frame is available
-            frame_data = image_queue.get()
-
-            # As usually done, exit loop when receiving None to stop the worker
-            if frame_data is None:
-                break
-
-            image, key, frame_index, episode_index, videos_dir = frame_data
-            futures.append(executor.submit(save_image, image, key, frame_index, episode_index, videos_dir))
-
-        # Before exiting function, wait for all threads to complete
-        with tqdm.tqdm(total=len(futures), desc="Writing images") as progress_bar:
-            concurrent.futures.wait(futures)
-            progress_bar.update(len(futures))
-
-
-def start_image_writer_processes(image_queue, num_processes, num_threads_per_process):
-    if num_processes < 1:
-        raise ValueError(f"Only `num_processes>=1` is supported, but {num_processes=} given.")
-
-    if num_threads_per_process < 1:
-        raise NotImplementedError(
-            "Only `num_threads_per_process>=1` is supported for now, but {num_threads_per_process=} given."
-        )
-
-    processes = []
-    for _ in range(num_processes):
-        process = multiprocessing.Process(
-            target=loop_to_save_images_in_threads,
-            args=(image_queue, num_threads_per_process),
-        )
-        process.start()
-        processes.append(process)
-    return processes
-
-
-def stop_processes(processes, queue, timeout):
-    # Send None to each process to signal them to stop
-    for _ in processes:
-        queue.put(None)
-
-    # Wait maximum 20 seconds for all processes to terminate
-    for process in processes:
-        process.join(timeout=timeout)
-
-    # If not terminated after 20 seconds, force termination
-    if process.is_alive():
-        process.terminate()
-
-    # Close the queue, no more items can be put in the queue
-    queue.close()
-
-    # Ensure all background queue threads have finished
-    queue.join_thread()
-
-
-def start_image_writer(num_processes, num_threads):
-    """This function abstract away the initialisation of processes or/and threads to
-    save images on disk asynchrounously, which is critical to control a robot and record data
-    at a high frame rate.
-
-    When `num_processes=0`, it returns a dictionary containing a threads pool of size `num_threads`.
-    When `num_processes>0`, it returns a dictionary containing a processes pool of size `num_processes`,
-    where each subprocess starts their own threads pool of size `num_threads`.
-
-    The optimal number of processes and threads depends on your computer capabilities.
-    We advise to use 4 threads per camera with 0 processes. If the fps is not stable, try to increase or lower
-    the number of threads. If it is still not stable, try to use 1 subprocess, or more.
-    """
-    image_writer = {}
-
-    if num_processes == 0:
-        futures = []
-        threads_pool = concurrent.futures.ThreadPoolExecutor(max_workers=num_threads)
-        image_writer["threads_pool"], image_writer["futures"] = threads_pool, futures
-    else:
-        # TODO(rcadene): When using num_processes>1, `multiprocessing.Manager().Queue()`
-        # might be better than `multiprocessing.Queue()`. Source: https://www.geeksforgeeks.org/python-multiprocessing-queue-vs-multiprocessing-manager-queue
-        image_queue = multiprocessing.Queue()
-        processes_pool = start_image_writer_processes(
-            image_queue, num_processes=num_processes, num_threads_per_process=num_threads
-        )
-        image_writer["processes_pool"], image_writer["image_queue"] = processes_pool, image_queue
-
-    return image_writer
-
-
-def async_save_image(image_writer, image, key, frame_index, episode_index, videos_dir):
-    """This function abstract away the saving of an image on disk asynchrounously. It uses a dictionary
-    called image writer which contains either a pool of processes or a pool of threads.
-    """
-    if "threads_pool" in image_writer:
-        threads_pool, futures = image_writer["threads_pool"], image_writer["futures"]
-        futures.append(threads_pool.submit(save_image, image, key, frame_index, episode_index, videos_dir))
-    else:
-        image_queue = image_writer["image_queue"]
-        image_queue.put((image, key, frame_index, episode_index, videos_dir))
-
-
-def stop_image_writer(image_writer, timeout):
-    if "threads_pool" in image_writer:
-        futures = image_writer["futures"]
-        # Before exiting function, wait for all threads to complete
-        with tqdm.tqdm(total=len(futures), desc="Writing images") as progress_bar:
-            concurrent.futures.wait(futures, timeout=timeout)
-            progress_bar.update(len(futures))
-    else:
-        processes_pool, image_queue = image_writer["processes_pool"], image_writer["image_queue"]
-        stop_processes(processes_pool, image_queue, timeout=timeout)
-
-
-########################################################################################
-# Functions to initialize, resume and populate a dataset
-########################################################################################
-
-
-def init_dataset(
-    repo_id,
-    root,
-    force_override,
-    fps,
-    video,
-    write_images,
-    num_image_writer_processes,
-    num_image_writer_threads,
-):
-    local_dir = Path(root) / repo_id
-    if local_dir.exists() and force_override:
-        shutil.rmtree(local_dir)
-
-    episodes_dir = local_dir / "episodes"
-    episodes_dir.mkdir(parents=True, exist_ok=True)
-
-    videos_dir = local_dir / "videos"
-    videos_dir.mkdir(parents=True, exist_ok=True)
-
-    # Logic to resume data recording
-    rec_info_path = episodes_dir / "data_recording_info.json"
-    if rec_info_path.exists():
-        with open(rec_info_path) as f:
-            rec_info = json.load(f)
-        num_episodes = rec_info["last_episode_index"] + 1
-    else:
-        num_episodes = 0
-
-    dataset = {
-        "repo_id": repo_id,
-        "local_dir": local_dir,
-        "videos_dir": videos_dir,
-        "episodes_dir": episodes_dir,
-        "fps": fps,
-        "video": video,
-        "rec_info_path": rec_info_path,
-        "num_episodes": num_episodes,
-    }
-
-    if write_images:
-        # Initialize processes or/and threads dedicated to save images on disk asynchronously,
-        # which is critical to control a robot and record data at a high frame rate.
-        image_writer = start_image_writer(
-            num_processes=num_image_writer_processes,
-            num_threads=num_image_writer_threads,
-        )
-        dataset["image_writer"] = image_writer
-
-    return dataset
-
-
-def add_frame(dataset, observation, action):
-    if "current_episode" not in dataset:
-        # initialize episode dictionary
-        ep_dict = {}
-        for key in observation:
-            if key not in ep_dict:
-                ep_dict[key] = []
-        for key in action:
-            if key not in ep_dict:
-                ep_dict[key] = []
-
-        ep_dict["episode_index"] = []
-        ep_dict["frame_index"] = []
-        ep_dict["timestamp"] = []
-        ep_dict["next.done"] = []
-
-        dataset["current_episode"] = ep_dict
-        dataset["current_frame_index"] = 0
-
-    ep_dict = dataset["current_episode"]
-    episode_index = dataset["num_episodes"]
-    frame_index = dataset["current_frame_index"]
-    videos_dir = dataset["videos_dir"]
-    video = dataset["video"]
-    fps = dataset["fps"]
-
-    ep_dict["episode_index"].append(episode_index)
-    ep_dict["frame_index"].append(frame_index)
-    ep_dict["timestamp"].append(frame_index / fps)
-    ep_dict["next.done"].append(False)
-
-    img_keys = [key for key in observation if "image" in key]
-    non_img_keys = [key for key in observation if "image" not in key]
-
-    # Save all observed modalities except images
-    for key in non_img_keys:
-        ep_dict[key].append(observation[key])
-
-    # Save actions
-    for key in action:
-        ep_dict[key].append(action[key])
-
-    if "image_writer" not in dataset:
-        dataset["current_frame_index"] += 1
-        return
-
-    # Save images
-    image_writer = dataset["image_writer"]
-    for key in img_keys:
-        imgs_dir = videos_dir / f"{key}_episode_{episode_index:06d}"
-        async_save_image(
-            image_writer,
-            image=observation[key],
-            key=key,
-            frame_index=frame_index,
-            episode_index=episode_index,
-            videos_dir=str(videos_dir),
-        )
-
-        if video:
-            fname = f"{key}_episode_{episode_index:06d}.mp4"
-            frame_info = {"path": f"videos/{fname}", "timestamp": frame_index / fps}
-        else:
-            frame_info = str(imgs_dir / f"frame_{frame_index:06d}.png")
-
-        ep_dict[key].append(frame_info)
-
-    dataset["current_frame_index"] += 1
-
-
-def delete_current_episode(dataset):
-    del dataset["current_episode"]
-    del dataset["current_frame_index"]
-
-    # delete temporary images
-    episode_index = dataset["num_episodes"]
-    videos_dir = dataset["videos_dir"]
-    for tmp_imgs_dir in videos_dir.glob(f"*_episode_{episode_index:06d}"):
-        shutil.rmtree(tmp_imgs_dir)
-
-
-def save_current_episode(dataset):
-    episode_index = dataset["num_episodes"]
-    ep_dict = dataset["current_episode"]
-    episodes_dir = dataset["episodes_dir"]
-    rec_info_path = dataset["rec_info_path"]
-
-    ep_dict["next.done"][-1] = True
-
-    for key in ep_dict:
-        if "observation" in key and "image" not in key:
-            ep_dict[key] = torch.stack(ep_dict[key])
-
-    ep_dict["action"] = torch.stack(ep_dict["action"])
-    ep_dict["episode_index"] = torch.tensor(ep_dict["episode_index"])
-    ep_dict["frame_index"] = torch.tensor(ep_dict["frame_index"])
-    ep_dict["timestamp"] = torch.tensor(ep_dict["timestamp"])
-    ep_dict["next.done"] = torch.tensor(ep_dict["next.done"])
-
-    ep_path = episodes_dir / f"episode_{episode_index}.pth"
-    torch.save(ep_dict, ep_path)
-
-    rec_info = {
-        "last_episode_index": episode_index,
-    }
-    with open(rec_info_path, "w") as f:
-        json.dump(rec_info, f)
-
-    # force re-initialization of episode dictionnary during add_frame
-    del dataset["current_episode"]
-
-    dataset["num_episodes"] += 1
-
-
-def encode_videos(dataset, image_keys, play_sounds):
-    log_say("Encoding videos", play_sounds)
-
-    num_episodes = dataset["num_episodes"]
-    videos_dir = dataset["videos_dir"]
-    local_dir = dataset["local_dir"]
-    fps = dataset["fps"]
-
-    # Use ffmpeg to convert frames stored as png into mp4 videos
-    for episode_index in tqdm.tqdm(range(num_episodes)):
-        for key in image_keys:
-            # key = f"observation.images.{name}"
-            tmp_imgs_dir = videos_dir / f"{key}_episode_{episode_index:06d}"
-            fname = f"{key}_episode_{episode_index:06d}.mp4"
-            video_path = local_dir / "videos" / fname
-            if video_path.exists():
-                # Skip if video is already encoded. Could be the case when resuming data recording.
-                continue
-            # note: `encode_video_frames` is a blocking call. Making it asynchronous shouldn't speedup encoding,
-            # since video encoding with ffmpeg is already using multithreading.
-            encode_video_frames(tmp_imgs_dir, video_path, fps, overwrite=True)
-            shutil.rmtree(tmp_imgs_dir)
-
-
-def from_dataset_to_lerobot_dataset(dataset, play_sounds):
-    log_say("Consolidate episodes", play_sounds)
-
-    num_episodes = dataset["num_episodes"]
-    episodes_dir = dataset["episodes_dir"]
-    videos_dir = dataset["videos_dir"]
-    video = dataset["video"]
-    fps = dataset["fps"]
-    repo_id = dataset["repo_id"]
-
-    ep_dicts = []
-    for episode_index in tqdm.tqdm(range(num_episodes)):
-        ep_path = episodes_dir / f"episode_{episode_index}.pth"
-        ep_dict = torch.load(ep_path)
-        ep_dicts.append(ep_dict)
-    data_dict = concatenate_episodes(ep_dicts)
-
-    if video:
-        image_keys = [key for key in data_dict if "image" in key]
-        encode_videos(dataset, image_keys, play_sounds)
-
-    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()
-
-    lerobot_dataset = LeRobotDataset.from_preloaded(
-        repo_id=repo_id,
-        hf_dataset=hf_dataset,
-        episode_data_index=episode_data_index,
-        info=info,
-        videos_dir=videos_dir,
-    )
-
-    return lerobot_dataset
-
-
-def save_lerobot_dataset_on_disk(lerobot_dataset):
-    hf_dataset = lerobot_dataset.hf_dataset
-    info = lerobot_dataset.info
-    stats = lerobot_dataset.stats
-    episode_data_index = lerobot_dataset.episode_data_index
-    local_dir = lerobot_dataset.videos_dir.parent
-    meta_data_dir = local_dir / "meta_data"
-
-    hf_dataset = hf_dataset.with_format(None)  # to remove transforms that cant be saved
-    hf_dataset.save_to_disk(str(local_dir / "train"))
-
-    save_meta_data(info, stats, episode_data_index, meta_data_dir)
-
-
-def push_lerobot_dataset_to_hub(lerobot_dataset, tags):
-    hf_dataset = lerobot_dataset.hf_dataset
-    local_dir = lerobot_dataset.videos_dir.parent
-    videos_dir = lerobot_dataset.videos_dir
-    repo_id = lerobot_dataset.repo_id
-    video = lerobot_dataset.video
-    meta_data_dir = local_dir / "meta_data"
-
-    if not (local_dir / "train").exists():
-        raise ValueError(
-            "You need to run `save_lerobot_dataset_on_disk(lerobot_dataset)` before pushing to the hub."
-        )
-
-    hf_dataset.push_to_hub(repo_id, revision="main")
-    push_meta_data_to_hub(repo_id, meta_data_dir, revision="main")
-    push_dataset_card_to_hub(repo_id, revision="main", tags=tags)
-    if video:
-        push_videos_to_hub(repo_id, videos_dir, revision="main")
-    create_branch(repo_id, repo_type="dataset", branch=CODEBASE_VERSION)
-
-
-def create_lerobot_dataset(dataset, run_compute_stats, push_to_hub, tags, play_sounds):
-    if "image_writer" in dataset:
-        logging.info("Waiting for image writer to terminate...")
-        image_writer = dataset["image_writer"]
-        stop_image_writer(image_writer, timeout=20)
-
-    lerobot_dataset = from_dataset_to_lerobot_dataset(dataset, play_sounds)
-
-    if run_compute_stats:
-        log_say("Computing dataset statistics", play_sounds)
-        lerobot_dataset.stats = compute_stats(lerobot_dataset)
-    else:
-        logging.info("Skipping computation of the dataset statistics")
-        lerobot_dataset.stats = {}
-
-    save_lerobot_dataset_on_disk(lerobot_dataset)
-
-    if push_to_hub:
-        push_lerobot_dataset_to_hub(lerobot_dataset, tags)
-
-    return lerobot_dataset

lerobot/common/datasets/push_dataset_to_hub/aloha_hdf5_format.py

@@ -30,12 +30,12 @@ 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 (
-    calculate_episode_data_index,
     hf_transform_to_torch,
 )
 from lerobot.common.datasets.video_utils import VideoFrame, encode_video_frames

lerobot/common/datasets/push_dataset_to_hub/cam_png_format.py

@@ -24,8 +24,11 @@ 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 concatenate_episodes
-from lerobot.common.datasets.utils import calculate_episode_data_index, hf_transform_to_torch
+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
 
 

lerobot/common/datasets/push_dataset_to_hub/dora_parquet_format.py

@@ -26,8 +26,8 @@ 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 (
-    calculate_episode_data_index,
     hf_transform_to_torch,
 )
 from lerobot.common.datasets.video_utils import VideoFrame

lerobot/common/datasets/push_dataset_to_hub/openx/configs.yaml

@@ -1,639 +0,0 @@
-OPENX_DATASET_CONFIGS:
-  fractal20220817_data:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - base_pose_tool_reached
-      - gripper_closed
-    fps: 3
-
-  kuka:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - clip_function_input/base_pose_tool_reached
-      - gripper_closed
-    fps: 10
-
-  bridge_openx:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - EEF_state
-      - gripper_state
-    fps: 5
-
-  taco_play:
-    image_obs_keys:
-      - rgb_static
-      - rgb_gripper
-    depth_obs_keys:
-      - depth_static
-      - depth_gripper
-    state_obs_keys:
-      - state_eef
-      - state_gripper
-    fps: 15
-
-  jaco_play:
-    image_obs_keys:
-      - image
-      - image_wrist
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state_eef
-      - state_gripper
-    fps: 10
-
-  berkeley_cable_routing:
-    image_obs_keys:
-      - image
-      - top_image
-      - wrist45_image
-      - wrist225_image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - robot_state
-    fps: 10
-
-  roboturk:
-    image_obs_keys:
-      - front_rgb
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - null
-    fps: 10
-
-  nyu_door_opening_surprising_effectiveness:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - null
-    fps: 3
-
-  viola:
-    image_obs_keys:
-      - agentview_rgb
-      - eye_in_hand_rgb
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - joint_states
-      - gripper_states
-    fps: 20
-
-  berkeley_autolab_ur5:
-    image_obs_keys:
-      - image
-      - hand_image
-    depth_obs_keys:
-      - image_with_depth
-    state_obs_keys:
-      - state
-    fps: 5
-
-  toto:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-    fps: 30
-
-  language_table:
-    image_obs_keys:
-      - rgb
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - effector_translation
-    fps: 10
-
-  columbia_cairlab_pusht_real:
-    image_obs_keys:
-      - image
-      - wrist_image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - robot_state
-    fps: 10
-
-  stanford_kuka_multimodal_dataset_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - depth_image
-    state_obs_keys:
-      - ee_position
-      - ee_orientation
-    fps: 20
-
-  nyu_rot_dataset_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - eef_state
-      - gripper_state
-    fps: 3
-
-  io_ai_tech:
-    image_obs_keys:
-      - image
-      - image_fisheye
-      - image_left_side
-      - image_right_side
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-    fps: 3
-
-  stanford_hydra_dataset_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-      - wrist_image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - eef_state
-      - gripper_state
-    fps: 10
-
-  austin_buds_dataset_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-      - wrist_image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-    fps: 20
-
-  nyu_franka_play_dataset_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-      - image_additional_view
-    depth_obs_keys:
-      - depth
-      - depth_additional_view
-    state_obs_keys:
-      - eef_state
-    fps: 3
-
-  maniskill_dataset_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-      - wrist_image
-    depth_obs_keys:
-      - depth
-      - wrist_depth
-    state_obs_keys:
-      - tcp_pose
-      - gripper_state
-    fps: 20
-
-  furniture_bench_dataset_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-      - wrist_image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-    fps: 10
-
-  cmu_franka_exploration_dataset_converted_externally_to_rlds:
-    image_obs_keys:
-      - highres_image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - null
-    fps: 10
-
-  ucsd_kitchen_dataset_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - joint_state
-    fps: 2
-
-  ucsd_pick_and_place_dataset_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - eef_state
-      - gripper_state
-    fps: 3
-
-  spoc:
-    image_obs_keys:
-      - image
-      - image_manipulation
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - null
-    fps: 3
-
-  austin_sailor_dataset_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-      - wrist_image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-    fps: 20
-
-  austin_sirius_dataset_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-      - wrist_image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-    fps: 20
-
-  bc_z:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - present/xyz
-      - present/axis_angle
-      - present/sensed_close
-    fps: 10
-
-  utokyo_pr2_opening_fridge_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - eef_state
-      - gripper_state
-    fps: 10
-
-  utokyo_pr2_tabletop_manipulation_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - eef_state
-      - gripper_state
-    fps: 10
-
-  utokyo_xarm_pick_and_place_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-      - image2
-      - hand_image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - end_effector_pose
-    fps: 10
-
-  utokyo_xarm_bimanual_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - pose_r
-    fps: 10
-
-  robo_net:
-    image_obs_keys:
-      - image
-      - image1
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - eef_state
-      - gripper_state
-    fps: 1
-
-  robo_set:
-    image_obs_keys:
-      - image_left
-      - image_right
-      - image_wrist
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-      - state_velocity
-    fps: 5
-
-  berkeley_mvp_converted_externally_to_rlds:
-    image_obs_keys:
-      - hand_image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - gripper
-      - pose
-      - joint_pos
-    fps: 5
-
-  berkeley_rpt_converted_externally_to_rlds:
-    image_obs_keys:
-      - hand_image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - joint_pos
-      - gripper
-    fps: 30
-
-  kaist_nonprehensile_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-    fps: 10
-
-  stanford_mask_vit_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - eef_state
-      - gripper_state
-
-  tokyo_u_lsmo_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - eef_state
-      - gripper_state
-    fps: 10
-
-  dlr_sara_pour_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-    fps: 10
-
-  dlr_sara_grid_clamp_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-    fps: 10
-
-  dlr_edan_shared_control_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-    fps: 5
-
-  asu_table_top_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - eef_state
-      - gripper_state
-    fps: 12.5
-
-  stanford_robocook_converted_externally_to_rlds:
-    image_obs_keys:
-      - image_1
-      - image_2
-    depth_obs_keys:
-      - depth_1
-      - depth_2
-    state_obs_keys:
-      - eef_state
-      - gripper_state
-    fps: 5
-
-  imperialcollege_sawyer_wrist_cam:
-    image_obs_keys:
-      - image
-      - wrist_image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-    fps: 10
-
-  iamlab_cmu_pickup_insert_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-      - wrist_image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - joint_state
-      - gripper_state
-    fps: 20
-
-  uiuc_d3field:
-    image_obs_keys:
-      - image_1
-      - image_2
-    depth_obs_keys:
-      - depth_1
-      - depth_2
-    state_obs_keys:
-      - null
-    fps: 1
-
-  utaustin_mutex:
-    image_obs_keys:
-      - image
-      - wrist_image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-    fps: 20
-
-  berkeley_fanuc_manipulation:
-    image_obs_keys:
-      - image
-      - wrist_image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - joint_state
-      - gripper_state
-    fps: 10
-
-  cmu_playing_with_food:
-    image_obs_keys:
-      - image
-      - finger_vision_1
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-    fps: 10
-
-  cmu_play_fusion:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-    fps: 5
-
-  cmu_stretch:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - eef_state
-      - gripper_state
-    fps: 10
-
-  berkeley_gnm_recon:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-      - position
-      - yaw
-    fps: 3
-
-  berkeley_gnm_cory_hall:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-      - position
-      - yaw
-    fps: 5
-
-  berkeley_gnm_sac_son:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-      - position
-      - yaw
-    fps: 10
-
-  droid:
-    image_obs_keys:
-      - exterior_image_1_left
-      - exterior_image_2_left
-      - wrist_image_left
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - proprio
-    fps: 15
-
-  droid_100:
-    image_obs_keys:
-      - exterior_image_1_left
-      - exterior_image_2_left
-      - wrist_image_left
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - proprio
-    fps: 15
-
-  fmb:
-    image_obs_keys:
-      - image_side_1
-      - image_side_2
-      - image_wrist_1
-      - image_wrist_2
-    depth_obs_keys:
-      - image_side_1_depth
-      - image_side_2_depth
-      - image_wrist_1_depth
-      - image_wrist_2_depth
-    state_obs_keys:
-      - proprio
-    fps: 10
-
-  dobbe:
-    image_obs_keys:
-      - wrist_image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - proprio
-    fps: 3.75
-
-  usc_cloth_sim_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - null
-    fps: 10
-
-  plex_robosuite:
-    image_obs_keys:
-      - image
-      - wrist_image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-    fps: 20
-
-  conq_hose_manipulation:
-    image_obs_keys:
-      - frontleft_fisheye_image
-      - frontright_fisheye_image
-      - hand_color_image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-    fps: 30

lerobot/common/datasets/push_dataset_to_hub/openx/data_utils.py

@@ -1,106 +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 Licens    e.
-# 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.
-"""
-NOTE(YL): Adapted from:
-    Octo: https://github.com/octo-models/octo/blob/main/octo/data/utils/data_utils.py
-
-data_utils.py
-
-Additional utils for data processing.
-"""
-
-from typing import Any, Dict, List
-
-import tensorflow as tf
-
-
-def binarize_gripper_actions(actions: tf.Tensor) -> tf.Tensor:
-    """
-    Converts gripper actions from continuous to binary values (0 and 1).
-
-    We exploit that fact that most of the time, the gripper is fully open (near 1.0) or fully closed (near 0.0). As it
-    transitions between the two, it sometimes passes through a few intermediate values. We relabel those intermediate
-    values based on the state that is reached _after_ those intermediate values.
-
-    In the edge case that the trajectory ends with an intermediate value, we give up on binarizing and relabel that
-    chunk of intermediate values as the last action in the trajectory.
-
-    The `scan_fn` implements the following logic:
-        new_actions = np.empty_like(actions)
-        carry = actions[-1]
-        for i in reversed(range(actions.shape[0])):
-            if in_between_mask[i]:
-                carry = carry
-            else:
-                carry = float(open_mask[i])
-            new_actions[i] = carry
-    """
-    open_mask, closed_mask = actions > 0.95, actions < 0.05
-    in_between_mask = tf.logical_not(tf.logical_or(open_mask, closed_mask))
-    is_open_float = tf.cast(open_mask, tf.float32)
-
-    def scan_fn(carry, i):
-        return tf.cond(in_between_mask[i], lambda: tf.cast(carry, tf.float32), lambda: is_open_float[i])
-
-    return tf.scan(scan_fn, tf.range(tf.shape(actions)[0]), actions[-1], reverse=True)
-
-
-def invert_gripper_actions(actions: tf.Tensor) -> tf.Tensor:
-    return 1 - actions
-
-
-def rel2abs_gripper_actions(actions: tf.Tensor) -> tf.Tensor:
-    """
-    Converts relative gripper actions (+1 for closing, -1 for opening) to absolute actions (0 = closed; 1 = open).
-
-    Assumes that the first relative gripper is not redundant (i.e. close when already closed)!
-    """
-    # Note =>> -1 for closing, 1 for opening, 0 for no change
-    opening_mask, closing_mask = actions < -0.1, actions > 0.1
-    thresholded_actions = tf.where(opening_mask, 1, tf.where(closing_mask, -1, 0))
-
-    def scan_fn(carry, i):
-        return tf.cond(thresholded_actions[i] == 0, lambda: carry, lambda: thresholded_actions[i])
-
-    # If no relative grasp, assumes open for whole trajectory
-    start = -1 * thresholded_actions[tf.argmax(thresholded_actions != 0, axis=0)]
-    start = tf.cond(start == 0, lambda: 1, lambda: start)
-
-    # Note =>> -1 for closed, 1 for open
-    new_actions = tf.scan(scan_fn, tf.range(tf.shape(actions)[0]), start)
-    new_actions = tf.cast(new_actions, tf.float32) / 2 + 0.5
-
-    return new_actions
-
-
-# === Bridge-V2 =>> Dataset-Specific Transform ===
-def relabel_bridge_actions(traj: Dict[str, Any]) -> Dict[str, Any]:
-    """Relabels actions to use reached proprioceptive state; discards last timestep (no-action)."""
-    movement_actions = traj["observation"]["state"][1:, :6] - traj["observation"]["state"][:-1, :6]
-    traj_truncated = tf.nest.map_structure(lambda x: x[:-1], traj)
-    traj_truncated["action"] = tf.concat([movement_actions, traj["action"][:-1, -1:]], axis=1)
-
-    return traj_truncated
-
-
-# === RLDS Dataset Initialization Utilities ===
-def pprint_data_mixture(dataset_kwargs_list: List[Dict[str, Any]], dataset_weights: List[int]) -> None:
-    print("\n######################################################################################")
-    print(f"# Loading the following {len(dataset_kwargs_list)} datasets (incl. sampling weight):{'': >24} #")
-    for dataset_kwargs, weight in zip(dataset_kwargs_list, dataset_weights, strict=False):
-        pad = 80 - len(dataset_kwargs["name"])
-        print(f"# {dataset_kwargs['name']}: {weight:=>{pad}f} #")
-    print("######################################################################################\n")

lerobot/common/datasets/push_dataset_to_hub/openx/droid_utils.py

@@ -1,200 +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.
-"""
-NOTE(YL): Adapted from:
-    OpenVLA: https://github.com/openvla/openvla
-
-Episode transforms for DROID dataset.
-"""
-
-from typing import Any, Dict
-
-import tensorflow as tf
-import tensorflow_graphics.geometry.transformation as tfg
-
-
-def rmat_to_euler(rot_mat):
-    return tfg.euler.from_rotation_matrix(rot_mat)
-
-
-def euler_to_rmat(euler):
-    return tfg.rotation_matrix_3d.from_euler(euler)
-
-
-def invert_rmat(rot_mat):
-    return tfg.rotation_matrix_3d.inverse(rot_mat)
-
-
-def rotmat_to_rot6d(mat):
-    """
-    Converts rotation matrix to R6 rotation representation (first two rows in rotation matrix).
-    Args:
-        mat: rotation matrix
-
-    Returns: 6d vector (first two rows of rotation matrix)
-
-    """
-    r6 = mat[..., :2, :]
-    r6_0, r6_1 = r6[..., 0, :], r6[..., 1, :]
-    r6_flat = tf.concat([r6_0, r6_1], axis=-1)
-    return r6_flat
-
-
-def velocity_act_to_wrist_frame(velocity, wrist_in_robot_frame):
-    """
-    Translates velocity actions (translation + rotation) from base frame of the robot to wrist frame.
-    Args:
-        velocity: 6d velocity action (3 x translation, 3 x rotation)
-        wrist_in_robot_frame: 6d pose of the end-effector in robot base frame
-
-    Returns: 9d velocity action in robot wrist frame (3 x translation, 6 x rotation as R6)
-
-    """
-    r_frame = euler_to_rmat(wrist_in_robot_frame[:, 3:6])
-    r_frame_inv = invert_rmat(r_frame)
-
-    # world to wrist: dT_pi = R^-1 dT_rbt
-    vel_t = (r_frame_inv @ velocity[:, :3][..., None])[..., 0]
-
-    # world to wrist: dR_pi = R^-1 dR_rbt R
-    dr_ = euler_to_rmat(velocity[:, 3:6])
-    dr_ = r_frame_inv @ (dr_ @ r_frame)
-    dr_r6 = rotmat_to_rot6d(dr_)
-    return tf.concat([vel_t, dr_r6], axis=-1)
-
-
-def rand_swap_exterior_images(img1, img2):
-    """
-    Randomly swaps the two exterior images (for training with single exterior input).
-    """
-    return tf.cond(tf.random.uniform(shape=[]) > 0.5, lambda: (img1, img2), lambda: (img2, img1))
-
-
-def droid_baseact_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    """
-    DROID dataset transformation for actions expressed in *base* frame of the robot.
-    """
-    dt = trajectory["action_dict"]["cartesian_velocity"][:, :3]
-    dr_ = trajectory["action_dict"]["cartesian_velocity"][:, 3:6]
-
-    trajectory["action"] = tf.concat(
-        (
-            dt,
-            dr_,
-            1 - trajectory["action_dict"]["gripper_position"],
-        ),
-        axis=-1,
-    )
-    trajectory["observation"]["exterior_image_1_left"], trajectory["observation"]["exterior_image_2_left"] = (
-        rand_swap_exterior_images(
-            trajectory["observation"]["exterior_image_1_left"],
-            trajectory["observation"]["exterior_image_2_left"],
-        )
-    )
-    trajectory["observation"]["proprio"] = tf.concat(
-        (
-            trajectory["observation"]["cartesian_position"],
-            trajectory["observation"]["gripper_position"],
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def droid_wristact_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    """
-    DROID dataset transformation for actions expressed in *wrist* frame of the robot.
-    """
-    wrist_act = velocity_act_to_wrist_frame(
-        trajectory["action_dict"]["cartesian_velocity"], trajectory["observation"]["cartesian_position"]
-    )
-    trajectory["action"] = tf.concat(
-        (
-            wrist_act,
-            trajectory["action_dict"]["gripper_position"],
-        ),
-        axis=-1,
-    )
-    trajectory["observation"]["exterior_image_1_left"], trajectory["observation"]["exterior_image_2_left"] = (
-        rand_swap_exterior_images(
-            trajectory["observation"]["exterior_image_1_left"],
-            trajectory["observation"]["exterior_image_2_left"],
-        )
-    )
-    trajectory["observation"]["proprio"] = tf.concat(
-        (
-            trajectory["observation"]["cartesian_position"],
-            trajectory["observation"]["gripper_position"],
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def droid_finetuning_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    """
-    DROID dataset transformation for actions expressed in *base* frame of the robot.
-    """
-    dt = trajectory["action_dict"]["cartesian_velocity"][:, :3]
-    dr_ = trajectory["action_dict"]["cartesian_velocity"][:, 3:6]
-    trajectory["action"] = tf.concat(
-        (
-            dt,
-            dr_,
-            1 - trajectory["action_dict"]["gripper_position"],
-        ),
-        axis=-1,
-    )
-    trajectory["observation"]["proprio"] = tf.concat(
-        (
-            trajectory["observation"]["cartesian_position"],
-            trajectory["observation"]["gripper_position"],
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def zero_action_filter(traj: Dict) -> bool:
-    """
-    Filters transitions whose actions are all-0 (only relative actions, no gripper action).
-    Note: this filter is applied *after* action normalization, so need to compare to "normalized 0".
-    """
-    droid_q01 = tf.convert_to_tensor(
-        [
-            -0.7776297926902771,
-            -0.5803514122962952,
-            -0.5795090794563293,
-            -0.6464047729969025,
-            -0.7041108310222626,
-            -0.8895104378461838,
-        ]
-    )
-    droid_q99 = tf.convert_to_tensor(
-        [
-            0.7597932070493698,
-            0.5726242214441299,
-            0.7351000607013702,
-            0.6705610305070877,
-            0.6464948207139969,
-            0.8897542208433151,
-        ]
-    )
-    droid_norm_0_act = (
-        2 * (tf.zeros_like(traj["action"][:, :6]) - droid_q01) / (droid_q99 - droid_q01 + 1e-8) - 1
-    )
-
-    return tf.reduce_any(tf.math.abs(traj["action"][:, :6] - droid_norm_0_act) > 1e-5)

lerobot/common/datasets/push_dataset_to_hub/openx/transforms.py

@@ -1,859 +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.
-"""
-NOTE(YL): Adapted from:
-    OpenVLA: https://github.com/openvla/openvla
-    Octo: https://github.com/octo-models/octo
-
-transforms.py
-
-Defines a registry of per-dataset standardization transforms for each dataset in Open-X Embodiment.
-
-Transforms adopt the following structure:
-    Input: Dictionary of *batched* features (i.e., has leading time dimension)
-    Output: Dictionary `step` =>> {
-        "observation": {
-            <image_keys, depth_image_keys>
-            State (in chosen state representation)
-        },
-        "action": Action (in chosen action representation),
-        "language_instruction": str
-    }
-"""
-
-from typing import Any, Dict
-
-import tensorflow as tf
-
-from lerobot.common.datasets.push_dataset_to_hub.openx.data_utils import (
-    binarize_gripper_actions,
-    invert_gripper_actions,
-    rel2abs_gripper_actions,
-    relabel_bridge_actions,
-)
-
-
-def droid_baseact_transform_fn():
-    from lerobot.common.datasets.push_dataset_to_hub.openx.droid_utils import droid_baseact_transform
-
-    return droid_baseact_transform
-
-
-def bridge_openx_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    """
-    Applies to version of Bridge V2 in Open X-Embodiment mixture.
-
-    Note =>> In original Bridge V2 dataset, the first timestep has an all-zero action, so we remove it!
-    """
-    for key in trajectory:
-        if key == "traj_metadata":
-            continue
-        elif key in ["observation", "action"]:
-            for key2 in trajectory[key]:
-                trajectory[key][key2] = trajectory[key][key2][1:]
-        else:
-            trajectory[key] = trajectory[key][1:]
-
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"]["world_vector"],
-            trajectory["action"]["rotation_delta"],
-            tf.cast(trajectory["action"]["open_gripper"][:, None], tf.float32),
-        ),
-        axis=-1,
-    )
-    trajectory["language_instruction"] = trajectory["observation"]["natural_language_instruction"]
-    trajectory = relabel_bridge_actions(trajectory)
-    trajectory["observation"]["EEF_state"] = trajectory["observation"]["state"][:, :6]
-    trajectory["observation"]["gripper_state"] = trajectory["observation"]["state"][:, -1:]
-    return trajectory
-
-
-def bridge_orig_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    """
-    Applies to original version of Bridge V2 from the official project website.
-
-    Note =>> In original Bridge V2 dataset, the first timestep has an all-zero action, so we remove it!
-    """
-    for key in trajectory:
-        if key == "traj_metadata":
-            continue
-        elif key == "observation":
-            for key2 in trajectory[key]:
-                trajectory[key][key2] = trajectory[key][key2][1:]
-        else:
-            trajectory[key] = trajectory[key][1:]
-
-    trajectory["action"] = tf.concat(
-        [
-            trajectory["action"][:, :6],
-            binarize_gripper_actions(trajectory["action"][:, -1])[:, None],
-        ],
-        axis=1,
-    )
-    trajectory = relabel_bridge_actions(trajectory)
-    trajectory["observation"]["EEF_state"] = trajectory["observation"]["state"][:, :6]
-    trajectory["observation"]["gripper_state"] = trajectory["observation"]["state"][:, -1:]
-    return trajectory
-
-
-def ppgm_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["action"] = tf.concat(
-        [
-            trajectory["action"][:, :6],
-            binarize_gripper_actions(trajectory["action"][:, -1])[:, None],
-        ],
-        axis=1,
-    )
-    trajectory["observation"]["EEF_state"] = trajectory["observation"]["cartesian_position"][:, :6]
-    trajectory["observation"]["gripper_state"] = trajectory["observation"]["gripper_position"][:, -1:]
-    return trajectory
-
-
-def rt1_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    # make gripper action absolute action, +1 = open, 0 = close
-    gripper_action = trajectory["action"]["gripper_closedness_action"][:, 0]
-    gripper_action = rel2abs_gripper_actions(gripper_action)
-
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"]["world_vector"],
-            trajectory["action"]["rotation_delta"],
-            gripper_action[:, None],
-        ),
-        axis=-1,
-    )
-    trajectory["language_instruction"] = trajectory["observation"]["natural_language_instruction"]
-    return trajectory
-
-
-def kuka_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    # make gripper action absolute action, +1 = open, 0 = close
-    gripper_action = trajectory["action"]["gripper_closedness_action"][:, 0]
-    gripper_action = rel2abs_gripper_actions(gripper_action)
-
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"]["world_vector"],
-            trajectory["action"]["rotation_delta"],
-            gripper_action[:, None],
-        ),
-        axis=-1,
-    )
-    # decode compressed state
-    eef_value = tf.io.decode_compressed(
-        trajectory["observation"]["clip_function_input/base_pose_tool_reached"],
-        compression_type="ZLIB",
-    )
-    eef_value = tf.io.decode_raw(eef_value, tf.float32)
-    trajectory["observation"]["clip_function_input/base_pose_tool_reached"] = tf.reshape(eef_value, (-1, 7))
-    gripper_value = tf.io.decode_compressed(
-        trajectory["observation"]["gripper_closed"], compression_type="ZLIB"
-    )
-    gripper_value = tf.io.decode_raw(gripper_value, tf.float32)
-    trajectory["observation"]["gripper_closed"] = tf.reshape(gripper_value, (-1, 1))
-    trajectory["language_instruction"] = trajectory["observation"]["natural_language_instruction"]
-    return trajectory
-
-
-def taco_play_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["state_eef"] = trajectory["observation"]["robot_obs"][:, :6]
-    trajectory["observation"]["state_gripper"] = trajectory["observation"]["robot_obs"][:, 7:8]
-    trajectory["action"] = trajectory["action"]["rel_actions_world"]
-
-    # invert gripper action + clip, +1 = open, 0 = close
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"][:, :6],
-            tf.clip_by_value(trajectory["action"][:, -1:], 0, 1),
-        ),
-        axis=-1,
-    )
-
-    trajectory["language_instruction"] = trajectory["observation"]["natural_language_instruction"]
-    return trajectory
-
-
-def jaco_play_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["state_eef"] = trajectory["observation"]["end_effector_cartesian_pos"][:, :6]
-    trajectory["observation"]["state_gripper"] = trajectory["observation"]["end_effector_cartesian_pos"][
-        :, -1:
-    ]
-
-    # make gripper action absolute action, +1 = open, 0 = close
-    gripper_action = trajectory["action"]["gripper_closedness_action"][:, 0]
-    gripper_action = rel2abs_gripper_actions(gripper_action)
-
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"]["world_vector"],
-            tf.zeros_like(trajectory["action"]["world_vector"]),
-            gripper_action[:, None],
-        ),
-        axis=-1,
-    )
-    trajectory["language_instruction"] = trajectory["observation"]["natural_language_instruction"]
-    return trajectory
-
-
-def berkeley_cable_routing_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"]["world_vector"],
-            trajectory["action"]["rotation_delta"],
-            tf.zeros_like(trajectory["action"]["world_vector"][:, :1]),
-        ),
-        axis=-1,
-    )
-    trajectory["language_instruction"] = trajectory["observation"]["natural_language_instruction"]
-    return trajectory
-
-
-def roboturk_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    # invert absolute gripper action, +1 = open, 0 = close
-    gripper_action = invert_gripper_actions(
-        tf.clip_by_value(trajectory["action"]["gripper_closedness_action"], 0, 1)
-    )
-
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"]["world_vector"],
-            trajectory["action"]["rotation_delta"],
-            gripper_action,
-        ),
-        axis=-1,
-    )
-    trajectory["language_instruction"] = trajectory["observation"]["natural_language_instruction"]
-    trajectory["language_embedding"] = trajectory["observation"]["natural_language_embedding"]
-    return trajectory
-
-
-def nyu_door_opening_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    # make gripper action absolute action, +1 = open, 0 = close
-    gripper_action = trajectory["action"]["gripper_closedness_action"][:, 0]
-    gripper_action = rel2abs_gripper_actions(gripper_action)
-
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"]["world_vector"],
-            trajectory["action"]["rotation_delta"],
-            gripper_action[:, None],
-        ),
-        axis=-1,
-    )
-    trajectory["language_instruction"] = trajectory["observation"]["natural_language_instruction"]
-    return trajectory
-
-
-def viola_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    # make gripper action, +1 = open, 0 = close
-    gripper_action = trajectory["action"]["gripper_closedness_action"][:, None]
-    gripper_action = tf.clip_by_value(gripper_action, 0, 1)
-    gripper_action = invert_gripper_actions(gripper_action)
-
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"]["world_vector"],
-            trajectory["action"]["rotation_delta"],
-            gripper_action,
-        ),
-        axis=-1,
-    )
-    trajectory["language_instruction"] = trajectory["observation"]["natural_language_instruction"]
-    return trajectory
-
-
-def berkeley_autolab_ur5_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["state"] = trajectory["observation"]["robot_state"][:, 6:14]
-
-    # make gripper action absolute action, +1 = open, 0 = close
-    gripper_action = trajectory["action"]["gripper_closedness_action"]
-    gripper_action = rel2abs_gripper_actions(gripper_action)
-
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"]["world_vector"],
-            trajectory["action"]["rotation_delta"],
-            gripper_action[:, None],
-        ),
-        axis=-1,
-    )
-    trajectory["language_instruction"] = trajectory["observation"]["natural_language_instruction"]
-    return trajectory
-
-
-def toto_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"]["world_vector"],
-            trajectory["action"]["rotation_delta"],
-            tf.cast(trajectory["action"]["open_gripper"][:, None], tf.float32),
-        ),
-        axis=-1,
-    )
-    trajectory["language_instruction"] = trajectory["observation"]["natural_language_instruction"]
-    return trajectory
-
-
-def language_table_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    # default to "open" gripper
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"],
-            tf.zeros_like(trajectory["action"]),
-            tf.zeros_like(trajectory["action"]),
-            tf.ones_like(trajectory["action"][:, :1]),
-        ),
-        axis=-1,
-    )
-
-    # decode language instruction
-    instruction_bytes = trajectory["observation"]["instruction"]
-    instruction_encoded = tf.strings.unicode_encode(instruction_bytes, output_encoding="UTF-8")
-    # Remove trailing padding --> convert RaggedTensor to regular Tensor.
-    trajectory["language_instruction"] = tf.strings.split(instruction_encoded, "\x00")[:, :1].to_tensor()[
-        :, 0
-    ]
-    return trajectory
-
-
-def pusht_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"]["world_vector"],
-            trajectory["action"]["rotation_delta"],
-            trajectory["action"]["gripper_closedness_action"][:, None],
-        ),
-        axis=-1,
-    )
-    trajectory["language_instruction"] = trajectory["observation"]["natural_language_instruction"]
-    return trajectory
-
-
-def stanford_kuka_multimodal_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["depth_image"] = trajectory["observation"]["depth_image"][..., 0]
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"][:, :3],
-            tf.zeros_like(trajectory["action"][:, :3]),
-            trajectory["action"][:, -1:],
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def nyu_rot_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["eef_state"] = trajectory["observation"]["state"][..., :6]
-    trajectory["observation"]["gripper_state"] = trajectory["observation"]["state"][..., -1:]
-    trajectory["action"] = trajectory["action"][..., :7]
-    return trajectory
-
-
-def stanford_hydra_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    # invert gripper action, +1 = open, 0 = close
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"][:, :6],
-            invert_gripper_actions(trajectory["action"][:, -1:]),
-        ),
-        axis=-1,
-    )
-
-    trajectory["observation"]["eef_state"] = tf.concat(
-        (
-            trajectory["observation"]["state"][:, :3],
-            trajectory["observation"]["state"][:, 7:10],
-        ),
-        axis=-1,
-    )
-    trajectory["observation"]["gripper_state"] = trajectory["observation"]["state"][:, -3:-2]
-    return trajectory
-
-
-def austin_buds_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    # invert gripper action + clip, +1 = open, 0 = close
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"][:, :6],
-            invert_gripper_actions(tf.clip_by_value(trajectory["action"][:, -1:], 0, 1)),
-        ),
-        axis=-1,
-    )
-
-    trajectory["observation"]["state"] = trajectory["observation"]["state"][:, :8]
-    return trajectory
-
-
-def nyu_franka_play_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["depth"] = tf.cast(trajectory["observation"]["depth"][..., 0], tf.float32)
-    trajectory["observation"]["depth_additional_view"] = tf.cast(
-        trajectory["observation"]["depth_additional_view"][..., 0], tf.float32
-    )
-    trajectory["observation"]["eef_state"] = trajectory["observation"]["state"][:, -6:]
-
-    # clip gripper action, +1 = open, 0 = close
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"][:, -8:-2],
-            tf.clip_by_value(trajectory["action"][:, -2:-1], 0, 1),
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def maniskill_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["gripper_state"] = trajectory["observation"]["state"][..., 7:8]
-    return trajectory
-
-
-def furniture_bench_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    import tensorflow_graphics.geometry.transformation as tft
-
-    trajectory["observation"]["state"] = tf.concat(
-        (
-            trajectory["observation"]["state"][:, :7],
-            trajectory["observation"]["state"][:, -1:],
-        ),
-        axis=-1,
-    )
-
-    # invert gripper action + clip, +1 = open, 0 = close
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"][:, :3],
-            tft.euler.from_quaternion(trajectory["action"][:, 3:7]),
-            invert_gripper_actions(tf.clip_by_value(trajectory["action"][:, -1:], 0, 1)),
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def cmu_franka_exploration_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["action"] = trajectory["action"][..., :-1]
-    return trajectory
-
-
-def ucsd_kitchen_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["joint_state"] = trajectory["observation"]["state"][:, :7]
-    trajectory["action"] = trajectory["action"][..., :-1]
-    return trajectory
-
-
-def ucsd_pick_place_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["eef_state"] = trajectory["observation"]["state"][:, :6]
-    trajectory["observation"]["gripper_state"] = trajectory["observation"]["state"][:, -1:]
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"][:, :3],
-            tf.zeros_like(trajectory["action"][:, :3]),
-            trajectory["action"][:, -1:],
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def austin_sailor_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    # invert gripper action + clip, +1 = open, 0 = close
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"][:, :6],
-            invert_gripper_actions(tf.clip_by_value(trajectory["action"][:, -1:], 0, 1)),
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def austin_sirius_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    # invert gripper action + clip, +1 = open, 0 = close
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"][:, :6],
-            invert_gripper_actions(tf.clip_by_value(trajectory["action"][:, -1:], 0, 1)),
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def bc_z_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"]["future/xyz_residual"][:, :3],
-            trajectory["action"]["future/axis_angle_residual"][:, :3],
-            invert_gripper_actions(tf.cast(trajectory["action"]["future/target_close"][:, :1], tf.float32)),
-        ),
-        axis=-1,
-    )
-    trajectory["language_instruction"] = trajectory["observation"]["natural_language_instruction"]
-    return trajectory
-
-
-def tokyo_pr2_opening_fridge_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["eef_state"] = trajectory["observation"]["state"][:, :6]
-    trajectory["observation"]["gripper_state"] = trajectory["observation"]["state"][:, -1:]
-    trajectory["action"] = trajectory["action"][..., :-1]
-    return trajectory
-
-
-def tokyo_pr2_tabletop_manipulation_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["eef_state"] = trajectory["observation"]["state"][:, :6]
-    trajectory["observation"]["gripper_state"] = trajectory["observation"]["state"][:, -1:]
-    trajectory["action"] = trajectory["action"][..., :-1]
-    return trajectory
-
-
-def utokyo_xarm_bimanual_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["action"] = trajectory["action"][..., -7:]
-    return trajectory
-
-
-def robo_net_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["eef_state"] = tf.concat(
-        (
-            trajectory["observation"]["state"][:, :4],
-            tf.zeros_like(trajectory["observation"]["state"][:, :2]),
-        ),
-        axis=-1,
-    )
-    trajectory["observation"]["gripper_state"] = trajectory["observation"]["state"][:, -1:]
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"][:, :4],
-            tf.zeros_like(trajectory["action"][:, :2]),
-            trajectory["action"][:, -1:],
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def berkeley_mvp_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    """
-    trajectory["observation"]["state"] = tf.concat((
-        tf.cast(trajectory["observation"]["gripper"][:, None], tf.float32),
-                        trajectory["observation"]["pose"],
-                        trajectory["observation"]["joint_pos"],),
-                        axis=-1,)
-    """
-    trajectory["observation"]["gripper"] = tf.cast(trajectory["observation"]["gripper"][:, None], tf.float32)
-    return trajectory
-
-
-def berkeley_rpt_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["gripper"] = tf.cast(trajectory["observation"]["gripper"][:, None], tf.float32)
-    return trajectory
-
-
-def kaist_nonprehensible_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["state"] = trajectory["observation"]["state"][:, -7:]
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"][:, :6],
-            tf.zeros_like(trajectory["action"][:, :1]),
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def stanford_mask_vit_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["eef_state"] = tf.concat(
-        (
-            trajectory["observation"]["end_effector_pose"][:, :4],
-            tf.zeros_like(trajectory["observation"]["end_effector_pose"][:, :2]),
-        ),
-        axis=-1,
-    )
-    trajectory["observation"]["gripper_state"] = trajectory["observation"]["end_effector_pose"][:, -1:]
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"][:, :4],
-            tf.zeros_like(trajectory["action"][:, :2]),
-            trajectory["action"][:, -1:],
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def tokyo_lsmo_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["eef_state"] = trajectory["observation"]["state"][:, :6]
-    trajectory["observation"]["gripper_state"] = trajectory["observation"]["state"][:, -1:]
-    return trajectory
-
-
-def dlr_sara_grid_clamp_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["state"] = trajectory["observation"]["state"][:, :6]
-    return trajectory
-
-
-def dlr_edan_shared_control_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    # invert gripper action, +1 = open, 0 = close
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"][:, :6],
-            invert_gripper_actions(trajectory["action"][:, -1:]),
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def asu_table_top_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["eef_state"] = trajectory["ground_truth_states"]["EE"]
-    trajectory["observation"]["gripper_state"] = trajectory["observation"]["state"][:, -1:]
-    return trajectory
-
-
-def robocook_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["eef_state"] = trajectory["observation"]["state"][:, :6]
-    trajectory["observation"]["gripper_state"] = trajectory["observation"]["state"][:, -1:]
-    return trajectory
-
-
-def imperial_wristcam_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["action"] = trajectory["action"][..., :-1]
-    return trajectory
-
-
-def iamlab_pick_insert_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    import tensorflow_graphics.geometry.transformation as tft
-
-    trajectory["observation"]["joint_state"] = trajectory["observation"]["state"][:, :7]
-    trajectory["observation"]["gripper_state"] = trajectory["observation"]["state"][:, 7:8]
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"][:, :3],
-            tft.euler.from_quaternion(trajectory["action"][:, 3:7]),
-            trajectory["action"][:, 7:8],
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def uiuc_d3field_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"],
-            tf.zeros_like(trajectory["action"]),
-            tf.zeros_like(trajectory["action"][:, :1]),
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def utaustin_mutex_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["state"] = trajectory["observation"]["state"][:, :8]
-
-    # invert gripper action + clip, +1 = open, 0 = close
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"][:, :6],
-            invert_gripper_actions(tf.clip_by_value(trajectory["action"][:, -1:], 0, 1)),
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def berkeley_fanuc_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["joint_state"] = trajectory["observation"]["state"][:, :6]
-    trajectory["observation"]["gripper_state"] = trajectory["observation"]["state"][:, 6:7]
-
-    # dataset does not store gripper actions, so use gripper state info, invert so +1 = open, 0 = close
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"],
-            invert_gripper_actions(trajectory["observation"]["gripper_state"]),
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def cmu_playing_with_food_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    import tensorflow_graphics.geometry.transformation as tft
-
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"][:, :3],
-            tft.euler.from_quaternion(trajectory["action"][:, 3:7]),
-            trajectory["action"][:, -1:],
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def playfusion_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"][:, :3],
-            trajectory["action"][:, -4:],
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def cmu_stretch_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["eef_state"] = tf.concat(
-        (
-            trajectory["observation"]["state"][:, :3],
-            tf.zeros_like(trajectory["observation"]["state"][:, :3]),
-        ),
-        axis=-1,
-    )
-    trajectory["observation"]["gripper_state"] = trajectory["observation"]["state"][:, -1:]
-    trajectory["action"] = trajectory["action"][..., :-1]
-    return trajectory
-
-
-def gnm_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["state"] = tf.concat(
-        (
-            trajectory["observation"]["position"],
-            tf.zeros_like(trajectory["observation"]["state"][:, :3]),
-            trajectory["observation"]["yaw"],
-        ),
-        axis=-1,
-    )
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"],
-            tf.zeros_like(trajectory["action"]),
-            tf.zeros_like(trajectory["action"]),
-            tf.zeros_like(trajectory["action"][:, :1]),
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def fmb_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    # every input feature is batched, ie has leading batch dimension
-    trajectory["observation"]["proprio"] = tf.concat(
-        (
-            trajectory["observation"]["eef_pose"],
-            trajectory["observation"]["state_gripper_pose"][..., None],
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def dobbe_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    # every input feature is batched, ie has leading batch dimension
-    trajectory["observation"]["proprio"] = trajectory["observation"]["state"]
-    return trajectory
-
-
-def robo_set_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    # gripper action is in -1...1 --> clip to 0...1, flip
-    gripper_action = trajectory["action"][:, -1:]
-    gripper_action = invert_gripper_actions(tf.clip_by_value(gripper_action, 0, 1))
-
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"][:, :7],
-            gripper_action,
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def identity_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    return trajectory
-
-
-# === Registry ===
-OPENX_STANDARDIZATION_TRANSFORMS = {
-    "bridge_openx": bridge_openx_dataset_transform,
-    "bridge_orig": bridge_orig_dataset_transform,
-    "bridge_dataset": bridge_orig_dataset_transform,
-    "ppgm": ppgm_dataset_transform,
-    "ppgm_static": ppgm_dataset_transform,
-    "ppgm_wrist": ppgm_dataset_transform,
-    "fractal20220817_data": rt1_dataset_transform,
-    "kuka": kuka_dataset_transform,
-    "taco_play": taco_play_dataset_transform,
-    "jaco_play": jaco_play_dataset_transform,
-    "berkeley_cable_routing": berkeley_cable_routing_dataset_transform,
-    "roboturk": roboturk_dataset_transform,
-    "nyu_door_opening_surprising_effectiveness": nyu_door_opening_dataset_transform,
-    "viola": viola_dataset_transform,
-    "berkeley_autolab_ur5": berkeley_autolab_ur5_dataset_transform,
-    "toto": toto_dataset_transform,
-    "language_table": language_table_dataset_transform,
-    "columbia_cairlab_pusht_real": pusht_dataset_transform,
-    "stanford_kuka_multimodal_dataset_converted_externally_to_rlds": stanford_kuka_multimodal_dataset_transform,
-    "nyu_rot_dataset_converted_externally_to_rlds": nyu_rot_dataset_transform,
-    "stanford_hydra_dataset_converted_externally_to_rlds": stanford_hydra_dataset_transform,
-    "austin_buds_dataset_converted_externally_to_rlds": austin_buds_dataset_transform,
-    "nyu_franka_play_dataset_converted_externally_to_rlds": nyu_franka_play_dataset_transform,
-    "maniskill_dataset_converted_externally_to_rlds": maniskill_dataset_transform,
-    "furniture_bench_dataset_converted_externally_to_rlds": furniture_bench_dataset_transform,
-    "cmu_franka_exploration_dataset_converted_externally_to_rlds": cmu_franka_exploration_dataset_transform,
-    "ucsd_kitchen_dataset_converted_externally_to_rlds": ucsd_kitchen_dataset_transform,
-    "ucsd_pick_and_place_dataset_converted_externally_to_rlds": ucsd_pick_place_dataset_transform,
-    "austin_sailor_dataset_converted_externally_to_rlds": austin_sailor_dataset_transform,
-    "austin_sirius_dataset_converted_externally_to_rlds": austin_sirius_dataset_transform,
-    "bc_z": bc_z_dataset_transform,
-    "utokyo_pr2_opening_fridge_converted_externally_to_rlds": tokyo_pr2_opening_fridge_dataset_transform,
-    "utokyo_pr2_tabletop_manipulation_converted_externally_to_rlds": tokyo_pr2_tabletop_manipulation_dataset_transform,
-    "utokyo_xarm_pick_and_place_converted_externally_to_rlds": identity_transform,
-    "utokyo_xarm_bimanual_converted_externally_to_rlds": utokyo_xarm_bimanual_dataset_transform,
-    "robo_net": robo_net_dataset_transform,
-    "berkeley_mvp_converted_externally_to_rlds": berkeley_mvp_dataset_transform,
-    "berkeley_rpt_converted_externally_to_rlds": berkeley_rpt_dataset_transform,
-    "kaist_nonprehensile_converted_externally_to_rlds": kaist_nonprehensible_dataset_transform,
-    "stanford_mask_vit_converted_externally_to_rlds": stanford_mask_vit_dataset_transform,
-    "tokyo_u_lsmo_converted_externally_to_rlds": tokyo_lsmo_dataset_transform,
-    "dlr_sara_pour_converted_externally_to_rlds": identity_transform,
-    "dlr_sara_grid_clamp_converted_externally_to_rlds": dlr_sara_grid_clamp_dataset_transform,
-    "dlr_edan_shared_control_converted_externally_to_rlds": dlr_edan_shared_control_dataset_transform,
-    "asu_table_top_converted_externally_to_rlds": asu_table_top_dataset_transform,
-    "stanford_robocook_converted_externally_to_rlds": robocook_dataset_transform,
-    "imperialcollege_sawyer_wrist_cam": imperial_wristcam_dataset_transform,
-    "iamlab_cmu_pickup_insert_converted_externally_to_rlds": iamlab_pick_insert_dataset_transform,
-    "uiuc_d3field": uiuc_d3field_dataset_transform,
-    "utaustin_mutex": utaustin_mutex_dataset_transform,
-    "berkeley_fanuc_manipulation": berkeley_fanuc_dataset_transform,
-    "cmu_playing_with_food": cmu_playing_with_food_dataset_transform,
-    "cmu_play_fusion": playfusion_dataset_transform,
-    "cmu_stretch": cmu_stretch_dataset_transform,
-    "berkeley_gnm_recon": gnm_dataset_transform,
-    "berkeley_gnm_cory_hall": gnm_dataset_transform,
-    "berkeley_gnm_sac_son": gnm_dataset_transform,
-    "droid": droid_baseact_transform_fn(),
-    "droid_100": droid_baseact_transform_fn(),  # first 100 episodes of droid
-    "fmb": fmb_transform,
-    "dobbe": dobbe_dataset_transform,
-    "robo_set": robo_set_dataset_transform,
-    "usc_cloth_sim_converted_externally_to_rlds": identity_transform,
-    "plex_robosuite": identity_transform,
-    "conq_hose_manipulation": identity_transform,
-    "io_ai_tech": identity_transform,
-    "spoc": identity_transform,
-}

lerobot/common/datasets/push_dataset_to_hub/openx_rlds_format.py

@@ -14,13 +14,16 @@
 # 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_datsets before running this script.
+
 Example:
     python lerobot/scripts/push_dataset_to_hub.py \
-        --raw-dir /hdd/tensorflow_datasets/bridge_dataset/1.0.0/ \
-        --repo-id youliangtan/sampled_bridge_data_v2 \
-        --raw-format openx_rlds.bridge_orig \
+        --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:
@@ -35,28 +38,21 @@ import tensorflow as tf
 import tensorflow_datasets as tfds
 import torch
 import tqdm
-import yaml
 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.openx.transforms import OPENX_STANDARDIZATION_TRANSFORMS
 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 (
-    calculate_episode_data_index,
     hf_transform_to_torch,
 )
 from lerobot.common.datasets.video_utils import VideoFrame, encode_video_frames
 
-with open("lerobot/common/datasets/push_dataset_to_hub/openx/configs.yaml") as f:
-    _openx_list = yaml.safe_load(f)
-
-OPENX_DATASET_CONFIGS = _openx_list["OPENX_DATASET_CONFIGS"]
-
 np.set_printoptions(precision=2)
 
 
@@ -108,7 +104,6 @@ def load_from_raw(
     video: bool,
     episodes: list[int] | None = None,
     encoding: dict | None = None,
-    openx_dataset_name: str | None = None,
 ):
     """
     Args:
@@ -136,16 +131,17 @@ def load_from_raw(
     # 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
-    if openx_dataset_name is not None:
-        print(" - applying standardization transform for dataset: ", openx_dataset_name)
-        assert openx_dataset_name in OPENX_STANDARDIZATION_TRANSFORMS
-        transform_fn = OPENX_STANDARDIZATION_TRANSFORMS[openx_dataset_name]
-        dataset = dataset.map(transform_fn)
-
-        image_keys = OPENX_DATASET_CONFIGS[openx_dataset_name]["image_obs_keys"]
-    else:
-        obs_keys = dataset_info.features["steps"]["observation"].keys()
-        image_keys = [key for key in obs_keys if "image" in key]
+    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
 
@@ -193,50 +189,31 @@ def load_from_raw(
 
         num_frames = episode["action"].shape[0]
 
-        ###########################################################
-        # Handle the episodic data
-
-        # last step of demonstration is considered done
-        done = torch.zeros(num_frames, dtype=torch.bool)
-        done[-1] = True
         ep_dict = {}
-        langs = []  # TODO: might be located in "observation"
+        for key in state_keys:
+            ep_dict[f"observation.{key}"] = tf_to_torch(episode["observation"][key])
 
-        image_array_dict = {key: [] for key in image_keys}
-
-        # We will create the state observation tensor by stacking the state
-        # obs keys defined in the openx/configs.py
-        if openx_dataset_name is not None:
-            state_obs_keys = OPENX_DATASET_CONFIGS[openx_dataset_name]["state_obs_keys"]
-            # stack the state observations, if is None, pad with zeros
-            states = []
-            for key in state_obs_keys:
-                if key in episode["observation"]:
-                    states.append(tf_to_torch(episode["observation"][key]))
-                else:
-                    states.append(torch.zeros(num_frames, 1))  # pad with zeros
-            states = torch.cat(states, dim=1)
-            # assert states.shape == (num_frames, 8), f"states shape: {states.shape}"
-        else:
-            states = tf_to_torch(episode["observation"]["state"])
-
-        actions = tf_to_torch(episode["action"])
-        rewards = tf_to_torch(episode["reward"]).float()
+        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:
-            langs = [str(x) for x in episode[lang_key]]
+            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]
 
-        # simple assertions
-        for item in [states, actions, rewards, done]:
-            assert len(item) == num_frames
-
-        ###########################################################
-
         # loop through all cameras
         for im_key in image_keys:
             img_key = f"observation.images.{im_key}"
@@ -262,17 +239,6 @@ def load_from_raw(
             else:
                 ep_dict[img_key] = [PILImage.fromarray(x) for x in imgs_array]
 
-        if lang_key is not None:
-            ep_dict["language_instruction"] = langs
-
-        ep_dict["observation.state"] = states
-        ep_dict["action"] = actions
-        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)
-        ep_dict["next.reward"] = rewards
-        ep_dict["next.done"] = done
-
         path_ep_dict = tmp_ep_dicts_dir.joinpath(
             "ep_dict_" + "0" * (10 - len(str(ep_idx))) + str(ep_idx) + ".pt"
         )
@@ -290,30 +256,28 @@ def load_from_raw(
 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()
+    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()
 
-    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)
-        )
     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)
@@ -333,19 +297,8 @@ def from_raw_to_lerobot_format(
     video: bool = True,
     episodes: list[int] | None = None,
     encoding: dict | None = None,
-    openx_dataset_name: str | None = None,
 ):
-    """This is a test impl for rlds conversion"""
-    if openx_dataset_name is None:
-        # set a default rlds frame rate if the dataset is not from openx
-        fps = 30
-    elif "fps" not in OPENX_DATASET_CONFIGS[openx_dataset_name]:
-        raise ValueError(
-            "fps for this dataset is not specified in openx/configs.py yet," "means it is not yet tested"
-        )
-    fps = OPENX_DATASET_CONFIGS[openx_dataset_name]["fps"]
-
-    data_dict = load_from_raw(raw_dir, videos_dir, fps, video, episodes, encoding, openx_dataset_name)
+    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 = {

lerobot/common/datasets/push_dataset_to_hub/pusht_zarr_format.py

@@ -27,12 +27,12 @@ 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 (
-    calculate_episode_data_index,
     hf_transform_to_torch,
 )
 from lerobot.common.datasets.video_utils import VideoFrame, encode_video_frames

lerobot/common/datasets/push_dataset_to_hub/umi_zarr_format.py

@@ -28,12 +28,12 @@ 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 (
-    calculate_episode_data_index,
     hf_transform_to_torch,
 )
 from lerobot.common.datasets.video_utils import VideoFrame, encode_video_frames

lerobot/common/datasets/push_dataset_to_hub/utils.py

@@ -16,7 +16,9 @@
 import inspect
 from concurrent.futures import ThreadPoolExecutor
 from pathlib import Path
+from typing import Dict
 
+import datasets
 import numpy
 import PIL
 import torch
@@ -72,3 +74,58 @@ def check_repo_id(repo_id: str) -> None:
             f"""`repo_id` is expected to contain a community or user id `/` the name of the dataset
             (e.g. 'lerobot/pusht'), but contains '{repo_id}'."""
         )
+
+
+# TODO(aliberts): remove
+def calculate_episode_data_index(hf_dataset: datasets.Dataset) -> Dict[str, torch.Tensor]:
+    """
+    Calculate episode data index for the provided HuggingFace Dataset. Relies on episode_index column of hf_dataset.
+
+    Parameters:
+    - hf_dataset (datasets.Dataset): A HuggingFace dataset containing the episode index.
+
+    Returns:
+    - episode_data_index: A dictionary containing the data index for each episode. The dictionary has two keys:
+        - "from": A tensor containing the starting index of each episode.
+        - "to": A tensor containing the ending index of each episode.
+    """
+    episode_data_index = {"from": [], "to": []}
+
+    current_episode = None
+    """
+    The episode_index is a list of integers, each representing the episode index of the corresponding example.
+    For instance, the following is a valid episode_index:
+      [0, 0, 0, 1, 1, 1, 1, 2, 2, 2, 2, 2]
+
+    Below, we iterate through the episode_index and populate the episode_data_index dictionary with the starting and
+    ending index of each episode. For the episode_index above, the episode_data_index dictionary will look like this:
+        {
+            "from": [0, 3, 7],
+            "to": [3, 7, 12]
+        }
+    """
+    if len(hf_dataset) == 0:
+        episode_data_index = {
+            "from": torch.tensor([]),
+            "to": torch.tensor([]),
+        }
+        return episode_data_index
+    for idx, episode_idx in enumerate(hf_dataset["episode_index"]):
+        if episode_idx != current_episode:
+            # We encountered a new episode, so we append its starting location to the "from" list
+            episode_data_index["from"].append(idx)
+            # If this is not the first episode, we append the ending location of the previous episode to the "to" list
+            if current_episode is not None:
+                episode_data_index["to"].append(idx)
+            # Let's keep track of the current episode index
+            current_episode = episode_idx
+        else:
+            # We are still in the same episode, so there is nothing for us to do here
+            pass
+    # We have reached the end of the dataset, so we append the ending location of the last episode to the "to" list
+    episode_data_index["to"].append(idx + 1)
+
+    for k in ["from", "to"]:
+        episode_data_index[k] = torch.tensor(episode_data_index[k])
+
+    return episode_data_index

lerobot/common/datasets/push_dataset_to_hub/xarm_pkl_format.py

@@ -27,12 +27,12 @@ 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 (
-    calculate_episode_data_index,
     hf_transform_to_torch,
 )
 from lerobot.common.datasets.video_utils import VideoFrame, encode_video_frames

lerobot/common/datasets/utils.py

@@ -13,31 +13,60 @@
 # 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.resources
 import json
-import re
-import warnings
-from functools import cache
+import logging
+import textwrap
+from collections.abc import Iterator
+from itertools import accumulate
 from pathlib import Path
-from typing import Dict
+from pprint import pformat
+from types import SimpleNamespace
+from typing import Any
 
 import datasets
+import jsonlines
+import numpy as np
+import pyarrow.compute as pc
 import torch
-from datasets import load_dataset, load_from_disk
-from huggingface_hub import DatasetCard, HfApi, hf_hub_download, snapshot_download
+from datasets.table import embed_table_storage
+from huggingface_hub import DatasetCard, DatasetCardData, HfApi
 from PIL import Image as PILImage
-from safetensors.torch import load_file
 from torchvision import transforms
 
+from lerobot.common.robot_devices.robots.utils import Robot
+
+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"
+TASKS_PATH = "meta/tasks.jsonl"
+
+DEFAULT_VIDEO_PATH = "videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4"
+DEFAULT_PARQUET_PATH = "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet"
+DEFAULT_IMAGE_PATH = "images/{image_key}/episode_{episode_index:06d}/frame_{frame_index:06d}.png"
+
 DATASET_CARD_TEMPLATE = """
 ---
 # Metadata will go there
 ---
 This dataset was created using [LeRobot](https://github.com/huggingface/lerobot).
 
+## {}
+
 """
 
+DEFAULT_FEATURES = {
+    "timestamp": {"dtype": "float32", "shape": (1,), "names": None},
+    "frame_index": {"dtype": "int64", "shape": (1,), "names": None},
+    "episode_index": {"dtype": "int64", "shape": (1,), "names": None},
+    "index": {"dtype": "int64", "shape": (1,), "names": None},
+    "task_index": {"dtype": "int64", "shape": (1,), "names": None},
+}
 
-def flatten_dict(d, parent_key="", sep="/"):
+
+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.
 
     For example:
@@ -56,7 +85,7 @@ def flatten_dict(d, parent_key="", sep="/"):
     return dict(items)
 
 
-def unflatten_dict(d, sep="/"):
+def unflatten_dict(d: dict, sep: str = "/") -> dict:
     outdict = {}
     for key, value in d.items():
         parts = key.split(sep)
@@ -69,6 +98,82 @@ def unflatten_dict(d, sep="/"):
     return outdict
 
 
+def serialize_dict(stats: dict[str, torch.Tensor | np.ndarray | dict]) -> dict:
+    serialized_dict = {key: value.tolist() for key, value in flatten_dict(stats).items()}
+    return unflatten_dict(serialized_dict)
+
+
+def write_parquet(dataset: datasets.Dataset, fpath: Path) -> None:
+    # Embed image bytes into the table before saving to parquet
+    format = dataset.format
+    dataset = dataset.with_format("arrow")
+    dataset = dataset.map(embed_table_storage, batched=False)
+    dataset = dataset.with_format(**format)
+    dataset.to_parquet(fpath)
+
+
+def load_json(fpath: Path) -> Any:
+    with open(fpath) as f:
+        return json.load(f)
+
+
+def write_json(data: dict, fpath: Path) -> None:
+    fpath.parent.mkdir(exist_ok=True, parents=True)
+    with open(fpath, "w") as f:
+        json.dump(data, f, indent=4, ensure_ascii=False)
+
+
+def load_jsonlines(fpath: Path) -> list[Any]:
+    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 load_info(local_dir: Path) -> dict:
+    info = load_json(local_dir / INFO_PATH)
+    for ft in info["features"].values():
+        ft["shape"] = tuple(ft["shape"])
+    return info
+
+
+def load_stats(local_dir: Path) -> dict:
+    if not (local_dir / STATS_PATH).exists():
+        return None
+    stats = load_json(local_dir / STATS_PATH)
+    stats = {key: torch.tensor(value) for key, value in flatten_dict(stats).items()}
+    return unflatten_dict(stats)
+
+
+def load_tasks(local_dir: Path) -> dict:
+    tasks = load_jsonlines(local_dir / TASKS_PATH)
+    return {item["task_index"]: item["task"] for item in sorted(tasks, key=lambda x: x["task_index"])}
+
+
+def load_episodes(local_dir: Path) -> dict:
+    return load_jsonlines(local_dir / EPISODES_PATH)
+
+
+def load_image_as_numpy(fpath: str | Path, dtype="float32", channel_first: bool = True) -> np.ndarray:
+    img = PILImage.open(fpath).convert("RGB")
+    img_array = np.array(img, dtype=dtype)
+    if channel_first:  # (H, W, C) -> (C, H, W)
+        img_array = np.transpose(img_array, (2, 0, 1))
+    if "float" in dtype:
+        img_array /= 255.0
+    return img_array
+
+
 def hf_transform_to_torch(items_dict: dict[torch.Tensor | None]):
     """Get a transform function that convert items from Hugging Face dataset (pyarrow)
     to torch tensors. Importantly, images are converted from PIL, which corresponds to
@@ -80,14 +185,6 @@ def hf_transform_to_torch(items_dict: dict[torch.Tensor | None]):
         if isinstance(first_item, PILImage.Image):
             to_tensor = transforms.ToTensor()
             items_dict[key] = [to_tensor(img) for img in items_dict[key]]
-        elif isinstance(first_item, str):
-            # TODO (michel-aractingi): add str2embedding via language tokenizer
-            # For now we leave this part up to the user to choose how to address
-            # language conditioned tasks
-            pass
-        elif isinstance(first_item, dict) and "path" in first_item and "timestamp" in first_item:
-            # video frame will be processed downstream
-            pass
         elif first_item is None:
             pass
         else:
@@ -95,19 +192,67 @@ def hf_transform_to_torch(items_dict: dict[torch.Tensor | None]):
     return items_dict
 
 
-@cache
-def get_hf_dataset_safe_version(repo_id: str, version: str) -> str:
+def _get_major_minor(version: str) -> tuple[int]:
+    split = version.strip("v").split(".")
+    return int(split[0]), int(split[1])
+
+
+class BackwardCompatibilityError(Exception):
+    def __init__(self, repo_id, version):
+        message = textwrap.dedent(f"""
+            BackwardCompatibilityError: The dataset you requested ({repo_id}) is in {version} format.
+
+            We introduced a new format since v2.0 which is not backward compatible with v1.x.
+            Please, use our conversion script. Modify the following command with your own task description:
+            ```
+            python lerobot/common/datasets/v2/convert_dataset_v1_to_v2.py \\
+                --repo-id {repo_id} \\
+                --single-task "TASK DESCRIPTION."  # <---- /!\\ Replace TASK DESCRIPTION /!\\
+            ```
+
+            A few examples to replace TASK DESCRIPTION: "Pick up the blue cube and place it into the bin.",
+            "Insert the peg into the socket.", "Slide open the ziploc bag.", "Take the elevator to the 1st floor.",
+            "Open the top cabinet, store the pot inside it then close the cabinet.", "Push the T-shaped block onto the T-shaped target.",
+            "Grab the spray paint on the shelf and place it in the bin on top of the robot dog.", "Fold the sweatshirt.", ...
+
+            If you encounter a problem, contact LeRobot maintainers on [Discord](https://discord.com/invite/s3KuuzsPFb)
+            or open an [issue on GitHub](https://github.com/huggingface/lerobot/issues/new/choose).
+        """)
+        super().__init__(message)
+
+
+def check_version_compatibility(
+    repo_id: str, version_to_check: str, current_version: str, enforce_breaking_major: bool = True
+) -> None:
+    current_major, _ = _get_major_minor(current_version)
+    major_to_check, _ = _get_major_minor(version_to_check)
+    if major_to_check < current_major and enforce_breaking_major:
+        raise BackwardCompatibilityError(repo_id, version_to_check)
+    elif float(version_to_check.strip("v")) < float(current_version.strip("v")):
+        logging.warning(
+            f"""The dataset you requested ({repo_id}) was created with a previous version ({version_to_check}) of the
+            codebase. The current codebase version is {current_version}. You should be fine since
+            backward compatibility is maintained. If you encounter a problem, contact LeRobot maintainers on
+            Discord ('https://discord.com/invite/s3KuuzsPFb') or open an issue on github.""",
+        )
+
+
+def get_hub_safe_version(repo_id: str, version: str) -> str:
     api = HfApi()
     dataset_info = api.list_repo_refs(repo_id, repo_type="dataset")
     branches = [b.name for b in dataset_info.branches]
     if version not in branches:
-        warnings.warn(
+        num_version = float(version.strip("v"))
+        hub_num_versions = [float(v.strip("v")) for v in branches if v.startswith("v")]
+        if num_version >= 2.0 and all(v < 2.0 for v in hub_num_versions):
+            raise BackwardCompatibilityError(repo_id, version)
+
+        logging.warning(
             f"""You are trying to load a dataset from {repo_id} created with a previous version of the
             codebase. The following versions are available: {branches}.
             The requested version ('{version}') is not found. You should be fine since
             backward compatibility is maintained. If you encounter a problem, contact LeRobot maintainers on
             Discord ('https://discord.com/invite/s3KuuzsPFb') or open an issue on github.""",
-            stacklevel=1,
         )
         if "main" not in branches:
             raise ValueError(f"Version 'main' not found on {repo_id}")
@@ -116,275 +261,184 @@ def get_hf_dataset_safe_version(repo_id: str, version: str) -> str:
         return version
 
 
-def load_hf_dataset(repo_id: str, version: str, root: Path, split: str) -> datasets.Dataset:
-    """hf_dataset contains all the observations, states, actions, rewards, etc."""
-    if root is not None:
-        hf_dataset = load_from_disk(str(Path(root) / repo_id / "train"))
-        # TODO(rcadene): clean this which enables getting a subset of dataset
-        if split != "train":
-            if "%" in split:
-                raise NotImplementedError(f"We dont support splitting based on percentage for now ({split}).")
-            match_from = re.search(r"train\[(\d+):\]", split)
-            match_to = re.search(r"train\[:(\d+)\]", split)
-            if match_from:
-                from_frame_index = int(match_from.group(1))
-                hf_dataset = hf_dataset.select(range(from_frame_index, len(hf_dataset)))
-            elif match_to:
-                to_frame_index = int(match_to.group(1))
-                hf_dataset = hf_dataset.select(range(to_frame_index))
-            else:
-                raise ValueError(
-                    f'`split` ({split}) should either be "train", "train[INT:]", or "train[:INT]"'
-                )
-    else:
-        safe_version = get_hf_dataset_safe_version(repo_id, version)
-        hf_dataset = load_dataset(repo_id, revision=safe_version, split=split)
-
-    hf_dataset.set_transform(hf_transform_to_torch)
-    return hf_dataset
-
-
-def load_episode_data_index(repo_id, version, root) -> dict[str, torch.Tensor]:
-    """episode_data_index contains the range of indices for each episode
-
-    Example:
-    ```python
-    from_id = episode_data_index["from"][episode_id].item()
-    to_id = episode_data_index["to"][episode_id].item()
-    episode_frames = [dataset[i] for i in range(from_id, to_id)]
-    ```
-    """
-    if root is not None:
-        path = Path(root) / repo_id / "meta_data" / "episode_data_index.safetensors"
-    else:
-        safe_version = get_hf_dataset_safe_version(repo_id, version)
-        path = hf_hub_download(
-            repo_id, "meta_data/episode_data_index.safetensors", repo_type="dataset", revision=safe_version
-        )
-
-    return load_file(path)
-
-
-def load_stats(repo_id, version, root) -> dict[str, dict[str, torch.Tensor]]:
-    """stats contains the statistics per modality computed over the full dataset, such as max, min, mean, std
-
-    Example:
-    ```python
-    normalized_action = (action - stats["action"]["mean"]) / stats["action"]["std"]
-    ```
-    """
-    if root is not None:
-        path = Path(root) / repo_id / "meta_data" / "stats.safetensors"
-    else:
-        safe_version = get_hf_dataset_safe_version(repo_id, version)
-        path = hf_hub_download(
-            repo_id, "meta_data/stats.safetensors", repo_type="dataset", revision=safe_version
-        )
-
-    stats = load_file(path)
-    return unflatten_dict(stats)
-
-
-def load_info(repo_id, version, root) -> dict:
-    """info contains useful information regarding the dataset that are not stored elsewhere
-
-    Example:
-    ```python
-    print("frame per second used to collect the video", info["fps"])
-    ```
-    """
-    if root is not None:
-        path = Path(root) / repo_id / "meta_data" / "info.json"
-    else:
-        safe_version = get_hf_dataset_safe_version(repo_id, version)
-        path = hf_hub_download(repo_id, "meta_data/info.json", repo_type="dataset", revision=safe_version)
-
-    with open(path) as f:
-        info = json.load(f)
-    return info
-
-
-def load_videos(repo_id, version, root) -> Path:
-    if root is not None:
-        path = Path(root) / repo_id / "videos"
-    else:
-        # TODO(rcadene): we download the whole repo here. see if we can avoid this
-        safe_version = get_hf_dataset_safe_version(repo_id, version)
-        repo_dir = snapshot_download(repo_id, repo_type="dataset", revision=safe_version)
-        path = Path(repo_dir) / "videos"
-
-    return path
-
-
-def load_previous_and_future_frames(
-    item: dict[str, torch.Tensor],
-    hf_dataset: datasets.Dataset,
-    episode_data_index: dict[str, torch.Tensor],
-    delta_timestamps: dict[str, list[float]],
-    tolerance_s: float,
-) -> dict[torch.Tensor]:
-    """
-    Given a current item in the dataset containing a timestamp (e.g. 0.6 seconds), and a list of time differences of
-    some modalities (e.g. delta_timestamps={"observation.image": [-0.8, -0.2, 0, 0.2]}), this function computes for each
-    given modality (e.g. "observation.image") a list of query timestamps (e.g. [-0.2, 0.4, 0.6, 0.8]) and loads the closest
-    frames in the dataset.
-
-    Importantly, when no frame can be found around a query timestamp within a specified tolerance window, this function
-    raises an AssertionError. When a timestamp is queried before the first available timestamp of the episode or after
-    the last available timestamp, the violation of the tolerance doesnt raise an AssertionError, and the function
-    populates a boolean array indicating which frames are outside of the episode range. For instance, this boolean array
-    is useful during batched training to not supervise actions associated to timestamps coming after the end of the
-    episode, or to pad the observations in a specific way. Note that by default the observation frames before the start
-    of the episode are the same as the first frame of the episode.
-
-    Parameters:
-    - item (dict): A dictionary containing all the data related to a frame. It is the result of `dataset[idx]`. Each key
-      corresponds to a different modality (e.g., "timestamp", "observation.image", "action").
-    - hf_dataset (datasets.Dataset): A dictionary containing the full dataset. Each key corresponds to a different
-      modality (e.g., "timestamp", "observation.image", "action").
-    - episode_data_index (dict): A dictionary containing two keys ("from" and "to") associated to dataset indices.
-      They indicate the start index and end index of each episode in the dataset.
-    - delta_timestamps (dict): A dictionary containing lists of delta timestamps for each possible modality to be
-      retrieved. These deltas are added to the item timestamp to form the query timestamps.
-    - tolerance_s (float, optional): The tolerance level (in seconds) used to determine if a data point is close enough to the query
-      timestamp by asserting `tol > difference`. It is suggested to set `tol` to a smaller value than the
-      smallest expected inter-frame period, but large enough to account for jitter.
-
-    Returns:
-    - The same item with the queried frames for each modality specified in delta_timestamps, with an additional key for
-      each modality (e.g. "observation.image_is_pad").
-
-    Raises:
-    - AssertionError: If any of the frames unexpectedly violate the tolerance level. This could indicate synchronization
-      issues with timestamps during data collection.
-    """
-    # get indices of the frames associated to the episode, and their timestamps
-    ep_id = item["episode_index"].item()
-    ep_data_id_from = episode_data_index["from"][ep_id].item()
-    ep_data_id_to = episode_data_index["to"][ep_id].item()
-    ep_data_ids = torch.arange(ep_data_id_from, ep_data_id_to, 1)
-
-    # load timestamps
-    ep_timestamps = hf_dataset.select_columns("timestamp")[ep_data_id_from:ep_data_id_to]["timestamp"]
-    ep_timestamps = torch.stack(ep_timestamps)
-
-    # we make the assumption that the timestamps are sorted
-    ep_first_ts = ep_timestamps[0]
-    ep_last_ts = ep_timestamps[-1]
-    current_ts = item["timestamp"].item()
-
-    for key in delta_timestamps:
-        # get timestamps used as query to retrieve data of previous/future frames
-        delta_ts = delta_timestamps[key]
-        query_ts = current_ts + torch.tensor(delta_ts)
-
-        # compute distances between each query timestamp and all timestamps of all the frames belonging to the episode
-        dist = torch.cdist(query_ts[:, None], ep_timestamps[:, None], p=1)
-        min_, argmin_ = dist.min(1)
-
-        # TODO(rcadene): synchronize timestamps + interpolation if needed
-
-        is_pad = min_ > tolerance_s
-
-        # check violated query timestamps are all outside the episode range
-        assert ((query_ts[is_pad] < ep_first_ts) | (ep_last_ts < query_ts[is_pad])).all(), (
-            f"One or several timestamps unexpectedly violate the tolerance ({min_} > {tolerance_s=}) inside episode range."
-            "This might be due to synchronization issues with timestamps during data collection."
-        )
-
-        # get dataset indices corresponding to frames to be loaded
-        data_ids = ep_data_ids[argmin_]
-
-        # load frames modality
-        item[key] = hf_dataset.select_columns(key)[data_ids][key]
-
-        if isinstance(item[key][0], dict) and "path" in item[key][0]:
-            # video mode where frame are expressed as dict of path and timestamp
-            item[key] = item[key]
+def get_hf_features_from_features(features: dict) -> datasets.Features:
+    hf_features = {}
+    for key, ft in features.items():
+        if ft["dtype"] == "video":
+            continue
+        elif ft["dtype"] == "image":
+            hf_features[key] = datasets.Image()
+        elif ft["shape"] == (1,):
+            hf_features[key] = datasets.Value(dtype=ft["dtype"])
         else:
-            item[key] = torch.stack(item[key])
+            assert len(ft["shape"]) == 1
+            hf_features[key] = datasets.Sequence(
+                length=ft["shape"][0], feature=datasets.Value(dtype=ft["dtype"])
+            )
 
-        item[f"{key}_is_pad"] = is_pad
-
-    return item
+    return datasets.Features(hf_features)
 
 
-def calculate_episode_data_index(hf_dataset: datasets.Dataset) -> Dict[str, torch.Tensor]:
-    """
-    Calculate episode data index for the provided HuggingFace Dataset. Relies on episode_index column of hf_dataset.
-
-    Parameters:
-    - hf_dataset (datasets.Dataset): A HuggingFace dataset containing the episode index.
-
-    Returns:
-    - episode_data_index: A dictionary containing the data index for each episode. The dictionary has two keys:
-        - "from": A tensor containing the starting index of each episode.
-        - "to": A tensor containing the ending index of each episode.
-    """
-    episode_data_index = {"from": [], "to": []}
-
-    current_episode = None
-    """
-    The episode_index is a list of integers, each representing the episode index of the corresponding example.
-    For instance, the following is a valid episode_index:
-      [0, 0, 0, 1, 1, 1, 1, 2, 2, 2, 2, 2]
-
-    Below, we iterate through the episode_index and populate the episode_data_index dictionary with the starting and
-    ending index of each episode. For the episode_index above, the episode_data_index dictionary will look like this:
-        {
-            "from": [0, 3, 7],
-            "to": [3, 7, 12]
+def get_features_from_robot(robot: Robot, use_videos: bool = True) -> dict:
+    camera_ft = {}
+    if robot.cameras:
+        camera_ft = {
+            key: {"dtype": "video" if use_videos else "image", **ft}
+            for key, ft in robot.camera_features.items()
         }
-    """
-    if len(hf_dataset) == 0:
-        episode_data_index = {
-            "from": torch.tensor([]),
-            "to": torch.tensor([]),
-        }
-        return episode_data_index
-    for idx, episode_idx in enumerate(hf_dataset["episode_index"]):
-        if episode_idx != current_episode:
-            # We encountered a new episode, so we append its starting location to the "from" list
-            episode_data_index["from"].append(idx)
-            # If this is not the first episode, we append the ending location of the previous episode to the "to" list
-            if current_episode is not None:
-                episode_data_index["to"].append(idx)
-            # Let's keep track of the current episode index
-            current_episode = episode_idx
-        else:
-            # We are still in the same episode, so there is nothing for us to do here
-            pass
-    # We have reached the end of the dataset, so we append the ending location of the last episode to the "to" list
-    episode_data_index["to"].append(idx + 1)
-
-    for k in ["from", "to"]:
-        episode_data_index[k] = torch.tensor(episode_data_index[k])
-
-    return episode_data_index
+    return {**robot.motor_features, **camera_ft, **DEFAULT_FEATURES}
 
 
-def reset_episode_index(hf_dataset: datasets.Dataset) -> datasets.Dataset:
-    """Reset the `episode_index` of the provided HuggingFace Dataset.
-
-    `episode_data_index` (and related functionality such as `load_previous_and_future_frames`) requires the
-    `episode_index` to be sorted, continuous (1,1,1 and not 1,2,1) and start at 0.
-
-    This brings the `episode_index` to the required format.
-    """
-    if len(hf_dataset) == 0:
-        return hf_dataset
-    unique_episode_idxs = torch.stack(hf_dataset["episode_index"]).unique().tolist()
-    episode_idx_to_reset_idx_mapping = {
-        ep_id: reset_ep_id for reset_ep_id, ep_id in enumerate(unique_episode_idxs)
+def create_empty_dataset_info(
+    codebase_version: str,
+    fps: int,
+    robot_type: str,
+    features: dict,
+    use_videos: bool,
+) -> dict:
+    return {
+        "codebase_version": codebase_version,
+        "robot_type": robot_type,
+        "total_episodes": 0,
+        "total_frames": 0,
+        "total_tasks": 0,
+        "total_videos": 0,
+        "total_chunks": 0,
+        "chunks_size": DEFAULT_CHUNK_SIZE,
+        "fps": fps,
+        "splits": {},
+        "data_path": DEFAULT_PARQUET_PATH,
+        "video_path": DEFAULT_VIDEO_PATH if use_videos else None,
+        "features": features,
     }
 
-    def modify_ep_idx_func(example):
-        example["episode_index"] = episode_idx_to_reset_idx_mapping[example["episode_index"].item()]
-        return example
 
-    hf_dataset = hf_dataset.map(modify_ep_idx_func)
+def get_episode_data_index(
+    episode_dicts: list[dict], episodes: list[int] | None = None
+) -> dict[str, torch.Tensor]:
+    episode_lengths = {ep_idx: ep_dict["length"] for ep_idx, ep_dict in enumerate(episode_dicts)}
+    if episodes is not None:
+        episode_lengths = {ep_idx: episode_lengths[ep_idx] for ep_idx in episodes}
 
-    return hf_dataset
+    cumulative_lenghts = list(accumulate(episode_lengths.values()))
+    return {
+        "from": torch.LongTensor([0] + cumulative_lenghts[:-1]),
+        "to": torch.LongTensor(cumulative_lenghts),
+    }
+
+
+def calculate_total_episode(
+    hf_dataset: datasets.Dataset, raise_if_not_contiguous: bool = True
+) -> dict[str, torch.Tensor]:
+    episode_indices = sorted(hf_dataset.unique("episode_index"))
+    total_episodes = len(episode_indices)
+    if raise_if_not_contiguous and episode_indices != list(range(total_episodes)):
+        raise ValueError("episode_index values are not sorted and contiguous.")
+    return total_episodes
+
+
+def calculate_episode_data_index(hf_dataset: datasets.Dataset) -> dict[str, torch.Tensor]:
+    episode_lengths = []
+    table = hf_dataset.data.table
+    total_episodes = calculate_total_episode(hf_dataset)
+    for ep_idx in range(total_episodes):
+        ep_table = table.filter(pc.equal(table["episode_index"], ep_idx))
+        episode_lengths.insert(ep_idx, len(ep_table))
+
+    cumulative_lenghts = list(accumulate(episode_lengths))
+    return {
+        "from": torch.LongTensor([0] + cumulative_lenghts[:-1]),
+        "to": torch.LongTensor(cumulative_lenghts),
+    }
+
+
+def check_timestamps_sync(
+    hf_dataset: datasets.Dataset,
+    episode_data_index: dict[str, torch.Tensor],
+    fps: int,
+    tolerance_s: float,
+    raise_value_error: bool = True,
+) -> bool:
+    """
+    This check is to make sure that each timestamps is separated to the next by 1/fps +/- tolerance to
+    account for possible numerical error.
+    """
+    timestamps = torch.stack(hf_dataset["timestamp"])
+    diffs = torch.diff(timestamps)
+    within_tolerance = torch.abs(diffs - 1 / fps) <= tolerance_s
+
+    # We mask differences between the timestamp at the end of an episode
+    # and the one at the start of the next episode since these are expected
+    # to be outside tolerance.
+    mask = torch.ones(len(diffs), dtype=torch.bool)
+    ignored_diffs = episode_data_index["to"][:-1] - 1
+    mask[ignored_diffs] = False
+    filtered_within_tolerance = within_tolerance[mask]
+
+    if not torch.all(filtered_within_tolerance):
+        # Track original indices before masking
+        original_indices = torch.arange(len(diffs))
+        filtered_indices = original_indices[mask]
+        outside_tolerance_filtered_indices = torch.nonzero(~filtered_within_tolerance)  # .squeeze()
+        outside_tolerance_indices = filtered_indices[outside_tolerance_filtered_indices]
+        episode_indices = torch.stack(hf_dataset["episode_index"])
+
+        outside_tolerances = []
+        for idx in outside_tolerance_indices:
+            entry = {
+                "timestamps": [timestamps[idx], timestamps[idx + 1]],
+                "diff": diffs[idx],
+                "episode_index": episode_indices[idx].item(),
+            }
+            outside_tolerances.append(entry)
+
+        if raise_value_error:
+            raise ValueError(
+                f"""One or several timestamps unexpectedly violate the tolerance inside episode range.
+                This might be due to synchronization issues with timestamps during data collection.
+                \n{pformat(outside_tolerances)}"""
+            )
+        return False
+
+    return True
+
+
+def check_delta_timestamps(
+    delta_timestamps: dict[str, list[float]], fps: int, tolerance_s: float, raise_value_error: bool = True
+) -> bool:
+    """This will check if all the values in delta_timestamps are multiples of 1/fps +/- tolerance.
+    This is to ensure that these delta_timestamps added to any timestamp from a dataset will themselves be
+    actual timestamps from the dataset.
+    """
+    outside_tolerance = {}
+    for key, delta_ts in delta_timestamps.items():
+        within_tolerance = [abs(ts * fps - round(ts * fps)) / fps <= tolerance_s for ts in delta_ts]
+        if not all(within_tolerance):
+            outside_tolerance[key] = [
+                ts for ts, is_within in zip(delta_ts, within_tolerance, strict=True) if not is_within
+            ]
+
+    if len(outside_tolerance) > 0:
+        if raise_value_error:
+            raise ValueError(
+                f"""
+                The following delta_timestamps are found outside of tolerance range.
+                Please make sure they are multiples of 1/{fps} +/- tolerance and adjust
+                their values accordingly.
+                \n{pformat(outside_tolerance)}
+                """
+            )
+        return False
+
+    return True
+
+
+def get_delta_indices(delta_timestamps: dict[str, list[float]], fps: int) -> dict[str, list[int]]:
+    delta_indices = {}
+    for key, delta_ts in delta_timestamps.items():
+        delta_indices[key] = (torch.tensor(delta_ts) * fps).long().tolist()
+
+    return delta_indices
 
 
 def cycle(iterable):
@@ -400,7 +454,7 @@ def cycle(iterable):
             iterator = iter(iterable)
 
 
-def create_branch(repo_id, *, branch: str, repo_type: str | None = None):
+def create_branch(repo_id, *, branch: str, repo_type: str | None = None) -> None:
     """Create a branch on a existing Hugging Face repo. Delete the branch if it already
     exists before creating it.
     """
@@ -415,12 +469,94 @@ def create_branch(repo_id, *, branch: str, repo_type: str | None = None):
     api.create_branch(repo_id, repo_type=repo_type, branch=branch)
 
 
-def create_lerobot_dataset_card(tags: list | None = None, text: str | None = None) -> DatasetCard:
-    card = DatasetCard(DATASET_CARD_TEMPLATE)
-    card.data.task_categories = ["robotics"]
-    card.data.tags = ["LeRobot"]
-    if tags is not None:
-        card.data.tags += tags
-    if text is not None:
-        card.text += text
-    return card
+def create_lerobot_dataset_card(
+    tags: list | None = None,
+    dataset_info: dict | None = None,
+    **kwargs,
+) -> DatasetCard:
+    """
+    Keyword arguments will be used to replace values in ./lerobot/common/datasets/card_template.md.
+    Note: If specified, license must be one of https://huggingface.co/docs/hub/repositories-licenses.
+    """
+    card_tags = ["LeRobot"]
+
+    if tags:
+        card_tags += tags
+    if dataset_info:
+        dataset_structure = "[meta/info.json](meta/info.json):\n"
+        dataset_structure += f"```json\n{json.dumps(dataset_info, indent=4)}\n```\n"
+        kwargs = {**kwargs, "dataset_structure": dataset_structure}
+    card_data = DatasetCardData(
+        license=kwargs.get("license"),
+        tags=card_tags,
+        task_categories=["robotics"],
+        configs=[
+            {
+                "config_name": "default",
+                "data_files": "data/*/*.parquet",
+            }
+        ],
+    )
+
+    card_template = (importlib.resources.files("lerobot.common.datasets") / "card_template.md").read_text()
+
+    return DatasetCard.from_template(
+        card_data=card_data,
+        template_str=card_template,
+        **kwargs,
+    )
+
+
+class IterableNamespace(SimpleNamespace):
+    """
+    A namespace object that supports both dictionary-like iteration and dot notation access.
+    Automatically converts nested dictionaries into IterableNamespaces.
+
+    This class extends SimpleNamespace to provide:
+    - Dictionary-style iteration over keys
+    - Access to items via both dot notation (obj.key) and brackets (obj["key"])
+    - Dictionary-like methods: items(), keys(), values()
+    - Recursive conversion of nested dictionaries
+
+    Args:
+        dictionary: Optional dictionary to initialize the namespace
+        **kwargs: Additional keyword arguments passed to SimpleNamespace
+
+    Examples:
+        >>> data = {"name": "Alice", "details": {"age": 25}}
+        >>> ns = IterableNamespace(data)
+        >>> ns.name
+        'Alice'
+        >>> ns.details.age
+        25
+        >>> list(ns.keys())
+        ['name', 'details']
+        >>> for key, value in ns.items():
+        ...     print(f"{key}: {value}")
+        name: Alice
+        details: IterableNamespace(age=25)
+    """
+
+    def __init__(self, dictionary: dict[str, Any] = None, **kwargs):
+        super().__init__(**kwargs)
+        if dictionary is not None:
+            for key, value in dictionary.items():
+                if isinstance(value, dict):
+                    setattr(self, key, IterableNamespace(value))
+                else:
+                    setattr(self, key, value)
+
+    def __iter__(self) -> Iterator[str]:
+        return iter(vars(self))
+
+    def __getitem__(self, key: str) -> Any:
+        return vars(self)[key]
+
+    def items(self):
+        return vars(self).items()
+
+    def values(self):
+        return vars(self).values()
+
+    def keys(self):
+        return vars(self).keys()

lerobot/common/datasets/v2/batch_convert_dataset_v1_to_v2.py

@@ -0,0 +1,882 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This script is for internal use to convert all datasets under the 'lerobot' hub user account to v2.
+
+Note: Since the original Aloha datasets don't use shadow motors, you need to comment those out in
+lerobot/configs/robot/aloha.yaml before running this script.
+"""
+
+import traceback
+from pathlib import Path
+from textwrap import dedent
+
+from lerobot import available_datasets
+from lerobot.common.datasets.v2.convert_dataset_v1_to_v2 import convert_dataset, parse_robot_config
+
+LOCAL_DIR = Path("data/")
+
+ALOHA_CONFIG = Path("lerobot/configs/robot/aloha.yaml")
+ALOHA_MOBILE_INFO = {
+    "robot_config": parse_robot_config(ALOHA_CONFIG),
+    "license": "mit",
+    "url": "https://mobile-aloha.github.io/",
+    "paper": "https://arxiv.org/abs/2401.02117",
+    "citation_bibtex": dedent(r"""
+        @inproceedings{fu2024mobile,
+            author    = {Fu, Zipeng and Zhao, Tony Z. and Finn, Chelsea},
+            title     = {Mobile ALOHA: Learning Bimanual Mobile Manipulation with Low-Cost Whole-Body Teleoperation},
+            booktitle = {arXiv},
+            year      = {2024},
+        }""").lstrip(),
+}
+ALOHA_STATIC_INFO = {
+    "robot_config": parse_robot_config(ALOHA_CONFIG),
+    "license": "mit",
+    "url": "https://tonyzhaozh.github.io/aloha/",
+    "paper": "https://arxiv.org/abs/2304.13705",
+    "citation_bibtex": dedent(r"""
+        @article{Zhao2023LearningFB,
+            title={Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware},
+            author={Tony Zhao and Vikash Kumar and Sergey Levine and Chelsea Finn},
+            journal={RSS},
+            year={2023},
+            volume={abs/2304.13705},
+            url={https://arxiv.org/abs/2304.13705}
+        }""").lstrip(),
+}
+PUSHT_INFO = {
+    "license": "mit",
+    "url": "https://diffusion-policy.cs.columbia.edu/",
+    "paper": "https://arxiv.org/abs/2303.04137v5",
+    "citation_bibtex": dedent(r"""
+        @article{chi2024diffusionpolicy,
+            author = {Cheng Chi and Zhenjia Xu and Siyuan Feng and Eric Cousineau and Yilun Du and Benjamin Burchfiel and Russ Tedrake and Shuran Song},
+            title ={Diffusion Policy: Visuomotor Policy Learning via Action Diffusion},
+            journal = {The International Journal of Robotics Research},
+            year = {2024},
+        }""").lstrip(),
+}
+XARM_INFO = {
+    "license": "mit",
+    "url": "https://www.nicklashansen.com/td-mpc/",
+    "paper": "https://arxiv.org/abs/2203.04955",
+    "citation_bibtex": dedent(r"""
+        @inproceedings{Hansen2022tdmpc,
+            title={Temporal Difference Learning for Model Predictive Control},
+            author={Nicklas Hansen and Xiaolong Wang and Hao Su},
+            booktitle={ICML},
+            year={2022}
+        }
+    """),
+}
+UNITREEH_INFO = {
+    "license": "apache-2.0",
+}
+
+DATASETS = {
+    "aloha_mobile_cabinet": {
+        "single_task": "Open the top cabinet, store the pot inside it then close the cabinet.",
+        **ALOHA_MOBILE_INFO,
+    },
+    "aloha_mobile_chair": {
+        "single_task": "Push the chairs in front of the desk to place them against it.",
+        **ALOHA_MOBILE_INFO,
+    },
+    "aloha_mobile_elevator": {
+        "single_task": "Take the elevator to the 1st floor.",
+        **ALOHA_MOBILE_INFO,
+    },
+    "aloha_mobile_shrimp": {
+        "single_task": "Sauté the raw shrimp on both sides, then serve it in the bowl.",
+        **ALOHA_MOBILE_INFO,
+    },
+    "aloha_mobile_wash_pan": {
+        "single_task": "Pick up the pan, rinse it in the sink and then place it in the drying rack.",
+        **ALOHA_MOBILE_INFO,
+    },
+    "aloha_mobile_wipe_wine": {
+        "single_task": "Pick up the wet cloth on the faucet and use it to clean the spilled wine on the table and underneath the glass.",
+        **ALOHA_MOBILE_INFO,
+    },
+    "aloha_static_battery": {
+        "single_task": "Place the battery into the slot of the remote controller.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_candy": {"single_task": "Pick up the candy and unwrap it.", **ALOHA_STATIC_INFO},
+    "aloha_static_coffee": {
+        "single_task": "Place the coffee capsule inside the capsule container, then place the cup onto the center of the cup tray, then push the 'Hot Water' and 'Travel Mug' buttons.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_coffee_new": {
+        "single_task": "Place the coffee capsule inside the capsule container, then place the cup onto the center of the cup tray.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_cups_open": {
+        "single_task": "Pick up the plastic cup and open its lid.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_fork_pick_up": {
+        "single_task": "Pick up the fork and place it on the plate.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_pingpong_test": {
+        "single_task": "Transfer one of the two balls in the right glass into the left glass, then transfer it back to the right glass.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_pro_pencil": {
+        "single_task": "Pick up the pencil with the right arm, hand it over to the left arm then place it back onto the table.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_screw_driver": {
+        "single_task": "Pick up the screwdriver with the right arm, hand it over to the left arm then place it into the cup.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_tape": {
+        "single_task": "Cut a small piece of tape from the tape dispenser then place it on the cardboard box's edge.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_thread_velcro": {
+        "single_task": "Pick up the velcro cable tie with the left arm, then insert the end of the velcro tie into the other end's loop with the right arm.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_towel": {
+        "single_task": "Pick up a piece of paper towel and place it on the spilled liquid.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_vinh_cup": {
+        "single_task": "Pick up the plastic cup with the right arm, then pop its lid open with the left arm.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_vinh_cup_left": {
+        "single_task": "Pick up the plastic cup with the left arm, then pop its lid open with the right arm.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_ziploc_slide": {"single_task": "Slide open the ziploc bag.", **ALOHA_STATIC_INFO},
+    "aloha_sim_insertion_scripted": {"single_task": "Insert the peg into the socket.", **ALOHA_STATIC_INFO},
+    "aloha_sim_insertion_scripted_image": {
+        "single_task": "Insert the peg into the socket.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_sim_insertion_human": {"single_task": "Insert the peg into the socket.", **ALOHA_STATIC_INFO},
+    "aloha_sim_insertion_human_image": {
+        "single_task": "Insert the peg into the socket.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_sim_transfer_cube_scripted": {
+        "single_task": "Pick up the cube with the right arm and transfer it to the left arm.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_sim_transfer_cube_scripted_image": {
+        "single_task": "Pick up the cube with the right arm and transfer it to the left arm.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_sim_transfer_cube_human": {
+        "single_task": "Pick up the cube with the right arm and transfer it to the left arm.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_sim_transfer_cube_human_image": {
+        "single_task": "Pick up the cube with the right arm and transfer it to the left arm.",
+        **ALOHA_STATIC_INFO,
+    },
+    "pusht": {"single_task": "Push the T-shaped block onto the T-shaped target.", **PUSHT_INFO},
+    "pusht_image": {"single_task": "Push the T-shaped block onto the T-shaped target.", **PUSHT_INFO},
+    "unitreeh1_fold_clothes": {"single_task": "Fold the sweatshirt.", **UNITREEH_INFO},
+    "unitreeh1_rearrange_objects": {"single_task": "Put the object into the bin.", **UNITREEH_INFO},
+    "unitreeh1_two_robot_greeting": {
+        "single_task": "Greet the other robot with a high five.",
+        **UNITREEH_INFO,
+    },
+    "unitreeh1_warehouse": {
+        "single_task": "Grab the spray paint on the shelf and place it in the bin on top of the robot dog.",
+        **UNITREEH_INFO,
+    },
+    "xarm_lift_medium": {"single_task": "Pick up the cube and lift it.", **XARM_INFO},
+    "xarm_lift_medium_image": {"single_task": "Pick up the cube and lift it.", **XARM_INFO},
+    "xarm_lift_medium_replay": {"single_task": "Pick up the cube and lift it.", **XARM_INFO},
+    "xarm_lift_medium_replay_image": {"single_task": "Pick up the cube and lift it.", **XARM_INFO},
+    "xarm_push_medium": {"single_task": "Push the cube onto the target.", **XARM_INFO},
+    "xarm_push_medium_image": {"single_task": "Push the cube onto the target.", **XARM_INFO},
+    "xarm_push_medium_replay": {"single_task": "Push the cube onto the target.", **XARM_INFO},
+    "xarm_push_medium_replay_image": {"single_task": "Push the cube onto the target.", **XARM_INFO},
+    "umi_cup_in_the_wild": {
+        "single_task": "Put the cup on the plate.",
+        "license": "apache-2.0",
+    },
+    "asu_table_top": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "paper": "https://link.springer.com/article/10.1007/s10514-023-10129-1",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{zhou2023modularity,
+                title={Modularity through Attention: Efficient Training and Transfer of Language-Conditioned Policies for Robot Manipulation},
+                author={Zhou, Yifan and Sonawani, Shubham and Phielipp, Mariano and Stepputtis, Simon and Amor, Heni},
+                booktitle={Conference on Robot Learning},
+                pages={1684--1695},
+                year={2023},
+                organization={PMLR}
+            }
+            @article{zhou2023learning,
+                title={Learning modular language-conditioned robot policies through attention},
+                author={Zhou, Yifan and Sonawani, Shubham and Phielipp, Mariano and Ben Amor, Heni and Stepputtis, Simon},
+                journal={Autonomous Robots},
+                pages={1--21},
+                year={2023},
+                publisher={Springer}
+            }""").lstrip(),
+    },
+    "austin_buds_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://ut-austin-rpl.github.io/BUDS-website/",
+        "paper": "https://arxiv.org/abs/2109.13841",
+        "citation_bibtex": dedent(r"""
+            @article{zhu2022bottom,
+                title={Bottom-Up Skill Discovery From Unsegmented Demonstrations for Long-Horizon Robot Manipulation},
+                author={Zhu, Yifeng and Stone, Peter and Zhu, Yuke},
+                journal={IEEE Robotics and Automation Letters},
+                volume={7},
+                number={2},
+                pages={4126--4133},
+                year={2022},
+                publisher={IEEE}
+            }""").lstrip(),
+    },
+    "austin_sailor_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://ut-austin-rpl.github.io/sailor/",
+        "paper": "https://arxiv.org/abs/2210.11435",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{nasiriany2022sailor,
+                title={Learning and Retrieval from Prior Data for Skill-based Imitation Learning},
+                author={Soroush Nasiriany and Tian Gao and Ajay Mandlekar and Yuke Zhu},
+                booktitle={Conference on Robot Learning (CoRL)},
+                year={2022}
+            }""").lstrip(),
+    },
+    "austin_sirius_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://ut-austin-rpl.github.io/sirius/",
+        "paper": "https://arxiv.org/abs/2211.08416",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{liu2022robot,
+                title = {Robot Learning on the Job: Human-in-the-Loop Autonomy and Learning During Deployment},
+                author = {Huihan Liu and Soroush Nasiriany and Lance Zhang and Zhiyao Bao and Yuke Zhu},
+                booktitle = {Robotics: Science and Systems (RSS)},
+                year = {2023}
+            }""").lstrip(),
+    },
+    "berkeley_autolab_ur5": {
+        "tasks_col": "language_instruction",
+        "license": "cc-by-4.0",
+        "url": "https://sites.google.com/view/berkeley-ur5/home",
+        "citation_bibtex": dedent(r"""
+            @misc{BerkeleyUR5Website,
+                title = {Berkeley {UR5} Demonstration Dataset},
+                author = {Lawrence Yunliang Chen and Simeon Adebola and Ken Goldberg},
+                howpublished = {https://sites.google.com/view/berkeley-ur5/home},
+            }""").lstrip(),
+    },
+    "berkeley_cable_routing": {
+        "tasks_col": "language_instruction",
+        "license": "cc-by-4.0",
+        "url": "https://sites.google.com/view/cablerouting/home",
+        "paper": "https://arxiv.org/abs/2307.08927",
+        "citation_bibtex": dedent(r"""
+            @article{luo2023multistage,
+                author    = {Jianlan Luo and Charles Xu and Xinyang Geng and Gilbert Feng and Kuan Fang and Liam Tan and Stefan Schaal and Sergey Levine},
+                title     = {Multi-Stage Cable Routing through Hierarchical Imitation Learning},
+                journal   = {arXiv pre-print},
+                year      = {2023},
+                url       = {https://arxiv.org/abs/2307.08927},
+            }""").lstrip(),
+    },
+    "berkeley_fanuc_manipulation": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://sites.google.com/berkeley.edu/fanuc-manipulation",
+        "citation_bibtex": dedent(r"""
+            @article{fanuc_manipulation2023,
+                title={Fanuc Manipulation: A Dataset for Learning-based Manipulation with FANUC Mate 200iD Robot},
+                author={Zhu, Xinghao and Tian, Ran and Xu, Chenfeng and Ding, Mingyu and Zhan, Wei and Tomizuka, Masayoshi},
+                year={2023},
+            }""").lstrip(),
+    },
+    "berkeley_gnm_cory_hall": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "paper": "https://arxiv.org/abs/1709.10489",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{kahn2018self,
+                title={Self-supervised deep reinforcement learning with generalized computation graphs for robot navigation},
+                author={Kahn, Gregory and Villaflor, Adam and Ding, Bosen and Abbeel, Pieter and Levine, Sergey},
+                booktitle={2018 IEEE international conference on robotics and automation (ICRA)},
+                pages={5129--5136},
+                year={2018},
+                organization={IEEE}
+            }""").lstrip(),
+    },
+    "berkeley_gnm_recon": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://sites.google.com/view/recon-robot",
+        "paper": "https://arxiv.org/abs/2104.05859",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{shah2021rapid,
+                title={Rapid Exploration for Open-World Navigation with Latent Goal Models},
+                author={Dhruv Shah and Benjamin Eysenbach and Nicholas Rhinehart and Sergey Levine},
+                booktitle={5th Annual Conference on Robot Learning },
+                year={2021},
+                url={https://openreview.net/forum?id=d_SWJhyKfVw}
+            }""").lstrip(),
+    },
+    "berkeley_gnm_sac_son": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://sites.google.com/view/SACSoN-review",
+        "paper": "https://arxiv.org/abs/2306.01874",
+        "citation_bibtex": dedent(r"""
+            @article{hirose2023sacson,
+                title={SACSoN: Scalable Autonomous Data Collection for Social Navigation},
+                author={Hirose, Noriaki and Shah, Dhruv and Sridhar, Ajay and Levine, Sergey},
+                journal={arXiv preprint arXiv:2306.01874},
+                year={2023}
+            }""").lstrip(),
+    },
+    "berkeley_mvp": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "paper": "https://arxiv.org/abs/2203.06173",
+        "citation_bibtex": dedent(r"""
+            @InProceedings{Radosavovic2022,
+                title = {Real-World Robot Learning with Masked Visual Pre-training},
+                author = {Ilija Radosavovic and Tete Xiao and Stephen James and Pieter Abbeel and Jitendra Malik and Trevor Darrell},
+                booktitle = {CoRL},
+                year = {2022}
+            }""").lstrip(),
+    },
+    "berkeley_rpt": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "paper": "https://arxiv.org/abs/2306.10007",
+        "citation_bibtex": dedent(r"""
+            @article{Radosavovic2023,
+                title={Robot Learning with Sensorimotor Pre-training},
+                author={Ilija Radosavovic and Baifeng Shi and Letian Fu and Ken Goldberg and Trevor Darrell and Jitendra Malik},
+                year={2023},
+                journal={arXiv:2306.10007}
+            }""").lstrip(),
+    },
+    "cmu_franka_exploration_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://human-world-model.github.io/",
+        "paper": "https://arxiv.org/abs/2308.10901",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{mendonca2023structured,
+                title={Structured World Models from Human Videos},
+                author={Mendonca, Russell  and Bahl, Shikhar and Pathak, Deepak},
+                journal={RSS},
+                year={2023}
+            }""").lstrip(),
+    },
+    "cmu_play_fusion": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://play-fusion.github.io/",
+        "paper": "https://arxiv.org/abs/2312.04549",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{chen2023playfusion,
+                title={PlayFusion: Skill Acquisition via Diffusion from Language-Annotated Play},
+                author={Chen, Lili and Bahl, Shikhar and Pathak, Deepak},
+                booktitle={CoRL},
+                year={2023}
+            }""").lstrip(),
+    },
+    "cmu_stretch": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://robo-affordances.github.io/",
+        "paper": "https://arxiv.org/abs/2304.08488",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{bahl2023affordances,
+                title={Affordances from Human Videos as a Versatile Representation for Robotics},
+                author={Bahl, Shikhar and Mendonca, Russell and Chen, Lili and Jain, Unnat and Pathak, Deepak},
+                booktitle={CVPR},
+                year={2023}
+            }
+                @article{mendonca2023structured,
+                title={Structured World Models from Human Videos},
+                author={Mendonca, Russell and Bahl, Shikhar and Pathak, Deepak},
+                journal={CoRL},
+                year={2023}
+            }""").lstrip(),
+    },
+    "columbia_cairlab_pusht_real": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://diffusion-policy.cs.columbia.edu/",
+        "paper": "https://arxiv.org/abs/2303.04137v5",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{chi2023diffusionpolicy,
+                title={Diffusion Policy: Visuomotor Policy Learning via Action Diffusion},
+                author={Chi, Cheng and Feng, Siyuan and Du, Yilun and Xu, Zhenjia and Cousineau, Eric and Burchfiel, Benjamin and Song, Shuran},
+                booktitle={Proceedings of Robotics: Science and Systems (RSS)},
+                year={2023}
+            }""").lstrip(),
+    },
+    "conq_hose_manipulation": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://sites.google.com/view/conq-hose-manipulation-dataset/home",
+        "citation_bibtex": dedent(r"""
+            @misc{ConqHoseManipData,
+                author={Peter Mitrano and Dmitry Berenson},
+                title={Conq Hose Manipulation Dataset, v1.15.0},
+                year={2024},
+                howpublished={https://sites.google.com/view/conq-hose-manipulation-dataset}
+            }""").lstrip(),
+    },
+    "dlr_edan_shared_control": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "paper": "https://ieeexplore.ieee.org/document/9341156",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{vogel_edan_2020,
+                title = {EDAN - an EMG-Controlled Daily Assistant to Help People with Physical Disabilities},
+                language = {en},
+                booktitle = {2020 {IEEE}/{RSJ} {International} {Conference} on {Intelligent} {Robots} and {Systems} ({IROS})},
+                author = {Vogel, Jörn and Hagengruber, Annette and Iskandar, Maged and Quere, Gabriel and Leipscher, Ulrike and Bustamante, Samuel and Dietrich, Alexander and Hoeppner, Hannes and Leidner, Daniel and Albu-Schäffer, Alin},
+                year = {2020}
+            }
+            @inproceedings{quere_shared_2020,
+                address = {Paris, France},
+                title = {Shared {Control} {Templates} for {Assistive} {Robotics}},
+                language = {en},
+                booktitle = {2020 {IEEE} {International} {Conference} on {Robotics} and {Automation} ({ICRA})},
+                author = {Quere, Gabriel and Hagengruber, Annette and Iskandar, Maged and Bustamante, Samuel and Leidner, Daniel and Stulp, Freek and Vogel, Joern},
+                year = {2020},
+                pages = {7},
+            }""").lstrip(),
+    },
+    "dlr_sara_grid_clamp": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "paper": "https://www.researchsquare.com/article/rs-3289569/v1",
+        "citation_bibtex": dedent(r"""
+            @article{padalkar2023guided,
+                title={A guided reinforcement learning approach using shared control templates for learning manipulation skills in the real world},
+                author={Padalkar, Abhishek and Quere, Gabriel and Raffin, Antonin and Silv{\'e}rio, Jo{\~a}o and Stulp, Freek},
+                journal={Research square preprint rs-3289569/v1},
+                year={2023}
+            }""").lstrip(),
+    },
+    "dlr_sara_pour": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "paper": "https://elib.dlr.de/193739/1/padalkar2023rlsct.pdf",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{padalkar2023guiding,
+                title={Guiding Reinforcement Learning with Shared Control Templates},
+                author={Padalkar, Abhishek and Quere, Gabriel and Steinmetz, Franz and Raffin, Antonin and Nieuwenhuisen, Matthias and Silv{\'e}rio, Jo{\~a}o and Stulp, Freek},
+                booktitle={40th IEEE International Conference on Robotics and Automation, ICRA 2023},
+                year={2023},
+                organization={IEEE}
+            }""").lstrip(),
+    },
+    "droid_100": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://droid-dataset.github.io/",
+        "paper": "https://arxiv.org/abs/2403.12945",
+        "citation_bibtex": dedent(r"""
+            @article{khazatsky2024droid,
+                title   = {DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset},
+                author  = {Alexander Khazatsky and Karl Pertsch and Suraj Nair and Ashwin Balakrishna and Sudeep Dasari and Siddharth Karamcheti and Soroush Nasiriany and Mohan Kumar Srirama and Lawrence Yunliang Chen and Kirsty Ellis and Peter David Fagan and Joey Hejna and Masha Itkina and Marion Lepert and Yecheng Jason Ma and Patrick Tree Miller and Jimmy Wu and Suneel Belkhale and Shivin Dass and Huy Ha and Arhan Jain and Abraham Lee and Youngwoon Lee and Marius Memmel and Sungjae Park and Ilija Radosavovic and Kaiyuan Wang and Albert Zhan and Kevin Black and Cheng Chi and Kyle Beltran Hatch and Shan Lin and Jingpei Lu and Jean Mercat and Abdul Rehman and Pannag R Sanketi and Archit Sharma and Cody Simpson and Quan Vuong and Homer Rich Walke and Blake Wulfe and Ted Xiao and Jonathan Heewon Yang and Arefeh Yavary and Tony Z. Zhao and Christopher Agia and Rohan Baijal and Mateo Guaman Castro and Daphne Chen and Qiuyu Chen and Trinity Chung and Jaimyn Drake and Ethan Paul Foster and Jensen Gao and David Antonio Herrera and Minho Heo and Kyle Hsu and Jiaheng Hu and Donovon Jackson and Charlotte Le and Yunshuang Li and Kevin Lin and Roy Lin and Zehan Ma and Abhiram Maddukuri and Suvir Mirchandani and Daniel Morton and Tony Nguyen and Abigail O'Neill and Rosario Scalise and Derick Seale and Victor Son and Stephen Tian and Emi Tran and Andrew E. Wang and Yilin Wu and Annie Xie and Jingyun Yang and Patrick Yin and Yunchu Zhang and Osbert Bastani and Glen Berseth and Jeannette Bohg and Ken Goldberg and Abhinav Gupta and Abhishek Gupta and Dinesh Jayaraman and Joseph J Lim and Jitendra Malik and Roberto Martín-Martín and Subramanian Ramamoorthy and Dorsa Sadigh and Shuran Song and Jiajun Wu and Michael C. Yip and Yuke Zhu and Thomas Kollar and Sergey Levine and Chelsea Finn},
+                year    = {2024},
+            }""").lstrip(),
+    },
+    "fmb": {
+        "tasks_col": "language_instruction",
+        "license": "cc-by-4.0",
+        "url": "https://functional-manipulation-benchmark.github.io/",
+        "paper": "https://arxiv.org/abs/2401.08553",
+        "citation_bibtex": dedent(r"""
+            @article{luo2024fmb,
+                title={FMB: a Functional Manipulation Benchmark for Generalizable Robotic Learning},
+                author={Luo, Jianlan and Xu, Charles and Liu, Fangchen and Tan, Liam and Lin, Zipeng and Wu, Jeffrey and Abbeel, Pieter and Levine, Sergey},
+                journal={arXiv preprint arXiv:2401.08553},
+                year={2024}
+            }""").lstrip(),
+    },
+    "iamlab_cmu_pickup_insert": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://openreview.net/forum?id=WuBv9-IGDUA",
+        "paper": "https://arxiv.org/abs/2401.14502",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{saxena2023multiresolution,
+                title={Multi-Resolution Sensing for Real-Time Control with Vision-Language Models},
+                author={Saumya Saxena and Mohit Sharma and Oliver Kroemer},
+                booktitle={7th Annual Conference on Robot Learning},
+                year={2023},
+                url={https://openreview.net/forum?id=WuBv9-IGDUA}
+            }""").lstrip(),
+    },
+    "imperialcollege_sawyer_wrist_cam": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+    },
+    "jaco_play": {
+        "tasks_col": "language_instruction",
+        "license": "cc-by-4.0",
+        "url": "https://github.com/clvrai/clvr_jaco_play_dataset",
+        "citation_bibtex": dedent(r"""
+            @software{dass2023jacoplay,
+                author = {Dass, Shivin and Yapeter, Jullian and Zhang, Jesse and Zhang, Jiahui
+                            and Pertsch, Karl and Nikolaidis, Stefanos and Lim, Joseph J.},
+                title = {CLVR Jaco Play Dataset},
+                url = {https://github.com/clvrai/clvr_jaco_play_dataset},
+                version = {1.0.0},
+                year = {2023}
+            }""").lstrip(),
+    },
+    "kaist_nonprehensile": {
+        "tasks_col": "language_instruction",
+        "license": "cc-by-4.0",
+        "url": "https://github.com/JaeHyung-Kim/rlds_dataset_builder",
+        "citation_bibtex": dedent(r"""
+            @article{kimpre,
+                title={Pre-and post-contact policy decomposition for non-prehensile manipulation with zero-shot sim-to-real transfer},
+                author={Kim, Minchan and Han, Junhyek and Kim, Jaehyung and Kim, Beomjoon},
+                booktitle={2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
+                year={2023},
+                organization={IEEE}
+            }""").lstrip(),
+    },
+    "nyu_door_opening_surprising_effectiveness": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://jyopari.github.io/VINN/",
+        "paper": "https://arxiv.org/abs/2112.01511",
+        "citation_bibtex": dedent(r"""
+            @misc{pari2021surprising,
+                title={The Surprising Effectiveness of Representation Learning for Visual Imitation},
+                author={Jyothish Pari and Nur Muhammad Shafiullah and Sridhar Pandian Arunachalam and Lerrel Pinto},
+                year={2021},
+                eprint={2112.01511},
+                archivePrefix={arXiv},
+                primaryClass={cs.RO}
+            }""").lstrip(),
+    },
+    "nyu_franka_play_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://play-to-policy.github.io/",
+        "paper": "https://arxiv.org/abs/2210.10047",
+        "citation_bibtex": dedent(r"""
+            @article{cui2022play,
+                title   = {From Play to Policy: Conditional Behavior Generation from Uncurated Robot Data},
+                author  = {Cui, Zichen Jeff and Wang, Yibin and Shafiullah, Nur Muhammad Mahi and Pinto, Lerrel},
+                journal = {arXiv preprint arXiv:2210.10047},
+                year    = {2022}
+            }""").lstrip(),
+    },
+    "nyu_rot_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://rot-robot.github.io/",
+        "paper": "https://arxiv.org/abs/2206.15469",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{haldar2023watch,
+                title={Watch and match: Supercharging imitation with regularized optimal transport},
+                author={Haldar, Siddhant and Mathur, Vaibhav and Yarats, Denis and Pinto, Lerrel},
+                booktitle={Conference on Robot Learning},
+                pages={32--43},
+                year={2023},
+                organization={PMLR}
+            }""").lstrip(),
+    },
+    "roboturk": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://roboturk.stanford.edu/dataset_real.html",
+        "paper": "PAPER",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{mandlekar2019scaling,
+                title={Scaling robot supervision to hundreds of hours with roboturk: Robotic manipulation dataset through human reasoning and dexterity},
+                author={Mandlekar, Ajay and Booher, Jonathan and Spero, Max and Tung, Albert and Gupta, Anchit and Zhu, Yuke and Garg, Animesh and Savarese, Silvio and Fei-Fei, Li},
+                booktitle={2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
+                pages={1048--1055},
+                year={2019},
+                organization={IEEE}
+            }""").lstrip(),
+    },
+    "stanford_hydra_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://sites.google.com/view/hydra-il-2023",
+        "paper": "https://arxiv.org/abs/2306.17237",
+        "citation_bibtex": dedent(r"""
+            @article{belkhale2023hydra,
+                title={HYDRA: Hybrid Robot Actions for Imitation Learning},
+                author={Belkhale, Suneel and Cui, Yuchen and Sadigh, Dorsa},
+                journal={arxiv},
+                year={2023}
+            }""").lstrip(),
+    },
+    "stanford_kuka_multimodal_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://sites.google.com/view/visionandtouch",
+        "paper": "https://arxiv.org/abs/1810.10191",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{lee2019icra,
+                title={Making sense of vision and touch: Self-supervised learning of multimodal representations for contact-rich tasks},
+                author={Lee, Michelle A and Zhu, Yuke and Srinivasan, Krishnan and Shah, Parth and Savarese, Silvio and Fei-Fei, Li and  Garg, Animesh and Bohg, Jeannette},
+                booktitle={2019 IEEE International Conference on Robotics and Automation (ICRA)},
+                year={2019},
+                url={https://arxiv.org/abs/1810.10191}
+            }""").lstrip(),
+    },
+    "stanford_robocook": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://hshi74.github.io/robocook/",
+        "paper": "https://arxiv.org/abs/2306.14447",
+        "citation_bibtex": dedent(r"""
+            @article{shi2023robocook,
+                title={RoboCook: Long-Horizon Elasto-Plastic Object Manipulation with Diverse Tools},
+                author={Shi, Haochen and Xu, Huazhe and Clarke, Samuel and Li, Yunzhu and Wu, Jiajun},
+                journal={arXiv preprint arXiv:2306.14447},
+                year={2023}
+            }""").lstrip(),
+    },
+    "taco_play": {
+        "tasks_col": "language_instruction",
+        "license": "cc-by-4.0",
+        "url": "https://www.kaggle.com/datasets/oiermees/taco-robot",
+        "paper": "https://arxiv.org/abs/2209.08959, https://arxiv.org/abs/2210.01911",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{rosete2022tacorl,
+                author = {Erick Rosete-Beas and Oier Mees and Gabriel Kalweit and Joschka Boedecker and Wolfram Burgard},
+                title = {Latent Plans for Task Agnostic Offline Reinforcement Learning},
+                journal = {Proceedings of the 6th Conference on Robot Learning (CoRL)},
+                year = {2022}
+            }
+            @inproceedings{mees23hulc2,
+                title={Grounding  Language  with  Visual  Affordances  over  Unstructured  Data},
+                author={Oier Mees and Jessica Borja-Diaz and Wolfram Burgard},
+                booktitle = {Proceedings of the IEEE International Conference on Robotics and Automation (ICRA)},
+                year={2023},
+                address = {London, UK}
+            }""").lstrip(),
+    },
+    "tokyo_u_lsmo": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "URL",
+        "paper": "https://arxiv.org/abs/2107.05842",
+        "citation_bibtex": dedent(r"""
+            @Article{Osa22,
+                author  = {Takayuki Osa},
+                journal = {The International Journal of Robotics Research},
+                title   = {Motion Planning by Learning the Solution Manifold in Trajectory Optimization},
+                year    = {2022},
+                number  = {3},
+                pages   = {291--311},
+                volume  = {41},
+            }""").lstrip(),
+    },
+    "toto": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://toto-benchmark.org/",
+        "paper": "https://arxiv.org/abs/2306.00942",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{zhou2023train,
+                author={Zhou, Gaoyue and Dean, Victoria and Srirama, Mohan Kumar and Rajeswaran, Aravind and Pari, Jyothish and Hatch, Kyle and Jain, Aryan and Yu, Tianhe and Abbeel, Pieter and Pinto, Lerrel and Finn, Chelsea and Gupta, Abhinav},
+                booktitle={2023 IEEE International Conference on Robotics and Automation (ICRA)},
+                title={Train Offline, Test Online: A Real Robot Learning Benchmark},
+                year={2023},
+            }""").lstrip(),
+    },
+    "ucsd_kitchen_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "citation_bibtex": dedent(r"""
+            @ARTICLE{ucsd_kitchens,
+                author = {Ge Yan, Kris Wu, and Xiaolong Wang},
+                title = {{ucsd kitchens Dataset}},
+                year = {2023},
+                month = {August}
+            }""").lstrip(),
+    },
+    "ucsd_pick_and_place_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://owmcorl.github.io/#",
+        "paper": "https://arxiv.org/abs/2310.16029",
+        "citation_bibtex": dedent(r"""
+            @preprint{Feng2023Finetuning,
+                title={Finetuning Offline World Models in the Real World},
+                author={Yunhai Feng, Nicklas Hansen, Ziyan Xiong, Chandramouli Rajagopalan, Xiaolong Wang},
+                year={2023}
+            }""").lstrip(),
+    },
+    "uiuc_d3field": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://robopil.github.io/d3fields/",
+        "paper": "https://arxiv.org/abs/2309.16118",
+        "citation_bibtex": dedent(r"""
+            @article{wang2023d3field,
+                title={D^3Field: Dynamic 3D Descriptor Fields for Generalizable Robotic Manipulation},
+                author={Wang, Yixuan and Li, Zhuoran and Zhang, Mingtong and Driggs-Campbell, Katherine and Wu, Jiajun and Fei-Fei, Li and Li, Yunzhu},
+                journal={arXiv preprint arXiv:},
+                year={2023},
+            }""").lstrip(),
+    },
+    "usc_cloth_sim": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://uscresl.github.io/dmfd/",
+        "paper": "https://arxiv.org/abs/2207.10148",
+        "citation_bibtex": dedent(r"""
+            @article{salhotra2022dmfd,
+                author={Salhotra, Gautam and Liu, I-Chun Arthur and Dominguez-Kuhne, Marcus and Sukhatme, Gaurav S.},
+                journal={IEEE Robotics and Automation Letters},
+                title={Learning Deformable Object Manipulation From Expert Demonstrations},
+                year={2022},
+                volume={7},
+                number={4},
+                pages={8775-8782},
+                doi={10.1109/LRA.2022.3187843}
+            }""").lstrip(),
+    },
+    "utaustin_mutex": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://ut-austin-rpl.github.io/MUTEX/",
+        "paper": "https://arxiv.org/abs/2309.14320",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{shah2023mutex,
+                title={{MUTEX}: Learning Unified Policies from Multimodal Task Specifications},
+                author={Rutav Shah and Roberto Mart{\'\i}n-Mart{\'\i}n and Yuke Zhu},
+                booktitle={7th Annual Conference on Robot Learning},
+                year={2023},
+                url={https://openreview.net/forum?id=PwqiqaaEzJ}
+            }""").lstrip(),
+    },
+    "utokyo_pr2_opening_fridge": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "citation_bibtex": dedent(r"""
+            @misc{oh2023pr2utokyodatasets,
+                author={Jihoon Oh and Naoaki Kanazawa and Kento Kawaharazuka},
+                title={X-Embodiment U-Tokyo PR2 Datasets},
+                year={2023},
+                url={https://github.com/ojh6404/rlds_dataset_builder},
+            }""").lstrip(),
+    },
+    "utokyo_pr2_tabletop_manipulation": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "citation_bibtex": dedent(r"""
+            @misc{oh2023pr2utokyodatasets,
+                author={Jihoon Oh and Naoaki Kanazawa and Kento Kawaharazuka},
+                title={X-Embodiment U-Tokyo PR2 Datasets},
+                year={2023},
+                url={https://github.com/ojh6404/rlds_dataset_builder},
+            }""").lstrip(),
+    },
+    "utokyo_saytap": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://saytap.github.io/",
+        "paper": "https://arxiv.org/abs/2306.07580",
+        "citation_bibtex": dedent(r"""
+            @article{saytap2023,
+                author = {Yujin Tang and Wenhao Yu and Jie Tan and Heiga Zen and Aleksandra Faust and
+                Tatsuya Harada},
+                title  = {SayTap: Language to Quadrupedal Locomotion},
+                eprint = {arXiv:2306.07580},
+                url    = {https://saytap.github.io},
+                note   = {https://saytap.github.io},
+                year   = {2023}
+            }""").lstrip(),
+    },
+    "utokyo_xarm_bimanual": {
+        "tasks_col": "language_instruction",
+        "license": "cc-by-4.0",
+        "citation_bibtex": dedent(r"""
+            @misc{matsushima2023weblab,
+                title={Weblab xArm Dataset},
+                author={Tatsuya Matsushima and Hiroki Furuta and Yusuke Iwasawa and Yutaka Matsuo},
+                year={2023},
+            }""").lstrip(),
+    },
+    "utokyo_xarm_pick_and_place": {
+        "tasks_col": "language_instruction",
+        "license": "cc-by-4.0",
+        "citation_bibtex": dedent(r"""
+            @misc{matsushima2023weblab,
+                title={Weblab xArm Dataset},
+                author={Tatsuya Matsushima and Hiroki Furuta and Yusuke Iwasawa and Yutaka Matsuo},
+                year={2023},
+            }""").lstrip(),
+    },
+    "viola": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://ut-austin-rpl.github.io/VIOLA/",
+        "paper": "https://arxiv.org/abs/2210.11339",
+        "citation_bibtex": dedent(r"""
+            @article{zhu2022viola,
+                title={VIOLA: Imitation Learning for Vision-Based Manipulation with Object Proposal Priors},
+                author={Zhu, Yifeng and Joshi, Abhishek and Stone, Peter and Zhu, Yuke},
+                journal={6th Annual Conference on Robot Learning (CoRL)},
+                year={2022}
+            }""").lstrip(),
+    },
+}
+
+
+def batch_convert():
+    status = {}
+    logfile = LOCAL_DIR / "conversion_log.txt"
+    assert set(DATASETS) == {id_.split("/")[1] for id_ in available_datasets}
+    for num, (name, kwargs) in enumerate(DATASETS.items()):
+        repo_id = f"lerobot/{name}"
+        print(f"\nConverting {repo_id} ({num}/{len(DATASETS)})")
+        print("---------------------------------------------------------")
+        try:
+            convert_dataset(repo_id, LOCAL_DIR, **kwargs)
+            status = f"{repo_id}: success."
+            with open(logfile, "a") as file:
+                file.write(status + "\n")
+        except Exception:
+            status = f"{repo_id}: failed\n    {traceback.format_exc()}"
+            with open(logfile, "a") as file:
+                file.write(status + "\n")
+            continue
+
+
+if __name__ == "__main__":
+    batch_convert()

lerobot/common/datasets/v2/convert_dataset_v1_to_v2.py

@@ -0,0 +1,665 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This script will help you convert any LeRobot dataset already pushed to the hub from codebase version 1.6 to
+2.0. You will be required to provide the 'tasks', which is a short but accurate description in plain English
+for each of the task performed in the dataset. This will allow to easily train models with task-conditionning.
+
+We support 3 different scenarios for these tasks (see instructions below):
+    1. Single task dataset: all episodes of your dataset have the same single task.
+    2. Single task episodes: the episodes of your dataset each contain a single task but they can differ from
+      one episode to the next.
+    3. Multi task episodes: episodes of your dataset may each contain several different tasks.
+
+
+Can you can also provide a robot config .yaml file (not mandatory) to this script via the option
+'--robot-config' so that it writes information about the robot (robot type, motors names) this dataset was
+recorded with. For now, only Aloha/Koch type robots are supported with this option.
+
+
+# 1. Single task dataset
+If your dataset contains a single task, you can simply provide it directly via the CLI with the
+'--single-task' option.
+
+Examples:
+
+```bash
+python lerobot/common/datasets/v2/convert_dataset_v1_to_v2.py \
+    --repo-id lerobot/aloha_sim_insertion_human_image \
+    --single-task "Insert the peg into the socket." \
+    --robot-config lerobot/configs/robot/aloha.yaml \
+    --local-dir data
+```
+
+```bash
+python lerobot/common/datasets/v2/convert_dataset_v1_to_v2.py \
+    --repo-id aliberts/koch_tutorial \
+    --single-task "Pick the Lego block and drop it in the box on the right." \
+    --robot-config lerobot/configs/robot/koch.yaml \
+    --local-dir data
+```
+
+
+# 2. Single task episodes
+If your dataset is a multi-task dataset, you have two options to provide the tasks to this script:
+
+- If your dataset already contains a language instruction column in its parquet file, you can simply provide
+  this column's name with the '--tasks-col' arg.
+
+    Example:
+
+    ```bash
+    python lerobot/common/datasets/v2/convert_dataset_v1_to_v2.py \
+        --repo-id lerobot/stanford_kuka_multimodal_dataset \
+        --tasks-col "language_instruction" \
+        --local-dir data
+    ```
+
+- If your dataset doesn't contain a language instruction, you should provide the path to a .json file with the
+  '--tasks-path' arg. This file should have the following structure where keys correspond to each
+  episode_index in the dataset, and values are the language instruction for that episode.
+
+    Example:
+
+    ```json
+    {
+        "0": "Do something",
+        "1": "Do something else",
+        "2": "Do something",
+        "3": "Go there",
+        ...
+    }
+    ```
+
+# 3. Multi task episodes
+If you have multiple tasks per episodes, your dataset should contain a language instruction column in its
+parquet file, and you must provide this column's name with the '--tasks-col' arg.
+
+Example:
+
+```bash
+python lerobot/common/datasets/v2/convert_dataset_v1_to_v2.py \
+    --repo-id lerobot/stanford_kuka_multimodal_dataset \
+    --tasks-col "language_instruction" \
+    --local-dir data
+```
+"""
+
+import argparse
+import contextlib
+import filecmp
+import json
+import logging
+import math
+import shutil
+import subprocess
+import tempfile
+from pathlib import Path
+
+import datasets
+import pyarrow.compute as pc
+import pyarrow.parquet as pq
+import torch
+from datasets import Dataset
+from huggingface_hub import HfApi
+from huggingface_hub.errors import EntryNotFoundError, HfHubHTTPError
+from safetensors.torch import load_file
+
+from lerobot.common.datasets.utils import (
+    DEFAULT_CHUNK_SIZE,
+    DEFAULT_PARQUET_PATH,
+    DEFAULT_VIDEO_PATH,
+    EPISODES_PATH,
+    INFO_PATH,
+    STATS_PATH,
+    TASKS_PATH,
+    create_branch,
+    create_lerobot_dataset_card,
+    flatten_dict,
+    get_hub_safe_version,
+    load_json,
+    unflatten_dict,
+    write_json,
+    write_jsonlines,
+)
+from lerobot.common.datasets.video_utils import (
+    VideoFrame,  # noqa: F401
+    get_image_pixel_channels,
+    get_video_info,
+)
+from lerobot.common.utils.utils import init_hydra_config
+
+V16 = "v1.6"
+V20 = "v2.0"
+
+GITATTRIBUTES_REF = "aliberts/gitattributes_reference"
+V1_VIDEO_FILE = "{video_key}_episode_{episode_index:06d}.mp4"
+V1_INFO_PATH = "meta_data/info.json"
+V1_STATS_PATH = "meta_data/stats.safetensors"
+
+
+def parse_robot_config(config_path: Path, config_overrides: list[str] | None = None) -> tuple[str, dict]:
+    robot_cfg = init_hydra_config(config_path, config_overrides)
+    if robot_cfg["robot_type"] in ["aloha", "koch"]:
+        state_names = [
+            f"{arm}_{motor}" if len(robot_cfg["follower_arms"]) > 1 else motor
+            for arm in robot_cfg["follower_arms"]
+            for motor in robot_cfg["follower_arms"][arm]["motors"]
+        ]
+        action_names = [
+            # f"{arm}_{motor}" for arm in ["left", "right"] for motor in robot_cfg["leader_arms"][arm]["motors"]
+            f"{arm}_{motor}" if len(robot_cfg["leader_arms"]) > 1 else motor
+            for arm in robot_cfg["leader_arms"]
+            for motor in robot_cfg["leader_arms"][arm]["motors"]
+        ]
+    # elif robot_cfg["robot_type"] == "stretch3": TODO
+    else:
+        raise NotImplementedError(
+            "Please provide robot_config={'robot_type': ..., 'names': ...} directly to convert_dataset()."
+        )
+
+    return {
+        "robot_type": robot_cfg["robot_type"],
+        "names": {
+            "observation.state": state_names,
+            "observation.effort": state_names,
+            "action": action_names,
+        },
+    }
+
+
+def convert_stats_to_json(v1_dir: Path, v2_dir: Path) -> None:
+    safetensor_path = v1_dir / V1_STATS_PATH
+    stats = load_file(safetensor_path)
+    serialized_stats = {key: value.tolist() for key, value in stats.items()}
+    serialized_stats = unflatten_dict(serialized_stats)
+
+    json_path = v2_dir / STATS_PATH
+    json_path.parent.mkdir(exist_ok=True, parents=True)
+    with open(json_path, "w") as f:
+        json.dump(serialized_stats, f, indent=4)
+
+    # Sanity check
+    with open(json_path) as f:
+        stats_json = json.load(f)
+
+    stats_json = flatten_dict(stats_json)
+    stats_json = {key: torch.tensor(value) for key, value in stats_json.items()}
+    for key in stats:
+        torch.testing.assert_close(stats_json[key], stats[key])
+
+
+def get_features_from_hf_dataset(dataset: Dataset, robot_config: dict | None = None) -> dict[str, list]:
+    features = {}
+    for key, ft in dataset.features.items():
+        if isinstance(ft, datasets.Value):
+            dtype = ft.dtype
+            shape = (1,)
+            names = None
+        if isinstance(ft, datasets.Sequence):
+            assert isinstance(ft.feature, datasets.Value)
+            dtype = ft.feature.dtype
+            shape = (ft.length,)
+            motor_names = (
+                robot_config["names"][key] if robot_config else [f"motor_{i}" for i in range(ft.length)]
+            )
+            assert len(motor_names) == shape[0]
+            names = {"motors": motor_names}
+        elif isinstance(ft, datasets.Image):
+            dtype = "image"
+            image = dataset[0][key]  # Assuming first row
+            channels = get_image_pixel_channels(image)
+            shape = (image.height, image.width, channels)
+            names = ["height", "width", "channel"]
+        elif ft._type == "VideoFrame":
+            dtype = "video"
+            shape = None  # Add shape later
+            names = ["height", "width", "channel"]
+
+        features[key] = {
+            "dtype": dtype,
+            "shape": shape,
+            "names": names,
+        }
+
+    return features
+
+
+def add_task_index_by_episodes(dataset: Dataset, tasks_by_episodes: dict) -> tuple[Dataset, list[str]]:
+    df = dataset.to_pandas()
+    tasks = list(set(tasks_by_episodes.values()))
+    tasks_to_task_index = {task: task_idx for task_idx, task in enumerate(tasks)}
+    episodes_to_task_index = {ep_idx: tasks_to_task_index[task] for ep_idx, task in tasks_by_episodes.items()}
+    df["task_index"] = df["episode_index"].map(episodes_to_task_index).astype(int)
+
+    features = dataset.features
+    features["task_index"] = datasets.Value(dtype="int64")
+    dataset = Dataset.from_pandas(df, features=features, split="train")
+    return dataset, tasks
+
+
+def add_task_index_from_tasks_col(
+    dataset: Dataset, tasks_col: str
+) -> tuple[Dataset, dict[str, list[str]], list[str]]:
+    df = dataset.to_pandas()
+
+    # HACK: This is to clean some of the instructions in our version of Open X datasets
+    prefix_to_clean = "tf.Tensor(b'"
+    suffix_to_clean = "', shape=(), dtype=string)"
+    df[tasks_col] = df[tasks_col].str.removeprefix(prefix_to_clean).str.removesuffix(suffix_to_clean)
+
+    # Create task_index col
+    tasks_by_episode = df.groupby("episode_index")[tasks_col].unique().apply(lambda x: x.tolist()).to_dict()
+    tasks = df[tasks_col].unique().tolist()
+    tasks_to_task_index = {task: idx for idx, task in enumerate(tasks)}
+    df["task_index"] = df[tasks_col].map(tasks_to_task_index).astype(int)
+
+    # Build the dataset back from df
+    features = dataset.features
+    features["task_index"] = datasets.Value(dtype="int64")
+    dataset = Dataset.from_pandas(df, features=features, split="train")
+    dataset = dataset.remove_columns(tasks_col)
+
+    return dataset, tasks, tasks_by_episode
+
+
+def split_parquet_by_episodes(
+    dataset: Dataset,
+    total_episodes: int,
+    total_chunks: int,
+    output_dir: Path,
+) -> list:
+    table = dataset.data.table
+    episode_lengths = []
+    for ep_chunk in range(total_chunks):
+        ep_chunk_start = DEFAULT_CHUNK_SIZE * ep_chunk
+        ep_chunk_end = min(DEFAULT_CHUNK_SIZE * (ep_chunk + 1), total_episodes)
+        chunk_dir = "/".join(DEFAULT_PARQUET_PATH.split("/")[:-1]).format(episode_chunk=ep_chunk)
+        (output_dir / chunk_dir).mkdir(parents=True, exist_ok=True)
+        for ep_idx in range(ep_chunk_start, ep_chunk_end):
+            ep_table = table.filter(pc.equal(table["episode_index"], ep_idx))
+            episode_lengths.insert(ep_idx, len(ep_table))
+            output_file = output_dir / DEFAULT_PARQUET_PATH.format(
+                episode_chunk=ep_chunk, episode_index=ep_idx
+            )
+            pq.write_table(ep_table, output_file)
+
+    return episode_lengths
+
+
+def move_videos(
+    repo_id: str,
+    video_keys: list[str],
+    total_episodes: int,
+    total_chunks: int,
+    work_dir: Path,
+    clean_gittatributes: Path,
+    branch: str = "main",
+) -> None:
+    """
+    HACK: Since HfApi() doesn't provide a way to move files directly in a repo, this function will run git
+    commands to fetch git lfs video files references to move them into subdirectories without having to
+    actually download them.
+    """
+    _lfs_clone(repo_id, work_dir, branch)
+
+    videos_moved = False
+    video_files = [str(f.relative_to(work_dir)) for f in work_dir.glob("videos*/*.mp4")]
+    if len(video_files) == 0:
+        video_files = [str(f.relative_to(work_dir)) for f in work_dir.glob("videos*/*/*/*.mp4")]
+        videos_moved = True  # Videos have already been moved
+
+    assert len(video_files) == total_episodes * len(video_keys)
+
+    lfs_untracked_videos = _get_lfs_untracked_videos(work_dir, video_files)
+
+    current_gittatributes = work_dir / ".gitattributes"
+    if not filecmp.cmp(current_gittatributes, clean_gittatributes, shallow=False):
+        fix_gitattributes(work_dir, current_gittatributes, clean_gittatributes)
+
+    if lfs_untracked_videos:
+        fix_lfs_video_files_tracking(work_dir, video_files)
+
+    if videos_moved:
+        return
+
+    video_dirs = sorted(work_dir.glob("videos*/"))
+    for ep_chunk in range(total_chunks):
+        ep_chunk_start = DEFAULT_CHUNK_SIZE * ep_chunk
+        ep_chunk_end = min(DEFAULT_CHUNK_SIZE * (ep_chunk + 1), total_episodes)
+        for vid_key in video_keys:
+            chunk_dir = "/".join(DEFAULT_VIDEO_PATH.split("/")[:-1]).format(
+                episode_chunk=ep_chunk, video_key=vid_key
+            )
+            (work_dir / chunk_dir).mkdir(parents=True, exist_ok=True)
+
+            for ep_idx in range(ep_chunk_start, ep_chunk_end):
+                target_path = DEFAULT_VIDEO_PATH.format(
+                    episode_chunk=ep_chunk, video_key=vid_key, episode_index=ep_idx
+                )
+                video_file = V1_VIDEO_FILE.format(video_key=vid_key, episode_index=ep_idx)
+                if len(video_dirs) == 1:
+                    video_path = video_dirs[0] / video_file
+                else:
+                    for dir in video_dirs:
+                        if (dir / video_file).is_file():
+                            video_path = dir / video_file
+                            break
+
+                video_path.rename(work_dir / target_path)
+
+    commit_message = "Move video files into chunk subdirectories"
+    subprocess.run(["git", "add", "."], cwd=work_dir, check=True)
+    subprocess.run(["git", "commit", "-m", commit_message], cwd=work_dir, check=True)
+    subprocess.run(["git", "push"], cwd=work_dir, check=True)
+
+
+def fix_lfs_video_files_tracking(work_dir: Path, lfs_untracked_videos: list[str]) -> None:
+    """
+    HACK: This function fixes the tracking by git lfs which was not properly set on some repos. In that case,
+    there's no other option than to download the actual files and reupload them with lfs tracking.
+    """
+    for i in range(0, len(lfs_untracked_videos), 100):
+        files = lfs_untracked_videos[i : i + 100]
+        try:
+            subprocess.run(["git", "rm", "--cached", *files], cwd=work_dir, capture_output=True, check=True)
+        except subprocess.CalledProcessError as e:
+            print("git rm --cached ERROR:")
+            print(e.stderr)
+        subprocess.run(["git", "add", *files], cwd=work_dir, check=True)
+
+    commit_message = "Track video files with git lfs"
+    subprocess.run(["git", "commit", "-m", commit_message], cwd=work_dir, check=True)
+    subprocess.run(["git", "push"], cwd=work_dir, check=True)
+
+
+def fix_gitattributes(work_dir: Path, current_gittatributes: Path, clean_gittatributes: Path) -> None:
+    shutil.copyfile(clean_gittatributes, current_gittatributes)
+    subprocess.run(["git", "add", ".gitattributes"], cwd=work_dir, check=True)
+    subprocess.run(["git", "commit", "-m", "Fix .gitattributes"], cwd=work_dir, check=True)
+    subprocess.run(["git", "push"], cwd=work_dir, check=True)
+
+
+def _lfs_clone(repo_id: str, work_dir: Path, branch: str) -> None:
+    subprocess.run(["git", "lfs", "install"], cwd=work_dir, check=True)
+    repo_url = f"https://huggingface.co/datasets/{repo_id}"
+    env = {"GIT_LFS_SKIP_SMUDGE": "1"}  # Prevent downloading LFS files
+    subprocess.run(
+        ["git", "clone", "--branch", branch, "--single-branch", "--depth", "1", repo_url, str(work_dir)],
+        check=True,
+        env=env,
+    )
+
+
+def _get_lfs_untracked_videos(work_dir: Path, video_files: list[str]) -> list[str]:
+    lfs_tracked_files = subprocess.run(
+        ["git", "lfs", "ls-files", "-n"], cwd=work_dir, capture_output=True, text=True, check=True
+    )
+    lfs_tracked_files = set(lfs_tracked_files.stdout.splitlines())
+    return [f for f in video_files if f not in lfs_tracked_files]
+
+
+def get_videos_info(repo_id: str, local_dir: Path, video_keys: list[str], branch: str) -> dict:
+    # Assumes first episode
+    video_files = [
+        DEFAULT_VIDEO_PATH.format(episode_chunk=0, video_key=vid_key, episode_index=0)
+        for vid_key in video_keys
+    ]
+    hub_api = HfApi()
+    hub_api.snapshot_download(
+        repo_id=repo_id, repo_type="dataset", local_dir=local_dir, revision=branch, allow_patterns=video_files
+    )
+    videos_info_dict = {}
+    for vid_key, vid_path in zip(video_keys, video_files, strict=True):
+        videos_info_dict[vid_key] = get_video_info(local_dir / vid_path)
+
+    return videos_info_dict
+
+
+def convert_dataset(
+    repo_id: str,
+    local_dir: Path,
+    single_task: str | None = None,
+    tasks_path: Path | None = None,
+    tasks_col: Path | None = None,
+    robot_config: dict | None = None,
+    test_branch: str | None = None,
+    **card_kwargs,
+):
+    v1 = get_hub_safe_version(repo_id, V16)
+    v1x_dir = local_dir / V16 / repo_id
+    v20_dir = local_dir / V20 / repo_id
+    v1x_dir.mkdir(parents=True, exist_ok=True)
+    v20_dir.mkdir(parents=True, exist_ok=True)
+
+    hub_api = HfApi()
+    hub_api.snapshot_download(
+        repo_id=repo_id, repo_type="dataset", revision=v1, local_dir=v1x_dir, ignore_patterns="videos*/"
+    )
+    branch = "main"
+    if test_branch:
+        branch = test_branch
+        create_branch(repo_id=repo_id, branch=test_branch, repo_type="dataset")
+
+    metadata_v1 = load_json(v1x_dir / V1_INFO_PATH)
+    dataset = datasets.load_dataset("parquet", data_dir=v1x_dir / "data", split="train")
+    features = get_features_from_hf_dataset(dataset, robot_config)
+    video_keys = [key for key, ft in features.items() if ft["dtype"] == "video"]
+
+    if single_task and "language_instruction" in dataset.column_names:
+        logging.warning(
+            "'single_task' provided but 'language_instruction' tasks_col found. Using 'language_instruction'.",
+        )
+        single_task = None
+        tasks_col = "language_instruction"
+
+    # Episodes & chunks
+    episode_indices = sorted(dataset.unique("episode_index"))
+    total_episodes = len(episode_indices)
+    assert episode_indices == list(range(total_episodes))
+    total_videos = total_episodes * len(video_keys)
+    total_chunks = total_episodes // DEFAULT_CHUNK_SIZE
+    if total_episodes % DEFAULT_CHUNK_SIZE != 0:
+        total_chunks += 1
+
+    # Tasks
+    if single_task:
+        tasks_by_episodes = {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:
+        tasks_by_episodes = load_json(tasks_path)
+        tasks_by_episodes = {int(ep_idx): task for ep_idx, task in tasks_by_episodes.items()}
+        dataset, tasks = add_task_index_by_episodes(dataset, tasks_by_episodes)
+        tasks_by_episodes = {ep_idx: [task] for ep_idx, task in tasks_by_episodes.items()}
+    elif tasks_col:
+        dataset, tasks, tasks_by_episodes = add_task_index_from_tasks_col(dataset, tasks_col)
+    else:
+        raise ValueError
+
+    assert set(tasks) == {task for ep_tasks in tasks_by_episodes.values() for task in ep_tasks}
+    tasks = [{"task_index": task_idx, "task": task} for task_idx, task in enumerate(tasks)]
+    write_jsonlines(tasks, v20_dir / TASKS_PATH)
+    features["task_index"] = {
+        "dtype": "int64",
+        "shape": (1,),
+        "names": None,
+    }
+
+    # Videos
+    if video_keys:
+        assert metadata_v1.get("video", False)
+        dataset = dataset.remove_columns(video_keys)
+        clean_gitattr = Path(
+            hub_api.hf_hub_download(
+                repo_id=GITATTRIBUTES_REF, repo_type="dataset", local_dir=local_dir, filename=".gitattributes"
+            )
+        ).absolute()
+        with tempfile.TemporaryDirectory() as tmp_video_dir:
+            move_videos(
+                repo_id, video_keys, total_episodes, total_chunks, Path(tmp_video_dir), clean_gitattr, branch
+            )
+        videos_info = get_videos_info(repo_id, v1x_dir, video_keys=video_keys, branch=branch)
+        for key in video_keys:
+            features[key]["shape"] = (
+                videos_info[key].pop("video.height"),
+                videos_info[key].pop("video.width"),
+                videos_info[key].pop("video.channels"),
+            )
+            features[key]["video_info"] = videos_info[key]
+            assert math.isclose(videos_info[key]["video.fps"], metadata_v1["fps"], rel_tol=1e-3)
+            if "encoding" in metadata_v1:
+                assert videos_info[key]["video.pix_fmt"] == metadata_v1["encoding"]["pix_fmt"]
+    else:
+        assert metadata_v1.get("video", 0) == 0
+        videos_info = None
+
+    # Split data into 1 parquet file by episode
+    episode_lengths = split_parquet_by_episodes(dataset, total_episodes, total_chunks, v20_dir)
+
+    if robot_config is not None:
+        robot_type = robot_config["robot_type"]
+        repo_tags = [robot_type]
+    else:
+        robot_type = "unknown"
+        repo_tags = None
+
+    # Episodes
+    episodes = [
+        {"episode_index": ep_idx, "tasks": tasks_by_episodes[ep_idx], "length": episode_lengths[ep_idx]}
+        for ep_idx in episode_indices
+    ]
+    write_jsonlines(episodes, v20_dir / EPISODES_PATH)
+
+    # Assemble metadata v2.0
+    metadata_v2_0 = {
+        "codebase_version": V20,
+        "robot_type": robot_type,
+        "total_episodes": total_episodes,
+        "total_frames": len(dataset),
+        "total_tasks": len(tasks),
+        "total_videos": total_videos,
+        "total_chunks": total_chunks,
+        "chunks_size": DEFAULT_CHUNK_SIZE,
+        "fps": metadata_v1["fps"],
+        "splits": {"train": f"0:{total_episodes}"},
+        "data_path": DEFAULT_PARQUET_PATH,
+        "video_path": DEFAULT_VIDEO_PATH if video_keys else None,
+        "features": features,
+    }
+    write_json(metadata_v2_0, v20_dir / INFO_PATH)
+    convert_stats_to_json(v1x_dir, v20_dir)
+    card = create_lerobot_dataset_card(tags=repo_tags, dataset_info=metadata_v2_0, **card_kwargs)
+
+    with contextlib.suppress(EntryNotFoundError, HfHubHTTPError):
+        hub_api.delete_folder(repo_id=repo_id, path_in_repo="data", repo_type="dataset", revision=branch)
+
+    with contextlib.suppress(EntryNotFoundError, HfHubHTTPError):
+        hub_api.delete_folder(repo_id=repo_id, path_in_repo="meta_data", repo_type="dataset", revision=branch)
+
+    with contextlib.suppress(EntryNotFoundError, HfHubHTTPError):
+        hub_api.delete_folder(repo_id=repo_id, path_in_repo="meta", repo_type="dataset", revision=branch)
+
+    hub_api.upload_folder(
+        repo_id=repo_id,
+        path_in_repo="data",
+        folder_path=v20_dir / "data",
+        repo_type="dataset",
+        revision=branch,
+    )
+    hub_api.upload_folder(
+        repo_id=repo_id,
+        path_in_repo="meta",
+        folder_path=v20_dir / "meta",
+        repo_type="dataset",
+        revision=branch,
+    )
+
+    card.push_to_hub(repo_id=repo_id, repo_type="dataset", revision=branch)
+
+    if not test_branch:
+        create_branch(repo_id=repo_id, branch=V20, repo_type="dataset")
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    task_args = parser.add_mutually_exclusive_group(required=True)
+
+    parser.add_argument(
+        "--repo-id",
+        type=str,
+        required=True,
+        help="Repository identifier on Hugging Face: a community or a user name `/` the name of the dataset (e.g. `lerobot/pusht`, `cadene/aloha_sim_insertion_human`).",
+    )
+    task_args.add_argument(
+        "--single-task",
+        type=str,
+        help="A short but accurate description of the single task performed in the dataset.",
+    )
+    task_args.add_argument(
+        "--tasks-col",
+        type=str,
+        help="The name of the column containing language instructions",
+    )
+    task_args.add_argument(
+        "--tasks-path",
+        type=Path,
+        help="The path to a .json file containing one language instruction for each episode_index",
+    )
+    parser.add_argument(
+        "--robot-config",
+        type=Path,
+        default=None,
+        help="Path to the robot's config yaml the dataset during conversion.",
+    )
+    parser.add_argument(
+        "--robot-overrides",
+        type=str,
+        nargs="*",
+        help="Any key=value arguments to override the robot config values (use dots for.nested=overrides)",
+    )
+    parser.add_argument(
+        "--local-dir",
+        type=Path,
+        default=None,
+        help="Local directory to store the dataset during conversion. Defaults to /tmp/lerobot_dataset_v2",
+    )
+    parser.add_argument(
+        "--license",
+        type=str,
+        default="apache-2.0",
+        help="Repo license. Must be one of https://huggingface.co/docs/hub/repositories-licenses. Defaults to mit.",
+    )
+    parser.add_argument(
+        "--test-branch",
+        type=str,
+        default=None,
+        help="Repo branch to test your conversion first (e.g. 'v2.0.test')",
+    )
+
+    args = parser.parse_args()
+    if not args.local_dir:
+        args.local_dir = Path("/tmp/lerobot_dataset_v2")
+
+    robot_config = parse_robot_config(args.robot_config, args.robot_overrides) if args.robot_config else None
+    del args.robot_config, args.robot_overrides
+
+    convert_dataset(**vars(args), robot_config=robot_config)
+
+
+if __name__ == "__main__":
+    main()

lerobot/common/datasets/video_utils.py

@@ -13,6 +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 json
 import logging
 import subprocess
 import warnings
@@ -25,47 +26,11 @@ import pyarrow as pa
 import torch
 import torchvision
 from datasets.features.features import register_feature
-
-
-def load_from_videos(
-    item: dict[str, torch.Tensor],
-    video_frame_keys: list[str],
-    videos_dir: Path,
-    tolerance_s: float,
-    backend: str = "pyav",
-):
-    """Note: When using data workers (e.g. DataLoader with num_workers>0), do not call this function
-    in the main process (e.g. by using a second Dataloader with num_workers=0). It will result in a Segmentation Fault.
-    This probably happens because a memory reference to the video loader is created in the main process and a
-    subprocess fails to access it.
-    """
-    # since video path already contains "videos" (e.g. videos_dir="data/videos", path="videos/episode_0.mp4")
-    data_dir = videos_dir.parent
-
-    for key in video_frame_keys:
-        if isinstance(item[key], list):
-            # load multiple frames at once (expected when delta_timestamps is not None)
-            timestamps = [frame["timestamp"] for frame in item[key]]
-            paths = [frame["path"] for frame in item[key]]
-            if len(set(paths)) > 1:
-                raise NotImplementedError("All video paths are expected to be the same for now.")
-            video_path = data_dir / paths[0]
-
-            frames = decode_video_frames_torchvision(video_path, timestamps, tolerance_s, backend)
-            item[key] = frames
-        else:
-            # load one frame
-            timestamps = [item[key]["timestamp"]]
-            video_path = data_dir / item[key]["path"]
-
-            frames = decode_video_frames_torchvision(video_path, timestamps, tolerance_s, backend)
-            item[key] = frames[0]
-
-    return item
+from PIL import Image
 
 
 def decode_video_frames_torchvision(
-    video_path: str,
+    video_path: Path | str,
     timestamps: list[float],
     tolerance_s: float,
     backend: str = "pyav",
@@ -163,8 +128,8 @@ def decode_video_frames_torchvision(
 
 
 def encode_video_frames(
-    imgs_dir: Path,
-    video_path: Path,
+    imgs_dir: Path | str,
+    video_path: Path | str,
     fps: int,
     vcodec: str = "libsvtav1",
     pix_fmt: str = "yuv420p",
@@ -247,3 +212,104 @@ with warnings.catch_warnings():
     )
     # to make VideoFrame available in HuggingFace `datasets`
     register_feature(VideoFrame, "VideoFrame")
+
+
+def get_audio_info(video_path: Path | str) -> dict:
+    ffprobe_audio_cmd = [
+        "ffprobe",
+        "-v",
+        "error",
+        "-select_streams",
+        "a:0",
+        "-show_entries",
+        "stream=channels,codec_name,bit_rate,sample_rate,bit_depth,channel_layout,duration",
+        "-of",
+        "json",
+        str(video_path),
+    ]
+    result = subprocess.run(ffprobe_audio_cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
+    if result.returncode != 0:
+        raise RuntimeError(f"Error running ffprobe: {result.stderr}")
+
+    info = json.loads(result.stdout)
+    audio_stream_info = info["streams"][0] if info.get("streams") else None
+    if audio_stream_info is None:
+        return {"has_audio": False}
+
+    # Return the information, defaulting to None if no audio stream is present
+    return {
+        "has_audio": True,
+        "audio.channels": audio_stream_info.get("channels", None),
+        "audio.codec": audio_stream_info.get("codec_name", None),
+        "audio.bit_rate": int(audio_stream_info["bit_rate"]) if audio_stream_info.get("bit_rate") else None,
+        "audio.sample_rate": int(audio_stream_info["sample_rate"])
+        if audio_stream_info.get("sample_rate")
+        else None,
+        "audio.bit_depth": audio_stream_info.get("bit_depth", None),
+        "audio.channel_layout": audio_stream_info.get("channel_layout", None),
+    }
+
+
+def get_video_info(video_path: Path | str) -> dict:
+    ffprobe_video_cmd = [
+        "ffprobe",
+        "-v",
+        "error",
+        "-select_streams",
+        "v:0",
+        "-show_entries",
+        "stream=r_frame_rate,width,height,codec_name,nb_frames,duration,pix_fmt",
+        "-of",
+        "json",
+        str(video_path),
+    ]
+    result = subprocess.run(ffprobe_video_cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
+    if result.returncode != 0:
+        raise RuntimeError(f"Error running ffprobe: {result.stderr}")
+
+    info = json.loads(result.stdout)
+    video_stream_info = info["streams"][0]
+
+    # Calculate fps from r_frame_rate
+    r_frame_rate = video_stream_info["r_frame_rate"]
+    num, denom = map(int, r_frame_rate.split("/"))
+    fps = num / denom
+
+    pixel_channels = get_video_pixel_channels(video_stream_info["pix_fmt"])
+
+    video_info = {
+        "video.fps": fps,
+        "video.height": video_stream_info["height"],
+        "video.width": video_stream_info["width"],
+        "video.channels": pixel_channels,
+        "video.codec": video_stream_info["codec_name"],
+        "video.pix_fmt": video_stream_info["pix_fmt"],
+        "video.is_depth_map": False,
+        **get_audio_info(video_path),
+    }
+
+    return video_info
+
+
+def get_video_pixel_channels(pix_fmt: str) -> int:
+    if "gray" in pix_fmt or "depth" in pix_fmt or "monochrome" in pix_fmt:
+        return 1
+    elif "rgba" in pix_fmt or "yuva" in pix_fmt:
+        return 4
+    elif "rgb" in pix_fmt or "yuv" in pix_fmt:
+        return 3
+    else:
+        raise ValueError("Unknown format")
+
+
+def get_image_pixel_channels(image: Image):
+    if image.mode == "L":
+        return 1  # Grayscale
+    elif image.mode == "LA":
+        return 2  # Grayscale + Alpha
+    elif image.mode == "RGB":
+        return 3  # RGB
+    elif image.mode == "RGBA":
+        return 4  # RGBA
+    else:
+        raise ValueError("Unknown format")

lerobot/common/envs/factory.py

@@ -14,9 +14,13 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import importlib
+from collections import deque
 
 import gymnasium as gym
+import numpy as np
+import torch
 from omegaconf import DictConfig
+from mani_skill.utils import common
 
 
 def make_env(cfg: DictConfig, n_envs: int | None = None) -> gym.vector.VectorEnv | None:
@@ -30,6 +34,10 @@ def make_env(cfg: DictConfig, n_envs: int | None = None) -> gym.vector.VectorEnv
     if cfg.env.name == "real_world":
         return
 
+    if "maniskill" in cfg.env.name:
+        env = make_maniskill_env(cfg, n_envs if n_envs is not None else cfg.eval.batch_size)
+        return env
+
     package_name = f"gym_{cfg.env.name}"
 
     try:
@@ -56,3 +64,86 @@ def make_env(cfg: DictConfig, n_envs: int | None = None) -> gym.vector.VectorEnv
     )
 
     return env
+
+
+def make_maniskill_env(cfg: DictConfig, n_envs: int | None = None) -> gym.vector.VectorEnv | None:
+    """Make ManiSkill3 gym environment"""
+    from mani_skill.vector.wrappers.gymnasium import ManiSkillVectorEnv
+
+    env = gym.make(
+        cfg.env.task,
+        obs_mode=cfg.env.obs,
+        control_mode=cfg.env.control_mode,
+        render_mode=cfg.env.render_mode,
+        sensor_configs=dict(width=cfg.env.image_size, height=cfg.env.image_size),
+        num_envs=n_envs,
+    )
+    # cfg.env_cfg.control_mode = cfg.eval_env_cfg.control_mode = env.control_mode
+    env = ManiSkillVectorEnv(env, ignore_terminations=True)
+    # state should have the size of 25
+    # env = ConvertToLeRobotEnv(env, n_envs)
+    # env = PixelWrapper(cfg, env, n_envs)
+    env._max_episode_steps = env.max_episode_steps = 50  # gym_utils.find_max_episode_steps_value(env)
+    env.unwrapped.metadata["render_fps"] = 20
+
+    return env
+
+
+class PixelWrapper(gym.Wrapper):
+    """
+    Wrapper for pixel observations. Works with Maniskill vectorized environments
+    """
+
+    def __init__(self, cfg, env, num_envs, num_frames=3):
+        super().__init__(env)
+        self.cfg = cfg
+        self.env = env
+        self.observation_space = gym.spaces.Box(
+            low=0,
+            high=255,
+            shape=(num_envs, num_frames * 3, cfg.env.render_size, cfg.env.render_size),
+            dtype=np.uint8,
+        )
+        self._frames = deque([], maxlen=num_frames)
+        self._render_size = cfg.env.render_size
+
+    def _get_obs(self, obs):
+        frame = obs["sensor_data"]["base_camera"]["rgb"].cpu().permute(0, 3, 1, 2)
+        self._frames.append(frame)
+        return {"pixels": torch.from_numpy(np.concatenate(self._frames, axis=1)).to(self.env.device)}
+
+    def reset(self, seed):
+        obs, info = self.env.reset()  # (seed=seed)
+        for _ in range(self._frames.maxlen):
+            obs_frames = self._get_obs(obs)
+        return obs_frames, info
+
+    def step(self, action):
+        obs, reward, terminated, truncated, info = self.env.step(action)
+        return self._get_obs(obs), reward, terminated, truncated, info
+
+class ConvertToLeRobotEnv(gym.Wrapper):
+    def __init__(self, env, num_envs):
+        super().__init__(env)
+    def reset(self, seed=None, options=None):
+        obs, info = self.env.reset(seed=seed, options={})
+        return self._get_obs(obs), info
+    def step(self, action):
+        obs, reward, terminated, truncated, info = self.env.step(action)
+        return self._get_obs(obs), reward, terminated, truncated, info
+    def _get_obs(self, observation):
+        sensor_data = observation.pop("sensor_data")
+        del observation["sensor_param"]
+        images = []
+        for cam_data in sensor_data.values():
+                images.append(cam_data["rgb"])
+
+        images = torch.concat(images, axis=-1)
+        # flatten the rest of the data which should just be state data
+        observation = common.flatten_state_dict(
+            observation, use_torch=True, device=self.base_env.device
+        )
+        ret = dict()
+        ret["state"] = observation
+        ret["pixels"] = images
+        return ret

lerobot/common/envs/utils.py

@@ -28,6 +28,9 @@ def preprocess_observation(observations: dict[str, np.ndarray]) -> dict[str, Ten
     """
     # map to expected inputs for the policy
     return_observations = {}
+    # TODO: You have to merge all tensors from agent key and extra key
+    # You don't keep sensor param key in the observation
+    # And you keep sensor data rgb
     if "pixels" in observations:
         if isinstance(observations["pixels"], dict):
             imgs = {f"observation.images.{key}": img for key, img in observations["pixels"].items()}
@@ -50,6 +53,8 @@ def preprocess_observation(observations: dict[str, np.ndarray]) -> dict[str, Ten
             img /= 255
 
             return_observations[imgkey] = img
+        # obs state agent qpos and qvel
+        # image
 
     if "environment_state" in observations:
         return_observations["observation.environment_state"] = torch.from_numpy(
@@ -60,3 +65,38 @@ def preprocess_observation(observations: dict[str, np.ndarray]) -> dict[str, Ten
     # requirement for "agent_pos"
     return_observations["observation.state"] = torch.from_numpy(observations["agent_pos"]).float()
     return return_observations
+
+
+def preprocess_maniskill_observation(observations: dict[str, np.ndarray]) -> dict[str, Tensor]:
+    """Convert environment observation to LeRobot format observation.
+    Args:
+        observation: Dictionary of observation batches from a Gym vector environment.
+    Returns:
+        Dictionary of observation batches with keys renamed to LeRobot format and values as tensors.
+    """
+    # map to expected inputs for the policy
+    return_observations = {}
+    # TODO: You have to merge all tensors from agent key and extra key
+    # You don't keep sensor param key in the observation
+    # And you keep sensor data rgb
+    q_pos = observations["agent"]["qpos"]
+    q_vel = observations["agent"]["qvel"]
+    tcp_pos = observations["extra"]["tcp_pose"]
+    img = observations["sensor_data"]["base_camera"]["rgb"]
+
+    _, h, w, c = img.shape
+    assert c < h and c < w, f"expect channel last images, but instead got {img.shape=}"
+
+    # sanity check that images are uint8
+    assert img.dtype == torch.uint8, f"expect torch.uint8, but instead {img.dtype=}"
+
+    # convert to channel first of type float32 in range [0,1]
+    img = einops.rearrange(img, "b h w c -> b c h w").contiguous()
+    img = img.type(torch.float32)
+    img /= 255
+
+    state = torch.cat([q_pos, q_vel, tcp_pos], dim=-1)
+
+    return_observations["observation.image"] = img
+    return_observations["observation.state"] = state
+    return return_observations

lerobot/common/logger.py

@@ -25,6 +25,7 @@ from glob import glob
 from pathlib import Path
 
 import torch
+import wandb
 from huggingface_hub.constants import SAFETENSORS_SINGLE_FILE
 from omegaconf import DictConfig, OmegaConf
 from termcolor import colored
@@ -107,8 +108,6 @@ class Logger:
             self._wandb = None
         else:
             os.environ["WANDB_SILENT"] = "true"
-            import wandb
-
             wandb_run_id = None
             if cfg.resume:
                 wandb_run_id = get_wandb_run_id_from_filesystem(self.checkpoints_dir)
@@ -232,7 +231,7 @@ class Logger:
         # TODO(alexander-soare): Add local text log.
         if self._wandb is not None:
             for k, v in d.items():
-                if not isinstance(v, (int, float, str)):
+                if not isinstance(v, (int, float, str, wandb.Table)):
                     logging.warning(
                         f'WandB logging of key "{k}" was ignored as its type is not handled by this wrapper.'
                     )

lerobot/common/policies/factory.py

@@ -66,6 +66,11 @@ def get_policy_and_config_classes(name: str) -> tuple[Policy, object]:
         from lerobot.common.policies.vqbet.modeling_vqbet import VQBeTPolicy
 
         return VQBeTPolicy, VQBeTConfig
+    elif name == "sac":
+        from lerobot.common.policies.sac.configuration_sac import SACConfig
+        from lerobot.common.policies.sac.modeling_sac import SACPolicy
+
+        return SACPolicy, SACConfig
     else:
         raise NotImplementedError(f"Policy with name {name} is not implemented.")
 
@@ -85,10 +90,10 @@ def make_policy(
             be provided when initializing a new policy, and must not be provided when loading a pretrained
             policy. Therefore, this argument is mutually exclusive with `pretrained_policy_name_or_path`.
     """
-    if not (pretrained_policy_name_or_path is None) ^ (dataset_stats is None):
-        raise ValueError(
-            "Exactly one of `pretrained_policy_name_or_path` and `dataset_stats` must be provided."
-        )
+    # if not (pretrained_policy_name_or_path is None) ^ (dataset_stats is None):
+    #     raise ValueError(
+    #         "Exactly one of `pretrained_policy_name_or_path` and `dataset_stats` must be provided."
+    #     )
 
     policy_cls, policy_cfg_class = get_policy_and_config_classes(hydra_cfg.policy.name)
 

lerobot/common/policies/hilserl/classifier/configuration_classifier.py

@@ -0,0 +1,35 @@
+import json
+import os
+from dataclasses import asdict, dataclass
+
+
+@dataclass
+class ClassifierConfig:
+    """Configuration for the Classifier model."""
+
+    num_classes: int = 2
+    hidden_dim: int = 256
+    dropout_rate: float = 0.1
+    model_name: str = "microsoft/resnet-50"
+    device: str = "cpu"
+    model_type: str = "cnn"  # "transformer" or "cnn"
+    num_cameras: int = 2
+
+    def save_pretrained(self, save_dir):
+        """Save config to json file."""
+        os.makedirs(save_dir, exist_ok=True)
+
+        # Convert to dict and save as JSON
+        config_dict = asdict(self)
+        with open(os.path.join(save_dir, "config.json"), "w") as f:
+            json.dump(config_dict, f, indent=2)
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path):
+        """Load config from json file."""
+        config_file = os.path.join(pretrained_model_name_or_path, "config.json")
+
+        with open(config_file) as f:
+            config_dict = json.load(f)
+
+        return cls(**config_dict)

lerobot/common/policies/hilserl/classifier/modeling_classifier.py

@@ -0,0 +1,151 @@
+import logging
+from typing import Optional
+
+import torch
+from huggingface_hub import PyTorchModelHubMixin
+from torch import Tensor, nn
+
+from .configuration_classifier import ClassifierConfig
+
+logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(name)s - %(levelname)s - %(message)s")
+logger = logging.getLogger(__name__)
+
+
+class ClassifierOutput:
+    """Wrapper for classifier outputs with additional metadata."""
+
+    def __init__(
+        self, logits: Tensor, probabilities: Optional[Tensor] = None, hidden_states: Optional[Tensor] = None
+    ):
+        self.logits = logits
+        self.probabilities = probabilities
+        self.hidden_states = hidden_states
+
+    def __repr__(self):
+        return (
+            f"ClassifierOutput(logits={self.logits}, "
+            f"probabilities={self.probabilities}, "
+            f"hidden_states={self.hidden_states})"
+        )
+
+
+class Classifier(
+    nn.Module,
+    PyTorchModelHubMixin,
+    # Add Hub metadata
+    library_name="lerobot",
+    repo_url="https://github.com/huggingface/lerobot",
+    tags=["robotics", "vision-classifier"],
+):
+    """Image classifier built on top of a pre-trained encoder."""
+
+    # Add name attribute for factory
+    name = "classifier"
+
+    def __init__(self, config: ClassifierConfig):
+        from transformers import AutoImageProcessor, AutoModel
+
+        super().__init__()
+        self.config = config
+        self.processor = AutoImageProcessor.from_pretrained(self.config.model_name, trust_remote_code=True)
+        encoder = AutoModel.from_pretrained(self.config.model_name, trust_remote_code=True)
+        # Extract vision model if we're given a multimodal model
+        if hasattr(encoder, "vision_model"):
+            logging.info("Multimodal model detected - using vision encoder only")
+            self.encoder = encoder.vision_model
+            self.vision_config = encoder.config.vision_config
+        else:
+            self.encoder = encoder
+            self.vision_config = getattr(encoder, "config", None)
+
+        # Model type from config
+        self.is_cnn = self.config.model_type == "cnn"
+
+        # For CNNs, initialize backbone
+        if self.is_cnn:
+            self._setup_cnn_backbone()
+
+        self._freeze_encoder()
+        self._build_classifier_head()
+
+    def _setup_cnn_backbone(self):
+        """Set up CNN encoder"""
+        if hasattr(self.encoder, "fc"):
+            self.feature_dim = self.encoder.fc.in_features
+            self.encoder = nn.Sequential(*list(self.encoder.children())[:-1])
+        elif hasattr(self.encoder.config, "hidden_sizes"):
+            self.feature_dim = self.encoder.config.hidden_sizes[-1]  # Last channel dimension
+        else:
+            raise ValueError("Unsupported CNN architecture")
+
+        self.encoder = self.encoder.to(self.config.device)
+
+    def _freeze_encoder(self) -> None:
+        """Freeze the encoder parameters."""
+        for param in self.encoder.parameters():
+            param.requires_grad = False
+
+    def _build_classifier_head(self) -> None:
+        """Initialize the classifier head architecture."""
+        # Get input dimension based on model type
+        if self.is_cnn:
+            input_dim = self.feature_dim
+        else:  # Transformer models
+            if hasattr(self.encoder.config, "hidden_size"):
+                input_dim = self.encoder.config.hidden_size
+            else:
+                raise ValueError("Unsupported transformer architecture since hidden_size is not found")
+
+        self.classifier_head = nn.Sequential(
+            nn.Linear(input_dim * self.config.num_cameras, self.config.hidden_dim),
+            nn.Dropout(self.config.dropout_rate),
+            nn.LayerNorm(self.config.hidden_dim),
+            nn.ReLU(),
+            nn.Linear(self.config.hidden_dim, 1 if self.config.num_classes == 2 else self.config.num_classes),
+        )
+        self.classifier_head = self.classifier_head.to(self.config.device)
+
+    def _get_encoder_output(self, x: torch.Tensor) -> torch.Tensor:
+        """Extract the appropriate output from the encoder."""
+        # Process images with the processor (handles resizing and normalization)
+        processed = self.processor(
+            images=x,  # LeRobotDataset already provides proper tensor format
+            return_tensors="pt",
+        )
+        processed = processed["pixel_values"].to(x.device)
+
+        with torch.no_grad():
+            if self.is_cnn:
+                # The HF ResNet applies pooling internally
+                outputs = self.encoder(processed)
+                # Get pooled output directly
+                features = outputs.pooler_output
+
+                if features.dim() > 2:
+                    features = features.squeeze(-1).squeeze(-1)
+                return features
+            else:  # Transformer models
+                outputs = self.encoder(processed)
+                if hasattr(outputs, "pooler_output") and outputs.pooler_output is not None:
+                    return outputs.pooler_output
+                return outputs.last_hidden_state[:, 0, :]
+
+    def forward(self, xs: torch.Tensor) -> ClassifierOutput:
+        """Forward pass of the classifier."""
+        # For training, we expect input to be a tensor directly from LeRobotDataset
+        encoder_outputs = torch.hstack([self._get_encoder_output(x) for x in xs])
+        logits = self.classifier_head(encoder_outputs)
+
+        if self.config.num_classes == 2:
+            logits = logits.squeeze(-1)
+            probabilities = torch.sigmoid(logits)
+        else:
+            probabilities = torch.softmax(logits, dim=-1)
+
+        return ClassifierOutput(logits=logits, probabilities=probabilities, hidden_states=encoder_outputs)
+
+    def predict_reward(self, x):
+        if self.config.num_classes == 2:
+            return (self.forward(x).probabilities > 0.5).float()
+        else:
+            return torch.argmax(self.forward(x).probabilities, dim=1)

lerobot/common/policies/hilserl/configuration_hilserl.py

@@ -1,6 +1,7 @@
 #!/usr/bin/env python
 
-# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+# 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.
@@ -14,23 +15,9 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from pathlib import Path
-
-import pytest
-
-from lerobot.scripts.visualize_dataset_html import visualize_dataset_html
+from dataclasses import dataclass
 
 
-@pytest.mark.parametrize(
-    "repo_id",
-    ["lerobot/pusht"],
-)
-def test_visualize_dataset_html(tmpdir, repo_id):
-    tmpdir = Path(tmpdir)
-    visualize_dataset_html(
-        repo_id,
-        episodes=[0],
-        output_dir=tmpdir,
-        serve=False,
-    )
-    assert (tmpdir / "static" / "episode_0.csv").exists()
+@dataclass
+class HILSerlConfig:
+    pass

lerobot/common/policies/hilserl/modeling_hilserl.py

@@ -0,0 +1,29 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team.
+# All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch.nn as nn
+from huggingface_hub import PyTorchModelHubMixin
+
+
+class HILSerlPolicy(
+    nn.Module,
+    PyTorchModelHubMixin,
+    library_name="lerobot",
+    repo_url="https://github.com/huggingface/lerobot",
+    tags=["robotics", "hilserl"],
+):
+    pass

lerobot/common/policies/sac/configuration_sac.py

@@ -0,0 +1,83 @@
+#!/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.
+
+from dataclasses import dataclass, field
+from typing import Any
+
+
+@dataclass
+class SACConfig:
+    input_shapes: dict[str, list[int]] = field(
+        default_factory=lambda: {
+            "observation.image": [3, 84, 84],
+            "observation.state": [4],
+        }
+    )
+    output_shapes: dict[str, list[int]] = field(
+        default_factory=lambda: {
+            "action": [2],
+        }
+    )
+    input_normalization_modes: dict[str, str] = field(
+        default_factory=lambda: {
+            "observation.image": "mean_std",
+            "observation.state": "min_max",
+            "observation.environment_state": "min_max",
+        }
+    )
+    output_normalization_modes: dict[str, str] = field(default_factory=lambda: {"action": "min_max"})
+    output_normalization_params: dict[str, dict[str, list[float]]] = field(
+        default_factory=lambda: {
+            "action": {"min": [-1, -1], "max": [1, 1]},
+        }
+    )
+    camera_number: int = 1
+    # Add type annotations for these fields:
+    image_encoder_hidden_dim: int = 32
+    shared_encoder: bool = False
+    discount: float = 0.99
+    temperature_init: float = 1.0
+    num_critics: int = 2
+    num_subsample_critics: int | None = None
+    critic_lr: float = 3e-4
+    actor_lr: float = 3e-4
+    temperature_lr: float = 3e-4
+    critic_target_update_weight: float = 0.005
+    utd_ratio: int = 1  # If you want enable utd_ratio, you need to set it to >1
+    state_encoder_hidden_dim: int = 256
+    latent_dim: int = 256
+    target_entropy: float | None = None
+    use_backup_entropy: bool = True
+    critic_network_kwargs: dict[str, Any] = field(
+        default_factory=lambda: {
+            "hidden_dims": [256, 256],
+            "activate_final": True,
+        }
+    )
+    actor_network_kwargs: dict[str, Any] = field(
+        default_factory=lambda: {
+            "hidden_dims": [256, 256],
+            "activate_final": True,
+        }
+    )
+    policy_kwargs: dict[str, Any] = field(
+        default_factory=lambda: {
+            "use_tanh_squash": True,
+            "log_std_min": -5,
+            "log_std_max": 2,
+        }
+    )

lerobot/common/policies/sac/modeling_sac.py

@@ -0,0 +1,571 @@
+#!/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.
+
+# TODO: (1) better device management
+
+from collections import deque
+from typing import Callable, Optional, Sequence, Tuple, Union
+
+import einops
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F  # noqa: N812
+from huggingface_hub import PyTorchModelHubMixin
+from torch import Tensor
+
+from lerobot.common.policies.normalize import Normalize, Unnormalize
+from lerobot.common.policies.sac.configuration_sac import SACConfig
+
+
+class SACPolicy(
+    nn.Module,
+    PyTorchModelHubMixin,
+    library_name="lerobot",
+    repo_url="https://github.com/huggingface/lerobot",
+    tags=["robotics", "RL", "SAC"],
+):
+    name = "sac"
+
+    def __init__(
+        self,
+        config: SACConfig | None = None,
+        dataset_stats: dict[str, dict[str, Tensor]] | None = None,
+        device: str = "cpu",
+    ):
+        super().__init__()
+
+        if config is None:
+            config = SACConfig()
+        self.config = config
+        if config.input_normalization_modes is not None:
+            self.normalize_inputs = Normalize(
+                config.input_shapes, config.input_normalization_modes, dataset_stats
+            )
+        else:
+            self.normalize_inputs = nn.Identity()
+
+        output_normalization_params = {}
+        for outer_key, inner_dict in config.output_normalization_params.items():
+            output_normalization_params[outer_key] = {}
+            for key, value in inner_dict.items():
+                output_normalization_params[outer_key][key] = torch.tensor(value)
+
+        # HACK: This is hacky and should be removed
+        dataset_stats = dataset_stats or output_normalization_params
+        self.normalize_targets = Normalize(
+            config.output_shapes, config.output_normalization_modes, dataset_stats
+        )
+        self.unnormalize_outputs = Unnormalize(
+            config.output_shapes, config.output_normalization_modes, dataset_stats
+        )
+
+        if config.shared_encoder:
+            encoder_critic = SACObservationEncoder(config)
+            encoder_actor: SACObservationEncoder = encoder_critic
+        else:
+            encoder_critic = SACObservationEncoder(config)
+            encoder_actor = SACObservationEncoder(config)
+        # Define networks
+        critic_nets = []
+        for _ in range(config.num_critics):
+            critic_net = Critic(
+                encoder=encoder_critic,
+                network=MLP(
+                    input_dim=encoder_critic.output_dim + config.output_shapes["action"][0],
+                    **config.critic_network_kwargs,
+                ),
+                device=device,
+            )
+            critic_nets.append(critic_net)
+
+        target_critic_nets = []
+        for _ in range(config.num_critics):
+            target_critic_net = Critic(
+                encoder=encoder_critic,
+                network=MLP(
+                    input_dim=encoder_critic.output_dim + config.output_shapes["action"][0],
+                    **config.critic_network_kwargs,
+                ),
+                device=device,
+            )
+            target_critic_nets.append(target_critic_net)
+
+        self.critic_ensemble = create_critic_ensemble(
+            critics=critic_nets, num_critics=config.num_critics, device=device
+        )
+        self.critic_target = create_critic_ensemble(
+            critics=target_critic_nets, num_critics=config.num_critics, device=device
+        )
+        self.critic_target.load_state_dict(self.critic_ensemble.state_dict())
+
+        self.actor = Policy(
+            encoder=encoder_actor,
+            network=MLP(input_dim=encoder_actor.output_dim, **config.actor_network_kwargs),
+            action_dim=config.output_shapes["action"][0],
+            device=device,
+            encoder_is_shared=config.shared_encoder,
+            **config.policy_kwargs,
+        )
+        if config.target_entropy is None:
+            config.target_entropy = -np.prod(config.output_shapes["action"][0]) / 2  # (-dim(A)/2)
+        # TODO: Handle the case where the temparameter is a fixed
+        self.log_alpha = torch.zeros(1, requires_grad=True, device=device)
+        self.temperature = self.log_alpha.exp().item()
+
+    def reset(self):
+        """Reset the policy"""
+        pass
+
+    @torch.no_grad()
+    def select_action(self, batch: dict[str, Tensor]) -> Tensor:
+        """Select action for inference/evaluation"""
+        actions, _, _ = self.actor(batch)
+        actions = self.unnormalize_outputs({"action": actions})["action"]
+        return actions
+    
+    def critic_forward(self, observations: dict[str, Tensor], actions: Tensor, use_target: bool = False) -> Tensor:
+        """Forward pass through a critic network ensemble
+        
+        Args:
+            observations: Dictionary of observations
+            actions: Action tensor
+            use_target: If True, use target critics, otherwise use ensemble critics
+        
+        Returns:
+            Tensor of Q-values from all critics
+        """
+        critics = self.critic_target if use_target else self.critic_ensemble
+        q_values = torch.stack([critic(observations, actions) for critic in critics])
+        return q_values
+
+
+    def critic_forward(
+        self, observations: dict[str, Tensor], actions: Tensor, use_target: bool = False
+    ) -> Tensor:
+        """Forward pass through a critic network ensemble
+
+        Args:
+            observations: Dictionary of observations
+            actions: Action tensor
+            use_target: If True, use target critics, otherwise use ensemble critics
+
+        Returns:
+            Tensor of Q-values from all critics
+        """
+        critics = self.critic_target if use_target else self.critic_ensemble
+        q_values = torch.stack([critic(observations, actions) for critic in critics])
+        return q_values
+
+    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor | float]: ...
+    def update_target_networks(self):
+        """Update target networks with exponential moving average"""
+        for target_critic, critic in zip(self.critic_target, self.critic_ensemble, strict=False):
+            for target_param, param in zip(target_critic.parameters(), critic.parameters(), strict=False):
+                target_param.data.copy_(
+                    param.data * self.config.critic_target_update_weight
+                    + target_param.data * (1.0 - self.config.critic_target_update_weight)
+                )
+
+    def compute_loss_critic(self, observations, actions, rewards, next_observations, done) -> Tensor:
+        temperature = self.log_alpha.exp().item()
+        with torch.no_grad():
+            next_action_preds, next_log_probs, _ = self.actor(next_observations)
+
+            # 2- compute q targets
+            q_targets = self.critic_forward(
+                observations=next_observations, actions=next_action_preds, use_target=True
+            )
+
+            # subsample critics to prevent overfitting if use high UTD (update to date)
+            if self.config.num_subsample_critics is not None:
+                indices = torch.randperm(self.config.num_critics)
+                indices = indices[: self.config.num_subsample_critics]
+                q_targets = q_targets[indices]
+
+            # critics subsample size
+            min_q, _ = q_targets.min(dim=0)  # Get values from min operation
+            if self.config.use_backup_entropy:
+                min_q = min_q - (temperature * next_log_probs)
+
+            td_target = rewards + (1 - done) * self.config.discount * min_q
+
+        # 3- compute predicted qs
+        q_preds = self.critic_forward(observations, actions, use_target=False)
+
+        # 4- Calculate loss
+        # Compute state-action value loss (TD loss) for all of the Q functions in the ensemble.
+        td_target_duplicate = einops.repeat(td_target, "b -> e b", e=q_preds.shape[0])
+        # You compute the mean loss of the batch for each critic and then to compute the final loss you sum them up
+        critics_loss = (
+            F.mse_loss(
+                input=q_preds,
+                target=td_target_duplicate,
+                reduction="none",
+            ).mean(1)
+        ).sum()
+        return critics_loss
+
+    def compute_loss_temperature(self, observations) -> Tensor:
+        """Compute the temperature loss"""
+        # calculate temperature loss
+        with torch.no_grad():
+            _, log_probs, _ = self.actor(observations)
+        temperature_loss = (-self.log_alpha.exp() * (log_probs + self.config.target_entropy)).mean()
+        return temperature_loss
+
+    def compute_loss_actor(self, observations) -> Tensor:
+        temperature = self.log_alpha.exp().item()
+
+        actions_pi, log_probs, _ = self.actor(observations)
+
+        q_preds = self.critic_forward(observations, actions_pi, use_target=False)
+        min_q_preds = q_preds.min(dim=0)[0]
+
+        actor_loss = ((temperature * log_probs) - min_q_preds).mean()
+        return actor_loss
+
+
+class MLP(nn.Module):
+    def __init__(
+        self,
+        input_dim: int,
+        hidden_dims: list[int],
+        activations: Callable[[torch.Tensor], torch.Tensor] | str = nn.SiLU(),
+        activate_final: bool = False,
+        dropout_rate: Optional[float] = None,
+    ):
+        super().__init__()
+        self.activate_final = activate_final
+        layers = []
+
+        # First layer uses input_dim
+        layers.append(nn.Linear(input_dim, hidden_dims[0]))
+
+        # Add activation after first layer
+        if dropout_rate is not None and dropout_rate > 0:
+            layers.append(nn.Dropout(p=dropout_rate))
+        layers.append(nn.LayerNorm(hidden_dims[0]))
+        layers.append(activations if isinstance(activations, nn.Module) else getattr(nn, activations)())
+
+        # Rest of the layers
+        for i in range(1, len(hidden_dims)):
+            layers.append(nn.Linear(hidden_dims[i - 1], hidden_dims[i]))
+
+            if i + 1 < len(hidden_dims) or activate_final:
+                if dropout_rate is not None and dropout_rate > 0:
+                    layers.append(nn.Dropout(p=dropout_rate))
+                layers.append(nn.LayerNorm(hidden_dims[i]))
+                layers.append(
+                    activations if isinstance(activations, nn.Module) else getattr(nn, activations)()
+                )
+
+        self.net = nn.Sequential(*layers)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.net(x)
+    
+    
+class Critic(nn.Module):
+    def __init__(
+        self,
+        encoder: Optional[nn.Module],
+        network: nn.Module,
+        init_final: Optional[float] = None,
+        device: str = "cpu",
+    ):
+        super().__init__()
+        self.device = torch.device(device)
+        self.encoder = encoder
+        self.network = network
+        self.init_final = init_final
+        
+        # Find the last Linear layer's output dimension
+        for layer in reversed(network.net):
+            if isinstance(layer, nn.Linear):
+                out_features = layer.out_features
+                break
+        
+        # Output layer
+        if init_final is not None:
+            self.output_layer = nn.Linear(out_features, 1)
+            nn.init.uniform_(self.output_layer.weight, -init_final, init_final)
+            nn.init.uniform_(self.output_layer.bias, -init_final, init_final)
+        else:
+            self.output_layer = nn.Linear(out_features, 1)
+            orthogonal_init()(self.output_layer.weight)
+        
+        self.to(self.device)
+
+    def forward(
+        self,
+        observations: dict[str, torch.Tensor],
+        actions: torch.Tensor,
+    ) -> torch.Tensor:
+        # Move each tensor in observations to device
+        observations = {k: v.to(self.device) for k, v in observations.items()}
+        actions = actions.to(self.device)
+        
+        obs_enc = observations if self.encoder is None else self.encoder(observations)
+            
+        inputs = torch.cat([obs_enc, actions], dim=-1)
+        x = self.network(inputs)
+        value = self.output_layer(x)
+        return value.squeeze(-1)
+
+class Policy(nn.Module):
+    def __init__(
+        self,
+        encoder: Optional[nn.Module],
+        network: nn.Module,
+        action_dim: int,
+        log_std_min: float = -5,
+        log_std_max: float = 2,
+        fixed_std: Optional[torch.Tensor] = None,
+        init_final: Optional[float] = None,
+        use_tanh_squash: bool = False,
+        device: str = "cpu",
+        encoder_is_shared: bool = False,
+    ):
+        super().__init__()
+        self.device = torch.device(device)
+        self.encoder = encoder
+        self.network = network
+        self.action_dim = action_dim
+        self.log_std_min = log_std_min
+        self.log_std_max = log_std_max
+        self.fixed_std = fixed_std.to(self.device) if fixed_std is not None else None
+        self.use_tanh_squash = use_tanh_squash
+        self.parameters_to_optimize = []
+
+        self.parameters_to_optimize += list(self.network.parameters())
+
+        if self.encoder is not None and not encoder_is_shared:
+            self.parameters_to_optimize += list(self.encoder.parameters())
+        # Find the last Linear layer's output dimension
+        for layer in reversed(network.net):
+            if isinstance(layer, nn.Linear):
+                out_features = layer.out_features
+                break
+        # Mean layer
+        self.mean_layer = nn.Linear(out_features, action_dim)
+        if init_final is not None:
+            nn.init.uniform_(self.mean_layer.weight, -init_final, init_final)
+            nn.init.uniform_(self.mean_layer.bias, -init_final, init_final)
+        else:
+            orthogonal_init()(self.mean_layer.weight)
+
+        self.parameters_to_optimize += list(self.mean_layer.parameters())
+        # Standard deviation layer or parameter
+        if fixed_std is None:
+            self.std_layer = nn.Linear(out_features, action_dim)
+            if init_final is not None:
+                nn.init.uniform_(self.std_layer.weight, -init_final, init_final)
+                nn.init.uniform_(self.std_layer.bias, -init_final, init_final)
+            else:
+                orthogonal_init()(self.std_layer.weight)
+            self.parameters_to_optimize += list(self.std_layer.parameters())
+
+        self.to(self.device)
+
+    def forward(
+        self, 
+        observations: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+                
+        # Encode observations if encoder exists
+        obs_enc = observations if self.encoder is None else self.encoder(observations)
+
+        # Get network outputs
+        outputs = self.network(obs_enc)
+        means = self.mean_layer(outputs)
+        
+        # Compute standard deviations
+        if self.fixed_std is None:
+            log_std = self.std_layer(outputs)
+            assert not torch.isnan(log_std).any(), "[ERROR] log_std became NaN after std_layer!"
+
+            if self.use_tanh_squash:
+                log_std = torch.tanh(log_std)
+                log_std = self.log_std_min + 0.5 * (self.log_std_max - self.log_std_min) * (log_std + 1.0)
+            else:
+                log_std = torch.clamp(log_std, self.log_std_min, self.log_std_max)
+        else:
+            log_std = self.fixed_std.expand_as(means)
+
+        # uses tanh activation function to squash the action to be in the range of [-1, 1]
+        normal = torch.distributions.Normal(means, torch.exp(log_std))
+        x_t = normal.rsample()  # Reparameterization trick (mean + std * N(0,1))
+        log_probs = normal.log_prob(x_t)  # Base log probability before Tanh
+
+        if self.use_tanh_squash:
+            actions = torch.tanh(x_t)
+            log_probs -= torch.log((1 - actions.pow(2)) + 1e-6)  # Adjust log-probs for Tanh
+        else:
+            actions = x_t  # No Tanh; raw Gaussian sample
+
+        log_probs = log_probs.sum(-1)  # Sum over action dimensions
+        means = torch.tanh(means) if self.use_tanh_squash else means
+        return actions, log_probs, means
+
+    def get_features(self, observations: torch.Tensor) -> torch.Tensor:
+        """Get encoded features from observations"""
+        observations = observations.to(self.device)
+        if self.encoder is not None:
+            with torch.inference_mode():
+                return self.encoder(observations)
+        return observations
+
+
+class SACObservationEncoder(nn.Module):
+    """Encode image and/or state vector observations.
+    TODO(ke-wang): The original work allows for (1) stacking multiple history frames and (2) using pretrained resnet encoders.
+    """
+
+    def __init__(self, config: SACConfig):
+        """
+        Creates encoders for pixel and/or state modalities.
+        """
+        super().__init__()
+        self.config = config
+        if "observation.image" in config.input_shapes:
+            self.image_enc_layers = nn.Sequential(
+                nn.Conv2d(
+                    in_channels=config.input_shapes["observation.image"][0],
+                    out_channels=config.image_encoder_hidden_dim,
+                    kernel_size=7,
+                    stride=2,
+                ),
+                nn.ReLU(),
+                nn.Conv2d(
+                    in_channels=config.image_encoder_hidden_dim,
+                    out_channels=config.image_encoder_hidden_dim,
+                    kernel_size=5,
+                    stride=2,
+                ),
+                nn.ReLU(),
+                nn.Conv2d(
+                    in_channels=config.image_encoder_hidden_dim,
+                    out_channels=config.image_encoder_hidden_dim,
+                    kernel_size=3,
+                    stride=2,
+                ),
+                nn.ReLU(),
+                nn.Conv2d(
+                    in_channels=config.image_encoder_hidden_dim,
+                    out_channels=config.image_encoder_hidden_dim,
+                    kernel_size=3,
+                    stride=2,
+                ),
+                nn.ReLU(),
+            )
+            self.camera_number = config.camera_number
+            self.aggregation_size: int = 0
+
+            dummy_batch = torch.zeros(1, *config.input_shapes["observation.image"])
+            with torch.inference_mode():
+                out_shape = self.image_enc_layers(dummy_batch).shape[1:]
+            self.image_enc_layers.extend(
+                sequential=nn.Sequential(
+                    nn.Flatten(),
+                    nn.Linear(
+                        in_features=np.prod(out_shape) * self.camera_number, out_features=config.latent_dim
+                    ),
+                    nn.LayerNorm(normalized_shape=config.latent_dim),
+                    nn.Tanh(),
+                )
+            )
+
+            self.aggregation_size += config.latent_dim * self.camera_number
+        if "observation.state" in config.input_shapes:
+            self.state_enc_layers = nn.Sequential(
+                nn.Linear(
+                    in_features=config.input_shapes["observation.state"][0], out_features=config.latent_dim
+                ),
+                nn.LayerNorm(normalized_shape=config.latent_dim),
+                nn.Tanh(),
+            )
+            self.aggregation_size += config.latent_dim
+
+        if "observation.environment_state" in config.input_shapes:
+            self.env_state_enc_layers = nn.Sequential(
+                nn.Linear(
+                    in_features=config.input_shapes["observation.environment_state"][0],
+                    out_features=config.latent_dim,
+                ),
+                nn.LayerNorm(normalized_shape=config.latent_dim),
+                nn.Tanh(),
+            )
+
+            self.aggregation_size += config.latent_dim
+        self.aggregation_layer = nn.Linear(in_features=self.aggregation_size, out_features=config.latent_dim)
+
+    def forward(self, obs_dict: dict[str, Tensor]) -> Tensor:
+        """Encode the image and/or state vector.
+
+        Each modality is encoded into a feature vector of size (latent_dim,) and then a uniform mean is taken
+        over all features.
+        """
+        feat = []
+        # Concatenate all images along the channel dimension.
+        image_keys = [k for k in self.config.input_shapes if k.startswith("observation.image")]
+        for image_key in image_keys:
+            feat.append(flatten_forward_unflatten(self.image_enc_layers, obs_dict[image_key]))
+        if "observation.environment_state" in self.config.input_shapes:
+            feat.append(self.env_state_enc_layers(obs_dict["observation.environment_state"]))
+        if "observation.state" in self.config.input_shapes:
+            feat.append(self.state_enc_layers(obs_dict["observation.state"]))
+        # TODO(ke-wang): currently average over all features, concatenate all features maybe a better way
+        # return torch.stack(feat, dim=0).mean(0)
+        features = torch.cat(tensors=feat, dim=-1)
+        features = self.aggregation_layer(features)
+
+        return features
+
+    @property
+    def output_dim(self) -> int:
+        """Returns the dimension of the encoder output"""
+        return self.config.latent_dim
+
+
+def orthogonal_init():
+    return lambda x: torch.nn.init.orthogonal_(x, gain=1.0)
+
+
+def create_critic_ensemble(critics: list[nn.Module], num_critics: int, device: str = "cpu") -> nn.ModuleList:
+    """Creates an ensemble of critic networks"""
+    assert len(critics) == num_critics, f"Expected {num_critics} critics, got {len(critics)}"
+    return nn.ModuleList(critics).to(device)
+
+# borrowed from tdmpc
+def flatten_forward_unflatten(fn: Callable[[Tensor], Tensor], image_tensor: Tensor) -> Tensor:
+    """Helper to temporarily flatten extra dims at the start of the image tensor.
+
+    Args:
+        fn: Callable that the image tensor will be passed to. It should accept (B, C, H, W) and return
+            (B, *), where * is any number of dimensions.
+        image_tensor: An image tensor of shape (**, C, H, W), where ** is any number of dimensions and 
+        can be more than 1 dimensions, generally different from *.
+    Returns:
+        A return value from the callable reshaped to (**, *).
+    """
+    if image_tensor.ndim == 4:
+        return fn(image_tensor)
+    start_dims = image_tensor.shape[:-3]
+    inp = torch.flatten(image_tensor, end_dim=-4)
+    flat_out = fn(inp)
+    return torch.reshape(flat_out, (*start_dims, *flat_out.shape[1:]))

lerobot/common/robot_devices/cameras/intelrealsense.py

@@ -168,6 +168,7 @@ class IntelRealSenseCameraConfig:
     width: int | None = None
     height: int | None = None
     color_mode: str = "rgb"
+    channels: int | None = None
     use_depth: bool = False
     force_hardware_reset: bool = True
     rotation: int | None = None
@@ -179,6 +180,8 @@ class IntelRealSenseCameraConfig:
                 f"`color_mode` is expected to be 'rgb' or 'bgr', but {self.color_mode} is provided."
             )
 
+        self.channels = 3
+
         at_least_one_is_not_none = self.fps is not None or self.width is not None or self.height is not None
         at_least_one_is_none = self.fps is None or self.width is None or self.height is None
         if at_least_one_is_not_none and at_least_one_is_none:
@@ -254,6 +257,7 @@ class IntelRealSenseCamera:
         self.fps = config.fps
         self.width = config.width
         self.height = config.height
+        self.channels = config.channels
         self.color_mode = config.color_mode
         self.use_depth = config.use_depth
         self.force_hardware_reset = config.force_hardware_reset

lerobot/common/robot_devices/cameras/opencv.py

@@ -192,6 +192,7 @@ class OpenCVCameraConfig:
     width: int | None = None
     height: int | None = None
     color_mode: str = "rgb"
+    channels: int | None = None
     rotation: int | None = None
     mock: bool = False
 
@@ -201,6 +202,8 @@ class OpenCVCameraConfig:
                 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})")
 
@@ -268,6 +271,7 @@ class OpenCVCamera:
         self.fps = config.fps
         self.width = config.width
         self.height = config.height
+        self.channels = config.channels
         self.color_mode = config.color_mode
         self.mock = config.mock
 

lerobot/common/robot_devices/control_utils.py

@@ -11,11 +11,15 @@ from copy import copy
 from functools import cache
 
 import cv2
+import numpy as np
 import torch
 import tqdm
+from deepdiff import DeepDiff
 from termcolor import colored
 
-from lerobot.common.datasets.populate_dataset import add_frame, safe_stop_image_writer
+from lerobot.common.datasets.image_writer import safe_stop_image_writer
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.common.datasets.utils import get_features_from_robot
 from lerobot.common.policies.factory import make_policy
 from lerobot.common.robot_devices.robots.utils import Robot
 from lerobot.common.robot_devices.utils import busy_wait
@@ -117,14 +121,22 @@ def predict_action(observation, policy, device, use_amp):
     return action
 
 
-def init_keyboard_listener():
-    # Allow to exit early while recording an episode or resetting the environment,
-    # by tapping the right arrow key '->'. This might require a sudo permission
-    # to allow your terminal to monitor keyboard events.
+def init_keyboard_listener(assign_rewards=False):
+    """
+    Initializes a keyboard listener to enable early termination of an episode
+    or environment reset by pressing the right arrow key ('->'). This may require
+    sudo permissions to allow the terminal to monitor keyboard events.
+
+    Args:
+        assign_rewards (bool): If True, allows annotating the collected trajectory
+        with a binary reward at the end of the episode to indicate success.
+    """
     events = {}
     events["exit_early"] = False
     events["rerecord_episode"] = False
     events["stop_recording"] = False
+    if assign_rewards:
+        events["next.reward"] = 0
 
     if is_headless():
         logging.warning(
@@ -149,6 +161,13 @@ def init_keyboard_listener():
                 print("Escape key pressed. Stopping data recording...")
                 events["stop_recording"] = True
                 events["exit_early"] = True
+            elif assign_rewards and key == keyboard.Key.space:
+                events["next.reward"] = 1 if events["next.reward"] == 0 else 0
+                print(
+                    "Space key pressed. Assigning new reward to the subsequent frames. New reward:",
+                    events["next.reward"],
+                )
+
         except Exception as e:
             print(f"Error handling key press: {e}")
 
@@ -181,7 +200,7 @@ def init_policy(pretrained_policy_name_or_path, policy_overrides):
 def warmup_record(
     robot,
     events,
-    enable_teloperation,
+    enable_teleoperation,
     warmup_time_s,
     display_cameras,
     fps,
@@ -192,7 +211,7 @@ def warmup_record(
         display_cameras=display_cameras,
         events=events,
         fps=fps,
-        teleoperate=enable_teloperation,
+        teleoperate=enable_teleoperation,
     )
 
 
@@ -227,7 +246,7 @@ def control_loop(
     control_time_s=None,
     teleoperate=False,
     display_cameras=False,
-    dataset=None,
+    dataset: LeRobotDataset | None = None,
     events=None,
     policy=None,
     device=None,
@@ -247,7 +266,7 @@ def control_loop(
     if teleoperate and policy is not None:
         raise ValueError("When `teleoperate` is True, `policy` should be None.")
 
-    if dataset is not None and fps is not None and dataset["fps"] != fps:
+    if dataset is not None and fps is not None and dataset.fps != fps:
         raise ValueError(f"The dataset fps should be equal to requested fps ({dataset['fps']} != {fps}).")
 
     timestamp = 0
@@ -268,7 +287,10 @@ def control_loop(
                 action = {"action": action}
 
         if dataset is not None:
-            add_frame(dataset, observation, action)
+            frame = {**observation, **action}
+            if "next.reward" in events:
+                frame["next.reward"] = events["next.reward"]
+            dataset.add_frame(frame)
 
         if display_cameras and not is_headless():
             image_keys = [key for key in observation if "image" in key]
@@ -297,6 +319,8 @@ def reset_environment(robot, events, reset_time_s):
 
     timestamp = 0
     start_vencod_t = time.perf_counter()
+    if "next.reward" in events:
+        events["next.reward"] = 0
 
     # Wait if necessary
     with tqdm.tqdm(total=reset_time_s, desc="Waiting") as pbar:
@@ -309,6 +333,14 @@ def reset_environment(robot, events, reset_time_s):
                 break
 
 
+def reset_follower_position(robot: Robot, target_position):
+    current_position = robot.follower_arms["main"].read("Present_Position")
+    trajectory = torch.from_numpy(np.linspace(current_position, target_position, 30)) # NOTE: 30 is just an aribtrary number 
+    for pose in trajectory:
+        robot.send_action(pose)
+        busy_wait(0.015)
+
+
 def stop_recording(robot, listener, display_cameras):
     robot.disconnect()
 
@@ -324,7 +356,40 @@ def sanity_check_dataset_name(repo_id, policy):
     _, dataset_name = repo_id.split("/")
     # either repo_id doesnt start with "eval_" and there is no policy
     # or repo_id starts with "eval_" and there is a policy
-    if dataset_name.startswith("eval_") == (policy is None):
+
+    # Check if dataset_name starts with "eval_" but policy is missing
+    if dataset_name.startswith("eval_") and policy is None:
         raise ValueError(
-            f"Your dataset name begins by 'eval_' ({dataset_name}) but no policy is provided ({policy})."
+            f"Your dataset name begins with 'eval_' ({dataset_name}), but no policy is provided."
+        )
+
+    # Check if dataset_name does not start with "eval_" but policy is provided
+    if not dataset_name.startswith("eval_") and policy is not None:
+        raise ValueError(
+            f"Your dataset name does not begin with 'eval_' ({dataset_name}), but a policy is provided ({policy})."
+        )
+
+
+def sanity_check_dataset_robot_compatibility(
+    dataset: LeRobotDataset, robot: Robot, fps: int, use_videos: bool, extra_features: dict = None
+) -> None:
+    features_from_robot = get_features_from_robot(robot, use_videos)
+    if extra_features is not None:
+        features_from_robot.update(extra_features)
+
+    fields = [
+        ("robot_type", dataset.meta.robot_type, robot.robot_type),
+        ("fps", dataset.fps, fps),
+        ("features", dataset.features, features_from_robot),
+    ]
+
+    mismatches = []
+    for field, dataset_value, present_value in fields:
+        diff = DeepDiff(dataset_value, present_value, exclude_regex_paths=[r".*\['info'\]$"])
+        if diff:
+            mismatches.append(f"{field}: expected {present_value}, got {dataset_value}")
+
+    if mismatches:
+        raise ValueError(
+            "Dataset metadata compatibility check failed with mismatches:\n" + "\n".join(mismatches)
         )

lerobot/common/robot_devices/robots/manipulator.py

@@ -226,6 +226,42 @@ class ManipulatorRobot:
         self.is_connected = False
         self.logs = {}
 
+    def get_motor_names(self, arm: dict[str, MotorsBus]) -> list:
+        return [f"{arm}_{motor}" for arm, bus in arm.items() for motor in bus.motors]
+
+    @property
+    def camera_features(self) -> dict:
+        cam_ft = {}
+        for cam_key, cam in self.cameras.items():
+            key = f"observation.images.{cam_key}"
+            cam_ft[key] = {
+                "shape": (cam.height, cam.width, cam.channels),
+                "names": ["height", "width", "channels"],
+                "info": None,
+            }
+        return cam_ft
+
+    @property
+    def motor_features(self) -> dict:
+        action_names = self.get_motor_names(self.leader_arms)
+        state_names = self.get_motor_names(self.leader_arms)
+        return {
+            "action": {
+                "dtype": "float32",
+                "shape": (len(action_names),),
+                "names": action_names,
+            },
+            "observation.state": {
+                "dtype": "float32",
+                "shape": (len(state_names),),
+                "names": state_names,
+            },
+        }
+
+    @property
+    def features(self):
+        return {**self.motor_features, **self.camera_features}
+
     @property
     def has_camera(self):
         return len(self.cameras) > 0

lerobot/common/robot_devices/robots/utils.py

@@ -11,6 +11,7 @@ def get_arm_id(name, arm_type):
 class Robot(Protocol):
     # TODO(rcadene, aliberts): Add unit test checking the protocol is implemented in the corresponding classes
     robot_type: str
+    features: dict
 
     def connect(self): ...
     def run_calibration(self): ...

lerobot/configs/policy/act_aloha_real.yaml

@@ -114,7 +114,7 @@ policy:
   n_vae_encoder_layers: 4
 
   # Inference.
-  temporal_ensemble_momentum: null
+  temporal_ensemble_coeff: null
 
   # Training and loss computation.
   dropout: 0.1

lerobot/configs/policy/act_koch_real.yaml

@@ -95,7 +95,7 @@ policy:
   n_vae_encoder_layers: 4
 
   # Inference.
-  temporal_ensemble_momentum: null
+  temporal_ensemble_coeff: null
 
   # Training and loss computation.
   dropout: 0.1

lerobot/configs/policy/act_moss_real.yaml

@@ -95,7 +95,7 @@ policy:
   n_vae_encoder_layers: 4
 
   # Inference.
-  temporal_ensemble_momentum: null
+  temporal_ensemble_coeff: null
 
   # Training and loss computation.
   dropout: 0.1

lerobot/configs/policy/act_so100_real.yaml

@@ -95,7 +95,7 @@ policy:
   n_vae_encoder_layers: 4
 
   # Inference.
-  temporal_ensemble_momentum: null
+  temporal_ensemble_coeff: null
 
   # Training and loss computation.
   dropout: 0.1

lerobot/configs/policy/hilserl_classifier.yaml

@@ -0,0 +1,48 @@
+# @package _global_
+
+defaults:
+  - _self_
+
+seed: 13
+dataset_repo_id: aractingi/pick_place_lego_cube_1
+train_split_proportion: 0.8
+
+# Required by logger
+env:
+  name: "classifier"
+  task: "binary_classification"
+
+
+training:
+  num_epochs: 5
+  batch_size: 16
+  learning_rate: 1e-4
+  num_workers: 4
+  grad_clip_norm: 10
+  use_amp: true
+  log_freq: 1
+  eval_freq: 1  # How often to run validation (in epochs)
+  save_freq: 1  # How often to save checkpoints (in epochs)
+  save_checkpoint: true
+  image_keys: ["observation.images.top", "observation.images.wrist"]
+  label_key: "next.reward"
+
+eval:
+  batch_size: 16
+  num_samples_to_log: 30  # Number of validation samples to log in the table
+
+policy:
+  name: "hilserl/classifier/pick_place_lego_cube_1"
+  model_name: "facebook/convnext-base-224"
+  model_type: "cnn"
+  num_cameras: 2 # Has to be len(training.image_keys)
+
+wandb:
+  enable: false
+  project: "classifier-training"
+  job_name: "classifier_training_0"
+  disable_artifact: false
+
+device: "mps"
+resume: false
+output_dir: "outputs/classifier"

lerobot/configs/policy/sac_manyskill.yaml

@@ -0,0 +1,97 @@
+# @package _global_
+
+# Train with:
+#
+# python lerobot/scripts/train.py \
+#   +dataset=lerobot/pusht_keypoints
+#   env=pusht \
+#   env.gym.obs_type=environment_state_agent_pos \
+
+seed: 1
+dataset_repo_id: null 
+
+
+training:
+  # Offline training dataloader
+  num_workers: 4
+
+  # batch_size: 256
+  batch_size: 512
+  grad_clip_norm: 10.0
+  lr: 3e-4
+
+  eval_freq: 2500
+  log_freq: 500
+  save_freq: 50000
+
+  online_steps: 1000000
+  online_rollout_n_episodes: 10
+  online_rollout_batch_size: 10
+  online_steps_between_rollouts: 1000
+  online_sampling_ratio: 1.0
+  online_env_seed: 10000
+  online_buffer_capacity: 1000000
+  online_buffer_seed_size: 0
+  online_step_before_learning: 5000
+  do_online_rollout_async: false
+  policy_update_freq: 1
+
+  # delta_timestamps:
+  #   observation.environment_state: "[i / ${fps} for i in range(${policy.horizon} + 1)]"
+  #   observation.state: "[i / ${fps} for i in range(${policy.horizon} + 1)]"
+  #   action: "[i / ${fps} for i in range(${policy.horizon})]"
+  #   next.reward: "[i / ${fps} for i in range(${policy.horizon})]"
+
+policy:
+  name: sac
+
+  pretrained_model_path:
+
+  # Input / output structure.
+  n_action_repeats: 1
+  horizon: 1
+  n_action_steps: 1
+
+  shared_encoder: true
+  input_shapes:
+    # # TODO(rcadene, alexander-soare): add variables for height and width from the dataset/env?
+    observation.state: ["${env.state_dim}"]
+    observation.image: [3, 64, 64]
+  output_shapes:
+    action: ["${env.action_dim}"]
+
+  # Normalization / Unnormalization
+  input_normalization_modes: null
+  output_normalization_modes:
+    action: min_max
+  output_normalization_params:
+    action:
+      min: [-1.0, -1.0, -1.0, -1.0, -1.0, -1.0, -1.0]
+      max: [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0]
+
+  # Architecture / modeling.
+  # Neural networks.
+  image_encoder_hidden_dim: 32
+  # discount: 0.99
+  discount: 0.80
+  temperature_init: 1.0
+  num_critics: 2
+  num_subsample_critics: null
+  critic_lr: 3e-4
+  actor_lr: 3e-4
+  temperature_lr: 3e-4
+  # critic_target_update_weight: 0.005
+  critic_target_update_weight: 0.01
+  utd_ratio: 1
+
+
+  # # Loss coefficients.
+  # reward_coeff: 0.5
+  # expectile_weight: 0.9
+  # value_coeff: 0.1
+  # consistency_coeff: 20.0
+  # advantage_scaling: 3.0
+  # pi_coeff: 0.5
+  # temporal_decay_coeff: 0.5
+  # # Target model.
+  # target_model_momentum: 0.995

lerobot/configs/policy/sac_pusht_keypoints.yaml

@@ -0,0 +1,89 @@
+# @package _global_
+
+# Train with:
+#
+# python lerobot/scripts/train.py \
+#   env=pusht \
+#   +dataset=lerobot/pusht_keypoints
+
+seed: 1
+dataset_repo_id: lerobot/pusht_keypoints
+
+training:
+  offline_steps: 0
+
+  # Offline training dataloader
+  num_workers: 4
+
+  batch_size: 128
+  grad_clip_norm: 10.0
+  lr: 3e-4
+
+  eval_freq: 50000
+  log_freq: 500
+  save_freq: 50000
+
+  online_steps: 1000000
+  online_rollout_n_episodes: 10
+  online_rollout_batch_size: 10
+  online_steps_between_rollouts: 1000
+  online_sampling_ratio: 1.0
+  online_env_seed: 10000
+  online_buffer_capacity: 40000
+  online_buffer_seed_size: 0
+  do_online_rollout_async: false
+
+  delta_timestamps:
+    observation.environment_state: "[i / ${fps} for i in range(${policy.horizon} + 1)]"
+    observation.state: "[i / ${fps} for i in range(${policy.horizon} + 1)]"
+    action: "[i / ${fps} for i in range(${policy.horizon})]"
+    next.reward: "[i / ${fps} for i in range(${policy.horizon})]"
+
+policy:
+  name: sac
+
+  pretrained_model_path:
+
+  # Input / output structure.
+  n_action_repeats: 1
+  horizon: 5
+  n_action_steps: 5
+
+  input_shapes:
+    # TODO(rcadene, alexander-soare): add variables for height and width from the dataset/env?
+    observation.environment_state: [16]
+    observation.state: ["${env.state_dim}"]
+  output_shapes:
+    action: ["${env.action_dim}"]
+
+  # Normalization / Unnormalization
+  input_normalization_modes:
+    observation.environment_state: min_max
+    observation.state: min_max
+  output_normalization_modes:
+    action: min_max
+
+  # Architecture / modeling.
+  # Neural networks.
+  # image_encoder_hidden_dim: 32
+  discount: 0.99
+  temperature_init: 1.0
+  num_critics: 2
+  num_subsample_critics: None
+  critic_lr: 3e-4
+  actor_lr: 3e-4
+  temperature_lr: 3e-4
+  critic_target_update_weight: 0.005
+  utd_ratio: 2
+
+
+  # # Loss coefficients.
+  # reward_coeff: 0.5
+  # expectile_weight: 0.9
+  # value_coeff: 0.1
+  # consistency_coeff: 20.0
+  # advantage_scaling: 3.0
+  # pi_coeff: 0.5
+  # temporal_decay_coeff: 0.5
+  # # Target model.
+  # target_model_momentum: 0.995

lerobot/configs/robot/koch.yaml

@@ -10,7 +10,7 @@ max_relative_target: null
 leader_arms:
   main:
     _target_: lerobot.common.robot_devices.motors.dynamixel.DynamixelMotorsBus
-    port: /dev/tty.usbmodem575E0031751
+    port: /dev/tty.usbmodem58760430441
     motors:
       # name: (index, model)
       shoulder_pan: [1, "xl330-m077"]
@@ -23,7 +23,7 @@ leader_arms:
 follower_arms:
   main:
     _target_: lerobot.common.robot_devices.motors.dynamixel.DynamixelMotorsBus
-    port: /dev/tty.usbmodem575E0032081
+    port: /dev/tty.usbmodem585A0083391
     motors:
       # name: (index, model)
       shoulder_pan: [1, "xl430-w250"]

lerobot/configs/robot/so100.yaml

@@ -18,7 +18,7 @@ max_relative_target: null
 leader_arms:
   main:
     _target_: lerobot.common.robot_devices.motors.feetech.FeetechMotorsBus
-    port: /dev/tty.usbmodem585A0077581
+    port: /dev/tty.usbmodem58760433331
     motors:
       # name: (index, model)
       shoulder_pan: [1, "sts3215"]

lerobot/scripts/control_robot.py

@@ -29,7 +29,6 @@ python lerobot/scripts/control_robot.py teleoperate \
 ```bash
 python lerobot/scripts/control_robot.py record \
     --fps 30 \
-    --root tmp/data \
     --repo-id $USER/koch_test \
     --num-episodes 1 \
     --run-compute-stats 0
@@ -38,7 +37,6 @@ python lerobot/scripts/control_robot.py record \
 - Visualize dataset:
 ```bash
 python lerobot/scripts/visualize_dataset.py \
-    --root tmp/data \
     --repo-id $USER/koch_test \
     --episode-index 0
 ```
@@ -47,7 +45,6 @@ python lerobot/scripts/visualize_dataset.py \
 ```bash
 python lerobot/scripts/control_robot.py replay \
     --fps 30 \
-    --root tmp/data \
     --repo-id $USER/koch_test \
     --episode 0
 ```
@@ -57,7 +54,6 @@ python lerobot/scripts/control_robot.py replay \
 ```bash
 python lerobot/scripts/control_robot.py record \
     --fps 30 \
-    --root data \
     --repo-id $USER/koch_pick_place_lego \
     --num-episodes 50 \
     --warmup-time-s 2 \
@@ -72,12 +68,12 @@ python lerobot/scripts/control_robot.py record \
 - Tap escape key 'esc' to stop the data recording.
 This might require a sudo permission to allow your terminal to monitor keyboard events.
 
-**NOTE**: You can resume/continue data recording by running the same data recording command twice.
-To avoid resuming by deleting the dataset, use `--force-override 1`.
+**NOTE**: You can resume/continue data recording by running the same data recording command and adding `--resume 1`.
+If the dataset you want to extend is not on the hub, you also need to add `--local-files-only 1`.
 
 - Train on this dataset with the ACT policy:
 ```bash
-DATA_DIR=data python lerobot/scripts/train.py \
+python lerobot/scripts/train.py \
     policy=act_koch_real \
     env=koch_real \
     dataset_repo_id=$USER/koch_pick_place_lego \
@@ -88,7 +84,6 @@ DATA_DIR=data python lerobot/scripts/train.py \
 ```bash
 python lerobot/scripts/control_robot.py record \
     --fps 30 \
-    --root data \
     --repo-id $USER/eval_act_koch_real \
     --num-episodes 10 \
     --warmup-time-s 2 \
@@ -106,12 +101,6 @@ from typing import List
 
 # from safetensors.torch import load_file, save_file
 from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.common.datasets.populate_dataset import (
-    create_lerobot_dataset,
-    delete_current_episode,
-    init_dataset,
-    save_current_episode,
-)
 from lerobot.common.robot_devices.control_utils import (
     control_loop,
     has_method,
@@ -120,7 +109,9 @@ from lerobot.common.robot_devices.control_utils import (
     log_control_info,
     record_episode,
     reset_environment,
+    reset_follower_position,
     sanity_check_dataset_name,
+    sanity_check_dataset_robot_compatibility,
     stop_recording,
     warmup_record,
 )
@@ -196,31 +187,44 @@ def teleoperate(
 @safe_disconnect
 def record(
     robot: Robot,
-    root: str,
+    root: Path,
     repo_id: str,
+    single_task: str,
     pretrained_policy_name_or_path: str | None = None,
     policy_overrides: List[str] | None = None,
+    assign_rewards: bool = False,
     fps: int | None = None,
-    warmup_time_s=2,
-    episode_time_s=10,
-    reset_time_s=5,
-    num_episodes=50,
-    video=True,
-    run_compute_stats=True,
-    push_to_hub=True,
-    tags=None,
-    num_image_writer_processes=0,
-    num_image_writer_threads_per_camera=4,
-    force_override=False,
-    display_cameras=True,
-    play_sounds=True,
-):
+    warmup_time_s: int | float = 2,
+    episode_time_s: int | float = 10,
+    reset_time_s: int | float = 5,
+    num_episodes: int = 50,
+    video: bool = True,
+    run_compute_stats: bool = True,
+    push_to_hub: bool = True,
+    tags: list[str] | None = None,
+    num_image_writer_processes: int = 0,
+    num_image_writer_threads_per_camera: int = 4,
+    display_cameras: bool = True,
+    play_sounds: bool = True,
+    reset_follower: bool = False, 
+    resume: bool = False,
+    # TODO(rcadene, aliberts): remove local_files_only when refactor with dataset as argument
+    local_files_only: bool = False,
+) -> LeRobotDataset:
     # TODO(rcadene): Add option to record logs
     listener = None
     events = None
     policy = None
     device = None
     use_amp = None
+    extra_features = (
+        {"next.reward": {"dtype": "int64", "shape": (1,), "names": None}} if assign_rewards else None
+    )
+
+    if single_task:
+        task = single_task
+    else:
+        raise NotImplementedError("Only single-task recording is supported for now")
 
     # Load pretrained policy
     if pretrained_policy_name_or_path is not None:
@@ -234,24 +238,38 @@ def record(
                 f"There is a mismatch between the provided fps ({fps}) and the one from policy config ({policy_fps})."
             )
 
-    # Create empty dataset or load existing saved episodes
-    sanity_check_dataset_name(repo_id, policy)
-    dataset = init_dataset(
-        repo_id,
-        root,
-        force_override,
-        fps,
-        video,
-        write_images=robot.has_camera,
-        num_image_writer_processes=num_image_writer_processes,
-        num_image_writer_threads=num_image_writer_threads_per_camera * robot.num_cameras,
-    )
+    if resume:
+        dataset = LeRobotDataset(
+            repo_id,
+            root=root,
+            local_files_only=local_files_only,
+        )
+        dataset.start_image_writer(
+            num_processes=num_image_writer_processes,
+            num_threads=num_image_writer_threads_per_camera * len(robot.cameras),
+        )
+        sanity_check_dataset_robot_compatibility(dataset, robot, fps, video, extra_features)
+    else:
+        # Create empty dataset or load existing saved episodes
+        sanity_check_dataset_name(repo_id, policy)
+        dataset = LeRobotDataset.create(
+            repo_id,
+            fps,
+            root=root,
+            robot=robot,
+            use_videos=video,
+            image_writer_processes=num_image_writer_processes,
+            image_writer_threads=num_image_writer_threads_per_camera * len(robot.cameras),
+            features=extra_features,
+        )
 
     if not robot.is_connected:
         robot.connect()
+    listener, events = init_keyboard_listener(assign_rewards=assign_rewards)
 
-    listener, events = init_keyboard_listener()
-
+    if reset_follower:
+        initial_position = robot.follower_arms["main"].read("Present_Position")
+        
     # Execute a few seconds without recording to:
     # 1. teleoperate the robot to move it in starting position if no policy provided,
     # 2. give times to the robot devices to connect and start synchronizing,
@@ -263,12 +281,17 @@ def record(
     if has_method(robot, "teleop_safety_stop"):
         robot.teleop_safety_stop()
 
+    recorded_episodes = 0
     while True:
-        if dataset["num_episodes"] >= num_episodes:
+        if recorded_episodes >= num_episodes:
             break
 
-        episode_index = dataset["num_episodes"]
-        log_say(f"Recording episode {episode_index}", play_sounds)
+        # TODO(aliberts): add task prompt for multitask here. Might need to temporarily disable event if
+        # input() messes with them.
+        # if multi_task:
+        #     task = input("Enter your task description: ")
+
+        log_say(f"Recording episode {dataset.num_episodes}", play_sounds)
         record_episode(
             dataset=dataset,
             robot=robot,
@@ -286,20 +309,22 @@ def record(
         # TODO(rcadene): add an option to enable teleoperation during reset
         # Skip reset for the last episode to be recorded
         if not events["stop_recording"] and (
-            (episode_index < num_episodes - 1) or events["rerecord_episode"]
+            (recorded_episodes < num_episodes - 1) or events["rerecord_episode"]
         ):
             log_say("Reset the environment", play_sounds)
+            if reset_follower:
+                reset_follower_position(robot, initial_position)
             reset_environment(robot, events, reset_time_s)
 
         if events["rerecord_episode"]:
             log_say("Re-record episode", play_sounds)
             events["rerecord_episode"] = False
             events["exit_early"] = False
-            delete_current_episode(dataset)
+            dataset.clear_episode_buffer()
             continue
 
-        # Increment by one dataset["current_episode_index"]
-        save_current_episode(dataset)
+        dataset.save_episode(task)
+        recorded_episodes += 1
 
         if events["stop_recording"]:
             break
@@ -307,35 +332,42 @@ def record(
     log_say("Stop recording", play_sounds, blocking=True)
     stop_recording(robot, listener, display_cameras)
 
-    lerobot_dataset = create_lerobot_dataset(dataset, run_compute_stats, push_to_hub, tags, play_sounds)
+    if run_compute_stats:
+        logging.info("Computing dataset statistics")
+
+    dataset.consolidate(run_compute_stats)
+
+    if push_to_hub:
+        dataset.push_to_hub(tags=tags)
 
     log_say("Exiting", play_sounds)
-    return lerobot_dataset
+    return dataset
 
 
 @safe_disconnect
 def replay(
-    robot: Robot, episode: int, fps: int | None = None, root="data", repo_id="lerobot/debug", play_sounds=True
+    robot: Robot,
+    root: Path,
+    repo_id: str,
+    episode: int,
+    fps: int | None = None,
+    play_sounds: bool = True,
+    local_files_only: bool = False,
 ):
     # TODO(rcadene, aliberts): refactor with control_loop, once `dataset` is an instance of LeRobotDataset
     # TODO(rcadene): Add option to record logs
-    local_dir = Path(root) / repo_id
-    if not local_dir.exists():
-        raise ValueError(local_dir)
 
-    dataset = LeRobotDataset(repo_id, root=root)
-    items = dataset.hf_dataset.select_columns("action")
-    from_idx = dataset.episode_data_index["from"][episode].item()
-    to_idx = dataset.episode_data_index["to"][episode].item()
+    dataset = LeRobotDataset(repo_id, root=root, episodes=[episode], local_files_only=local_files_only)
+    actions = dataset.hf_dataset.select_columns("action")
 
     if not robot.is_connected:
         robot.connect()
 
     log_say("Replaying episode", play_sounds, blocking=True)
-    for idx in range(from_idx, to_idx):
+    for idx in range(dataset.num_frames):
         start_episode_t = time.perf_counter()
 
-        action = items[idx]["action"]
+        action = actions[idx]["action"]
         robot.send_action(action)
 
         dt_s = time.perf_counter() - start_episode_t
@@ -384,14 +416,26 @@ if __name__ == "__main__":
     )
 
     parser_record = subparsers.add_parser("record", parents=[base_parser])
+    task_args = parser_record.add_mutually_exclusive_group(required=True)
     parser_record.add_argument(
         "--fps", type=none_or_int, default=None, help="Frames per second (set to None to disable)"
     )
+    task_args.add_argument(
+        "--single-task",
+        type=str,
+        help="A short but accurate description of the task performed during the recording.",
+    )
+    # TODO(aliberts): add multi-task support
+    # task_args.add_argument(
+    #     "--multi-task",
+    #     type=int,
+    #     help="You will need to enter the task performed at the start of each episode.",
+    # )
     parser_record.add_argument(
         "--root",
         type=Path,
-        default="data",
-        help="Root directory where the dataset will be stored locally at '{root}/{repo_id}' (e.g. 'data/hf_username/dataset_name').",
+        default=None,
+        help="Root directory where the dataset will be stored (e.g. 'dataset/path').",
     )
     parser_record.add_argument(
         "--repo-id",
@@ -399,6 +443,12 @@ if __name__ == "__main__":
         default="lerobot/test",
         help="Dataset identifier. By convention it should match '{hf_username}/{dataset_name}' (e.g. `lerobot/test`).",
     )
+    parser_record.add_argument(
+        "--local-files-only",
+        type=int,
+        default=0,
+        help="Use local files only. By default, this script will try to fetch the dataset from the hub if it exists.",
+    )
     parser_record.add_argument(
         "--warmup-time-s",
         type=int,
@@ -430,12 +480,12 @@ if __name__ == "__main__":
         default=1,
         help="Upload dataset to Hugging Face hub.",
     )
-    parser_record.add_argument(
-        "--tags",
-        type=str,
-        nargs="*",
-        help="Add tags to your dataset on the hub.",
-    )
+    # parser_record.add_argument(
+    #     "--tags",
+    #     type=str,
+    #     nargs="*",
+    #     help="Add tags to your dataset on the hub.",
+    # )
     parser_record.add_argument(
         "--num-image-writer-processes",
         type=int,
@@ -458,10 +508,10 @@ if __name__ == "__main__":
         ),
     )
     parser_record.add_argument(
-        "--force-override",
+        "--resume",
         type=int,
         default=0,
-        help="By default, data recording is resumed. When set to 1, delete the local directory and start data recording from scratch.",
+        help="Resume recording on an existing dataset.",
     )
     parser_record.add_argument(
         "-p",
@@ -478,6 +528,18 @@ if __name__ == "__main__":
         nargs="*",
         help="Any key=value arguments to override config values (use dots for.nested=overrides)",
     )
+    parser_record.add_argument(
+        "--assign-rewards",
+        type=int,
+        default=0,
+        help="Enables the assignation of rewards to frames (by default no assignation). When enabled, assign a 0 reward to frames until the space bar is pressed which assign a 1 reward. Press the space bar a second time to assign a 0 reward. The reward assigned is reset to 0 when the episode ends.",
+    )
+    parser_record.add_argument(
+        "--reset-follower",
+        type=int,
+        default=0,
+        help="Resets the follower to the initial position during while reseting the evironment, this is to avoid having the follower start at an awkward position in the next episode",
+    )
 
     parser_replay = subparsers.add_parser("replay", parents=[base_parser])
     parser_replay.add_argument(
@@ -486,8 +548,8 @@ if __name__ == "__main__":
     parser_replay.add_argument(
         "--root",
         type=Path,
-        default="data",
-        help="Root directory where the dataset will be stored locally at '{root}/{repo_id}' (e.g. 'data/hf_username/dataset_name').",
+        default=None,
+        help="Root directory where the dataset will be stored (e.g. 'dataset/path').",
     )
     parser_replay.add_argument(
         "--repo-id",
@@ -495,6 +557,12 @@ if __name__ == "__main__":
         default="lerobot/test",
         help="Dataset identifier. By convention it should match '{hf_username}/{dataset_name}' (e.g. `lerobot/test`).",
     )
+    parser_replay.add_argument(
+        "--local-files-only",
+        type=int,
+        default=0,
+        help="Use local files only. By default, this script will try to fetch the dataset from the hub if it exists.",
+    )
     parser_replay.add_argument("--episode", type=int, default=0, help="Index of the episode to replay.")
 
     args = parser.parse_args()

lerobot/scripts/control_sim_robot.py

@@ -0,0 +1,567 @@
+"""
+Utilities to control a robot in simulation.
+
+Useful to record a dataset, replay a recorded episode and record an evaluation dataset.
+
+Examples of usage:
+
+
+- Unlimited teleoperation at a limited frequency of 30 Hz, to simulate data recording frequency.
+  You can modify this value depending on how fast your simulation can run:
+```bash
+python lerobot/scripts/control_robot.py teleoperate \
+    --fps 30 \
+    --robot-path lerobot/configs/robot/your_robot_config.yaml \
+    --sim-config lerobot/configs/env/your_sim_config.yaml
+```
+
+- Record one episode in order to test replay:
+```bash
+python lerobot/scripts/control_sim_robot.py record \
+    --robot-path lerobot/configs/robot/your_robot_config.yaml \
+    --sim-config lerobot/configs/env/your_sim_config.yaml \
+    --fps 30 \
+    --repo-id $USER/robot_sim_test \
+    --num-episodes 1 \
+    --run-compute-stats 0
+```
+
+Enable the --push-to-hub 1 to push the recorded dataset to the huggingface hub.
+
+- Visualize dataset:
+```bash
+python lerobot/scripts/visualize_dataset.py \
+    --repo-id $USER/robot_sim_test \
+    --episode-index 0
+```
+
+- Replay a sequence of test episodes:
+```bash
+python lerobot/scripts/control_sim_robot.py replay \
+    --robot-path lerobot/configs/robot/your_robot_config.yaml \
+    --sim-config lerobot/configs/env/your_sim_config.yaml \
+    --fps 30 \
+    --repo-id $USER/robot_sim_test \
+    --episode 0
+```
+Note: The seed is saved, therefore, during replay we can load the same environment state as the one during collection.
+
+- Record a full dataset in order to train a policy,
+30 seconds of recording for each episode, and 10 seconds to reset the environment in between episodes:
+```bash
+python lerobot/scripts/control_sim_robot.py record \
+    --robot-path lerobot/configs/robot/your_robot_config.yaml \
+    --sim-config lerobot/configs/env/your_sim_config.yaml \
+    --fps 30 \
+    --repo-id $USER/robot_sim_test \
+    --num-episodes 50 \
+    --episode-time-s 30 \
+```
+
+**NOTE**: You can use your keyboard to control data recording flow.
+- Tap right arrow key '->' to early exit while recording an episode and go to reseting the environment.
+- Tap right arrow key '->' to early exit while reseting the environment and got to recording the next episode.
+- Tap left arrow key '<-' to early exit and re-record the current episode.
+- Tap escape key 'esc' to stop the data recording.
+This might require a sudo permission to allow your terminal to monitor keyboard events.
+
+**NOTE**: You can resume/continue data recording by running the same data recording command twice.
+"""
+
+import argparse
+import importlib
+import logging
+import time
+from pathlib import Path
+
+import cv2
+import gymnasium as gym
+import numpy as np
+import torch
+
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.common.robot_devices.control_utils import (
+    init_keyboard_listener,
+    init_policy,
+    is_headless,
+    log_control_info,
+    predict_action,
+    sanity_check_dataset_name,
+    sanity_check_dataset_robot_compatibility,
+    stop_recording,
+)
+from lerobot.common.robot_devices.robots.factory import make_robot
+from lerobot.common.robot_devices.robots.utils import Robot
+from lerobot.common.robot_devices.utils import busy_wait
+from lerobot.common.utils.utils import init_hydra_config, init_logging, log_say
+
+DEFAULT_FEATURES = {
+    "next.reward": {
+        "dtype": "float32",
+        "shape": (1,),
+        "names": None,
+    },
+    "next.success": {
+        "dtype": "bool",
+        "shape": (1,),
+        "names": None,
+    },
+    "seed": {
+        "dtype": "int64",
+        "shape": (1,),
+        "names": None,
+    },
+    "timestamp": {
+        "dtype": "float32",
+        "shape": (1,),
+        "names": None,
+    },
+}
+
+
+########################################################################################
+# Utilities
+########################################################################################
+def none_or_int(value):
+    if value == "None":
+        return None
+    return int(value)
+
+
+def init_sim_calibration(robot, cfg):
+    # Constants necessary for transforming the joint pos of the real robot to the sim
+    # depending on the robot discription used in that sim.
+    start_pos = np.array(robot.leader_arms.main.calibration["start_pos"])
+    axis_directions = np.array(cfg.get("axis_directions", [1]))
+    offsets = np.array(cfg.get("offsets", [0])) * np.pi
+
+    return {"start_pos": start_pos, "axis_directions": axis_directions, "offsets": offsets}
+
+
+def real_positions_to_sim(real_positions, axis_directions, start_pos, offsets):
+    """Counts - starting position -> radians -> align axes -> offset"""
+    return axis_directions * (real_positions - start_pos) * 2.0 * np.pi / 4096 + offsets
+
+
+########################################################################################
+# Control modes
+########################################################################################
+
+
+def teleoperate(env, robot: Robot, process_action_fn, teleop_time_s=None):
+    env = env()
+    env.reset()
+    start_teleop_t = time.perf_counter()
+    while True:
+        leader_pos = robot.leader_arms.main.read("Present_Position")
+        action = process_action_fn(leader_pos)
+        env.step(np.expand_dims(action, 0))
+        if teleop_time_s is not None and time.perf_counter() - start_teleop_t > teleop_time_s:
+            print("Teleoperation processes finished.")
+            break
+
+
+def record(
+    env,
+    robot: Robot,
+    process_action_from_leader,
+    root: Path,
+    repo_id: str,
+    task: str,
+    fps: int | None = None,
+    tags: list[str] | None = None,
+    pretrained_policy_name_or_path: str = None,
+    policy_overrides: bool | None = None,
+    episode_time_s: int = 30,
+    num_episodes: int = 50,
+    video: bool = True,
+    push_to_hub: bool = True,
+    num_image_writer_processes: int = 0,
+    num_image_writer_threads_per_camera: int = 4,
+    display_cameras: bool = False,
+    play_sounds: bool = True,
+    resume: bool = False,
+    local_files_only: bool = False,
+    run_compute_stats: bool = True,
+    assign_rewards: bool = False,
+) -> LeRobotDataset:
+    # Load pretrained policy
+
+    extra_features = (
+        {"next.reward": {"dtype": "int64", "shape": (1,), "names": None}} if assign_rewards else None
+    )
+
+    policy = None
+    if pretrained_policy_name_or_path is not None:
+        policy, policy_fps, device, use_amp = init_policy(pretrained_policy_name_or_path, policy_overrides)
+
+        if fps is None:
+            fps = policy_fps
+            logging.warning(f"No fps provided, so using the fps from policy config ({policy_fps}).")
+
+    if policy is None and process_action_from_leader is None:
+        raise ValueError("Either policy or process_action_fn has to be set to enable control in sim.")
+
+    # initialize listener before sim env
+    listener, events = init_keyboard_listener(assign_rewards=assign_rewards)
+
+    # create sim env
+    env = env()
+
+    # Create empty dataset or load existing saved episodes
+    num_cameras = sum([1 if "image" in key else 0 for key in env.observation_space])
+
+    # get image keys
+    image_keys = [key for key in env.observation_space if "image" in key]
+    state_keys_dict = env_cfg.state_keys
+
+    if resume:
+        dataset = LeRobotDataset(
+            repo_id,
+            root=root,
+            local_files_only=local_files_only,
+        )
+        dataset.start_image_writer(
+            num_processes=num_image_writer_processes,
+            num_threads=num_image_writer_threads_per_camera * num_cameras,
+        )
+        sanity_check_dataset_robot_compatibility(dataset, robot, fps, video)
+    else:
+        features = DEFAULT_FEATURES
+        # add image keys to features
+        for key in image_keys:
+            shape = env.observation_space[key].shape
+            if not key.startswith("observation.image."):
+                key = "observation.image." + key
+            features[key] = {"dtype": "video", "names": ["channel", "height", "width"], "shape": shape}
+
+        for key, obs_key in state_keys_dict.items():
+            features[key] = {
+                "dtype": "float32",
+                "names": None,
+                "shape": env.observation_space[obs_key].shape,
+            }
+
+        features["action"] = {"dtype": "float32", "shape": env.action_space.shape, "names": None}
+        features = {**features, **extra_features}
+
+        # Create empty dataset or load existing saved episodes
+        sanity_check_dataset_name(repo_id, policy)
+        dataset = LeRobotDataset.create(
+            repo_id,
+            fps,
+            root=root,
+            features=features,
+            use_videos=video,
+            image_writer_processes=num_image_writer_processes,
+            image_writer_threads=num_image_writer_threads_per_camera * num_cameras,
+        )
+
+    recorded_episodes = 0
+    while True:
+        log_say(f"Recording episode {dataset.num_episodes}", play_sounds)
+
+        if events is None:
+            events = {"exit_early": False}
+
+        if episode_time_s is None:
+            episode_time_s = float("inf")
+
+        timestamp = 0
+        start_episode_t = time.perf_counter()
+
+        seed = np.random.randint(0, 1e5)
+        observation, info = env.reset(seed=seed)
+
+        while timestamp < episode_time_s:
+            start_loop_t = time.perf_counter()
+
+            if policy is not None:
+                action = predict_action(observation, policy, device, use_amp)
+            else:
+                leader_pos = robot.leader_arms.main.read("Present_Position")
+                action = process_action_from_leader(leader_pos)
+
+            observation, reward, terminated, _, info = env.step(action)
+
+            success = info.get("is_success", False)
+            env_timestamp = info.get("timestamp", dataset.episode_buffer["size"] / fps)
+
+            frame = {
+                "action": torch.from_numpy(action),
+                "next.reward": reward,
+                "next.success": success,
+                "seed": seed,
+                "timestamp": env_timestamp,
+            }
+
+            # Overwrite environment reward with manually assigned reward
+            if assign_rewards:
+                frame["next.reward"] = events["next.reward"]
+
+                # Should success always be false to match what we do in control_utils?
+                frame["next.success"] = False
+
+            for key in image_keys:
+                if not key.startswith("observation.image"):
+                    frame["observation.image." + key] = observation[key]
+                else:
+                    frame[key] = observation[key]
+
+            for key, obs_key in state_keys_dict.items():
+                frame[key] = torch.from_numpy(observation[obs_key])
+
+            dataset.add_frame(frame)
+
+            if display_cameras and not is_headless():
+                for key in image_keys:
+                    cv2.imshow(key, cv2.cvtColor(observation[key], cv2.COLOR_RGB2BGR))
+                cv2.waitKey(1)
+
+            if fps is not None:
+                dt_s = time.perf_counter() - start_loop_t
+                busy_wait(1 / fps - dt_s)
+
+            dt_s = time.perf_counter() - start_loop_t
+            log_control_info(robot, dt_s, fps=fps)
+
+            timestamp = time.perf_counter() - start_episode_t
+            if events["exit_early"] or terminated:
+                events["exit_early"] = False
+                break
+
+        if events["rerecord_episode"]:
+            log_say("Re-record episode", play_sounds)
+            events["rerecord_episode"] = False
+            events["exit_early"] = False
+            dataset.clear_episode_buffer()
+            continue
+
+        dataset.save_episode(task=task)
+        recorded_episodes += 1
+
+        if events["stop_recording"] or recorded_episodes >= num_episodes:
+            break
+        else:
+            logging.info("Waiting for a few seconds before starting next episode recording...")
+            busy_wait(3)
+
+    log_say("Stop recording", play_sounds, blocking=True)
+    stop_recording(robot, listener, display_cameras)
+
+    if run_compute_stats:
+        logging.info("Computing dataset statistics")
+    dataset.consolidate(run_compute_stats)
+
+    if push_to_hub:
+        dataset.push_to_hub(tags=tags)
+
+    log_say("Exiting", play_sounds)
+    return dataset
+
+
+def replay(
+    env, root: Path, repo_id: str, episode: int, fps: int | None = None, local_files_only: bool = True
+):
+    env = env()
+
+    local_dir = Path(root) / repo_id
+    if not local_dir.exists():
+        raise ValueError(local_dir)
+
+    dataset = LeRobotDataset(repo_id, root=root, local_files_only=local_files_only)
+    items = dataset.hf_dataset.select_columns("action")
+    seeds = dataset.hf_dataset.select_columns("seed")["seed"]
+
+    from_idx = dataset.episode_data_index["from"][episode].item()
+    to_idx = dataset.episode_data_index["to"][episode].item()
+    env.reset(seed=seeds[from_idx].item())
+    logging.info("Replaying episode")
+    log_say("Replaying episode", play_sounds=True)
+    for idx in range(from_idx, to_idx):
+        start_episode_t = time.perf_counter()
+        action = items[idx]["action"]
+        env.step(action.unsqueeze(0).numpy())
+        dt_s = time.perf_counter() - start_episode_t
+        busy_wait(1 / fps - dt_s)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    subparsers = parser.add_subparsers(dest="mode", required=True)
+
+    # Set common options for all the subparsers
+    base_parser = argparse.ArgumentParser(add_help=False)
+    base_parser.add_argument(
+        "--robot-path",
+        type=str,
+        default="lerobot/configs/robot/koch.yaml",
+        help="Path to robot yaml file used to instantiate the robot using `make_robot` factory function.",
+    )
+
+    base_parser.add_argument(
+        "--sim-config",
+        help="Path to a yaml config you want to use for initializing a sim environment based on gym ",
+    )
+
+    parser_record = subparsers.add_parser("teleoperate", parents=[base_parser])
+
+    parser_record = subparsers.add_parser("record", parents=[base_parser])
+    parser_record.add_argument(
+        "--fps", type=none_or_int, default=None, help="Frames per second (set to None to disable)"
+    )
+    parser_record.add_argument(
+        "--root",
+        type=Path,
+        default=None,
+        help="Root directory where the dataset will be stored locally at '{root}/{repo_id}' (e.g. 'data/hf_username/dataset_name').",
+    )
+    parser_record.add_argument(
+        "--repo-id",
+        type=str,
+        default="lerobot/test",
+        help="Dataset identifier. By convention it should match '{hf_username}/{dataset_name}' (e.g. `lerobot/test`).",
+    )
+    parser_record.add_argument(
+        "--episode-time-s",
+        type=int,
+        default=60,
+        help="Number of seconds for data recording for each episode.",
+    )
+    parser_record.add_argument(
+        "--task",
+        type=str,
+        required=True,
+        help="A description of the task preformed during recording that can be used as a language instruction.",
+    )
+    parser_record.add_argument("--num-episodes", type=int, default=50, help="Number of episodes to record.")
+    parser_record.add_argument(
+        "--run-compute-stats",
+        type=int,
+        default=1,
+        help="By default, run the computation of the data statistics at the end of data collection. Compute intensive and not required to just replay an episode.",
+    )
+    parser_record.add_argument(
+        "--push-to-hub",
+        type=int,
+        default=1,
+        help="Upload dataset to Hugging Face hub.",
+    )
+    parser_record.add_argument(
+        "--tags",
+        type=str,
+        nargs="*",
+        help="Add tags to your dataset on the hub.",
+    )
+    parser_record.add_argument(
+        "--num-image-writer-processes",
+        type=int,
+        default=0,
+        help=(
+            "Number of subprocesses handling the saving of frames as PNGs. Set to 0 to use threads only; "
+            "set to ≥1 to use subprocesses, each using threads to write images. The best number of processes "
+            "and threads depends on your system. We recommend 4 threads per camera with 0 processes. "
+            "If fps is unstable, adjust the thread count. If still unstable, try using 1 or more subprocesses."
+        ),
+    )
+    parser_record.add_argument(
+        "--num-image-writer-threads-per-camera",
+        type=int,
+        default=4,
+        help=(
+            "Number of threads writing the frames as png images on disk, per camera. "
+            "Too much threads might cause unstable teleoperation fps due to main thread being blocked. "
+            "Not enough threads might cause low camera fps."
+        ),
+    )
+    parser_record.add_argument(
+        "--display-cameras",
+        type=int,
+        default=0,
+        help="Visualize image observations with opencv.",
+    )
+    parser_record.add_argument(
+        "--resume",
+        type=int,
+        default=0,
+        help="Resume recording on an existing dataset.",
+    )
+    parser_record.add_argument(
+        "--assign-rewards",
+        type=int,
+        default=0,
+        help="Enables the assignation of rewards to frames (by default no assignation). When enabled, assign a 0 reward to frames until the space bar is pressed which assign a 1 reward. Press the space bar a second time to assign a 0 reward. The reward assigned is reset to 0 when the episode ends.",
+    )
+
+    parser_replay = subparsers.add_parser("replay", parents=[base_parser])
+    parser_replay.add_argument(
+        "--fps", type=none_or_int, default=None, help="Frames per second (set to None to disable)"
+    )
+    parser_replay.add_argument(
+        "--root",
+        type=Path,
+        default=None,
+        help="Root directory where the dataset will be stored locally (e.g. 'data/hf_username/dataset_name'). By default, stored in cache folder.",
+    )
+    parser_replay.add_argument(
+        "--repo-id",
+        type=str,
+        default="lerobot/test",
+        help="Dataset identifier. By convention it should match '{hf_username}/{dataset_name}' (e.g. `lerobot/test`).",
+    )
+    parser_replay.add_argument("--episode", type=int, default=0, help="Index of the episodes to replay.")
+
+    args = parser.parse_args()
+
+    init_logging()
+
+    control_mode = args.mode
+    robot_path = args.robot_path
+    env_config_path = args.sim_config
+    kwargs = vars(args)
+    del kwargs["mode"]
+    del kwargs["robot_path"]
+    del kwargs["sim_config"]
+
+    # make gym env
+    env_cfg = init_hydra_config(env_config_path)
+    importlib.import_module(f"gym_{env_cfg.env.name}")
+
+    def env_constructor():
+        return gym.make(env_cfg.env.handle, disable_env_checker=True, **env_cfg.env.gym)
+
+    robot = None
+    process_leader_actions_fn = None
+
+    if control_mode in ["teleoperate", "record"]:
+        # make robot
+        robot_overrides = ["~cameras", "~follower_arms"]
+        robot_cfg = init_hydra_config(robot_path, robot_overrides)
+        robot = make_robot(robot_cfg)
+        robot.connect()
+
+        calib_kwgs = init_sim_calibration(robot, env_cfg.calibration)
+
+        def process_leader_actions_fn(action):
+            return real_positions_to_sim(action, **calib_kwgs)
+
+        robot.leader_arms.main.calibration = None
+
+    if control_mode == "teleoperate":
+        teleoperate(env_constructor, robot, process_leader_actions_fn)
+
+    elif control_mode == "record":
+        record(env_constructor, robot, process_leader_actions_fn, **kwargs)
+
+    elif control_mode == "replay":
+        replay(env_constructor, **kwargs)
+
+    else:
+        raise ValueError(
+            f"Invalid control mode: '{control_mode}', only valid modes are teleoperate, record and replay."
+        )
+
+    if robot and robot.is_connected:
+        # Disconnect manually to avoid a "Core dump" during process
+        # termination due to camera threads not properly exiting.
+        robot.disconnect()

lerobot/scripts/eval.py

@@ -484,7 +484,7 @@ def main(
         policy = make_policy(hydra_cfg=hydra_cfg, pretrained_policy_name_or_path=str(pretrained_policy_path))
     else:
         # Note: We need the dataset stats to pass to the policy's normalization modules.
-        policy = make_policy(hydra_cfg=hydra_cfg, dataset_stats=make_dataset(hydra_cfg).stats)
+        policy = make_policy(hydra_cfg=hydra_cfg, dataset_stats=make_dataset(hydra_cfg).meta.stats)
 
     assert isinstance(policy, nn.Module)
     policy.eval()

lerobot/scripts/eval_on_robot.py

@@ -0,0 +1,433 @@
+#!/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.
+"""Evaluate a policy by running rollouts on the real robot and computing metrics.
+
+This script supports performing human interventions during rollouts. 
+Human interventions allow the user to take control of the robot from the policy 
+and correct its behavior. It is specifically designed for reinforcement learning 
+experiments and HIL-SERL (human-in-the-loop reinforcement learning) methods.
+
+### How to Use
+
+To rollout a policy on the robot:
+```
+python lerobot/scripts/eval_on_robot.py \
+    --robot-path lerobot/configs/robot/so100.yaml \
+    --pretrained-policy-path-or-name path/to/pretrained_model \
+    --policy-config path/to/policy/config.yaml \
+    --display-cameras 1
+```
+
+If you trained a reward classifier on your task, you can also evaluate it using this script.
+You can annotate the collection with a pre-trained reward classifier by running:
+
+```
+python lerobot/scripts/eval_on_robot.py \
+    --robot-path lerobot/configs/robot/so100.yaml \
+    --pretrained-policy-path-or-name path/to/pretrained_model \
+    --policy-config path/to/policy/config.yaml \
+    --reward-classifier-pretrained-path outputs/classifier/checkpoints/best/pretrained_model \
+    --reward-classifier-config-file lerobot/configs/policy/hilserl_classifier.yaml \
+    --display-cameras 1
+```
+"""
+
+import argparse
+import logging
+import time
+
+import cv2
+import numpy as np
+import torch
+from tqdm import trange
+
+from lerobot.common.policies.policy_protocol import Policy
+from lerobot.common.policies.utils import get_device_from_parameters
+from lerobot.common.robot_devices.control_utils import busy_wait, is_headless, reset_follower_position, predict_action
+from lerobot.common.robot_devices.robots.factory import Robot, make_robot
+from lerobot.common.utils.utils import (
+    init_hydra_config,
+    init_logging,
+    log_say,
+)
+
+
+def get_classifier(pretrained_path, config_path):
+    if pretrained_path is None or config_path is None:
+        return
+
+    from lerobot.common.policies.factory import _policy_cfg_from_hydra_cfg
+    from lerobot.common.policies.hilserl.classifier.configuration_classifier import ClassifierConfig
+    from lerobot.common.policies.hilserl.classifier.modeling_classifier import Classifier
+
+    cfg = init_hydra_config(config_path)
+
+    classifier_config = _policy_cfg_from_hydra_cfg(ClassifierConfig, cfg)
+    classifier_config.num_cameras = len(cfg.training.image_keys)  # TODO automate these paths
+    model = Classifier(classifier_config)
+    model.load_state_dict(Classifier.from_pretrained(pretrained_path).state_dict())
+    return model
+
+
+def rollout(
+    robot: Robot,
+    policy: Policy,
+    reward_classifier,
+    fps: int,
+    control_time_s: float = 20,
+    use_amp: bool = True,
+    display_cameras: bool = False,
+    device: str = "cpu"
+) -> dict:
+    """Run a batched policy rollout on the real robot.
+
+    This function executes a rollout using the provided policy and robot interface, 
+    simulating batched interactions for a fixed control duration. 
+
+    The returned dictionary contains rollout statistics, which can be used for analysis and debugging.
+
+    Args:
+        "robot": The robot interface for interacting with the real robot hardware.
+        "policy": The policy to execute. Must be a PyTorch `nn.Module` object.
+        "reward_classifier": A module to classify rewards during the rollout.
+        "fps": The control frequency at which the policy is executed.
+        "control_time_s": The total control duration of the rollout in seconds.
+        "use_amp": Whether to use automatic mixed precision (AMP) for policy evaluation.
+        "display_cameras": If True, displays camera streams during the rollout.
+        "device": The device to use for computations (e.g., "cpu", "cuda" or "mps").
+
+    Returns:
+        Dictionary of the statisitcs collected during rollouts.
+    """
+    # define keyboard listener
+    listener, events = init_keyboard_listener()
+
+    # Reset the policy. TODO (michel-aractingi) add real policy evaluation once the code is ready.
+    if policy is not None:
+        policy.reset()  
+
+    # NOTE: sorting to make sure the key sequence is the same during training and testing.
+    observation = robot.capture_observation()
+    image_keys = [key for key in observation if "image" in key]
+    image_keys.sort() # CG{T}
+
+    
+    all_actions = []
+    all_rewards = []
+    
+    indices_from_policy = []
+
+    start_episode_t = time.perf_counter()
+    init_pos = robot.follower_arms["main"].read("Present_Position")
+    timestamp = 0.0
+    while timestamp < control_time_s:
+        start_loop_t = time.perf_counter()
+
+        # Apply the next action.
+        while events["pause_policy"] and not events["human_intervention_step"]:
+            busy_wait(0.5)
+
+        if events["human_intervention_step"]:
+            # take over the robot's actions
+            observation, action = robot.teleop_step(record_data=True)
+            action = action["action"]  # teleop step returns torch tensors but in a dict
+        else:
+            # explore with policy
+            with torch.inference_mode():
+                # TODO (michel-aractingi) in placy temporarly for testing purposes
+                if policy is None:
+                    action = robot.follower_arms["main"].read("Present_Position")
+                    action = torch.from_numpy(action)
+                    indices_from_policy.append(False)
+                else:
+                    action = predict_action(observation, policy, device, use_amp)
+                    indices_from_policy.append(True)
+                
+                robot.send_action(action)
+                observation = robot.capture_observation()
+                
+
+        images = []
+        for key in image_keys:
+            if display_cameras:
+                cv2.imshow(key, cv2.cvtColor(observation[key].numpy(), cv2.COLOR_RGB2BGR))
+                cv2.waitKey(1)
+            images.append(observation[key].to(device))
+
+        reward = reward_classifier.predict_reward(images) if reward_classifier is not None else 0.0
+
+        # TODO send data through the server as soon as you have it       
+
+        all_rewards.append(reward)
+        all_actions.append(action)
+
+        dt_s = time.perf_counter() - start_loop_t
+        busy_wait(1 / fps - dt_s)
+        timestamp = time.perf_counter() - start_episode_t
+        if events["exit_early"]:
+            events["exit_early"] = False
+            events["human_intervention_step"] = False
+            events["pause_policy"] = False
+            break
+
+    reset_follower_position(robot, target_position=init_pos)
+
+    dones = torch.tensor([False] * len(all_actions))
+    dones[-1] = True
+    # Stack the sequence along the first dimension so that we have (batch, sequence, *) tensors.
+    ret = {
+        "action": torch.stack(all_actions, dim=1),
+        "next.reward": torch.stack(all_rewards, dim=1),
+        "done": dones,
+    }
+
+    listener.stop()
+
+    return ret
+
+
+def eval_policy(
+    robot: Robot,
+    policy: torch.nn.Module,
+    fps: float,
+    n_episodes: int,
+    control_time_s: int = 20,
+    use_amp: bool = True,
+    display_cameras: bool = False,
+    reward_classifier_pretrained_path: str | None = None,
+    reward_classifier_config_file: str | None = None,
+    device: str | None = None,
+) -> dict:
+    """
+    Evaluate a policy on a real robot by running multiple episodes and collecting metrics.
+
+    This function executes rollouts of the specified policy on the robot, computes metrics 
+    for the rollouts, and optionally evaluates a reward classifier if provided.
+
+    Args:
+        "robot": The robot interface used to interact with the real robot hardware.
+        "policy": The policy to be evaluated. Must be a PyTorch neural network module.
+        "fps": Frames per second (control frequency) for running the policy.
+        "n_episodes": The number of episodes to evaluate the policy.
+        "control_time_s": The max duration for each episode in seconds.
+        "use_amp": Whether to use automatic mixed precision (AMP) for policy evaluation.
+        "display_cameras": Whether to display camera streams during rollouts.
+        "reward_classifier_pretrained_path": Path to the pretrained reward classifier. 
+            If provided, the reward classifier will be evaluated during rollouts.
+        "reward_classifier_config_file": Path to the configuration file for the reward classifier. 
+            Required if `reward_classifier_pretrained_path` is provided.
+        "device": The device for computations (e.g., "cpu", "cuda" or "mps"). 
+
+    Returns:
+        "dict": A dictionary containing the following rollout metrics and data:
+            - "metrics": Evaluation metrics such as cumulative rewards, success rates, etc.
+            - "rollout_data": Detailed data from the rollouts, including observations, actions, rewards, and done flags.
+    """
+    # TODO (michel-aractingi) comment this out for testing with a fixed policy
+    # assert isinstance(policy, Policy)
+    # policy.eval()
+
+    sum_rewards = []
+    max_rewards = []
+    rollouts = []
+
+    start_eval = time.perf_counter()
+    progbar = trange(n_episodes, desc="Evaluating policy on real robot")
+    reward_classifier = get_classifier(reward_classifier_pretrained_path, reward_classifier_config_file).to(device)
+
+    device = get_device_from_parameters(policy) if device is None else device
+
+    for _ in progbar:
+        rollout_data = rollout(
+            robot, policy, reward_classifier, fps, control_time_s, use_amp, display_cameras, device
+        )
+
+        rollouts.append(rollout_data)
+        sum_rewards.append(sum(rollout_data["next.reward"]))
+        max_rewards.append(max(rollout_data["next.reward"]))
+
+    info = {
+        "per_episode": [
+            {
+                "episode_ix": i,
+                "sum_reward": sum_reward,
+                "max_reward": max_reward,
+            }
+            for i, (sum_reward, max_reward) in enumerate(
+                zip(
+                    sum_rewards[:n_episodes],
+                    max_rewards[:n_episodes],
+                    strict=False,
+                )
+            )
+        ],
+        "aggregated": {
+            "avg_sum_reward":    float(np.nanmean(torch.cat(sum_rewards[:n_episodes]))),
+            "avg_max_reward": float(np.nanmean(torch.cat(max_rewards[:n_episodes]))),
+            "eval_s": time.time() - start_eval,
+            "eval_ep_s": (time.time() - start_eval) / n_episodes,
+        },
+    }
+
+    if robot.is_connected:
+        robot.disconnect()
+
+    return info
+
+
+def init_keyboard_listener():
+    """
+    Initialize a keyboard listener for controlling the recording and human intervention process.
+
+    Keyboard controls: (Note that this might require sudo permissions to monitor keyboard events)
+        - Right Arrow Key ('->'): Stops the current recording and exits early, useful for ending an episode 
+          and moving the next episode recording. 
+        - Left Arrow Key ('<-'): Re-records the current episode, allowing the user to start over.
+        - Space Bar: Controls the human intervention process in three steps:
+            1. First press pauses the policy and prompts the user to position the leader similar to the follower.
+            2. Second press initiates human interventions, allowing teleop control of the robot.
+            3. Third press resumes the policy rollout.
+    """
+    events = {}
+    events["exit_early"] = False
+    events["rerecord_episode"] = False
+    events["pause_policy"] = False
+    events["human_intervention_step"] = False
+
+    if is_headless():
+        logging.warning(
+            "Headless environment detected. On-screen cameras display and keyboard inputs will not be available."
+        )
+        listener = None
+        return listener, events
+
+    # Only import pynput if not in a headless environment
+    from pynput import keyboard
+
+    def on_press(key):
+        try:
+            if key == keyboard.Key.right:
+                print("Right arrow key pressed. Exiting loop...")
+                events["exit_early"] = True
+            elif key == keyboard.Key.left:
+                print("Left arrow key pressed. Exiting loop and rerecord the last episode...")
+                events["rerecord_episode"] = True
+                events["exit_early"] = True
+            elif key == keyboard.Key.space:
+                # check if first space press then pause the policy for the user to get ready
+                # if second space press then the user is ready to start intervention
+                if not events["pause_policy"]:
+                    print(
+                        "Space key pressed. Human intervention required.\n"
+                        "Place the leader in similar pose to the follower and press space again."
+                    )
+                    events["pause_policy"] = True
+                    log_say("Human intervention stage. Get ready to take over.", play_sounds=True)
+                elif events["pause_policy"] and not events["human_intervention_step"]:
+                    events["human_intervention_step"] = True
+                    print("Space key pressed. Human intervention starting.")
+                    log_say("Starting human intervention.", play_sounds=True)
+                elif events["human_intervention_step"]:
+                    events["human_intervention_step"] = False
+                    events["pause_policy"] = False
+                    print("Space key pressed. Human intervention ending, policy resumes control.")
+                    log_say("Policy resuming.", play_sounds=True)
+
+        except Exception as e:
+            print(f"Error handling key press: {e}")
+
+    listener = keyboard.Listener(on_press=on_press)
+    listener.start()
+
+    return listener, events
+
+
+if __name__ == "__main__":
+    init_logging()
+
+    parser = argparse.ArgumentParser(
+        description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter
+    )
+    group = parser.add_mutually_exclusive_group(required=True)
+    group.add_argument(
+        "--robot-path",
+        type=str,
+        default="lerobot/configs/robot/koch.yaml",
+        help="Path to robot yaml file used to instantiate the robot using `make_robot` factory function.",
+    )
+    group.add_argument(
+        "--robot-overrides",
+        type=str,
+        nargs="*",
+        help="Any key=value arguments to override config values (use dots for.nested=overrides)",
+    )
+    group.add_argument(
+        "-p",
+        "--pretrained-policy-name-or-path",
+        help=(
+            "Either the repo ID of a model hosted on the Hub or a path to a directory containing weights "
+            "saved using `Policy.save_pretrained`. If not provided, the policy is initialized from scratch "
+            "(useful for debugging). This argument is mutually exclusive with `--config`."
+        ),
+    )
+    group.add_argument(
+        "--config",
+        help=(
+            "Path to a yaml config you want to use for initializing a policy from scratch (useful for "
+            "debugging). This argument is mutually exclusive with `--pretrained-policy-name-or-path` (`-p`)."
+        ),
+    )
+    parser.add_argument("--revision", help="Optionally provide the Hugging Face Hub revision ID.")
+    parser.add_argument(
+        "--out-dir",
+        help=(
+            "Where to save the evaluation outputs. If not provided, outputs are saved in "
+            "outputs/eval/{timestamp}_{env_name}_{policy_name}"
+        ),
+    )
+    parser.add_argument(
+        "--display-cameras", help=("Whether to display the camera feed while the rollout is happening")
+    )
+    parser.add_argument(
+        "--reward-classifier-pretrained-path",
+        type=str,
+        default=None,
+        help="Path to the pretrained classifier weights.",
+    )
+    parser.add_argument(
+        "--reward-classifier-config-file",
+        type=str,
+        default=None,
+        help="Path to a yaml config file that is necessary to build the reward classifier model.",
+    )
+
+    args = parser.parse_args()
+
+    robot_cfg = init_hydra_config(args.robot_path, args.robot_overrides)
+    robot = make_robot(robot_cfg)
+    if not robot.is_connected:
+        robot.connect()
+
+    eval_policy(
+        robot,
+        None,
+        fps=40,
+        n_episodes=2,
+        control_time_s=100,
+        display_cameras=args.display_cameras,
+        reward_classifier_config_file=args.reward_classifier_config_file,
+        reward_classifier_pretrained_path=args.reward_classifier_pretrained_path,
+    )

lerobot/scripts/find_motors_bus_port.py

@@ -1,30 +1,36 @@
+import os
 import time
 from pathlib import Path
 
+from serial.tools import list_ports  # Part of pyserial library
+
 
 def find_available_ports():
-    ports = []
-    for path in Path("/dev").glob("tty*"):
-        ports.append(str(path))
+    if os.name == "nt":  # Windows
+        # List COM ports using pyserial
+        ports = [port.device for port in list_ports.comports()]
+    else:  # Linux/macOS
+        # List /dev/tty* ports for Unix-based systems
+        ports = [str(path) for path in Path("/dev").glob("tty*")]
     return ports
 
 
 def find_port():
     print("Finding all available ports for the MotorsBus.")
     ports_before = find_available_ports()
-    print(ports_before)
+    print("Ports before disconnecting:", ports_before)
 
-    print("Remove the usb cable from your MotorsBus and press Enter when done.")
-    input()
+    print("Remove the USB cable from your MotorsBus and press Enter when done.")
+    input()  # Wait for user to disconnect the device
 
-    time.sleep(0.5)
+    time.sleep(0.5)  # Allow some time for port to be released
     ports_after = find_available_ports()
     ports_diff = list(set(ports_before) - set(ports_after))
 
     if len(ports_diff) == 1:
         port = ports_diff[0]
         print(f"The port of this MotorsBus is '{port}'")
-        print("Reconnect the usb cable.")
+        print("Reconnect the USB cable.")
     elif len(ports_diff) == 0:
         raise OSError(f"Could not detect the port. No difference was found ({ports_diff}).")
     else:
@@ -32,5 +38,5 @@ def find_port():
 
 
 if __name__ == "__main__":
-    # Helper to find the usb port associated to all your MotorsBus.
+    # Helper to find the USB port associated with your MotorsBus.
     find_port()

lerobot/scripts/push_dataset_to_hub.py

@@ -66,7 +66,7 @@ def get_from_raw_to_lerobot_format_fn(raw_format: str):
         from lerobot.common.datasets.push_dataset_to_hub.umi_zarr_format import from_raw_to_lerobot_format
     elif raw_format == "aloha_hdf5":
         from lerobot.common.datasets.push_dataset_to_hub.aloha_hdf5_format import from_raw_to_lerobot_format
-    elif "openx_rlds" in raw_format:
+    elif raw_format in ["rlds", "openx"]:
         from lerobot.common.datasets.push_dataset_to_hub.openx_rlds_format import from_raw_to_lerobot_format
     elif raw_format == "dora_parquet":
         from lerobot.common.datasets.push_dataset_to_hub.dora_parquet_format import from_raw_to_lerobot_format
@@ -117,10 +117,14 @@ def push_meta_data_to_hub(repo_id: str, meta_data_dir: str | Path, revision: str
 
 
 def push_dataset_card_to_hub(
-    repo_id: str, revision: str | None, tags: list | None = None, text: str | None = None
+    repo_id: str,
+    revision: str | None,
+    tags: list | None = None,
+    license: str = "apache-2.0",
+    **card_kwargs,
 ):
     """Creates and pushes a LeRobotDataset Card with appropriate tags to easily find it on the hub."""
-    card = create_lerobot_dataset_card(tags=tags, text=text)
+    card = create_lerobot_dataset_card(tags=tags, license=license, **card_kwargs)
     card.push_to_hub(repo_id=repo_id, repo_type="dataset", revision=revision)
 
 
@@ -200,24 +204,14 @@ def push_dataset_to_hub(
     # convert dataset from original raw format to LeRobot format
     from_raw_to_lerobot_format = get_from_raw_to_lerobot_format_fn(raw_format)
 
-    fmt_kwgs = {
-        "raw_dir": raw_dir,
-        "videos_dir": videos_dir,
-        "fps": fps,
-        "video": video,
-        "episodes": episodes,
-        "encoding": encoding,
-    }
-
-    if "openx_rlds." in raw_format:
-        # Support for official OXE dataset name inside `raw_format`.
-        # For instance, `raw_format="oxe_rlds"` uses the default formating (TODO what does that mean?),
-        # and `raw_format="oxe_rlds.bridge_orig"` uses the brdige_orig formating
-        _, openx_dataset_name = raw_format.split(".")
-        print(f"Converting dataset [{openx_dataset_name}] from 'openx_rlds' to LeRobot format.")
-        fmt_kwgs["openx_dataset_name"] = openx_dataset_name
-
-    hf_dataset, episode_data_index, info = from_raw_to_lerobot_format(**fmt_kwgs)
+    hf_dataset, episode_data_index, info = from_raw_to_lerobot_format(
+        raw_dir,
+        videos_dir,
+        fps,
+        video,
+        episodes,
+        encoding,
+    )
 
     lerobot_dataset = LeRobotDataset.from_preloaded(
         repo_id=repo_id,
@@ -260,7 +254,7 @@ def push_dataset_to_hub(
         episode_index = 0
         tests_videos_dir = tests_data_dir / repo_id / "videos"
         tests_videos_dir.mkdir(parents=True, exist_ok=True)
-        for key in lerobot_dataset.video_frame_keys:
+        for key in lerobot_dataset.camera_keys:
             fname = f"{key}_episode_{episode_index:06d}.mp4"
             shutil.copy(videos_dir / fname, tests_videos_dir / fname)
 
@@ -286,7 +280,7 @@ def main():
         "--raw-format",
         type=str,
         required=True,
-        help="Dataset type (e.g. `pusht_zarr`, `umi_zarr`, `aloha_hdf5`, `xarm_pkl`, `dora_parquet`, `openx_rlds`).",
+        help="Dataset type (e.g. `pusht_zarr`, `umi_zarr`, `aloha_hdf5`, `xarm_pkl`, `dora_parquet`, `rlds`, `openx`).",
     )
     parser.add_argument(
         "--repo-id",

lerobot/scripts/train.py

@@ -93,6 +93,17 @@ def make_optimizer_and_scheduler(cfg, policy):
     elif policy.name == "tdmpc":
         optimizer = torch.optim.Adam(policy.parameters(), cfg.training.lr)
         lr_scheduler = None
+
+    elif policy.name == "sac":
+        optimizer = torch.optim.Adam(
+            [
+                {"params": policy.actor.parameters(), "lr": policy.config.actor_lr},
+                {"params": policy.critic_ensemble.parameters(), "lr": policy.config.critic_lr},
+                {"params": policy.temperature.parameters(), "lr": policy.config.temperature_lr},
+            ]
+        )
+        lr_scheduler = None
+
     elif cfg.policy.name == "vqbet":
         from lerobot.common.policies.vqbet.modeling_vqbet import VQBeTOptimizer, VQBeTScheduler
 
@@ -171,9 +182,9 @@ def log_train_info(logger: Logger, info, step, cfg, dataset, is_online):
     # A sample is an (observation,action) pair, where observation and action
     # can be on multiple timestamps. In a batch, we have `batch_size`` number of samples.
     num_samples = (step + 1) * cfg.training.batch_size
-    avg_samples_per_ep = dataset.num_samples / dataset.num_episodes
+    avg_samples_per_ep = dataset.num_frames / dataset.num_episodes
     num_episodes = num_samples / avg_samples_per_ep
-    num_epochs = num_samples / dataset.num_samples
+    num_epochs = num_samples / dataset.num_frames
     log_items = [
         f"step:{format_big_number(step)}",
         # number of samples seen during training
@@ -208,9 +219,9 @@ def log_eval_info(logger, info, step, cfg, dataset, is_online):
     # A sample is an (observation,action) pair, where observation and action
     # can be on multiple timestamps. In a batch, we have `batch_size`` number of samples.
     num_samples = (step + 1) * cfg.training.batch_size
-    avg_samples_per_ep = dataset.num_samples / dataset.num_episodes
+    avg_samples_per_ep = dataset.num_frames / dataset.num_episodes
     num_episodes = num_samples / avg_samples_per_ep
-    num_epochs = num_samples / dataset.num_samples
+    num_epochs = num_samples / dataset.num_frames
     log_items = [
         f"step:{format_big_number(step)}",
         # number of samples seen during training
@@ -311,6 +322,11 @@ def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = No
 
     logging.info("make_dataset")
     offline_dataset = make_dataset(cfg)
+    # TODO (michel-aractingi): temporary fix to avoid datasets with task_index key that doesn't exist in online environment
+    # i.e., pusht
+    if "task_index" in offline_dataset.hf_dataset[0]:
+        offline_dataset.hf_dataset = offline_dataset.hf_dataset.remove_columns(["task_index"])
+
     if isinstance(offline_dataset, MultiLeRobotDataset):
         logging.info(
             "Multiple datasets were provided. Applied the following index mapping to the provided datasets: "
@@ -328,7 +344,7 @@ def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = No
     logging.info("make_policy")
     policy = make_policy(
         hydra_cfg=cfg,
-        dataset_stats=offline_dataset.stats if not cfg.resume else None,
+        dataset_stats=offline_dataset.meta.stats if not cfg.resume else None,
         pretrained_policy_name_or_path=str(logger.last_pretrained_model_dir) if cfg.resume else None,
     )
     assert isinstance(policy, nn.Module)
@@ -349,7 +365,7 @@ def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = No
     logging.info(f"{cfg.env.task=}")
     logging.info(f"{cfg.training.offline_steps=} ({format_big_number(cfg.training.offline_steps)})")
     logging.info(f"{cfg.training.online_steps=}")
-    logging.info(f"{offline_dataset.num_samples=} ({format_big_number(offline_dataset.num_samples)})")
+    logging.info(f"{offline_dataset.num_frames=} ({format_big_number(offline_dataset.num_frames)})")
     logging.info(f"{offline_dataset.num_episodes=}")
     logging.info(f"{num_learnable_params=} ({format_big_number(num_learnable_params)})")
     logging.info(f"{num_total_params=} ({format_big_number(num_total_params)})")
@@ -573,7 +589,7 @@ def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = No
                     online_drop_n_last_frames=cfg.training.get("drop_n_last_frames", 0) + 1,
                     online_sampling_ratio=cfg.training.online_sampling_ratio,
                 )
-                sampler.num_samples = len(concat_dataset)
+                sampler.num_frames = len(concat_dataset)
 
                 update_online_buffer_s = time.perf_counter() - start_update_buffer_time
 

lerobot/scripts/train_hilserl_classifier.py

@@ -0,0 +1,320 @@
+#!/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 logging
+import time
+from contextlib import nullcontext
+from pathlib import Path
+from pprint import pformat
+
+import hydra
+import torch
+import torch.nn as nn
+import wandb
+from deepdiff import DeepDiff
+from omegaconf import DictConfig, OmegaConf
+from termcolor import colored
+from torch import optim
+from torch.cuda.amp import GradScaler
+from torch.utils.data import DataLoader, WeightedRandomSampler, random_split
+from tqdm import tqdm
+
+from lerobot.common.datasets.factory import resolve_delta_timestamps
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.common.logger import Logger
+from lerobot.common.policies.factory import _policy_cfg_from_hydra_cfg
+from lerobot.common.policies.hilserl.classifier.configuration_classifier import ClassifierConfig
+from lerobot.common.policies.hilserl.classifier.modeling_classifier import Classifier
+from lerobot.common.utils.utils import (
+    format_big_number,
+    get_safe_torch_device,
+    init_hydra_config,
+    set_global_seed,
+)
+
+
+def get_model(cfg, logger):  # noqa I001
+    classifier_config = _policy_cfg_from_hydra_cfg(ClassifierConfig, cfg)
+    model = Classifier(classifier_config)
+    if cfg.resume:
+        model.load_state_dict(Classifier.from_pretrained(str(logger.last_pretrained_model_dir)).state_dict())
+    return model
+
+
+def create_balanced_sampler(dataset, cfg):
+    # Creates a weighted sampler to handle class imbalance
+
+    labels = torch.tensor([item[cfg.training.label_key] for item in dataset])
+    _, counts = torch.unique(labels, return_counts=True)
+    class_weights = 1.0 / counts.float()
+    sample_weights = class_weights[labels]
+
+    return WeightedRandomSampler(weights=sample_weights, num_samples=len(sample_weights), replacement=True)
+
+
+def support_amp(device: torch.device, cfg: DictConfig) -> bool:
+    # Check if the device supports AMP
+    # Here is an example of the issue that says that MPS doesn't support AMP properply
+    return cfg.training.use_amp and device.type in ("cuda", "cpu")
+
+
+def train_epoch(model, train_loader, criterion, optimizer, grad_scaler, device, logger, step, cfg):
+    # Single epoch training loop with AMP support and progress tracking
+    model.train()
+    correct = 0
+    total = 0
+
+    pbar = tqdm(train_loader, desc="Training")
+    for batch_idx, batch in enumerate(pbar):
+        start_time = time.perf_counter()
+        images = [batch[img_key].to(device) for img_key in cfg.training.image_keys]
+        labels = batch[cfg.training.label_key].float().to(device)
+
+        # Forward pass with optional AMP
+        with torch.autocast(device_type=device.type) if support_amp(device, cfg) else nullcontext():
+            outputs = model(images)
+            loss = criterion(outputs.logits, labels)
+
+        # Backward pass with gradient scaling if AMP enabled
+        optimizer.zero_grad()
+        if cfg.training.use_amp:
+            grad_scaler.scale(loss).backward()
+            grad_scaler.step(optimizer)
+            grad_scaler.update()
+        else:
+            loss.backward()
+            optimizer.step()
+
+        # Track metrics
+        if model.config.num_classes == 2:
+            predictions = (torch.sigmoid(outputs.logits) > 0.5).float()
+        else:
+            predictions = torch.argmax(outputs.logits, dim=1)
+        correct += (predictions == labels).sum().item()
+        total += labels.size(0)
+
+        current_acc = 100 * correct / total
+        train_info = {
+            "loss": loss.item(),
+            "accuracy": current_acc,
+            "dataloading_s": time.perf_counter() - start_time,
+        }
+
+        logger.log_dict(train_info, step + batch_idx, mode="train")
+        pbar.set_postfix({"loss": f"{loss.item():.4f}", "acc": f"{current_acc:.2f}%"})
+
+
+def validate(model, val_loader, criterion, device, logger, cfg, num_samples_to_log=8):
+    # Validation loop with metric tracking and sample logging
+    model.eval()
+    correct = 0
+    total = 0
+    batch_start_time = time.perf_counter()
+    samples = []
+    running_loss = 0
+
+    with (
+        torch.no_grad(),
+        torch.autocast(device_type=device.type) if support_amp(device, cfg) else nullcontext(),
+    ):
+        for batch in tqdm(val_loader, desc="Validation"):
+            images = [batch[img_key].to(device) for img_key in cfg.training.image_keys]
+            labels = batch[cfg.training.label_key].float().to(device)
+
+            outputs = model(images)
+            loss = criterion(outputs.logits, labels)
+
+            # Track metrics
+            if model.config.num_classes == 2:
+                predictions = (torch.sigmoid(outputs.logits) > 0.5).float()
+            else:
+                predictions = torch.argmax(outputs.logits, dim=1)
+            correct += (predictions == labels).sum().item()
+            total += labels.size(0)
+            running_loss += loss.item()
+
+            # Log sample predictions for visualization
+            if len(samples) < num_samples_to_log:
+                for i in range(min(num_samples_to_log - len(samples), len(images))):
+                    if model.config.num_classes == 2:
+                        confidence = round(outputs.probabilities[i].item(), 3)
+                    else:
+                        confidence = [round(prob, 3) for prob in outputs.probabilities[i].tolist()]
+                    samples.append(
+                        {
+                            "image": wandb.Image(images[i].cpu()),
+                            "true_label": labels[i].item(),
+                            "predicted": predictions[i].item(),
+                            "confidence": confidence,
+                        }
+                    )
+
+    accuracy = 100 * correct / total
+    avg_loss = running_loss / len(val_loader)
+    print(f"Average validation loss {avg_loss}, and accuracy {accuracy}")
+
+    eval_info = {
+        "loss": avg_loss,
+        "accuracy": accuracy,
+        "eval_s": time.perf_counter() - batch_start_time,
+        "eval/prediction_samples": wandb.Table(
+            data=[[s["image"], s["true_label"], s["predicted"], f"{s['confidence']}"] for s in samples],
+            columns=["Image", "True Label", "Predicted", "Confidence"],
+        )
+        if logger._cfg.wandb.enable
+        else None,
+    }
+
+    return accuracy, eval_info
+
+
+@hydra.main(version_base="1.2", config_path="../configs/policy", config_name="hilserl_classifier")
+def train(cfg: DictConfig) -> None:
+    # Main training pipeline with support for resuming training
+    logging.info(OmegaConf.to_yaml(cfg))
+
+    # Initialize training environment
+    device = get_safe_torch_device(cfg.device, log=True)
+    set_global_seed(cfg.seed)
+
+    out_dir = Path(cfg.output_dir)
+    out_dir.mkdir(parents=True, exist_ok=True)
+    logger = Logger(cfg, out_dir, cfg.wandb.job_name if cfg.wandb.enable else None)
+
+    # Setup dataset and dataloaders
+    dataset = LeRobotDataset(cfg.dataset_repo_id)
+    logging.info(f"Dataset size: {len(dataset)}")
+
+    train_size = int(cfg.train_split_proportion * len(dataset))
+    val_size = len(dataset) - train_size
+    train_dataset, val_dataset = random_split(dataset, [train_size, val_size])
+
+    sampler = create_balanced_sampler(train_dataset, cfg)
+    train_loader = DataLoader(
+        train_dataset,
+        batch_size=cfg.training.batch_size,
+        num_workers=cfg.training.num_workers,
+        sampler=sampler,
+        pin_memory=True,
+    )
+
+    val_loader = DataLoader(
+        val_dataset,
+        batch_size=cfg.eval.batch_size,
+        shuffle=False,
+        num_workers=cfg.training.num_workers,
+        pin_memory=True,
+    )
+
+    # Resume training if requested
+    step = 0
+    best_val_acc = 0
+
+    if cfg.resume:
+        if not Logger.get_last_checkpoint_dir(out_dir).exists():
+            raise RuntimeError(
+                "You have set resume=True, but there is no model checkpoint in "
+                f"{Logger.get_last_checkpoint_dir(out_dir)}"
+            )
+        checkpoint_cfg_path = str(Logger.get_last_pretrained_model_dir(out_dir) / "config.yaml")
+        logging.info(
+            colored(
+                "You have set resume=True, indicating that you wish to resume a run",
+                color="yellow",
+                attrs=["bold"],
+            )
+        )
+        # Load and validate checkpoint configuration
+        checkpoint_cfg = init_hydra_config(checkpoint_cfg_path)
+        # Check for differences between the checkpoint configuration and provided configuration.
+        # Hack to resolve the delta_timestamps ahead of time in order to properly diff.
+        resolve_delta_timestamps(cfg)
+        diff = DeepDiff(OmegaConf.to_container(checkpoint_cfg), OmegaConf.to_container(cfg))
+        # Ignore the `resume` and parameters.
+        if "values_changed" in diff and "root['resume']" in diff["values_changed"]:
+            del diff["values_changed"]["root['resume']"]
+        if len(diff) > 0:
+            logging.warning(
+                "At least one difference was detected between the checkpoint configuration and "
+                f"the provided configuration: \n{pformat(diff)}\nNote that the checkpoint configuration "
+                "takes precedence.",
+            )
+        # Use the checkpoint config instead of the provided config (but keep `resume` parameter).
+        cfg = checkpoint_cfg
+        cfg.resume = True
+
+    # Initialize model and training components
+    model = get_model(cfg=cfg, logger=logger).to(device)
+
+    optimizer = optim.AdamW(model.parameters(), lr=cfg.training.learning_rate)
+    # Use BCEWithLogitsLoss for binary classification and CrossEntropyLoss for multi-class
+    criterion = nn.BCEWithLogitsLoss() if model.config.num_classes == 2 else nn.CrossEntropyLoss()
+    grad_scaler = GradScaler(enabled=cfg.training.use_amp)
+
+    # Log model parameters
+    num_learnable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
+    num_total_params = sum(p.numel() for p in model.parameters())
+    logging.info(f"Learnable parameters: {format_big_number(num_learnable_params)}")
+    logging.info(f"Total parameters: {format_big_number(num_total_params)}")
+
+    if cfg.resume:
+        step = logger.load_last_training_state(optimizer, None)
+
+    # Training loop with validation and checkpointing
+    for epoch in range(cfg.training.num_epochs):
+        logging.info(f"\nEpoch {epoch+1}/{cfg.training.num_epochs}")
+
+        train_epoch(model, train_loader, criterion, optimizer, grad_scaler, device, logger, step, cfg)
+
+        # Periodic validation
+        if cfg.training.eval_freq > 0 and (epoch + 1) % cfg.training.eval_freq == 0:
+            val_acc, eval_info = validate(
+                model,
+                val_loader,
+                criterion,
+                device,
+                logger,
+                cfg,
+            )
+            logger.log_dict(eval_info, step + len(train_loader), mode="eval")
+
+            # Save best model
+            if val_acc > best_val_acc:
+                best_val_acc = val_acc
+                logger.save_checkpoint(
+                    train_step=step + len(train_loader),
+                    policy=model,
+                    optimizer=optimizer,
+                    scheduler=None,
+                    identifier="best",
+                )
+
+        # Periodic checkpointing
+        if cfg.training.save_checkpoint and (epoch + 1) % cfg.training.save_freq == 0:
+            logger.save_checkpoint(
+                train_step=step + len(train_loader),
+                policy=model,
+                optimizer=optimizer,
+                scheduler=None,
+                identifier=f"{epoch+1:06d}",
+            )
+
+        step += len(train_loader)
+
+    logging.info("Training completed")
+
+
+if __name__ == "__main__":
+    train()

lerobot/scripts/train_sac.py

@@ -0,0 +1,586 @@
+#!/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 logging
+import functools
+from pprint import pformat
+import random
+from typing import Optional, Sequence, TypedDict, Callable
+
+import hydra
+import torch
+import torch.nn.functional as F
+from torch import nn
+from tqdm import tqdm
+from deepdiff import DeepDiff
+from omegaconf import DictConfig, OmegaConf
+
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+
+# TODO: Remove the import of maniskill
+from lerobot.common.datasets.factory import make_dataset
+from lerobot.common.envs.factory import make_env, make_maniskill_env
+from lerobot.common.envs.utils import preprocess_observation, preprocess_maniskill_observation
+from lerobot.common.logger import Logger, log_output_dir
+from lerobot.common.policies.factory import make_policy
+from lerobot.common.policies.sac.modeling_sac import SACPolicy
+from lerobot.common.policies.utils import get_device_from_parameters
+from lerobot.common.utils.utils import (
+    format_big_number,
+    get_safe_torch_device,
+    init_hydra_config,
+    init_logging,
+    set_global_seed,
+)
+from lerobot.scripts.eval import eval_policy
+
+
+def make_optimizers_and_scheduler(cfg, policy):
+    optimizer_actor = torch.optim.Adam(
+        # NOTE: Handle the case of shared encoder where the encoder weights are not optimized with the gradient of the actor
+        params=policy.actor.parameters_to_optimize,
+        lr=policy.config.actor_lr,
+    )
+    optimizer_critic = torch.optim.Adam(
+        params=policy.critic_ensemble.parameters(), lr=policy.config.critic_lr
+    )
+    # We wrap policy log temperature in list because this is a torch tensor and not a nn.Module
+    optimizer_temperature = torch.optim.Adam(params=[policy.log_alpha], lr=policy.config.critic_lr)
+    lr_scheduler = None
+    optimizers = {
+        "actor": optimizer_actor,
+        "critic": optimizer_critic,
+        "temperature": optimizer_temperature,
+    }
+    return optimizers, lr_scheduler
+
+
+class Transition(TypedDict):
+    state: dict[str, torch.Tensor]
+    action: torch.Tensor
+    reward: float
+    next_state: dict[str, torch.Tensor]
+    done: bool
+    complementary_info: dict[str, torch.Tensor] = None
+
+
+class BatchTransition(TypedDict):
+    state: dict[str, torch.Tensor]
+    action: torch.Tensor
+    reward: torch.Tensor
+    next_state: dict[str, torch.Tensor]
+    done: torch.Tensor
+
+
+def random_crop_vectorized(images: torch.Tensor, output_size: tuple) -> torch.Tensor:
+    """
+    Perform a per-image random crop over a batch of images in a vectorized way.
+    (Same as shown previously.)
+    """
+    B, C, H, W = images.shape
+    crop_h, crop_w = output_size
+
+    if crop_h > H or crop_w > W:
+        raise ValueError(
+            f"Requested crop size ({crop_h}, {crop_w}) is bigger than the image size ({H}, {W})."
+        )
+
+    tops = torch.randint(0, H - crop_h + 1, (B,), device=images.device)
+    lefts = torch.randint(0, W - crop_w + 1, (B,), device=images.device)
+
+    rows = torch.arange(crop_h, device=images.device).unsqueeze(0) + tops.unsqueeze(1)
+    cols = torch.arange(crop_w, device=images.device).unsqueeze(0) + lefts.unsqueeze(1)
+
+    rows = rows.unsqueeze(2).expand(-1, -1, crop_w)  # (B, crop_h, crop_w)
+    cols = cols.unsqueeze(1).expand(-1, crop_h, -1)  # (B, crop_h, crop_w)
+
+    images_hwcn = images.permute(0, 2, 3, 1)  # (B, H, W, C)
+
+    # Gather pixels
+    cropped_hwcn = images_hwcn[torch.arange(B, device=images.device).view(B, 1, 1), rows, cols, :]
+    # cropped_hwcn => (B, crop_h, crop_w, C)
+
+    cropped = cropped_hwcn.permute(0, 3, 1, 2)  # (B, C, crop_h, crop_w)
+    return cropped
+
+
+def random_shift(images: torch.Tensor, pad: int = 4):
+    """Vectorized random shift, imgs: (B,C,H,W), pad: #pixels"""
+    _, _, h, w = images.shape
+    images = F.pad(input=images, pad=(pad, pad, pad, pad), mode="replicate")
+    return random_crop_vectorized(images=images, output_size=(h, w))
+
+
+class ReplayBuffer:
+    def __init__(
+        self,
+        capacity: int,
+        device: str = "cuda:0",
+        state_keys: Optional[Sequence[str]] = None,
+        image_augmentation_function: Optional[Callable] = None,
+        use_drq: bool = True,
+    ):
+        """
+        Args:
+            capacity (int): Maximum number of transitions to store in the buffer.
+            device (str): The device where the tensors will be moved ("cuda:0" or "cpu").
+            state_keys (List[str]): The list of keys that appear in `state` and `next_state`.
+            image_augmentation_function (Optional[Callable]): A function that takes a batch of images
+                and returns a batch of augmented images. If None, a default augmentation function is used.
+            use_drq (bool): Whether to use the default DRQ image augmentation style, when sampling in the buffer.
+        """
+        self.capacity = capacity
+        self.device = device
+        self.memory: list[Transition] = []
+        self.position = 0
+
+        # If no state_keys provided, default to an empty list
+        # (you can handle this differently if needed)
+        self.state_keys = state_keys if state_keys is not None else []
+        if image_augmentation_function is None:
+            self.image_augmentation_function = functools.partial(random_shift, pad=4)
+        self.use_drq = use_drq
+
+    def add(
+        self,
+        state: dict[str, torch.Tensor],
+        action: torch.Tensor,
+        reward: float,
+        next_state: dict[str, torch.Tensor],
+        done: bool,
+        complementary_info: Optional[dict[str, torch.Tensor]] = None,
+    ):
+        """Saves a transition."""
+        if len(self.memory) < self.capacity:
+            self.memory.append(None)
+
+        # Create and store the Transition
+        self.memory[self.position] = Transition(
+            state=state,
+            action=action,
+            reward=reward,
+            next_state=next_state,
+            done=done,
+            complementary_info=complementary_info,
+        )
+        self.position: int = (self.position + 1) % self.capacity
+
+    @classmethod
+    def from_lerobot_dataset(
+        cls,
+        lerobot_dataset: LeRobotDataset,
+        device: str = "cuda:0",
+        state_keys: Optional[Sequence[str]] = None,
+    ) -> "ReplayBuffer":
+        """
+        Convert a LeRobotDataset into a ReplayBuffer.
+
+        Args:
+            lerobot_dataset (LeRobotDataset): The dataset to convert.
+            device (str): The device . Defaults to "cuda:0".
+            state_keys (Optional[Sequence[str]], optional): The list of keys that appear in `state` and `next_state`.
+            Defaults to None.
+
+        Returns:
+            ReplayBuffer: The replay buffer with offline dataset transitions.
+        """
+        # We convert the LeRobotDataset into a replay buffer, because it is more efficient to sample from
+        # a replay buffer than from a lerobot dataset.
+        replay_buffer = cls(capacity=len(lerobot_dataset), device=device, state_keys=state_keys)
+        list_transition = cls._lerobotdataset_to_transitions(dataset=lerobot_dataset, state_keys=state_keys)
+        # Fill the replay buffer with the lerobot dataset transitions
+        for data in list_transition:
+            replay_buffer.add(
+                state=data["state"],
+                action=data["action"],
+                reward=data["reward"],
+                next_state=data["next_state"],
+                done=data["done"],
+            )
+        return replay_buffer
+
+    @staticmethod
+    def _lerobotdataset_to_transitions(
+        dataset: LeRobotDataset,
+        state_keys: Optional[Sequence[str]] = None,
+    ) -> list[Transition]:
+        """
+        Convert a LeRobotDataset into a list of RL (s, a, r, s', done) transitions.
+
+        Args:
+            dataset (LeRobotDataset):
+                The dataset to convert. Each item in the dataset is expected to have
+                at least the following keys:
+                {
+                    "action": ...
+                    "next.reward": ...
+                    "next.done": ...
+                    "episode_index": ...
+                }
+                plus whatever your 'state_keys' specify.
+
+            state_keys (Optional[Sequence[str]]):
+                The dataset keys to include in 'state' and 'next_state'. Their names
+                will be kept as-is in the output transitions. E.g.
+                ["observation.state", "observation.environment_state"].
+                If None, you must handle or define default keys.
+
+        Returns:
+            transitions (List[Transition]):
+                A list of Transition dictionaries with the same length as `dataset`.
+        """
+
+        # If not provided, you can either raise an error or define a default:
+        if state_keys is None:
+            raise ValueError("You must provide a list of keys in `state_keys` that define your 'state'.")
+
+        transitions: list[Transition] = []
+        num_frames = len(dataset)
+
+        for i in tqdm(range(num_frames)):
+            current_sample = dataset[i]
+
+            # ----- 1) Current state -----
+            current_state: dict[str, torch.Tensor] = {}
+            for key in state_keys:
+                val = current_sample[key]
+                current_state[key] = val.unsqueeze(0)  # Add batch dimension
+
+            # ----- 2) Action -----
+            action = current_sample["action"].unsqueeze(0)  # Add batch dimension
+
+            # ----- 3) Reward and done -----
+            reward = float(current_sample["next.reward"].item())  # ensure float
+            done = bool(current_sample["next.done"].item())  # ensure bool
+
+            # ----- 4) Next state -----
+            # If not done and the next sample is in the same episode, we pull the next sample's state.
+            # Otherwise (done=True or next sample crosses to a new episode), next_state = current_state.
+            next_state = current_state  # default
+            if not done and (i < num_frames - 1):
+                next_sample = dataset[i + 1]
+                if next_sample["episode_index"] == current_sample["episode_index"]:
+                    # Build next_state from the same keys
+                    next_state_data: dict[str, torch.Tensor] = {}
+                    for key in state_keys:
+                        val = next_sample[key]
+                        next_state_data[key] = val.unsqueeze(0)  # Add batch dimension
+                    next_state = next_state_data
+
+            # ----- Construct the Transition -----
+            transition = Transition(
+                state=current_state,
+                action=action,
+                reward=reward,
+                next_state=next_state,
+                done=done,
+            )
+            transitions.append(transition)
+
+        return transitions
+
+    def sample(self, batch_size: int) -> BatchTransition:
+        """Sample a random batch of transitions and collate them into batched tensors."""
+        list_of_transitions = random.sample(self.memory, batch_size)
+
+        # -- Build batched states --
+        batch_state = {}
+        for key in self.state_keys:
+            batch_state[key] = torch.cat([t["state"][key] for t in list_of_transitions], dim=0).to(
+                self.device
+            )
+            if key.startswith("observation.image") and self.use_drq:
+                batch_state[key] = self.image_augmentation_function(batch_state[key])
+
+        # -- Build batched actions --
+        batch_actions = torch.cat([t["action"] for t in list_of_transitions]).to(self.device)
+
+        # -- Build batched rewards --
+        batch_rewards = torch.tensor([t["reward"] for t in list_of_transitions], dtype=torch.float32).to(
+            self.device
+        )
+
+        # -- Build batched next states --
+        batch_next_state = {}
+        for key in self.state_keys:
+            batch_next_state[key] = torch.cat([t["next_state"][key] for t in list_of_transitions], dim=0).to(
+                self.device
+            )
+            if key.startswith("observation.image") and self.use_drq:
+                batch_next_state[key] = self.image_augmentation_function(batch_next_state[key])
+
+        # -- Build batched dones --
+        batch_dones = torch.tensor([t["done"] for t in list_of_transitions], dtype=torch.float32).to(
+            self.device
+        )
+        batch_dones = torch.tensor([t["done"] for t in list_of_transitions], dtype=torch.float32).to(
+            self.device
+        )
+
+        # Return a BatchTransition typed dict
+        return BatchTransition(
+            state=batch_state,
+            action=batch_actions,
+            reward=batch_rewards,
+            next_state=batch_next_state,
+            done=batch_dones,
+        )
+
+
+def concatenate_batch_transitions(
+    left_batch_transitions: BatchTransition, right_batch_transition: BatchTransition
+) -> BatchTransition:
+    """NOTE: Be careful it change the left_batch_transitions in place"""
+    left_batch_transitions["state"] = {
+        key: torch.cat([left_batch_transitions["state"][key], right_batch_transition["state"][key]], dim=0)
+        for key in left_batch_transitions["state"]
+    }
+    left_batch_transitions["action"] = torch.cat(
+        [left_batch_transitions["action"], right_batch_transition["action"]], dim=0
+    )
+    left_batch_transitions["reward"] = torch.cat(
+        [left_batch_transitions["reward"], right_batch_transition["reward"]], dim=0
+    )
+    left_batch_transitions["next_state"] = {
+        key: torch.cat(
+            [left_batch_transitions["next_state"][key], right_batch_transition["next_state"][key]], dim=0
+        )
+        for key in left_batch_transitions["next_state"]
+    }
+    left_batch_transitions["done"] = torch.cat(
+        [left_batch_transitions["done"], right_batch_transition["done"]], dim=0
+    )
+    return left_batch_transitions
+
+
+def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = None):
+    if out_dir is None:
+        raise NotImplementedError()
+    if job_name is None:
+        raise NotImplementedError()
+
+    init_logging()
+    logging.info(pformat(OmegaConf.to_container(cfg)))
+
+    # Create an env dedicated to online episodes collection from policy rollout.
+    # online_env = make_env(cfg, n_envs=cfg.training.online_rollout_batch_size)
+    # NOTE: Off policy algorithm are efficient enought to use a single environment
+    logging.info("make_env online")
+    # online_env = make_env(cfg, n_envs=1)
+    # TODO: Remove the import of maniskill and unifiy with make env
+    online_env = make_maniskill_env(cfg, n_envs=1)
+    if cfg.training.eval_freq > 0:
+        logging.info("make_env eval")
+        # eval_env = make_env(cfg, n_envs=1)
+        # TODO: Remove the import of maniskill and unifiy with make env
+        eval_env = make_maniskill_env(cfg, n_envs=1)
+
+    # TODO: Add a way to resume training
+
+    # log metrics to terminal and wandb
+    logger = Logger(cfg, out_dir, wandb_job_name=job_name)
+
+    set_global_seed(cfg.seed)
+
+    # Check device is available
+    device = get_safe_torch_device(cfg.device, log=True)
+
+    torch.backends.cudnn.benchmark = True
+    torch.backends.cuda.matmul.allow_tf32 = True
+
+    logging.info("make_policy")
+    # TODO: At some point we should just need make sac policy
+    policy: SACPolicy = make_policy(
+        hydra_cfg=cfg,
+        # dataset_stats=offline_dataset.meta.stats if not cfg.resume else None,
+        # Hack: But if we do online traning, we do not need dataset_stats
+        dataset_stats=None,
+        pretrained_policy_name_or_path=str(logger.last_pretrained_model_dir) if cfg.resume else None,
+        device=device,
+    )
+    assert isinstance(policy, nn.Module)
+
+    optimizers, lr_scheduler = make_optimizers_and_scheduler(cfg, policy)
+
+    # TODO: Handle resume
+
+    num_learnable_params = sum(p.numel() for p in policy.parameters() if p.requires_grad)
+    num_total_params = sum(p.numel() for p in policy.parameters())
+
+    log_output_dir(out_dir)
+    logging.info(f"{cfg.env.task=}")
+    logging.info(f"{cfg.training.online_steps=}")
+    logging.info(f"{num_learnable_params=} ({format_big_number(num_learnable_params)})")
+    logging.info(f"{num_total_params=} ({format_big_number(num_total_params)})")
+
+    obs, info = online_env.reset()
+
+    # HACK for maniskill
+    # obs = preprocess_observation(obs)
+    obs = preprocess_maniskill_observation(obs)
+    obs = {key: obs[key].to(device, non_blocking=True) for key in obs}
+
+    replay_buffer = ReplayBuffer(
+        capacity=cfg.training.online_buffer_capacity, device=device, state_keys=cfg.policy.input_shapes.keys()
+    )
+
+    batch_size = cfg.training.batch_size
+
+    if cfg.dataset_repo_id is not None:
+        logging.info("make_dataset offline buffer")
+        offline_dataset = make_dataset(cfg)
+        logging.info("Convertion to a offline replay buffer")
+        offline_replay_buffer = ReplayBuffer.from_lerobot_dataset(
+            offline_dataset, device=device, state_keys=cfg.policy.input_shapes.keys()
+        )
+        batch_size: int = batch_size // 2  # We will sample from both replay buffer
+
+    # NOTE: For the moment we will solely handle the case of a single environment
+    sum_reward_episode = 0
+
+    for interaction_step in range(cfg.training.online_steps):
+        # NOTE: At some point we should use a  wrapper to handle the observation
+
+        if interaction_step >= cfg.training.online_step_before_learning:
+            action = policy.select_action(batch=obs)
+            next_obs, reward, done, truncated, info = online_env.step(action.cpu().numpy())
+        else:
+            action = online_env.action_space.sample()
+            next_obs, reward, done, truncated, info = online_env.step(action)
+            # HACK
+            action = torch.tensor(action, dtype=torch.float32).to(device, non_blocking=True)
+
+        # HACK: For maniskill
+        # next_obs = preprocess_observation(next_obs)
+        next_obs = preprocess_maniskill_observation(next_obs)
+        next_obs = {key: next_obs[key].to(device, non_blocking=True) for key in obs}
+        sum_reward_episode += float(reward[0])
+        # Because we are using a single environment
+        # we can safely assume that the episode is done
+        if done[0] or truncated[0]:
+            logging.info(f"Global step {interaction_step}: Episode reward: {sum_reward_episode}")
+            logger.log_dict({"Sum episode reward": sum_reward_episode}, interaction_step)
+            sum_reward_episode = 0
+            # HACK: This is for maniskill
+            logging.info(
+                f"global step {interaction_step}: episode success: {info['success'].float().item()} \n"
+            )
+            logger.log_dict({"Episode success": info["success"].float().item()}, interaction_step)
+
+        replay_buffer.add(
+            state=obs,
+            action=action,
+            reward=float(reward[0]),
+            next_state=next_obs,
+            done=done[0],
+        )
+        obs = next_obs
+
+        if interaction_step < cfg.training.online_step_before_learning:
+            continue
+        for _ in range(cfg.policy.utd_ratio - 1):
+            batch = replay_buffer.sample(batch_size)
+            if cfg.dataset_repo_id is not None:
+                batch_offline = offline_replay_buffer.sample(batch_size)
+                batch = concatenate_batch_transitions(batch, batch_offline)
+
+            actions = batch["action"]
+            rewards = batch["reward"]
+            observations = batch["state"]
+            next_observations = batch["next_state"]
+            done = batch["done"]
+
+            loss_critic = policy.compute_loss_critic(
+                observations=observations,
+                actions=actions,
+                rewards=rewards,
+                next_observations=next_observations,
+                done=done,
+            )
+            optimizers["critic"].zero_grad()
+            loss_critic.backward()
+            optimizers["critic"].step()
+
+        batch = replay_buffer.sample(batch_size)
+        if cfg.dataset_repo_id is not None:
+            batch_offline = offline_replay_buffer.sample(batch_size)
+            batch = concatenate_batch_transitions(
+                left_batch_transitions=batch, right_batch_transition=batch_offline
+            )
+
+        actions = batch["action"]
+        rewards = batch["reward"]
+        observations = batch["state"]
+        next_observations = batch["next_state"]
+        done = batch["done"]
+
+        loss_critic = policy.compute_loss_critic(
+            observations=observations,
+            actions=actions,
+            rewards=rewards,
+            next_observations=next_observations,
+            done=done,
+        )
+        optimizers["critic"].zero_grad()
+        loss_critic.backward()
+        optimizers["critic"].step()
+
+        training_infos = {}
+        training_infos["loss_critic"] = loss_critic.item()
+
+        if interaction_step % cfg.training.policy_update_freq == 0:
+            # TD3 Trick
+            for _ in range(cfg.training.policy_update_freq):
+                loss_actor = policy.compute_loss_actor(observations=observations)
+
+                optimizers["actor"].zero_grad()
+                loss_actor.backward()
+                optimizers["actor"].step()
+
+                training_infos["loss_actor"] = loss_actor.item()
+
+                loss_temperature = policy.compute_loss_temperature(observations=observations)
+                optimizers["temperature"].zero_grad()
+                loss_temperature.backward()
+                optimizers["temperature"].step()
+
+                training_infos["loss_temperature"] = loss_temperature.item()
+
+        if interaction_step % cfg.training.log_freq == 0:
+            logger.log_dict(training_infos, interaction_step, mode="train")
+
+        policy.update_target_networks()
+
+
+@hydra.main(version_base="1.2", config_name="default", config_path="../configs")
+def train_cli(cfg: dict):
+    train(
+        cfg,
+        out_dir=hydra.core.hydra_config.HydraConfig.get().run.dir,
+        job_name=hydra.core.hydra_config.HydraConfig.get().job.name,
+    )
+
+
+def train_notebook(out_dir=None, job_name=None, config_name="default", config_path="../configs"):
+    from hydra import compose, initialize
+
+    hydra.core.global_hydra.GlobalHydra.instance().clear()
+    initialize(config_path=config_path)
+    cfg = compose(config_name=config_name)
+    train(cfg, out_dir=out_dir, job_name=job_name)
+
+
+if __name__ == "__main__":
+    train_cli()

lerobot/scripts/visualize_dataset.py

@@ -100,7 +100,7 @@ def to_hwc_uint8_numpy(chw_float32_torch: torch.Tensor) -> np.ndarray:
 
 
 def visualize_dataset(
-    repo_id: str,
+    dataset: LeRobotDataset,
     episode_index: int,
     batch_size: int = 32,
     num_workers: int = 0,
@@ -108,7 +108,6 @@ def visualize_dataset(
     web_port: int = 9090,
     ws_port: int = 9087,
     save: bool = False,
-    root: Path | None = None,
     output_dir: Path | None = None,
 ) -> Path | None:
     if save:
@@ -116,8 +115,7 @@ def visualize_dataset(
             output_dir is not None
         ), "Set an output directory where to write .rrd files with `--output-dir path/to/directory`."
 
-    logging.info("Loading dataset")
-    dataset = LeRobotDataset(repo_id, root=root)
+    repo_id = dataset.repo_id
 
     logging.info("Loading dataloader")
     episode_sampler = EpisodeSampler(dataset, episode_index)
@@ -153,7 +151,7 @@ def visualize_dataset(
             rr.set_time_seconds("timestamp", batch["timestamp"][i].item())
 
             # display each camera image
-            for key in dataset.camera_keys:
+            for key in dataset.meta.camera_keys:
                 # TODO(rcadene): add `.compress()`? is it lossless?
                 rr.log(key, rr.Image(to_hwc_uint8_numpy(batch[key][i])))
 
@@ -209,11 +207,17 @@ def main():
         required=True,
         help="Episode to visualize.",
     )
+    parser.add_argument(
+        "--local-files-only",
+        type=int,
+        default=0,
+        help="Use local files only. By default, this script will try to fetch the dataset from the hub if it exists.",
+    )
     parser.add_argument(
         "--root",
         type=Path,
         default=None,
-        help="Root directory for a dataset stored locally (e.g. `--root data`). By default, the dataset will be loaded from hugging face cache folder, or downloaded from the hub if available.",
+        help="Root directory for the dataset stored locally (e.g. `--root data`). By default, the dataset will be loaded from hugging face cache folder, or downloaded from the hub if available.",
     )
     parser.add_argument(
         "--output-dir",
@@ -268,7 +272,15 @@ def main():
     )
 
     args = parser.parse_args()
-    visualize_dataset(**vars(args))
+    kwargs = vars(args)
+    repo_id = kwargs.pop("repo_id")
+    root = kwargs.pop("root")
+    local_files_only = kwargs.pop("local_files_only")
+
+    logging.info("Loading dataset")
+    dataset = LeRobotDataset(repo_id, root=root, local_files_only=local_files_only)
+
+    visualize_dataset(dataset, **vars(args))
 
 
 if __name__ == "__main__":

lerobot/scripts/visualize_dataset_html.py

@@ -53,20 +53,29 @@ python lerobot/scripts/visualize_dataset_html.py \
 """
 
 import argparse
+import csv
+import json
 import logging
+import re
 import shutil
+import tempfile
+from io import StringIO
 from pathlib import Path
 
-import tqdm
-from flask import Flask, redirect, render_template, url_for
+import numpy as np
+import pandas as pd
+import requests
+from flask import Flask, redirect, render_template, request, url_for
 
+from lerobot import available_datasets
 from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.common.datasets.utils import IterableNamespace
 from lerobot.common.utils.utils import init_logging
 
 
 def run_server(
-    dataset: LeRobotDataset,
-    episodes: list[int],
+    dataset: LeRobotDataset | IterableNamespace | None,
+    episodes: list[int] | None,
     host: str,
     port: str,
     static_folder: Path,
@@ -76,10 +85,50 @@ def run_server(
     app.config["SEND_FILE_MAX_AGE_DEFAULT"] = 0  # specifying not to cache
 
     @app.route("/")
-    def index():
-        # home page redirects to the first episode page
-        [dataset_namespace, dataset_name] = dataset.repo_id.split("/")
-        first_episode_id = episodes[0]
+    def hommepage(dataset=dataset):
+        if dataset:
+            dataset_namespace, dataset_name = dataset.repo_id.split("/")
+            return redirect(
+                url_for(
+                    "show_episode",
+                    dataset_namespace=dataset_namespace,
+                    dataset_name=dataset_name,
+                    episode_id=0,
+                )
+            )
+
+        dataset_param, episode_param = None, None
+        all_params = request.args
+        if "dataset" in all_params:
+            dataset_param = all_params["dataset"]
+        if "episode" in all_params:
+            episode_param = int(all_params["episode"])
+
+        if dataset_param:
+            dataset_namespace, dataset_name = dataset_param.split("/")
+            return redirect(
+                url_for(
+                    "show_episode",
+                    dataset_namespace=dataset_namespace,
+                    dataset_name=dataset_name,
+                    episode_id=episode_param if episode_param is not None else 0,
+                )
+            )
+
+        featured_datasets = [
+            "lerobot/aloha_static_cups_open",
+            "lerobot/columbia_cairlab_pusht_real",
+            "lerobot/taco_play",
+        ]
+        return render_template(
+            "visualize_dataset_homepage.html",
+            featured_datasets=featured_datasets,
+            lerobot_datasets=available_datasets,
+        )
+
+    @app.route("/<string:dataset_namespace>/<string:dataset_name>")
+    def show_first_episode(dataset_namespace, dataset_name):
+        first_episode_id = 0
         return redirect(
             url_for(
                 "show_episode",
@@ -90,31 +139,85 @@ def run_server(
         )
 
     @app.route("/<string:dataset_namespace>/<string:dataset_name>/episode_<int:episode_id>")
-    def show_episode(dataset_namespace, dataset_name, episode_id):
+    def show_episode(dataset_namespace, dataset_name, episode_id, dataset=dataset, episodes=episodes):
+        repo_id = f"{dataset_namespace}/{dataset_name}"
+        try:
+            if dataset is None:
+                dataset = get_dataset_info(repo_id)
+        except FileNotFoundError:
+            return (
+                "Make sure to convert your LeRobotDataset to v2 & above. See how to convert your dataset at https://github.com/huggingface/lerobot/pull/461",
+                400,
+            )
+        dataset_version = (
+            dataset.meta._version if isinstance(dataset, LeRobotDataset) else dataset.codebase_version
+        )
+        match = re.search(r"v(\d+)\.", dataset_version)
+        if match:
+            major_version = int(match.group(1))
+            if major_version < 2:
+                return "Make sure to convert your LeRobotDataset to v2 & above."
+
+        episode_data_csv_str, columns = get_episode_data(dataset, episode_id)
         dataset_info = {
-            "repo_id": dataset.repo_id,
-            "num_samples": dataset.num_samples,
-            "num_episodes": dataset.num_episodes,
+            "repo_id": f"{dataset_namespace}/{dataset_name}",
+            "num_samples": dataset.num_frames
+            if isinstance(dataset, LeRobotDataset)
+            else dataset.total_frames,
+            "num_episodes": dataset.num_episodes
+            if isinstance(dataset, LeRobotDataset)
+            else dataset.total_episodes,
             "fps": dataset.fps,
         }
-        video_paths = get_episode_video_paths(dataset, episode_id)
-        language_instruction = get_episode_language_instruction(dataset, episode_id)
-        videos_info = [
-            {"url": url_for("static", filename=video_path), "filename": Path(video_path).name}
-            for video_path in video_paths
-        ]
-        if language_instruction:
-            videos_info[0]["language_instruction"] = language_instruction
+        if isinstance(dataset, LeRobotDataset):
+            video_paths = [
+                dataset.meta.get_video_file_path(episode_id, key) for key in dataset.meta.video_keys
+            ]
+            videos_info = [
+                {"url": url_for("static", filename=video_path), "filename": video_path.parent.name}
+                for video_path in video_paths
+            ]
+            tasks = dataset.meta.episodes[episode_id]["tasks"]
+        else:
+            video_keys = [key for key, ft in dataset.features.items() if ft["dtype"] == "video"]
+            videos_info = [
+                {
+                    "url": f"https://huggingface.co/datasets/{repo_id}/resolve/main/"
+                    + dataset.video_path.format(
+                        episode_chunk=int(episode_id) // dataset.chunks_size,
+                        video_key=video_key,
+                        episode_index=episode_id,
+                    ),
+                    "filename": video_key,
+                }
+                for video_key in video_keys
+            ]
+
+            response = requests.get(
+                f"https://huggingface.co/datasets/{repo_id}/resolve/main/meta/episodes.jsonl"
+            )
+            response.raise_for_status()
+            # Split into lines and parse each line as JSON
+            tasks_jsonl = [json.loads(line) for line in response.text.splitlines() if line.strip()]
+
+            filtered_tasks_jsonl = [row for row in tasks_jsonl if row["episode_index"] == episode_id]
+            tasks = filtered_tasks_jsonl[0]["tasks"]
+
+        videos_info[0]["language_instruction"] = tasks
+
+        if episodes is None:
+            episodes = list(
+                range(dataset.num_episodes if isinstance(dataset, LeRobotDataset) else dataset.total_episodes)
+            )
 
-        ep_csv_url = url_for("static", filename=get_ep_csv_fname(episode_id))
         return render_template(
             "visualize_dataset_template.html",
             episode_id=episode_id,
             episodes=episodes,
             dataset_info=dataset_info,
             videos_info=videos_info,
-            ep_csv_url=ep_csv_url,
-            has_policy=False,
+            episode_data_csv_str=episode_data_csv_str,
+            columns=columns,
         )
 
     app.run(host=host, port=port)
@@ -125,46 +228,69 @@ def get_ep_csv_fname(episode_id: int):
     return ep_csv_fname
 
 
-def write_episode_data_csv(output_dir, file_name, episode_index, dataset):
-    """Write a csv file containg timeseries data of an episode (e.g. state and action).
+def get_episode_data(dataset: LeRobotDataset | IterableNamespace, episode_index):
+    """Get a csv str containing timeseries data of an episode (e.g. state and action).
     This file will be loaded by Dygraph javascript to plot data in real time."""
-    from_idx = dataset.episode_data_index["from"][episode_index]
-    to_idx = dataset.episode_data_index["to"][episode_index]
+    columns = []
 
-    has_state = "observation.state" in dataset.hf_dataset.features
-    has_action = "action" in dataset.hf_dataset.features
+    selected_columns = [col for col, ft in dataset.features.items() if ft["dtype"] == "float32"]
+    selected_columns.remove("timestamp")
 
     # init header of csv with state and action names
     header = ["timestamp"]
-    if has_state:
-        dim_state = len(dataset.hf_dataset["observation.state"][0])
-        header += [f"state_{i}" for i in range(dim_state)]
-    if has_action:
-        dim_action = len(dataset.hf_dataset["action"][0])
-        header += [f"action_{i}" for i in range(dim_action)]
 
-    columns = ["timestamp"]
-    if has_state:
-        columns += ["observation.state"]
-    if has_action:
-        columns += ["action"]
+    for column_name in selected_columns:
+        dim_state = (
+            dataset.meta.shapes[column_name][0]
+            if isinstance(dataset, LeRobotDataset)
+            else dataset.features[column_name].shape[0]
+        )
+        header += [f"{column_name}_{i}" for i in range(dim_state)]
 
-    rows = []
-    data = dataset.hf_dataset.select_columns(columns)
-    for i in range(from_idx, to_idx):
-        row = [data[i]["timestamp"].item()]
-        if has_state:
-            row += data[i]["observation.state"].tolist()
-        if has_action:
-            row += data[i]["action"].tolist()
-        rows.append(row)
+        if "names" in dataset.features[column_name] and dataset.features[column_name]["names"]:
+            column_names = dataset.features[column_name]["names"]
+            while not isinstance(column_names, list):
+                column_names = list(column_names.values())[0]
+        else:
+            column_names = [f"motor_{i}" for i in range(dim_state)]
+        columns.append({"key": column_name, "value": column_names})
 
-    output_dir.mkdir(parents=True, exist_ok=True)
-    with open(output_dir / file_name, "w") as f:
-        f.write(",".join(header) + "\n")
-        for row in rows:
-            row_str = [str(col) for col in row]
-            f.write(",".join(row_str) + "\n")
+    selected_columns.insert(0, "timestamp")
+
+    if isinstance(dataset, LeRobotDataset):
+        from_idx = dataset.episode_data_index["from"][episode_index]
+        to_idx = dataset.episode_data_index["to"][episode_index]
+        data = (
+            dataset.hf_dataset.select(range(from_idx, to_idx))
+            .select_columns(selected_columns)
+            .with_format("pandas")
+        )
+    else:
+        repo_id = dataset.repo_id
+
+        url = f"https://huggingface.co/datasets/{repo_id}/resolve/main/" + dataset.data_path.format(
+            episode_chunk=int(episode_index) // dataset.chunks_size, episode_index=episode_index
+        )
+        df = pd.read_parquet(url)
+        data = df[selected_columns]  # Select specific columns
+
+    rows = np.hstack(
+        (
+            np.expand_dims(data["timestamp"], axis=1),
+            *[np.vstack(data[col]) for col in selected_columns[1:]],
+        )
+    ).tolist()
+
+    # Convert data to CSV string
+    csv_buffer = StringIO()
+    csv_writer = csv.writer(csv_buffer)
+    # Write header
+    csv_writer.writerow(header)
+    # Write data rows
+    csv_writer.writerows(rows)
+    csv_string = csv_buffer.getvalue()
+
+    return csv_string, columns
 
 
 def get_episode_video_paths(dataset: LeRobotDataset, ep_index: int) -> list[str]:
@@ -172,13 +298,13 @@ def get_episode_video_paths(dataset: LeRobotDataset, ep_index: int) -> list[str]
     first_frame_idx = dataset.episode_data_index["from"][ep_index].item()
     return [
         dataset.hf_dataset.select_columns(key)[first_frame_idx][key]["path"]
-        for key in dataset.video_frame_keys
+        for key in dataset.meta.video_keys
     ]
 
 
 def get_episode_language_instruction(dataset: LeRobotDataset, ep_index: int) -> list[str]:
     # check if the dataset has language instructions
-    if "language_instruction" not in dataset.hf_dataset.features:
+    if "language_instruction" not in dataset.features:
         return None
 
     # get first frame index
@@ -190,10 +316,17 @@ def get_episode_language_instruction(dataset: LeRobotDataset, ep_index: int) ->
     return language_instruction.removeprefix("tf.Tensor(b'").removesuffix("', shape=(), dtype=string)")
 
 
+def get_dataset_info(repo_id: str) -> IterableNamespace:
+    response = requests.get(f"https://huggingface.co/datasets/{repo_id}/resolve/main/meta/info.json")
+    response.raise_for_status()  # Raises an HTTPError for bad responses
+    dataset_info = response.json()
+    dataset_info["repo_id"] = repo_id
+    return IterableNamespace(dataset_info)
+
+
 def visualize_dataset_html(
-    repo_id: str,
-    root: Path | None = None,
-    episodes: list[int] = None,
+    dataset: LeRobotDataset | None,
+    episodes: list[int] | None = None,
     output_dir: Path | None = None,
     serve: bool = True,
     host: str = "127.0.0.1",
@@ -202,13 +335,11 @@ def visualize_dataset_html(
 ) -> Path | None:
     init_logging()
 
-    dataset = LeRobotDataset(repo_id, root=root)
-
-    if not dataset.video:
-        raise NotImplementedError(f"Image datasets ({dataset.video=}) are currently not supported.")
+    template_dir = Path(__file__).resolve().parent.parent / "templates"
 
     if output_dir is None:
-        output_dir = f"outputs/visualize_dataset_html/{repo_id}"
+        # Create a temporary directory that will be automatically cleaned up
+        output_dir = tempfile.mkdtemp(prefix="lerobot_visualize_dataset_")
 
     output_dir = Path(output_dir)
     if output_dir.exists():
@@ -219,28 +350,29 @@ def visualize_dataset_html(
 
     output_dir.mkdir(parents=True, exist_ok=True)
 
-    # Create a simlink from the dataset video folder containg mp4 files to the output directory
-    # so that the http server can get access to the mp4 files.
     static_dir = output_dir / "static"
     static_dir.mkdir(parents=True, exist_ok=True)
-    ln_videos_dir = static_dir / "videos"
-    if not ln_videos_dir.exists():
-        ln_videos_dir.symlink_to(dataset.videos_dir.resolve())
 
-    template_dir = Path(__file__).resolve().parent.parent / "templates"
+    if dataset is None:
+        if serve:
+            run_server(
+                dataset=None,
+                episodes=None,
+                host=host,
+                port=port,
+                static_folder=static_dir,
+                template_folder=template_dir,
+            )
+    else:
+        # Create a simlink from the dataset video folder containg mp4 files to the output directory
+        # so that the http server can get access to the mp4 files.
+        if isinstance(dataset, LeRobotDataset):
+            ln_videos_dir = static_dir / "videos"
+            if not ln_videos_dir.exists():
+                ln_videos_dir.symlink_to((dataset.root / "videos").resolve())
 
-    if episodes is None:
-        episodes = list(range(dataset.num_episodes))
-
-    logging.info("Writing CSV files")
-    for episode_index in tqdm.tqdm(episodes):
-        # write states and actions in a csv (it can be slow for big datasets)
-        ep_csv_fname = get_ep_csv_fname(episode_index)
-        # TODO(rcadene): speedup script by loading directly from dataset, pyarrow, parquet, safetensors?
-        write_episode_data_csv(static_dir, ep_csv_fname, episode_index, dataset)
-
-    if serve:
-        run_server(dataset, episodes, host, port, static_dir, template_dir)
+        if serve:
+            run_server(dataset, episodes, host, port, static_dir, template_dir)
 
 
 def main():
@@ -249,15 +381,27 @@ def main():
     parser.add_argument(
         "--repo-id",
         type=str,
-        required=True,
+        default=None,
         help="Name of hugging face repositery containing a LeRobotDataset dataset (e.g. `lerobot/pusht` for https://huggingface.co/datasets/lerobot/pusht).",
     )
+    parser.add_argument(
+        "--local-files-only",
+        type=int,
+        default=0,
+        help="Use local files only. By default, this script will try to fetch the dataset from the hub if it exists.",
+    )
     parser.add_argument(
         "--root",
         type=Path,
         default=None,
         help="Root directory for a dataset stored locally (e.g. `--root data`). By default, the dataset will be loaded from hugging face cache folder, or downloaded from the hub if available.",
     )
+    parser.add_argument(
+        "--load-from-hf-hub",
+        type=int,
+        default=0,
+        help="Load videos and parquet files from HF Hub rather than local system.",
+    )
     parser.add_argument(
         "--episodes",
         type=int,
@@ -297,7 +441,21 @@ def main():
     )
 
     args = parser.parse_args()
-    visualize_dataset_html(**vars(args))
+    kwargs = vars(args)
+    repo_id = kwargs.pop("repo_id")
+    load_from_hf_hub = kwargs.pop("load_from_hf_hub")
+    root = kwargs.pop("root")
+    local_files_only = kwargs.pop("local_files_only")
+
+    dataset = None
+    if repo_id:
+        dataset = (
+            LeRobotDataset(repo_id, root=root, local_files_only=local_files_only)
+            if not load_from_hf_hub
+            else get_dataset_info(repo_id)
+        )
+
+    visualize_dataset_html(dataset, **vars(args))
 
 
 if __name__ == "__main__":

lerobot/scripts/visualize_image_transforms.py

@@ -157,7 +157,7 @@ def visualize_transforms(cfg, output_dir: Path, n_examples: int = 5):
     output_dir.mkdir(parents=True, exist_ok=True)
 
     # Get 1st frame from 1st camera of 1st episode
-    original_frame = dataset[0][dataset.camera_keys[0]]
+    original_frame = dataset[0][dataset.meta.camera_keys[0]]
     to_pil(original_frame).save(output_dir / "original_frame.png", quality=100)
     print("\nOriginal frame saved to:")
     print(f"    {output_dir / 'original_frame.png'}.")

lerobot/templates/visualize_dataset_homepage.html

@@ -0,0 +1,68 @@
+<!DOCTYPE html>
+<html lang="en">
+<head>
+    <meta charset="UTF-8">
+    <meta name="viewport" content="width=device-width, initial-scale=1.0">
+    <title>Interactive Video Background Page</title>
+    <script src="https://cdn.tailwindcss.com"></script>
+    <script defer src="https://cdn.jsdelivr.net/npm/alpinejs@3.x.x/dist/cdn.min.js"></script>
+</head>
+<body class="h-screen overflow-hidden font-mono text-white" x-data="{ 
+    inputValue: '',
+    navigateToDataset() {
+        const trimmedValue = this.inputValue.trim();
+        if (trimmedValue) {
+            window.location.href = `/${trimmedValue}`;
+        }
+    }
+}">
+    <div class="fixed inset-0 w-full h-full overflow-hidden">
+        <video class="absolute min-w-full min-h-full w-auto h-auto top-1/2 left-1/2 transform -translate-x-1/2 -translate-y-1/2" autoplay muted loop>
+            <source src="https://huggingface.co/datasets/cadene/koch_bimanual_folding/resolve/v1.6/videos/observation.images.phone_episode_000037.mp4" type="video/mp4">
+            Your browser does not support HTML5 video.
+        </video>
+    </div>
+    <div class="fixed inset-0 bg-black bg-opacity-80"></div>
+    <div class="relative z-10 flex flex-col items-center justify-center h-screen">
+        <div class="text-center mb-8">
+            <h1 class="text-4xl font-bold mb-4">LeRobot Dataset Visualizer</h1>
+
+            <a href="https://x.com/RemiCadene/status/1825455895561859185" target="_blank" rel="noopener noreferrer" class="underline">create & train your own robots</a>
+
+            <p class="text-xl mb-4"></p>
+            <div class="text-left inline-block">
+                <h3 class="font-semibold mb-2 mt-4">Example Datasets:</h3>
+                <ul class="list-disc list-inside">
+                    {% for dataset in featured_datasets %}
+                        <li><a href="/{{ dataset }}" class="text-blue-300 hover:text-blue-100 hover:underline">{{ dataset }}</a></li>
+                    {% endfor %}
+                </ul>
+            </div>
+        </div>
+        <div class="flex w-full max-w-lg px-4 mb-4">
+            <input 
+                type="text" 
+                x-model="inputValue"
+                @keyup.enter="navigateToDataset"
+                placeholder="enter dataset id (ex: lerobot/droid_100)"
+                class="flex-grow px-4 py-2 rounded-l bg-white bg-opacity-20 text-white placeholder-gray-300 focus:outline-none focus:ring-2 focus:ring-blue-300"
+            >
+            <button 
+                @click="navigateToDataset"
+                class="px-4 py-2 bg-blue-500 text-white rounded-r hover:bg-blue-600 focus:outline-none focus:ring-2 focus:ring-blue-300"
+            >
+                Go
+            </button>
+        </div>
+
+        <details class="mt-4 max-w-full px-4">
+            <summary>More example datasets</summary>
+            <ul class="list-disc list-inside max-h-28 overflow-y-auto break-all">
+                {% for dataset in lerobot_datasets %}
+                    <li><a href="/{{ dataset }}" class="text-blue-300 hover:text-blue-100 hover:underline">{{ dataset }}</a></li>
+                {% endfor %}
+            </ul>
+        </details>
+    </div>
+</body>
+</html>

lerobot/templates/visualize_dataset_template.html

@@ -31,7 +31,12 @@
 }">
     <!-- Sidebar -->
     <div x-ref="sidebar" class="bg-slate-900 p-5 break-words overflow-y-auto shrink-0 md:shrink md:w-60 md:max-h-screen">
-        <h1 class="mb-4 text-xl font-semibold">{{ dataset_info.repo_id }}</h1>
+        <a href="https://github.com/huggingface/lerobot" target="_blank" class="hidden md:block">
+            <img src="https://github.com/huggingface/lerobot/raw/main/media/lerobot-logo-thumbnail.png">
+        </a>
+        <a href="https://huggingface.co/datasets/{{ dataset_info.repo_id }}" target="_blank">
+            <h1 class="mb-4 text-xl font-semibold">{{ dataset_info.repo_id }}</h1>
+        </a>
 
         <ul>
             <li>
@@ -93,10 +98,35 @@
         </div>
 
         <!-- Videos -->
-        <div class="flex flex-wrap gap-1">
+        <div  class="max-w-32 relative text-sm mb-4 select-none"
+            @click.outside="isVideosDropdownOpen = false">
+            <div
+                @click="isVideosDropdownOpen = !isVideosDropdownOpen"
+                class="p-2 border border-slate-500 rounded flex justify-between items-center cursor-pointer"
+            >
+            <span class="truncate">filter videos</span>
+            <div class="transition-transform" :class="{ 'rotate-180': isVideosDropdownOpen }">🔽</div>
+            </div>
+    
+            <div x-show="isVideosDropdownOpen" 
+                class="absolute mt-1 border border-slate-500 rounded shadow-lg z-10">
+            <div>
+                <template x-for="option in videosKeys" :key="option">
+                <div
+                    @click="videosKeysSelected = videosKeysSelected.includes(option) ? videosKeysSelected.filter(v => v !== option) : [...videosKeysSelected, option]"
+                    class="p-2 cursor-pointer bg-slate-900"
+                    :class="{ 'bg-slate-700': videosKeysSelected.includes(option) }"
+                    x-text="option"
+                ></div>
+                </template>
+            </div>
+            </div>
+        </div>
+
+        <div class="flex flex-wrap gap-x-2 gap-y-6">
             {% for video_info in videos_info %}
-            <div x-show="!videoCodecError" class="max-w-96">
-                <p class="text-sm text-gray-300 bg-gray-800 px-2 rounded-t-xl truncate">{{ video_info.filename }}</p>
+            <div x-show="!videoCodecError && videosKeysSelected.includes('{{ video_info.filename }}')" class="max-w-96 relative">
+                <p class="absolute inset-x-0 -top-4 text-sm text-gray-300 bg-gray-800 px-2 rounded-t-xl truncate">{{ video_info.filename }}</p>
                 <video muted loop type="video/mp4" class="object-contain w-full h-full" @canplaythrough="videoCanPlay" @timeupdate="() => {
                     if (video.duration) {
                       const time = video.currentTime;
@@ -182,12 +212,12 @@
                 <thead>
                     <tr>
                         <th></th>
-                        <template x-for="(_, colIndex) in Array.from({length: nColumns}, (_, index) => index)">
+                        <template x-for="(_, colIndex) in Array.from({length: columns.length}, (_, index) => index)">
                             <th class="border border-slate-700">
                                 <div class="flex gap-x-2 justify-between px-2">
                                     <input type="checkbox" :checked="isColumnChecked(colIndex)"
                                         @change="toggleColumn(colIndex)">
-                                    <p x-text="`${columnNames[colIndex]}`"></p>
+                                    <p x-text="`${columns[colIndex].key}`"></p>
                                 </div>
                             </th>
                         </template>
@@ -197,10 +227,10 @@
                     <template x-for="(row, rowIndex) in rows">
                         <tr class="odd:bg-gray-800 even:bg-gray-900">
                             <td class="border border-slate-700">
-                                <div class="flex gap-x-2 w-24 font-semibold px-1">
+                                <div class="flex gap-x-2 max-w-64 font-semibold px-1 break-all">
                                     <input type="checkbox" :checked="isRowChecked(rowIndex)"
                                         @change="toggleRow(rowIndex)">
-                                    <p x-text="`Motor ${rowIndex}`"></p>
+                                    <p x-text="`${rowLabels[rowIndex]}`"></p>
                                 </div>
                             </td>
                             <template x-for="(cell, colIndex) in row">
@@ -222,16 +252,20 @@
         </div>
     </div>
 
+    <script>
+        const parentOrigin = "https://huggingface.co";
+        const searchParams = new URLSearchParams();
+        searchParams.set("dataset", "{{ dataset_info.repo_id }}");
+        searchParams.set("episode", "{{ episode_id }}");
+		window.parent.postMessage({ queryString: searchParams.toString() }, parentOrigin);
+    </script>
+
     <script>
         function createAlpineData() {
             return {
                 // state
                 dygraph: null,
                 currentFrameData: null,
-                columnNames: ["state", "action", "pred action"],
-                nColumns: 2,
-                nStates: 0,
-                nActions: 0,
                 checked: [],
                 dygraphTime: 0.0,
                 dygraphIndex: 0,
@@ -241,6 +275,11 @@
                 nVideos: {{ videos_info | length }},
                 nVideoReadyToPlay: 0,
                 videoCodecError: false,
+                isVideosDropdownOpen: false,
+                videosKeys: {{ videos_info | map(attribute='filename') | list | tojson }},
+                videosKeysSelected: [],
+                columns: {{ columns | tojson }},
+                rowLabels: {{ columns | tojson }}.reduce((colA, colB) => colA.value.length > colB.value.length ? colA : colB).value,
 
                 // alpine initialization
                 init() {
@@ -250,11 +289,19 @@
                     if(!canPlayVideos){
                         this.videoCodecError = true;
                     }
+                    this.videosKeysSelected = this.videosKeys.map(opt => opt)
+
+                    // process CSV data
+                    const csvDataStr = {{ episode_data_csv_str|tojson|safe }};
+                    // Create a Blob with the CSV data
+                    const blob = new Blob([csvDataStr], { type: 'text/csv;charset=utf-8;' });
+                    // Create a URL for the Blob
+                    const csvUrl = URL.createObjectURL(blob);
 
                     // process CSV data
                     this.videos = document.querySelectorAll('video');
                     this.video = this.videos[0];
-                    this.dygraph = new Dygraph(document.getElementById("graph"), '{{ ep_csv_url }}', {
+                    this.dygraph = new Dygraph(document.getElementById("graph"), csvUrl, {
                         pixelsPerPoint: 0.01,
                         legend: 'always',
                         labelsDiv: document.getElementById('labels'),
@@ -275,21 +322,17 @@
                                 this.colors = this.dygraph.getColors();
                                 this.checked = Array(this.colors.length).fill(true);
 
-                                const seriesNames = this.dygraph.getLabels().slice(1);
-                                this.nStates = seriesNames.findIndex(item => item.startsWith('action_'));
-                                this.nActions = seriesNames.length - this.nStates;
                                 const colors = [];
-                                const LIGHTNESS = [30, 65, 85]; // state_lightness, action_lightness, pred_action_lightness
-                                // colors for "state" lines
-                                for (let hue = 0; hue < 360; hue += parseInt(360/this.nStates)) {
-                                    const color = `hsl(${hue}, 100%, ${LIGHTNESS[0]}%)`;
-                                    colors.push(color);
-                                }
-                                // colors for "action" lines
-                                for (let hue = 0; hue < 360; hue += parseInt(360/this.nActions)) {
-                                    const color = `hsl(${hue}, 100%, ${LIGHTNESS[1]}%)`;
-                                    colors.push(color);
+                                let lightness = 30; // const LIGHTNESS = [30, 65, 85]; // state_lightness, action_lightness, pred_action_lightness
+                                for(const column of this.columns){
+                                    const nValues = column.value.length;
+                                    for (let hue = 0; hue < 360; hue += parseInt(360/nValues)) {
+                                        const color = `hsl(${hue}, 100%, ${lightness}%)`;
+                                        colors.push(color);
+                                    }
+                                    lightness += 35;
                                 }
+
                                 this.dygraph.updateOptions({ colors });
                                 this.colors = colors;
 
@@ -316,17 +359,19 @@
                         return [];
                     }
                     const rows = [];
-                    const nRows = Math.max(this.nStates, this.nActions);
+                    const nRows = Math.max(...this.columns.map(column => column.value.length));
                     let rowIndex = 0;
                     while(rowIndex < nRows){
                         const row = [];
                         // number of states may NOT match number of actions. In this case, we null-pad the 2D array to make a fully rectangular 2d array
                         const nullCell = { isNull: true };
-                        const stateValueIdx = rowIndex;
-                        const actionValueIdx = stateValueIdx + this.nStates; // because this.currentFrameData = [state0, state1, ..., stateN, action0, action1, ..., actionN]
                         // row consists of [state value, action value]
-                        row.push(rowIndex < this.nStates ? this.currentFrameData[stateValueIdx] : nullCell); // push "state value" to row
-                        row.push(rowIndex < this.nActions ? this.currentFrameData[actionValueIdx] : nullCell); // push "action value" to row
+                        let idx = rowIndex;
+                        for(const column of this.columns){
+                            const nColumn = column.value.length;
+                            row.push(rowIndex < nColumn ? this.currentFrameData[idx] : nullCell);
+                            idx += nColumn; // because this.currentFrameData = [state0, state1, ..., stateN, action0, action1, ..., actionN]
+                        }
                         rowIndex += 1;
                         rows.push(row);
                     }

pyproject.toml

@@ -43,9 +43,8 @@ opencv-python = ">=4.9.0"
 diffusers = ">=0.27.2"
 torchvision = ">=0.17.1"
 h5py = ">=3.10.0"
-huggingface-hub = {extras = ["hf-transfer", "cli"], version = ">=0.25.0"}
-# TODO(rcadene, aliberts): Make gym 1.0.0 work
-gymnasium = "==0.29.1"
+huggingface-hub = {extras = ["hf-transfer", "cli"], version = ">=0.25.2"}
+gymnasium = "==0.29.1" # TODO(rcadene, aliberts): Make gym 1.0.0 work
 cmake = ">=3.29.0.1"
 gym-dora = { git = "https://github.com/dora-rs/dora-lerobot.git", subdirectory = "gym_dora", optional = true }
 gym-pusht = { version = ">=0.1.5", optional = true}
@@ -71,6 +70,9 @@ pyrealsense2 = {version = ">=2.55.1.6486", markers = "sys_platform != 'darwin'",
 pyrender = {git = "https://github.com/mmatl/pyrender.git", markers = "sys_platform == 'linux'", optional = true}
 hello-robot-stretch-body = {version = ">=0.7.27", markers = "sys_platform == 'linux'", optional = true}
 pyserial = {version = ">=3.5", optional = true}
+jsonlines = ">=4.0.0"
+transformers = {version = ">=4.47.0", optional = true}
+torchmetrics = {version = ">=1.6.0", optional = true}
 
 
 [tool.poetry.extras]
@@ -86,6 +88,7 @@ dynamixel = ["dynamixel-sdk", "pynput"]
 feetech = ["feetech-servo-sdk", "pynput"]
 intelrealsense = ["pyrealsense2"]
 stretch = ["hello-robot-stretch-body", "pyrender", "pyrealsense2", "pynput"]
+hilserl = ["transformers", "torchmetrics"]
 
 [tool.ruff]
 line-length = 110

tests/conftest.py

@@ -14,15 +14,24 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import random
 import traceback
 
 import pytest
+import torch
 from serial import SerialException
 
 from lerobot import available_cameras, available_motors, available_robots
 from lerobot.common.utils.utils import init_hydra_config
 from tests.utils import DEVICE, ROBOT_CONFIG_PATH_TEMPLATE, make_camera, make_motors_bus
 
+# Import fixture modules as plugins
+pytest_plugins = [
+    "tests.fixtures.dataset_factories",
+    "tests.fixtures.files",
+    "tests.fixtures.hub",
+]
+
 
 def pytest_collection_finish():
     print(f"\nTesting with {DEVICE=}")
@@ -117,3 +126,14 @@ def patch_builtins_input(monkeypatch):
             print(text)
 
     monkeypatch.setattr("builtins.input", print_text)
+
+
+def pytest_addoption(parser):
+    parser.addoption("--seed", action="store", default="42", help="Set random seed for reproducibility")
+
+
+@pytest.fixture(autouse=True)
+def set_random_seed(request):
+    seed = int(request.config.getoption("--seed"))
+    random.seed(seed)  # Python random
+    torch.manual_seed(seed)  # PyTorch

tests/fixtures/constants.py

@@ -0,0 +1,29 @@
+from lerobot.common.datasets.lerobot_dataset import LEROBOT_HOME
+
+LEROBOT_TEST_DIR = LEROBOT_HOME / "_testing"
+DUMMY_REPO_ID = "dummy/repo"
+DUMMY_ROBOT_TYPE = "dummy_robot"
+DUMMY_MOTOR_FEATURES = {
+    "action": {
+        "dtype": "float32",
+        "shape": (6,),
+        "names": ["shoulder_pan", "shoulder_lift", "elbow_flex", "wrist_flex", "wrist_roll", "gripper"],
+    },
+    "state": {
+        "dtype": "float32",
+        "shape": (6,),
+        "names": ["shoulder_pan", "shoulder_lift", "elbow_flex", "wrist_flex", "wrist_roll", "gripper"],
+    },
+}
+DUMMY_CAMERA_FEATURES = {
+    "laptop": {"shape": (480, 640, 3), "names": ["height", "width", "channels"], "info": None},
+    "phone": {"shape": (480, 640, 3), "names": ["height", "width", "channels"], "info": None},
+}
+DEFAULT_FPS = 30
+DUMMY_VIDEO_INFO = {
+    "video.fps": DEFAULT_FPS,
+    "video.codec": "av1",
+    "video.pix_fmt": "yuv420p",
+    "video.is_depth_map": False,
+    "has_audio": False,
+}

tests/fixtures/dataset_factories.py

@@ -0,0 +1,396 @@
+import random
+from pathlib import Path
+from unittest.mock import patch
+
+import datasets
+import numpy as np
+import PIL.Image
+import pytest
+import torch
+
+from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.common.datasets.utils import (
+    DEFAULT_CHUNK_SIZE,
+    DEFAULT_FEATURES,
+    DEFAULT_PARQUET_PATH,
+    DEFAULT_VIDEO_PATH,
+    get_hf_features_from_features,
+    hf_transform_to_torch,
+)
+from tests.fixtures.constants import (
+    DEFAULT_FPS,
+    DUMMY_CAMERA_FEATURES,
+    DUMMY_MOTOR_FEATURES,
+    DUMMY_REPO_ID,
+    DUMMY_ROBOT_TYPE,
+    DUMMY_VIDEO_INFO,
+)
+
+
+def get_task_index(task_dicts: dict, task: str) -> int:
+    tasks = {d["task_index"]: d["task"] for d in task_dicts}
+    task_to_task_index = {task: task_idx for task_idx, task in tasks.items()}
+    return task_to_task_index[task]
+
+
+@pytest.fixture(scope="session")
+def img_tensor_factory():
+    def _create_img_tensor(height=100, width=100, channels=3, dtype=torch.float32) -> torch.Tensor:
+        return torch.rand((channels, height, width), dtype=dtype)
+
+    return _create_img_tensor
+
+
+@pytest.fixture(scope="session")
+def img_array_factory():
+    def _create_img_array(height=100, width=100, channels=3, dtype=np.uint8) -> np.ndarray:
+        if np.issubdtype(dtype, np.unsignedinteger):
+            # Int array in [0, 255] range
+            img_array = np.random.randint(0, 256, size=(height, width, channels), dtype=dtype)
+        elif np.issubdtype(dtype, np.floating):
+            # Float array in [0, 1] range
+            img_array = np.random.rand(height, width, channels).astype(dtype)
+        else:
+            raise ValueError(dtype)
+        return img_array
+
+    return _create_img_array
+
+
+@pytest.fixture(scope="session")
+def img_factory(img_array_factory):
+    def _create_img(height=100, width=100) -> PIL.Image.Image:
+        img_array = img_array_factory(height=height, width=width)
+        return PIL.Image.fromarray(img_array)
+
+    return _create_img
+
+
+@pytest.fixture(scope="session")
+def features_factory():
+    def _create_features(
+        motor_features: dict = DUMMY_MOTOR_FEATURES,
+        camera_features: dict = DUMMY_CAMERA_FEATURES,
+        use_videos: bool = True,
+    ) -> dict:
+        if use_videos:
+            camera_ft = {
+                key: {"dtype": "video", **ft, **DUMMY_VIDEO_INFO} for key, ft in camera_features.items()
+            }
+        else:
+            camera_ft = {key: {"dtype": "image", **ft} for key, ft in camera_features.items()}
+        return {
+            **motor_features,
+            **camera_ft,
+            **DEFAULT_FEATURES,
+        }
+
+    return _create_features
+
+
+@pytest.fixture(scope="session")
+def info_factory(features_factory):
+    def _create_info(
+        codebase_version: str = CODEBASE_VERSION,
+        fps: int = DEFAULT_FPS,
+        robot_type: str = DUMMY_ROBOT_TYPE,
+        total_episodes: int = 0,
+        total_frames: int = 0,
+        total_tasks: int = 0,
+        total_videos: int = 0,
+        total_chunks: int = 0,
+        chunks_size: int = DEFAULT_CHUNK_SIZE,
+        data_path: str = DEFAULT_PARQUET_PATH,
+        video_path: str = DEFAULT_VIDEO_PATH,
+        motor_features: dict = DUMMY_MOTOR_FEATURES,
+        camera_features: dict = DUMMY_CAMERA_FEATURES,
+        use_videos: bool = True,
+    ) -> dict:
+        features = features_factory(motor_features, camera_features, use_videos)
+        return {
+            "codebase_version": codebase_version,
+            "robot_type": robot_type,
+            "total_episodes": total_episodes,
+            "total_frames": total_frames,
+            "total_tasks": total_tasks,
+            "total_videos": total_videos,
+            "total_chunks": total_chunks,
+            "chunks_size": chunks_size,
+            "fps": fps,
+            "splits": {},
+            "data_path": data_path,
+            "video_path": video_path if use_videos else None,
+            "features": features,
+        }
+
+    return _create_info
+
+
+@pytest.fixture(scope="session")
+def stats_factory():
+    def _create_stats(
+        features: dict[str] | None = None,
+    ) -> dict:
+        stats = {}
+        for key, ft in features.items():
+            shape = ft["shape"]
+            dtype = ft["dtype"]
+            if dtype in ["image", "video"]:
+                stats[key] = {
+                    "max": np.full((3, 1, 1), 1, dtype=np.float32).tolist(),
+                    "mean": np.full((3, 1, 1), 0.5, dtype=np.float32).tolist(),
+                    "min": np.full((3, 1, 1), 0, dtype=np.float32).tolist(),
+                    "std": np.full((3, 1, 1), 0.25, dtype=np.float32).tolist(),
+                }
+            else:
+                stats[key] = {
+                    "max": np.full(shape, 1, dtype=dtype).tolist(),
+                    "mean": np.full(shape, 0.5, dtype=dtype).tolist(),
+                    "min": np.full(shape, 0, dtype=dtype).tolist(),
+                    "std": np.full(shape, 0.25, dtype=dtype).tolist(),
+                }
+        return stats
+
+    return _create_stats
+
+
+@pytest.fixture(scope="session")
+def tasks_factory():
+    def _create_tasks(total_tasks: int = 3) -> int:
+        tasks_list = []
+        for i in range(total_tasks):
+            task_dict = {"task_index": i, "task": f"Perform action {i}."}
+            tasks_list.append(task_dict)
+        return tasks_list
+
+    return _create_tasks
+
+
+@pytest.fixture(scope="session")
+def episodes_factory(tasks_factory):
+    def _create_episodes(
+        total_episodes: int = 3,
+        total_frames: int = 400,
+        tasks: dict | None = None,
+        multi_task: bool = False,
+    ):
+        if total_episodes <= 0 or total_frames <= 0:
+            raise ValueError("num_episodes and total_length must be positive integers.")
+        if total_frames < total_episodes:
+            raise ValueError("total_length must be greater than or equal to num_episodes.")
+
+        if not tasks:
+            min_tasks = 2 if multi_task else 1
+            total_tasks = random.randint(min_tasks, total_episodes)
+            tasks = tasks_factory(total_tasks)
+
+        if total_episodes < len(tasks) and not multi_task:
+            raise ValueError("The number of tasks should be less than the number of episodes.")
+
+        # Generate random lengths that sum up to total_length
+        lengths = np.random.multinomial(total_frames, [1 / total_episodes] * total_episodes).tolist()
+
+        tasks_list = [task_dict["task"] for task_dict in tasks]
+        num_tasks_available = len(tasks_list)
+
+        episodes_list = []
+        remaining_tasks = tasks_list.copy()
+        for ep_idx in range(total_episodes):
+            num_tasks_in_episode = random.randint(1, min(3, num_tasks_available)) if multi_task else 1
+            tasks_to_sample = remaining_tasks if remaining_tasks else tasks_list
+            episode_tasks = random.sample(tasks_to_sample, min(num_tasks_in_episode, len(tasks_to_sample)))
+            if remaining_tasks:
+                for task in episode_tasks:
+                    remaining_tasks.remove(task)
+
+            episodes_list.append(
+                {
+                    "episode_index": ep_idx,
+                    "tasks": episode_tasks,
+                    "length": lengths[ep_idx],
+                }
+            )
+
+        return episodes_list
+
+    return _create_episodes
+
+
+@pytest.fixture(scope="session")
+def hf_dataset_factory(features_factory, tasks_factory, episodes_factory, img_array_factory):
+    def _create_hf_dataset(
+        features: dict | None = None,
+        tasks: list[dict] | None = None,
+        episodes: list[dict] | None = None,
+        fps: int = DEFAULT_FPS,
+    ) -> datasets.Dataset:
+        if not tasks:
+            tasks = tasks_factory()
+        if not episodes:
+            episodes = episodes_factory()
+        if not features:
+            features = features_factory()
+
+        timestamp_col = np.array([], dtype=np.float32)
+        frame_index_col = np.array([], dtype=np.int64)
+        episode_index_col = np.array([], dtype=np.int64)
+        task_index = np.array([], dtype=np.int64)
+        for ep_dict in episodes:
+            timestamp_col = np.concatenate((timestamp_col, np.arange(ep_dict["length"]) / fps))
+            frame_index_col = np.concatenate((frame_index_col, np.arange(ep_dict["length"], dtype=int)))
+            episode_index_col = np.concatenate(
+                (episode_index_col, np.full(ep_dict["length"], ep_dict["episode_index"], dtype=int))
+            )
+            ep_task_index = get_task_index(tasks, ep_dict["tasks"][0])
+            task_index = np.concatenate((task_index, np.full(ep_dict["length"], ep_task_index, dtype=int)))
+
+        index_col = np.arange(len(episode_index_col))
+
+        robot_cols = {}
+        for key, ft in features.items():
+            if ft["dtype"] == "image":
+                robot_cols[key] = [
+                    img_array_factory(height=ft["shapes"][1], width=ft["shapes"][0])
+                    for _ in range(len(index_col))
+                ]
+            elif ft["shape"][0] > 1 and ft["dtype"] != "video":
+                robot_cols[key] = np.random.random((len(index_col), ft["shape"][0])).astype(ft["dtype"])
+
+        hf_features = get_hf_features_from_features(features)
+        dataset = datasets.Dataset.from_dict(
+            {
+                **robot_cols,
+                "timestamp": timestamp_col,
+                "frame_index": frame_index_col,
+                "episode_index": episode_index_col,
+                "index": index_col,
+                "task_index": task_index,
+            },
+            features=hf_features,
+        )
+        dataset.set_transform(hf_transform_to_torch)
+        return dataset
+
+    return _create_hf_dataset
+
+
+@pytest.fixture(scope="session")
+def lerobot_dataset_metadata_factory(
+    info_factory,
+    stats_factory,
+    tasks_factory,
+    episodes_factory,
+    mock_snapshot_download_factory,
+):
+    def _create_lerobot_dataset_metadata(
+        root: Path,
+        repo_id: str = DUMMY_REPO_ID,
+        info: dict | None = None,
+        stats: dict | None = None,
+        tasks: list[dict] | None = None,
+        episodes: list[dict] | None = None,
+        local_files_only: bool = False,
+    ) -> LeRobotDatasetMetadata:
+        if not info:
+            info = info_factory()
+        if not stats:
+            stats = stats_factory(features=info["features"])
+        if not tasks:
+            tasks = tasks_factory(total_tasks=info["total_tasks"])
+        if not episodes:
+            episodes = episodes_factory(
+                total_episodes=info["total_episodes"], total_frames=info["total_frames"], tasks=tasks
+            )
+
+        mock_snapshot_download = mock_snapshot_download_factory(
+            info=info,
+            stats=stats,
+            tasks=tasks,
+            episodes=episodes,
+        )
+        with (
+            patch(
+                "lerobot.common.datasets.lerobot_dataset.get_hub_safe_version"
+            ) as mock_get_hub_safe_version_patch,
+            patch(
+                "lerobot.common.datasets.lerobot_dataset.snapshot_download"
+            ) as mock_snapshot_download_patch,
+        ):
+            mock_get_hub_safe_version_patch.side_effect = lambda repo_id, version: version
+            mock_snapshot_download_patch.side_effect = mock_snapshot_download
+
+            return LeRobotDatasetMetadata(repo_id=repo_id, root=root, local_files_only=local_files_only)
+
+    return _create_lerobot_dataset_metadata
+
+
+@pytest.fixture(scope="session")
+def lerobot_dataset_factory(
+    info_factory,
+    stats_factory,
+    tasks_factory,
+    episodes_factory,
+    hf_dataset_factory,
+    mock_snapshot_download_factory,
+    lerobot_dataset_metadata_factory,
+):
+    def _create_lerobot_dataset(
+        root: Path,
+        repo_id: str = DUMMY_REPO_ID,
+        total_episodes: int = 3,
+        total_frames: int = 150,
+        total_tasks: int = 1,
+        multi_task: bool = False,
+        info: dict | None = None,
+        stats: dict | None = None,
+        tasks: list[dict] | None = None,
+        episode_dicts: list[dict] | None = None,
+        hf_dataset: datasets.Dataset | None = None,
+        **kwargs,
+    ) -> LeRobotDataset:
+        if not info:
+            info = info_factory(
+                total_episodes=total_episodes, total_frames=total_frames, total_tasks=total_tasks
+            )
+        if not stats:
+            stats = stats_factory(features=info["features"])
+        if not tasks:
+            tasks = tasks_factory(total_tasks=info["total_tasks"])
+        if not episode_dicts:
+            episode_dicts = episodes_factory(
+                total_episodes=info["total_episodes"],
+                total_frames=info["total_frames"],
+                tasks=tasks,
+                multi_task=multi_task,
+            )
+        if not hf_dataset:
+            hf_dataset = hf_dataset_factory(tasks=tasks, episodes=episode_dicts, fps=info["fps"])
+
+        mock_snapshot_download = mock_snapshot_download_factory(
+            info=info,
+            stats=stats,
+            tasks=tasks,
+            episodes=episode_dicts,
+            hf_dataset=hf_dataset,
+        )
+        mock_metadata = lerobot_dataset_metadata_factory(
+            root=root,
+            repo_id=repo_id,
+            info=info,
+            stats=stats,
+            tasks=tasks,
+            episodes=episode_dicts,
+            local_files_only=kwargs.get("local_files_only", False),
+        )
+        with (
+            patch("lerobot.common.datasets.lerobot_dataset.LeRobotDatasetMetadata") as mock_metadata_patch,
+            patch(
+                "lerobot.common.datasets.lerobot_dataset.snapshot_download"
+            ) as mock_snapshot_download_patch,
+        ):
+            mock_metadata_patch.return_value = mock_metadata
+            mock_snapshot_download_patch.side_effect = mock_snapshot_download
+
+            return LeRobotDataset(repo_id=repo_id, root=root, **kwargs)
+
+    return _create_lerobot_dataset

tests/fixtures/files.py

@@ -0,0 +1,114 @@
+import json
+from pathlib import Path
+
+import datasets
+import jsonlines
+import pyarrow.compute as pc
+import pyarrow.parquet as pq
+import pytest
+
+from lerobot.common.datasets.utils import EPISODES_PATH, INFO_PATH, STATS_PATH, TASKS_PATH
+
+
+@pytest.fixture(scope="session")
+def info_path(info_factory):
+    def _create_info_json_file(dir: Path, info: dict | None = None) -> Path:
+        if not info:
+            info = info_factory()
+        fpath = dir / INFO_PATH
+        fpath.parent.mkdir(parents=True, exist_ok=True)
+        with open(fpath, "w") as f:
+            json.dump(info, f, indent=4, ensure_ascii=False)
+        return fpath
+
+    return _create_info_json_file
+
+
+@pytest.fixture(scope="session")
+def stats_path(stats_factory):
+    def _create_stats_json_file(dir: Path, stats: dict | None = None) -> Path:
+        if not stats:
+            stats = stats_factory()
+        fpath = dir / STATS_PATH
+        fpath.parent.mkdir(parents=True, exist_ok=True)
+        with open(fpath, "w") as f:
+            json.dump(stats, f, indent=4, ensure_ascii=False)
+        return fpath
+
+    return _create_stats_json_file
+
+
+@pytest.fixture(scope="session")
+def tasks_path(tasks_factory):
+    def _create_tasks_jsonl_file(dir: Path, tasks: list | None = None) -> Path:
+        if not tasks:
+            tasks = tasks_factory()
+        fpath = dir / TASKS_PATH
+        fpath.parent.mkdir(parents=True, exist_ok=True)
+        with jsonlines.open(fpath, "w") as writer:
+            writer.write_all(tasks)
+        return fpath
+
+    return _create_tasks_jsonl_file
+
+
+@pytest.fixture(scope="session")
+def episode_path(episodes_factory):
+    def _create_episodes_jsonl_file(dir: Path, episodes: list | None = None) -> Path:
+        if not episodes:
+            episodes = episodes_factory()
+        fpath = dir / EPISODES_PATH
+        fpath.parent.mkdir(parents=True, exist_ok=True)
+        with jsonlines.open(fpath, "w") as writer:
+            writer.write_all(episodes)
+        return fpath
+
+    return _create_episodes_jsonl_file
+
+
+@pytest.fixture(scope="session")
+def single_episode_parquet_path(hf_dataset_factory, info_factory):
+    def _create_single_episode_parquet(
+        dir: Path, ep_idx: int = 0, hf_dataset: datasets.Dataset | None = None, info: dict | None = None
+    ) -> Path:
+        if not info:
+            info = info_factory()
+        if hf_dataset is None:
+            hf_dataset = hf_dataset_factory()
+
+        data_path = info["data_path"]
+        chunks_size = info["chunks_size"]
+        ep_chunk = ep_idx // chunks_size
+        fpath = dir / data_path.format(episode_chunk=ep_chunk, episode_index=ep_idx)
+        fpath.parent.mkdir(parents=True, exist_ok=True)
+        table = hf_dataset.data.table
+        ep_table = table.filter(pc.equal(table["episode_index"], ep_idx))
+        pq.write_table(ep_table, fpath)
+        return fpath
+
+    return _create_single_episode_parquet
+
+
+@pytest.fixture(scope="session")
+def multi_episode_parquet_path(hf_dataset_factory, info_factory):
+    def _create_multi_episode_parquet(
+        dir: Path, hf_dataset: datasets.Dataset | None = None, info: dict | None = None
+    ) -> Path:
+        if not info:
+            info = info_factory()
+        if hf_dataset is None:
+            hf_dataset = hf_dataset_factory()
+
+        data_path = info["data_path"]
+        chunks_size = info["chunks_size"]
+        total_episodes = info["total_episodes"]
+        for ep_idx in range(total_episodes):
+            ep_chunk = ep_idx // chunks_size
+            fpath = dir / data_path.format(episode_chunk=ep_chunk, episode_index=ep_idx)
+            fpath.parent.mkdir(parents=True, exist_ok=True)
+            table = hf_dataset.data.table
+            ep_table = table.filter(pc.equal(table["episode_index"], ep_idx))
+            pq.write_table(ep_table, fpath)
+        return dir / "data"
+
+    return _create_multi_episode_parquet

tests/fixtures/hub.py

@@ -0,0 +1,105 @@
+from pathlib import Path
+
+import datasets
+import pytest
+from huggingface_hub.utils import filter_repo_objects
+
+from lerobot.common.datasets.utils import EPISODES_PATH, INFO_PATH, STATS_PATH, TASKS_PATH
+from tests.fixtures.constants import LEROBOT_TEST_DIR
+
+
+@pytest.fixture(scope="session")
+def mock_snapshot_download_factory(
+    info_factory,
+    info_path,
+    stats_factory,
+    stats_path,
+    tasks_factory,
+    tasks_path,
+    episodes_factory,
+    episode_path,
+    single_episode_parquet_path,
+    hf_dataset_factory,
+):
+    """
+    This factory allows to patch snapshot_download such that when called, it will create expected files rather
+    than making calls to the hub api. Its design allows to pass explicitly files which you want to be created.
+    """
+
+    def _mock_snapshot_download_func(
+        info: dict | None = None,
+        stats: dict | None = None,
+        tasks: list[dict] | None = None,
+        episodes: list[dict] | None = None,
+        hf_dataset: datasets.Dataset | None = None,
+    ):
+        if not info:
+            info = info_factory()
+        if not stats:
+            stats = stats_factory(features=info["features"])
+        if not tasks:
+            tasks = tasks_factory(total_tasks=info["total_tasks"])
+        if not episodes:
+            episodes = episodes_factory(
+                total_episodes=info["total_episodes"], total_frames=info["total_frames"], tasks=tasks
+            )
+        if not hf_dataset:
+            hf_dataset = hf_dataset_factory(tasks=tasks, episodes=episodes, fps=info["fps"])
+
+        def _extract_episode_index_from_path(fpath: str) -> int:
+            path = Path(fpath)
+            if path.suffix == ".parquet" and path.stem.startswith("episode_"):
+                episode_index = int(path.stem[len("episode_") :])  # 'episode_000000' -> 0
+                return episode_index
+            else:
+                return None
+
+        def _mock_snapshot_download(
+            repo_id: str,
+            local_dir: str | Path | None = None,
+            allow_patterns: str | list[str] | None = None,
+            ignore_patterns: str | list[str] | None = None,
+            *args,
+            **kwargs,
+        ) -> str:
+            if not local_dir:
+                local_dir = LEROBOT_TEST_DIR
+
+            # List all possible files
+            all_files = []
+            meta_files = [INFO_PATH, STATS_PATH, TASKS_PATH, EPISODES_PATH]
+            all_files.extend(meta_files)
+
+            data_files = []
+            for episode_dict in episodes:
+                ep_idx = episode_dict["episode_index"]
+                ep_chunk = ep_idx // info["chunks_size"]
+                data_path = info["data_path"].format(episode_chunk=ep_chunk, episode_index=ep_idx)
+                data_files.append(data_path)
+            all_files.extend(data_files)
+
+            allowed_files = filter_repo_objects(
+                all_files, allow_patterns=allow_patterns, ignore_patterns=ignore_patterns
+            )
+
+            # Create allowed files
+            for rel_path in allowed_files:
+                if rel_path.startswith("data/"):
+                    episode_index = _extract_episode_index_from_path(rel_path)
+                    if episode_index is not None:
+                        _ = single_episode_parquet_path(local_dir, episode_index, hf_dataset, info)
+                if rel_path == INFO_PATH:
+                    _ = info_path(local_dir, info)
+                elif rel_path == STATS_PATH:
+                    _ = stats_path(local_dir, stats)
+                elif rel_path == TASKS_PATH:
+                    _ = tasks_path(local_dir, tasks)
+                elif rel_path == EPISODES_PATH:
+                    _ = episode_path(local_dir, episodes)
+                else:
+                    pass
+            return str(local_dir)
+
+        return _mock_snapshot_download
+
+    return _mock_snapshot_download_func

tests/lerobot/common/datasets/test_utils.py

@@ -0,0 +1,38 @@
+#!/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.
+
+from huggingface_hub import DatasetCard
+
+from lerobot.common.datasets.utils import create_lerobot_dataset_card
+
+
+def test_default_parameters():
+    card = create_lerobot_dataset_card()
+    assert isinstance(card, DatasetCard)
+    assert card.data.tags == ["LeRobot"]
+    assert card.data.task_categories == ["robotics"]
+    assert card.data.configs == [
+        {
+            "config_name": "default",
+            "data_files": "data/*/*.parquet",
+        }
+    ]
+
+
+def test_with_tags():
+    tags = ["tag1", "tag2"]
+    card = create_lerobot_dataset_card(tags=tags)
+    assert card.data.tags == ["LeRobot", "tag1", "tag2"]

tests/policies/hilserl/classifier/check_hiserl_reward_classifier.py

@@ -0,0 +1,244 @@
+#!/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 logging
+
+import numpy as np
+import torch
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss
+from torch.optim import Adam
+from torch.utils.data import DataLoader
+from torchmetrics import AUROC, Accuracy, F1Score, Precision, Recall
+from torchvision.datasets import CIFAR10
+from torchvision.transforms import ToTensor
+
+from lerobot.common.policies.hilserl.classifier.modeling_classifier import Classifier, ClassifierConfig
+
+BATCH_SIZE = 1000
+LR = 0.1
+EPOCH_NUM = 2
+
+if torch.cuda.is_available():
+    DEVICE = torch.device("cuda")
+elif torch.backends.mps.is_available():
+    DEVICE = torch.device("mps")
+else:
+    DEVICE = torch.device("cpu")
+
+
+def train_evaluate_multiclass_classifier():
+    logging.info(
+        f"Start multiclass classifier train eval with {DEVICE} device, batch size {BATCH_SIZE}, learning rate {LR}"
+    )
+    multiclass_config = ClassifierConfig(model_name="microsoft/resnet-18", device=DEVICE, num_classes=10)
+    multiclass_classifier = Classifier(multiclass_config)
+
+    trainset = CIFAR10(root="data", train=True, download=True, transform=ToTensor())
+    testset = CIFAR10(root="data", train=False, download=True, transform=ToTensor())
+
+    trainloader = DataLoader(trainset, batch_size=BATCH_SIZE, shuffle=True)
+    testloader = DataLoader(testset, batch_size=BATCH_SIZE, shuffle=False)
+
+    multiclass_num_classes = 10
+    epoch = 1
+
+    criterion = CrossEntropyLoss()
+    optimizer = Adam(multiclass_classifier.parameters(), lr=LR)
+
+    multiclass_classifier.train()
+
+    logging.info("Start multiclass classifier training")
+
+    # Training loop
+    while epoch < EPOCH_NUM:  # loop over the dataset multiple times
+        for i, data in enumerate(trainloader):
+            inputs, labels = data
+            inputs, labels = inputs.to(DEVICE), labels.to(DEVICE)
+
+            # Zero the parameter gradients
+            optimizer.zero_grad()
+
+            # Forward pass
+            outputs = multiclass_classifier(inputs)
+
+            loss = criterion(outputs.logits, labels)
+            loss.backward()
+            optimizer.step()
+
+            if i % 10 == 0:  # print every 10 mini-batches
+                logging.info(f"[Epoch {epoch}, Batch {i}] loss: {loss.item():.3f}")
+
+        epoch += 1
+
+    print("Multiclass classifier training finished")
+
+    multiclass_classifier.eval()
+
+    test_loss = 0.0
+    test_labels = []
+    test_pridections = []
+    test_probs = []
+
+    with torch.no_grad():
+        for data in testloader:
+            images, labels = data
+            images, labels = images.to(DEVICE), labels.to(DEVICE)
+            outputs = multiclass_classifier(images)
+            loss = criterion(outputs.logits, labels)
+            test_loss += loss.item() * BATCH_SIZE
+
+            _, predicted = torch.max(outputs.logits, 1)
+            test_labels.extend(labels.cpu())
+            test_pridections.extend(predicted.cpu())
+            test_probs.extend(outputs.probabilities.cpu())
+
+    test_loss = test_loss / len(testset)
+
+    logging.info(f"Multiclass classifier test loss {test_loss:.3f}")
+
+    test_labels = torch.stack(test_labels)
+    test_predictions = torch.stack(test_pridections)
+    test_probs = torch.stack(test_probs)
+
+    accuracy = Accuracy(task="multiclass", num_classes=multiclass_num_classes)
+    precision = Precision(task="multiclass", average="weighted", num_classes=multiclass_num_classes)
+    recall = Recall(task="multiclass", average="weighted", num_classes=multiclass_num_classes)
+    f1 = F1Score(task="multiclass", average="weighted", num_classes=multiclass_num_classes)
+    auroc = AUROC(task="multiclass", num_classes=multiclass_num_classes, average="weighted")
+
+    # Calculate metrics
+    acc = accuracy(test_predictions, test_labels)
+    prec = precision(test_predictions, test_labels)
+    rec = recall(test_predictions, test_labels)
+    f1_score = f1(test_predictions, test_labels)
+    auroc_score = auroc(test_probs, test_labels)
+
+    logging.info(f"Accuracy: {acc:.2f}")
+    logging.info(f"Precision: {prec:.2f}")
+    logging.info(f"Recall: {rec:.2f}")
+    logging.info(f"F1 Score: {f1_score:.2f}")
+    logging.info(f"AUROC Score: {auroc_score:.2f}")
+
+
+def train_evaluate_binary_classifier():
+    logging.info(
+        f"Start binary classifier train eval with {DEVICE} device, batch size {BATCH_SIZE}, learning rate {LR}"
+    )
+
+    target_binary_class = 3
+
+    def one_vs_rest(dataset, target_class):
+        new_targets = []
+        for _, label in dataset:
+            new_label = float(1.0) if label == target_class else float(0.0)
+            new_targets.append(new_label)
+
+        dataset.targets = new_targets  # Replace the original labels with the binary ones
+        return dataset
+
+    binary_train_dataset = CIFAR10(root="data", train=True, download=True, transform=ToTensor())
+    binary_test_dataset = CIFAR10(root="data", train=False, download=True, transform=ToTensor())
+
+    # Apply one-vs-rest labeling
+    binary_train_dataset = one_vs_rest(binary_train_dataset, target_binary_class)
+    binary_test_dataset = one_vs_rest(binary_test_dataset, target_binary_class)
+
+    binary_trainloader = DataLoader(binary_train_dataset, batch_size=BATCH_SIZE, shuffle=True)
+    binary_testloader = DataLoader(binary_test_dataset, batch_size=BATCH_SIZE, shuffle=False)
+
+    binary_epoch = 1
+
+    binary_config = ClassifierConfig(model_name="microsoft/resnet-50", device=DEVICE)
+    binary_classifier = Classifier(binary_config)
+
+    class_counts = np.bincount(binary_train_dataset.targets)
+    n = len(binary_train_dataset)
+    w0 = n / (2.0 * class_counts[0])
+    w1 = n / (2.0 * class_counts[1])
+
+    binary_criterion = BCEWithLogitsLoss(pos_weight=torch.tensor(w1 / w0))
+    binary_optimizer = Adam(binary_classifier.parameters(), lr=LR)
+
+    binary_classifier.train()
+
+    logging.info("Start binary classifier training")
+
+    # Training loop
+    while binary_epoch < EPOCH_NUM:  # loop over the dataset multiple times
+        for i, data in enumerate(binary_trainloader):
+            inputs, labels = data
+            inputs, labels = inputs.to(DEVICE), labels.to(torch.float32).to(DEVICE)
+
+            # Zero the parameter gradients
+            binary_optimizer.zero_grad()
+
+            # Forward pass
+            outputs = binary_classifier(inputs)
+            loss = binary_criterion(outputs.logits, labels)
+            loss.backward()
+            binary_optimizer.step()
+
+            if i % 10 == 0:  # print every 10 mini-batches
+                print(f"[Epoch {binary_epoch}, Batch {i}] loss: {loss.item():.3f}")
+        binary_epoch += 1
+
+    logging.info("Binary classifier training finished")
+    logging.info("Start binary classifier evaluation")
+
+    binary_classifier.eval()
+
+    test_loss = 0.0
+    test_labels = []
+    test_pridections = []
+    test_probs = []
+
+    with torch.no_grad():
+        for data in binary_testloader:
+            images, labels = data
+            images, labels = images.to(DEVICE), labels.to(torch.float32).to(DEVICE)
+            outputs = binary_classifier(images)
+            loss = binary_criterion(outputs.logits, labels)
+            test_loss += loss.item() * BATCH_SIZE
+
+            test_labels.extend(labels.cpu())
+            test_pridections.extend(outputs.logits.cpu())
+            test_probs.extend(outputs.probabilities.cpu())
+
+    test_loss = test_loss / len(binary_test_dataset)
+
+    logging.info(f"Binary classifier test loss {test_loss:.3f}")
+
+    test_labels = torch.stack(test_labels)
+    test_predictions = torch.stack(test_pridections)
+    test_probs = torch.stack(test_probs)
+
+    # Calculate metrics
+    acc = Accuracy(task="binary")(test_predictions, test_labels)
+    prec = Precision(task="binary", average="weighted")(test_predictions, test_labels)
+    rec = Recall(task="binary", average="weighted")(test_predictions, test_labels)
+    f1_score = F1Score(task="binary", average="weighted")(test_predictions, test_labels)
+    auroc_score = AUROC(task="binary", average="weighted")(test_probs, test_labels)
+
+    logging.info(f"Accuracy: {acc:.2f}")
+    logging.info(f"Precision: {prec:.2f}")
+    logging.info(f"Recall: {rec:.2f}")
+    logging.info(f"F1 Score: {f1_score:.2f}")
+    logging.info(f"AUROC Score: {auroc_score:.2f}")
+
+
+if __name__ == "__main__":
+    train_evaluate_multiclass_classifier()
+    train_evaluate_binary_classifier()

tests/policies/hilserl/classifier/test_modelling_classifier.py

@@ -0,0 +1,85 @@
+import torch
+
+from lerobot.common.policies.hilserl.classifier.modeling_classifier import (
+    ClassifierConfig,
+    ClassifierOutput,
+)
+from tests.utils import require_package
+
+
+def test_classifier_output():
+    output = ClassifierOutput(
+        logits=torch.tensor([1, 2, 3]), probabilities=torch.tensor([0.1, 0.2, 0.3]), hidden_states=None
+    )
+
+    assert (
+        f"{output}"
+        == "ClassifierOutput(logits=tensor([1, 2, 3]), probabilities=tensor([0.1000, 0.2000, 0.3000]), hidden_states=None)"
+    )
+
+
+@require_package("transformers")
+def test_binary_classifier_with_default_params():
+    from lerobot.common.policies.hilserl.classifier.modeling_classifier import Classifier
+
+    config = ClassifierConfig()
+    classifier = Classifier(config)
+
+    batch_size = 10
+
+    input = torch.rand(batch_size, 3, 224, 224)
+    output = classifier(input)
+
+    assert output is not None
+    assert output.logits.shape == torch.Size([batch_size])
+    assert not torch.isnan(output.logits).any(), "Tensor contains NaN values"
+    assert output.probabilities.shape == torch.Size([batch_size])
+    assert not torch.isnan(output.probabilities).any(), "Tensor contains NaN values"
+    assert output.hidden_states.shape == torch.Size([batch_size, 2048])
+    assert not torch.isnan(output.hidden_states).any(), "Tensor contains NaN values"
+
+
+@require_package("transformers")
+def test_multiclass_classifier():
+    from lerobot.common.policies.hilserl.classifier.modeling_classifier import Classifier
+
+    num_classes = 5
+    config = ClassifierConfig(num_classes=num_classes)
+    classifier = Classifier(config)
+
+    batch_size = 10
+
+    input = torch.rand(batch_size, 3, 224, 224)
+    output = classifier(input)
+
+    assert output is not None
+    assert output.logits.shape == torch.Size([batch_size, num_classes])
+    assert not torch.isnan(output.logits).any(), "Tensor contains NaN values"
+    assert output.probabilities.shape == torch.Size([batch_size, num_classes])
+    assert not torch.isnan(output.probabilities).any(), "Tensor contains NaN values"
+    assert output.hidden_states.shape == torch.Size([batch_size, 2048])
+    assert not torch.isnan(output.hidden_states).any(), "Tensor contains NaN values"
+
+
+@require_package("transformers")
+def test_default_device():
+    from lerobot.common.policies.hilserl.classifier.modeling_classifier import Classifier
+
+    config = ClassifierConfig()
+    assert config.device == "cpu"
+
+    classifier = Classifier(config)
+    for p in classifier.parameters():
+        assert p.device == torch.device("cpu")
+
+
+@require_package("transformers")
+def test_explicit_device_setup():
+    from lerobot.common.policies.hilserl.classifier.modeling_classifier import Classifier
+
+    config = ClassifierConfig(device="meta")
+    assert config.device == "meta"
+
+    classifier = Classifier(config)
+    for p in classifier.parameters():
+        assert p.device == torch.device("meta")

tests/scripts/save_image_transforms_to_safetensors.py

@@ -76,7 +76,7 @@ def main():
     dataset = LeRobotDataset(DATASET_REPO_ID, image_transforms=None)
     output_dir = Path(ARTIFACT_DIR)
     output_dir.mkdir(parents=True, exist_ok=True)
-    original_frame = dataset[0][dataset.camera_keys[0]]
+    original_frame = dataset[0][dataset.meta.camera_keys[0]]
 
     save_single_transforms(original_frame, output_dir)
     save_default_config_transform(original_frame, output_dir)

tests/scripts/save_policy_to_safetensors.py

@@ -38,7 +38,7 @@ def get_policy_stats(env_name, policy_name, extra_overrides):
     )
     set_global_seed(1337)
     dataset = make_dataset(cfg)
-    policy = make_policy(cfg, dataset_stats=dataset.stats)
+    policy = make_policy(cfg, dataset_stats=dataset.meta.stats)
     policy.train()
     optimizer, _ = make_optimizer_and_scheduler(cfg, policy)
 

tests/test_control_robot.py

@@ -29,7 +29,6 @@ from unittest.mock import patch
 
 import pytest
 
-from lerobot.common.datasets.populate_dataset import add_frame, init_dataset
 from lerobot.common.logger import Logger
 from lerobot.common.policies.factory import make_policy
 from lerobot.common.utils.utils import init_hydra_config
@@ -93,8 +92,9 @@ def test_record_without_cameras(tmpdir, request, robot_type, mock):
         mock_calibration_dir(calibration_dir)
         overrides.append(f"calibration_dir={calibration_dir}")
 
-    root = Path(tmpdir) / "data"
     repo_id = "lerobot/debug"
+    root = Path(tmpdir) / "data" / repo_id
+    single_task = "Do something."
 
     robot = make_robot(robot_type, overrides=overrides, mock=mock)
     record(
@@ -102,6 +102,7 @@ def test_record_without_cameras(tmpdir, request, robot_type, mock):
         fps=30,
         root=root,
         repo_id=repo_id,
+        single_task=single_task,
         warmup_time_s=1,
         episode_time_s=1,
         num_episodes=2,
@@ -132,17 +133,18 @@ def test_record_and_replay_and_policy(tmpdir, request, robot_type, mock):
     env_name = "koch_real"
     policy_name = "act_koch_real"
 
-    root = tmpdir / "data"
     repo_id = "lerobot/debug"
-    eval_repo_id = "lerobot/eval_debug"
+    root = tmpdir / "data" / repo_id
+    single_task = "Do something."
 
     robot = make_robot(robot_type, overrides=overrides, mock=mock)
     dataset = record(
         robot,
         root,
         repo_id,
+        single_task,
         fps=1,
-        warmup_time_s=1,
+        warmup_time_s=0.5,
         episode_time_s=1,
         reset_time_s=1,
         num_episodes=2,
@@ -153,10 +155,10 @@ def test_record_and_replay_and_policy(tmpdir, request, robot_type, mock):
         display_cameras=False,
         play_sounds=False,
     )
-    assert dataset.num_episodes == 2
+    assert dataset.meta.total_episodes == 2
     assert len(dataset) == 2
 
-    replay(robot, episode=0, fps=1, root=root, repo_id=repo_id, play_sounds=False)
+    replay(robot, episode=0, fps=1, root=root, repo_id=repo_id, play_sounds=False, local_files_only=True)
 
     # TODO(rcadene, aliberts): rethink this design
     if robot_type == "aloha":
@@ -191,7 +193,7 @@ def test_record_and_replay_and_policy(tmpdir, request, robot_type, mock):
         overrides=overrides,
     )
 
-    policy = make_policy(hydra_cfg=cfg, dataset_stats=dataset.stats)
+    policy = make_policy(hydra_cfg=cfg, dataset_stats=dataset.meta.stats)
     optimizer, lr_scheduler = make_optimizer_and_scheduler(cfg, policy)
     out_dir = tmpdir / "logger"
     logger = Logger(cfg, out_dir, wandb_job_name="debug")
@@ -225,10 +227,14 @@ def test_record_and_replay_and_policy(tmpdir, request, robot_type, mock):
     else:
         num_image_writer_processes = 0
 
-    record(
+    eval_repo_id = "lerobot/eval_debug"
+    eval_root = tmpdir / "data" / eval_repo_id
+
+    dataset = record(
         robot,
-        root,
+        eval_root,
         eval_repo_id,
+        single_task,
         pretrained_policy_name_or_path,
         warmup_time_s=1,
         episode_time_s=1,
@@ -265,51 +271,36 @@ def test_resume_record(tmpdir, request, robot_type, mock):
 
     robot = make_robot(robot_type, overrides=overrides, mock=mock)
 
-    root = Path(tmpdir) / "data"
     repo_id = "lerobot/debug"
+    root = Path(tmpdir) / "data" / repo_id
+    single_task = "Do something."
 
-    dataset = record(
-        robot,
-        root,
-        repo_id,
-        fps=1,
-        warmup_time_s=0,
-        episode_time_s=1,
-        num_episodes=1,
-        push_to_hub=False,
-        video=False,
-        display_cameras=False,
-        play_sounds=False,
-        run_compute_stats=False,
-    )
-    assert len(dataset) == 1, "`dataset` should contain only 1 frame"
+    record_kwargs = {
+        "robot": robot,
+        "root": root,
+        "repo_id": repo_id,
+        "single_task": single_task,
+        "fps": 1,
+        "warmup_time_s": 0,
+        "episode_time_s": 1,
+        "push_to_hub": False,
+        "video": False,
+        "display_cameras": False,
+        "play_sounds": False,
+        "run_compute_stats": False,
+        "local_files_only": True,
+        "num_episodes": 1,
+    }
 
-    init_dataset_return_value = {}
+    dataset = record(**record_kwargs)
+    assert len(dataset) == 1, f"`dataset` should contain 1 frame, not {len(dataset)}"
 
-    def wrapped_init_dataset(*args, **kwargs):
-        nonlocal init_dataset_return_value
-        init_dataset_return_value = init_dataset(*args, **kwargs)
-        return init_dataset_return_value
+    with pytest.raises(FileExistsError):
+        # Dataset already exists, but resume=False by default
+        record(**record_kwargs)
 
-    with patch("lerobot.scripts.control_robot.init_dataset", wraps=wrapped_init_dataset):
-        dataset = record(
-            robot,
-            root,
-            repo_id,
-            fps=1,
-            warmup_time_s=0,
-            episode_time_s=1,
-            num_episodes=2,
-            push_to_hub=False,
-            video=False,
-            display_cameras=False,
-            play_sounds=False,
-            run_compute_stats=False,
-        )
-        assert len(dataset) == 2, "`dataset` should contain only 1 frame"
-        assert (
-            init_dataset_return_value["num_episodes"] == 2
-        ), "`init_dataset` should load the previous episode"
+    dataset = record(**record_kwargs, resume=True)
+    assert len(dataset) == 2, f"`dataset` should contain 2 frames, not {len(dataset)}"
 
 
 @pytest.mark.parametrize("robot_type, mock", [("koch", True)])
@@ -328,23 +319,22 @@ def test_record_with_event_rerecord_episode(tmpdir, request, robot_type, mock):
         overrides = []
 
     robot = make_robot(robot_type, overrides=overrides, mock=mock)
-    with (
-        patch("lerobot.scripts.control_robot.init_keyboard_listener") as mock_listener,
-        patch("lerobot.common.robot_devices.control_utils.add_frame", wraps=add_frame) as mock_add_frame,
-    ):
+    with patch("lerobot.scripts.control_robot.init_keyboard_listener") as mock_listener:
         mock_events = {}
         mock_events["exit_early"] = True
         mock_events["rerecord_episode"] = True
         mock_events["stop_recording"] = False
         mock_listener.return_value = (None, mock_events)
 
-        root = Path(tmpdir) / "data"
         repo_id = "lerobot/debug"
+        root = Path(tmpdir) / "data" / repo_id
+        single_task = "Do something."
 
         dataset = record(
             robot,
             root,
             repo_id,
+            single_task,
             fps=1,
             warmup_time_s=0,
             episode_time_s=1,
@@ -358,7 +348,6 @@ def test_record_with_event_rerecord_episode(tmpdir, request, robot_type, mock):
 
         assert not mock_events["rerecord_episode"], "`rerecord_episode` wasn't properly reset to False"
         assert not mock_events["exit_early"], "`exit_early` wasn't properly reset to False"
-        assert mock_add_frame.call_count == 2, "`add_frame` should have been called 2 times"
         assert len(dataset) == 1, "`dataset` should contain only 1 frame"
 
 
@@ -378,23 +367,22 @@ def test_record_with_event_exit_early(tmpdir, request, robot_type, mock):
         overrides = []
 
     robot = make_robot(robot_type, overrides=overrides, mock=mock)
-    with (
-        patch("lerobot.scripts.control_robot.init_keyboard_listener") as mock_listener,
-        patch("lerobot.common.robot_devices.control_utils.add_frame", wraps=add_frame) as mock_add_frame,
-    ):
+    with patch("lerobot.scripts.control_robot.init_keyboard_listener") as mock_listener:
         mock_events = {}
         mock_events["exit_early"] = True
         mock_events["rerecord_episode"] = False
         mock_events["stop_recording"] = False
         mock_listener.return_value = (None, mock_events)
 
-        root = Path(tmpdir) / "data"
         repo_id = "lerobot/debug"
+        root = Path(tmpdir) / "data" / repo_id
+        single_task = "Do something."
 
         dataset = record(
             robot,
             fps=2,
             root=root,
+            single_task=single_task,
             repo_id=repo_id,
             warmup_time_s=0,
             episode_time_s=1,
@@ -407,7 +395,6 @@ def test_record_with_event_exit_early(tmpdir, request, robot_type, mock):
         )
 
         assert not mock_events["exit_early"], "`exit_early` wasn't properly reset to False"
-        assert mock_add_frame.call_count == 1, "`add_frame` should have been called 1 time"
         assert len(dataset) == 1, "`dataset` should contain only 1 frame"
 
 
@@ -429,23 +416,22 @@ def test_record_with_event_stop_recording(tmpdir, request, robot_type, mock, num
         overrides = []
 
     robot = make_robot(robot_type, overrides=overrides, mock=mock)
-    with (
-        patch("lerobot.scripts.control_robot.init_keyboard_listener") as mock_listener,
-        patch("lerobot.common.robot_devices.control_utils.add_frame", wraps=add_frame) as mock_add_frame,
-    ):
+    with patch("lerobot.scripts.control_robot.init_keyboard_listener") as mock_listener:
         mock_events = {}
         mock_events["exit_early"] = True
         mock_events["rerecord_episode"] = False
         mock_events["stop_recording"] = True
         mock_listener.return_value = (None, mock_events)
 
-        root = Path(tmpdir) / "data"
         repo_id = "lerobot/debug"
+        root = Path(tmpdir) / "data" / repo_id
+        single_task = "Do something."
 
         dataset = record(
             robot,
             root,
             repo_id,
+            single_task=single_task,
             fps=1,
             warmup_time_s=0,
             episode_time_s=1,
@@ -459,5 +445,4 @@ def test_record_with_event_stop_recording(tmpdir, request, robot_type, mock, num
         )
 
         assert not mock_events["exit_early"], "`exit_early` wasn't properly reset to False"
-        assert mock_add_frame.call_count == 1, "`add_frame` should have been called 1 time"
         assert len(dataset) == 1, "`dataset` should contain only 1 frame"

tests/test_datasets.py

@@ -33,18 +33,72 @@ from lerobot.common.datasets.compute_stats import (
     get_stats_einops_patterns,
 )
 from lerobot.common.datasets.factory import make_dataset
-from lerobot.common.datasets.lerobot_dataset import LeRobotDataset, MultiLeRobotDataset
+from lerobot.common.datasets.lerobot_dataset import (
+    LeRobotDataset,
+    MultiLeRobotDataset,
+)
 from lerobot.common.datasets.utils import (
     create_branch,
     flatten_dict,
     hf_transform_to_torch,
-    load_previous_and_future_frames,
     unflatten_dict,
 )
 from lerobot.common.utils.utils import init_hydra_config, seeded_context
-from tests.utils import DEFAULT_CONFIG_PATH, DEVICE
+from tests.fixtures.constants import DUMMY_REPO_ID
+from tests.utils import DEFAULT_CONFIG_PATH, DEVICE, make_robot
 
 
+def test_same_attributes_defined(lerobot_dataset_factory, tmp_path):
+    """
+    Instantiate a LeRobotDataset both ways with '__init__()' and 'create()' and verify that instantiated
+    objects have the same sets of attributes defined.
+    """
+    # Instantiate both ways
+    robot = make_robot("koch", mock=True)
+    root_create = tmp_path / "create"
+    dataset_create = LeRobotDataset.create(repo_id=DUMMY_REPO_ID, fps=30, robot=robot, root=root_create)
+
+    root_init = tmp_path / "init"
+    dataset_init = lerobot_dataset_factory(root=root_init)
+
+    # Access the '_hub_version' cached_property in both instances to force its creation
+    _ = dataset_init.meta._hub_version
+    _ = dataset_create.meta._hub_version
+
+    init_attr = set(vars(dataset_init).keys())
+    create_attr = set(vars(dataset_create).keys())
+
+    assert init_attr == create_attr
+
+
+def test_dataset_initialization(lerobot_dataset_factory, tmp_path):
+    kwargs = {
+        "repo_id": DUMMY_REPO_ID,
+        "total_episodes": 10,
+        "total_frames": 400,
+        "episodes": [2, 5, 6],
+    }
+    dataset = lerobot_dataset_factory(root=tmp_path, **kwargs)
+
+    assert dataset.repo_id == kwargs["repo_id"]
+    assert dataset.meta.total_episodes == kwargs["total_episodes"]
+    assert dataset.meta.total_frames == kwargs["total_frames"]
+    assert dataset.episodes == kwargs["episodes"]
+    assert dataset.num_episodes == len(kwargs["episodes"])
+    assert dataset.num_frames == len(dataset)
+
+
+# TODO(aliberts):
+# - [ ] test various attributes & state from init and create
+# - [ ] test init with episodes and check num_frames
+# - [ ] test add_frame
+# - [ ] test add_episode
+# - [ ] test consolidate
+# - [ ] test push_to_hub
+# - [ ] test smaller methods
+
+
+@pytest.mark.skip("TODO after v2 migration / removing hydra")
 @pytest.mark.parametrize(
     "env_name, repo_id, policy_name",
     lerobot.env_dataset_policy_triplets
@@ -67,7 +121,7 @@ def test_factory(env_name, repo_id, policy_name):
     )
     dataset = make_dataset(cfg)
     delta_timestamps = dataset.delta_timestamps
-    camera_keys = dataset.camera_keys
+    camera_keys = dataset.meta.camera_keys
 
     item = dataset[0]
 
@@ -117,6 +171,7 @@ def test_factory(env_name, repo_id, policy_name):
 
 
 # TODO(alexander-soare): If you're hunting for savings on testing time, this takes about 5 seconds.
+@pytest.mark.skip("TODO after v2 migration / removing hydra")
 def test_multilerobotdataset_frames():
     """Check that all dataset frames are incorporated."""
     # Note: use the image variants of the dataset to make the test approx 3x faster.
@@ -130,7 +185,7 @@ def test_multilerobotdataset_frames():
     sub_datasets = [LeRobotDataset(repo_id) for repo_id in repo_ids]
     dataset = MultiLeRobotDataset(repo_ids)
     assert len(dataset) == sum(len(d) for d in sub_datasets)
-    assert dataset.num_samples == sum(d.num_samples for d in sub_datasets)
+    assert dataset.num_frames == sum(d.num_frames for d in sub_datasets)
     assert dataset.num_episodes == sum(d.num_episodes for d in sub_datasets)
 
     # Run through all items of the LeRobotDatasets in parallel with the items of the MultiLerobotDataset and
@@ -149,6 +204,8 @@ def test_multilerobotdataset_frames():
             assert torch.equal(sub_dataset_item[k], dataset_item[k])
 
 
+# TODO(aliberts, rcadene): Refactor and move this to a tests/test_compute_stats.py
+@pytest.mark.skip("TODO after v2 migration / removing hydra")
 def test_compute_stats_on_xarm():
     """Check that the statistics are computed correctly according to the stats_patterns property.
 
@@ -197,7 +254,7 @@ def test_compute_stats_on_xarm():
         assert torch.allclose(computed_stats[k]["max"], expected_stats[k]["max"])
 
     # load stats used during training which are expected to match the ones returned by computed_stats
-    loaded_stats = dataset.stats  # noqa: F841
+    loaded_stats = dataset.meta.stats  # noqa: F841
 
     # TODO(rcadene): we can't test this because expected_stats is computed on a subset
     # # test loaded stats match expected stats
@@ -208,72 +265,7 @@ def test_compute_stats_on_xarm():
     #     assert torch.allclose(loaded_stats[k]["max"], expected_stats[k]["max"])
 
 
-def test_load_previous_and_future_frames_within_tolerance():
-    hf_dataset = Dataset.from_dict(
-        {
-            "timestamp": [0.1, 0.2, 0.3, 0.4, 0.5],
-            "index": [0, 1, 2, 3, 4],
-            "episode_index": [0, 0, 0, 0, 0],
-        }
-    )
-    hf_dataset.set_transform(hf_transform_to_torch)
-    episode_data_index = {
-        "from": torch.tensor([0]),
-        "to": torch.tensor([5]),
-    }
-    delta_timestamps = {"index": [-0.2, 0, 0.139]}
-    tol = 0.04
-    item = hf_dataset[2]
-    item = load_previous_and_future_frames(item, hf_dataset, episode_data_index, delta_timestamps, tol)
-    data, is_pad = item["index"], item["index_is_pad"]
-    assert torch.equal(data, torch.tensor([0, 2, 3])), "Data does not match expected values"
-    assert not is_pad.any(), "Unexpected padding detected"
-
-
-def test_load_previous_and_future_frames_outside_tolerance_inside_episode_range():
-    hf_dataset = Dataset.from_dict(
-        {
-            "timestamp": [0.1, 0.2, 0.3, 0.4, 0.5],
-            "index": [0, 1, 2, 3, 4],
-            "episode_index": [0, 0, 0, 0, 0],
-        }
-    )
-    hf_dataset.set_transform(hf_transform_to_torch)
-    episode_data_index = {
-        "from": torch.tensor([0]),
-        "to": torch.tensor([5]),
-    }
-    delta_timestamps = {"index": [-0.2, 0, 0.141]}
-    tol = 0.04
-    item = hf_dataset[2]
-    with pytest.raises(AssertionError):
-        load_previous_and_future_frames(item, hf_dataset, episode_data_index, delta_timestamps, tol)
-
-
-def test_load_previous_and_future_frames_outside_tolerance_outside_episode_range():
-    hf_dataset = Dataset.from_dict(
-        {
-            "timestamp": [0.1, 0.2, 0.3, 0.4, 0.5],
-            "index": [0, 1, 2, 3, 4],
-            "episode_index": [0, 0, 0, 0, 0],
-        }
-    )
-    hf_dataset.set_transform(hf_transform_to_torch)
-    episode_data_index = {
-        "from": torch.tensor([0]),
-        "to": torch.tensor([5]),
-    }
-    delta_timestamps = {"index": [-0.3, -0.24, 0, 0.26, 0.3]}
-    tol = 0.04
-    item = hf_dataset[2]
-    item = load_previous_and_future_frames(item, hf_dataset, episode_data_index, delta_timestamps, tol)
-    data, is_pad = item["index"], item["index_is_pad"]
-    assert torch.equal(data, torch.tensor([0, 0, 2, 4, 4])), "Data does not match expected values"
-    assert torch.equal(
-        is_pad, torch.tensor([True, False, False, True, True])
-    ), "Padding does not match expected values"
-
-
+# TODO(aliberts): Move to more appropriate location
 def test_flatten_unflatten_dict():
     d = {
         "obs": {
@@ -297,6 +289,7 @@ def test_flatten_unflatten_dict():
     assert json.dumps(original_d, sort_keys=True) == json.dumps(d, sort_keys=True), f"{original_d} != {d}"
 
 
+@pytest.mark.skip("TODO after v2 migration / removing hydra")
 @pytest.mark.parametrize(
     "repo_id",
     [
@@ -368,6 +361,7 @@ def test_backward_compatibility(repo_id):
     # load_and_compare(i - 1)
 
 
+@pytest.mark.skip("TODO after v2 migration / removing hydra")
 def test_aggregate_stats():
     """Makes 3 basic datasets and checks that aggregate stats are computed correctly."""
     with seeded_context(0):

tests/test_delta_timestamps.py

@@ -0,0 +1,256 @@
+import pytest
+import torch
+from datasets import Dataset
+
+from lerobot.common.datasets.utils import (
+    calculate_episode_data_index,
+    check_delta_timestamps,
+    check_timestamps_sync,
+    get_delta_indices,
+    hf_transform_to_torch,
+)
+from tests.fixtures.constants import DUMMY_MOTOR_FEATURES
+
+
+@pytest.fixture(scope="module")
+def synced_hf_dataset_factory(hf_dataset_factory):
+    def _create_synced_hf_dataset(fps: int = 30) -> Dataset:
+        return hf_dataset_factory(fps=fps)
+
+    return _create_synced_hf_dataset
+
+
+@pytest.fixture(scope="module")
+def unsynced_hf_dataset_factory(synced_hf_dataset_factory):
+    def _create_unsynced_hf_dataset(fps: int = 30, tolerance_s: float = 1e-4) -> Dataset:
+        hf_dataset = synced_hf_dataset_factory(fps=fps)
+        features = hf_dataset.features
+        df = hf_dataset.to_pandas()
+        dtype = df["timestamp"].dtype  # This is to avoid pandas type warning
+        # Modify a single timestamp just outside tolerance
+        df.at[30, "timestamp"] = dtype.type(df.at[30, "timestamp"] + (tolerance_s * 1.1))
+        unsynced_hf_dataset = Dataset.from_pandas(df, features=features)
+        unsynced_hf_dataset.set_transform(hf_transform_to_torch)
+        return unsynced_hf_dataset
+
+    return _create_unsynced_hf_dataset
+
+
+@pytest.fixture(scope="module")
+def slightly_off_hf_dataset_factory(synced_hf_dataset_factory):
+    def _create_slightly_off_hf_dataset(fps: int = 30, tolerance_s: float = 1e-4) -> Dataset:
+        hf_dataset = synced_hf_dataset_factory(fps=fps)
+        features = hf_dataset.features
+        df = hf_dataset.to_pandas()
+        dtype = df["timestamp"].dtype  # This is to avoid pandas type warning
+        # Modify a single timestamp just inside tolerance
+        df.at[30, "timestamp"] = dtype.type(df.at[30, "timestamp"] + (tolerance_s * 0.9))
+        unsynced_hf_dataset = Dataset.from_pandas(df, features=features)
+        unsynced_hf_dataset.set_transform(hf_transform_to_torch)
+        return unsynced_hf_dataset
+
+    return _create_slightly_off_hf_dataset
+
+
+@pytest.fixture(scope="module")
+def valid_delta_timestamps_factory():
+    def _create_valid_delta_timestamps(fps: int = 30, keys: list = DUMMY_MOTOR_FEATURES) -> dict:
+        delta_timestamps = {key: [i * (1 / fps) for i in range(-10, 10)] for key in keys}
+        return delta_timestamps
+
+    return _create_valid_delta_timestamps
+
+
+@pytest.fixture(scope="module")
+def invalid_delta_timestamps_factory(valid_delta_timestamps_factory):
+    def _create_invalid_delta_timestamps(
+        fps: int = 30, tolerance_s: float = 1e-4, keys: list = DUMMY_MOTOR_FEATURES
+    ) -> dict:
+        delta_timestamps = valid_delta_timestamps_factory(fps, keys)
+        # Modify a single timestamp just outside tolerance
+        for key in keys:
+            delta_timestamps[key][3] += tolerance_s * 1.1
+        return delta_timestamps
+
+    return _create_invalid_delta_timestamps
+
+
+@pytest.fixture(scope="module")
+def slightly_off_delta_timestamps_factory(valid_delta_timestamps_factory):
+    def _create_slightly_off_delta_timestamps(
+        fps: int = 30, tolerance_s: float = 1e-4, keys: list = DUMMY_MOTOR_FEATURES
+    ) -> dict:
+        delta_timestamps = valid_delta_timestamps_factory(fps, keys)
+        # Modify a single timestamp just inside tolerance
+        for key in delta_timestamps:
+            delta_timestamps[key][3] += tolerance_s * 0.9
+            delta_timestamps[key][-3] += tolerance_s * 0.9
+        return delta_timestamps
+
+    return _create_slightly_off_delta_timestamps
+
+
+@pytest.fixture(scope="module")
+def delta_indices(keys: list = DUMMY_MOTOR_FEATURES) -> dict:
+    return {key: list(range(-10, 10)) for key in keys}
+
+
+def test_check_timestamps_sync_synced(synced_hf_dataset_factory):
+    fps = 30
+    tolerance_s = 1e-4
+    synced_hf_dataset = synced_hf_dataset_factory(fps)
+    episode_data_index = calculate_episode_data_index(synced_hf_dataset)
+    result = check_timestamps_sync(
+        hf_dataset=synced_hf_dataset,
+        episode_data_index=episode_data_index,
+        fps=fps,
+        tolerance_s=tolerance_s,
+    )
+    assert result is True
+
+
+def test_check_timestamps_sync_unsynced(unsynced_hf_dataset_factory):
+    fps = 30
+    tolerance_s = 1e-4
+    unsynced_hf_dataset = unsynced_hf_dataset_factory(fps, tolerance_s)
+    episode_data_index = calculate_episode_data_index(unsynced_hf_dataset)
+    with pytest.raises(ValueError):
+        check_timestamps_sync(
+            hf_dataset=unsynced_hf_dataset,
+            episode_data_index=episode_data_index,
+            fps=fps,
+            tolerance_s=tolerance_s,
+        )
+
+
+def test_check_timestamps_sync_unsynced_no_exception(unsynced_hf_dataset_factory):
+    fps = 30
+    tolerance_s = 1e-4
+    unsynced_hf_dataset = unsynced_hf_dataset_factory(fps, tolerance_s)
+    episode_data_index = calculate_episode_data_index(unsynced_hf_dataset)
+    result = check_timestamps_sync(
+        hf_dataset=unsynced_hf_dataset,
+        episode_data_index=episode_data_index,
+        fps=fps,
+        tolerance_s=tolerance_s,
+        raise_value_error=False,
+    )
+    assert result is False
+
+
+def test_check_timestamps_sync_slightly_off(slightly_off_hf_dataset_factory):
+    fps = 30
+    tolerance_s = 1e-4
+    slightly_off_hf_dataset = slightly_off_hf_dataset_factory(fps, tolerance_s)
+    episode_data_index = calculate_episode_data_index(slightly_off_hf_dataset)
+    result = check_timestamps_sync(
+        hf_dataset=slightly_off_hf_dataset,
+        episode_data_index=episode_data_index,
+        fps=fps,
+        tolerance_s=tolerance_s,
+    )
+    assert result is True
+
+
+def test_check_timestamps_sync_single_timestamp():
+    single_timestamp_hf_dataset = Dataset.from_dict({"timestamp": [0.0], "episode_index": [0]})
+    single_timestamp_hf_dataset.set_transform(hf_transform_to_torch)
+    episode_data_index = {"to": torch.tensor([1]), "from": torch.tensor([0])}
+    fps = 30
+    tolerance_s = 1e-4
+    result = check_timestamps_sync(
+        hf_dataset=single_timestamp_hf_dataset,
+        episode_data_index=episode_data_index,
+        fps=fps,
+        tolerance_s=tolerance_s,
+    )
+    assert result is True
+
+
+# TODO(aliberts): Change behavior of hf_transform_to_torch so that it can work with empty dataset
+@pytest.mark.skip("TODO: fix")
+def test_check_timestamps_sync_empty_dataset():
+    fps = 30
+    tolerance_s = 1e-4
+    empty_hf_dataset = Dataset.from_dict({"timestamp": [], "episode_index": []})
+    empty_hf_dataset.set_transform(hf_transform_to_torch)
+    episode_data_index = {
+        "to": torch.tensor([], dtype=torch.int64),
+        "from": torch.tensor([], dtype=torch.int64),
+    }
+    result = check_timestamps_sync(
+        hf_dataset=empty_hf_dataset,
+        episode_data_index=episode_data_index,
+        fps=fps,
+        tolerance_s=tolerance_s,
+    )
+    assert result is True
+
+
+def test_check_delta_timestamps_valid(valid_delta_timestamps_factory):
+    fps = 30
+    tolerance_s = 1e-4
+    valid_delta_timestamps = valid_delta_timestamps_factory(fps)
+    result = check_delta_timestamps(
+        delta_timestamps=valid_delta_timestamps,
+        fps=fps,
+        tolerance_s=tolerance_s,
+    )
+    assert result is True
+
+
+def test_check_delta_timestamps_slightly_off(slightly_off_delta_timestamps_factory):
+    fps = 30
+    tolerance_s = 1e-4
+    slightly_off_delta_timestamps = slightly_off_delta_timestamps_factory(fps, tolerance_s)
+    result = check_delta_timestamps(
+        delta_timestamps=slightly_off_delta_timestamps,
+        fps=fps,
+        tolerance_s=tolerance_s,
+    )
+    assert result is True
+
+
+def test_check_delta_timestamps_invalid(invalid_delta_timestamps_factory):
+    fps = 30
+    tolerance_s = 1e-4
+    invalid_delta_timestamps = invalid_delta_timestamps_factory(fps, tolerance_s)
+    with pytest.raises(ValueError):
+        check_delta_timestamps(
+            delta_timestamps=invalid_delta_timestamps,
+            fps=fps,
+            tolerance_s=tolerance_s,
+        )
+
+
+def test_check_delta_timestamps_invalid_no_exception(invalid_delta_timestamps_factory):
+    fps = 30
+    tolerance_s = 1e-4
+    invalid_delta_timestamps = invalid_delta_timestamps_factory(fps, tolerance_s)
+    result = check_delta_timestamps(
+        delta_timestamps=invalid_delta_timestamps,
+        fps=fps,
+        tolerance_s=tolerance_s,
+        raise_value_error=False,
+    )
+    assert result is False
+
+
+def test_check_delta_timestamps_empty():
+    delta_timestamps = {}
+    fps = 30
+    tolerance_s = 1e-4
+    result = check_delta_timestamps(
+        delta_timestamps=delta_timestamps,
+        fps=fps,
+        tolerance_s=tolerance_s,
+    )
+    assert result is True
+
+
+def test_delta_indices(valid_delta_timestamps_factory, delta_indices):
+    fps = 30
+    delta_timestamps = valid_delta_timestamps_factory(fps)
+    expected_delta_indices = delta_indices
+    actual_delta_indices = get_delta_indices(delta_timestamps, fps)
+    assert expected_delta_indices == actual_delta_indices

tests/test_examples.py

@@ -13,12 +13,15 @@
 # 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.
-# TODO(aliberts): Mute logging for these tests
+
 import io
 import subprocess
 import sys
 from pathlib import Path
 
+import pytest
+
+from tests.fixtures.constants import DUMMY_REPO_ID
 from tests.utils import require_package
 
 
@@ -29,6 +32,7 @@ def _find_and_replace(text: str, finds_and_replaces: list[tuple[str, str]]) -> s
     return text
 
 
+# TODO(aliberts): Remove usage of subprocess calls and patch code with fixtures
 def _run_script(path):
     subprocess.run([sys.executable, path], check=True)
 
@@ -38,12 +42,26 @@ def _read_file(path):
         return file.read()
 
 
-def test_example_1():
+@pytest.mark.skip("TODO Fix and remove subprocess / excec calls")
+def test_example_1(tmp_path, lerobot_dataset_factory):
+    _ = lerobot_dataset_factory(root=tmp_path, repo_id=DUMMY_REPO_ID)
     path = "examples/1_load_lerobot_dataset.py"
-    _run_script(path)
+    file_contents = _read_file(path)
+    file_contents = _find_and_replace(
+        file_contents,
+        [
+            ('repo_id = "lerobot/pusht"', f'repo_id = "{DUMMY_REPO_ID}"'),
+            (
+                "LeRobotDataset(repo_id",
+                f"LeRobotDataset(repo_id, root='{str(tmp_path)}', local_files_only=True",
+            ),
+        ],
+    )
+    exec(file_contents, {})
     assert Path("outputs/examples/1_load_lerobot_dataset/episode_0.mp4").exists()
 
 
+@pytest.mark.skip("TODO Fix and remove subprocess / excec calls")
 @require_package("gym_pusht")
 def test_examples_basic2_basic3_advanced1():
     """
@@ -111,7 +129,8 @@ def test_examples_basic2_basic3_advanced1():
                 '# pretrained_policy_path = Path("outputs/train/example_pusht_diffusion")',
                 'pretrained_policy_path = Path("outputs/train/example_pusht_diffusion")',
             ),
-            ('split=f"train[{first_val_frame_index}:]"', 'split="train[30:]"'),
+            ("train_episodes = episodes[:num_train_episodes]", "train_episodes = [0]"),
+            ("val_episodes = episodes[num_train_episodes:]", "val_episodes = [1]"),
             ("num_workers=4", "num_workers=0"),
             ('device = torch.device("cuda")', 'device = torch.device("cpu")'),
             ("batch_size=64", "batch_size=1"),

tests/test_image_transforms.py

@@ -15,15 +15,12 @@
 # limitations under the License.
 from pathlib import Path
 
-import numpy as np
 import pytest
 import torch
-from PIL import Image
 from safetensors.torch import load_file
 from torchvision.transforms import v2
 from torchvision.transforms.v2 import functional as F  # noqa: N812
 
-from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.common.datasets.transforms import RandomSubsetApply, SharpnessJitter, get_image_transforms
 from lerobot.common.utils.utils import init_hydra_config, seeded_context
 from lerobot.scripts.visualize_image_transforms import visualize_transforms
@@ -33,21 +30,6 @@ ARTIFACT_DIR = Path("tests/data/save_image_transforms_to_safetensors")
 DATASET_REPO_ID = "lerobot/aloha_mobile_shrimp"
 
 
-def load_png_to_tensor(path: Path):
-    return torch.from_numpy(np.array(Image.open(path).convert("RGB"))).permute(2, 0, 1)
-
-
-@pytest.fixture
-def img():
-    dataset = LeRobotDataset(DATASET_REPO_ID)
-    return dataset[0][dataset.camera_keys[0]]
-
-
-@pytest.fixture
-def img_random():
-    return torch.rand(3, 480, 640)
-
-
 @pytest.fixture
 def color_jitters():
     return [
@@ -67,47 +49,54 @@ def default_transforms():
     return load_file(ARTIFACT_DIR / "default_transforms.safetensors")
 
 
-def test_get_image_transforms_no_transform(img):
+def test_get_image_transforms_no_transform(img_tensor_factory):
+    img_tensor = img_tensor_factory()
     tf_actual = get_image_transforms(brightness_min_max=(0.5, 0.5), max_num_transforms=0)
-    torch.testing.assert_close(tf_actual(img), img)
+    torch.testing.assert_close(tf_actual(img_tensor), img_tensor)
 
 
 @pytest.mark.parametrize("min_max", [(0.5, 0.5), (2.0, 2.0)])
-def test_get_image_transforms_brightness(img, min_max):
+def test_get_image_transforms_brightness(img_tensor_factory, min_max):
+    img_tensor = img_tensor_factory()
     tf_actual = get_image_transforms(brightness_weight=1.0, brightness_min_max=min_max)
     tf_expected = v2.ColorJitter(brightness=min_max)
-    torch.testing.assert_close(tf_actual(img), tf_expected(img))
+    torch.testing.assert_close(tf_actual(img_tensor), tf_expected(img_tensor))
 
 
 @pytest.mark.parametrize("min_max", [(0.5, 0.5), (2.0, 2.0)])
-def test_get_image_transforms_contrast(img, min_max):
+def test_get_image_transforms_contrast(img_tensor_factory, min_max):
+    img_tensor = img_tensor_factory()
     tf_actual = get_image_transforms(contrast_weight=1.0, contrast_min_max=min_max)
     tf_expected = v2.ColorJitter(contrast=min_max)
-    torch.testing.assert_close(tf_actual(img), tf_expected(img))
+    torch.testing.assert_close(tf_actual(img_tensor), tf_expected(img_tensor))
 
 
 @pytest.mark.parametrize("min_max", [(0.5, 0.5), (2.0, 2.0)])
-def test_get_image_transforms_saturation(img, min_max):
+def test_get_image_transforms_saturation(img_tensor_factory, min_max):
+    img_tensor = img_tensor_factory()
     tf_actual = get_image_transforms(saturation_weight=1.0, saturation_min_max=min_max)
     tf_expected = v2.ColorJitter(saturation=min_max)
-    torch.testing.assert_close(tf_actual(img), tf_expected(img))
+    torch.testing.assert_close(tf_actual(img_tensor), tf_expected(img_tensor))
 
 
 @pytest.mark.parametrize("min_max", [(-0.25, -0.25), (0.25, 0.25)])
-def test_get_image_transforms_hue(img, min_max):
+def test_get_image_transforms_hue(img_tensor_factory, min_max):
+    img_tensor = img_tensor_factory()
     tf_actual = get_image_transforms(hue_weight=1.0, hue_min_max=min_max)
     tf_expected = v2.ColorJitter(hue=min_max)
-    torch.testing.assert_close(tf_actual(img), tf_expected(img))
+    torch.testing.assert_close(tf_actual(img_tensor), tf_expected(img_tensor))
 
 
 @pytest.mark.parametrize("min_max", [(0.5, 0.5), (2.0, 2.0)])
-def test_get_image_transforms_sharpness(img, min_max):
+def test_get_image_transforms_sharpness(img_tensor_factory, min_max):
+    img_tensor = img_tensor_factory()
     tf_actual = get_image_transforms(sharpness_weight=1.0, sharpness_min_max=min_max)
     tf_expected = SharpnessJitter(sharpness=min_max)
-    torch.testing.assert_close(tf_actual(img), tf_expected(img))
+    torch.testing.assert_close(tf_actual(img_tensor), tf_expected(img_tensor))
 
 
-def test_get_image_transforms_max_num_transforms(img):
+def test_get_image_transforms_max_num_transforms(img_tensor_factory):
+    img_tensor = img_tensor_factory()
     tf_actual = get_image_transforms(
         brightness_min_max=(0.5, 0.5),
         contrast_min_max=(0.5, 0.5),
@@ -125,12 +114,13 @@ def test_get_image_transforms_max_num_transforms(img):
             SharpnessJitter(sharpness=(0.5, 0.5)),
         ]
     )
-    torch.testing.assert_close(tf_actual(img), tf_expected(img))
+    torch.testing.assert_close(tf_actual(img_tensor), tf_expected(img_tensor))
 
 
 @require_x86_64_kernel
-def test_get_image_transforms_random_order(img):
+def test_get_image_transforms_random_order(img_tensor_factory):
     out_imgs = []
+    img_tensor = img_tensor_factory()
     tf = get_image_transforms(
         brightness_min_max=(0.5, 0.5),
         contrast_min_max=(0.5, 0.5),
@@ -141,13 +131,14 @@ def test_get_image_transforms_random_order(img):
     )
     with seeded_context(1337):
         for _ in range(10):
-            out_imgs.append(tf(img))
+            out_imgs.append(tf(img_tensor))
 
     for i in range(1, len(out_imgs)):
         with pytest.raises(AssertionError):
             torch.testing.assert_close(out_imgs[0], out_imgs[i])
 
 
+@pytest.mark.skip("TODO after v2 migration / removing hydra")
 @pytest.mark.parametrize(
     "transform, min_max_values",
     [
@@ -158,21 +149,24 @@ def test_get_image_transforms_random_order(img):
         ("sharpness", [(0.5, 0.5), (2.0, 2.0)]),
     ],
 )
-def test_backward_compatibility_torchvision(transform, min_max_values, img, single_transforms):
+def test_backward_compatibility_torchvision(img_tensor_factory, transform, min_max_values, single_transforms):
+    img_tensor = img_tensor_factory()
     for min_max in min_max_values:
         kwargs = {
             f"{transform}_weight": 1.0,
             f"{transform}_min_max": min_max,
         }
         tf = get_image_transforms(**kwargs)
-        actual = tf(img)
+        actual = tf(img_tensor)
         key = f"{transform}_{min_max[0]}_{min_max[1]}"
         expected = single_transforms[key]
         torch.testing.assert_close(actual, expected)
 
 
+@pytest.mark.skip("TODO after v2 migration / removing hydra")
 @require_x86_64_kernel
-def test_backward_compatibility_default_config(img, default_transforms):
+def test_backward_compatibility_default_config(img_tensor_factory, default_transforms):
+    img_tensor = img_tensor_factory()
     cfg = init_hydra_config(DEFAULT_CONFIG_PATH)
     cfg_tf = cfg.training.image_transforms
     default_tf = get_image_transforms(
@@ -191,7 +185,7 @@ def test_backward_compatibility_default_config(img, default_transforms):
     )
 
     with seeded_context(1337):
-        actual = default_tf(img)
+        actual = default_tf(img_tensor)
 
     expected = default_transforms["default"]
 
@@ -199,33 +193,36 @@ def test_backward_compatibility_default_config(img, default_transforms):
 
 
 @pytest.mark.parametrize("p", [[0, 1], [1, 0]])
-def test_random_subset_apply_single_choice(p, img):
+def test_random_subset_apply_single_choice(img_tensor_factory, p):
+    img_tensor = img_tensor_factory()
     flips = [v2.RandomHorizontalFlip(p=1), v2.RandomVerticalFlip(p=1)]
     random_choice = RandomSubsetApply(flips, p=p, n_subset=1, random_order=False)
-    actual = random_choice(img)
+    actual = random_choice(img_tensor)
 
     p_horz, _ = p
     if p_horz:
-        torch.testing.assert_close(actual, F.horizontal_flip(img))
+        torch.testing.assert_close(actual, F.horizontal_flip(img_tensor))
     else:
-        torch.testing.assert_close(actual, F.vertical_flip(img))
+        torch.testing.assert_close(actual, F.vertical_flip(img_tensor))
 
 
-def test_random_subset_apply_random_order(img):
+def test_random_subset_apply_random_order(img_tensor_factory):
+    img_tensor = img_tensor_factory()
     flips = [v2.RandomHorizontalFlip(p=1), v2.RandomVerticalFlip(p=1)]
     random_order = RandomSubsetApply(flips, p=[0.5, 0.5], n_subset=2, random_order=True)
     # We can't really check whether the transforms are actually applied in random order. However,
     # horizontal and vertical flip are commutative. Meaning, even under the assumption that the transform
     # applies them in random order, we can use a fixed order to compute the expected value.
-    actual = random_order(img)
-    expected = v2.Compose(flips)(img)
+    actual = random_order(img_tensor)
+    expected = v2.Compose(flips)(img_tensor)
     torch.testing.assert_close(actual, expected)
 
 
-def test_random_subset_apply_valid_transforms(color_jitters, img):
+def test_random_subset_apply_valid_transforms(img_tensor_factory, color_jitters):
+    img_tensor = img_tensor_factory()
     transform = RandomSubsetApply(color_jitters)
-    output = transform(img)
-    assert output.shape == img.shape
+    output = transform(img_tensor)
+    assert output.shape == img_tensor.shape
 
 
 def test_random_subset_apply_probability_length_mismatch(color_jitters):
@@ -239,16 +236,18 @@ def test_random_subset_apply_invalid_n_subset(color_jitters, n_subset):
         RandomSubsetApply(color_jitters, n_subset=n_subset)
 
 
-def test_sharpness_jitter_valid_range_tuple(img):
+def test_sharpness_jitter_valid_range_tuple(img_tensor_factory):
+    img_tensor = img_tensor_factory()
     tf = SharpnessJitter((0.1, 2.0))
-    output = tf(img)
-    assert output.shape == img.shape
+    output = tf(img_tensor)
+    assert output.shape == img_tensor.shape
 
 
-def test_sharpness_jitter_valid_range_float(img):
+def test_sharpness_jitter_valid_range_float(img_tensor_factory):
+    img_tensor = img_tensor_factory()
     tf = SharpnessJitter(0.5)
-    output = tf(img)
-    assert output.shape == img.shape
+    output = tf(img_tensor)
+    assert output.shape == img_tensor.shape
 
 
 def test_sharpness_jitter_invalid_range_min_negative():
@@ -261,6 +260,7 @@ def test_sharpness_jitter_invalid_range_max_smaller():
         SharpnessJitter((2.0, 0.1))
 
 
+@pytest.mark.skip("TODO after v2 migration / removing hydra")
 @pytest.mark.parametrize(
     "repo_id, n_examples",
     [

tests/test_image_writer.py

@@ -0,0 +1,359 @@
+import queue
+import time
+from multiprocessing import queues
+from unittest.mock import MagicMock, patch
+
+import numpy as np
+import pytest
+from PIL import Image
+
+from lerobot.common.datasets.image_writer import (
+    AsyncImageWriter,
+    image_array_to_image,
+    safe_stop_image_writer,
+    write_image,
+)
+
+DUMMY_IMAGE = "test_image.png"
+
+
+def test_init_threading():
+    writer = AsyncImageWriter(num_processes=0, num_threads=2)
+    try:
+        assert writer.num_processes == 0
+        assert writer.num_threads == 2
+        assert isinstance(writer.queue, queue.Queue)
+        assert len(writer.threads) == 2
+        assert len(writer.processes) == 0
+        assert all(t.is_alive() for t in writer.threads)
+    finally:
+        writer.stop()
+
+
+def test_init_multiprocessing():
+    writer = AsyncImageWriter(num_processes=2, num_threads=2)
+    try:
+        assert writer.num_processes == 2
+        assert writer.num_threads == 2
+        assert isinstance(writer.queue, queues.JoinableQueue)
+        assert len(writer.threads) == 0
+        assert len(writer.processes) == 2
+        assert all(p.is_alive() for p in writer.processes)
+    finally:
+        writer.stop()
+
+
+def test_zero_threads():
+    with pytest.raises(ValueError):
+        AsyncImageWriter(num_processes=0, num_threads=0)
+
+
+def test_image_array_to_image_rgb(img_array_factory):
+    img_array = img_array_factory(100, 100)
+    result_image = image_array_to_image(img_array)
+    assert isinstance(result_image, Image.Image)
+    assert result_image.size == (100, 100)
+    assert result_image.mode == "RGB"
+
+
+def test_image_array_to_image_pytorch_format(img_array_factory):
+    img_array = img_array_factory(100, 100).transpose(2, 0, 1)
+    result_image = image_array_to_image(img_array)
+    assert isinstance(result_image, Image.Image)
+    assert result_image.size == (100, 100)
+    assert result_image.mode == "RGB"
+
+
+@pytest.mark.skip("TODO: implement")
+def test_image_array_to_image_single_channel(img_array_factory):
+    img_array = img_array_factory(channels=1)
+    result_image = image_array_to_image(img_array)
+    assert isinstance(result_image, Image.Image)
+    assert result_image.size == (100, 100)
+    assert result_image.mode == "L"
+
+
+def test_image_array_to_image_float_array(img_array_factory):
+    img_array = img_array_factory(dtype=np.float32)
+    result_image = image_array_to_image(img_array)
+    assert isinstance(result_image, Image.Image)
+    assert result_image.size == (100, 100)
+    assert result_image.mode == "RGB"
+    assert np.array(result_image).dtype == np.uint8
+
+
+def test_image_array_to_image_out_of_bounds_float():
+    # Float array with values out of [0, 1]
+    img_array = np.random.uniform(-1, 2, size=(100, 100, 3)).astype(np.float32)
+    result_image = image_array_to_image(img_array)
+    assert isinstance(result_image, Image.Image)
+    assert result_image.size == (100, 100)
+    assert result_image.mode == "RGB"
+    assert np.array(result_image).dtype == np.uint8
+    assert np.array(result_image).min() >= 0 and np.array(result_image).max() <= 255
+
+
+def test_write_image_numpy(tmp_path, img_array_factory):
+    image_array = img_array_factory()
+    fpath = tmp_path / DUMMY_IMAGE
+    write_image(image_array, fpath)
+    assert fpath.exists()
+    saved_image = np.array(Image.open(fpath))
+    assert np.array_equal(image_array, saved_image)
+
+
+def test_write_image_image(tmp_path, img_factory):
+    image_pil = img_factory()
+    fpath = tmp_path / DUMMY_IMAGE
+    write_image(image_pil, fpath)
+    assert fpath.exists()
+    saved_image = Image.open(fpath)
+    assert list(saved_image.getdata()) == list(image_pil.getdata())
+    assert np.array_equal(image_pil, saved_image)
+
+
+def test_write_image_exception(tmp_path):
+    image_array = "invalid data"
+    fpath = tmp_path / DUMMY_IMAGE
+    with patch("builtins.print") as mock_print:
+        write_image(image_array, fpath)
+        mock_print.assert_called()
+        assert not fpath.exists()
+
+
+def test_save_image_numpy(tmp_path, img_array_factory):
+    writer = AsyncImageWriter()
+    try:
+        image_array = img_array_factory()
+        fpath = tmp_path / DUMMY_IMAGE
+        fpath.parent.mkdir(parents=True, exist_ok=True)
+        writer.save_image(image_array, fpath)
+        writer.wait_until_done()
+        assert fpath.exists()
+        saved_image = np.array(Image.open(fpath))
+        assert np.array_equal(image_array, saved_image)
+    finally:
+        writer.stop()
+
+
+def test_save_image_numpy_multiprocessing(tmp_path, img_array_factory):
+    writer = AsyncImageWriter(num_processes=2, num_threads=2)
+    try:
+        image_array = img_array_factory()
+        fpath = tmp_path / DUMMY_IMAGE
+        writer.save_image(image_array, fpath)
+        writer.wait_until_done()
+        assert fpath.exists()
+        saved_image = np.array(Image.open(fpath))
+        assert np.array_equal(image_array, saved_image)
+    finally:
+        writer.stop()
+
+
+def test_save_image_torch(tmp_path, img_tensor_factory):
+    writer = AsyncImageWriter()
+    try:
+        image_tensor = img_tensor_factory()
+        fpath = tmp_path / DUMMY_IMAGE
+        fpath.parent.mkdir(parents=True, exist_ok=True)
+        writer.save_image(image_tensor, fpath)
+        writer.wait_until_done()
+        assert fpath.exists()
+        saved_image = np.array(Image.open(fpath))
+        expected_image = (image_tensor.permute(1, 2, 0).cpu().numpy() * 255).astype(np.uint8)
+        assert np.array_equal(expected_image, saved_image)
+    finally:
+        writer.stop()
+
+
+def test_save_image_torch_multiprocessing(tmp_path, img_tensor_factory):
+    writer = AsyncImageWriter(num_processes=2, num_threads=2)
+    try:
+        image_tensor = img_tensor_factory()
+        fpath = tmp_path / DUMMY_IMAGE
+        writer.save_image(image_tensor, fpath)
+        writer.wait_until_done()
+        assert fpath.exists()
+        saved_image = np.array(Image.open(fpath))
+        expected_image = (image_tensor.permute(1, 2, 0).cpu().numpy() * 255).astype(np.uint8)
+        assert np.array_equal(expected_image, saved_image)
+    finally:
+        writer.stop()
+
+
+def test_save_image_pil(tmp_path, img_factory):
+    writer = AsyncImageWriter()
+    try:
+        image_pil = img_factory()
+        fpath = tmp_path / DUMMY_IMAGE
+        fpath.parent.mkdir(parents=True, exist_ok=True)
+        writer.save_image(image_pil, fpath)
+        writer.wait_until_done()
+        assert fpath.exists()
+        saved_image = Image.open(fpath)
+        assert list(saved_image.getdata()) == list(image_pil.getdata())
+    finally:
+        writer.stop()
+
+
+def test_save_image_pil_multiprocessing(tmp_path, img_factory):
+    writer = AsyncImageWriter(num_processes=2, num_threads=2)
+    try:
+        image_pil = img_factory()
+        fpath = tmp_path / DUMMY_IMAGE
+        writer.save_image(image_pil, fpath)
+        writer.wait_until_done()
+        assert fpath.exists()
+        saved_image = Image.open(fpath)
+        assert list(saved_image.getdata()) == list(image_pil.getdata())
+    finally:
+        writer.stop()
+
+
+def test_save_image_invalid_data(tmp_path):
+    writer = AsyncImageWriter()
+    try:
+        image_array = "invalid data"
+        fpath = tmp_path / DUMMY_IMAGE
+        fpath.parent.mkdir(parents=True, exist_ok=True)
+        with patch("builtins.print") as mock_print:
+            writer.save_image(image_array, fpath)
+            writer.wait_until_done()
+            mock_print.assert_called()
+            assert not fpath.exists()
+    finally:
+        writer.stop()
+
+
+def test_save_image_after_stop(tmp_path, img_array_factory):
+    writer = AsyncImageWriter()
+    writer.stop()
+    image_array = img_array_factory()
+    fpath = tmp_path / DUMMY_IMAGE
+    writer.save_image(image_array, fpath)
+    time.sleep(1)
+    assert not fpath.exists()
+
+
+def test_stop():
+    writer = AsyncImageWriter(num_processes=0, num_threads=2)
+    writer.stop()
+    assert not any(t.is_alive() for t in writer.threads)
+
+
+def test_stop_multiprocessing():
+    writer = AsyncImageWriter(num_processes=2, num_threads=2)
+    writer.stop()
+    assert not any(p.is_alive() for p in writer.processes)
+
+
+def test_multiple_stops():
+    writer = AsyncImageWriter()
+    writer.stop()
+    writer.stop()  # Should not raise an exception
+    assert not any(t.is_alive() for t in writer.threads)
+
+
+def test_multiple_stops_multiprocessing():
+    writer = AsyncImageWriter(num_processes=2, num_threads=2)
+    writer.stop()
+    writer.stop()  # Should not raise an exception
+    assert not any(t.is_alive() for t in writer.threads)
+
+
+def test_wait_until_done(tmp_path, img_array_factory):
+    writer = AsyncImageWriter(num_processes=0, num_threads=4)
+    try:
+        num_images = 100
+        image_arrays = [img_array_factory(height=500, width=500) for _ in range(num_images)]
+        fpaths = [tmp_path / f"frame_{i:06d}.png" for i in range(num_images)]
+        for image_array, fpath in zip(image_arrays, fpaths, strict=True):
+            fpath.parent.mkdir(parents=True, exist_ok=True)
+            writer.save_image(image_array, fpath)
+        writer.wait_until_done()
+        for i, fpath in enumerate(fpaths):
+            assert fpath.exists()
+            saved_image = np.array(Image.open(fpath))
+            assert np.array_equal(saved_image, image_arrays[i])
+    finally:
+        writer.stop()
+
+
+def test_wait_until_done_multiprocessing(tmp_path, img_array_factory):
+    writer = AsyncImageWriter(num_processes=2, num_threads=2)
+    try:
+        num_images = 100
+        image_arrays = [img_array_factory() for _ in range(num_images)]
+        fpaths = [tmp_path / f"frame_{i:06d}.png" for i in range(num_images)]
+        for image_array, fpath in zip(image_arrays, fpaths, strict=True):
+            fpath.parent.mkdir(parents=True, exist_ok=True)
+            writer.save_image(image_array, fpath)
+        writer.wait_until_done()
+        for i, fpath in enumerate(fpaths):
+            assert fpath.exists()
+            saved_image = np.array(Image.open(fpath))
+            assert np.array_equal(saved_image, image_arrays[i])
+    finally:
+        writer.stop()
+
+
+def test_exception_handling(tmp_path, img_array_factory):
+    writer = AsyncImageWriter()
+    try:
+        image_array = img_array_factory()
+        with (
+            patch.object(writer.queue, "put", side_effect=queue.Full("Queue is full")),
+            pytest.raises(queue.Full) as exc_info,
+        ):
+            writer.save_image(image_array, tmp_path / "test.png")
+        assert str(exc_info.value) == "Queue is full"
+    finally:
+        writer.stop()
+
+
+def test_with_different_image_formats(tmp_path, img_array_factory):
+    writer = AsyncImageWriter()
+    try:
+        image_array = img_array_factory()
+        formats = ["png", "jpeg", "bmp"]
+        for fmt in formats:
+            fpath = tmp_path / f"test_image.{fmt}"
+            write_image(image_array, fpath)
+            assert fpath.exists()
+    finally:
+        writer.stop()
+
+
+def test_safe_stop_image_writer_decorator():
+    class MockDataset:
+        def __init__(self):
+            self.image_writer = MagicMock(spec=AsyncImageWriter)
+
+    @safe_stop_image_writer
+    def function_that_raises_exception(dataset=None):
+        raise Exception("Test exception")
+
+    dataset = MockDataset()
+
+    with pytest.raises(Exception) as exc_info:
+        function_that_raises_exception(dataset=dataset)
+
+    assert str(exc_info.value) == "Test exception"
+    dataset.image_writer.stop.assert_called_once()
+
+
+def test_main_process_time(tmp_path, img_tensor_factory):
+    writer = AsyncImageWriter()
+    try:
+        image_tensor = img_tensor_factory()
+        fpath = tmp_path / DUMMY_IMAGE
+        start_time = time.perf_counter()
+        writer.save_image(image_tensor, fpath)
+        end_time = time.perf_counter()
+        time_spent = end_time - start_time
+        # Might need to adjust this threshold depending on hardware
+        assert time_spent < 0.01, f"Main process time exceeded threshold: {time_spent}s"
+        writer.wait_until_done()
+        assert fpath.exists()
+    finally:
+        writer.stop()