Fix record

This reverts commit 8d624643aa.

.cache/calibration/aloha_default/left_follower.json

@@ -0,0 +1,68 @@
+{
+    "homing_offset": [
+        2048,
+        3072,
+        3072,
+        -1024,
+        -1024,
+        2048,
+        -2048,
+        2048,
+        -2048
+    ],
+    "drive_mode": [
+        1,
+        1,
+        1,
+        0,
+        0,
+        1,
+        0,
+        1,
+        0
+    ],
+    "start_pos": [
+        2015,
+        3058,
+        3061,
+        1071,
+        1071,
+        2035,
+        2152,
+        2029,
+        2499
+    ],
+    "end_pos": [
+        -1008,
+        -1963,
+        -1966,
+        2141,
+        2143,
+        -971,
+        3043,
+        -1077,
+        3144
+    ],
+    "calib_mode": [
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "LINEAR"
+    ],
+    "motor_names": [
+        "waist",
+        "shoulder",
+        "shoulder_shadow",
+        "elbow",
+        "elbow_shadow",
+        "forearm_roll",
+        "wrist_angle",
+        "wrist_rotate",
+        "gripper"
+    ]
+}

.cache/calibration/aloha_default/left_leader.json

@@ -0,0 +1,68 @@
+{
+    "homing_offset": [
+        2048,
+        3072,
+        3072,
+        -1024,
+        -1024,
+        2048,
+        -2048,
+        2048,
+        -1024
+    ],
+    "drive_mode": [
+        1,
+        1,
+        1,
+        0,
+        0,
+        1,
+        0,
+        1,
+        0
+    ],
+    "start_pos": [
+        2035,
+        3024,
+        3019,
+        979,
+        981,
+        1982,
+        2166,
+        2124,
+        1968
+    ],
+    "end_pos": [
+        -990,
+        -2017,
+        -2015,
+        2078,
+        2076,
+        -1030,
+        3117,
+        -1016,
+        2556
+    ],
+    "calib_mode": [
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "LINEAR"
+    ],
+    "motor_names": [
+        "waist",
+        "shoulder",
+        "shoulder_shadow",
+        "elbow",
+        "elbow_shadow",
+        "forearm_roll",
+        "wrist_angle",
+        "wrist_rotate",
+        "gripper"
+    ]
+}

.cache/calibration/aloha_default/right_follower.json

@@ -0,0 +1,68 @@
+{
+    "homing_offset": [
+        2048,
+        3072,
+        3072,
+        -1024,
+        -1024,
+        2048,
+        -2048,
+        2048,
+        -2048
+    ],
+    "drive_mode": [
+        1,
+        1,
+        1,
+        0,
+        0,
+        1,
+        0,
+        1,
+        0
+    ],
+    "start_pos": [
+        2056,
+        2895,
+        2896,
+        1191,
+        1190,
+        2018,
+        2051,
+        2056,
+        2509
+    ],
+    "end_pos": [
+        -1040,
+        -2004,
+        -2006,
+        2126,
+        2127,
+        -1010,
+        3050,
+        -1117,
+        3143
+    ],
+    "calib_mode": [
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "LINEAR"
+    ],
+    "motor_names": [
+        "waist",
+        "shoulder",
+        "shoulder_shadow",
+        "elbow",
+        "elbow_shadow",
+        "forearm_roll",
+        "wrist_angle",
+        "wrist_rotate",
+        "gripper"
+    ]
+}

.cache/calibration/aloha_default/right_leader.json

@@ -0,0 +1,68 @@
+{
+    "homing_offset": [
+        2048,
+        3072,
+        3072,
+        -1024,
+        -1024,
+        2048,
+        -2048,
+        2048,
+        -2048
+    ],
+    "drive_mode": [
+        1,
+        1,
+        1,
+        0,
+        0,
+        1,
+        0,
+        1,
+        0
+    ],
+    "start_pos": [
+        2068,
+        3034,
+        3030,
+        1038,
+        1041,
+        1991,
+        1948,
+        2090,
+        1985
+    ],
+    "end_pos": [
+        -1025,
+        -2014,
+        -2015,
+        2058,
+        2060,
+        -955,
+        3091,
+        -940,
+        2576
+    ],
+    "calib_mode": [
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "LINEAR"
+    ],
+    "motor_names": [
+        "waist",
+        "shoulder",
+        "shoulder_shadow",
+        "elbow",
+        "elbow_shadow",
+        "forearm_roll",
+        "wrist_angle",
+        "wrist_rotate",
+        "gripper"
+    ]
+}

.dockerignore

@@ -65,7 +65,6 @@ htmlcov/
 .nox/
 .coverage
 .coverage.*
-.cache
 nosetests.xml
 coverage.xml
 *.cover
@@ -73,6 +72,11 @@ coverage.xml
 .hypothesis/
 .pytest_cache/
 
+# Ignore .cache except calibration
+.cache/*
+!.cache/calibration/
+!.cache/calibration/**
+
 # Translations
 *.mo
 *.pot

.gitattributes

@@ -3,4 +3,4 @@
 *.safetensors filter=lfs diff=lfs merge=lfs -text
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.arrow filter=lfs diff=lfs merge=lfs -text
-*.json filter=lfs diff=lfs merge=lfs -text
+*.json !text !filter !merge !diff

.github/workflows/test.yml

@@ -11,6 +11,7 @@ on:
       - ".github/**"
       - "poetry.lock"
       - "Makefile"
+      - ".cache/**"
   push:
     branches:
       - main
@@ -21,6 +22,7 @@ on:
       - ".github/**"
       - "poetry.lock"
       - "Makefile"
+      - ".cache/**"
 
 jobs:
   pytest:
@@ -35,13 +37,17 @@ jobs:
           lfs: true  # Ensure LFS files are pulled
 
       - name: Install apt dependencies
-        run: sudo apt-get update && sudo apt-get install -y libegl1-mesa-dev ffmpeg
+      # portaudio19-dev is needed to install pyaudio
+        run: |
+          sudo apt-get update && \
+          sudo apt-get install -y libegl1-mesa-dev ffmpeg portaudio19-dev
 
       - name: Install poetry
         run: |
           pipx install poetry && poetry config virtualenvs.in-project true
           echo "${{ github.workspace }}/.venv/bin" >> $GITHUB_PATH
 
+      # TODO(rcadene, aliberts): python 3.12 seems to be used in the tests, not python 3.10
       - name: Set up Python 3.10
         uses: actions/setup-python@v5
         with:
@@ -60,7 +66,6 @@ jobs:
             -W ignore::UserWarning:gymnasium.utils.env_checker:247 \
             && rm -rf tests/outputs outputs
 
-
   pytest-minimal:
     name: Pytest (minimal install)
     runs-on: ubuntu-latest
@@ -80,6 +85,7 @@ jobs:
           pipx install poetry && poetry config virtualenvs.in-project true
           echo "${{ github.workspace }}/.venv/bin" >> $GITHUB_PATH
 
+      # TODO(rcadene, aliberts): python 3.12 seems to be used in the tests, not python 3.10
       - name: Set up Python 3.10
         uses: actions/setup-python@v5
         with:
@@ -110,7 +116,10 @@ jobs:
           lfs: true  # Ensure LFS files are pulled
 
       - name: Install apt dependencies
-        run: sudo apt-get update && sudo apt-get install -y libegl1-mesa-dev
+      # 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: |

.gitignore

@@ -66,7 +66,6 @@ htmlcov/
 .nox/
 .coverage
 .coverage.*
-.cache
 nosetests.xml
 coverage.xml
 *.cover
@@ -74,6 +73,11 @@ coverage.xml
 .hypothesis/
 .pytest_cache/
 
+# Ignore .cache except calibration
+.cache/*
+!.cache/calibration/
+!.cache/calibration/**
+
 # Translations
 *.mo
 *.pot

CONTRIBUTING.md

@@ -20,7 +20,7 @@ Some of the ways you can contribute to 🤗 LeRobot:
 * Contributing to the examples or to the documentation.
 * Submitting issues related to bugs or desired new features.
 
-Following the guides below, feel free to open issues and PRs and to coordinate your efforts with the community on our [Discord Channel](https://discord.gg/VjFz58wn3R). For specific inquiries, reach out to [Remi Cadene](remi.cadene@huggingface.co).
+Following the guides below, feel free to open issues and PRs and to coordinate your efforts with the community on our [Discord Channel](https://discord.gg/VjFz58wn3R). For specific inquiries, reach out to [Remi Cadene](mailto:remi.cadene@huggingface.co).
 
 If you are not sure how to contribute or want to know the next features we working on, look on this project page: [LeRobot TODO](https://github.com/orgs/huggingface/projects/46)
 

README.md

@@ -31,12 +31,13 @@
     <p>We just dropped an in-depth tutorial on how to build your own robot!</p>
     <p>Teach it new skills by showing it a few moves with just a laptop.</p>
     <p>Then watch your homemade robot act autonomously 🤯</p>
-    <p>For more info, see [our thread on X](https://x.com/RemiCadene/status/1825455895561859185) or [our tutorial page](https://github.com/huggingface/lerobot/blob/main/examples/7_get_started_with_real_robot.md).</p>
+    <p>For more info, see <a href="https://x.com/RemiCadene/status/1825455895561859185">our thread on X</a> or <a href="https://github.com/huggingface/lerobot/blob/main/examples/7_get_started_with_real_robot.md">our tutorial page</a>.</p>
 </div>
 
+<br/>
 
 <h3 align="center">
-    <p>State-of-the-art AI for real-world robotics</p>
+    <p>LeRobot: State-of-the-art AI for real-world robotics</p>
 </h3>
 
 ---
@@ -266,13 +267,20 @@ checkpoints
 │   └── training_state.pth  # optimizer/scheduler/rng state and training step
 ```
 
+To resume training from a checkpoint, you can add these to the `train.py` python command:
+```bash
+    hydra.run.dir=your/original/experiment/dir resume=true
+```
+
+It will load the pretrained model, optimizer and scheduler states for training. For more information please see our tutorial on training resumption [here](https://github.com/huggingface/lerobot/blob/main/examples/5_resume_training.md).
+
 To use wandb for logging training and evaluation curves, make sure you've run `wandb login` as a one-time setup step. Then, when running the training command above, enable WandB in the configuration by adding:
 
 ```bash
     wandb.enable=true
 ```
 
-A link to the wandb logs for the run will also show up in yellow in your terminal. Here is an example of what they look like in your browser:
+A link to the wandb logs for the run will also show up in yellow in your terminal. Here is an example of what they look like in your browser. Please also check [here](https://github.com/huggingface/lerobot/blob/main/examples/4_train_policy_with_script.md#typical-logs-and-metrics) for the explaination of some commonly used metrics in logs.
 
 ![](media/wandb.png)
 

docker/lerobot-cpu/Dockerfile

@@ -22,7 +22,7 @@ RUN echo "source /opt/venv/bin/activate" >> /root/.bashrc
 COPY . /lerobot
 WORKDIR /lerobot
 RUN pip install --upgrade --no-cache-dir pip
-RUN pip install --no-cache-dir ".[test, aloha, xarm, pusht, koch]" \
+RUN pip install --no-cache-dir ".[test, aloha, xarm, pusht, dynamixel]" \
     --extra-index-url https://download.pytorch.org/whl/cpu
 
 # Set EGL as the rendering backend for MuJoCo

docker/lerobot-gpu/Dockerfile

@@ -24,7 +24,7 @@ RUN echo "source /opt/venv/bin/activate" >> /root/.bashrc
 COPY . /lerobot
 WORKDIR /lerobot
 RUN pip install --upgrade --no-cache-dir pip
-RUN pip install --no-cache-dir ".[test, aloha, xarm, pusht, koch]"
+RUN pip install --no-cache-dir ".[test, aloha, xarm, pusht, dynamixel]"
 
 # Set EGL as the rendering backend for MuJoCo
 ENV MUJOCO_GL="egl"

examples/10_use_so100.md

@@ -0,0 +1,208 @@
+This tutorial explains how to use [SO-100](https://github.com/TheRobotStudio/SO-ARM100) with LeRobot.
+
+## 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 install LeRobot. We will next provide a tutorial for assembly.
+
+## Install LeRobot
+
+On your computer:
+
+1. [Install Miniconda](https://docs.anaconda.com/miniconda/#quick-command-line-install):
+```bash
+mkdir -p ~/miniconda3
+wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-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`
+
+3. Create and activate a fresh conda environment for lerobot
+```bash
+conda create -y -n lerobot python=3.10 && conda activate lerobot
+```
+
+4. Clone LeRobot:
+```bash
+git clone https://github.com/huggingface/lerobot.git ~/lerobot
+```
+
+5. Install LeRobot with dependencies for the feetech motors:
+```bash
+cd ~/lerobot && pip install -e ".[feetech]"
+```
+
+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
+
+Run this script two times to find the ports (e.g. "/dev/tty.usbmodem58760432961") of your motor buses:
+```bash
+python lerobot/scripts/find_motors_bus_port.py
+```
+
+Then plug your first motor, corresponding to "shoulder_pan" and run this script to set its ID to 1 and set its present position and offset to ~2048 (useful for calibration).
+```bash
+python lerobot/scripts/configure_motor.py \
+  --port /dev/tty.usbmodem58760432961 \
+  --brand feetech \
+  --model sts3215 \
+  --baudrate 1000000 \
+  --ID 1
+```
+
+Then unplug your motor and plug the second motor, corresponding to "shoulder lift", and set its ID to 2.
+```bash
+python lerobot/scripts/configure_motor.py \
+  --port /dev/tty.usbmodem58760432961 \
+  --brand feetech \
+  --model sts3215 \
+  --baudrate 1000000 \
+  --ID 2
+```
+
+Redo the process for all your motors until the gripper with ID 6. Do the same for the motors of the leader arm, starting for ID 1 up to 6.
+
+## Assemble the arms
+
+TODO
+
+## Calibrate
+
+```bash
+python lerobot/scripts/control_robot.py calibrate \
+    --robot-path lerobot/configs/robot/so100.yaml \
+    --robot-overrides '~cameras'
+```
+
+## Teleoperate
+
+Without displaying the cameras:
+```bash
+python lerobot/scripts/control_robot.py teleoperate \
+    --robot-path lerobot/configs/robot/so100.yaml \
+    --robot-overrides '~cameras' \
+    --display-cameras 0
+```
+
+With displaying the cameras:
+```bash
+python lerobot/scripts/control_robot.py teleoperate \
+    --robot-path lerobot/configs/robot/so100.yaml
+```
+
+## Record a dataset
+
+Once you're familiar with teleoperation, you can record your first dataset with so100.
+
+If you want to use the Hugging Face hub features for uploading your dataset and you haven't previously done it, make sure you've logged in using a write-access token, which can be generated from the [Hugging Face settings](https://huggingface.co/settings/tokens):
+```bash
+huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+```
+
+Store your Hugging Face repository name in a variable to run these commands:
+```bash
+HF_USER=$(huggingface-cli whoami | head -n 1)
+echo $HF_USER
+```
+
+Record 2 episodes and upload your dataset to the hub:
+```bash
+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 \
+    --episode-time-s 40 \
+    --reset-time-s 10 \
+    --num-episodes 2 \
+    --push-to-hub 1
+```
+
+## 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
+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
+
+Now try to replay the first episode on your robot:
+```bash
+DATA_DIR=data python lerobot/scripts/control_robot.py replay \
+    --robot-path lerobot/configs/robot/so100.yaml \
+    --fps 30 \
+    --root data \
+    --repo-id ${HF_USER}/so100_test \
+    --episode 0
+```
+
+## Train a policy
+
+To train a policy to control your robot, use the [`python lerobot/scripts/train.py`](../lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
+```bash
+DATA_DIR=data python lerobot/scripts/train.py \
+  dataset_repo_id=${HF_USER}/so100_test \
+  policy=act_so100_real \
+  env=so100_real \
+  hydra.run.dir=outputs/train/act_so100_test \
+  hydra.job.name=act_so100_test \
+  device=cuda \
+  wandb.enable=true
+```
+
+Let's explain it:
+1. We provided the dataset as argument with `dataset_repo_id=${HF_USER}/so100_test`.
+2. We provided the policy with `policy=act_so100_real`. This loads configurations from [`lerobot/configs/policy/act_so100_real.yaml`](../lerobot/configs/policy/act_so100_real.yaml). Importantly, this policy uses 2 cameras as input `laptop`, `phone`.
+3. We provided an environment as argument with `env=so100_real`. This loads configurations from [`lerobot/configs/env/so100_real.yaml`](../lerobot/configs/env/so100_real.yaml).
+4. We provided `device=cuda` since we are training on a Nvidia GPU, but you can also use `device=mps` if you are using a Mac with Apple silicon, or `device=cpu` otherwise.
+5. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
+6. We added `DATA_DIR=data` to access your dataset stored in your local `data` directory. If you dont provide `DATA_DIR`, your dataset will be downloaded from Hugging Face hub to your cache folder `$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
+
+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 \
+  --episode-time-s 40 \
+  --reset-time-s 10 \
+  --num-episodes 10 \
+  -p outputs/train/act_so100_test/checkpoints/last/pretrained_model
+```
+
+As you can see, it's almost the same command as previously used to record your training dataset. Two things changed:
+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
+
+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 explaination.
+
+If you have any question or need help, please reach out on Discord in the channel `#so100-arm`.

examples/4_train_policy_with_script.md

@@ -170,6 +170,36 @@ python lerobot/scripts/train.py --config-dir outputs/train/my_experiment/checkpo
 
 Note that you may still use the regular syntax for config parameter overrides (eg: by adding `training.offline_steps=200000`).
 
+## 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:
+
+```
+INFO 2024-08-14 13:35:12 ts/train.py:192 step:0 smpl:64 ep:1 epch:0.00 loss:1.112 grdn:15.387 lr:2.0e-07 updt_s:1.738 data_s:4.774
+```
+
+or evaluation log like:
+
+```
+INFO 2024-08-14 13:38:45 ts/train.py:226 step:100 smpl:6K ep:52 epch:0.25 ∑rwrd:20.693 success:0.0% eval_s:120.266
+```
+
+These logs will also be saved in wandb if `wandb.enable` is set to `true`. Here are the meaning of some abbreviations:
+
+- `smpl`: number of samples seen during training.
+- `ep`: number of episodes seen during training. An episode contains multiple samples in a complete manipulation task.
+- `epch`: number of time all unique samples are seen (epoch).
+- `grdn`: gradient norm.
+- `∑rwrd`: compute the sum of rewards in every evaluation episode and then take an average of them.
+- `success`: average success rate of eval episodes. Reward and success are usually different except for the sparsing reward setting, where reward=1 only when the task is completed successfully.
+- `eval_s`: time to evaluate the policy in the environment, in second.
+- `updt_s`: time to update the network parameters, in second.
+- `data_s`: time to load a batch of data, in second.
+
+Some metrics are useful for initial performance profiling. For example, if you find the current GPU utilization is low via the `nvidia-smi` command and `data_s` sometimes is too high, you may need to modify batch size or number of dataloading workers to accelerate dataloading. We also recommend [pytorch profiler](https://github.com/huggingface/lerobot?tab=readme-ov-file#improve-your-code-with-profiling) for detailed performance probing.
+
 ---
 
 So far we've seen how to train Diffusion Policy for PushT and ACT for ALOHA. Now, what if we want to train ACT for PushT? Well, there are aspects of the ACT configuration that are specific to the ALOHA environments, and these happen to be incompatible with PushT. Therefore, trying to run the following will almost certainly raise an exception of sorts (eg: feature dimension mismatch):

examples/7_get_started_with_real_robot.md

@@ -11,7 +11,7 @@ This tutorial will guide you through the process of setting up and training a ne
 
 By following these steps, you'll be able to replicate tasks like picking up a Lego block and placing it in a bin with a high success rate, as demonstrated in [this video](https://x.com/RemiCadene/status/1814680760592572934).
 
-Although this tutorial is general and can be easily adapted to various types of robots by changing the configuration, it is specifically based on the [Koch v1.1](https://github.com/jess-moss/koch-v1-1), an affordable robot. The Koch v1.1 consists of a leader arm and a follower arm, each with 6 motors. It can work with one or several cameras to record the scene, which serve as visual sensors for the robot.
+This tutorial is specifically made for the affordable [Koch v1.1](https://github.com/jess-moss/koch-v1-1) robot, but it contains additional information to be easily adapted to various types of robots like [Aloha bimanual robot](https://aloha-2.github.io) by changing some configurations. The Koch v1.1 consists of a leader arm and a follower arm, each with 6 motors. It can work with one or several cameras to record the scene, which serve as visual sensors for the robot.
 
 During the data collection phase, you will control the follower arm by moving the leader arm. This process is known as "teleoperation." This technique is used to collect robot trajectories. Afterward, you'll train a neural network to imitate these trajectories and deploy the network to enable your robot to operate autonomously.
 
@@ -29,16 +29,23 @@ 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 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.
 
 Using `pip`:
 ```bash
-pip install -e ".[koch]"
+pip install -e ".[dynamixel]"
 ```
 
 Or using `poetry`:
 ```bash
-poetry install --sync --extras "koch"
+poetry install --sync --extras "dynamixel"
+```
+
+/!\ For Linux only, ffmpeg and opencv requires conda install for now. Run this exact sequence of commands:
+```bash
+conda install -c conda-forge ffmpeg
+pip uninstall opencv-python
+conda install -c conda-forge "opencv>=4.10.0"
 ```
 
 You are now ready to plug the 5V power supply to the motor bus of the leader arm (the smaller one) since all its motors only require 5V.
@@ -71,12 +78,12 @@ To begin, create two instances of the  [`DynamixelMotorsBus`](../lerobot/common/
 
 To find the correct ports for each arm, run the utility script twice:
 ```bash
-python lerobot/common/robot_devices/motors/dynamixel.py
+python lerobot/scripts/find_motors_bus_port.py
 ```
 
 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 DynamixelMotorsBus.
+Finding all available ports for the MotorBus.
 ['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
 Remove the usb cable from your DynamixelMotorsBus and press Enter when done.
 
@@ -88,7 +95,7 @@ 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 DynamixelMotorsBus.
+Finding all available ports for the MotorBus.
 ['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
 Remove the usb cable from your DynamixelMotorsBus and press Enter when done.
 
@@ -147,6 +154,7 @@ follower_arm = DynamixelMotorsBus(
 Next, update the port values in the YAML configuration file for the Koch robot at [`lerobot/configs/robot/koch.yaml`](../lerobot/configs/robot/koch.yaml) with the ports you've identified:
 ```yaml
 [...]
+robot_type: koch
 leader_arms:
   main:
     _target_: lerobot.common.robot_devices.motors.dynamixel.DynamixelMotorsBus
@@ -174,6 +182,8 @@ follower_arms:
 [...]
 ```
 
+Don't forget to set `robot_type: aloha` if you follow this tutorial with [Aloha bimanual robot](aloha-2.github.io) instead of Koch v1.1
+
 This configuration file is used to instantiate your robot across all scripts. We'll cover how this works later on.
 
 **Connect and Configure your Motors**
@@ -298,32 +308,37 @@ Alternatively, you can unplug the power cord, which will automatically disable t
 
 */!\ Warning*: These motors tend to overheat, especially under torque or if left plugged in for too long. Unplug after use.
 
-### b. Teleoperate your Koch v1.1 with KochRobot
+### b. Teleoperate your Koch v1.1 with ManipulatorRobot
 
-**Instantiate the KochRobot**
+**Instantiate the ManipulatorRobot**
 
-Before you can teleoperate your robot, you need to instantiate the  [`KochRobot`](../lerobot/common/robot_devices/robots/koch.py) using the previously defined `leader_arm` and `follower_arm`.
+Before you can teleoperate your robot, you need to instantiate the  [`ManipulatorRobot`](../lerobot/common/robot_devices/robots/manipulator.py) using the previously defined `leader_arm` and `follower_arm`.
 
-For the Koch robot, we only have one leader, so we refer to it as `"main"` and define it as `leader_arms={"main": leader_arm}`. We do the same for the follower arm. For other robots (like the Aloha), which may have two pairs of leader and follower arms, you would define them like this: `leader_arms={"left": left_leader_arm, "right": right_leader_arm},`. Same thing for the follower arms.
+For the Koch v1.1 robot, we only have one leader, so we refer to it as `"main"` and define it as `leader_arms={"main": leader_arm}`. We do the same for the follower arm. For other robots (like the Aloha), which may have two pairs of leader and follower arms, you would define them like this: `leader_arms={"left": left_leader_arm, "right": right_leader_arm},`. Same thing for the follower arms.
 
-You also need to provide a path to a calibration file, such as  `calibration_path=".cache/calibration/koch.pkl"`. More on this in the next section.
+You also need to provide a path to a calibration directory, such as  `calibration_dir=".cache/calibration/koch"`. More on this in the next section.
 
-Run the following code to instantiate your Koch robot:
+Run the following code to instantiate your manipulator robot:
 ```python
-from lerobot.common.robot_devices.robots.koch import KochRobot
+from lerobot.common.robot_devices.robots.manipulator import ManipulatorRobot
 
-robot = KochRobot(
+robot = ManipulatorRobot(
+    robot_type="koch",
     leader_arms={"main": leader_arm},
     follower_arms={"main": follower_arm},
-    calibration_path=".cache/calibration/koch.pkl",
+    calibration_dir=".cache/calibration/koch",
 )
 ```
 
-**Calibrate and Connect the KochRobot**
+The `robot_type="koch"` is used to set the associated settings and calibration process. For instance, we activate the torque of the gripper of the leader Koch v1.1 arm and position it at a 40 degree angle to use it as a trigger.
 
-Next, you'll need to calibrate your 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 Koch robot to work on another.
+For the [Aloha bimanual robot](https://aloha-2.github.io), we would use `robot_type="aloha"` to set different settings such as a secondary ID for shadow joints (shoulder, elbow). Specific to Aloha, LeRobot comes with default calibration files stored in in `.cache/calibration/aloha_default`. Assuming the motors have been properly assembled, no manual calibration step is expected. If you need to run manual calibration, simply update `calibration_dir` to `.cache/calibration/aloha`.
 
-When you connect your robot for the first time, the [`KochRobot`](../lerobot/common/robot_devices/robots/koch.py) will detect if the calibration file is missing and trigger the calibration procedure. During this process, you will be guided to move each arm to three different positions.
+**Calibrate and Connect the ManipulatorRobot**
+
+Next, you'll need to calibrate your Koch 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 Koch robot to work on another.
+
+When you connect your robot for the first time, the [`ManipulatorRobot`](../lerobot/common/robot_devices/robots/manipulator.py) will detect if the calibration file is missing and trigger the calibration procedure. During this process, you will be guided to move each arm to three different positions.
 
 Here are the positions you'll move the follower arm to:
 
@@ -354,27 +369,26 @@ The output will look like this:
 ```
 Connecting main follower arm
 Connecting main leader arm
-Missing calibration file '.cache/calibration/koch.pkl'. Starting calibration procedure.
-
-Running calibration of main follower...
 
+Missing calibration file '.cache/calibration/koch/main_follower.json'
+Running calibration of koch main follower...
 Move arm to zero position
 [...]
 Move arm to rotated position
 [...]
 Move arm to rest position
 [...]
+Calibration is done! Saving calibration file '.cache/calibration/koch/main_follower.json'
 
-Running calibration of main leader...
-
+Missing calibration file '.cache/calibration/koch/main_leader.json'
+Running calibration of koch main leader...
 Move arm to zero position
 [...]
 Move arm to rotated position
 [...]
 Move arm to rest position
 [...]
-
-Calibration is done! Saving calibration file '.cache/calibration/koch.pkl'
+Calibration is done! Saving calibration file '.cache/calibration/koch/main_leader.json'
 ```
 
 *Verifying Calibration*
@@ -414,7 +428,7 @@ for _ in tqdm.tqdm(range(seconds*frequency)):
 
 *Using `teleop_step` for Teleoperation*
 
-Alternatively, you can teleoperate the robot using the `teleop_step` method from [`KochRobot`](../lerobot/common/robot_devices/robots/koch.py).
+Alternatively, you can teleoperate the robot using the `teleop_step` method from [`ManipulatorRobot`](../lerobot/common/robot_devices/robots/manipulator.py).
 
 Run this code to teleoperate:
 ```python
@@ -607,10 +621,10 @@ Additionaly, you can set up your robot to work with your cameras.
 
 Modify the following Python code with the appropriate camera names and configurations:
 ```python
-robot = KochRobot(
+robot = ManipulatorRobot(
     leader_arms={"main": leader_arm},
     follower_arms={"main": follower_arm},
-    calibration_path=".cache/calibration/koch.pkl",
+    calibration_dir=".cache/calibration/koch",
     cameras={
         "laptop": OpenCVCamera(0, fps=30, width=640, height=480),
         "phone": OpenCVCamera(1, fps=30, width=640, height=480),
@@ -752,7 +766,7 @@ Before trying `record`, if you want to push your dataset to the hub, make sure y
 ```bash
 huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
 ```
-Also, store your Hugging Face repositery name in a variable (e.g. `cadene` or `lerobot`). For instance, run this to use your Hugging Face user name as repositery:
+Also, store your Hugging Face repository name in a variable (e.g. `cadene` or `lerobot`). For instance, run this to use your Hugging Face user name as repository:
 ```bash
 HF_USER=$(huggingface-cli whoami | head -n 1)
 echo $HF_USER
@@ -925,7 +939,7 @@ huggingface-cli upload ${HF_USER}/act_koch_test_${CKPT} \
 
 ## 5. Evaluate your policy
 
-Now that you have a policy checkpoint, you can easily control your robot with it using methods from [`KochRobot`](../lerobot/common/robot_devices/robots/koch.py) and the policy.
+Now that you have a policy checkpoint, you can easily control your robot with it using methods from [`ManipulatorRobot`](../lerobot/common/robot_devices/robots/manipulator.py) and the policy.
 
 Try this code for running inference for 60 seconds at 30 fps:
 ```python

examples/8_use_stretch.md

@@ -0,0 +1,158 @@
+This tutorial explains how to use [Stretch 3](https://hello-robot.com/stretch-3-product) with LeRobot.
+
+## Setup
+
+Familiarize yourself with Stretch by following its [tutorials](https://docs.hello-robot.com/0.3/getting_started/hello_robot/) (recommended).
+
+To use LeRobot on Stretch, 3 options are available:
+- [tethered setup](https://docs.hello-robot.com/0.3/getting_started/connecting_to_stretch/#tethered-setup)
+- [untethered setup](https://docs.hello-robot.com/0.3/getting_started/connecting_to_stretch/#untethered-setup)
+- ssh directly into Stretch (you will first need to install and configure openssh-server on stretch using one of the two above setups)
+
+
+## Install LeRobot
+
+On Stretch's CLI, follow these steps:
+
+1. [Install Miniconda](https://docs.anaconda.com/miniconda/#quick-command-line-install):
+```bash
+mkdir -p ~/miniconda3
+wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-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. Comment out these lines in `~/.profile` (this can mess up paths used by conda and ~/.local/bin should already be in your PATH)
+```
+# set PATH so it includes user's private bin if it exists
+if [ -d "$HOME/.local/bin" ] ; then
+    PATH="$HOME/.local/bin:$PATH"
+fi
+```
+
+3. Restart shell or `source ~/.bashrc`
+
+4. Create and activate a fresh conda environment for lerobot
+```bash
+conda create -y -n lerobot python=3.10 && conda activate lerobot
+```
+
+5. Clone LeRobot:
+```bash
+git clone https://github.com/huggingface/lerobot.git ~/lerobot
+```
+
+6. Install LeRobot with stretch dependencies:
+```bash
+cd ~/lerobot && pip install -e ".[stretch]"
+```
+
+> **Note:** If you get this message, you can ignore it: `ERROR: pip's dependency resolver does not currently take into account all the packages that are installed.`
+
+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"
+```
+
+7. Run a [system check](https://docs.hello-robot.com/0.3/getting_started/stretch_hardware_overview/#system-check) to make sure your robot is ready:
+```bash
+stretch_system_check.py
+```
+
+> **Note:** You may need to free the "robot process" after booting Stretch by running `stretch_free_robot_process.py`. For more info this Stretch's [doc](https://docs.hello-robot.com/0.3/getting_started/stretch_hardware_overview/#turning-off-gamepad-teleoperation).
+
+You should get something like this:
+```bash
+For use with S T R E T C H (R) from Hello Robot Inc.
+---------------------------------------------------------------------
+
+Model = Stretch 3
+Tool = DexWrist 3 w/ Gripper
+Serial Number = stretch-se3-3054
+
+---- Checking Hardware ----
+[Pass] Comms are ready
+[Pass] Actuators are ready
+[Warn] Sensors not ready (IMU AZ = -10.19 out of range -10.1 to -9.5)
+[Pass] Battery voltage is 13.6 V
+
+---- Checking Software ----
+[Pass] Ubuntu 22.04 is ready
+[Pass] All APT pkgs are setup correctly
+[Pass] Firmware is up-to-date
+[Pass] Python pkgs are up-to-date
+[Pass] ROS2 Humble is ready
+```
+
+## Teleoperate, record a dataset and run a policy
+
+**Calibrate (Optional)**
+Before operating Stretch, you need to [home](https://docs.hello-robot.com/0.3/getting_started/stretch_hardware_overview/#homing) it first. Be mindful about giving Stretch some space as this procedure will move the robot's arm and gripper. Now run this command:
+```bash
+python lerobot/scripts/control_robot.py calibrate \
+    --robot-path lerobot/configs/robot/stretch.yaml
+```
+This is equivalent to running `stretch_robot_home.py`
+
+> **Note:** If you run any of the LeRobot scripts below and Stretch is not poperly homed, it will automatically home/calibrate first.
+
+**Teleoperate**
+Before trying teleoperation, you need activate the gamepad controller by pressing the middle button. For more info, see Stretch's [doc](https://docs.hello-robot.com/0.3/getting_started/hello_robot/#gamepad-teleoperation).
+
+Now try out teleoperation (see above documentation to learn about the gamepad controls):
+```bash
+python lerobot/scripts/control_robot.py teleoperate \
+    --robot-path lerobot/configs/robot/stretch.yaml
+```
+This is essentially the same as running `stretch_gamepad_teleop.py`
+
+**Record a dataset**
+Once you're familiar with the gamepad controls and after a bit of practice, you can try to record your first dataset with Stretch.
+
+If you want to use the Hugging Face hub features for uploading your dataset and you haven't previously done it, make sure you've logged in using a write-access token, which can be generated from the [Hugging Face settings](https://huggingface.co/settings/tokens):
+```bash
+huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+```
+
+Store your Hugging Face repository name in a variable to run these commands:
+```bash
+HF_USER=$(huggingface-cli whoami | head -n 1)
+echo $HF_USER
+```
+
+Record one episode:
+```bash
+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 \
+    --episode-time-s 40 \
+    --reset-time-s 10 \
+    --num-episodes 1 \
+    --push-to-hub 0
+```
+
+> **Note:** If you're using ssh to connect to Stretch and run this script, you won't be able to visualize its cameras feed (though they will still be recording). To see the cameras stream, use [tethered](https://docs.hello-robot.com/0.3/getting_started/connecting_to_stretch/#tethered-setup) or [untethered setup](https://docs.hello-robot.com/0.3/getting_started/connecting_to_stretch/#untethered-setup).
+
+**Replay an episode**
+Now try to replay this episode (make sure the robot's initial position is the same):
+```bash
+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
+```
+
+Follow [previous tutorial](https://github.com/huggingface/lerobot/blob/main/examples/7_get_started_with_real_robot.md#4-train-a-policy-on-your-data) to train a policy on your data and run inference on your robot. You will need to adapt the code for Stretch.
+
+> TODO(rcadene, aliberts): Add already setup environment and policy yaml configuration files
+
+If you need help, please reach out on Discord in the channel `#stretch3-mobile-arm`.

examples/9_use_aloha.md

@@ -0,0 +1,179 @@
+This tutorial explains how to use [Aloha and Aloha 2 stationary](https://www.trossenrobotics.com/aloha-stationary) with LeRobot.
+
+## Setup
+
+Follow the [documentation from Trossen Robotics](https://docs.trossenrobotics.com/aloha_docs/getting_started/stationary/hardware_setup.html) for setting up the hardware and plugging the 4 arms and 4 cameras to your computer.
+
+
+## Install LeRobot
+
+On your computer:
+
+1. [Install Miniconda](https://docs.anaconda.com/miniconda/#quick-command-line-install):
+```bash
+mkdir -p ~/miniconda3
+wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-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`
+
+3. Create and activate a fresh conda environment for lerobot
+```bash
+conda create -y -n lerobot python=3.10 && conda activate lerobot
+```
+
+4. Clone LeRobot:
+```bash
+git clone https://github.com/huggingface/lerobot.git ~/lerobot
+```
+
+5. Install LeRobot with dependencies for the Aloha motors (dynamixel) and cameras (intelrealsense):
+```bash
+cd ~/lerobot && pip install -e ".[dynamixel, intelrealsense]"
+```
+
+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"
+```
+
+## Teleoperate
+
+**/!\ FOR SAFETY, READ THIS /!\**
+Teleoperation consists in manually operating the leader arms to move the follower arms. Importantly:
+1. Make sure your leader arms are in the same position as the follower arms, so that the follower arms don't move too fast to match the leader arms,
+2. Our code assumes that your robot has been assembled following Trossen Robotics instructions. This allows us to skip calibration, as we use the pre-defined calibration files in `.cache/calibration/aloha_default`. If you replace a motor, make sure you follow the exact instructions from Trossen Robotics.
+
+By running the following code, you can start your first **SAFE** teleoperation:
+```bash
+python lerobot/scripts/control_robot.py teleoperate \
+    --robot-path lerobot/configs/robot/aloha.yaml \
+    --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:
+```bash
+python lerobot/scripts/control_robot.py teleoperate \
+    --robot-path lerobot/configs/robot/aloha.yaml \
+    --robot-overrides max_relative_target=null
+```
+
+## Record a dataset
+
+Once you're familiar with teleoperation, you can record your first dataset with Aloha.
+
+If you want to use the Hugging Face hub features for uploading your dataset and you haven't previously done it, make sure you've logged in using a write-access token, which can be generated from the [Hugging Face settings](https://huggingface.co/settings/tokens):
+```bash
+huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+```
+
+Store your Hugging Face repository name in a variable to run these commands:
+```bash
+HF_USER=$(huggingface-cli whoami | head -n 1)
+echo $HF_USER
+```
+
+Record 2 episodes and upload your dataset to the hub:
+```bash
+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 \
+    --episode-time-s 40 \
+    --reset-time-s 10 \
+    --num-episodes 2 \
+    --push-to-hub 1
+```
+
+## 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
+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
+```
+
+## Replay an episode
+
+**/!\ FOR SAFETY, READ THIS /!\**
+Replay consists in automatically replaying the sequence of actions (i.e. goal positions for your motors) recorded in a given dataset episode. Make sure the current initial position of your robot is similar to the one in your episode, so that your follower arms don't move too fast to go to the first goal positions. For safety, you might want to add `--robot-overrides max_relative_target=5` to your command line as explained above.
+
+Now try to replay the first episode on your robot:
+```bash
+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 ${HF_USER}/aloha_test \
+    --episode 0
+```
+
+## Train a policy
+
+To train a policy to control your robot, use the [`python lerobot/scripts/train.py`](../lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
+```bash
+DATA_DIR=data python lerobot/scripts/train.py \
+  dataset_repo_id=${HF_USER}/aloha_test \
+  policy=act_aloha_real \
+  env=aloha_real \
+  hydra.run.dir=outputs/train/act_aloha_test \
+  hydra.job.name=act_aloha_test \
+  device=cuda \
+  wandb.enable=true
+```
+
+Let's explain it:
+1. We provided the dataset as argument with `dataset_repo_id=${HF_USER}/aloha_test`.
+2. We provided the policy with `policy=act_aloha_real`. This loads configurations from [`lerobot/configs/policy/act_aloha_real.yaml`](../lerobot/configs/policy/act_aloha_real.yaml). Importantly, this policy uses 4 cameras as input `cam_right_wrist`, `cam_left_wrist`, `cam_high`, and `cam_low`.
+3. We provided an environment as argument with `env=aloha_real`. This loads configurations from [`lerobot/configs/env/aloha_real.yaml`](../lerobot/configs/env/aloha_real.yaml). Note: this yaml defines 18 dimensions for the `state_dim` and `action_dim`, corresponding to 18 motors, not 14 motors as used in previous Aloha work. This is because, we include the `shoulder_shadow` and `elbow_shadow` motors for simplicity.
+4. We provided `device=cuda` since we are training on a Nvidia GPU.
+5. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
+6. We added `DATA_DIR=data` to access your dataset stored in your local `data` directory. If you dont provide `DATA_DIR`, your dataset will be downloaded from Hugging Face hub to your cache folder `$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`.
+
+## 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/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 \
+  --episode-time-s 40 \
+  --reset-time-s 10 \
+  --num-episodes 10 \
+  --num-image-writer-processes 1 \
+  -p outputs/train/act_aloha_test/checkpoints/last/pretrained_model
+```
+
+As you can see, it's almost the same command as previously used to record your training dataset. Two things changed:
+1. There is an additional `-p` argument which indicates the path to your policy checkpoint with  (e.g. `-p outputs/train/eval_aloha_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_aloha_test`).
+2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `--repo-id ${HF_USER}/eval_act_aloha_test`).
+3. We use `--num-image-writer-processes 1` instead of the default value (`0`). On our computer, using a dedicated process to write images from the 4 cameras on disk allows to reach constent 30 fps during inference. Feel free to explore different values for `--num-image-writer-processes`.
+
+## More
+
+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 explaination.
+
+If you have any question or need help, please reach out on Discord in the channel `#aloha-arm`.

lerobot/__init__.py

@@ -27,6 +27,9 @@ Example:
         print(lerobot.available_real_world_datasets)
         print(lerobot.available_policies)
         print(lerobot.available_policies_per_env)
+        print(lerobot.available_robots)
+        print(lerobot.available_cameras)
+        print(lerobot.available_motors)
     ```
 
 When implementing a new dataset loadable with LeRobotDataset follow these steps:
@@ -129,13 +132,60 @@ available_real_world_datasets = [
     "lerobot/unitreeh1_rearrange_objects",
     "lerobot/unitreeh1_two_robot_greeting",
     "lerobot/unitreeh1_warehouse",
+    "lerobot/nyu_rot_dataset",
+    "lerobot/utokyo_saytap",
+    "lerobot/imperialcollege_sawyer_wrist_cam",
+    "lerobot/utokyo_xarm_bimanual",
+    "lerobot/tokyo_u_lsmo",
+    "lerobot/utokyo_pr2_opening_fridge",
+    "lerobot/cmu_franka_exploration_dataset",
+    "lerobot/cmu_stretch",
+    "lerobot/asu_table_top",
+    "lerobot/utokyo_pr2_tabletop_manipulation",
+    "lerobot/utokyo_xarm_pick_and_place",
+    "lerobot/ucsd_kitchen_dataset",
+    "lerobot/austin_buds_dataset",
+    "lerobot/dlr_sara_grid_clamp",
+    "lerobot/conq_hose_manipulation",
+    "lerobot/columbia_cairlab_pusht_real",
+    "lerobot/dlr_sara_pour",
+    "lerobot/dlr_edan_shared_control",
+    "lerobot/ucsd_pick_and_place_dataset",
+    "lerobot/berkeley_cable_routing",
+    "lerobot/nyu_franka_play_dataset",
+    "lerobot/austin_sirius_dataset",
+    "lerobot/cmu_play_fusion",
+    "lerobot/berkeley_gnm_sac_son",
+    "lerobot/nyu_door_opening_surprising_effectiveness",
+    "lerobot/berkeley_fanuc_manipulation",
+    "lerobot/jaco_play",
+    "lerobot/viola",
+    "lerobot/kaist_nonprehensile",
+    "lerobot/berkeley_mvp",
+    "lerobot/uiuc_d3field",
+    "lerobot/berkeley_gnm_recon",
+    "lerobot/austin_sailor_dataset",
+    "lerobot/utaustin_mutex",
+    "lerobot/roboturk",
+    "lerobot/stanford_hydra_dataset",
+    "lerobot/berkeley_autolab_ur5",
+    "lerobot/stanford_robocook",
+    "lerobot/toto",
+    "lerobot/fmb",
+    "lerobot/droid_100",
+    "lerobot/berkeley_rpt",
+    "lerobot/stanford_kuka_multimodal_dataset",
+    "lerobot/iamlab_cmu_pickup_insert",
+    "lerobot/taco_play",
+    "lerobot/berkeley_gnm_cory_hall",
+    "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` by their class attribute: `name`.
+# lists all available policies from `lerobot/common/policies`
 available_policies = [
     "act",
     "diffusion",
@@ -143,12 +193,35 @@ available_policies = [
     "vqbet",
 ]
 
+# lists all available robots from `lerobot/common/robot_devices/robots`
+available_robots = [
+    "koch",
+    "koch_bimanual",
+    "aloha",
+    "so100",
+    "moss",
+]
+
+# lists all available cameras from `lerobot/common/robot_devices/cameras`
+available_cameras = [
+    "opencv",
+    "intelrealsense",
+]
+
+# lists all available motors from `lerobot/common/robot_devices/motors`
+available_motors = [
+    "dynamixel",
+    "feetech",
+]
+
 # keys and values refer to yaml files
 available_policies_per_env = {
     "aloha": ["act"],
     "pusht": ["diffusion", "vqbet"],
     "xarm": ["tdmpc"],
-    "dora_aloha_real": ["act_real"],
+    "koch_real": ["act_koch_real"],
+    "aloha_real": ["act_aloha_real"],
+    "dora_aloha_real": ["act_aloha_real"],
 }
 
 env_task_pairs = [(env, task) for env, tasks in available_tasks_per_env.items() for task in tasks]

lerobot/common/datasets/compute_stats.py

@@ -40,6 +40,10 @@ def get_stats_einops_patterns(dataset, num_workers=0):
 
     stats_patterns = {}
     for key, feats_type in dataset.features.items():
+        # NOTE: skip language_instruction embedding in stats computation
+        if key == "language_instruction":
+            continue
+
         # sanity check that tensors are not float64
         assert batch[key].dtype != torch.float64
 
@@ -64,7 +68,7 @@ def get_stats_einops_patterns(dataset, num_workers=0):
     return stats_patterns
 
 
-def compute_stats(dataset, batch_size=32, num_workers=16, max_num_samples=None):
+def compute_stats(dataset, batch_size=8, num_workers=8, max_num_samples=None):
     """Compute mean/std and min/max statistics of all data keys in a LeRobotDataset."""
     if max_num_samples is None:
         max_num_samples = len(dataset)

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,

lerobot/common/datasets/image_writer.py

@@ -0,0 +1,134 @@
+#!/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
+from concurrent.futures import ThreadPoolExecutor, wait
+from pathlib import Path
+
+import torch
+import tqdm
+from PIL import Image
+
+DEFAULT_IMAGE_PATH = "{image_key}/episode_{episode_index:06d}/frame_{frame_index:06d}.png"
+
+
+def safe_stop_image_writer(func):
+    def wrapper(*args, **kwargs):
+        try:
+            return func(*args, **kwargs)
+        except Exception as e:
+            dataset = kwargs.get("dataset", None)
+            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
+
+
+class ImageWriter:
+    """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, write_dir: Path, num_processes: int = 0, num_threads: int = 1):
+        self.dir = write_dir
+        self.dir.mkdir(parents=True, exist_ok=True)
+        self.image_path = DEFAULT_IMAGE_PATH
+        self.num_processes = num_processes
+        self.num_threads = self.num_threads_per_process = num_threads
+
+        if self.num_processes <= 0:
+            self.type = "threads"
+            self.threads = ThreadPoolExecutor(max_workers=self.num_threads)
+            self.futures = []
+        else:
+            self.type = "processes"
+            self.num_threads_per_process = self.num_threads
+            self.image_queue = multiprocessing.Queue()
+            self.processes: list[multiprocessing.Process] = []
+            for _ in range(num_processes):
+                process = multiprocessing.Process(target=self._loop_to_save_images_in_threads)
+                process.start()
+                self.processes.append(process)
+
+    def _loop_to_save_images_in_threads(self) -> None:
+        with ThreadPoolExecutor(max_workers=self.num_threads) as executor:
+            futures = []
+            while True:
+                frame_data = self.image_queue.get()
+                if frame_data is None:
+                    break
+
+                image, file_path = frame_data
+                futures.append(executor.submit(self._save_image, image, file_path))
+
+            with tqdm.tqdm(total=len(futures), desc="Writing images") as progress_bar:
+                wait(futures)
+                progress_bar.update(len(futures))
+
+    def async_save_image(self, image: torch.Tensor, file_path: Path) -> None:
+        """Save an image asynchronously using threads or processes."""
+        if self.type == "threads":
+            self.futures.append(self.threads.submit(self._save_image, image, file_path))
+        else:
+            self.image_queue.put((image, file_path))
+
+    def _save_image(self, image: torch.Tensor, file_path: Path) -> None:
+        img = Image.fromarray(image.numpy())
+        img.save(str(file_path), quality=100)
+
+    def get_image_file_path(self, episode_index: int, image_key: str, frame_index: int) -> Path:
+        fpath = self.image_path.format(
+            image_key=image_key, episode_index=episode_index, frame_index=frame_index
+        )
+        return self.dir / fpath
+
+    def get_episode_dir(self, episode_index: int, image_key: str) -> Path:
+        return self.get_image_file_path(
+            episode_index=episode_index, image_key=image_key, frame_index=0
+        ).parent
+
+    def stop(self, timeout=20) -> None:
+        """Stop the image writer, waiting for all processes or threads to finish."""
+        if self.type == "threads":
+            with tqdm.tqdm(total=len(self.futures), desc="Writing images") as progress_bar:
+                wait(self.futures, timeout=timeout)
+                progress_bar.update(len(self.futures))
+        else:
+            self._stop_processes(timeout)
+
+    def _stop_processes(self, timeout) -> None:
+        for _ in self.processes:
+            self.image_queue.put(None)
+
+        for process in self.processes:
+            process.join(timeout=timeout)
+
+        if process.is_alive():
+            process.terminate()
+
+        self.image_queue.close()
+        self.image_queue.join_thread()

lerobot/common/datasets/lerobot_dataset.py

@@ -13,63 +13,312 @@
 # 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 os
+import shutil
+from functools import cached_property
 from pathlib import Path
 from typing import Callable
 
 import datasets
+import pyarrow.parquet as pq
 import torch
 import torch.utils
+from datasets import load_dataset
+from huggingface_hub import snapshot_download, upload_folder
 
-from lerobot.common.datasets.compute_stats import aggregate_stats
+from lerobot.common.datasets.compute_stats import aggregate_stats, compute_stats
+from lerobot.common.datasets.image_writer import ImageWriter
 from lerobot.common.datasets.utils import (
-    calculate_episode_data_index,
-    load_episode_data_index,
-    load_hf_dataset,
+    EPISODES_PATH,
+    INFO_PATH,
+    STATS_PATH,
+    TASKS_PATH,
+    _get_info_from_robot,
+    append_jsonl,
+    check_delta_timestamps,
+    check_timestamps_sync,
+    check_version_compatibility,
+    create_branch,
+    create_empty_dataset_info,
+    get_delta_indices,
+    get_episode_data_index,
+    get_hub_safe_version,
+    hf_transform_to_torch,
+    load_episode_dicts,
     load_info,
-    load_previous_and_future_frames,
     load_stats,
-    load_videos,
-    reset_episode_index,
+    load_tasks,
+    write_json,
+    write_stats,
 )
-from lerobot.common.datasets.video_utils import VideoFrame, load_from_videos
+from lerobot.common.datasets.video_utils import (
+    VideoFrame,
+    decode_video_frames_torchvision,
+    encode_video_frames,
+)
+from lerobot.common.robot_devices.robots.utils import Robot
 
 # For maintainers, see lerobot/common/datasets/push_dataset_to_hub/CODEBASE_VERSION.md
-CODEBASE_VERSION = "v1.6"
-DATA_DIR = Path(os.environ["DATA_DIR"]) if "DATA_DIR" in os.environ else None
+CODEBASE_VERSION = "v2.0"
+LEROBOT_HOME = Path(os.getenv("LEROBOT_HOME", "~/.cache/huggingface/lerobot")).expanduser()
 
 
 class LeRobotDataset(torch.utils.data.Dataset):
     def __init__(
         self,
         repo_id: str,
-        root: Path | None = DATA_DIR,
-        split: str = "train",
+        root: Path | None = None,
+        episodes: list[int] | None = None,
         image_transforms: Callable | None = None,
         delta_timestamps: dict[list[float]] | None = None,
+        tolerance_s: float = 1e-4,
+        download_videos: bool = True,
+        local_files_only: bool = False,
         video_backend: str | None = None,
+        image_writer: ImageWriter | None = None,
     ):
+        """LeRobotDataset encapsulates 3 main things:
+            - metadata:
+                - info contains various information about the dataset like shapes, keys, fps etc.
+                - stats stores the dataset statistics of the different modalities for normalization
+                - tasks contains the prompts for each task of the dataset, which can be used for
+                  task-conditionned training.
+            - hf_dataset (from datasets.Dataset), which will read any values from parquet files.
+            - (optional) videos from which frames are loaded to be synchronous with data from parquet files.
+
+        3 modes are available for this class, depending on 3 different use cases:
+
+        1. Your dataset already exists on the Hugging Face Hub at the address
+        https://huggingface.co/datasets/{repo_id} and is not on your local disk in the 'root' folder:
+            Instantiating this class with this 'repo_id' will download the dataset from that address and load
+            it, pending your dataset is compliant with codebase_version v2.0. If your dataset has been created
+            before this new format, you will be prompted to convert it using our conversion script from v1.6
+            to v2.0, which you can find at lerobot/common/datasets/v2/convert_dataset_v1_to_v2.py.
+
+        2. Your dataset already exists on your local disk in the 'root' folder:
+            This is typically the case when you recorded your dataset locally and you may or may not have
+            pushed it to the hub yet. Instantiating this class with 'root' will load your dataset directly
+            from disk. This can happen while you're offline (no internet connection).
+
+        3. Your dataset doesn't already exists (either on local disk or on the Hub):
+            [TODO(aliberts): add classmethod for this case?]
+
+
+        In terms of files, a typical LeRobotDataset looks like this from its root path:
+        .
+        ├── data
+        │   ├── chunk-000
+        │   │   ├── episode_000000.parquet
+        │   │   ├── episode_000001.parquet
+        │   │   ├── episode_000002.parquet
+        │   │   └── ...
+        │   ├── chunk-001
+        │   │   ├── episode_001000.parquet
+        │   │   ├── episode_001001.parquet
+        │   │   ├── episode_001002.parquet
+        │   │   └── ...
+        │   └── ...
+        ├── meta
+        │   ├── episodes.jsonl
+        │   ├── info.json
+        │   ├── stats.json
+        │   └── tasks.jsonl
+        └── videos (optional)
+            ├── chunk-000
+            │   ├── observation.images.laptop
+            │   │   ├── episode_000000.mp4
+            │   │   ├── episode_000001.mp4
+            │   │   ├── episode_000002.mp4
+            │   │   └── ...
+            │   ├── observation.images.phone
+            │   │   ├── episode_000000.mp4
+            │   │   ├── episode_000001.mp4
+            │   │   ├── episode_000002.mp4
+            │   │   └── ...
+            ├── chunk-001
+            └── ...
+
+        Note that this file-based structure is designed to be as versatile as possible. The files are split by
+        episodes which allows a more granular control over which episodes one wants to use and download. The
+        structure of the dataset is entirely described in the info.json file, which can be easily downloaded
+        or viewed directly on the hub before downloading any actual data. The type of files used are very
+        simple and do not need complex tools to be read, it only uses .parquet, .json and .mp4 files (and .md
+        for the README).
+
+        Args:
+            repo_id (str): This is the repo id that will be used to fetch the dataset. Locally, the dataset
+                will be stored under root/repo_id.
+            root (Path | None, optional): Local directory to use for downloading/writing files. You can also
+                set the LEROBOT_HOME environment variable to point to a different location. Defaults to
+                '~/.cache/huggingface/lerobot'.
+            episodes (list[int] | None, optional): If specified, this will only load episodes specified by
+                their episode_index in this list. Defaults to None.
+            image_transforms (Callable | None, optional): You can pass standard v2 image transforms from
+                torchvision.transforms.v2 here which will be applied to visual modalities (whether they come
+                from videos or images). Defaults to None.
+            delta_timestamps (dict[list[float]] | None, optional): _description_. Defaults to None.
+            tolerance_s (float, optional): Tolerance in seconds used to ensure data timestamps are actually in
+                sync with the fps value. It is used at the init of the dataset to make sure that each
+                timestamps is separated to the next by 1/fps +/- tolerance_s. This also applies to frames
+                decoded from video files. It is also used to check that `delta_timestamps` (when provided) are
+                multiples of 1/fps. Defaults to 1e-4.
+            download_videos (bool, optional): Flag to download the videos. Note that when set to True but the
+                video files are already present on local disk, they won't be downloaded again. Defaults to
+                True.
+            video_backend (str | None, optional): Video backend to use for decoding videos. There is currently
+                a single option which is the pyav decoder used by Torchvision. Defaults to pyav.
+        """
         super().__init__()
         self.repo_id = repo_id
-        self.root = root
-        self.split = split
+        self.root = root if root is not None else LEROBOT_HOME / repo_id
         self.image_transforms = image_transforms
         self.delta_timestamps = delta_timestamps
-        # load data from hub or locally when root is provided
+        self.episodes = episodes
+        self.tolerance_s = tolerance_s
+        self.video_backend = video_backend if video_backend is not None else "pyav"
+        self.delta_indices = None
+        self.local_files_only = local_files_only
+        self.consolidated = True
+
+        # Unused attributes
+        self.image_writer = None
+        self.episode_buffer = {}
+
+        # Load metadata
+        self.root.mkdir(exist_ok=True, parents=True)
+        self.pull_from_repo(allow_patterns="meta/")
+        self.info = load_info(self.root)
+        self.stats = load_stats(self.root)
+        self.tasks = load_tasks(self.root)
+        self.episode_dicts = load_episode_dicts(self.root)
+
+        # Check version
+        check_version_compatibility(self.repo_id, self._version, CODEBASE_VERSION)
+
+        # Load actual data
+        self.download_episodes(download_videos)
+        self.hf_dataset = self.load_hf_dataset()
+        self.episode_data_index = get_episode_data_index(self.episodes, self.episode_dicts)
+
+        # Check timestamps
+        check_timestamps_sync(self.hf_dataset, self.episode_data_index, self.fps, self.tolerance_s)
+
+        # Setup delta_indices
+        if self.delta_timestamps is not None:
+            check_delta_timestamps(self.delta_timestamps, self.fps, self.tolerance_s)
+            self.delta_indices = get_delta_indices(self.delta_timestamps, self.fps)
+
+        # TODO(aliberts):
+        # - [X] Move delta_timestamp logic outside __get_item__
+        # - [X] Update __get_item__
+        # - [/] Add doc
+        # - [ ] Add self.add_frame()
+        # - [ ] Add self.consolidate() for:
+        #     - [X] Check timestamps sync
+        #     - [ ] Sanity checks (episodes num, shapes, files, etc.)
+        #     - [ ] Update episode_index (arg update=True)
+        #     - [ ] Update info.json (arg update=True)
+
+    @cached_property
+    def _hub_version(self) -> str | None:
+        return None if self.local_files_only else get_hub_safe_version(self.repo_id, CODEBASE_VERSION)
+
+    @property
+    def _version(self) -> str:
+        """Codebase version used to create this dataset."""
+        return self.info["codebase_version"]
+
+    def push_to_hub(self, push_videos: bool = True) -> None:
+        if not self.consolidated:
+            raise RuntimeError(
+                "You are trying to upload to the hub a LeRobotDataset that has not been consolidated yet."
+                "Please call the dataset 'consolidate()' method first."
+            )
+        ignore_patterns = ["images/"]
+        if not push_videos:
+            ignore_patterns.append("videos/")
+
+        upload_folder(
+            repo_id=self.repo_id,
+            folder_path=self.root,
+            repo_type="dataset",
+            ignore_patterns=ignore_patterns,
+        )
+        create_branch(repo_id=self.repo_id, branch=CODEBASE_VERSION, repo_type="dataset")
+
+    def pull_from_repo(
+        self,
+        allow_patterns: list[str] | str | None = None,
+        ignore_patterns: list[str] | str | None = None,
+    ) -> None:
+        snapshot_download(
+            self.repo_id,
+            repo_type="dataset",
+            revision=self._hub_version,
+            local_dir=self.root,
+            allow_patterns=allow_patterns,
+            ignore_patterns=ignore_patterns,
+            local_files_only=self.local_files_only,
+        )
+
+    def download_episodes(self, download_videos: bool = True) -> None:
+        """Downloads the dataset from the given 'repo_id' at the provided version. If 'episodes' is given, this
+        will only download those episodes (selected by their episode_index). If 'episodes' is None, the whole
+        dataset will be downloaded. Thanks to the behavior of snapshot_download, if the files are already present
+        in 'local_dir', they won't be downloaded again.
+        """
         # TODO(rcadene, aliberts): implement faster transfer
         # https://huggingface.co/docs/huggingface_hub/en/guides/download#faster-downloads
-        self.hf_dataset = load_hf_dataset(repo_id, CODEBASE_VERSION, root, split)
-        if split == "train":
-            self.episode_data_index = load_episode_data_index(repo_id, CODEBASE_VERSION, root)
+        files = None
+        ignore_patterns = None if download_videos else "videos/"
+        if self.episodes is not None:
+            files = [str(self.get_data_file_path(ep_idx)) for ep_idx in self.episodes]
+            if len(self.video_keys) > 0 and download_videos:
+                video_files = [
+                    str(self.get_video_file_path(ep_idx, vid_key))
+                    for vid_key in self.video_keys
+                    for ep_idx in self.episodes
+                ]
+                files += video_files
+
+        self.pull_from_repo(allow_patterns=files, ignore_patterns=ignore_patterns)
+
+    def load_hf_dataset(self) -> datasets.Dataset:
+        """hf_dataset contains all the observations, states, actions, rewards, etc."""
+        if self.episodes is None:
+            path = str(self.root / "data")
+            hf_dataset = load_dataset("parquet", data_dir=path, split="train")
         else:
-            self.episode_data_index = calculate_episode_data_index(self.hf_dataset)
-            self.hf_dataset = reset_episode_index(self.hf_dataset)
-        self.stats = load_stats(repo_id, CODEBASE_VERSION, root)
-        self.info = load_info(repo_id, CODEBASE_VERSION, root)
-        if self.video:
-            self.videos_dir = load_videos(repo_id, CODEBASE_VERSION, root)
-            self.video_backend = video_backend if video_backend is not None else "pyav"
+            files = [str(self.root / self.get_data_file_path(ep_idx)) for ep_idx in self.episodes]
+            hf_dataset = load_dataset("parquet", data_files=files, split="train")
+
+        hf_dataset.set_transform(hf_transform_to_torch)
+        return hf_dataset
+
+    def get_data_file_path(self, ep_index: int) -> Path:
+        ep_chunk = self.get_episode_chunk(ep_index)
+        fpath = self.data_path.format(episode_chunk=ep_chunk, episode_index=ep_index)
+        return Path(fpath)
+
+    def get_video_file_path(self, ep_index: int, vid_key: str) -> Path:
+        ep_chunk = self.get_episode_chunk(ep_index)
+        fpath = self.videos_path.format(episode_chunk=ep_chunk, video_key=vid_key, episode_index=ep_index)
+        return Path(fpath)
+
+    def get_episode_chunk(self, ep_index: int) -> int:
+        return ep_index // self.chunks_size
+
+    @property
+    def data_path(self) -> str:
+        """Formattable string for the parquet files."""
+        return self.info["data_path"]
+
+    @property
+    def videos_path(self) -> str | None:
+        """Formattable string for the video files."""
+        return self.info["videos"]["videos_path"] if len(self.video_keys) > 0 else None
 
     @property
     def fps(self) -> int:
@@ -77,140 +326,495 @@ class LeRobotDataset(torch.utils.data.Dataset):
         return self.info["fps"]
 
     @property
-    def video(self) -> bool:
-        """Returns True if this dataset loads video frames from mp4 files.
-        Returns False if it only loads images from png files.
-        """
-        return self.info.get("video", False)
+    def keys(self) -> list[str]:
+        """Keys to access non-image data (state, actions etc.)."""
+        return self.info["keys"]
 
     @property
-    def features(self) -> datasets.Features:
-        return self.hf_dataset.features
+    def image_keys(self) -> list[str]:
+        """Keys to access visual modalities stored as images."""
+        return self.info["image_keys"]
+
+    @property
+    def video_keys(self) -> list[str]:
+        """Keys to access visual modalities stored as videos."""
+        return self.info["video_keys"]
 
     @property
     def camera_keys(self) -> list[str]:
-        """Keys to access image and video stream from cameras."""
-        keys = []
-        for key, feats in self.hf_dataset.features.items():
-            if isinstance(feats, (datasets.Image, VideoFrame)):
-                keys.append(key)
-        return keys
+        """Keys to access visual modalities (regardless of their storage method)."""
+        return self.image_keys + self.video_keys
 
     @property
-    def video_frame_keys(self) -> list[str]:
-        """Keys to access video frames that requires to be decoded into images.
-
-        Note: It is empty if the dataset contains images only,
-        or equal to `self.cameras` if the dataset contains videos only,
-        or can even be a subset of `self.cameras` in a case of a mixed image/video dataset.
-        """
-        video_frame_keys = []
-        for key, feats in self.hf_dataset.features.items():
-            if isinstance(feats, VideoFrame):
-                video_frame_keys.append(key)
-        return video_frame_keys
+    def names(self) -> dict[list[str]]:
+        """Names of the various dimensions of vector modalities."""
+        return self.info["names"]
 
     @property
     def num_samples(self) -> int:
-        """Number of samples/frames."""
+        """Number of samples/frames in selected episodes."""
         return len(self.hf_dataset)
 
     @property
     def num_episodes(self) -> int:
-        """Number of episodes."""
-        return len(self.hf_dataset.unique("episode_index"))
+        """Number of episodes selected."""
+        return len(self.episodes) if self.episodes is not None else self.total_episodes
 
     @property
-    def tolerance_s(self) -> float:
-        """Tolerance in seconds used to discard loaded frames when their timestamps
-        are not close enough from the requested frames. It is only used when `delta_timestamps`
-        is provided or when loading video frames from mp4 files.
+    def total_episodes(self) -> int:
+        """Total number of episodes available."""
+        return self.info["total_episodes"]
+
+    @property
+    def total_frames(self) -> int:
+        """Total number of frames saved in this dataset."""
+        return self.info["total_frames"]
+
+    @property
+    def total_tasks(self) -> int:
+        """Total number of different tasks performed in this dataset."""
+        return self.info["total_tasks"]
+
+    @property
+    def total_chunks(self) -> int:
+        """Total number of chunks (groups of episodes)."""
+        return self.info["total_chunks"]
+
+    @property
+    def chunks_size(self) -> int:
+        """Max number of episodes per chunk."""
+        return self.info["chunks_size"]
+
+    @property
+    def shapes(self) -> dict:
+        """Shapes for the different features."""
+        return self.info["shapes"]
+
+    @property
+    def features(self) -> datasets.Features:
+        """Features of the hf_dataset."""
+        if self.hf_dataset is not None:
+            return self.hf_dataset.features
+        elif self.episode_buffer is None:
+            raise NotImplementedError(
+                "Dataset features must be infered from an existing hf_dataset or episode_buffer."
+            )
+
+        features = {}
+        for key in self.episode_buffer:
+            if key in ["episode_index", "frame_index", "index", "task_index"]:
+                features[key] = datasets.Value(dtype="int64")
+            elif key in ["next.done", "next.success"]:
+                features[key] = datasets.Value(dtype="bool")
+            elif key in ["timestamp", "next.reward"]:
+                features[key] = datasets.Value(dtype="float32")
+            elif key in self.image_keys:
+                features[key] = datasets.Image()
+            elif key in self.keys:
+                features[key] = datasets.Sequence(
+                    length=self.shapes[key], feature=datasets.Value(dtype="float32")
+                )
+
+        return datasets.Features(features)
+
+    @property
+    def task_to_task_index(self) -> dict:
+        return {task: task_idx for task_idx, task in self.tasks.items()}
+
+    def get_task_index(self, task: str) -> int:
         """
-        # 1e-4 to account for possible numerical error
-        return 1 / self.fps - 1e-4
+        Given a task in natural language, returns its task_index if the task already exists in the dataset,
+        otherwise creates a new task_index.
+        """
+        task_index = self.task_to_task_index.get(task, None)
+        return task_index if task_index is not None else self.total_tasks
+
+    def current_episode_index(self, idx: int) -> int:
+        episode_index = self.hf_dataset["episode_index"][idx]
+        if self.episodes is not None:
+            # get episode_index from selected episodes
+            episode_index = self.episodes.index(episode_index)
+
+        return episode_index
+
+    def episode_length(self, episode_index) -> int:
+        """Number of samples/frames for given episode."""
+        return self.info["episodes"][episode_index]["length"]
+
+    def _get_query_indices(self, idx: int, ep_idx: int) -> tuple[dict[str, list[int | bool]]]:
+        ep_start = self.episode_data_index["from"][ep_idx]
+        ep_end = self.episode_data_index["to"][ep_idx]
+        query_indices = {
+            key: [max(ep_start.item(), min(ep_end.item() - 1, idx + delta)) for delta in delta_idx]
+            for key, delta_idx in self.delta_indices.items()
+        }
+        padding = {  # Pad values outside of current episode range
+            f"{key}_is_pad": torch.BoolTensor(
+                [(idx + delta < ep_start.item()) | (idx + delta >= ep_end.item()) for delta in delta_idx]
+            )
+            for key, delta_idx in self.delta_indices.items()
+        }
+        return query_indices, padding
+
+    def _get_query_timestamps(
+        self,
+        current_ts: float,
+        query_indices: dict[str, list[int]] | None = None,
+    ) -> dict[str, list[float]]:
+        query_timestamps = {}
+        for key in self.video_keys:
+            if query_indices is not None and key in query_indices:
+                timestamps = self.hf_dataset.select(query_indices[key])["timestamp"]
+                query_timestamps[key] = torch.stack(timestamps).tolist()
+            else:
+                query_timestamps[key] = [current_ts]
+
+        return query_timestamps
+
+    def _query_hf_dataset(self, query_indices: dict[str, list[int]]) -> dict:
+        return {
+            key: torch.stack(self.hf_dataset.select(q_idx)[key])
+            for key, q_idx in query_indices.items()
+            if key not in self.video_keys
+        }
+
+    def _query_videos(self, query_timestamps: dict[str, list[float]], ep_idx: int) -> dict:
+        """Note: When using data workers (e.g. DataLoader with num_workers>0), do not call this function
+        in the main process (e.g. by using a second Dataloader with num_workers=0). It will result in a
+        Segmentation Fault. This probably happens because a memory reference to the video loader is created in
+        the main process and a subprocess fails to access it.
+        """
+        item = {}
+        for vid_key, query_ts in query_timestamps.items():
+            video_path = self.root / self.get_video_file_path(ep_idx, vid_key)
+            frames = decode_video_frames_torchvision(
+                video_path, query_ts, self.tolerance_s, self.video_backend
+            )
+            item[vid_key] = frames.squeeze(0)
+
+        return item
+
+    def _add_padding_keys(self, item: dict, padding: dict[str, list[bool]]) -> dict:
+        for key, val in padding.items():
+            item[key] = torch.BoolTensor(val)
+        return item
 
     def __len__(self):
         return self.num_samples
 
-    def __getitem__(self, idx):
+    def __getitem__(self, idx) -> dict:
         item = self.hf_dataset[idx]
+        ep_idx = item["episode_index"].item()
 
-        if self.delta_timestamps is not None:
-            item = load_previous_and_future_frames(
-                item,
-                self.hf_dataset,
-                self.episode_data_index,
-                self.delta_timestamps,
-                self.tolerance_s,
-            )
+        query_indices = None
+        if self.delta_indices is not None:
+            current_ep_idx = self.episodes.index(ep_idx) if self.episodes is not None else ep_idx
+            query_indices, padding = self._get_query_indices(idx, current_ep_idx)
+            query_result = self._query_hf_dataset(query_indices)
+            item = {**item, **padding}
+            for key, val in query_result.items():
+                item[key] = val
 
-        if self.video:
-            item = load_from_videos(
-                item,
-                self.video_frame_keys,
-                self.videos_dir,
-                self.tolerance_s,
-                self.video_backend,
-            )
+        if len(self.video_keys) > 0:
+            current_ts = item["timestamp"].item()
+            query_timestamps = self._get_query_timestamps(current_ts, query_indices)
+            video_frames = self._query_videos(query_timestamps, ep_idx)
+            item = {**video_frames, **item}
 
         if self.image_transforms is not None:
-            for cam in self.camera_keys:
+            image_keys = self.camera_keys
+            for cam in image_keys:
                 item[cam] = self.image_transforms(item[cam])
 
         return item
 
     def __repr__(self):
         return (
-            f"{self.__class__.__name__}(\n"
+            f"{self.__class__.__name__}\n"
             f"  Repository ID: '{self.repo_id}',\n"
-            f"  Split: '{self.split}',\n"
-            f"  Number of Samples: {self.num_samples},\n"
-            f"  Number of Episodes: {self.num_episodes},\n"
-            f"  Type: {'video (.mp4)' if self.video else 'image (.png)'},\n"
-            f"  Recorded Frames per Second: {self.fps},\n"
-            f"  Camera Keys: {self.camera_keys},\n"
-            f"  Video Frame Keys: {self.video_frame_keys if self.video else 'N/A'},\n"
-            f"  Transformations: {self.image_transforms},\n"
-            f"  Codebase Version: {self.info.get('codebase_version', '< v1.6')},\n"
-            f")"
+            f"  Selected episodes: {self.episodes},\n"
+            f"  Number of selected episodes: {self.num_episodes},\n"
+            f"  Number of selected samples: {self.num_samples},\n"
+            f"\n{json.dumps(self.info, indent=4)}\n"
         )
 
-    @classmethod
-    def from_preloaded(
-        cls,
-        repo_id: str = "from_preloaded",
-        root: Path | None = None,
-        split: str = "train",
-        transform: callable = None,
-        delta_timestamps: dict[list[float]] | None = None,
-        # additional preloaded attributes
-        hf_dataset=None,
-        episode_data_index=None,
-        stats=None,
-        info=None,
-        videos_dir=None,
-        video_backend=None,
-    ) -> "LeRobotDataset":
-        """Create a LeRobot Dataset from existing data and attributes instead of loading from the filesystem.
+    def _create_episode_buffer(self, episode_index: int | None = None) -> dict:
+        # TODO(aliberts): Handle resume
+        return {
+            "size": 0,
+            "episode_index": self.total_episodes if episode_index is None else episode_index,
+            "task_index": None,
+            "frame_index": [],
+            "timestamp": [],
+            "next.done": [],
+            **{key: [] for key in self.keys},
+            **{key: [] for key in self.image_keys},
+        }
 
-        It is especially useful when converting raw data into LeRobotDataset before saving the dataset
-        on the filesystem or uploading to the hub.
-
-        Note: Meta-data attributes like `repo_id`, `version`, `root`, etc are optional and potentially
-        meaningless depending on the downstream usage of the return dataset.
+    def add_frame(self, frame: dict) -> None:
         """
-        # create an empty object of type LeRobotDataset
+        This function only adds the frame to the episode_buffer. Apart from images — which are written in a
+        temporary directory — nothing is written to disk. To save those frames, the 'add_episode()' method
+        then needs to be called.
+        """
+        frame_index = self.episode_buffer["size"]
+        self.episode_buffer["frame_index"].append(frame_index)
+        self.episode_buffer["timestamp"].append(frame_index / self.fps)
+        self.episode_buffer["next.done"].append(False)
+
+        # Save all observed modalities except images
+        for key in self.keys:
+            self.episode_buffer[key].append(frame[key])
+
+        self.episode_buffer["size"] += 1
+
+        if self.image_writer is None:
+            return
+
+        # Save images
+        for cam_key in self.camera_keys:
+            img_path = self.image_writer.get_image_file_path(
+                episode_index=self.episode_buffer["episode_index"], image_key=cam_key, frame_index=frame_index
+            )
+            if frame_index == 0:
+                img_path.parent.mkdir(parents=True, exist_ok=True)
+
+            self.image_writer.async_save_image(
+                image=frame[cam_key],
+                file_path=img_path,
+            )
+            if cam_key in self.image_keys:
+                self.episode_buffer[cam_key].append(str(img_path))
+
+    def add_episode(self, task: str, encode_videos: bool = False) -> None:
+        """
+        This will save to disk the current episode in self.episode_buffer. Note that since it affects files on
+        disk, it sets self.consolidated to False to ensure proper consolidation later on before uploading to
+        the hub.
+
+        Use 'encode_videos' if you want to encode videos during the saving of each episode. Otherwise,
+        you can do it later with dataset.consolidate(). This is to give more flexibility on when to spend
+        time for video encoding.
+        """
+        episode_length = self.episode_buffer.pop("size")
+        episode_index = self.episode_buffer["episode_index"]
+        if episode_index != self.total_episodes:
+            # TODO(aliberts): Add option to use existing episode_index
+            raise NotImplementedError()
+
+        task_index = self.get_task_index(task)
+        self.episode_buffer["next.done"][-1] = True
+
+        for key in self.episode_buffer:
+            if key in self.image_keys:
+                continue
+            if key in self.keys:
+                self.episode_buffer[key] = torch.stack(self.episode_buffer[key])
+            elif key == "episode_index":
+                self.episode_buffer[key] = torch.full((episode_length,), episode_index)
+            elif key == "task_index":
+                self.episode_buffer[key] = torch.full((episode_length,), task_index)
+            else:
+                self.episode_buffer[key] = torch.tensor(self.episode_buffer[key])
+
+        self.episode_buffer["index"] = torch.arange(self.total_frames, self.total_frames + episode_length)
+        self._save_episode_to_metadata(episode_index, episode_length, task, task_index)
+        self._save_episode_table(episode_index)
+
+        if encode_videos and len(self.video_keys) > 0:
+            self.encode_videos()
+
+        # Reset the buffer
+        self.episode_buffer = self._create_episode_buffer()
+        self.consolidated = False
+
+    def _save_episode_table(self, episode_index: int) -> None:
+        ep_dataset = datasets.Dataset.from_dict(self.episode_buffer, features=self.features, split="train")
+        ep_table = ep_dataset._data.table
+        ep_data_path = self.root / self.get_data_file_path(ep_index=episode_index)
+        ep_data_path.parent.mkdir(parents=True, exist_ok=True)
+        pq.write_table(ep_table, ep_data_path)
+
+    def _save_episode_to_metadata(
+        self, episode_index: int, episode_length: int, task: str, task_index: int
+    ) -> None:
+        self.info["total_episodes"] += 1
+        self.info["total_frames"] += episode_length
+
+        if task_index not in self.tasks:
+            self.info["total_tasks"] += 1
+            self.tasks[task_index] = task
+            task_dict = {
+                "task_index": task_index,
+                "task": task,
+            }
+            append_jsonl(task_dict, self.root / TASKS_PATH)
+
+        chunk = self.get_episode_chunk(episode_index)
+        if chunk >= self.total_chunks:
+            self.info["total_chunks"] += 1
+
+        self.info["splits"] = {"train": f"0:{self.info['total_episodes']}"}
+        self.info["total_videos"] += len(self.video_keys)
+        write_json(self.info, self.root / INFO_PATH)
+
+        episode_dict = {
+            "episode_index": episode_index,
+            "tasks": [task],
+            "length": episode_length,
+        }
+        self.episode_dicts.append(episode_dict)
+        append_jsonl(episode_dict, self.root / EPISODES_PATH)
+
+    def clear_episode_buffer(self) -> None:
+        episode_index = self.episode_buffer["episode_index"]
+        if self.image_writer is not None:
+            for cam_key in self.camera_keys:
+                img_dir = self.image_writer.get_episode_dir(episode_index, cam_key)
+                if img_dir.is_dir():
+                    shutil.rmtree(img_dir)
+
+        # Reset the buffer
+        self.episode_buffer = self._create_episode_buffer()
+
+    def start_image_writter(self, num_processes: int = 0, num_threads: int = 1) -> None:
+        if isinstance(self.image_writer, ImageWriter):
+            logging.warning(
+                "You are starting a new ImageWriter that is replacing an already exising one in the dataset."
+            )
+
+        self.image_writer = ImageWriter(
+            write_dir=self.root / "images",
+            num_processes=num_processes,
+            num_threads=num_threads,
+        )
+
+    def stop_image_writter(self) -> None:
+        """
+        Whenever wrapping this dataset inside a parallelized DataLoader, this needs to be called first to
+        remove the image_write in order for the LeRobotDataset object to be pickleable and parallelized.
+        """
+        if self.image_writer is not None:
+            self.image_writer.stop()
+            self.image_writer = None
+
+    def encode_videos(self) -> None:
+        # Use ffmpeg to convert frames stored as png into mp4 videos
+        for episode_index in range(self.num_episodes):
+            for key in self.video_keys:
+                # TODO: create video_buffer to store the state of encoded/unencoded videos and remove the need
+                # to call self.image_writer here
+                tmp_imgs_dir = self.image_writer.get_episode_dir(episode_index, key)
+                video_path = self.root / self.get_video_file_path(episode_index, key)
+                if video_path.is_file():
+                    # 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, self.fps, overwrite=True)
+
+    def consolidate(self, run_compute_stats: bool = True, keep_image_files: bool = False) -> None:
+        self.hf_dataset = self.load_hf_dataset()
+        self.episode_data_index = get_episode_data_index(self.episodes, self.episode_dicts)
+        check_timestamps_sync(self.hf_dataset, self.episode_data_index, self.fps, self.tolerance_s)
+
+        if len(self.video_keys) > 0:
+            self.encode_videos()
+
+        if not keep_image_files and self.image_writer is not None:
+            shutil.rmtree(self.image_writer.dir)
+
+        if run_compute_stats:
+            self.stop_image_writter()
+            self.stats = compute_stats(self)
+            write_stats(self.stats, self.root / STATS_PATH)
+            self.consolidated = True
+        else:
+            logging.warning(
+                "Skipping computation of the dataset statistics, dataset is not fully consolidated."
+            )
+
+        # TODO(aliberts)
+        # - [ ] add video info in info.json
+        # Sanity checks:
+        # - [ ] shapes
+        # - [ ] ep_lenghts
+        # - [ ] number of files
+
+    @classmethod
+    def create(
+        cls,
+        repo_id: str,
+        fps: int,
+        root: Path | None = None,
+        robot: Robot | None = None,
+        robot_type: str | None = None,
+        keys: list[str] | None = None,
+        image_keys: list[str] | None = None,
+        video_keys: list[str] = None,
+        shapes: dict | None = None,
+        names: dict | None = None,
+        tolerance_s: float = 1e-4,
+        image_writer_processes: int = 0,
+        image_writer_threads_per_camera: int = 0,
+        use_videos: bool = True,
+        video_backend: str | None = None,
+    ) -> "LeRobotDataset":
+        """Create a LeRobot Dataset from scratch in order to record data."""
         obj = cls.__new__(cls)
         obj.repo_id = repo_id
-        obj.root = root
-        obj.split = split
-        obj.image_transforms = transform
-        obj.delta_timestamps = delta_timestamps
-        obj.hf_dataset = hf_dataset
-        obj.episode_data_index = episode_data_index
-        obj.stats = stats
-        obj.info = info if info is not None else {}
-        obj.videos_dir = videos_dir
+        obj.root = root if root is not None else LEROBOT_HOME / repo_id
+        obj.tolerance_s = tolerance_s
+        obj.image_writer = None
+
+        if robot is not None:
+            robot_type, keys, image_keys, video_keys, shapes, names = _get_info_from_robot(robot, use_videos)
+            if not all(cam.fps == fps for cam in robot.cameras.values()):
+                logging.warning(
+                    f"Some cameras in your {robot.robot_type} robot don't have an fps matching the fps of your dataset."
+                    "In this case, frames from lower fps cameras will be repeated to fill in the blanks"
+                )
+            if len(robot.cameras) > 0 and (image_writer_processes or image_writer_threads_per_camera):
+                obj.start_image_writter(
+                    image_writer_processes, image_writer_threads_per_camera * robot.num_cameras
+                )
+        elif (
+            robot_type is None
+            or keys is None
+            or image_keys is None
+            or video_keys is None
+            or shapes is None
+            or names is None
+        ):
+            raise ValueError(
+                "Dataset info (robot_type, keys, shapes...) must either come from a Robot or explicitly passed upon creation."
+            )
+
+        if len(video_keys) > 0 and not use_videos:
+            raise ValueError
+
+        obj.tasks, obj.stats, obj.episode_dicts = {}, {}, []
+        obj.info = create_empty_dataset_info(
+            CODEBASE_VERSION, fps, robot_type, keys, image_keys, video_keys, shapes, names
+        )
+        write_json(obj.info, obj.root / INFO_PATH)
+
+        # TODO(aliberts, rcadene, alexander-soare): Merge this with OnlineBuffer/DataBuffer
+        obj.episode_buffer = obj._create_episode_buffer()
+
+        # This bool indicates that the current LeRobotDataset instance is in sync with the files on disk. It
+        # is used to know when certain operations are need (for instance, computing dataset statistics). In
+        # order to be able to push the dataset to the hub, it needs to be consolidated first by calling
+        # self.consolidate().
+        obj.consolidated = True
+
+        obj.episodes = None
+        obj.hf_dataset = None
+        obj.image_transforms = None
+        obj.delta_timestamps = None
+        obj.delta_indices = None
+        obj.local_files_only = True
+        obj.episode_data_index = None
         obj.video_backend = video_backend if video_backend is not None else "pyav"
         return obj
 
@@ -225,8 +829,8 @@ class MultiLeRobotDataset(torch.utils.data.Dataset):
     def __init__(
         self,
         repo_ids: list[str],
-        root: Path | None = DATA_DIR,
-        split: str = "train",
+        root: Path | None = None,
+        episodes: dict | None = None,
         image_transforms: Callable | None = None,
         delta_timestamps: dict[list[float]] | None = None,
         video_backend: str | None = None,
@@ -238,8 +842,8 @@ class MultiLeRobotDataset(torch.utils.data.Dataset):
         self._datasets = [
             LeRobotDataset(
                 repo_id,
-                root=root,
-                split=split,
+                root=root / repo_id if root is not None else None,
+                episodes=episodes[repo_id] if episodes is not None else None,
                 delta_timestamps=delta_timestamps,
                 image_transforms=image_transforms,
                 video_backend=video_backend,
@@ -275,7 +879,6 @@ class MultiLeRobotDataset(torch.utils.data.Dataset):
             self.disabled_data_keys.update(extra_keys)
 
         self.root = root
-        self.split = split
         self.image_transforms = image_transforms
         self.delta_timestamps = delta_timestamps
         self.stats = aggregate_stats(self._datasets)
@@ -389,7 +992,6 @@ class MultiLeRobotDataset(torch.utils.data.Dataset):
         return (
             f"{self.__class__.__name__}(\n"
             f"  Repository IDs: '{self.repo_ids}',\n"
-            f"  Split: '{self.split}',\n"
             f"  Number of Samples: {self.num_samples},\n"
             f"  Number of Episodes: {self.num_episodes},\n"
             f"  Type: {'video (.mp4)' if self.video else 'image (.png)'},\n"

lerobot/common/datasets/push_dataset_to_hub/_download_raw.py

@@ -60,8 +60,8 @@ AVAILABLE_RAW_REPO_IDS = {
     "lerobot-raw/aloha_static_vinh_cup_left_raw": "aloha_hdf5",
     "lerobot-raw/aloha_static_vinh_cup_raw": "aloha_hdf5",
     "lerobot-raw/aloha_static_ziploc_slide_raw": "aloha_hdf5",
-    "lerobot-raw/pusht_raw": "pusht_zarr",
     "lerobot-raw/umi_cup_in_the_wild_raw": "umi_zarr",
+    "lerobot-raw/pusht_raw": "pusht_zarr",
     "lerobot-raw/unitreeh1_fold_clothes_raw": "aloha_hdf5",
     "lerobot-raw/unitreeh1_rearrange_objects_raw": "aloha_hdf5",
     "lerobot-raw/unitreeh1_two_robot_greeting_raw": "aloha_hdf5",
@@ -70,6 +70,74 @@ AVAILABLE_RAW_REPO_IDS = {
     "lerobot-raw/xarm_lift_medium_replay_raw": "xarm_pkl",
     "lerobot-raw/xarm_push_medium_raw": "xarm_pkl",
     "lerobot-raw/xarm_push_medium_replay_raw": "xarm_pkl",
+    "lerobot-raw/fractal20220817_data_raw": "openx_rlds.fractal20220817_data",
+    "lerobot-raw/kuka_raw": "openx_rlds.kuka",
+    "lerobot-raw/bridge_openx_raw": "openx_rlds.bridge_openx",
+    "lerobot-raw/taco_play_raw": "openx_rlds.taco_play",
+    "lerobot-raw/jaco_play_raw": "openx_rlds.jaco_play",
+    "lerobot-raw/berkeley_cable_routing_raw": "openx_rlds.berkeley_cable_routing",
+    "lerobot-raw/roboturk_raw": "openx_rlds.roboturk",
+    "lerobot-raw/nyu_door_opening_surprising_effectiveness_raw": "openx_rlds.nyu_door_opening_surprising_effectiveness",
+    "lerobot-raw/viola_raw": "openx_rlds.viola",
+    "lerobot-raw/berkeley_autolab_ur5_raw": "openx_rlds.berkeley_autolab_ur5",
+    "lerobot-raw/toto_raw": "openx_rlds.toto",
+    "lerobot-raw/language_table_raw": "openx_rlds.language_table",
+    "lerobot-raw/columbia_cairlab_pusht_real_raw": "openx_rlds.columbia_cairlab_pusht_real",
+    "lerobot-raw/stanford_kuka_multimodal_dataset_raw": "openx_rlds.stanford_kuka_multimodal_dataset",
+    "lerobot-raw/nyu_rot_dataset_raw": "openx_rlds.nyu_rot_dataset",
+    "lerobot-raw/io_ai_tech_raw": "openx_rlds.io_ai_tech",
+    "lerobot-raw/stanford_hydra_dataset_raw": "openx_rlds.stanford_hydra_dataset",
+    "lerobot-raw/austin_buds_dataset_raw": "openx_rlds.austin_buds_dataset",
+    "lerobot-raw/nyu_franka_play_dataset_raw": "openx_rlds.nyu_franka_play_dataset",
+    "lerobot-raw/maniskill_dataset_raw": "openx_rlds.maniskill_dataset",
+    "lerobot-raw/furniture_bench_dataset_raw": "openx_rlds.furniture_bench_dataset",
+    "lerobot-raw/cmu_franka_exploration_dataset_raw": "openx_rlds.cmu_franka_exploration_dataset",
+    "lerobot-raw/ucsd_kitchen_dataset_raw": "openx_rlds.ucsd_kitchen_dataset",
+    "lerobot-raw/ucsd_pick_and_place_dataset_raw": "openx_rlds.ucsd_pick_and_place_dataset",
+    "lerobot-raw/spoc_raw": "openx_rlds.spoc",
+    "lerobot-raw/austin_sailor_dataset_raw": "openx_rlds.austin_sailor_dataset",
+    "lerobot-raw/austin_sirius_dataset_raw": "openx_rlds.austin_sirius_dataset",
+    "lerobot-raw/bc_z_raw": "openx_rlds.bc_z",
+    "lerobot-raw/utokyo_pr2_opening_fridge_raw": "openx_rlds.utokyo_pr2_opening_fridge",
+    "lerobot-raw/utokyo_pr2_tabletop_manipulation_raw": "openx_rlds.utokyo_pr2_tabletop_manipulation",
+    "lerobot-raw/utokyo_xarm_pick_and_place_raw": "openx_rlds.utokyo_xarm_pick_and_place",
+    "lerobot-raw/utokyo_xarm_bimanual_raw": "openx_rlds.utokyo_xarm_bimanual",
+    "lerobot-raw/utokyo_saytap_raw": "openx_rlds.utokyo_saytap",
+    "lerobot-raw/robo_net_raw": "openx_rlds.robo_net",
+    "lerobot-raw/robo_set_raw": "openx_rlds.robo_set",
+    "lerobot-raw/berkeley_mvp_raw": "openx_rlds.berkeley_mvp",
+    "lerobot-raw/berkeley_rpt_raw": "openx_rlds.berkeley_rpt",
+    "lerobot-raw/kaist_nonprehensile_raw": "openx_rlds.kaist_nonprehensile",
+    "lerobot-raw/stanford_mask_vit_raw": "openx_rlds.stanford_mask_vit",
+    "lerobot-raw/tokyo_u_lsmo_raw": "openx_rlds.tokyo_u_lsmo",
+    "lerobot-raw/dlr_sara_pour_raw": "openx_rlds.dlr_sara_pour",
+    "lerobot-raw/dlr_sara_grid_clamp_raw": "openx_rlds.dlr_sara_grid_clamp",
+    "lerobot-raw/dlr_edan_shared_control_raw": "openx_rlds.dlr_edan_shared_control",
+    "lerobot-raw/asu_table_top_raw": "openx_rlds.asu_table_top",
+    "lerobot-raw/stanford_robocook_raw": "openx_rlds.stanford_robocook",
+    "lerobot-raw/imperialcollege_sawyer_wrist_cam_raw": "openx_rlds.imperialcollege_sawyer_wrist_cam",
+    "lerobot-raw/iamlab_cmu_pickup_insert_raw": "openx_rlds.iamlab_cmu_pickup_insert",
+    "lerobot-raw/uiuc_d3field_raw": "openx_rlds.uiuc_d3field",
+    "lerobot-raw/utaustin_mutex_raw": "openx_rlds.utaustin_mutex",
+    "lerobot-raw/berkeley_fanuc_manipulation_raw": "openx_rlds.berkeley_fanuc_manipulation",
+    "lerobot-raw/cmu_playing_with_food_raw": "openx_rlds.cmu_playing_with_food",
+    "lerobot-raw/cmu_play_fusion_raw": "openx_rlds.cmu_play_fusion",
+    "lerobot-raw/cmu_stretch_raw": "openx_rlds.cmu_stretch",
+    "lerobot-raw/berkeley_gnm_recon_raw": "openx_rlds.berkeley_gnm_recon",
+    "lerobot-raw/berkeley_gnm_cory_hall_raw": "openx_rlds.berkeley_gnm_cory_hall",
+    "lerobot-raw/berkeley_gnm_sac_son_raw": "openx_rlds.berkeley_gnm_sac_son",
+    "lerobot-raw/droid_raw": "openx_rlds.droid",
+    "lerobot-raw/droid_100_raw": "openx_rlds.droid100",
+    "lerobot-raw/fmb_raw": "openx_rlds.fmb",
+    "lerobot-raw/dobbe_raw": "openx_rlds.dobbe",
+    "lerobot-raw/usc_cloth_sim_raw": "openx_rlds.usc_cloth_sim",
+    "lerobot-raw/plex_robosuite_raw": "openx_rlds.plex_robosuite",
+    "lerobot-raw/conq_hose_manipulation_raw": "openx_rlds.conq_hose_manipulation",
+    "lerobot-raw/vima_raw": "openx_rlds.vima",
+    "lerobot-raw/robot_vqa_raw": "openx_rlds.robot_vqa",
+    "lerobot-raw/mimic_play_raw": "openx_rlds.mimic_play",
+    "lerobot-raw/tidybot_raw": "openx_rlds.tidybot",
+    "lerobot-raw/eth_agent_affordances_raw": "openx_rlds.eth_agent_affordances",
 }
 
 
@@ -110,7 +178,7 @@ def download_all_raw_datasets(data_dir: Path | None = None):
 def main():
     parser = argparse.ArgumentParser(
         description=f"""A script to download raw datasets from Hugging Face hub to a local directory. Here is a
-            non exhaustive list of available repositories to use in `--repo-id`: {AVAILABLE_RAW_REPO_IDS}""",
+            non exhaustive list of available repositories to use in `--repo-id`: {list(AVAILABLE_RAW_REPO_IDS.keys())}""",
     )
 
     parser.add_argument(

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

@@ -0,0 +1,639 @@
+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

@@ -0,0 +1,106 @@
+#!/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

@@ -0,0 +1,200 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+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

@@ -0,0 +1,859 @@
+#!/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

@@ -0,0 +1,359 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+For https://github.com/google-deepmind/open_x_embodiment (OPENX) datasets.
+
+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 \
+        --episodes 3 4 5 8 9
+
+Exact dataset fps defined in openx/config.py, obtained from:
+    https://docs.google.com/spreadsheets/d/1rPBD77tk60AEIGZrGSODwyyzs5FgCU9Uz3h-3_t2A9g/edit?gid=0#gid=0&range=R:R
+"""
+
+import shutil
+from pathlib import Path
+
+import numpy as np
+import tensorflow as tf
+import tensorflow_datasets as tfds
+import torch
+import tqdm
+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 (
+    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)
+
+
+def tf_to_torch(data):
+    return torch.from_numpy(data.numpy())
+
+
+def tf_img_convert(img):
+    if img.dtype == tf.string:
+        img = tf.io.decode_image(img, expand_animations=False, dtype=tf.uint8)
+    elif img.dtype != tf.uint8:
+        raise ValueError(f"Unsupported image dtype: found with dtype {img.dtype}")
+    return img.numpy()
+
+
+def _broadcast_metadata_rlds(i: tf.Tensor, traj: dict) -> dict:
+    """
+    In the RLDS format, each trajectory has some top-level metadata that is explicitly separated out, and a "steps"
+    entry. This function moves the "steps" entry to the top level, broadcasting any metadata to the length of the
+    trajectory. This function also adds the extra metadata fields `_len`, `_traj_index`, and `_frame_index`.
+
+    NOTE: adapted from DLimp library https://github.com/kvablack/dlimp/
+    """
+    steps = traj.pop("steps")
+
+    traj_len = tf.shape(tf.nest.flatten(steps)[0])[0]
+
+    # broadcast metadata to the length of the trajectory
+    metadata = tf.nest.map_structure(lambda x: tf.repeat(x, traj_len), traj)
+
+    # put steps back in
+    assert "traj_metadata" not in steps
+    traj = {**steps, "traj_metadata": metadata}
+
+    assert "_len" not in traj
+    assert "_traj_index" not in traj
+    assert "_frame_index" not in traj
+    traj["_len"] = tf.repeat(traj_len, traj_len)
+    traj["_traj_index"] = tf.repeat(i, traj_len)
+    traj["_frame_index"] = tf.range(traj_len)
+
+    return traj
+
+
+def load_from_raw(
+    raw_dir: Path,
+    videos_dir: Path,
+    fps: int,
+    video: bool,
+    episodes: list[int] | None = None,
+    encoding: dict | None = None,
+    openx_dataset_name: str | None = None,
+):
+    """
+    Args:
+        raw_dir (Path): _description_
+        videos_dir (Path): _description_
+        fps (int): _description_
+        video (bool): _description_
+        episodes (list[int] | None, optional): _description_. Defaults to None.
+    """
+    ds_builder = tfds.builder_from_directory(str(raw_dir))
+    dataset = ds_builder.as_dataset(
+        split="all",
+        decoders={"steps": tfds.decode.SkipDecoding()},
+    )
+
+    dataset_info = ds_builder.info
+    print("dataset_info: ", dataset_info)
+
+    ds_length = len(dataset)
+    dataset = dataset.take(ds_length)
+    # "flatten" the dataset as such we can apply trajectory level map() easily
+    # each [obs][key] has a shape of (frame_size, ...)
+    dataset = dataset.enumerate().map(_broadcast_metadata_rlds)
+
+    # we will apply the standardization transform if the dataset_name is provided
+    # if the dataset name is not provided and the goal is to convert any rlds formatted dataset
+    # search for 'image' keys in the observations
+    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]
+
+    lang_key = "language_instruction" if "language_instruction" in dataset.element_spec else None
+
+    print(" - image_keys: ", image_keys)
+    print(" - lang_key: ", lang_key)
+
+    it = iter(dataset)
+
+    ep_dicts = []
+    # Init temp path to save ep_dicts in case of crash
+    tmp_ep_dicts_dir = videos_dir.parent.joinpath("ep_dicts")
+    tmp_ep_dicts_dir.mkdir(parents=True, exist_ok=True)
+
+    # check if ep_dicts have already been saved in /tmp
+    starting_ep_idx = 0
+    saved_ep_dicts = [ep.__str__() for ep in tmp_ep_dicts_dir.iterdir()]
+    if len(saved_ep_dicts) > 0:
+        saved_ep_dicts.sort()
+        # get last ep_idx number
+        starting_ep_idx = int(saved_ep_dicts[-1][-13:-3]) + 1
+        for i in range(starting_ep_idx):
+            episode = next(it)
+            ep_dicts.append(torch.load(saved_ep_dicts[i]))
+
+    # if we user specified episodes, skip the ones not in the list
+    if episodes is not None:
+        if ds_length == 0:
+            raise ValueError("No episodes found.")
+        # convert episodes index to sorted list
+        episodes = sorted(episodes)
+
+    for ep_idx in tqdm.tqdm(range(starting_ep_idx, ds_length)):
+        episode = next(it)
+
+        # if user specified episodes, skip the ones not in the list
+        if episodes is not None:
+            if len(episodes) == 0:
+                break
+            if ep_idx == episodes[0]:
+                # process this episode
+                print(" selecting episode idx: ", ep_idx)
+                episodes.pop(0)
+            else:
+                continue  # skip
+
+        num_frames = episode["action"].shape[0]
+
+        ###########################################################
+        # 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"
+
+        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()
+
+        # 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]]
+
+        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}"
+            imgs_array = image_array_dict[im_key]
+            imgs_array = np.array(imgs_array)
+            if video:
+                # save png images in temporary directory
+                tmp_imgs_dir = videos_dir / "tmp_images"
+                save_images_concurrently(imgs_array, tmp_imgs_dir)
+
+                # encode images to a mp4 video
+                fname = f"{img_key}_episode_{ep_idx:06d}.mp4"
+                video_path = videos_dir / fname
+                encode_video_frames(tmp_imgs_dir, video_path, fps, **(encoding or {}))
+
+                # clean temporary images directory
+                shutil.rmtree(tmp_imgs_dir)
+
+                # store the reference to the video frame
+                ep_dict[img_key] = [
+                    {"path": f"videos/{fname}", "timestamp": i / fps} for i in range(num_frames)
+                ]
+            else:
+                ep_dict[img_key] = [PILImage.fromarray(x) for x in imgs_array]
+
+        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"
+        )
+        torch.save(ep_dict, path_ep_dict)
+
+        ep_dicts.append(ep_dict)
+
+    data_dict = concatenate_episodes(ep_dicts)
+
+    total_frames = data_dict["frame_index"].shape[0]
+    data_dict["index"] = torch.arange(0, total_frames, 1)
+    return data_dict
+
+
+def to_hf_dataset(data_dict, video) -> Dataset:
+    features = {}
+
+    keys = [key for key in data_dict if "observation.images." in key]
+    for key in keys:
+        if video:
+            features[key] = VideoFrame()
+        else:
+            features[key] = Image()
+
+    features["observation.state"] = Sequence(
+        length=data_dict["observation.state"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    if "observation.velocity" in data_dict:
+        features["observation.velocity"] = Sequence(
+            length=data_dict["observation.velocity"].shape[1], feature=Value(dtype="float32", id=None)
+        )
+    if "observation.effort" in data_dict:
+        features["observation.effort"] = Sequence(
+            length=data_dict["observation.effort"].shape[1], feature=Value(dtype="float32", id=None)
+        )
+    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["episode_index"] = Value(dtype="int64", id=None)
+    features["frame_index"] = Value(dtype="int64", id=None)
+    features["timestamp"] = Value(dtype="float32", id=None)
+    features["next.reward"] = Value(dtype="float32", id=None)
+    features["next.done"] = Value(dtype="bool", id=None)
+    features["index"] = Value(dtype="int64", id=None)
+
+    hf_dataset = Dataset.from_dict(data_dict, features=Features(features))
+    hf_dataset.set_transform(hf_transform_to_torch)
+    return hf_dataset
+
+
+def from_raw_to_lerobot_format(
+    raw_dir: Path,
+    videos_dir: Path,
+    fps: int | None = None,
+    video: bool = True,
+    episodes: list[int] | None = None,
+    encoding: dict | None = None,
+    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)
+    hf_dataset = to_hf_dataset(data_dict, video)
+    episode_data_index = calculate_episode_data_index(hf_dataset)
+    info = {
+        "codebase_version": CODEBASE_VERSION,
+        "fps": fps,
+        "video": video,
+    }
+    if video:
+        info["encoding"] = get_default_encoding()
+
+    return hf_dataset, episode_data_index, info

lerobot/common/datasets/utils.py

@@ -14,30 +14,41 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import json
-import re
 import warnings
-from functools import cache
+from itertools import accumulate
 from pathlib import Path
+from pprint import pformat
 from typing import Dict
 
 import datasets
+import jsonlines
 import torch
-from datasets import load_dataset, load_from_disk
-from huggingface_hub import DatasetCard, HfApi, hf_hub_download, snapshot_download
+from huggingface_hub import DatasetCard, 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"
+
 DATASET_CARD_TEMPLATE = """
 ---
 # Metadata will go there
 ---
-This dataset was created using [🤗 LeRobot](https://github.com/huggingface/lerobot).
+This dataset was created using [LeRobot](https://github.com/huggingface/lerobot).
 
 """
 
 
-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 +67,7 @@ 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 +80,24 @@ def unflatten_dict(d, sep="/"):
     return outdict
 
 
+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 append_jsonl(data: dict, fpath: Path) -> None:
+    fpath.parent.mkdir(exist_ok=True, parents=True)
+    with jsonlines.open(fpath, "a") as writer:
+        writer.write(data)
+
+
+def write_stats(stats: dict[str, torch.Tensor | dict], fpath: Path) -> None:
+    serialized_stats = {key: value.tolist() for key, value in flatten_dict(stats).items()}
+    serialized_stats = unflatten_dict(serialized_stats)
+    write_json(serialized_stats, fpath)
+
+
 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
@@ -80,9 +109,6 @@ 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, 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:
@@ -90,8 +116,40 @@ 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])
+
+
+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 ValueError(
+            f"""The dataset you requested ({repo_id}) is in {version_to_check} format. We introduced a new
+            format with v2.0 that is not backward compatible. Please use our conversion script
+            first (convert_dataset_v1_to_v2.py) to convert your dataset to this new format."""
+        )
+    elif float(version_to_check.strip("v")) < float(current_version.strip("v")):
+        warnings.warn(
+            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.""",
+            stacklevel=1,
+        )
+
+
+def get_hub_safe_version(repo_id: str, version: str, enforce_v2: bool = True) -> str:
+    num_version = float(version.strip("v"))
+    if num_version < 2 and enforce_v2:
+        raise ValueError(
+            f"""The dataset you requested ({repo_id}) is in {version} format. We introduced a new
+            format with v2.0 that is not backward compatible. Please use our conversion script
+            first (convert_dataset_v1_to_v2.py) to convert your dataset to this new format."""
+        )
     api = HfApi()
     dataset_info = api.list_repo_refs(repo_id, repo_type="dataset")
     branches = [b.name for b in dataset_info.branches]
@@ -111,106 +169,185 @@ 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_info(local_dir: Path) -> dict:
+    with open(local_dir / INFO_PATH) as f:
+        return json.load(f)
 
 
-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)
+def load_stats(local_dir: Path) -> dict:
+    with open(local_dir / STATS_PATH) as f:
+        stats = json.load(f)
+    stats = {key: torch.tensor(value) for key, value in flatten_dict(stats).items()}
     return unflatten_dict(stats)
 
 
-def load_info(repo_id, version, root) -> dict:
-    """info contains useful information regarding the dataset that are not stored elsewhere
+def load_tasks(local_dir: Path) -> dict:
+    with jsonlines.open(local_dir / TASKS_PATH, "r") as reader:
+        tasks = list(reader)
 
-    Example:
-    ```python
-    print("frame per second used to collect the video", info["fps"])
-    ```
+    return {item["task_index"]: item["task"] for item in sorted(tasks, key=lambda x: x["task_index"])}
+
+
+def load_episode_dicts(local_dir: Path) -> dict:
+    with jsonlines.open(local_dir / EPISODES_PATH, "r") as reader:
+        return list(reader)
+
+
+def _get_info_from_robot(robot: Robot, use_videos: bool) -> tuple[list | dict]:
+    shapes = {key: len(names) for key, names in robot.names.items()}
+    camera_shapes = {}
+    for key, cam in robot.cameras.items():
+        video_key = f"observation.images.{key}"
+        camera_shapes[video_key] = {
+            "width": cam.width,
+            "height": cam.height,
+            "channels": cam.channels,
+        }
+    keys = list(robot.names)
+    image_keys = [] if use_videos else list(camera_shapes)
+    video_keys = list(camera_shapes) if use_videos else []
+    shapes = {**shapes, **camera_shapes}
+    names = robot.names
+    robot_type = robot.robot_type
+
+    return robot_type, keys, image_keys, video_keys, shapes, names
+
+
+def create_empty_dataset_info(
+    codebase_version: str,
+    fps: int,
+    robot_type: str,
+    keys: list[str],
+    image_keys: list[str],
+    video_keys: list[str],
+    shapes: dict,
+    names: dict,
+) -> dict:
+    return {
+        "codebase_version": codebase_version,
+        "data_path": DEFAULT_PARQUET_PATH,
+        "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": {},
+        "keys": keys,
+        "video_keys": video_keys,
+        "image_keys": image_keys,
+        "shapes": shapes,
+        "names": names,
+        "videos": {"videos_path": DEFAULT_VIDEO_PATH} if len(video_keys) > 0 else None,
+    }
+
+
+def get_episode_data_index(episodes: list, episode_dicts: list[dict]) -> 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}
+
+    cumulative_lenghts = list(accumulate(episode_lengths.values()))
+    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:
     """
-    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)
+    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"])
+    # timestamps[2] += tolerance_s  # TODO delete
+    # timestamps[-2] += tolerance_s/2  # TODO delete
+    diffs = torch.diff(timestamps)
+    within_tolerance = torch.abs(diffs - 1 / fps) <= tolerance_s
 
-    with open(path) as f:
-        info = json.load(f)
-    return info
+    # We mask differences between the timestamp at the end of an episode
+    # and the one 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 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"
+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)) <= 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
+            ]
 
-    return path
+    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
+
+
+# TODO(aliberts): remove
 def load_previous_and_future_frames(
     item: dict[str, torch.Tensor],
     hf_dataset: datasets.Dataset,
@@ -304,6 +441,7 @@ def load_previous_and_future_frames(
     return item
 
 
+# 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.
@@ -358,6 +496,7 @@ def calculate_episode_data_index(hf_dataset: datasets.Dataset) -> Dict[str, torc
     return episode_data_index
 
 
+# TODO(aliberts): remove
 def reset_episode_index(hf_dataset: datasets.Dataset) -> datasets.Dataset:
     """Reset the `episode_index` of the provided HuggingFace Dataset.
 
@@ -395,7 +534,7 @@ 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.
     """
@@ -410,12 +549,17 @@ 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:
+def create_lerobot_dataset_card(
+    tags: list | None = None, text: str | None = None, info: dict | 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
+        card.text += f"{text}\n"
+    if info is not None:
+        card.text += "[meta/info.json](meta/info.json)\n"
+        card.text += f"```json\n{json.dumps(info, indent=4)}\n```"
     return card

lerobot/common/datasets/v2/batch_convert_dataset_v1_to_v2.py

@@ -0,0 +1,106 @@
+#!/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 traceback
+from pathlib import Path
+
+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_SINGLE_TASKS_REAL = {
+    "aloha_mobile_cabinet": "Open the top cabinet, store the pot inside it then close the cabinet.",
+    "aloha_mobile_chair": "Push the chairs in front of the desk to place them against it.",
+    "aloha_mobile_elevator": "Take the elevator to the 1st floor.",
+    "aloha_mobile_shrimp": "Sauté the raw shrimp on both sides, then serve it in the bowl.",
+    "aloha_mobile_wash_pan": "Pick up the pan, rinse it in the sink and then place it in the drying rack.",
+    "aloha_mobile_wipe_wine": "Pick up the wet cloth on the faucet and use it to clean the spilled wine on the table and underneath the glass.",
+    "aloha_static_battery": "Place the battery into the slot of the remote controller.",
+    "aloha_static_candy": "Pick up the candy and unwrap it.",
+    "aloha_static_coffee": "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_coffee_new": "Place the coffee capsule inside the capsule container, then place the cup onto the center of the cup tray.",
+    "aloha_static_cups_open": "Pick up the plastic cup and open its lid.",
+    "aloha_static_fork_pick_up": "Pick up the fork and place it on the plate.",
+    "aloha_static_pingpong_test": "Transfer one of the two balls in the right glass into the left glass, then transfer it back to the right glass.",
+    "aloha_static_pro_pencil": "Pick up the pencil with the right arm, hand it over to the left arm then place it back onto the table.",
+    "aloha_static_screw_driver": "Pick up the screwdriver with the right arm, hand it over to the left arm then place it into the cup.",
+    "aloha_static_tape": "Cut a small piece of tape from the tape dispenser then place it on the cardboard box's edge.",
+    "aloha_static_thread_velcro": "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_towel": "Pick up a piece of paper towel and place it on the spilled liquid.",
+    "aloha_static_vinh_cup": "Pick up the platic cup with the right arm, then pop its lid open with the left arm.",
+    "aloha_static_vinh_cup_left": "Pick up the platic cup with the left arm, then pop its lid open with the right arm.",
+    "aloha_static_ziploc_slide": "Slide open the ziploc bag.",
+}
+ALOHA_CONFIG = Path("lerobot/configs/robot/aloha.yaml")
+
+
+def batch_convert():
+    status = {}
+    logfile = LOCAL_DIR / "conversion_log.txt"
+    for num, repo_id in enumerate(available_datasets):
+        print(f"\nConverting {repo_id} ({num}/{len(available_datasets)})")
+        print("---------------------------------------------------------")
+        name = repo_id.split("/")[1]
+        single_task, tasks_col, robot_config = None, None, None
+
+        if "aloha" in name:
+            robot_config = parse_robot_config(ALOHA_CONFIG)
+            if "sim_insertion" in name:
+                single_task = "Insert the peg into the socket."
+            elif "sim_transfer" in name:
+                single_task = "Pick up the cube with the right arm and transfer it to the left arm."
+            else:
+                single_task = ALOHA_SINGLE_TASKS_REAL[name]
+        elif "unitreeh1" in name:
+            if "fold_clothes" in name:
+                single_task = "Fold the sweatshirt."
+            elif "rearrange_objects" in name or "rearrange_objects" in name:
+                single_task = "Put the object into the bin."
+            elif "two_robot_greeting" in name:
+                single_task = "Greet the other robot with a high five."
+            elif "warehouse" in name:
+                single_task = (
+                    "Grab the spray paint on the shelf and place it in the bin on top of the robot dog."
+                )
+        elif name != "columbia_cairlab_pusht_real" and "pusht" in name:
+            single_task = "Push the T-shaped block onto the T-shaped target."
+        elif "xarm_lift" in name or "xarm_push" in name:
+            single_task = "Pick up the cube and lift it."
+        elif name == "umi_cup_in_the_wild":
+            single_task = "Put the cup on the plate."
+        else:
+            tasks_col = "language_instruction"
+
+        try:
+            convert_dataset(
+                repo_id=repo_id,
+                local_dir=LOCAL_DIR,
+                single_task=single_task,
+                tasks_col=tasks_col,
+                robot_config=robot_config,
+            )
+            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 +1,814 @@
+#!/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 math
+import shutil
+import subprocess
+import warnings
+from pathlib import Path
+
+import datasets
+import jsonlines
+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
+from PIL import Image
+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,
+    unflatten_dict,
+)
+from lerobot.common.datasets.video_utils import VideoFrame  # noqa: F401
+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,
+            "action": action_names,
+        },
+    }
+
+
+def load_json(fpath: Path) -> dict:
+    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 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 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_keys(dataset: Dataset) -> dict[str, list]:
+    sequence_keys, image_keys, video_keys = [], [], []
+    for key, ft in dataset.features.items():
+        if isinstance(ft, datasets.Sequence):
+            sequence_keys.append(key)
+        elif isinstance(ft, datasets.Image):
+            image_keys.append(key)
+        elif ft._type == "VideoFrame":
+            video_keys.append(key)
+
+    return {
+        "sequence": sequence_keys,
+        "image": image_keys,
+        "video": video_keys,
+    }
+
+
+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,
+    keys: dict[str, list],
+    total_episodes: int,
+    total_chunks: int,
+    output_dir: Path,
+) -> list:
+    table = dataset.remove_columns(keys["video"])._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_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.width": video_stream_info["width"],
+        "video.height": video_stream_info["height"],
+        "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_videos_info(repo_id: str, local_dir: Path, video_keys: list[str], branch: str) -> dict:
+    hub_api = HfApi()
+    videos_info_dict = {"videos_path": DEFAULT_VIDEO_PATH}
+
+    # 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.snapshot_download(
+        repo_id=repo_id, repo_type="dataset", local_dir=local_dir, revision=branch, allow_patterns=video_files
+    )
+    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 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")
+
+
+def get_video_shapes(videos_info: dict, video_keys: list) -> dict:
+    video_shapes = {}
+    for img_key in video_keys:
+        channels = get_video_pixel_channels(videos_info[img_key]["video.pix_fmt"])
+        video_shapes[img_key] = {
+            "width": videos_info[img_key]["video.width"],
+            "height": videos_info[img_key]["video.height"],
+            "channels": channels,
+        }
+
+    return video_shapes
+
+
+def get_image_shapes(dataset: Dataset, image_keys: list) -> dict:
+    image_shapes = {}
+    for img_key in image_keys:
+        image = dataset[0][img_key]  # Assuming first row
+        channels = get_image_pixel_channels(image)
+        image_shapes[img_key] = {
+            "width": image.width,
+            "height": image.height,
+            "channels": channels,
+        }
+
+    return image_shapes
+
+
+def get_generic_motor_names(sequence_shapes: dict) -> dict:
+    return {key: [f"motor_{i}" for i in range(length)] for key, length in sequence_shapes.items()}
+
+
+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,
+):
+    v1 = get_hub_safe_version(repo_id, V16, enforce_v2=False)
+    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")
+    keys = get_keys(dataset)
+
+    if single_task and "language_instruction" in dataset.column_names:
+        warnings.warn(
+            "'single_task' provided but 'language_instruction' tasks_col found. Using 'language_instruction'.",
+            stacklevel=1,
+        )
+        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(keys["video"])
+    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()}
+        # tasks = list(set(tasks_by_episodes.values()))
+        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)
+
+    # Shapes
+    sequence_shapes = {key: dataset.features[key].length for key in keys["sequence"]}
+    image_shapes = get_image_shapes(dataset, keys["image"]) if len(keys["image"]) > 0 else {}
+
+    # Videos
+    if len(keys["video"]) > 0:
+        assert metadata_v1.get("video", False)
+        tmp_video_dir = local_dir / "videos" / V20 / repo_id
+        tmp_video_dir.mkdir(parents=True, exist_ok=True)
+        clean_gitattr = Path(
+            hub_api.hf_hub_download(
+                repo_id=GITATTRIBUTES_REF, repo_type="dataset", local_dir=local_dir, filename=".gitattributes"
+            )
+        ).absolute()
+        move_videos(
+            repo_id, keys["video"], total_episodes, total_chunks, tmp_video_dir, clean_gitattr, branch
+        )
+        videos_info = get_videos_info(repo_id, v1x_dir, video_keys=keys["video"], branch=branch)
+        video_shapes = get_video_shapes(videos_info, keys["video"])
+        for img_key in keys["video"]:
+            assert math.isclose(videos_info[img_key]["video.fps"], metadata_v1["fps"], rel_tol=1e-3)
+            if "encoding" in metadata_v1:
+                assert videos_info[img_key]["video.pix_fmt"] == metadata_v1["encoding"]["pix_fmt"]
+    else:
+        assert metadata_v1.get("video", 0) == 0
+        videos_info = None
+        video_shapes = {}
+
+    # Split data into 1 parquet file by episode
+    episode_lengths = split_parquet_by_episodes(dataset, keys, total_episodes, total_chunks, v20_dir)
+
+    # Names
+    if robot_config is not None:
+        robot_type = robot_config["robot_type"]
+        names = robot_config["names"]
+        if "observation.effort" in keys["sequence"]:
+            names["observation.effort"] = names["observation.state"]
+        if "observation.velocity" in keys["sequence"]:
+            names["observation.velocity"] = names["observation.state"]
+        repo_tags = [robot_type]
+    else:
+        robot_type = "unknown"
+        names = get_generic_motor_names(sequence_shapes)
+        repo_tags = None
+
+    assert set(names) == set(keys["sequence"])
+    for key in sequence_shapes:
+        assert len(names[key]) == sequence_shapes[key]
+
+    # 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,
+        "data_path": DEFAULT_PARQUET_PATH,
+        "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}"},
+        "keys": keys["sequence"],
+        "video_keys": keys["video"],
+        "image_keys": keys["image"],
+        "shapes": {**sequence_shapes, **video_shapes, **image_shapes},
+        "names": names,
+        "videos": videos_info,
+    }
+    write_json(metadata_v2_0, v20_dir / INFO_PATH)
+    convert_stats_to_json(v1x_dir, v20_dir)
+
+    with contextlib.suppress(EntryNotFoundError):
+        hub_api.delete_folder(repo_id=repo_id, path_in_repo="data", repo_type="dataset", revision=branch)
+
+    with contextlib.suppress(EntryNotFoundError):
+        hub_api.delete_folder(repo_id=repo_id, path_in_repo="meta_data", repo_type="dataset", revision=branch)
+
+    with contextlib.suppress(EntryNotFoundError):
+        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 = create_lerobot_dataset_card(tags=repo_tags, info=metadata_v2_0)
+    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")
+
+    # TODO:
+    # - [X] Add shapes
+    # - [X] Add keys
+    # - [X] Add paths
+    # - [X] convert stats.json
+    # - [X] Add task.json
+    # - [X] Add names
+    # - [X] Add robot_type
+    # - [X] Add splits
+    # - [X] Push properly to branch v2.0 and delete v1.6 stuff from that branch
+    # - [X] Handle multitask datasets
+    # - [X] Handle hf hub repo limits (add chunks logic)
+    # - [X] Add test-branch
+    # - [X] Use jsonlines for episodes
+    # - [X] Add sanity checks (encoding, shapes)
+
+
+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(
+        "--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__":
+    from time import sleep
+
+    sleep(1)
+    main()

lerobot/common/datasets/video_utils.py

@@ -27,45 +27,8 @@ 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
-
-
 def decode_video_frames_torchvision(
-    video_path: str,
+    video_path: Path | str,
     timestamps: list[float],
     tolerance_s: float,
     backend: str = "pyav",
@@ -163,8 +126,8 @@ 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",

lerobot/common/envs/utils.py

@@ -39,7 +39,7 @@ def preprocess_observation(observations: dict[str, np.ndarray]) -> dict[str, Ten
 
             # sanity check that images are channel last
             _, h, w, c = img.shape
-            assert c < h and c < w, f"expect channel first images, but instead {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=}"

lerobot/common/logger.py

@@ -189,7 +189,7 @@ class Logger:
             training_state["scheduler"] = scheduler.state_dict()
         torch.save(training_state, save_dir / self.training_state_file_name)
 
-    def save_checkpont(
+    def save_checkpoint(
         self,
         train_step: int,
         policy: Policy,

lerobot/common/policies/act/modeling_act.py

@@ -296,7 +296,7 @@ class ACT(nn.Module):
         self.use_images = any(k.startswith("observation.image") for k in config.input_shapes)
         self.use_env_state = "observation.environment_state" in config.input_shapes
         if self.config.use_vae:
-            self.vae_encoder = ACTEncoder(config)
+            self.vae_encoder = ACTEncoder(config, is_vae_encoder=True)
             self.vae_encoder_cls_embed = nn.Embedding(1, config.dim_model)
             # Projection layer for joint-space configuration to hidden dimension.
             if self.use_robot_state:
@@ -521,9 +521,11 @@ class ACT(nn.Module):
 class ACTEncoder(nn.Module):
     """Convenience module for running multiple encoder layers, maybe followed by normalization."""
 
-    def __init__(self, config: ACTConfig):
+    def __init__(self, config: ACTConfig, is_vae_encoder: bool = False):
         super().__init__()
-        self.layers = nn.ModuleList([ACTEncoderLayer(config) for _ in range(config.n_encoder_layers)])
+        self.is_vae_encoder = is_vae_encoder
+        num_layers = config.n_vae_encoder_layers if self.is_vae_encoder else config.n_encoder_layers
+        self.layers = nn.ModuleList([ACTEncoderLayer(config) for _ in range(num_layers)])
         self.norm = nn.LayerNorm(config.dim_model) if config.pre_norm else nn.Identity()
 
     def forward(

lerobot/common/policies/diffusion/configuration_diffusion.py

@@ -196,3 +196,12 @@ class DiffusionConfig:
                 f"`noise_scheduler_type` must be one of {supported_noise_schedulers}. "
                 f"Got {self.noise_scheduler_type}."
             )
+
+        # Check that the horizon size and U-Net downsampling is compatible.
+        # U-Net downsamples by 2 with each stage.
+        downsampling_factor = 2 ** len(self.down_dims)
+        if self.horizon % downsampling_factor != 0:
+            raise ValueError(
+                "The horizon should be an integer multiple of the downsampling factor (which is determined "
+                f"by `len(down_dims)`). Got {self.horizon=} and {self.down_dims=}"
+            )

lerobot/common/policies/vqbet/modeling_vqbet.py

@@ -350,17 +350,22 @@ class VQBeTModel(nn.Module):
 
         # get action features (pass through GPT)
         features = self.policy(input_tokens)
-        # len(self.config.input_shapes) is the number of different observation modes. this line gets the index of action prompt tokens.
+        # len(self.config.input_shapes) is the number of different observation modes.
+        # this line gets the index of action prompt tokens.
         historical_act_pred_index = np.arange(0, n_obs_steps) * (len(self.config.input_shapes) + 1) + len(
             self.config.input_shapes
         )
 
         # only extract the output tokens at the position of action query:
-        # Behavior Transformer (BeT), and VQ-BeT are both sequence-to-sequence prediction models, mapping sequential observation to sequential action (please refer to section 2.2 in BeT paper https://arxiv.org/pdf/2206.11251).
-        # Thus, it predict historical action sequence, in addition to current and future actions (predicting future actions : optional).
-        features = torch.cat(
-            [features[:, historical_act_pred_index], features[:, -len_additional_action_token:]], dim=1
-        )
+        # Behavior Transformer (BeT), and VQ-BeT are both sequence-to-sequence prediction models,
+        # mapping sequential observation to sequential action (please refer to section 2.2 in BeT paper https://arxiv.org/pdf/2206.11251).
+        # Thus, it predicts a historical action sequence, in addition to current and future actions (predicting future actions : optional).
+        if len_additional_action_token > 0:
+            features = torch.cat(
+                [features[:, historical_act_pred_index], features[:, -len_additional_action_token:]], dim=1
+            )
+        else:
+            features = features[:, historical_act_pred_index]
         # pass through action head
         action_head_output = self.action_head(features)
         # if rollout, VQ-BeT don't calculate loss

lerobot/common/robot_devices/cameras/intelrealsense.py

@@ -0,0 +1,562 @@
+"""
+This file contains utilities for recording frames from Intel Realsense cameras.
+"""
+
+import argparse
+import concurrent.futures
+import logging
+import math
+import shutil
+import threading
+import time
+import traceback
+from collections import Counter
+from dataclasses import dataclass, replace
+from pathlib import Path
+from threading import Thread
+
+import numpy as np
+from PIL import Image
+
+from lerobot.common.robot_devices.utils import (
+    RobotDeviceAlreadyConnectedError,
+    RobotDeviceNotConnectedError,
+    busy_wait,
+)
+from lerobot.common.utils.utils import capture_timestamp_utc
+
+SERIAL_NUMBER_INDEX = 1
+
+
+def find_cameras(raise_when_empty=True, mock=False) -> list[dict]:
+    """
+    Find the names and the serial numbers of the Intel RealSense cameras
+    connected to the computer.
+    """
+    if mock:
+        import tests.mock_pyrealsense2 as rs
+    else:
+        import pyrealsense2 as rs
+
+    cameras = []
+    for device in rs.context().query_devices():
+        serial_number = int(device.get_info(rs.camera_info(SERIAL_NUMBER_INDEX)))
+        name = device.get_info(rs.camera_info.name)
+        cameras.append(
+            {
+                "serial_number": serial_number,
+                "name": name,
+            }
+        )
+
+    if raise_when_empty and len(cameras) == 0:
+        raise OSError(
+            "Not a single camera was detected. Try re-plugging, or re-installing `librealsense` and its python wrapper `pyrealsense2`, or updating the firmware."
+        )
+
+    return cameras
+
+
+def save_image(img_array, serial_number, frame_index, images_dir):
+    try:
+        img = Image.fromarray(img_array)
+        path = images_dir / f"camera_{serial_number}_frame_{frame_index:06d}.png"
+        path.parent.mkdir(parents=True, exist_ok=True)
+        img.save(str(path), quality=100)
+        logging.info(f"Saved image: {path}")
+    except Exception as e:
+        logging.error(f"Failed to save image for camera {serial_number} frame {frame_index}: {e}")
+
+
+def save_images_from_cameras(
+    images_dir: Path,
+    serial_numbers: list[int] | None = None,
+    fps=None,
+    width=None,
+    height=None,
+    record_time_s=2,
+    mock=False,
+):
+    """
+    Initializes all the cameras and saves images to the directory. Useful to visually identify the camera
+    associated to a given serial number.
+    """
+    if serial_numbers is None or len(serial_numbers) == 0:
+        camera_infos = find_cameras(mock=mock)
+        serial_numbers = [cam["serial_number"] for cam in camera_infos]
+
+    if mock:
+        import tests.mock_cv2 as cv2
+    else:
+        import cv2
+
+    print("Connecting cameras")
+    cameras = []
+    for cam_sn in serial_numbers:
+        print(f"{cam_sn=}")
+        camera = IntelRealSenseCamera(cam_sn, fps=fps, width=width, height=height, mock=mock)
+        camera.connect()
+        print(
+            f"IntelRealSenseCamera({camera.serial_number}, fps={camera.fps}, width={camera.width}, height={camera.height}, color_mode={camera.color_mode})"
+        )
+        cameras.append(camera)
+
+    images_dir = Path(images_dir)
+    if images_dir.exists():
+        shutil.rmtree(
+            images_dir,
+        )
+    images_dir.mkdir(parents=True, exist_ok=True)
+
+    print(f"Saving images to {images_dir}")
+    frame_index = 0
+    start_time = time.perf_counter()
+    try:
+        with concurrent.futures.ThreadPoolExecutor(max_workers=1) as executor:
+            while True:
+                now = time.perf_counter()
+
+                for camera in cameras:
+                    # If we use async_read when fps is None, the loop will go full speed, and we will end up
+                    # saving the same images from the cameras multiple times until the RAM/disk is full.
+                    image = camera.read() if fps is None else camera.async_read()
+                    if image is None:
+                        print("No Frame")
+
+                    bgr_converted_image = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
+
+                    executor.submit(
+                        save_image,
+                        bgr_converted_image,
+                        camera.serial_number,
+                        frame_index,
+                        images_dir,
+                    )
+
+                if fps is not None:
+                    dt_s = time.perf_counter() - now
+                    busy_wait(1 / fps - dt_s)
+
+                if time.perf_counter() - start_time > record_time_s:
+                    break
+
+                print(f"Frame: {frame_index:04d}\tLatency (ms): {(time.perf_counter() - now) * 1000:.2f}")
+
+                frame_index += 1
+    finally:
+        print(f"Images have been saved to {images_dir}")
+        for camera in cameras:
+            camera.disconnect()
+
+
+@dataclass
+class IntelRealSenseCameraConfig:
+    """
+    Example of tested options for Intel Real Sense D405:
+
+    ```python
+    IntelRealSenseCameraConfig(30, 640, 480)
+    IntelRealSenseCameraConfig(60, 640, 480)
+    IntelRealSenseCameraConfig(90, 640, 480)
+    IntelRealSenseCameraConfig(30, 1280, 720)
+    IntelRealSenseCameraConfig(30, 640, 480, use_depth=True)
+    IntelRealSenseCameraConfig(30, 640, 480, rotation=90)
+    ```
+    """
+
+    fps: int | None = None
+    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
+    mock: bool = False
+
+    def __post_init__(self):
+        if self.color_mode not in ["rgb", "bgr"]:
+            raise ValueError(
+                f"`color_mode` is expected to be 'rgb' or 'bgr', but {self.color_mode} is provided."
+            )
+
+        self.channels = 3
+
+        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:
+            raise ValueError(
+                "For `fps`, `width` and `height`, either all of them need to be set, or none of them, "
+                f"but {self.fps=}, {self.width=}, {self.height=} were provided."
+            )
+
+        if self.rotation not in [-90, None, 90, 180]:
+            raise ValueError(f"`rotation` must be in [-90, None, 90, 180] (got {self.rotation})")
+
+
+class IntelRealSenseCamera:
+    """
+    The IntelRealSenseCamera class is similar to OpenCVCamera class but adds additional features for Intel Real Sense cameras:
+    - is instantiated with the serial number of the camera - won't randomly change as it can be the case of OpenCVCamera for Linux,
+    - can also be instantiated with the camera's name — if it's unique — using IntelRealSenseCamera.init_from_name(),
+    - depth map can be returned.
+
+    To find the camera indices of your cameras, you can run our utility script that will save a few frames for each camera:
+    ```bash
+    python lerobot/common/robot_devices/cameras/intelrealsense.py \
+    --images-dir outputs/images_from_intelrealsense_cameras
+    ```
+
+    When an IntelRealSenseCamera is instantiated, if no specific config is provided, the default fps, width, height and color_mode
+    of the given camera will be used.
+
+    Example of usage:
+    ```python
+    # Instantiate with its serial number
+    camera = IntelRealSenseCamera(128422271347)
+    # Or by its name if it's unique
+    camera = IntelRealSenseCamera.init_from_name("Intel RealSense D405")
+    camera.connect()
+    color_image = camera.read()
+    # when done using the camera, consider disconnecting
+    camera.disconnect()
+    ```
+
+    Example of changing default fps, width, height and color_mode:
+    ```python
+    camera = IntelRealSenseCamera(serial_number, fps=30, width=1280, height=720)
+    camera = connect()  # applies the settings, might error out if these settings are not compatible with the camera
+
+    camera = IntelRealSenseCamera(serial_number, fps=90, width=640, height=480)
+    camera = connect()
+
+    camera = IntelRealSenseCamera(serial_number, fps=90, width=640, height=480, color_mode="bgr")
+    camera = connect()
+    ```
+
+    Example of returning depth:
+    ```python
+    camera = IntelRealSenseCamera(serial_number, use_depth=True)
+    camera.connect()
+    color_image, depth_map = camera.read()
+    ```
+    """
+
+    def __init__(
+        self,
+        serial_number: int,
+        config: IntelRealSenseCameraConfig | None = None,
+        **kwargs,
+    ):
+        if config is None:
+            config = IntelRealSenseCameraConfig()
+
+        # Overwrite the config arguments using kwargs
+        config = replace(config, **kwargs)
+
+        self.serial_number = serial_number
+        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
+        self.mock = config.mock
+
+        self.camera = None
+        self.is_connected = False
+        self.thread = None
+        self.stop_event = None
+        self.color_image = None
+        self.depth_map = None
+        self.logs = {}
+
+        if self.mock:
+            import tests.mock_cv2 as cv2
+        else:
+            import cv2
+
+        # TODO(alibets): Do we keep original width/height or do we define them after rotation?
+        self.rotation = None
+        if config.rotation == -90:
+            self.rotation = cv2.ROTATE_90_COUNTERCLOCKWISE
+        elif config.rotation == 90:
+            self.rotation = cv2.ROTATE_90_CLOCKWISE
+        elif config.rotation == 180:
+            self.rotation = cv2.ROTATE_180
+
+    @classmethod
+    def init_from_name(cls, name: str, config: IntelRealSenseCameraConfig | None = None, **kwargs):
+        camera_infos = find_cameras()
+        camera_names = [cam["name"] for cam in camera_infos]
+        this_name_count = Counter(camera_names)[name]
+        if this_name_count > 1:
+            # TODO(aliberts): Test this with multiple identical cameras (Aloha)
+            raise ValueError(
+                f"Multiple {name} cameras have been detected. Please use their serial number to instantiate them."
+            )
+
+        name_to_serial_dict = {cam["name"]: cam["serial_number"] for cam in camera_infos}
+        cam_sn = name_to_serial_dict[name]
+
+        if config is None:
+            config = IntelRealSenseCameraConfig()
+
+        # Overwrite the config arguments using kwargs
+        config = replace(config, **kwargs)
+
+        return cls(serial_number=cam_sn, config=config, **kwargs)
+
+    def connect(self):
+        if self.is_connected:
+            raise RobotDeviceAlreadyConnectedError(
+                f"IntelRealSenseCamera({self.serial_number}) is already connected."
+            )
+
+        if self.mock:
+            import tests.mock_pyrealsense2 as rs
+        else:
+            import pyrealsense2 as rs
+
+        config = rs.config()
+        config.enable_device(str(self.serial_number))
+
+        if self.fps and self.width and self.height:
+            # TODO(rcadene): can we set rgb8 directly?
+            config.enable_stream(rs.stream.color, self.width, self.height, rs.format.rgb8, self.fps)
+        else:
+            config.enable_stream(rs.stream.color)
+
+        if self.use_depth:
+            if self.fps and self.width and self.height:
+                config.enable_stream(rs.stream.depth, self.width, self.height, rs.format.z16, self.fps)
+            else:
+                config.enable_stream(rs.stream.depth)
+
+        self.camera = rs.pipeline()
+        try:
+            profile = self.camera.start(config)
+            is_camera_open = True
+        except RuntimeError:
+            is_camera_open = False
+            traceback.print_exc()
+
+        # If the camera doesn't work, display the camera indices corresponding to
+        # valid cameras.
+        if not is_camera_open:
+            # Verify that the provided `serial_number` is valid before printing the traceback
+            camera_infos = find_cameras()
+            serial_numbers = [cam["serial_number"] for cam in camera_infos]
+            if self.serial_number not in serial_numbers:
+                raise ValueError(
+                    f"`serial_number` is expected to be one of these available cameras {serial_numbers}, but {self.serial_number} is provided instead. "
+                    "To find the serial number you should use, run `python lerobot/common/robot_devices/cameras/intelrealsense.py`."
+                )
+
+            raise OSError(f"Can't access IntelRealSenseCamera({self.serial_number}).")
+
+        color_stream = profile.get_stream(rs.stream.color)
+        color_profile = color_stream.as_video_stream_profile()
+        actual_fps = color_profile.fps()
+        actual_width = color_profile.width()
+        actual_height = color_profile.height()
+
+        # Using `math.isclose` since actual fps can be a float (e.g. 29.9 instead of 30)
+        if self.fps is not None and not math.isclose(self.fps, actual_fps, rel_tol=1e-3):
+            # Using `OSError` since it's a broad that encompasses issues related to device communication
+            raise OSError(
+                f"Can't set {self.fps=} for IntelRealSenseCamera({self.serial_number}). Actual value is {actual_fps}."
+            )
+        if self.width is not None and self.width != actual_width:
+            raise OSError(
+                f"Can't set {self.width=} for IntelRealSenseCamera({self.serial_number}). Actual value is {actual_width}."
+            )
+        if self.height is not None and self.height != actual_height:
+            raise OSError(
+                f"Can't set {self.height=} for IntelRealSenseCamera({self.serial_number}). Actual value is {actual_height}."
+            )
+
+        self.fps = round(actual_fps)
+        self.width = round(actual_width)
+        self.height = round(actual_height)
+
+        self.is_connected = True
+
+    def read(self, temporary_color: str | None = None) -> np.ndarray | tuple[np.ndarray, np.ndarray]:
+        """Read a frame from the camera returned in the format height x width x channels (e.g. 480 x 640 x 3)
+        of type `np.uint8`, contrarily to the pytorch format which is float channel first.
+
+        When `use_depth=True`, returns a tuple `(color_image, depth_map)` with a depth map in the format
+        height x width (e.g. 480 x 640) of type np.uint16.
+
+        Note: Reading a frame is done every `camera.fps` times per second, and it is blocking.
+        If you are reading data from other sensors, we advise to use `camera.async_read()` which is non blocking version of `camera.read()`.
+        """
+        if not self.is_connected:
+            raise RobotDeviceNotConnectedError(
+                f"IntelRealSenseCamera({self.serial_number}) is not connected. Try running `camera.connect()` first."
+            )
+
+        if self.mock:
+            import tests.mock_cv2 as cv2
+        else:
+            import cv2
+
+        start_time = time.perf_counter()
+
+        frame = self.camera.wait_for_frames(timeout_ms=5000)
+
+        color_frame = frame.get_color_frame()
+
+        if not color_frame:
+            raise OSError(f"Can't capture color image from IntelRealSenseCamera({self.serial_number}).")
+
+        color_image = np.asanyarray(color_frame.get_data())
+
+        requested_color_mode = self.color_mode if temporary_color is None else temporary_color
+        if requested_color_mode not in ["rgb", "bgr"]:
+            raise ValueError(
+                f"Expected color values are 'rgb' or 'bgr', but {requested_color_mode} is provided."
+            )
+
+        # IntelRealSense uses RGB format as default (red, green, blue).
+        if requested_color_mode == "bgr":
+            color_image = cv2.cvtColor(color_image, cv2.COLOR_RGB2BGR)
+
+        h, w, _ = color_image.shape
+        if h != self.height or w != self.width:
+            raise OSError(
+                f"Can't capture color image with expected height and width ({self.height} x {self.width}). ({h} x {w}) returned instead."
+            )
+
+        if self.rotation is not None:
+            color_image = cv2.rotate(color_image, self.rotation)
+
+        # log the number of seconds it took to read the image
+        self.logs["delta_timestamp_s"] = time.perf_counter() - start_time
+
+        # log the utc time at which the image was received
+        self.logs["timestamp_utc"] = capture_timestamp_utc()
+
+        if self.use_depth:
+            depth_frame = frame.get_depth_frame()
+            if not depth_frame:
+                raise OSError(f"Can't capture depth image from IntelRealSenseCamera({self.serial_number}).")
+
+            depth_map = np.asanyarray(depth_frame.get_data())
+
+            h, w = depth_map.shape
+            if h != self.height or w != self.width:
+                raise OSError(
+                    f"Can't capture depth map with expected height and width ({self.height} x {self.width}). ({h} x {w}) returned instead."
+                )
+
+            if self.rotation is not None:
+                depth_map = cv2.rotate(depth_map, self.rotation)
+
+            return color_image, depth_map
+        else:
+            return color_image
+
+    def read_loop(self):
+        while not self.stop_event.is_set():
+            if self.use_depth:
+                self.color_image, self.depth_map = self.read()
+            else:
+                self.color_image = self.read()
+
+    def async_read(self):
+        """Access the latest color image"""
+        if not self.is_connected:
+            raise RobotDeviceNotConnectedError(
+                f"IntelRealSenseCamera({self.serial_number}) is not connected. Try running `camera.connect()` first."
+            )
+
+        if self.thread is None:
+            self.stop_event = threading.Event()
+            self.thread = Thread(target=self.read_loop, args=())
+            self.thread.daemon = True
+            self.thread.start()
+
+        num_tries = 0
+        while self.color_image is None:
+            # TODO(rcadene, aliberts): intelrealsense has diverged compared to opencv over here
+            num_tries += 1
+            time.sleep(1 / self.fps)
+            if num_tries > self.fps and (self.thread.ident is None or not self.thread.is_alive()):
+                raise Exception(
+                    "The thread responsible for `self.async_read()` took too much time to start. There might be an issue. Verify that `self.thread.start()` has been called."
+                )
+
+        if self.use_depth:
+            return self.color_image, self.depth_map
+        else:
+            return self.color_image
+
+    def disconnect(self):
+        if not self.is_connected:
+            raise RobotDeviceNotConnectedError(
+                f"IntelRealSenseCamera({self.serial_number}) is not connected. Try running `camera.connect()` first."
+            )
+
+        if self.thread is not None and self.thread.is_alive():
+            # wait for the thread to finish
+            self.stop_event.set()
+            self.thread.join()
+            self.thread = None
+            self.stop_event = None
+
+        self.camera.stop()
+        self.camera = None
+
+        self.is_connected = False
+
+    def __del__(self):
+        if getattr(self, "is_connected", False):
+            self.disconnect()
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(
+        description="Save a few frames using `IntelRealSenseCamera` for all cameras connected to the computer, or a selected subset."
+    )
+    parser.add_argument(
+        "--serial-numbers",
+        type=int,
+        nargs="*",
+        default=None,
+        help="List of serial numbers used to instantiate the `IntelRealSenseCamera`. If not provided, find and use all available camera indices.",
+    )
+    parser.add_argument(
+        "--fps",
+        type=int,
+        default=30,
+        help="Set the number of frames recorded per seconds for all cameras. If not provided, use the default fps of each camera.",
+    )
+    parser.add_argument(
+        "--width",
+        type=str,
+        default=640,
+        help="Set the width for all cameras. If not provided, use the default width of each camera.",
+    )
+    parser.add_argument(
+        "--height",
+        type=str,
+        default=480,
+        help="Set the height for all cameras. If not provided, use the default height of each camera.",
+    )
+    parser.add_argument(
+        "--images-dir",
+        type=Path,
+        default="outputs/images_from_intelrealsense_cameras",
+        help="Set directory to save a few frames for each camera.",
+    )
+    parser.add_argument(
+        "--record-time-s",
+        type=float,
+        default=2.0,
+        help="Set the number of seconds used to record the frames. By default, 2 seconds.",
+    )
+    args = parser.parse_args()
+    save_images_from_cameras(**vars(args))

lerobot/common/robot_devices/cameras/opencv.py

@@ -13,17 +13,15 @@ from dataclasses import dataclass, replace
 from pathlib import Path
 from threading import Thread
 
-import cv2
 import numpy as np
 from PIL import Image
 
-from lerobot.common.robot_devices.utils import RobotDeviceAlreadyConnectedError, RobotDeviceNotConnectedError
+from lerobot.common.robot_devices.utils import (
+    RobotDeviceAlreadyConnectedError,
+    RobotDeviceNotConnectedError,
+    busy_wait,
+)
 from lerobot.common.utils.utils import capture_timestamp_utc
-from lerobot.scripts.control_robot import busy_wait
-
-# Use 1 thread to avoid blocking the main thread. Especially useful during data collection
-# when other threads are used to save the images.
-cv2.setNumThreads(1)
 
 # The maximum opencv device index depends on your operating system. For instance,
 # if you have 3 cameras, they should be associated to index 0, 1, and 2. This is the case
@@ -33,20 +31,44 @@ cv2.setNumThreads(1)
 MAX_OPENCV_INDEX = 60
 
 
-def find_camera_indices(raise_when_empty=False, max_index_search_range=MAX_OPENCV_INDEX):
+def find_cameras(raise_when_empty=False, max_index_search_range=MAX_OPENCV_INDEX, mock=False) -> list[dict]:
+    cameras = []
     if platform.system() == "Linux":
-        # Linux uses camera ports
         print("Linux detected. Finding available camera indices through scanning '/dev/video*' ports")
-        possible_camera_ids = []
-        for port in Path("/dev").glob("video*"):
-            camera_idx = int(str(port).replace("/dev/video", ""))
-            possible_camera_ids.append(camera_idx)
+        possible_ports = [str(port) for port in Path("/dev").glob("video*")]
+        ports = _find_cameras(possible_ports, mock=mock)
+        for port in ports:
+            cameras.append(
+                {
+                    "port": port,
+                    "index": int(port.removeprefix("/dev/video")),
+                }
+            )
     else:
         print(
             "Mac or Windows detected. Finding available camera indices through "
             f"scanning all indices from 0 to {MAX_OPENCV_INDEX}"
         )
-        possible_camera_ids = range(max_index_search_range)
+        possible_indices = range(max_index_search_range)
+        indices = _find_cameras(possible_indices, mock=mock)
+        for index in indices:
+            cameras.append(
+                {
+                    "port": None,
+                    "index": index,
+                }
+            )
+
+    return cameras
+
+
+def _find_cameras(
+    possible_camera_ids: list[int | str], raise_when_empty=False, mock=False
+) -> list[int | str]:
+    if mock:
+        import tests.mock_cv2 as cv2
+    else:
+        import cv2
 
     camera_ids = []
     for camera_idx in possible_camera_ids:
@@ -67,6 +89,16 @@ def find_camera_indices(raise_when_empty=False, max_index_search_range=MAX_OPENC
     return camera_ids
 
 
+def is_valid_unix_path(path: str) -> bool:
+    """Note: if 'path' points to a symlink, this will return True only if the target exists"""
+    p = Path(path)
+    return p.is_absolute() and p.exists()
+
+
+def get_camera_index_from_unix_port(port: Path) -> int:
+    return int(str(port.resolve()).removeprefix("/dev/video"))
+
+
 def save_image(img_array, camera_index, frame_index, images_dir):
     img = Image.fromarray(img_array)
     path = images_dir / f"camera_{camera_index:02d}_frame_{frame_index:06d}.png"
@@ -75,15 +107,26 @@ def save_image(img_array, camera_index, frame_index, images_dir):
 
 
 def save_images_from_cameras(
-    images_dir: Path, camera_ids: list[int] | None = None, fps=None, width=None, height=None, record_time_s=2
+    images_dir: Path,
+    camera_ids: list | None = None,
+    fps=None,
+    width=None,
+    height=None,
+    record_time_s=2,
+    mock=False,
 ):
-    if camera_ids is None:
-        camera_ids = find_camera_indices()
+    """
+    Initializes all the cameras and saves images to the directory. Useful to visually identify the camera
+    associated to a given camera index.
+    """
+    if camera_ids is None or len(camera_ids) == 0:
+        camera_infos = find_cameras(mock=mock)
+        camera_ids = [cam["index"] for cam in camera_infos]
 
     print("Connecting cameras")
     cameras = []
     for cam_idx in camera_ids:
-        camera = OpenCVCamera(cam_idx, fps=fps, width=width, height=height)
+        camera = OpenCVCamera(cam_idx, fps=fps, width=width, height=height, mock=mock)
         camera.connect()
         print(
             f"OpenCVCamera({camera.camera_index}, fps={camera.fps}, width={camera.width}, "
@@ -101,7 +144,7 @@ def save_images_from_cameras(
     print(f"Saving images to {images_dir}")
     frame_index = 0
     start_time = time.perf_counter()
-    with concurrent.futures.ThreadPoolExecutor(max_workers=4) as executor:
+    with concurrent.futures.ThreadPoolExecutor(max_workers=1) as executor:
         while True:
             now = time.perf_counter()
 
@@ -122,11 +165,11 @@ def save_images_from_cameras(
                 dt_s = time.perf_counter() - now
                 busy_wait(1 / fps - dt_s)
 
+            print(f"Frame: {frame_index:04d}\tLatency (ms): {(time.perf_counter() - now) * 1000:.2f}")
+
             if time.perf_counter() - start_time > record_time_s:
                 break
 
-            print(f"Frame: {frame_index:04d}\tLatency (ms): {(time.perf_counter() - now) * 1000:.2f}")
-
             frame_index += 1
 
     print(f"Images have been saved to {images_dir}")
@@ -149,13 +192,21 @@ 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
 
     def __post_init__(self):
         if self.color_mode not in ["rgb", "bgr"]:
             raise ValueError(
-                f"Expected color_mode values are 'rgb' or 'bgr', but {self.color_mode} is provided."
+                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})")
+
 
 class OpenCVCamera:
     """
@@ -168,7 +219,8 @@ class OpenCVCamera:
 
     To find the camera indices of your cameras, you can run our utility script that will be save a few frames for each camera:
     ```bash
-    python lerobot/common/robot_devices/cameras/opencv.py --images-dir outputs/images_from_opencv_cameras
+    python lerobot/common/robot_devices/cameras/opencv.py \
+    --images-dir outputs/images_from_opencv_cameras
     ```
 
     When an OpenCVCamera is instantiated, if no specific config is provided, the default fps, width, height and color_mode
@@ -196,17 +248,33 @@ class OpenCVCamera:
     ```
     """
 
-    def __init__(self, camera_index: int, config: OpenCVCameraConfig | None = None, **kwargs):
+    def __init__(self, camera_index: int | str, config: OpenCVCameraConfig | None = None, **kwargs):
         if config is None:
             config = OpenCVCameraConfig()
+
         # Overwrite config arguments using kwargs
         config = replace(config, **kwargs)
 
         self.camera_index = camera_index
+        self.port = None
+
+        # Linux uses ports for connecting to cameras
+        if platform.system() == "Linux":
+            if isinstance(self.camera_index, int):
+                self.port = Path(f"/dev/video{self.camera_index}")
+            elif isinstance(self.camera_index, str) and is_valid_unix_path(self.camera_index):
+                self.port = Path(self.camera_index)
+                # Retrieve the camera index from a potentially symlinked path
+                self.camera_index = get_camera_index_from_unix_port(self.port)
+            else:
+                raise ValueError(f"Please check the provided camera_index: {camera_index}")
+
         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
 
         self.camera = None
         self.is_connected = False
@@ -215,43 +283,60 @@ class OpenCVCamera:
         self.color_image = None
         self.logs = {}
 
+        if self.mock:
+            import tests.mock_cv2 as cv2
+        else:
+            import cv2
+
+        # TODO(aliberts): Do we keep original width/height or do we define them after rotation?
+        self.rotation = None
+        if config.rotation == -90:
+            self.rotation = cv2.ROTATE_90_COUNTERCLOCKWISE
+        elif config.rotation == 90:
+            self.rotation = cv2.ROTATE_90_CLOCKWISE
+        elif config.rotation == 180:
+            self.rotation = cv2.ROTATE_180
+
     def connect(self):
         if self.is_connected:
-            raise RobotDeviceAlreadyConnectedError(f"Camera {self.camera_index} is already connected.")
+            raise RobotDeviceAlreadyConnectedError(f"OpenCVCamera({self.camera_index}) is already connected.")
 
+        if self.mock:
+            import tests.mock_cv2 as cv2
+        else:
+            import cv2
+
+            # Use 1 thread to avoid blocking the main thread. Especially useful during data collection
+            # when other threads are used to save the images.
+            cv2.setNumThreads(1)
+
+        camera_idx = f"/dev/video{self.camera_index}" if platform.system() == "Linux" else self.camera_index
         # First create a temporary camera trying to access `camera_index`,
         # and verify it is a valid camera by calling `isOpened`.
-
-        if platform.system() == "Linux":
-            # Linux uses ports for connecting to cameras
-            tmp_camera = cv2.VideoCapture(f"/dev/video{self.camera_index}")
-        else:
-            tmp_camera = cv2.VideoCapture(self.camera_index)
-
+        tmp_camera = cv2.VideoCapture(camera_idx)
         is_camera_open = tmp_camera.isOpened()
         # Release camera to make it accessible for `find_camera_indices`
+        tmp_camera.release()
         del tmp_camera
 
         # If the camera doesn't work, display the camera indices corresponding to
         # valid cameras.
         if not is_camera_open:
             # Verify that the provided `camera_index` is valid before printing the traceback
-            available_cam_ids = find_camera_indices()
+            cameras_info = find_cameras()
+            available_cam_ids = [cam["index"] for cam in cameras_info]
             if self.camera_index not in available_cam_ids:
                 raise ValueError(
                     f"`camera_index` is expected to be one of these available cameras {available_cam_ids}, but {self.camera_index} is provided instead. "
-                    "To find the camera index you should use, run `python lerobot/common/robot_devices/cameras/opencv.py`."
+                    "To find the camera index you should use, run `python lerobot/lerobot/common/robot_devices/cameras/opencv.py`."
                 )
 
-            raise OSError(f"Can't access camera {self.camera_index}.")
+            raise OSError(f"Can't access OpenCVCamera({camera_idx}).")
 
         # Secondly, create the camera that will be used downstream.
         # Note: For some unknown reason, calling `isOpened` blocks the camera which then
         # needs to be re-created.
-        if platform.system() == "Linux":
-            self.camera = cv2.VideoCapture(f"/dev/video{self.camera_index}")
-        else:
-            self.camera = cv2.VideoCapture(self.camera_index)
+        self.camera = cv2.VideoCapture(camera_idx)
 
         if self.fps is not None:
             self.camera.set(cv2.CAP_PROP_FPS, self.fps)
@@ -264,28 +349,30 @@ class OpenCVCamera:
         actual_width = self.camera.get(cv2.CAP_PROP_FRAME_WIDTH)
         actual_height = self.camera.get(cv2.CAP_PROP_FRAME_HEIGHT)
 
+        # Using `math.isclose` since actual fps can be a float (e.g. 29.9 instead of 30)
         if self.fps is not None and not math.isclose(self.fps, actual_fps, rel_tol=1e-3):
+            # Using `OSError` since it's a broad that encompasses issues related to device communication
             raise OSError(
-                f"Can't set {self.fps=} for camera {self.camera_index}. Actual value is {actual_fps}."
+                f"Can't set {self.fps=} for OpenCVCamera({self.camera_index}). Actual value is {actual_fps}."
             )
-        if self.width is not None and self.width != actual_width:
+        if self.width is not None and not math.isclose(self.width, actual_width, rel_tol=1e-3):
             raise OSError(
-                f"Can't set {self.width=} for camera {self.camera_index}. Actual value is {actual_width}."
+                f"Can't set {self.width=} for OpenCVCamera({self.camera_index}). Actual value is {actual_width}."
             )
-        if self.height is not None and self.height != actual_height:
+        if self.height is not None and not math.isclose(self.height, actual_height, rel_tol=1e-3):
             raise OSError(
-                f"Can't set {self.height=} for camera {self.camera_index}. Actual value is {actual_height}."
+                f"Can't set {self.height=} for OpenCVCamera({self.camera_index}). Actual value is {actual_height}."
             )
 
-        self.fps = actual_fps
-        self.width = actual_width
-        self.height = actual_height
+        self.fps = round(actual_fps)
+        self.width = round(actual_width)
+        self.height = round(actual_height)
 
         self.is_connected = True
 
     def read(self, temporary_color_mode: str | None = None) -> np.ndarray:
         """Read a frame from the camera returned in the format (height, width, channels)
-        (e.g. (640, 480, 3)), contrarily to the pytorch format which is channel first.
+        (e.g. 480 x 640 x 3), contrarily to the pytorch format which is channel first.
 
         Note: Reading a frame is done every `camera.fps` times per second, and it is blocking.
         If you are reading data from other sensors, we advise to use `camera.async_read()` which is non blocking version of `camera.read()`.
@@ -298,6 +385,7 @@ class OpenCVCamera:
         start_time = time.perf_counter()
 
         ret, color_image = self.camera.read()
+
         if not ret:
             raise OSError(f"Can't capture color image from camera {self.camera_index}.")
 
@@ -308,10 +396,15 @@ class OpenCVCamera:
                 f"Expected color values are 'rgb' or 'bgr', but {requested_color_mode} is provided."
             )
 
-        # OpenCV uses BGR format as default (blue, green red) for all operations, including displaying images.
+        # OpenCV uses BGR format as default (blue, green, red) for all operations, including displaying images.
         # However, Deep Learning framework such as LeRobot uses RGB format as default to train neural networks,
         # so we convert the image color from BGR to RGB.
         if requested_color_mode == "rgb":
+            if self.mock:
+                import tests.mock_cv2 as cv2
+            else:
+                import cv2
+
             color_image = cv2.cvtColor(color_image, cv2.COLOR_BGR2RGB)
 
         h, w, _ = color_image.shape
@@ -320,17 +413,25 @@ class OpenCVCamera:
                 f"Can't capture color image with expected height and width ({self.height} x {self.width}). ({h} x {w}) returned instead."
             )
 
+        if self.rotation is not None:
+            color_image = cv2.rotate(color_image, self.rotation)
+
         # log the number of seconds it took to read the image
         self.logs["delta_timestamp_s"] = time.perf_counter() - start_time
 
         # log the utc time at which the image was received
         self.logs["timestamp_utc"] = capture_timestamp_utc()
 
+        self.color_image = color_image
+
         return color_image
 
     def read_loop(self):
-        while self.stop_event is None or not self.stop_event.is_set():
-            self.color_image = self.read()
+        while not self.stop_event.is_set():
+            try:
+                self.color_image = self.read()
+            except Exception as e:
+                print(f"Error reading in thread: {e}")
 
     def async_read(self):
         if not self.is_connected:
@@ -345,15 +446,14 @@ class OpenCVCamera:
             self.thread.start()
 
         num_tries = 0
-        while self.color_image is None:
-            num_tries += 1
-            time.sleep(1 / self.fps)
-            if num_tries > self.fps and (self.thread.ident is None or not self.thread.is_alive()):
-                raise Exception(
-                    "The thread responsible for `self.async_read()` took too much time to start. There might be an issue. Verify that `self.thread.start()` has been called."
-                )
+        while True:
+            if self.color_image is not None:
+                return self.color_image
 
-        return self.color_image
+            time.sleep(1 / self.fps)
+            num_tries += 1
+            if num_tries > self.fps * 2:
+                raise TimeoutError("Timed out waiting for async_read() to start.")
 
     def disconnect(self):
         if not self.is_connected:
@@ -361,16 +461,14 @@ class OpenCVCamera:
                 f"OpenCVCamera({self.camera_index}) is not connected. Try running `camera.connect()` first."
             )
 
-        if self.thread is not None and self.thread.is_alive():
-            # wait for the thread to finish
+        if self.thread is not None:
             self.stop_event.set()
-            self.thread.join()
+            self.thread.join()  # wait for the thread to finish
             self.thread = None
             self.stop_event = None
 
         self.camera.release()
         self.camera = None
-
         self.is_connected = False
 
     def __del__(self):
@@ -416,7 +514,7 @@ if __name__ == "__main__":
     parser.add_argument(
         "--record-time-s",
         type=float,
-        default=2.0,
+        default=4.0,
         help="Set the number of seconds used to record the frames. By default, 2 seconds.",
     )
     args = parser.parse_args()

lerobot/common/robot_devices/cameras/utils.py

@@ -1,55 +1,8 @@
-from pathlib import Path
 from typing import Protocol
 
-import cv2
-import einops
 import numpy as np
 
 
-def write_shape_on_image_inplace(image):
-    height, width = image.shape[:2]
-    text = f"Width: {width} Height: {height}"
-
-    # Define the font, scale, color, and thickness
-    font = cv2.FONT_HERSHEY_SIMPLEX
-    font_scale = 1
-    color = (255, 0, 0)  # Blue in BGR
-    thickness = 2
-
-    position = (10, height - 10)  # 10 pixels from the bottom-left corner
-    cv2.putText(image, text, position, font, font_scale, color, thickness)
-
-
-def save_color_image(image, path, write_shape=False):
-    path = Path(path)
-    path.parent.mkdir(parents=True, exist_ok=True)
-    if write_shape:
-        write_shape_on_image_inplace(image)
-    cv2.imwrite(str(path), image)
-
-
-def save_depth_image(depth, path, write_shape=False):
-    path = Path(path)
-    path.parent.mkdir(parents=True, exist_ok=True)
-
-    # Apply colormap on depth image (image must be converted to 8-bit per pixel first)
-    depth_image = cv2.applyColorMap(cv2.convertScaleAbs(depth, alpha=0.03), cv2.COLORMAP_JET)
-
-    if write_shape:
-        write_shape_on_image_inplace(depth_image)
-    cv2.imwrite(str(path), depth_image)
-
-
-def convert_torch_image_to_cv2(tensor, rgb_to_bgr=True):
-    assert tensor.ndim == 3
-    c, h, w = tensor.shape
-    assert c < h and c < w
-    color_image = einops.rearrange(tensor, "c h w -> h w c").numpy()
-    if rgb_to_bgr:
-        color_image = cv2.cvtColor(color_image, cv2.COLOR_RGB2BGR)
-    return color_image
-
-
 # Defines a camera type
 class Camera(Protocol):
     def connect(self): ...

lerobot/common/robot_devices/control_utils.py

@@ -0,0 +1,332 @@
+########################################################################################
+# Utilities
+########################################################################################
+
+
+import logging
+import time
+import traceback
+from contextlib import nullcontext
+from copy import copy
+from functools import cache
+
+import cv2
+import torch
+import tqdm
+from termcolor import colored
+
+from lerobot.common.datasets.image_writer import safe_stop_image_writer
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+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
+from lerobot.common.utils.utils import get_safe_torch_device, init_hydra_config, set_global_seed
+from lerobot.scripts.eval import get_pretrained_policy_path
+
+
+def log_control_info(robot: Robot, dt_s, episode_index=None, frame_index=None, fps=None):
+    log_items = []
+    if episode_index is not None:
+        log_items.append(f"ep:{episode_index}")
+    if frame_index is not None:
+        log_items.append(f"frame:{frame_index}")
+
+    def log_dt(shortname, dt_val_s):
+        nonlocal log_items, fps
+        info_str = f"{shortname}:{dt_val_s * 1000:5.2f} ({1/ dt_val_s:3.1f}hz)"
+        if fps is not None:
+            actual_fps = 1 / dt_val_s
+            if actual_fps < fps - 1:
+                info_str = colored(info_str, "yellow")
+        log_items.append(info_str)
+
+    # total step time displayed in milliseconds and its frequency
+    log_dt("dt", dt_s)
+
+    # TODO(aliberts): move robot-specific logs logic in robot.print_logs()
+    if not robot.robot_type.startswith("stretch"):
+        for name in robot.leader_arms:
+            key = f"read_leader_{name}_pos_dt_s"
+            if key in robot.logs:
+                log_dt("dtRlead", robot.logs[key])
+
+        for name in robot.follower_arms:
+            key = f"write_follower_{name}_goal_pos_dt_s"
+            if key in robot.logs:
+                log_dt("dtWfoll", robot.logs[key])
+
+            key = f"read_follower_{name}_pos_dt_s"
+            if key in robot.logs:
+                log_dt("dtRfoll", robot.logs[key])
+
+        for name in robot.cameras:
+            key = f"read_camera_{name}_dt_s"
+            if key in robot.logs:
+                log_dt(f"dtR{name}", robot.logs[key])
+
+    info_str = " ".join(log_items)
+    logging.info(info_str)
+
+
+@cache
+def is_headless():
+    """Detects if python is running without a monitor."""
+    try:
+        import pynput  # noqa
+
+        return False
+    except Exception:
+        print(
+            "Error trying to import pynput. Switching to headless mode. "
+            "As a result, the video stream from the cameras won't be shown, "
+            "and you won't be able to change the control flow with keyboards. "
+            "For more info, see traceback below.\n"
+        )
+        traceback.print_exc()
+        print()
+        return True
+
+
+def has_method(_object: object, method_name: str):
+    return hasattr(_object, method_name) and callable(getattr(_object, method_name))
+
+
+def predict_action(observation, policy, device, use_amp):
+    observation = copy(observation)
+    with (
+        torch.inference_mode(),
+        torch.autocast(device_type=device.type) if device.type == "cuda" and use_amp else nullcontext(),
+    ):
+        # Convert to pytorch format: channel first and float32 in [0,1] with batch dimension
+        for name in observation:
+            if "image" in name:
+                observation[name] = observation[name].type(torch.float32) / 255
+                observation[name] = observation[name].permute(2, 0, 1).contiguous()
+            observation[name] = observation[name].unsqueeze(0)
+            observation[name] = observation[name].to(device)
+
+        # Compute the next action with the policy
+        # based on the current observation
+        action = policy.select_action(observation)
+
+        # Remove batch dimension
+        action = action.squeeze(0)
+
+        # Move to cpu, if not already the case
+        action = action.to("cpu")
+
+    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.
+    events = {}
+    events["exit_early"] = False
+    events["rerecord_episode"] = False
+    events["stop_recording"] = 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.esc:
+                print("Escape key pressed. Stopping data recording...")
+                events["stop_recording"] = True
+                events["exit_early"] = True
+        except Exception as e:
+            print(f"Error handling key press: {e}")
+
+    listener = keyboard.Listener(on_press=on_press)
+    listener.start()
+
+    return listener, events
+
+
+def init_policy(pretrained_policy_name_or_path, policy_overrides):
+    """Instantiate the policy and load fps, device and use_amp from config yaml"""
+    pretrained_policy_path = get_pretrained_policy_path(pretrained_policy_name_or_path)
+    hydra_cfg = init_hydra_config(pretrained_policy_path / "config.yaml", policy_overrides)
+    policy = make_policy(hydra_cfg=hydra_cfg, pretrained_policy_name_or_path=pretrained_policy_path)
+
+    # Check device is available
+    device = get_safe_torch_device(hydra_cfg.device, log=True)
+    use_amp = hydra_cfg.use_amp
+    policy_fps = hydra_cfg.env.fps
+
+    policy.eval()
+    policy.to(device)
+
+    torch.backends.cudnn.benchmark = True
+    torch.backends.cuda.matmul.allow_tf32 = True
+    set_global_seed(hydra_cfg.seed)
+    return policy, policy_fps, device, use_amp
+
+
+def warmup_record(
+    robot,
+    events,
+    enable_teloperation,
+    warmup_time_s,
+    display_cameras,
+    fps,
+):
+    control_loop(
+        robot=robot,
+        control_time_s=warmup_time_s,
+        display_cameras=display_cameras,
+        events=events,
+        fps=fps,
+        teleoperate=enable_teloperation,
+    )
+
+
+def record_episode(
+    robot,
+    dataset,
+    events,
+    episode_time_s,
+    display_cameras,
+    policy,
+    device,
+    use_amp,
+    fps,
+):
+    control_loop(
+        robot=robot,
+        control_time_s=episode_time_s,
+        display_cameras=display_cameras,
+        dataset=dataset,
+        events=events,
+        policy=policy,
+        device=device,
+        use_amp=use_amp,
+        fps=fps,
+        teleoperate=policy is None,
+    )
+
+
+@safe_stop_image_writer
+def control_loop(
+    robot,
+    control_time_s=None,
+    teleoperate=False,
+    display_cameras=False,
+    dataset: LeRobotDataset | None = None,
+    events=None,
+    policy=None,
+    device=None,
+    use_amp=None,
+    fps=None,
+):
+    # TODO(rcadene): Add option to record logs
+    if not robot.is_connected:
+        robot.connect()
+
+    if events is None:
+        events = {"exit_early": False}
+
+    if control_time_s is None:
+        control_time_s = float("inf")
+
+    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:
+        raise ValueError(f"The dataset fps should be equal to requested fps ({dataset['fps']} != {fps}).")
+
+    timestamp = 0
+    start_episode_t = time.perf_counter()
+    while timestamp < control_time_s:
+        start_loop_t = time.perf_counter()
+
+        if teleoperate:
+            observation, action = robot.teleop_step(record_data=True)
+        else:
+            observation = robot.capture_observation()
+
+            if policy is not None:
+                pred_action = predict_action(observation, policy, device, use_amp)
+                # Action can eventually be clipped using `max_relative_target`,
+                # so action actually sent is saved in the dataset.
+                action = robot.send_action(pred_action)
+                action = {"action": action}
+
+        if dataset is not None:
+            frame = {**observation, **action}
+            dataset.add_frame(frame)
+
+        if display_cameras and not is_headless():
+            image_keys = [key for key in observation if "image" in key]
+            for key in image_keys:
+                cv2.imshow(key, cv2.cvtColor(observation[key].numpy(), 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"]:
+            events["exit_early"] = False
+            break
+
+
+def reset_environment(robot, events, reset_time_s):
+    # TODO(rcadene): refactor warmup_record and reset_environment
+    # TODO(alibets): allow for teleop during reset
+    if has_method(robot, "teleop_safety_stop"):
+        robot.teleop_safety_stop()
+
+    timestamp = 0
+    start_vencod_t = time.perf_counter()
+
+    # Wait if necessary
+    with tqdm.tqdm(total=reset_time_s, desc="Waiting") as pbar:
+        while timestamp < reset_time_s:
+            time.sleep(1)
+            timestamp = time.perf_counter() - start_vencod_t
+            pbar.update(1)
+            if events["exit_early"]:
+                events["exit_early"] = False
+                break
+
+
+def stop_recording(robot, listener, display_cameras):
+    robot.disconnect()
+
+    if not is_headless():
+        if listener is not None:
+            listener.stop()
+
+        if display_cameras:
+            cv2.destroyAllWindows()
+
+
+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):
+        raise ValueError(
+            f"Your dataset name begins by 'eval_' ({dataset_name}) but no policy is provided ({policy})."
+        )

lerobot/common/robot_devices/motors/dynamixel.py

@@ -1,22 +1,12 @@
 import enum
+import logging
+import math
 import time
 import traceback
 from copy import deepcopy
-from pathlib import Path
 
 import numpy as np
 import tqdm
-from dynamixel_sdk import (
-    COMM_SUCCESS,
-    DXL_HIBYTE,
-    DXL_HIWORD,
-    DXL_LOBYTE,
-    DXL_LOWORD,
-    GroupSyncRead,
-    GroupSyncWrite,
-    PacketHandler,
-    PortHandler,
-)
 
 from lerobot.common.robot_devices.utils import RobotDeviceAlreadyConnectedError, RobotDeviceNotConnectedError
 from lerobot.common.utils.utils import capture_timestamp_utc
@@ -27,11 +17,28 @@ TIMEOUT_MS = 1000
 
 MAX_ID_RANGE = 252
 
+# The following bounds define the lower and upper joints range (after calibration).
+# For joints in degree (i.e. revolute joints), their nominal range is [-180, 180] degrees
+# which corresponds to a half rotation on the left and half rotation on the right.
+# Some joints might require higher range, so we allow up to [-270, 270] degrees until
+# an error is raised.
+LOWER_BOUND_DEGREE = -270
+UPPER_BOUND_DEGREE = 270
+# For joints in percentage (i.e. joints that move linearly like the prismatic joint of a gripper),
+# their nominal range is [0, 100] %. For instance, for Aloha gripper, 0% is fully
+# closed, and 100% is fully open. To account for slight calibration issue, we allow up to
+# [-10, 110] until an error is raised.
+LOWER_BOUND_LINEAR = -10
+UPPER_BOUND_LINEAR = 110
+
+HALF_TURN_DEGREE = 180
+
 # https://emanual.robotis.com/docs/en/dxl/x/xl330-m077
 # https://emanual.robotis.com/docs/en/dxl/x/xl330-m288
 # https://emanual.robotis.com/docs/en/dxl/x/xl430-w250
 # https://emanual.robotis.com/docs/en/dxl/x/xm430-w350
 # https://emanual.robotis.com/docs/en/dxl/x/xm540-w270
+# https://emanual.robotis.com/docs/en/dxl/x/xc430-w150
 
 # data_name: (address, size_byte)
 X_SERIES_CONTROL_TABLE = {
@@ -109,6 +116,7 @@ MODEL_CONTROL_TABLE = {
     "xl430-w250": X_SERIES_CONTROL_TABLE,
     "xm430-w350": X_SERIES_CONTROL_TABLE,
     "xm540-w270": X_SERIES_CONTROL_TABLE,
+    "xc430-w150": X_SERIES_CONTROL_TABLE,
 }
 
 MODEL_RESOLUTION = {
@@ -118,6 +126,7 @@ MODEL_RESOLUTION = {
     "xl430-w250": 4096,
     "xm430-w350": 4096,
     "xm540-w270": 4096,
+    "xc430-w150": 4096,
 }
 
 MODEL_BAUDRATE_TABLE = {
@@ -127,44 +136,47 @@ MODEL_BAUDRATE_TABLE = {
     "xl430-w250": X_SERIES_BAUDRATE_TABLE,
     "xm430-w350": X_SERIES_BAUDRATE_TABLE,
     "xm540-w270": X_SERIES_BAUDRATE_TABLE,
+    "xc430-w150": X_SERIES_BAUDRATE_TABLE,
 }
 
 NUM_READ_RETRY = 10
 NUM_WRITE_RETRY = 10
 
 
-def convert_degrees_to_steps(degrees: float | np.ndarray, models: str | list[str]):
-    """This function convert the degree range to the step range for indicating motors rotation.
-    It assums a motor achieves a full rotation by going from -180 degree position to +180.
+def convert_degrees_to_steps(degrees: float | np.ndarray, models: str | list[str]) -> np.ndarray:
+    """This function converts the degree range to the step range for indicating motors rotation.
+    It assumes a motor achieves a full rotation by going from -180 degree position to +180.
     The motor resolution (e.g. 4096) corresponds to the number of steps needed to achieve a full rotation.
     """
-    if isinstance(degrees, float):
-        degrees = np.array(degrees)
-
     resolutions = [MODEL_RESOLUTION[model] for model in models]
     steps = degrees / 180 * np.array(resolutions) / 2
     steps = steps.astype(int)
     return steps
 
 
-def convert_to_bytes(value, bytes):
+def convert_to_bytes(value, bytes, mock=False):
+    if mock:
+        return value
+
+    import dynamixel_sdk as dxl
+
     # Note: No need to convert back into unsigned int, since this byte preprocessing
     # already handles it for us.
     if bytes == 1:
         data = [
-            DXL_LOBYTE(DXL_LOWORD(value)),
+            dxl.DXL_LOBYTE(dxl.DXL_LOWORD(value)),
         ]
     elif bytes == 2:
         data = [
-            DXL_LOBYTE(DXL_LOWORD(value)),
-            DXL_HIBYTE(DXL_LOWORD(value)),
+            dxl.DXL_LOBYTE(dxl.DXL_LOWORD(value)),
+            dxl.DXL_HIBYTE(dxl.DXL_LOWORD(value)),
         ]
     elif bytes == 4:
         data = [
-            DXL_LOBYTE(DXL_LOWORD(value)),
-            DXL_HIBYTE(DXL_LOWORD(value)),
-            DXL_LOBYTE(DXL_HIWORD(value)),
-            DXL_HIBYTE(DXL_HIWORD(value)),
+            dxl.DXL_LOBYTE(dxl.DXL_LOWORD(value)),
+            dxl.DXL_HIBYTE(dxl.DXL_LOWORD(value)),
+            dxl.DXL_LOBYTE(dxl.DXL_HIWORD(value)),
+            dxl.DXL_HIBYTE(dxl.DXL_HIWORD(value)),
         ]
     else:
         raise NotImplementedError(
@@ -216,54 +228,29 @@ def assert_same_address(model_ctrl_table, motor_models, data_name):
         )
 
 
-def find_available_ports():
-    ports = []
-    for path in Path("/dev").glob("tty*"):
-        ports.append(str(path))
-    return ports
-
-
-def find_port():
-    print("Finding all available ports for the DynamixelMotorsBus.")
-    ports_before = find_available_ports()
-    print(ports_before)
-
-    print("Remove the usb cable from your DynamixelMotorsBus and press Enter when done.")
-    input()
-
-    time.sleep(0.5)
-    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 DynamixelMotorsBus is '{port}'")
-        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:
-        raise OSError(f"Could not detect the port. More than one port was found ({ports_diff}).")
-
-
 class TorqueMode(enum.Enum):
     ENABLED = 1
     DISABLED = 0
 
 
-class OperatingMode(enum.Enum):
-    VELOCITY = 1
-    POSITION = 3
-    EXTENDED_POSITION = 4
-    CURRENT_CONTROLLED_POSITION = 5
-    PWM = 16
-    UNKNOWN = -1
-
-
 class DriveMode(enum.Enum):
     NON_INVERTED = 0
     INVERTED = 1
 
 
+class CalibrationMode(enum.Enum):
+    # Joints with rotational motions are expressed in degrees in nominal range of [-180, 180]
+    DEGREE = 0
+    # Joints with linear motions (like gripper of Aloha) are experessed in nominal range of [0, 100]
+    LINEAR = 1
+
+
+class JointOutOfRangeError(Exception):
+    def __init__(self, message="Joint is out of range"):
+        self.message = message
+        super().__init__(self.message)
+
+
 class DynamixelMotorsBus:
     # TODO(rcadene): Add a script to find the motor indices without DynamixelWizzard2
     """
@@ -273,8 +260,8 @@ class DynamixelMotorsBus:
     A DynamixelMotorsBus instance requires a port (e.g. `DynamixelMotorsBus(port="/dev/tty.usbmodem575E0031751"`)).
     To find the port, you can run our utility script:
     ```bash
-    python lerobot/common/robot_devices/motors/dynamixel.py
-    >>> Finding all available ports for the DynamixelMotorsBus.
+    python lerobot/scripts/find_motors_bus_port.py
+    >>> Finding all available ports for the MotorBus.
     >>> ['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
     >>> Remove the usb cable from your DynamixelMotorsBus and press Enter when done.
     >>> The port of this DynamixelMotorsBus is /dev/tty.usbmodem575E0031751.
@@ -310,9 +297,11 @@ class DynamixelMotorsBus:
         motors: dict[str, tuple[int, str]],
         extra_model_control_table: dict[str, list[tuple]] | None = None,
         extra_model_resolution: dict[str, int] | None = None,
+        mock=False,
     ):
         self.port = port
         self.motors = motors
+        self.mock = mock
 
         self.model_ctrl_table = deepcopy(MODEL_CONTROL_TABLE)
         if extra_model_control_table:
@@ -336,8 +325,13 @@ class DynamixelMotorsBus:
                 f"DynamixelMotorsBus({self.port}) is already connected. Do not call `motors_bus.connect()` twice."
             )
 
-        self.port_handler = PortHandler(self.port)
-        self.packet_handler = PacketHandler(PROTOCOL_VERSION)
+        if self.mock:
+            import tests.mock_dynamixel_sdk as dxl
+        else:
+            import dynamixel_sdk as dxl
+
+        self.port_handler = dxl.PortHandler(self.port)
+        self.packet_handler = dxl.PacketHandler(PROTOCOL_VERSION)
 
         try:
             if not self.port_handler.openPort():
@@ -345,7 +339,7 @@ class DynamixelMotorsBus:
         except Exception:
             traceback.print_exc()
             print(
-                "\nTry running `python lerobot/common/robot_devices/motors/dynamixel.py` to make sure you are using the correct port.\n"
+                "\nTry running `python lerobot/scripts/find_motors_bus_port.py` to make sure you are using the correct port.\n"
             )
             raise
 
@@ -354,25 +348,18 @@ class DynamixelMotorsBus:
 
         self.port_handler.setPacketTimeoutMillis(TIMEOUT_MS)
 
-        # Set expected baudrate for the bus
-        self.set_bus_baudrate(BAUDRATE)
-
-        if not self.are_motors_configured():
-            input(
-                "\n/!\\ A configuration issue has been detected with your motors: \n"
-                "If it's the first time that you use these motors, press enter to configure your motors... but before "
-                "verify that all the cables are connected the proper way. If you find an issue, before making a modification, "
-                "kill the python process, unplug the power cord to not damage the motors, rewire correctly, then plug the power "
-                "again and relaunch the script.\n"
-            )
-            print()
-            self.configure_motors()
-
     def reconnect(self):
-        self.port_handler = PortHandler(self.port)
-        self.packet_handler = PacketHandler(PROTOCOL_VERSION)
+        if self.mock:
+            import tests.mock_dynamixel_sdk as dxl
+        else:
+            import dynamixel_sdk as dxl
+
+        self.port_handler = dxl.PortHandler(self.port)
+        self.packet_handler = dxl.PacketHandler(PROTOCOL_VERSION)
+
         if not self.port_handler.openPort():
             raise OSError(f"Failed to open port '{self.port}'.")
+
         self.is_connected = True
 
     def are_motors_configured(self):
@@ -384,120 +371,14 @@ class DynamixelMotorsBus:
             print(e)
             return False
 
-    def configure_motors(self):
-        # TODO(rcadene): This script assumes motors follow the X_SERIES baudrates
-        # TODO(rcadene): Refactor this function with intermediate high-level functions
-
-        print("Scanning all baudrates and motor indices")
-        all_baudrates = set(X_SERIES_BAUDRATE_TABLE.values())
-        ids_per_baudrate = {}
-        for baudrate in all_baudrates:
-            self.set_bus_baudrate(baudrate)
-            present_ids = self.find_motor_indices()
-            if len(present_ids) > 0:
-                ids_per_baudrate[baudrate] = present_ids
-        print(f"Motor indices detected: {ids_per_baudrate}")
-        print()
-
-        possible_baudrates = list(ids_per_baudrate.keys())
-        possible_ids = list({idx for sublist in ids_per_baudrate.values() for idx in sublist})
-        untaken_ids = list(set(range(MAX_ID_RANGE)) - set(possible_ids) - set(self.motor_indices))
-
-        # Connect successively one motor to the chain and write a unique random index for each
-        for i in range(len(self.motors)):
-            self.disconnect()
-            input(
-                "1. Unplug the power cord\n"
-                "2. Plug/unplug minimal number of cables to only have the first "
-                f"{i+1} motor(s) ({self.motor_names[:i+1]}) connected.\n"
-                "3. Re-plug the power cord\n"
-                "Press Enter to continue..."
-            )
-            print()
-            self.reconnect()
-
-            if i > 0:
-                try:
-                    self._read_with_motor_ids(self.motor_models, untaken_ids[:i], "ID")
-                except ConnectionError:
-                    print(f"Failed to read from {untaken_ids[:i+1]}. Make sure the power cord is plugged in.")
-                    input("Press Enter to continue...")
-                    print()
-                    self.reconnect()
-
-            print("Scanning possible baudrates and motor indices")
-            motor_found = False
-            for baudrate in possible_baudrates:
-                self.set_bus_baudrate(baudrate)
-                present_ids = self.find_motor_indices(possible_ids)
-                if len(present_ids) == 1:
-                    present_idx = present_ids[0]
-                    print(f"Detected motor with index {present_idx}")
-
-                    if baudrate != BAUDRATE:
-                        print(f"Setting its baudrate to {BAUDRATE}")
-                        baudrate_idx = list(X_SERIES_BAUDRATE_TABLE.values()).index(BAUDRATE)
-
-                        # The write can fail, so we allow retries
-                        for _ in range(NUM_WRITE_RETRY):
-                            self._write_with_motor_ids(
-                                self.motor_models, present_idx, "Baud_Rate", baudrate_idx
-                            )
-                            time.sleep(0.5)
-                            self.set_bus_baudrate(BAUDRATE)
-                            try:
-                                present_baudrate_idx = self._read_with_motor_ids(
-                                    self.motor_models, present_idx, "Baud_Rate"
-                                )
-                            except ConnectionError:
-                                print("Failed to write baudrate. Retrying.")
-                                self.set_bus_baudrate(baudrate)
-                                continue
-                            break
-                        else:
-                            raise
-
-                        if present_baudrate_idx != baudrate_idx:
-                            raise OSError("Failed to write baudrate.")
-
-                    print(f"Setting its index to a temporary untaken index ({untaken_ids[i]})")
-                    self._write_with_motor_ids(self.motor_models, present_idx, "ID", untaken_ids[i])
-
-                    present_idx = self._read_with_motor_ids(self.motor_models, untaken_ids[i], "ID")
-                    if present_idx != untaken_ids[i]:
-                        raise OSError("Failed to write index.")
-
-                    motor_found = True
-                    break
-                elif len(present_ids) > 1:
-                    raise OSError(f"More than one motor detected ({present_ids}), but only one was expected.")
-
-            if not motor_found:
-                raise OSError(
-                    "No motor found, but one new motor expected. Verify power cord is plugged in and retry."
-                )
-            print()
-
-        print(f"Setting expected motor indices: {self.motor_indices}")
-        self.set_bus_baudrate(BAUDRATE)
-        self._write_with_motor_ids(
-            self.motor_models, untaken_ids[: len(self.motors)], "ID", self.motor_indices
-        )
-        print()
-
-        if (self.read("ID") != self.motor_indices).any():
-            raise OSError("Failed to write motors indices.")
-
-        print("Configuration is done!")
-
-    def find_motor_indices(self, possible_ids=None):
+    def find_motor_indices(self, possible_ids=None, num_retry=2):
         if possible_ids is None:
             possible_ids = range(MAX_ID_RANGE)
 
         indices = []
         for idx in tqdm.tqdm(possible_ids):
             try:
-                present_idx = self._read_with_motor_ids(self.motor_models, [idx], "ID")[0]
+                present_idx = self.read_with_motor_ids(self.motor_models, [idx], "ID", num_retry=num_retry)[0]
             except ConnectionError:
                 continue
 
@@ -531,9 +412,22 @@ class DynamixelMotorsBus:
     def motor_indices(self) -> list[int]:
         return [idx for idx, _ in self.motors.values()]
 
-    def set_calibration(self, calibration: dict[str, tuple[int, bool]]):
+    def set_calibration(self, calibration: dict[str, list]):
         self.calibration = calibration
 
+    def apply_calibration_autocorrect(self, values: np.ndarray | list, motor_names: list[str] | None):
+        """This function applies the calibration, automatically detects out of range errors for motors values and attempts to correct.
+
+        For more info, see docstring of `apply_calibration` and `autocorrect_calibration`.
+        """
+        try:
+            values = self.apply_calibration(values, motor_names)
+        except JointOutOfRangeError as e:
+            print(e)
+            self.autocorrect_calibration(values, motor_names)
+            values = self.apply_calibration(values, motor_names)
+        return values
+
     def apply_calibration(self, values: np.ndarray | list, motor_names: list[str] | None):
         """Convert from unsigned int32 joint position range [0, 2**32[ to the universal float32 nominal degree range ]-180.0, 180.0[ with
         a "zero position" at 0 degree.
@@ -551,56 +445,205 @@ class DynamixelMotorsBus:
         if motor_names is None:
             motor_names = self.motor_names
 
-        # Convert from unsigned int32 original range [0, 2**32[ to centered signed int32 range [-2**31, 2**31[
-        values = values.astype(np.int32)
-
-        for i, name in enumerate(motor_names):
-            homing_offset, drive_mode = self.calibration[name]
-
-            # Update direction of rotation of the motor to match between leader and follower. In fact, the motor of the leader for a given joint
-            # can be assembled in an opposite direction in term of rotation than the motor of the follower on the same joint.
-            if drive_mode:
-                values[i] *= -1
-
-            # Convert from range [-2**31, 2**31[ to nominal range ]-resolution, resolution[ (e.g. ]-2048, 2048[)
-            values[i] += homing_offset
-
-        # Convert from range ]-resolution, resolution[ to the universal float32 centered degree range ]-180, 180[
+        # Convert from unsigned int32 original range [0, 2**32] to signed float32 range
         values = values.astype(np.float32)
+
         for i, name in enumerate(motor_names):
-            _, model = self.motors[name]
-            resolution = self.model_resolution[model]
-            values[i] = values[i] / (resolution // 2) * 180
+            calib_idx = self.calibration["motor_names"].index(name)
+            calib_mode = self.calibration["calib_mode"][calib_idx]
+
+            if CalibrationMode[calib_mode] == CalibrationMode.DEGREE:
+                drive_mode = self.calibration["drive_mode"][calib_idx]
+                homing_offset = self.calibration["homing_offset"][calib_idx]
+                _, model = self.motors[name]
+                resolution = self.model_resolution[model]
+
+                # Update direction of rotation of the motor to match between leader and follower.
+                # In fact, the motor of the leader for a given joint can be assembled in an
+                # opposite direction in term of rotation than the motor of the follower on the same joint.
+                if drive_mode:
+                    values[i] *= -1
+
+                # Convert from range [-2**31, 2**31] to
+                # nominal range [-resolution//2, resolution//2] (e.g. [-2048, 2048])
+                values[i] += homing_offset
+
+                # Convert from range [-resolution//2, resolution//2] to
+                # universal float32 centered degree range [-180, 180]
+                # (e.g. 2048 / (4096 // 2) * 180 = 180)
+                values[i] = values[i] / (resolution // 2) * HALF_TURN_DEGREE
+
+                if (values[i] < LOWER_BOUND_DEGREE) or (values[i] > UPPER_BOUND_DEGREE):
+                    raise JointOutOfRangeError(
+                        f"Wrong motor position range detected for {name}. "
+                        f"Expected to be in nominal range of [-{HALF_TURN_DEGREE}, {HALF_TURN_DEGREE}] degrees (a full rotation), "
+                        f"with a maximum range of [{LOWER_BOUND_DEGREE}, {UPPER_BOUND_DEGREE}] degrees to account for joints that can rotate a bit more, "
+                        f"but present value is {values[i]} degree. "
+                        "This might be due to a cable connection issue creating an artificial 360 degrees jump in motor values. "
+                        "You need to recalibrate by running: `python lerobot/scripts/control_robot.py calibrate`"
+                    )
+
+            elif CalibrationMode[calib_mode] == CalibrationMode.LINEAR:
+                start_pos = self.calibration["start_pos"][calib_idx]
+                end_pos = self.calibration["end_pos"][calib_idx]
+
+                # Rescale the present position to a nominal range [0, 100] %,
+                # useful for joints with linear motions like Aloha gripper
+                values[i] = (values[i] - start_pos) / (end_pos - start_pos) * 100
+
+                if (values[i] < LOWER_BOUND_LINEAR) or (values[i] > UPPER_BOUND_LINEAR):
+                    raise JointOutOfRangeError(
+                        f"Wrong motor position range detected for {name}. "
+                        f"Expected to be in nominal range of [0, 100] % (a full linear translation), "
+                        f"with a maximum range of [{LOWER_BOUND_LINEAR}, {UPPER_BOUND_LINEAR}] % to account for some imprecision during calibration, "
+                        f"but present value is {values[i]} %. "
+                        "This might be due to a cable connection issue creating an artificial jump in motor values. "
+                        "You need to recalibrate by running: `python lerobot/scripts/control_robot.py calibrate`"
+                    )
 
         return values
 
+    def autocorrect_calibration(self, values: np.ndarray | list, motor_names: list[str] | None):
+        """This function automatically detects issues with values of motors after calibration, and correct for these issues.
+
+        Some motors might have values outside of expected maximum bounds after calibration.
+        For instance, for a joint in degree, its value can be outside [-270, 270] degrees, which is totally unexpected given
+        a nominal range of [-180, 180] degrees, which represents half a turn to the left or right starting from zero position.
+
+        Known issues:
+        #1: Motor value randomly shifts of a full turn, caused by hardware/connection errors.
+        #2: Motor internal homing offset is shifted by a full turn, caused by using default calibration (e.g Aloha).
+        #3: motor internal homing offset is shifted by less or more than a full turn, caused by using default calibration
+            or by human error during manual calibration.
+
+        Issues #1 and #2 can be solved by shifting the calibration homing offset by a full turn.
+        Issue #3 will be visually detected by user and potentially captured by the safety feature `max_relative_target`,
+        that will slow down the motor, raise an error asking to recalibrate. Manual recalibrating will solve the issue.
+
+        Note: A full turn corresponds to 360 degrees but also to 4096 steps for a motor resolution of 4096.
+        """
+        if motor_names is None:
+            motor_names = self.motor_names
+
+        # Convert from unsigned int32 original range [0, 2**32] to signed float32 range
+        values = values.astype(np.float32)
+
+        for i, name in enumerate(motor_names):
+            calib_idx = self.calibration["motor_names"].index(name)
+            calib_mode = self.calibration["calib_mode"][calib_idx]
+
+            if CalibrationMode[calib_mode] == CalibrationMode.DEGREE:
+                drive_mode = self.calibration["drive_mode"][calib_idx]
+                homing_offset = self.calibration["homing_offset"][calib_idx]
+                _, model = self.motors[name]
+                resolution = self.model_resolution[model]
+
+                # Update direction of rotation of the motor to match between leader and follower.
+                # In fact, the motor of the leader for a given joint can be assembled in an
+                # opposite direction in term of rotation than the motor of the follower on the same joint.
+                if drive_mode:
+                    values[i] *= -1
+
+                # Convert from initial range to range [-180, 180] degrees
+                calib_val = (values[i] + homing_offset) / (resolution // 2) * HALF_TURN_DEGREE
+                in_range = (calib_val > LOWER_BOUND_DEGREE) and (calib_val < UPPER_BOUND_DEGREE)
+
+                # Solve this inequality to find the factor to shift the range into [-180, 180] degrees
+                # values[i] = (values[i] + homing_offset + resolution * factor) / (resolution // 2) * HALF_TURN_DEGREE
+                # - HALF_TURN_DEGREE <= (values[i] + homing_offset + resolution * factor) / (resolution // 2) * HALF_TURN_DEGREE <= HALF_TURN_DEGREE
+                # (- (resolution // 2) - values[i] - homing_offset) / resolution <= factor <= ((resolution // 2) - values[i] - homing_offset) / resolution
+                low_factor = (-(resolution // 2) - values[i] - homing_offset) / resolution
+                upp_factor = ((resolution // 2) - values[i] - homing_offset) / resolution
+
+            elif CalibrationMode[calib_mode] == CalibrationMode.LINEAR:
+                start_pos = self.calibration["start_pos"][calib_idx]
+                end_pos = self.calibration["end_pos"][calib_idx]
+
+                # Convert from initial range to range [0, 100] in %
+                calib_val = (values[i] - start_pos) / (end_pos - start_pos) * 100
+                in_range = (calib_val > LOWER_BOUND_LINEAR) and (calib_val < UPPER_BOUND_LINEAR)
+
+                # Solve this inequality to find the factor to shift the range into [0, 100] %
+                # values[i] = (values[i] - start_pos + resolution * factor) / (end_pos + resolution * factor - start_pos - resolution * factor) * 100
+                # values[i] = (values[i] - start_pos + resolution * factor) / (end_pos - start_pos) * 100
+                # 0 <= (values[i] - start_pos + resolution * factor) / (end_pos - start_pos) * 100 <= 100
+                # (start_pos - values[i]) / resolution <= factor <= (end_pos - values[i]) / resolution
+                low_factor = (start_pos - values[i]) / resolution
+                upp_factor = (end_pos - values[i]) / resolution
+
+            if not in_range:
+                # Get first integer between the two bounds
+                if low_factor < upp_factor:
+                    factor = math.ceil(low_factor)
+
+                    if factor > upp_factor:
+                        raise ValueError(f"No integer found between bounds [{low_factor=}, {upp_factor=}]")
+                else:
+                    factor = math.ceil(upp_factor)
+
+                    if factor > low_factor:
+                        raise ValueError(f"No integer found between bounds [{low_factor=}, {upp_factor=}]")
+
+                if CalibrationMode[calib_mode] == CalibrationMode.DEGREE:
+                    out_of_range_str = f"{LOWER_BOUND_DEGREE} < {calib_val} < {UPPER_BOUND_DEGREE} degrees"
+                    in_range_str = f"{LOWER_BOUND_DEGREE} < {calib_val} < {UPPER_BOUND_DEGREE} degrees"
+                elif CalibrationMode[calib_mode] == CalibrationMode.LINEAR:
+                    out_of_range_str = f"{LOWER_BOUND_LINEAR} < {calib_val} < {UPPER_BOUND_LINEAR} %"
+                    in_range_str = f"{LOWER_BOUND_LINEAR} < {calib_val} < {UPPER_BOUND_LINEAR} %"
+
+                logging.warning(
+                    f"Auto-correct calibration of motor '{name}' by shifting value by {abs(factor)} full turns, "
+                    f"from '{out_of_range_str}' to '{in_range_str}'."
+                )
+
+                # A full turn corresponds to 360 degrees but also to 4096 steps for a motor resolution of 4096.
+                self.calibration["homing_offset"][calib_idx] += resolution * factor
+
     def revert_calibration(self, values: np.ndarray | list, motor_names: list[str] | None):
         """Inverse of `apply_calibration`."""
         if motor_names is None:
             motor_names = self.motor_names
 
-        # Convert from the universal float32 centered degree range ]-180, 180[ to resolution range ]-resolution, resolution[
         for i, name in enumerate(motor_names):
-            _, model = self.motors[name]
-            resolution = self.model_resolution[model]
-            values[i] = values[i] / 180 * (resolution // 2)
+            calib_idx = self.calibration["motor_names"].index(name)
+            calib_mode = self.calibration["calib_mode"][calib_idx]
+
+            if CalibrationMode[calib_mode] == CalibrationMode.DEGREE:
+                drive_mode = self.calibration["drive_mode"][calib_idx]
+                homing_offset = self.calibration["homing_offset"][calib_idx]
+                _, model = self.motors[name]
+                resolution = self.model_resolution[model]
+
+                # Convert from nominal 0-centered degree range [-180, 180] to
+                # 0-centered resolution range (e.g. [-2048, 2048] for resolution=4096)
+                values[i] = values[i] / HALF_TURN_DEGREE * (resolution // 2)
+
+                # Substract the homing offsets to come back to actual motor range of values
+                # which can be arbitrary.
+                values[i] -= homing_offset
+
+                # Remove drive mode, which is the rotation direction of the motor, to come back to
+                # actual motor rotation direction which can be arbitrary.
+                if drive_mode:
+                    values[i] *= -1
+
+            elif CalibrationMode[calib_mode] == CalibrationMode.LINEAR:
+                start_pos = self.calibration["start_pos"][calib_idx]
+                end_pos = self.calibration["end_pos"][calib_idx]
+
+                # Convert from nominal lnear range of [0, 100] % to
+                # actual motor range of values which can be arbitrary.
+                values[i] = values[i] / 100 * (end_pos - start_pos) + start_pos
 
         values = np.round(values).astype(np.int32)
-
-        # Convert from nominal range ]-resolution, resolution[ to centered signed int32 range [-2**31, 2**31[
-        for i, name in enumerate(motor_names):
-            homing_offset, drive_mode = self.calibration[name]
-            values[i] -= homing_offset
-
-            # Update direction of rotation of the motor that was matching between leader and follower to their original direction.
-            # In fact, the motor of the leader for a given joint can be assembled in an opposite direction in term of rotation
-            # than the motor of the follower on the same joint.
-            if drive_mode:
-                values[i] *= -1
-
         return values
 
-    def _read_with_motor_ids(self, motor_models, motor_ids, data_name):
+    def read_with_motor_ids(self, motor_models, motor_ids, data_name, num_retry=NUM_READ_RETRY):
+        if self.mock:
+            import tests.mock_dynamixel_sdk as dxl
+        else:
+            import dynamixel_sdk as dxl
+
         return_list = True
         if not isinstance(motor_ids, list):
             return_list = False
@@ -608,12 +651,16 @@ class DynamixelMotorsBus:
 
         assert_same_address(self.model_ctrl_table, self.motor_models, data_name)
         addr, bytes = self.model_ctrl_table[motor_models[0]][data_name]
-        group = GroupSyncRead(self.port_handler, self.packet_handler, addr, bytes)
+        group = dxl.GroupSyncRead(self.port_handler, self.packet_handler, addr, bytes)
         for idx in motor_ids:
             group.addParam(idx)
 
-        comm = group.txRxPacket()
-        if comm != COMM_SUCCESS:
+        for _ in range(num_retry):
+            comm = group.txRxPacket()
+            if comm == dxl.COMM_SUCCESS:
+                break
+
+        if comm != dxl.COMM_SUCCESS:
             raise ConnectionError(
                 f"Read failed due to communication error on port {self.port_handler.port_name} for indices {motor_ids}: "
                 f"{self.packet_handler.getTxRxResult(comm)}"
@@ -637,6 +684,11 @@ class DynamixelMotorsBus:
 
         start_time = time.perf_counter()
 
+        if self.mock:
+            import tests.mock_dynamixel_sdk as dxl
+        else:
+            import dynamixel_sdk as dxl
+
         if motor_names is None:
             motor_names = self.motor_names
 
@@ -656,16 +708,18 @@ class DynamixelMotorsBus:
 
         if data_name not in self.group_readers:
             # create new group reader
-            self.group_readers[group_key] = GroupSyncRead(self.port_handler, self.packet_handler, addr, bytes)
+            self.group_readers[group_key] = dxl.GroupSyncRead(
+                self.port_handler, self.packet_handler, addr, bytes
+            )
             for idx in motor_ids:
                 self.group_readers[group_key].addParam(idx)
 
         for _ in range(NUM_READ_RETRY):
             comm = self.group_readers[group_key].txRxPacket()
-            if comm == COMM_SUCCESS:
+            if comm == dxl.COMM_SUCCESS:
                 break
 
-        if comm != COMM_SUCCESS:
+        if comm != dxl.COMM_SUCCESS:
             raise ConnectionError(
                 f"Read failed due to communication error on port {self.port} for group_key {group_key}: "
                 f"{self.packet_handler.getTxRxResult(comm)}"
@@ -683,19 +737,7 @@ class DynamixelMotorsBus:
             values = values.astype(np.int32)
 
         if data_name in CALIBRATION_REQUIRED and self.calibration is not None:
-            values = self.apply_calibration(values, motor_names)
-
-            # We expect our motors to stay in a nominal range of [-180, 180] degrees
-            # which corresponds to a half turn rotation.
-            # However, some motors can turn a bit more, hence we extend the nominal range to [-270, 270]
-            # which is less than a full 360 degree rotation.
-            if not np.all((values > -270) & (values < 270)):
-                raise ValueError(
-                    f"Wrong motor position range detected. "
-                    f"Expected to be in [-270, +270] but in [{values.min()}, {values.max()}]. "
-                    "This might be due to a cable connection issue creating an artificial 360 degrees jump in motor values. "
-                    "You need to recalibrate by running: `python lerobot/scripts/control_robot.py calibrate`"
-                )
+            values = self.apply_calibration_autocorrect(values, motor_names)
 
         # log the number of seconds it took to read the data from the motors
         delta_ts_name = get_log_name("delta_timestamp_s", "read", data_name, motor_names)
@@ -707,7 +749,12 @@ class DynamixelMotorsBus:
 
         return values
 
-    def _write_with_motor_ids(self, motor_models, motor_ids, data_name, values):
+    def write_with_motor_ids(self, motor_models, motor_ids, data_name, values, num_retry=NUM_WRITE_RETRY):
+        if self.mock:
+            import tests.mock_dynamixel_sdk as dxl
+        else:
+            import dynamixel_sdk as dxl
+
         if not isinstance(motor_ids, list):
             motor_ids = [motor_ids]
         if not isinstance(values, list):
@@ -715,13 +762,17 @@ class DynamixelMotorsBus:
 
         assert_same_address(self.model_ctrl_table, motor_models, data_name)
         addr, bytes = self.model_ctrl_table[motor_models[0]][data_name]
-        group = GroupSyncWrite(self.port_handler, self.packet_handler, addr, bytes)
+        group = dxl.GroupSyncWrite(self.port_handler, self.packet_handler, addr, bytes)
         for idx, value in zip(motor_ids, values, strict=True):
-            data = convert_to_bytes(value, bytes)
+            data = convert_to_bytes(value, bytes, self.mock)
             group.addParam(idx, data)
 
-        comm = group.txPacket()
-        if comm != COMM_SUCCESS:
+        for _ in range(num_retry):
+            comm = group.txPacket()
+            if comm == dxl.COMM_SUCCESS:
+                break
+
+        if comm != dxl.COMM_SUCCESS:
             raise ConnectionError(
                 f"Write failed due to communication error on port {self.port_handler.port_name} for indices {motor_ids}: "
                 f"{self.packet_handler.getTxRxResult(comm)}"
@@ -735,6 +786,11 @@ class DynamixelMotorsBus:
 
         start_time = time.perf_counter()
 
+        if self.mock:
+            import tests.mock_dynamixel_sdk as dxl
+        else:
+            import dynamixel_sdk as dxl
+
         if motor_names is None:
             motor_names = self.motor_names
 
@@ -764,19 +820,19 @@ class DynamixelMotorsBus:
 
         init_group = data_name not in self.group_readers
         if init_group:
-            self.group_writers[group_key] = GroupSyncWrite(
+            self.group_writers[group_key] = dxl.GroupSyncWrite(
                 self.port_handler, self.packet_handler, addr, bytes
             )
 
         for idx, value in zip(motor_ids, values, strict=True):
-            data = convert_to_bytes(value, bytes)
+            data = convert_to_bytes(value, bytes, self.mock)
             if init_group:
                 self.group_writers[group_key].addParam(idx, data)
             else:
                 self.group_writers[group_key].changeParam(idx, data)
 
         comm = self.group_writers[group_key].txPacket()
-        if comm != COMM_SUCCESS:
+        if comm != dxl.COMM_SUCCESS:
             raise ConnectionError(
                 f"Write failed due to communication error on port {self.port} for group_key {group_key}: "
                 f"{self.packet_handler.getTxRxResult(comm)}"
@@ -809,8 +865,3 @@ class DynamixelMotorsBus:
     def __del__(self):
         if getattr(self, "is_connected", False):
             self.disconnect()
-
-
-if __name__ == "__main__":
-    # Helper to find the usb port associated to all your DynamixelMotorsBus.
-    find_port()

lerobot/common/robot_devices/motors/feetech.py

@@ -0,0 +1,897 @@
+import enum
+import logging
+import math
+import time
+import traceback
+from copy import deepcopy
+
+import numpy as np
+import tqdm
+
+from lerobot.common.robot_devices.utils import RobotDeviceAlreadyConnectedError, RobotDeviceNotConnectedError
+from lerobot.common.utils.utils import capture_timestamp_utc
+
+PROTOCOL_VERSION = 0
+BAUDRATE = 1_000_000
+TIMEOUT_MS = 1000
+
+MAX_ID_RANGE = 252
+
+# The following bounds define the lower and upper joints range (after calibration).
+# For joints in degree (i.e. revolute joints), their nominal range is [-180, 180] degrees
+# which corresponds to a half rotation on the left and half rotation on the right.
+# Some joints might require higher range, so we allow up to [-270, 270] degrees until
+# an error is raised.
+LOWER_BOUND_DEGREE = -270
+UPPER_BOUND_DEGREE = 270
+# For joints in percentage (i.e. joints that move linearly like the prismatic joint of a gripper),
+# their nominal range is [0, 100] %. For instance, for Aloha gripper, 0% is fully
+# closed, and 100% is fully open. To account for slight calibration issue, we allow up to
+# [-10, 110] until an error is raised.
+LOWER_BOUND_LINEAR = -10
+UPPER_BOUND_LINEAR = 110
+
+HALF_TURN_DEGREE = 180
+
+
+# See this link for STS3215 Memory Table:
+# https://docs.google.com/spreadsheets/d/1GVs7W1VS1PqdhA1nW-abeyAHhTUxKUdR/edit?usp=sharing&ouid=116566590112741600240&rtpof=true&sd=true
+# data_name: (address, size_byte)
+SCS_SERIES_CONTROL_TABLE = {
+    "Model": (3, 2),
+    "ID": (5, 1),
+    "Baud_Rate": (6, 1),
+    "Return_Delay": (7, 1),
+    "Response_Status_Level": (8, 1),
+    "Min_Angle_Limit": (9, 2),
+    "Max_Angle_Limit": (11, 2),
+    "Max_Temperature_Limit": (13, 1),
+    "Max_Voltage_Limit": (14, 1),
+    "Min_Voltage_Limit": (15, 1),
+    "Max_Torque_Limit": (16, 2),
+    "Phase": (18, 1),
+    "Unloading_Condition": (19, 1),
+    "LED_Alarm_Condition": (20, 1),
+    "P_Coefficient": (21, 1),
+    "D_Coefficient": (22, 1),
+    "I_Coefficient": (23, 1),
+    "Minimum_Startup_Force": (24, 2),
+    "CW_Dead_Zone": (26, 1),
+    "CCW_Dead_Zone": (27, 1),
+    "Protection_Current": (28, 2),
+    "Angular_Resolution": (30, 1),
+    "Offset": (31, 2),
+    "Mode": (33, 1),
+    "Protective_Torque": (34, 1),
+    "Protection_Time": (35, 1),
+    "Overload_Torque": (36, 1),
+    "Speed_closed_loop_P_proportional_coefficient": (37, 1),
+    "Over_Current_Protection_Time": (38, 1),
+    "Velocity_closed_loop_I_integral_coefficient": (39, 1),
+    "Torque_Enable": (40, 1),
+    "Acceleration": (41, 1),
+    "Goal_Position": (42, 2),
+    "Goal_Time": (44, 2),
+    "Goal_Speed": (46, 2),
+    "Torque_Limit": (48, 2),
+    "Lock": (55, 1),
+    "Present_Position": (56, 2),
+    "Present_Speed": (58, 2),
+    "Present_Load": (60, 2),
+    "Present_Voltage": (62, 1),
+    "Present_Temperature": (63, 1),
+    "Status": (65, 1),
+    "Moving": (66, 1),
+    "Present_Current": (69, 2),
+    # Not in the Memory Table
+    "Maximum_Acceleration": (85, 2),
+}
+
+SCS_SERIES_BAUDRATE_TABLE = {
+    0: 1_000_000,
+    1: 500_000,
+    2: 250_000,
+    3: 128_000,
+    4: 115_200,
+    5: 57_600,
+    6: 38_400,
+    7: 19_200,
+}
+
+CALIBRATION_REQUIRED = ["Goal_Position", "Present_Position"]
+CONVERT_UINT32_TO_INT32_REQUIRED = ["Goal_Position", "Present_Position"]
+
+
+MODEL_CONTROL_TABLE = {
+    "scs_series": SCS_SERIES_CONTROL_TABLE,
+    "sts3215": SCS_SERIES_CONTROL_TABLE,
+}
+
+MODEL_RESOLUTION = {
+    "scs_series": 4096,
+    "sts3215": 4096,
+}
+
+MODEL_BAUDRATE_TABLE = {
+    "scs_series": SCS_SERIES_BAUDRATE_TABLE,
+    "sts3215": SCS_SERIES_BAUDRATE_TABLE,
+}
+
+# High number of retries is needed for feetech compared to dynamixel motors.
+NUM_READ_RETRY = 20
+NUM_WRITE_RETRY = 20
+
+
+def convert_degrees_to_steps(degrees: float | np.ndarray, models: str | list[str]) -> np.ndarray:
+    """This function converts the degree range to the step range for indicating motors rotation.
+    It assumes a motor achieves a full rotation by going from -180 degree position to +180.
+    The motor resolution (e.g. 4096) corresponds to the number of steps needed to achieve a full rotation.
+    """
+    resolutions = [MODEL_RESOLUTION[model] for model in models]
+    steps = degrees / 180 * np.array(resolutions) / 2
+    steps = steps.astype(int)
+    return steps
+
+
+def convert_to_bytes(value, bytes, mock=False):
+    if mock:
+        return value
+
+    import scservo_sdk as scs
+
+    # Note: No need to convert back into unsigned int, since this byte preprocessing
+    # already handles it for us.
+    if bytes == 1:
+        data = [
+            scs.SCS_LOBYTE(scs.SCS_LOWORD(value)),
+        ]
+    elif bytes == 2:
+        data = [
+            scs.SCS_LOBYTE(scs.SCS_LOWORD(value)),
+            scs.SCS_HIBYTE(scs.SCS_LOWORD(value)),
+        ]
+    elif bytes == 4:
+        data = [
+            scs.SCS_LOBYTE(scs.SCS_LOWORD(value)),
+            scs.SCS_HIBYTE(scs.SCS_LOWORD(value)),
+            scs.SCS_LOBYTE(scs.SCS_HIWORD(value)),
+            scs.SCS_HIBYTE(scs.SCS_HIWORD(value)),
+        ]
+    else:
+        raise NotImplementedError(
+            f"Value of the number of bytes to be sent is expected to be in [1, 2, 4], but "
+            f"{bytes} is provided instead."
+        )
+    return data
+
+
+def get_group_sync_key(data_name, motor_names):
+    group_key = f"{data_name}_" + "_".join(motor_names)
+    return group_key
+
+
+def get_result_name(fn_name, data_name, motor_names):
+    group_key = get_group_sync_key(data_name, motor_names)
+    rslt_name = f"{fn_name}_{group_key}"
+    return rslt_name
+
+
+def get_queue_name(fn_name, data_name, motor_names):
+    group_key = get_group_sync_key(data_name, motor_names)
+    queue_name = f"{fn_name}_{group_key}"
+    return queue_name
+
+
+def get_log_name(var_name, fn_name, data_name, motor_names):
+    group_key = get_group_sync_key(data_name, motor_names)
+    log_name = f"{var_name}_{fn_name}_{group_key}"
+    return log_name
+
+
+def assert_same_address(model_ctrl_table, motor_models, data_name):
+    all_addr = []
+    all_bytes = []
+    for model in motor_models:
+        addr, bytes = model_ctrl_table[model][data_name]
+        all_addr.append(addr)
+        all_bytes.append(bytes)
+
+    if len(set(all_addr)) != 1:
+        raise NotImplementedError(
+            f"At least two motor models use a different address for `data_name`='{data_name}' ({list(zip(motor_models, all_addr, strict=False))}). Contact a LeRobot maintainer."
+        )
+
+    if len(set(all_bytes)) != 1:
+        raise NotImplementedError(
+            f"At least two motor models use a different bytes representation for `data_name`='{data_name}' ({list(zip(motor_models, all_bytes, strict=False))}). Contact a LeRobot maintainer."
+        )
+
+
+class TorqueMode(enum.Enum):
+    ENABLED = 1
+    DISABLED = 0
+
+
+class DriveMode(enum.Enum):
+    NON_INVERTED = 0
+    INVERTED = 1
+
+
+class CalibrationMode(enum.Enum):
+    # Joints with rotational motions are expressed in degrees in nominal range of [-180, 180]
+    DEGREE = 0
+    # Joints with linear motions (like gripper of Aloha) are experessed in nominal range of [0, 100]
+    LINEAR = 1
+
+
+class JointOutOfRangeError(Exception):
+    def __init__(self, message="Joint is out of range"):
+        self.message = message
+        super().__init__(self.message)
+
+
+class FeetechMotorsBus:
+    """
+    The FeetechMotorsBus class allows to efficiently read and write to the attached motors. It relies on
+    the python feetech sdk to communicate with the motors. For more info, see the [feetech SDK Documentation](https://emanual.robotis.com/docs/en/software/feetech/feetech_sdk/sample_code/python_read_write_protocol_2_0/#python-read-write-protocol-20).
+
+    A FeetechMotorsBus instance requires a port (e.g. `FeetechMotorsBus(port="/dev/tty.usbmodem575E0031751"`)).
+    To find the port, you can run our utility script:
+    ```bash
+    python lerobot/scripts/find_motors_bus_port.py
+    >>> Finding all available ports for the MotorsBus.
+    >>> ['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
+    >>> Remove the usb cable from your FeetechMotorsBus and press Enter when done.
+    >>> The port of this FeetechMotorsBus is /dev/tty.usbmodem575E0031751.
+    >>> Reconnect the usb cable.
+    ```
+
+    Example of usage for 1 motor connected to the bus:
+    ```python
+    motor_name = "gripper"
+    motor_index = 6
+    motor_model = "sts3215"
+
+    motors_bus = FeetechMotorsBus(
+        port="/dev/tty.usbmodem575E0031751",
+        motors={motor_name: (motor_index, motor_model)},
+    )
+    motors_bus.connect()
+
+    position = motors_bus.read("Present_Position")
+
+    # move from a few motor steps as an example
+    few_steps = 30
+    motors_bus.write("Goal_Position", position + few_steps)
+
+    # when done, consider disconnecting
+    motors_bus.disconnect()
+    ```
+    """
+
+    def __init__(
+        self,
+        port: str,
+        motors: dict[str, tuple[int, str]],
+        extra_model_control_table: dict[str, list[tuple]] | None = None,
+        extra_model_resolution: dict[str, int] | None = None,
+        mock=False,
+    ):
+        self.port = port
+        self.motors = motors
+        self.mock = mock
+
+        self.model_ctrl_table = deepcopy(MODEL_CONTROL_TABLE)
+        if extra_model_control_table:
+            self.model_ctrl_table.update(extra_model_control_table)
+
+        self.model_resolution = deepcopy(MODEL_RESOLUTION)
+        if extra_model_resolution:
+            self.model_resolution.update(extra_model_resolution)
+
+        self.port_handler = None
+        self.packet_handler = None
+        self.calibration = None
+        self.is_connected = False
+        self.group_readers = {}
+        self.group_writers = {}
+        self.logs = {}
+
+        self.track_positions = {}
+        self.present_pos = {
+            "prev": [None] * len(self.motor_names),
+            "below_zero": [None] * len(self.motor_names),
+            "above_max": [None] * len(self.motor_names),
+        }
+        self.goal_pos = {
+            "prev": [None] * len(self.motor_names),
+            "below_zero": [None] * len(self.motor_names),
+            "above_max": [None] * len(self.motor_names),
+        }
+
+    def connect(self):
+        if self.is_connected:
+            raise RobotDeviceAlreadyConnectedError(
+                f"FeetechMotorsBus({self.port}) is already connected. Do not call `motors_bus.connect()` twice."
+            )
+
+        if self.mock:
+            import tests.mock_scservo_sdk as scs
+        else:
+            import scservo_sdk as scs
+
+        self.port_handler = scs.PortHandler(self.port)
+        self.packet_handler = scs.PacketHandler(PROTOCOL_VERSION)
+
+        try:
+            if not self.port_handler.openPort():
+                raise OSError(f"Failed to open port '{self.port}'.")
+        except Exception:
+            traceback.print_exc()
+            print(
+                "\nTry running `python lerobot/scripts/find_motors_bus_port.py` to make sure you are using the correct port.\n"
+            )
+            raise
+
+        # Allow to read and write
+        self.is_connected = True
+
+        self.port_handler.setPacketTimeoutMillis(TIMEOUT_MS)
+
+    def reconnect(self):
+        if self.mock:
+            import tests.mock_scservo_sdk as scs
+        else:
+            import scservo_sdk as scs
+
+        self.port_handler = scs.PortHandler(self.port)
+        self.packet_handler = scs.PacketHandler(PROTOCOL_VERSION)
+
+        if not self.port_handler.openPort():
+            raise OSError(f"Failed to open port '{self.port}'.")
+
+        self.is_connected = True
+
+    def are_motors_configured(self):
+        # Only check the motor indices and not baudrate, since if the motor baudrates are incorrect,
+        # a ConnectionError will be raised anyway.
+        try:
+            return (self.motor_indices == self.read("ID")).all()
+        except ConnectionError as e:
+            print(e)
+            return False
+
+    def find_motor_indices(self, possible_ids=None, num_retry=2):
+        if possible_ids is None:
+            possible_ids = range(MAX_ID_RANGE)
+
+        indices = []
+        for idx in tqdm.tqdm(possible_ids):
+            try:
+                present_idx = self.read_with_motor_ids(self.motor_models, [idx], "ID", num_retry=num_retry)[0]
+            except ConnectionError:
+                continue
+
+            if idx != present_idx:
+                # sanity check
+                raise OSError(
+                    "Motor index used to communicate through the bus is not the same as the one present in the motor memory. The motor memory might be damaged."
+                )
+            indices.append(idx)
+
+        return indices
+
+    def set_bus_baudrate(self, baudrate):
+        present_bus_baudrate = self.port_handler.getBaudRate()
+        if present_bus_baudrate != baudrate:
+            print(f"Setting bus baud rate to {baudrate}. Previously {present_bus_baudrate}.")
+            self.port_handler.setBaudRate(baudrate)
+
+            if self.port_handler.getBaudRate() != baudrate:
+                raise OSError("Failed to write bus baud rate.")
+
+    @property
+    def motor_names(self) -> list[str]:
+        return list(self.motors.keys())
+
+    @property
+    def motor_models(self) -> list[str]:
+        return [model for _, model in self.motors.values()]
+
+    @property
+    def motor_indices(self) -> list[int]:
+        return [idx for idx, _ in self.motors.values()]
+
+    def set_calibration(self, calibration: dict[str, list]):
+        self.calibration = calibration
+
+    def apply_calibration_autocorrect(self, values: np.ndarray | list, motor_names: list[str] | None):
+        """This function apply the calibration, automatically detects out of range errors for motors values and attempt to correct.
+
+        For more info, see docstring of `apply_calibration` and `autocorrect_calibration`.
+        """
+        try:
+            values = self.apply_calibration(values, motor_names)
+        except JointOutOfRangeError as e:
+            print(e)
+            self.autocorrect_calibration(values, motor_names)
+            values = self.apply_calibration(values, motor_names)
+        return values
+
+    def apply_calibration(self, values: np.ndarray | list, motor_names: list[str] | None):
+        """Convert from unsigned int32 joint position range [0, 2**32[ to the universal float32 nominal degree range ]-180.0, 180.0[ with
+        a "zero position" at 0 degree.
+
+        Note: We say "nominal degree range" since the motors can take values outside this range. For instance, 190 degrees, if the motor
+        rotate more than a half a turn from the zero position. However, most motors can't rotate more than 180 degrees and will stay in this range.
+
+        Joints values are original in [0, 2**32[ (unsigned int32). Each motor are expected to complete a full rotation
+        when given a goal position that is + or - their resolution. For instance, feetech xl330-m077 have a resolution of 4096, and
+        at any position in their original range, let's say the position 56734, they complete a full rotation clockwise by moving to 60830,
+        or anticlockwise by moving to 52638. The position in the original range is arbitrary and might change a lot between each motor.
+        To harmonize between motors of the same model, different robots, or even models of different brands, we propose to work
+        in the centered nominal degree range ]-180, 180[.
+        """
+        if motor_names is None:
+            motor_names = self.motor_names
+
+        # Convert from unsigned int32 original range [0, 2**32] to signed float32 range
+        values = values.astype(np.float32)
+
+        for i, name in enumerate(motor_names):
+            calib_idx = self.calibration["motor_names"].index(name)
+            calib_mode = self.calibration["calib_mode"][calib_idx]
+
+            if CalibrationMode[calib_mode] == CalibrationMode.DEGREE:
+                drive_mode = self.calibration["drive_mode"][calib_idx]
+                homing_offset = self.calibration["homing_offset"][calib_idx]
+                _, model = self.motors[name]
+                resolution = self.model_resolution[model]
+
+                # Update direction of rotation of the motor to match between leader and follower.
+                # In fact, the motor of the leader for a given joint can be assembled in an
+                # opposite direction in term of rotation than the motor of the follower on the same joint.
+                if drive_mode:
+                    values[i] *= -1
+
+                # Convert from range [-2**31, 2**31[ to
+                # nominal range ]-resolution, resolution[ (e.g. ]-2048, 2048[)
+                values[i] += homing_offset
+
+                # Convert from range ]-resolution, resolution[ to
+                # universal float32 centered degree range ]-180, 180[
+                values[i] = values[i] / (resolution // 2) * HALF_TURN_DEGREE
+
+                if (values[i] < LOWER_BOUND_DEGREE) or (values[i] > UPPER_BOUND_DEGREE):
+                    raise JointOutOfRangeError(
+                        f"Wrong motor position range detected for {name}. "
+                        f"Expected to be in nominal range of [-{HALF_TURN_DEGREE}, {HALF_TURN_DEGREE}] degrees (a full rotation), "
+                        f"with a maximum range of [{LOWER_BOUND_DEGREE}, {UPPER_BOUND_DEGREE}] degrees to account for joints that can rotate a bit more, "
+                        f"but present value is {values[i]} degree. "
+                        "This might be due to a cable connection issue creating an artificial 360 degrees jump in motor values. "
+                        "You need to recalibrate by running: `python lerobot/scripts/control_robot.py calibrate`"
+                    )
+
+            elif CalibrationMode[calib_mode] == CalibrationMode.LINEAR:
+                start_pos = self.calibration["start_pos"][calib_idx]
+                end_pos = self.calibration["end_pos"][calib_idx]
+
+                # Rescale the present position to a nominal range [0, 100] %,
+                # useful for joints with linear motions like Aloha gripper
+                values[i] = (values[i] - start_pos) / (end_pos - start_pos) * 100
+
+                if (values[i] < LOWER_BOUND_LINEAR) or (values[i] > UPPER_BOUND_LINEAR):
+                    raise JointOutOfRangeError(
+                        f"Wrong motor position range detected for {name}. "
+                        f"Expected to be in nominal range of [0, 100] % (a full linear translation), "
+                        f"with a maximum range of [{LOWER_BOUND_LINEAR}, {UPPER_BOUND_LINEAR}] % to account for some imprecision during calibration, "
+                        f"but present value is {values[i]} %. "
+                        "This might be due to a cable connection issue creating an artificial jump in motor values. "
+                        "You need to recalibrate by running: `python lerobot/scripts/control_robot.py calibrate`"
+                    )
+
+        return values
+
+    def autocorrect_calibration(self, values: np.ndarray | list, motor_names: list[str] | None):
+        """This function automatically detects issues with values of motors after calibration, and correct for these issues.
+
+        Some motors might have values outside of expected maximum bounds after calibration.
+        For instance, for a joint in degree, its value can be outside [-270, 270] degrees, which is totally unexpected given
+        a nominal range of [-180, 180] degrees, which represents half a turn to the left or right starting from zero position.
+
+        Known issues:
+        #1: Motor value randomly shifts of a full turn, caused by hardware/connection errors.
+        #2: Motor internal homing offset is shifted of a full turn, caused by using default calibration (e.g Aloha).
+        #3: motor internal homing offset is shifted of less or more than a full turn, caused by using default calibration
+            or by human error during manual calibration.
+
+        Issues #1 and #2 can be solved by shifting the calibration homing offset by a full turn.
+        Issue #3 will be visually detected by user and potentially captured by the safety feature `max_relative_target`,
+        that will slow down the motor, raise an error asking to recalibrate. Manual recalibrating will solve the issue.
+
+        Note: A full turn corresponds to 360 degrees but also to 4096 steps for a motor resolution of 4096.
+        """
+        if motor_names is None:
+            motor_names = self.motor_names
+
+        # Convert from unsigned int32 original range [0, 2**32] to signed float32 range
+        values = values.astype(np.float32)
+
+        for i, name in enumerate(motor_names):
+            calib_idx = self.calibration["motor_names"].index(name)
+            calib_mode = self.calibration["calib_mode"][calib_idx]
+
+            if CalibrationMode[calib_mode] == CalibrationMode.DEGREE:
+                drive_mode = self.calibration["drive_mode"][calib_idx]
+                homing_offset = self.calibration["homing_offset"][calib_idx]
+                _, model = self.motors[name]
+                resolution = self.model_resolution[model]
+
+                if drive_mode:
+                    values[i] *= -1
+
+                # Convert from initial range to range [-180, 180] degrees
+                calib_val = (values[i] + homing_offset) / (resolution // 2) * HALF_TURN_DEGREE
+                in_range = (calib_val > LOWER_BOUND_DEGREE) and (calib_val < UPPER_BOUND_DEGREE)
+
+                # Solve this inequality to find the factor to shift the range into [-180, 180] degrees
+                # values[i] = (values[i] + homing_offset + resolution * factor) / (resolution // 2) * HALF_TURN_DEGREE
+                # - HALF_TURN_DEGREE <= (values[i] + homing_offset + resolution * factor) / (resolution // 2) * HALF_TURN_DEGREE <= HALF_TURN_DEGREE
+                # (- HALF_TURN_DEGREE / HALF_TURN_DEGREE * (resolution // 2) - values[i] - homing_offset) / resolution <= factor <= (HALF_TURN_DEGREE / 180 * (resolution // 2) - values[i] - homing_offset) / resolution
+                low_factor = (
+                    -HALF_TURN_DEGREE / HALF_TURN_DEGREE * (resolution // 2) - values[i] - homing_offset
+                ) / resolution
+                upp_factor = (
+                    HALF_TURN_DEGREE / HALF_TURN_DEGREE * (resolution // 2) - values[i] - homing_offset
+                ) / resolution
+
+            elif CalibrationMode[calib_mode] == CalibrationMode.LINEAR:
+                start_pos = self.calibration["start_pos"][calib_idx]
+                end_pos = self.calibration["end_pos"][calib_idx]
+
+                # Convert from initial range to range [0, 100] in %
+                calib_val = (values[i] - start_pos) / (end_pos - start_pos) * 100
+                in_range = (calib_val > LOWER_BOUND_LINEAR) and (calib_val < UPPER_BOUND_LINEAR)
+
+                # Solve this inequality to find the factor to shift the range into [0, 100] %
+                # values[i] = (values[i] - start_pos + resolution * factor) / (end_pos + resolution * factor - start_pos - resolution * factor) * 100
+                # values[i] = (values[i] - start_pos + resolution * factor) / (end_pos - start_pos) * 100
+                # 0 <= (values[i] - start_pos + resolution * factor) / (end_pos - start_pos) * 100 <= 100
+                # (start_pos - values[i]) / resolution <= factor <= (end_pos - values[i]) / resolution
+                low_factor = (start_pos - values[i]) / resolution
+                upp_factor = (end_pos - values[i]) / resolution
+
+            if not in_range:
+                # Get first integer between the two bounds
+                if low_factor < upp_factor:
+                    factor = math.ceil(low_factor)
+
+                    if factor > upp_factor:
+                        raise ValueError(f"No integer found between bounds [{low_factor=}, {upp_factor=}]")
+                else:
+                    factor = math.ceil(upp_factor)
+
+                    if factor > low_factor:
+                        raise ValueError(f"No integer found between bounds [{low_factor=}, {upp_factor=}]")
+
+                if CalibrationMode[calib_mode] == CalibrationMode.DEGREE:
+                    out_of_range_str = f"{LOWER_BOUND_DEGREE} < {calib_val} < {UPPER_BOUND_DEGREE} degrees"
+                    in_range_str = f"{LOWER_BOUND_DEGREE} < {calib_val} < {UPPER_BOUND_DEGREE} degrees"
+                elif CalibrationMode[calib_mode] == CalibrationMode.LINEAR:
+                    out_of_range_str = f"{LOWER_BOUND_LINEAR} < {calib_val} < {UPPER_BOUND_LINEAR} %"
+                    in_range_str = f"{LOWER_BOUND_LINEAR} < {calib_val} < {UPPER_BOUND_LINEAR} %"
+
+                logging.warning(
+                    f"Auto-correct calibration of motor '{name}' by shifting value by {abs(factor)} full turns, "
+                    f"from '{out_of_range_str}' to '{in_range_str}'."
+                )
+
+                # A full turn corresponds to 360 degrees but also to 4096 steps for a motor resolution of 4096.
+                self.calibration["homing_offset"][calib_idx] += resolution * factor
+
+    def revert_calibration(self, values: np.ndarray | list, motor_names: list[str] | None):
+        """Inverse of `apply_calibration`."""
+        if motor_names is None:
+            motor_names = self.motor_names
+
+        for i, name in enumerate(motor_names):
+            calib_idx = self.calibration["motor_names"].index(name)
+            calib_mode = self.calibration["calib_mode"][calib_idx]
+
+            if CalibrationMode[calib_mode] == CalibrationMode.DEGREE:
+                drive_mode = self.calibration["drive_mode"][calib_idx]
+                homing_offset = self.calibration["homing_offset"][calib_idx]
+                _, model = self.motors[name]
+                resolution = self.model_resolution[model]
+
+                # Convert from nominal 0-centered degree range [-180, 180] to
+                # 0-centered resolution range (e.g. [-2048, 2048] for resolution=4096)
+                values[i] = values[i] / HALF_TURN_DEGREE * (resolution // 2)
+
+                # Substract the homing offsets to come back to actual motor range of values
+                # which can be arbitrary.
+                values[i] -= homing_offset
+
+                # Remove drive mode, which is the rotation direction of the motor, to come back to
+                # actual motor rotation direction which can be arbitrary.
+                if drive_mode:
+                    values[i] *= -1
+
+            elif CalibrationMode[calib_mode] == CalibrationMode.LINEAR:
+                start_pos = self.calibration["start_pos"][calib_idx]
+                end_pos = self.calibration["end_pos"][calib_idx]
+
+                # Convert from nominal lnear range of [0, 100] % to
+                # actual motor range of values which can be arbitrary.
+                values[i] = values[i] / 100 * (end_pos - start_pos) + start_pos
+
+        values = np.round(values).astype(np.int32)
+        return values
+
+    def avoid_rotation_reset(self, values, motor_names, data_name):
+        if data_name not in self.track_positions:
+            self.track_positions[data_name] = {
+                "prev": [None] * len(self.motor_names),
+                # Assume False at initialization
+                "below_zero": [False] * len(self.motor_names),
+                "above_max": [False] * len(self.motor_names),
+            }
+
+        track = self.track_positions[data_name]
+
+        if motor_names is None:
+            motor_names = self.motor_names
+
+        for i, name in enumerate(motor_names):
+            idx = self.motor_names.index(name)
+
+            if track["prev"][idx] is None:
+                track["prev"][idx] = values[i]
+                continue
+
+            # Detect a full rotation occured
+            if abs(track["prev"][idx] - values[i]) > 2048:
+                # Position went below 0 and got reset to 4095
+                if track["prev"][idx] < values[i]:
+                    # So we set negative value by adding a full rotation
+                    values[i] -= 4096
+
+                # Position went above 4095 and got reset to 0
+                elif track["prev"][idx] > values[i]:
+                    # So we add a full rotation
+                    values[i] += 4096
+
+            track["prev"][idx] = values[i]
+
+        return values
+
+    def read_with_motor_ids(self, motor_models, motor_ids, data_name, num_retry=NUM_READ_RETRY):
+        if self.mock:
+            import tests.mock_scservo_sdk as scs
+        else:
+            import scservo_sdk as scs
+
+        return_list = True
+        if not isinstance(motor_ids, list):
+            return_list = False
+            motor_ids = [motor_ids]
+
+        assert_same_address(self.model_ctrl_table, self.motor_models, data_name)
+        addr, bytes = self.model_ctrl_table[motor_models[0]][data_name]
+        group = scs.GroupSyncRead(self.port_handler, self.packet_handler, addr, bytes)
+        for idx in motor_ids:
+            group.addParam(idx)
+
+        for _ in range(num_retry):
+            comm = group.txRxPacket()
+            if comm == scs.COMM_SUCCESS:
+                break
+
+        if comm != scs.COMM_SUCCESS:
+            raise ConnectionError(
+                f"Read failed due to communication error on port {self.port_handler.port_name} for indices {motor_ids}: "
+                f"{self.packet_handler.getTxRxResult(comm)}"
+            )
+
+        values = []
+        for idx in motor_ids:
+            value = group.getData(idx, addr, bytes)
+            values.append(value)
+
+        if return_list:
+            return values
+        else:
+            return values[0]
+
+    def read(self, data_name, motor_names: str | list[str] | None = None):
+        if self.mock:
+            import tests.mock_scservo_sdk as scs
+        else:
+            import scservo_sdk as scs
+
+        if not self.is_connected:
+            raise RobotDeviceNotConnectedError(
+                f"FeetechMotorsBus({self.port}) is not connected. You need to run `motors_bus.connect()`."
+            )
+
+        start_time = time.perf_counter()
+
+        if motor_names is None:
+            motor_names = self.motor_names
+
+        if isinstance(motor_names, str):
+            motor_names = [motor_names]
+
+        motor_ids = []
+        models = []
+        for name in motor_names:
+            motor_idx, model = self.motors[name]
+            motor_ids.append(motor_idx)
+            models.append(model)
+
+        assert_same_address(self.model_ctrl_table, models, data_name)
+        addr, bytes = self.model_ctrl_table[model][data_name]
+        group_key = get_group_sync_key(data_name, motor_names)
+
+        if data_name not in self.group_readers:
+            # create new group reader
+            self.group_readers[group_key] = scs.GroupSyncRead(
+                self.port_handler, self.packet_handler, addr, bytes
+            )
+            for idx in motor_ids:
+                self.group_readers[group_key].addParam(idx)
+
+        for _ in range(NUM_READ_RETRY):
+            comm = self.group_readers[group_key].txRxPacket()
+            if comm == scs.COMM_SUCCESS:
+                break
+
+        if comm != scs.COMM_SUCCESS:
+            raise ConnectionError(
+                f"Read failed due to communication error on port {self.port} for group_key {group_key}: "
+                f"{self.packet_handler.getTxRxResult(comm)}"
+            )
+
+        values = []
+        for idx in motor_ids:
+            value = self.group_readers[group_key].getData(idx, addr, bytes)
+            values.append(value)
+
+        values = np.array(values)
+
+        # Convert to signed int to use range [-2048, 2048] for our motor positions.
+        if data_name in CONVERT_UINT32_TO_INT32_REQUIRED:
+            values = values.astype(np.int32)
+
+        if data_name in CALIBRATION_REQUIRED:
+            values = self.avoid_rotation_reset(values, motor_names, data_name)
+
+        if data_name in CALIBRATION_REQUIRED and self.calibration is not None:
+            values = self.apply_calibration_autocorrect(values, motor_names)
+
+        # log the number of seconds it took to read the data from the motors
+        delta_ts_name = get_log_name("delta_timestamp_s", "read", data_name, motor_names)
+        self.logs[delta_ts_name] = time.perf_counter() - start_time
+
+        # log the utc time at which the data was received
+        ts_utc_name = get_log_name("timestamp_utc", "read", data_name, motor_names)
+        self.logs[ts_utc_name] = capture_timestamp_utc()
+
+        return values
+
+    def write_with_motor_ids(self, motor_models, motor_ids, data_name, values, num_retry=NUM_WRITE_RETRY):
+        if self.mock:
+            import tests.mock_scservo_sdk as scs
+        else:
+            import scservo_sdk as scs
+
+        if not isinstance(motor_ids, list):
+            motor_ids = [motor_ids]
+        if not isinstance(values, list):
+            values = [values]
+
+        assert_same_address(self.model_ctrl_table, motor_models, data_name)
+        addr, bytes = self.model_ctrl_table[motor_models[0]][data_name]
+        group = scs.GroupSyncWrite(self.port_handler, self.packet_handler, addr, bytes)
+        for idx, value in zip(motor_ids, values, strict=True):
+            data = convert_to_bytes(value, bytes, self.mock)
+            group.addParam(idx, data)
+
+        for _ in range(num_retry):
+            comm = group.txPacket()
+            if comm == scs.COMM_SUCCESS:
+                break
+
+        if comm != scs.COMM_SUCCESS:
+            raise ConnectionError(
+                f"Write failed due to communication error on port {self.port_handler.port_name} for indices {motor_ids}: "
+                f"{self.packet_handler.getTxRxResult(comm)}"
+            )
+
+    def write(self, data_name, values: int | float | np.ndarray, motor_names: str | list[str] | None = None):
+        if not self.is_connected:
+            raise RobotDeviceNotConnectedError(
+                f"FeetechMotorsBus({self.port}) is not connected. You need to run `motors_bus.connect()`."
+            )
+
+        start_time = time.perf_counter()
+
+        if self.mock:
+            import tests.mock_scservo_sdk as scs
+        else:
+            import scservo_sdk as scs
+
+        if motor_names is None:
+            motor_names = self.motor_names
+
+        if isinstance(motor_names, str):
+            motor_names = [motor_names]
+
+        if isinstance(values, (int, float, np.integer)):
+            values = [int(values)] * len(motor_names)
+
+        values = np.array(values)
+
+        motor_ids = []
+        models = []
+        for name in motor_names:
+            motor_idx, model = self.motors[name]
+            motor_ids.append(motor_idx)
+            models.append(model)
+
+        if data_name in CALIBRATION_REQUIRED and self.calibration is not None:
+            values = self.revert_calibration(values, motor_names)
+
+        values = values.tolist()
+
+        assert_same_address(self.model_ctrl_table, models, data_name)
+        addr, bytes = self.model_ctrl_table[model][data_name]
+        group_key = get_group_sync_key(data_name, motor_names)
+
+        init_group = data_name not in self.group_readers
+        if init_group:
+            self.group_writers[group_key] = scs.GroupSyncWrite(
+                self.port_handler, self.packet_handler, addr, bytes
+            )
+
+        for idx, value in zip(motor_ids, values, strict=True):
+            data = convert_to_bytes(value, bytes, self.mock)
+            if init_group:
+                self.group_writers[group_key].addParam(idx, data)
+            else:
+                self.group_writers[group_key].changeParam(idx, data)
+
+        comm = self.group_writers[group_key].txPacket()
+        if comm != scs.COMM_SUCCESS:
+            raise ConnectionError(
+                f"Write failed due to communication error on port {self.port} for group_key {group_key}: "
+                f"{self.packet_handler.getTxRxResult(comm)}"
+            )
+
+        # log the number of seconds it took to write the data to the motors
+        delta_ts_name = get_log_name("delta_timestamp_s", "write", data_name, motor_names)
+        self.logs[delta_ts_name] = time.perf_counter() - start_time
+
+        # TODO(rcadene): should we log the time before sending the write command?
+        # log the utc time when the write has been completed
+        ts_utc_name = get_log_name("timestamp_utc", "write", data_name, motor_names)
+        self.logs[ts_utc_name] = capture_timestamp_utc()
+
+    def disconnect(self):
+        if not self.is_connected:
+            raise RobotDeviceNotConnectedError(
+                f"FeetechMotorsBus({self.port}) is not connected. Try running `motors_bus.connect()` first."
+            )
+
+        if self.port_handler is not None:
+            self.port_handler.closePort()
+            self.port_handler = None
+
+        self.packet_handler = None
+        self.group_readers = {}
+        self.group_writers = {}
+        self.is_connected = False
+
+    def __del__(self):
+        if getattr(self, "is_connected", False):
+            self.disconnect()

lerobot/common/robot_devices/robots/dynamixel_calibration.py

@@ -0,0 +1,130 @@
+"""Logic to calibrate a robot arm built with dynamixel motors"""
+# TODO(rcadene, aliberts): move this logic into the robot code when refactoring
+
+import numpy as np
+
+from lerobot.common.robot_devices.motors.dynamixel import (
+    CalibrationMode,
+    TorqueMode,
+    convert_degrees_to_steps,
+)
+from lerobot.common.robot_devices.motors.utils import MotorsBus
+
+URL_TEMPLATE = (
+    "https://raw.githubusercontent.com/huggingface/lerobot/main/media/{robot}/{arm}_{position}.webp"
+)
+
+# The following positions are provided in nominal degree range ]-180, +180[
+# For more info on these constants, see comments in the code where they get used.
+ZERO_POSITION_DEGREE = 0
+ROTATED_POSITION_DEGREE = 90
+
+
+def assert_drive_mode(drive_mode):
+    # `drive_mode` is in [0,1] with 0 means original rotation direction for the motor, and 1 means inverted.
+    if not np.all(np.isin(drive_mode, [0, 1])):
+        raise ValueError(f"`drive_mode` contains values other than 0 or 1: ({drive_mode})")
+
+
+def apply_drive_mode(position, drive_mode):
+    assert_drive_mode(drive_mode)
+    # Convert `drive_mode` from [0, 1] with 0 indicates original rotation direction and 1 inverted,
+    # to [-1, 1] with 1 indicates original rotation direction and -1 inverted.
+    signed_drive_mode = -(drive_mode * 2 - 1)
+    position *= signed_drive_mode
+    return position
+
+
+def compute_nearest_rounded_position(position, models):
+    delta_turn = convert_degrees_to_steps(ROTATED_POSITION_DEGREE, models)
+    nearest_pos = np.round(position.astype(float) / delta_turn) * delta_turn
+    return nearest_pos.astype(position.dtype)
+
+
+def run_arm_calibration(arm: MotorsBus, robot_type: str, arm_name: str, arm_type: str):
+    """This function ensures that a neural network trained on data collected on a given robot
+    can work on another robot. For instance before calibration, setting a same goal position
+    for each motor of two different robots will get two very different positions. But after calibration,
+    the two robots will move to the same position.To this end, this function computes the homing offset
+    and the drive mode for each motor of a given robot.
+
+    Homing offset is used to shift the motor position to a ]-2048, +2048[ nominal range (when the motor uses 2048 steps
+    to complete a half a turn). This range is set around an arbitrary "zero position" corresponding to all motor positions
+    being 0. During the calibration process, you will need to manually move the robot to this "zero position".
+
+    Drive mode is used to invert the rotation direction of the motor. This is useful when some motors have been assembled
+    in the opposite orientation for some robots. During the calibration process, you will need to manually move the robot
+    to the "rotated position".
+
+    After calibration, the homing offsets and drive modes are stored in a cache.
+
+    Example of usage:
+    ```python
+    run_arm_calibration(arm, "koch", "left", "follower")
+    ```
+    """
+    if (arm.read("Torque_Enable") != TorqueMode.DISABLED.value).any():
+        raise ValueError("To run calibration, the torque must be disabled on all motors.")
+
+    print(f"\nRunning calibration of {robot_type} {arm_name} {arm_type}...")
+
+    print("\nMove arm to zero position")
+    print("See: " + URL_TEMPLATE.format(robot=robot_type, arm=arm_type, position="zero"))
+    input("Press Enter to continue...")
+
+    # We arbitrarily chose our zero target position to be a straight horizontal position with gripper upwards and closed.
+    # It is easy to identify and all motors are in a "quarter turn" position. Once calibration is done, this position will
+    # correspond to every motor angle being 0. If you set all 0 as Goal Position, the arm will move in this position.
+    zero_target_pos = convert_degrees_to_steps(ZERO_POSITION_DEGREE, arm.motor_models)
+
+    # Compute homing offset so that `present_position + homing_offset ~= target_position`.
+    zero_pos = arm.read("Present_Position")
+    zero_nearest_pos = compute_nearest_rounded_position(zero_pos, arm.motor_models)
+    homing_offset = zero_target_pos - zero_nearest_pos
+
+    # The rotated target position corresponds to a rotation of a quarter turn from the zero position.
+    # This allows to identify the rotation direction of each motor.
+    # For instance, if the motor rotates 90 degree, and its value is -90 after applying the homing offset, then we know its rotation direction
+    # is inverted. However, for the calibration being successful, we need everyone to follow the same target position.
+    # Sometimes, there is only one possible rotation direction. For instance, if the gripper is closed, there is only one direction which
+    # corresponds to opening the gripper. When the rotation direction is ambiguous, we arbitrarely rotate clockwise from the point of view
+    # of the previous motor in the kinetic chain.
+    print("\nMove arm to rotated target position")
+    print("See: " + URL_TEMPLATE.format(robot=robot_type, arm=arm_type, position="rotated"))
+    input("Press Enter to continue...")
+
+    rotated_target_pos = convert_degrees_to_steps(ROTATED_POSITION_DEGREE, arm.motor_models)
+
+    # Find drive mode by rotating each motor by a quarter of a turn.
+    # Drive mode indicates if the motor rotation direction should be inverted (=1) or not (=0).
+    rotated_pos = arm.read("Present_Position")
+    drive_mode = (rotated_pos < zero_pos).astype(np.int32)
+
+    # Re-compute homing offset to take into account drive mode
+    rotated_drived_pos = apply_drive_mode(rotated_pos, drive_mode)
+    rotated_nearest_pos = compute_nearest_rounded_position(rotated_drived_pos, arm.motor_models)
+    homing_offset = rotated_target_pos - rotated_nearest_pos
+
+    print("\nMove arm to rest position")
+    print("See: " + URL_TEMPLATE.format(robot=robot_type, arm=arm_type, position="rest"))
+    input("Press Enter to continue...")
+    print()
+
+    # Joints with rotational motions are expressed in degrees in nominal range of [-180, 180]
+    calib_mode = [CalibrationMode.DEGREE.name] * len(arm.motor_names)
+
+    # TODO(rcadene): make type of joints (DEGREE or LINEAR) configurable from yaml?
+    if robot_type in ["aloha"] and "gripper" in arm.motor_names:
+        # Joints with linear motions (like gripper of Aloha) are experessed in nominal range of [0, 100]
+        calib_idx = arm.motor_names.index("gripper")
+        calib_mode[calib_idx] = CalibrationMode.LINEAR.name
+
+    calib_data = {
+        "homing_offset": homing_offset.tolist(),
+        "drive_mode": drive_mode.tolist(),
+        "start_pos": zero_pos.tolist(),
+        "end_pos": rotated_pos.tolist(),
+        "calib_mode": calib_mode,
+        "motor_names": arm.motor_names,
+    }
+    return calib_data

lerobot/common/robot_devices/robots/factory.py

@@ -1,7 +1,9 @@
 import hydra
 from omegaconf import DictConfig
 
+from lerobot.common.robot_devices.robots.utils import Robot
 
-def make_robot(cfg: DictConfig):
+
+def make_robot(cfg: DictConfig) -> Robot:
     robot = hydra.utils.instantiate(cfg)
     return robot

lerobot/common/robot_devices/robots/feetech_calibration.py

@@ -0,0 +1,485 @@
+"""Logic to calibrate a robot arm built with feetech motors"""
+# TODO(rcadene, aliberts): move this logic into the robot code when refactoring
+
+import time
+
+import numpy as np
+
+from lerobot.common.robot_devices.motors.feetech import (
+    CalibrationMode,
+    TorqueMode,
+    convert_degrees_to_steps,
+)
+from lerobot.common.robot_devices.motors.utils import MotorsBus
+
+URL_TEMPLATE = (
+    "https://raw.githubusercontent.com/huggingface/lerobot/main/media/{robot}/{arm}_{position}.webp"
+)
+
+# The following positions are provided in nominal degree range ]-180, +180[
+# For more info on these constants, see comments in the code where they get used.
+ZERO_POSITION_DEGREE = 0
+ROTATED_POSITION_DEGREE = 90
+
+
+def assert_drive_mode(drive_mode):
+    # `drive_mode` is in [0,1] with 0 means original rotation direction for the motor, and 1 means inverted.
+    if not np.all(np.isin(drive_mode, [0, 1])):
+        raise ValueError(f"`drive_mode` contains values other than 0 or 1: ({drive_mode})")
+
+
+def apply_drive_mode(position, drive_mode):
+    assert_drive_mode(drive_mode)
+    # Convert `drive_mode` from [0, 1] with 0 indicates original rotation direction and 1 inverted,
+    # to [-1, 1] with 1 indicates original rotation direction and -1 inverted.
+    signed_drive_mode = -(drive_mode * 2 - 1)
+    position *= signed_drive_mode
+    return position
+
+
+def move_until_block(arm, motor_name, positive_direction=True, while_move_hook=None):
+    count = 0
+    while True:
+        present_pos = arm.read("Present_Position", motor_name)
+        if positive_direction:
+            # Move +100 steps every time. Lower the steps to lower the speed at which the arm moves.
+            arm.write("Goal_Position", present_pos + 100, motor_name)
+        else:
+            arm.write("Goal_Position", present_pos - 100, motor_name)
+
+        if while_move_hook is not None:
+            while_move_hook()
+
+        present_pos = arm.read("Present_Position", motor_name).item()
+        present_speed = arm.read("Present_Speed", motor_name).item()
+        present_current = arm.read("Present_Current", motor_name).item()
+        # present_load = arm.read("Present_Load", motor_name).item()
+        # present_voltage = arm.read("Present_Voltage", motor_name).item()
+        # present_temperature = arm.read("Present_Temperature", motor_name).item()
+
+        # print(f"{present_pos=}")
+        # print(f"{present_speed=}")
+        # print(f"{present_current=}")
+        # print(f"{present_load=}")
+        # print(f"{present_voltage=}")
+        # print(f"{present_temperature=}")
+
+        if present_speed == 0 and present_current > 50:
+            count += 1
+            if count > 100 or present_current > 300:
+                return present_pos
+        else:
+            count = 0
+
+
+def move_to_calibrate(
+    arm,
+    motor_name,
+    invert_drive_mode=False,
+    positive_first=True,
+    in_between_move_hook=None,
+    while_move_hook=None,
+):
+    initial_pos = arm.read("Present_Position", motor_name)
+
+    if positive_first:
+        p_present_pos = move_until_block(
+            arm, motor_name, positive_direction=True, while_move_hook=while_move_hook
+        )
+    else:
+        n_present_pos = move_until_block(
+            arm, motor_name, positive_direction=False, while_move_hook=while_move_hook
+        )
+
+    if in_between_move_hook is not None:
+        in_between_move_hook()
+
+    if positive_first:
+        n_present_pos = move_until_block(
+            arm, motor_name, positive_direction=False, while_move_hook=while_move_hook
+        )
+    else:
+        p_present_pos = move_until_block(
+            arm, motor_name, positive_direction=True, while_move_hook=while_move_hook
+        )
+
+    zero_pos = (n_present_pos + p_present_pos) / 2
+
+    calib_data = {
+        "initial_pos": initial_pos,
+        "homing_offset": zero_pos if invert_drive_mode else -zero_pos,
+        "invert_drive_mode": invert_drive_mode,
+        "drive_mode": -1 if invert_drive_mode else 0,
+        "zero_pos": zero_pos,
+        "start_pos": n_present_pos if invert_drive_mode else p_present_pos,
+        "end_pos": p_present_pos if invert_drive_mode else n_present_pos,
+    }
+    return calib_data
+
+
+def apply_offset(calib, offset):
+    calib["zero_pos"] += offset
+    if calib["drive_mode"]:
+        calib["homing_offset"] += offset
+    else:
+        calib["homing_offset"] -= offset
+    return calib
+
+
+def run_arm_auto_calibration(arm: MotorsBus, robot_type: str, arm_name: str, arm_type: str):
+    if robot_type == "so100":
+        return run_arm_auto_calibration_so100(arm, robot_type, arm_name, arm_type)
+    elif robot_type == "moss":
+        return run_arm_auto_calibration_moss(arm, robot_type, arm_name, arm_type)
+    else:
+        raise ValueError(robot_type)
+
+
+def run_arm_auto_calibration_so100(arm: MotorsBus, robot_type: str, arm_name: str, arm_type: str):
+    """All the offsets and magic numbers are hand tuned, and are unique to SO-100 follower arms"""
+    if (arm.read("Torque_Enable") != TorqueMode.DISABLED.value).any():
+        raise ValueError("To run calibration, the torque must be disabled on all motors.")
+
+    if not (robot_type == "so100" and arm_type == "follower"):
+        raise NotImplementedError("Auto calibration only supports the follower of so100 arms for now.")
+
+    print(f"\nRunning calibration of {robot_type} {arm_name} {arm_type}...")
+
+    print("\nMove arm to initial position")
+    print("See: " + URL_TEMPLATE.format(robot=robot_type, arm=arm_type, position="initial"))
+    input("Press Enter to continue...")
+
+    # Lower the acceleration of the motors (in [0,254])
+    initial_acceleration = arm.read("Acceleration")
+    arm.write("Lock", 0)
+    arm.write("Acceleration", 10)
+    time.sleep(1)
+
+    arm.write("Torque_Enable", TorqueMode.ENABLED.value)
+
+    print(f'{arm.read("Present_Position", "elbow_flex")=}')
+
+    calib = {}
+
+    init_wf_pos = arm.read("Present_Position", "wrist_flex")
+    init_sl_pos = arm.read("Present_Position", "shoulder_lift")
+    init_ef_pos = arm.read("Present_Position", "elbow_flex")
+    arm.write("Goal_Position", init_wf_pos - 800, "wrist_flex")
+    arm.write("Goal_Position", init_sl_pos + 150 + 1024, "shoulder_lift")
+    arm.write("Goal_Position", init_ef_pos - 2048, "elbow_flex")
+    time.sleep(2)
+
+    print("Calibrate shoulder_pan")
+    calib["shoulder_pan"] = move_to_calibrate(arm, "shoulder_pan")
+    arm.write("Goal_Position", calib["shoulder_pan"]["zero_pos"], "shoulder_pan")
+    time.sleep(1)
+
+    print("Calibrate gripper")
+    calib["gripper"] = move_to_calibrate(arm, "gripper", invert_drive_mode=True)
+    time.sleep(1)
+
+    print("Calibrate wrist_flex")
+    calib["wrist_flex"] = move_to_calibrate(arm, "wrist_flex")
+    calib["wrist_flex"] = apply_offset(calib["wrist_flex"], offset=80)
+
+    def in_between_move_hook():
+        nonlocal arm, calib
+        time.sleep(2)
+        ef_pos = arm.read("Present_Position", "elbow_flex")
+        sl_pos = arm.read("Present_Position", "shoulder_lift")
+        arm.write("Goal_Position", ef_pos + 1024, "elbow_flex")
+        arm.write("Goal_Position", sl_pos - 1024, "shoulder_lift")
+        time.sleep(2)
+
+    print("Calibrate elbow_flex")
+    calib["elbow_flex"] = move_to_calibrate(
+        arm, "elbow_flex", positive_first=False, in_between_move_hook=in_between_move_hook
+    )
+    calib["elbow_flex"] = apply_offset(calib["elbow_flex"], offset=80 - 1024)
+
+    arm.write("Goal_Position", calib["elbow_flex"]["zero_pos"] + 1024 + 512, "elbow_flex")
+    time.sleep(1)
+
+    def in_between_move_hook():
+        nonlocal arm, calib
+        arm.write("Goal_Position", calib["elbow_flex"]["zero_pos"], "elbow_flex")
+
+    print("Calibrate shoulder_lift")
+    calib["shoulder_lift"] = move_to_calibrate(
+        arm,
+        "shoulder_lift",
+        invert_drive_mode=True,
+        positive_first=False,
+        in_between_move_hook=in_between_move_hook,
+    )
+    # add an 30 steps as offset to align with body
+    calib["shoulder_lift"] = apply_offset(calib["shoulder_lift"], offset=1024 - 50)
+
+    def while_move_hook():
+        nonlocal arm, calib
+        positions = {
+            "shoulder_lift": round(calib["shoulder_lift"]["zero_pos"] - 1600),
+            "elbow_flex": round(calib["elbow_flex"]["zero_pos"] + 1700),
+            "wrist_flex": round(calib["wrist_flex"]["zero_pos"] + 800),
+            "gripper": round(calib["gripper"]["end_pos"]),
+        }
+        arm.write("Goal_Position", list(positions.values()), list(positions.keys()))
+
+    arm.write("Goal_Position", round(calib["shoulder_lift"]["zero_pos"] - 1600), "shoulder_lift")
+    time.sleep(2)
+    arm.write("Goal_Position", round(calib["elbow_flex"]["zero_pos"] + 1700), "elbow_flex")
+    time.sleep(2)
+    arm.write("Goal_Position", round(calib["wrist_flex"]["zero_pos"] + 800), "wrist_flex")
+    time.sleep(2)
+    arm.write("Goal_Position", round(calib["gripper"]["end_pos"]), "gripper")
+    time.sleep(2)
+
+    print("Calibrate wrist_roll")
+    calib["wrist_roll"] = move_to_calibrate(
+        arm, "wrist_roll", invert_drive_mode=True, positive_first=False, while_move_hook=while_move_hook
+    )
+
+    arm.write("Goal_Position", calib["wrist_roll"]["zero_pos"], "wrist_roll")
+    time.sleep(1)
+    arm.write("Goal_Position", calib["gripper"]["start_pos"], "gripper")
+    time.sleep(1)
+    arm.write("Goal_Position", calib["wrist_flex"]["zero_pos"], "wrist_flex")
+    time.sleep(1)
+    arm.write("Goal_Position", calib["elbow_flex"]["zero_pos"] + 2048, "elbow_flex")
+    arm.write("Goal_Position", calib["shoulder_lift"]["zero_pos"] - 2048, "shoulder_lift")
+    time.sleep(1)
+    arm.write("Goal_Position", calib["shoulder_pan"]["zero_pos"], "shoulder_pan")
+    time.sleep(1)
+
+    calib_modes = []
+    for name in arm.motor_names:
+        if name == "gripper":
+            calib_modes.append(CalibrationMode.LINEAR.name)
+        else:
+            calib_modes.append(CalibrationMode.DEGREE.name)
+
+    calib_dict = {
+        "homing_offset": [calib[name]["homing_offset"] for name in arm.motor_names],
+        "drive_mode": [calib[name]["drive_mode"] for name in arm.motor_names],
+        "start_pos": [calib[name]["start_pos"] for name in arm.motor_names],
+        "end_pos": [calib[name]["end_pos"] for name in arm.motor_names],
+        "calib_mode": calib_modes,
+        "motor_names": arm.motor_names,
+    }
+
+    # Re-enable original accerlation
+    arm.write("Lock", 0)
+    arm.write("Acceleration", initial_acceleration)
+    time.sleep(1)
+
+    return calib_dict
+
+
+def run_arm_auto_calibration_moss(arm: MotorsBus, robot_type: str, arm_name: str, arm_type: str):
+    """All the offsets and magic numbers are hand tuned, and are unique to SO-100 follower arms"""
+    if (arm.read("Torque_Enable") != TorqueMode.DISABLED.value).any():
+        raise ValueError("To run calibration, the torque must be disabled on all motors.")
+
+    if not (robot_type == "moss" and arm_type == "follower"):
+        raise NotImplementedError("Auto calibration only supports the follower of moss arms for now.")
+
+    print(f"\nRunning calibration of {robot_type} {arm_name} {arm_type}...")
+
+    print("\nMove arm to initial position")
+    print("See: " + URL_TEMPLATE.format(robot=robot_type, arm=arm_type, position="initial"))
+    input("Press Enter to continue...")
+
+    # Lower the acceleration of the motors (in [0,254])
+    initial_acceleration = arm.read("Acceleration")
+    arm.write("Lock", 0)
+    arm.write("Acceleration", 10)
+    time.sleep(1)
+
+    arm.write("Torque_Enable", TorqueMode.ENABLED.value)
+
+    sl_pos = arm.read("Present_Position", "shoulder_lift")
+    arm.write("Goal_Position", sl_pos - 1024 - 450, "shoulder_lift")
+    ef_pos = arm.read("Present_Position", "elbow_flex")
+    arm.write("Goal_Position", ef_pos + 1024 + 450, "elbow_flex")
+    time.sleep(2)
+
+    calib = {}
+
+    print("Calibrate shoulder_pan")
+    calib["shoulder_pan"] = move_to_calibrate(arm, "shoulder_pan", load_threshold=350, count_threshold=200)
+    arm.write("Goal_Position", calib["shoulder_pan"]["zero_pos"], "shoulder_pan")
+    time.sleep(1)
+
+    print("Calibrate gripper")
+    calib["gripper"] = move_to_calibrate(arm, "gripper", invert_drive_mode=True, count_threshold=200)
+    time.sleep(1)
+
+    print("Calibrate wrist_flex")
+    calib["wrist_flex"] = move_to_calibrate(arm, "wrist_flex", invert_drive_mode=True, count_threshold=200)
+    calib["wrist_flex"] = apply_offset(calib["wrist_flex"], offset=-210 + 1024)
+
+    wr_pos = arm.read("Present_Position", "wrist_roll")
+    arm.write("Goal_Position", calib["wrist_flex"]["zero_pos"] - 1024, "wrist_flex")
+    time.sleep(1)
+    arm.write("Goal_Position", wr_pos - 1024, "wrist_roll")
+    time.sleep(1)
+    arm.write("Goal_Position", calib["wrist_flex"]["zero_pos"] - 2048, "wrist_flex")
+    time.sleep(1)
+    arm.write("Goal_Position", calib["gripper"]["end_pos"], "gripper")
+    time.sleep(1)
+
+    print("Calibrate wrist_roll")
+    calib["wrist_roll"] = move_to_calibrate(arm, "wrist_roll", invert_drive_mode=True, count_threshold=200)
+    calib["wrist_roll"] = apply_offset(calib["wrist_roll"], offset=790)
+
+    arm.write("Goal_Position", calib["wrist_roll"]["zero_pos"] - 1024, "wrist_roll")
+    arm.write("Goal_Position", calib["gripper"]["start_pos"], "gripper")
+    arm.write("Goal_Position", calib["wrist_flex"]["zero_pos"] - 1024, "wrist_flex")
+    time.sleep(1)
+    arm.write("Goal_Position", calib["wrist_roll"]["zero_pos"], "wrist_roll")
+    arm.write("Goal_Position", calib["wrist_flex"]["zero_pos"] - 2048, "wrist_flex")
+
+    def in_between_move_elbow_flex_hook():
+        nonlocal arm, calib
+        arm.write("Goal_Position", calib["wrist_flex"]["zero_pos"], "wrist_flex")
+
+    print("Calibrate elbow_flex")
+    calib["elbow_flex"] = move_to_calibrate(
+        arm,
+        "elbow_flex",
+        invert_drive_mode=True,
+        count_threshold=200,
+        in_between_move_hook=in_between_move_elbow_flex_hook,
+    )
+    arm.write("Goal_Position", calib["wrist_flex"]["zero_pos"] - 1024, "wrist_flex")
+
+    def in_between_move_shoulder_lift_hook():
+        nonlocal arm, calib
+        sl = arm.read("Present_Position", "shoulder_lift")
+        arm.write("Goal_Position", sl - 1500, "shoulder_lift")
+        time.sleep(1)
+        arm.write("Goal_Position", calib["elbow_flex"]["zero_pos"] + 1536, "elbow_flex")
+        time.sleep(1)
+        arm.write("Goal_Position", calib["wrist_flex"]["start_pos"], "wrist_flex")
+        time.sleep(1)
+
+    print("Calibrate shoulder_lift")
+    calib["shoulder_lift"] = move_to_calibrate(
+        arm, "shoulder_lift", in_between_move_hook=in_between_move_shoulder_lift_hook
+    )
+    calib["shoulder_lift"] = apply_offset(calib["shoulder_lift"], offset=-1024)
+
+    arm.write("Goal_Position", calib["wrist_flex"]["zero_pos"] - 1024, "wrist_flex")
+    time.sleep(1)
+    arm.write("Goal_Position", calib["shoulder_lift"]["zero_pos"] + 2048, "shoulder_lift")
+    arm.write("Goal_Position", calib["elbow_flex"]["zero_pos"] - 1024 - 400, "elbow_flex")
+    time.sleep(2)
+
+    calib_modes = []
+    for name in arm.motor_names:
+        if name == "gripper":
+            calib_modes.append(CalibrationMode.LINEAR.name)
+        else:
+            calib_modes.append(CalibrationMode.DEGREE.name)
+
+    calib_dict = {
+        "homing_offset": [calib[name]["homing_offset"] for name in arm.motor_names],
+        "drive_mode": [calib[name]["drive_mode"] for name in arm.motor_names],
+        "start_pos": [calib[name]["start_pos"] for name in arm.motor_names],
+        "end_pos": [calib[name]["end_pos"] for name in arm.motor_names],
+        "calib_mode": calib_modes,
+        "motor_names": arm.motor_names,
+    }
+
+    # Re-enable original accerlation
+    arm.write("Lock", 0)
+    arm.write("Acceleration", initial_acceleration)
+    time.sleep(1)
+
+    return calib_dict
+
+
+def run_arm_manual_calibration(arm: MotorsBus, robot_type: str, arm_name: str, arm_type: str):
+    """This function ensures that a neural network trained on data collected on a given robot
+    can work on another robot. For instance before calibration, setting a same goal position
+    for each motor of two different robots will get two very different positions. But after calibration,
+    the two robots will move to the same position.To this end, this function computes the homing offset
+    and the drive mode for each motor of a given robot.
+
+    Homing offset is used to shift the motor position to a ]-2048, +2048[ nominal range (when the motor uses 2048 steps
+    to complete a half a turn). This range is set around an arbitrary "zero position" corresponding to all motor positions
+    being 0. During the calibration process, you will need to manually move the robot to this "zero position".
+
+    Drive mode is used to invert the rotation direction of the motor. This is useful when some motors have been assembled
+    in the opposite orientation for some robots. During the calibration process, you will need to manually move the robot
+    to the "rotated position".
+
+    After calibration, the homing offsets and drive modes are stored in a cache.
+
+    Example of usage:
+    ```python
+    run_arm_calibration(arm, "so100", "left", "follower")
+    ```
+    """
+    if (arm.read("Torque_Enable") != TorqueMode.DISABLED.value).any():
+        raise ValueError("To run calibration, the torque must be disabled on all motors.")
+
+    print(f"\nRunning calibration of {robot_type} {arm_name} {arm_type}...")
+
+    print("\nMove arm to zero position")
+    print("See: " + URL_TEMPLATE.format(robot=robot_type, arm=arm_type, position="zero"))
+    input("Press Enter to continue...")
+
+    # We arbitrarily chose our zero target position to be a straight horizontal position with gripper upwards and closed.
+    # It is easy to identify and all motors are in a "quarter turn" position. Once calibration is done, this position will
+    # correspond to every motor angle being 0. If you set all 0 as Goal Position, the arm will move in this position.
+    zero_target_pos = convert_degrees_to_steps(ZERO_POSITION_DEGREE, arm.motor_models)
+
+    # Compute homing offset so that `present_position + homing_offset ~= target_position`.
+    zero_pos = arm.read("Present_Position")
+    homing_offset = zero_target_pos - zero_pos
+
+    # The rotated target position corresponds to a rotation of a quarter turn from the zero position.
+    # This allows to identify the rotation direction of each motor.
+    # For instance, if the motor rotates 90 degree, and its value is -90 after applying the homing offset, then we know its rotation direction
+    # is inverted. However, for the calibration being successful, we need everyone to follow the same target position.
+    # Sometimes, there is only one possible rotation direction. For instance, if the gripper is closed, there is only one direction which
+    # corresponds to opening the gripper. When the rotation direction is ambiguous, we arbitrarely rotate clockwise from the point of view
+    # of the previous motor in the kinetic chain.
+    print("\nMove arm to rotated target position")
+    print("See: " + URL_TEMPLATE.format(robot=robot_type, arm=arm_type, position="rotated"))
+    input("Press Enter to continue...")
+
+    rotated_target_pos = convert_degrees_to_steps(ROTATED_POSITION_DEGREE, arm.motor_models)
+
+    # Find drive mode by rotating each motor by a quarter of a turn.
+    # Drive mode indicates if the motor rotation direction should be inverted (=1) or not (=0).
+    rotated_pos = arm.read("Present_Position")
+    drive_mode = (rotated_pos < zero_pos).astype(np.int32)
+
+    # Re-compute homing offset to take into account drive mode
+    rotated_drived_pos = apply_drive_mode(rotated_pos, drive_mode)
+    homing_offset = rotated_target_pos - rotated_drived_pos
+
+    print("\nMove arm to rest position")
+    print("See: " + URL_TEMPLATE.format(robot=robot_type, arm=arm_type, position="rest"))
+    input("Press Enter to continue...")
+    print()
+
+    # Joints with rotational motions are expressed in degrees in nominal range of [-180, 180]
+    calib_modes = []
+    for name in arm.motor_names:
+        if name == "gripper":
+            calib_modes.append(CalibrationMode.LINEAR.name)
+        else:
+            calib_modes.append(CalibrationMode.DEGREE.name)
+
+    calib_dict = {
+        "homing_offset": homing_offset.tolist(),
+        "drive_mode": drive_mode.tolist(),
+        "start_pos": zero_pos.tolist(),
+        "end_pos": rotated_pos.tolist(),
+        "calib_mode": calib_modes,
+        "motor_names": arm.motor_names,
+    }
+    return calib_dict

lerobot/common/robot_devices/robots/koch.py

@@ -1,515 +0,0 @@
-import pickle
-import time
-from dataclasses import dataclass, field, replace
-from pathlib import Path
-
-import numpy as np
-import torch
-
-from lerobot.common.robot_devices.cameras.utils import Camera
-from lerobot.common.robot_devices.motors.dynamixel import (
-    OperatingMode,
-    TorqueMode,
-    convert_degrees_to_steps,
-)
-from lerobot.common.robot_devices.motors.utils import MotorsBus
-from lerobot.common.robot_devices.utils import RobotDeviceAlreadyConnectedError, RobotDeviceNotConnectedError
-
-########################################################################
-# Calibration logic
-########################################################################
-
-URL_TEMPLATE = (
-    "https://raw.githubusercontent.com/huggingface/lerobot/main/media/{robot}/{arm}_{position}.webp"
-)
-
-# In nominal degree range ]-180, +180[
-ZERO_POSITION_DEGREE = 0
-ROTATED_POSITION_DEGREE = 90
-GRIPPER_OPEN_DEGREE = 35.156
-
-
-def assert_drive_mode(drive_mode):
-    # `drive_mode` is in [0,1] with 0 means original rotation direction for the motor, and 1 means inverted.
-    if not np.all(np.isin(drive_mode, [0, 1])):
-        raise ValueError(f"`drive_mode` contains values other than 0 or 1: ({drive_mode})")
-
-
-def apply_drive_mode(position, drive_mode):
-    assert_drive_mode(drive_mode)
-    # Convert `drive_mode` from [0, 1] with 0 indicates original rotation direction and 1 inverted,
-    # to [-1, 1] with 1 indicates original rotation direction and -1 inverted.
-    signed_drive_mode = -(drive_mode * 2 - 1)
-    position *= signed_drive_mode
-    return position
-
-
-def reset_torque_mode(arm: MotorsBus):
-    # To be configured, all servos must be in "torque disable" mode
-    arm.write("Torque_Enable", TorqueMode.DISABLED.value)
-
-    # Use 'extended position mode' for all motors except gripper, because in joint mode the servos can't
-    # rotate more than 360 degrees (from 0 to 4095) And some mistake can happen while assembling the arm,
-    # you could end up with a servo with a position 0 or 4095 at a crucial point See [
-    # https://emanual.robotis.com/docs/en/dxl/x/x_series/#operating-mode11]
-    all_motors_except_gripper = [name for name in arm.motor_names if name != "gripper"]
-    if len(all_motors_except_gripper) > 0:
-        arm.write("Operating_Mode", OperatingMode.EXTENDED_POSITION.value, all_motors_except_gripper)
-
-    # Use 'position control current based' for gripper to be limited by the limit of the current.
-    # For the follower gripper, it means it can grasp an object without forcing too much even tho,
-    # it's goal position is a complete grasp (both gripper fingers are ordered to join and reach a touch).
-    # For the leader gripper, it means we can use it as a physical trigger, since we can force with our finger
-    # to make it move, and it will move back to its original target position when we release the force.
-    arm.write("Operating_Mode", OperatingMode.CURRENT_CONTROLLED_POSITION.value, "gripper")
-
-
-def run_arm_calibration(arm: MotorsBus, name: str, arm_type: str):
-    """This function ensures that a neural network trained on data collected on a given robot
-    can work on another robot. For instance before calibration, setting a same goal position
-    for each motor of two different robots will get two very different positions. But after calibration,
-    the two robots will move to the same position.To this end, this function computes the homing offset
-    and the drive mode for each motor of a given robot.
-
-    Homing offset is used to shift the motor position to a ]-2048, +2048[ nominal range (when the motor uses 2048 steps
-    to complete a half a turn). This range is set around an arbitrary "zero position" corresponding to all motor positions
-    being 0. During the calibration process, you will need to manually move the robot to this "zero position".
-
-    Drive mode is used to invert the rotation direction of the motor. This is useful when some motors have been assembled
-    in the opposite orientation for some robots. During the calibration process, you will need to manually move the robot
-    to the "rotated position".
-
-    After calibration, the homing offsets and drive modes are stored in a cache.
-
-    Example of usage:
-    ```python
-    run_arm_calibration(arm, "left", "follower")
-    ```
-    """
-    reset_torque_mode(arm)
-
-    print(f"\nRunning calibration of {name} {arm_type}...")
-
-    print("\nMove arm to zero position")
-    print("See: " + URL_TEMPLATE.format(robot="koch", arm=arm_type, position="zero"))
-    input("Press Enter to continue...")
-
-    # We arbitrarely choosed our zero target position to be a straight horizontal position with gripper upwards and closed.
-    # It is easy to identify and all motors are in a "quarter turn" position. Once calibration is done, this position will
-    # corresponds to every motor angle being 0. If you set all 0 as Goal Position, the arm will move in this position.
-    zero_position = convert_degrees_to_steps(ZERO_POSITION_DEGREE, arm.motor_models)
-
-    def _compute_nearest_rounded_position(position, models):
-        # TODO(rcadene): Rework this function since some motors cant physically rotate a quarter turn
-        # (e.g. the gripper of Aloha arms can only rotate ~50 degree)
-        quarter_turn_degree = 90
-        quarter_turn = convert_degrees_to_steps(quarter_turn_degree, models)
-        nearest_pos = np.round(position.astype(float) / quarter_turn) * quarter_turn
-        return nearest_pos.astype(position.dtype)
-
-    # Compute homing offset so that `present_position + homing_offset ~= target_position`.
-    position = arm.read("Present_Position")
-    position = _compute_nearest_rounded_position(position, arm.motor_models)
-    homing_offset = zero_position - position
-
-    print("\nMove arm to rotated target position")
-    print("See: " + URL_TEMPLATE.format(robot="koch", arm=arm_type, position="rotated"))
-    input("Press Enter to continue...")
-
-    # The rotated target position corresponds to a rotation of a quarter turn from the zero position.
-    # This allows to identify the rotation direction of each motor.
-    # For instance, if the motor rotates 90 degree, and its value is -90 after applying the homing offset, then we know its rotation direction
-    # is inverted. However, for the calibration being successful, we need everyone to follow the same target position.
-    # Sometimes, there is only one possible rotation direction. For instance, if the gripper is closed, there is only one direction which
-    # corresponds to opening the gripper. When the rotation direction is ambiguous, we arbitrarely rotate clockwise from the point of view
-    # of the previous motor in the kinetic chain.
-    rotated_position = convert_degrees_to_steps(ROTATED_POSITION_DEGREE, arm.motor_models)
-
-    # Find drive mode by rotating each motor by a quarter of a turn.
-    # Drive mode indicates if the motor rotation direction should be inverted (=1) or not (=0).
-    position = arm.read("Present_Position")
-    position += homing_offset
-    position = _compute_nearest_rounded_position(position, arm.motor_models)
-    drive_mode = (position != rotated_position).astype(np.int32)
-
-    # Re-compute homing offset to take into account drive mode
-    position = arm.read("Present_Position")
-    position = apply_drive_mode(position, drive_mode)
-    position = _compute_nearest_rounded_position(position, arm.motor_models)
-    homing_offset = rotated_position - position
-
-    print("\nMove arm to rest position")
-    print("See: " + URL_TEMPLATE.format(robot="koch", arm=arm_type, position="rest"))
-    input("Press Enter to continue...")
-    print()
-
-    return homing_offset, drive_mode
-
-
-########################################################################
-# Alexander Koch robot arm
-########################################################################
-
-
-@dataclass
-class KochRobotConfig:
-    """
-    Example of usage:
-    ```python
-    KochRobotConfig()
-    ```
-    """
-
-    # Define all components of the robot
-    leader_arms: dict[str, MotorsBus] = field(default_factory=lambda: {})
-    follower_arms: dict[str, MotorsBus] = field(default_factory=lambda: {})
-    cameras: dict[str, Camera] = field(default_factory=lambda: {})
-
-
-class KochRobot:
-    # TODO(rcadene): Implement force feedback
-    """This class allows to control any Koch robot of various number of motors.
-
-    A few versions are available:
-    - [Koch v1.0](https://github.com/AlexanderKoch-Koch/low_cost_robot), with and without the wrist-to-elbow expansion, which was developed
-    by Alexander Koch from [Tau Robotics](https://tau-robotics.com): [Github for sourcing and assembly](
-    - [Koch v1.1])https://github.com/jess-moss/koch-v1-1), which was developed by Jess Moss.
-
-    Example of highest frequency teleoperation without camera:
-    ```python
-    # Defines how to communicate with the motors of the leader and follower arms
-    leader_arms = {
-        "main": DynamixelMotorsBus(
-            port="/dev/tty.usbmodem575E0031751",
-            motors={
-                # name: (index, model)
-                "shoulder_pan": (1, "xl330-m077"),
-                "shoulder_lift": (2, "xl330-m077"),
-                "elbow_flex": (3, "xl330-m077"),
-                "wrist_flex": (4, "xl330-m077"),
-                "wrist_roll": (5, "xl330-m077"),
-                "gripper": (6, "xl330-m077"),
-            },
-        ),
-    }
-    follower_arms = {
-        "main": DynamixelMotorsBus(
-            port="/dev/tty.usbmodem575E0032081",
-            motors={
-                # name: (index, model)
-                "shoulder_pan": (1, "xl430-w250"),
-                "shoulder_lift": (2, "xl430-w250"),
-                "elbow_flex": (3, "xl330-m288"),
-                "wrist_flex": (4, "xl330-m288"),
-                "wrist_roll": (5, "xl330-m288"),
-                "gripper": (6, "xl330-m288"),
-            },
-        ),
-    }
-    robot = KochRobot(leader_arms, follower_arms)
-
-    # Connect motors buses and cameras if any (Required)
-    robot.connect()
-
-    while True:
-        robot.teleop_step()
-    ```
-
-    Example of highest frequency data collection without camera:
-    ```python
-    # Assumes leader and follower arms have been instantiated already (see first example)
-    robot = KochRobot(leader_arms, follower_arms)
-    robot.connect()
-    while True:
-        observation, action = robot.teleop_step(record_data=True)
-    ```
-
-    Example of highest frequency data collection with cameras:
-    ```python
-    # Defines how to communicate with 2 cameras connected to the computer.
-    # Here, the webcam of the laptop and the phone (connected in USB to the laptop)
-    # can be reached respectively using the camera indices 0 and 1. These indices can be
-    # arbitrary. See the documentation of `OpenCVCamera` to find your own camera indices.
-    cameras = {
-        "laptop": OpenCVCamera(camera_index=0, fps=30, width=640, height=480),
-        "phone": OpenCVCamera(camera_index=1, fps=30, width=640, height=480),
-    }
-
-    # Assumes leader and follower arms have been instantiated already (see first example)
-    robot = KochRobot(leader_arms, follower_arms, cameras)
-    robot.connect()
-    while True:
-        observation, action = robot.teleop_step(record_data=True)
-    ```
-
-    Example of controlling the robot with a policy (without running multiple policies in parallel to ensure highest frequency):
-    ```python
-    # Assumes leader and follower arms + cameras have been instantiated already (see previous example)
-    robot = KochRobot(leader_arms, follower_arms, cameras)
-    robot.connect()
-    while True:
-        # Uses the follower arms and cameras to capture an observation
-        observation = robot.capture_observation()
-
-        # Assumes a policy has been instantiated
-        with torch.inference_mode():
-            action = policy.select_action(observation)
-
-        # Orders the robot to move
-        robot.send_action(action)
-    ```
-
-    Example of disconnecting which is not mandatory since we disconnect when the object is deleted:
-    ```python
-    robot.disconnect()
-    ```
-    """
-
-    def __init__(
-        self,
-        config: KochRobotConfig | None = None,
-        calibration_path: Path = ".cache/calibration/koch.pkl",
-        **kwargs,
-    ):
-        if config is None:
-            config = KochRobotConfig()
-        # Overwrite config arguments using kwargs
-        self.config = replace(config, **kwargs)
-        self.calibration_path = Path(calibration_path)
-
-        self.leader_arms = self.config.leader_arms
-        self.follower_arms = self.config.follower_arms
-        self.cameras = self.config.cameras
-        self.is_connected = False
-        self.logs = {}
-
-    def connect(self):
-        if self.is_connected:
-            raise RobotDeviceAlreadyConnectedError(
-                "KochRobot is already connected. Do not run `robot.connect()` twice."
-            )
-
-        if not self.leader_arms and not self.follower_arms and not self.cameras:
-            raise ValueError(
-                "KochRobot doesn't have any device to connect. See example of usage in docstring of the class."
-            )
-
-        # Connect the arms
-        for name in self.follower_arms:
-            print(f"Connecting {name} follower arm.")
-            self.follower_arms[name].connect()
-            print(f"Connecting {name} leader arm.")
-            self.leader_arms[name].connect()
-
-        # Reset the arms and load or run calibration
-        if self.calibration_path.exists():
-            # Reset all arms before setting calibration
-            for name in self.follower_arms:
-                reset_torque_mode(self.follower_arms[name])
-            for name in self.leader_arms:
-                reset_torque_mode(self.leader_arms[name])
-
-            with open(self.calibration_path, "rb") as f:
-                calibration = pickle.load(f)
-        else:
-            print(f"Missing calibration file '{self.calibration_path}'. Starting calibration precedure.")
-            # Run calibration process which begins by reseting all arms
-            calibration = self.run_calibration()
-
-            print(f"Calibration is done! Saving calibration file '{self.calibration_path}'")
-            self.calibration_path.parent.mkdir(parents=True, exist_ok=True)
-            with open(self.calibration_path, "wb") as f:
-                pickle.dump(calibration, f)
-
-        # Set calibration
-        for name in self.follower_arms:
-            self.follower_arms[name].set_calibration(calibration[f"follower_{name}"])
-        for name in self.leader_arms:
-            self.leader_arms[name].set_calibration(calibration[f"leader_{name}"])
-
-        # Set better PID values to close the gap between recored states and actions
-        # TODO(rcadene): Implement an automatic procedure to set optimial PID values for each motor
-        for name in self.follower_arms:
-            self.follower_arms[name].write("Position_P_Gain", 1500, "elbow_flex")
-            self.follower_arms[name].write("Position_I_Gain", 0, "elbow_flex")
-            self.follower_arms[name].write("Position_D_Gain", 600, "elbow_flex")
-
-        # Enable torque on all motors of the follower arms
-        for name in self.follower_arms:
-            print(f"Activating torque on {name} follower arm.")
-            self.follower_arms[name].write("Torque_Enable", 1)
-
-        # Enable torque on the gripper of the leader arms, and move it to 45 degrees,
-        # so that we can use it as a trigger to close the gripper of the follower arms.
-        for name in self.leader_arms:
-            self.leader_arms[name].write("Torque_Enable", 1, "gripper")
-            self.leader_arms[name].write("Goal_Position", GRIPPER_OPEN_DEGREE, "gripper")
-
-        # Connect the cameras
-        for name in self.cameras:
-            self.cameras[name].connect()
-
-        self.is_connected = True
-
-    def run_calibration(self):
-        calibration = {}
-
-        for name in self.follower_arms:
-            homing_offset, drive_mode = run_arm_calibration(self.follower_arms[name], name, "follower")
-
-            calibration[f"follower_{name}"] = {}
-            for idx, motor_name in enumerate(self.follower_arms[name].motor_names):
-                calibration[f"follower_{name}"][motor_name] = (homing_offset[idx], drive_mode[idx])
-
-        for name in self.leader_arms:
-            homing_offset, drive_mode = run_arm_calibration(self.leader_arms[name], name, "leader")
-
-            calibration[f"leader_{name}"] = {}
-            for idx, motor_name in enumerate(self.leader_arms[name].motor_names):
-                calibration[f"leader_{name}"][motor_name] = (homing_offset[idx], drive_mode[idx])
-
-        return calibration
-
-    def teleop_step(
-        self, record_data=False
-    ) -> None | tuple[dict[str, torch.Tensor], dict[str, torch.Tensor]]:
-        if not self.is_connected:
-            raise RobotDeviceNotConnectedError(
-                "KochRobot is not connected. You need to run `robot.connect()`."
-            )
-
-        # Prepare to assign the position of the leader to the follower
-        leader_pos = {}
-        for name in self.leader_arms:
-            before_lread_t = time.perf_counter()
-            leader_pos[name] = self.leader_arms[name].read("Present_Position")
-            self.logs[f"read_leader_{name}_pos_dt_s"] = time.perf_counter() - before_lread_t
-
-        follower_goal_pos = {}
-        for name in self.leader_arms:
-            follower_goal_pos[name] = leader_pos[name]
-
-        # Send action
-        for name in self.follower_arms:
-            before_fwrite_t = time.perf_counter()
-            self.follower_arms[name].write("Goal_Position", follower_goal_pos[name])
-            self.logs[f"write_follower_{name}_goal_pos_dt_s"] = time.perf_counter() - before_fwrite_t
-
-        # Early exit when recording data is not requested
-        if not record_data:
-            return
-
-        # TODO(rcadene): Add velocity and other info
-        # Read follower position
-        follower_pos = {}
-        for name in self.follower_arms:
-            before_fread_t = time.perf_counter()
-            follower_pos[name] = self.follower_arms[name].read("Present_Position")
-            self.logs[f"read_follower_{name}_pos_dt_s"] = time.perf_counter() - before_fread_t
-
-        # Create state by concatenating follower current position
-        state = []
-        for name in self.follower_arms:
-            if name in follower_pos:
-                state.append(follower_pos[name])
-        state = np.concatenate(state)
-
-        # Create action by concatenating follower goal position
-        action = []
-        for name in self.follower_arms:
-            if name in follower_goal_pos:
-                action.append(follower_goal_pos[name])
-        action = np.concatenate(action)
-
-        # Capture images from cameras
-        images = {}
-        for name in self.cameras:
-            before_camread_t = time.perf_counter()
-            images[name] = self.cameras[name].async_read()
-            self.logs[f"read_camera_{name}_dt_s"] = self.cameras[name].logs["delta_timestamp_s"]
-            self.logs[f"async_read_camera_{name}_dt_s"] = time.perf_counter() - before_camread_t
-
-        # Populate output dictionnaries and format to pytorch
-        obs_dict, action_dict = {}, {}
-        obs_dict["observation.state"] = torch.from_numpy(state)
-        action_dict["action"] = torch.from_numpy(action)
-        for name in self.cameras:
-            obs_dict[f"observation.images.{name}"] = torch.from_numpy(images[name])
-
-        return obs_dict, action_dict
-
-    def capture_observation(self):
-        """The returned observations do not have a batch dimension."""
-        if not self.is_connected:
-            raise RobotDeviceNotConnectedError(
-                "KochRobot is not connected. You need to run `robot.connect()`."
-            )
-
-        # Read follower position
-        follower_pos = {}
-        for name in self.follower_arms:
-            before_fread_t = time.perf_counter()
-            follower_pos[name] = self.follower_arms[name].read("Present_Position")
-            self.logs[f"read_follower_{name}_pos_dt_s"] = time.perf_counter() - before_fread_t
-
-        # Create state by concatenating follower current position
-        state = []
-        for name in self.follower_arms:
-            if name in follower_pos:
-                state.append(follower_pos[name])
-        state = np.concatenate(state)
-
-        # Capture images from cameras
-        images = {}
-        for name in self.cameras:
-            before_camread_t = time.perf_counter()
-            images[name] = self.cameras[name].async_read()
-            self.logs[f"read_camera_{name}_dt_s"] = self.cameras[name].logs["delta_timestamp_s"]
-            self.logs[f"async_read_camera_{name}_dt_s"] = time.perf_counter() - before_camread_t
-
-        # Populate output dictionnaries and format to pytorch
-        obs_dict = {}
-        obs_dict["observation.state"] = torch.from_numpy(state)
-        for name in self.cameras:
-            obs_dict[f"observation.images.{name}"] = torch.from_numpy(images[name])
-        return obs_dict
-
-    def send_action(self, action: torch.Tensor):
-        """The provided action is expected to be a vector."""
-        if not self.is_connected:
-            raise RobotDeviceNotConnectedError(
-                "KochRobot is not connected. You need to run `robot.connect()`."
-            )
-
-        from_idx = 0
-        to_idx = 0
-        follower_goal_pos = {}
-        for name in self.follower_arms:
-            if name in self.follower_arms:
-                to_idx += len(self.follower_arms[name].motor_names)
-                follower_goal_pos[name] = action[from_idx:to_idx].numpy()
-                from_idx = to_idx
-
-        for name in self.follower_arms:
-            self.follower_arms[name].write("Goal_Position", follower_goal_pos[name].astype(np.int32))
-
-    def disconnect(self):
-        if not self.is_connected:
-            raise RobotDeviceNotConnectedError(
-                "KochRobot is not connected. You need to run `robot.connect()` before disconnecting."
-            )
-
-        for name in self.follower_arms:
-            self.follower_arms[name].disconnect()
-
-        for name in self.leader_arms:
-            self.leader_arms[name].disconnect()
-
-        for name in self.cameras:
-            self.cameras[name].disconnect()
-
-        self.is_connected = False
-
-    def __del__(self):
-        if getattr(self, "is_connected", False):
-            self.disconnect()

lerobot/common/robot_devices/robots/manipulator.py

@@ -0,0 +1,663 @@
+"""Contains logic to instantiate a robot, read information from its motors and cameras,
+and send orders to its motors.
+"""
+# TODO(rcadene, aliberts): reorganize the codebase into one file per robot, with the associated
+# calibration procedure, to make it easy for people to add their own robot.
+
+import json
+import logging
+import time
+import warnings
+from dataclasses import dataclass, field, replace
+from pathlib import Path
+from typing import Sequence
+
+import numpy as np
+import torch
+
+from lerobot.common.robot_devices.cameras.utils import Camera
+from lerobot.common.robot_devices.motors.utils import MotorsBus
+from lerobot.common.robot_devices.robots.utils import get_arm_id
+from lerobot.common.robot_devices.utils import RobotDeviceAlreadyConnectedError, RobotDeviceNotConnectedError
+
+
+def ensure_safe_goal_position(
+    goal_pos: torch.Tensor, present_pos: torch.Tensor, max_relative_target: float | list[float]
+):
+    # Cap relative action target magnitude for safety.
+    diff = goal_pos - present_pos
+    max_relative_target = torch.tensor(max_relative_target)
+    safe_diff = torch.minimum(diff, max_relative_target)
+    safe_diff = torch.maximum(safe_diff, -max_relative_target)
+    safe_goal_pos = present_pos + safe_diff
+
+    if not torch.allclose(goal_pos, safe_goal_pos):
+        logging.warning(
+            "Relative goal position magnitude had to be clamped to be safe.\n"
+            f"  requested relative goal position target: {diff}\n"
+            f"    clamped relative goal position target: {safe_diff}"
+        )
+
+    return safe_goal_pos
+
+
+@dataclass
+class ManipulatorRobotConfig:
+    """
+    Example of usage:
+    ```python
+    ManipulatorRobotConfig()
+    ```
+    """
+
+    # Define all components of the robot
+    robot_type: str = "koch"
+    leader_arms: dict[str, MotorsBus] = field(default_factory=lambda: {})
+    follower_arms: dict[str, MotorsBus] = field(default_factory=lambda: {})
+    cameras: dict[str, Camera] = field(default_factory=lambda: {})
+
+    # Optionally limit the magnitude of the relative positional target vector for safety purposes.
+    # Set this to a positive scalar to have the same value for all motors, or a list that is the same length
+    # as the number of motors in your follower arms (assumes all follower arms have the same number of
+    # motors).
+    max_relative_target: list[float] | float | None = None
+
+    # Optionally set the leader arm in torque mode with the gripper motor set to this angle. This makes it
+    # possible to squeeze the gripper and have it spring back to an open position on its own. If None, the
+    # gripper is not put in torque mode.
+    gripper_open_degree: float | None = None
+
+    def __setattr__(self, prop: str, val):
+        if prop == "max_relative_target" and val is not None and isinstance(val, Sequence):
+            for name in self.follower_arms:
+                if len(self.follower_arms[name].motors) != len(val):
+                    raise ValueError(
+                        f"len(max_relative_target)={len(val)} but the follower arm with name {name} has "
+                        f"{len(self.follower_arms[name].motors)} motors. Please make sure that the "
+                        f"`max_relative_target` list has as many parameters as there are motors per arm. "
+                        "Note: This feature does not yet work with robots where different follower arms have "
+                        "different numbers of motors."
+                    )
+        super().__setattr__(prop, val)
+
+    def __post_init__(self):
+        if self.robot_type not in ["koch", "koch_bimanual", "aloha", "so100", "moss"]:
+            raise ValueError(f"Provided robot type ({self.robot_type}) is not supported.")
+
+
+class ManipulatorRobot:
+    # TODO(rcadene): Implement force feedback
+    """This class allows to control any manipulator robot of various number of motors.
+
+    Non exaustive list of robots:
+    - [Koch v1.0](https://github.com/AlexanderKoch-Koch/low_cost_robot), with and without the wrist-to-elbow expansion, developed
+    by Alexander Koch from [Tau Robotics](https://tau-robotics.com)
+    - [Koch v1.1](https://github.com/jess-moss/koch-v1-1) developed by Jess Moss
+    - [Aloha](https://www.trossenrobotics.com/aloha-kits) developed by Trossen Robotics
+
+    Example of highest frequency teleoperation without camera:
+    ```python
+    # Defines how to communicate with the motors of the leader and follower arms
+    leader_arms = {
+        "main": DynamixelMotorsBus(
+            port="/dev/tty.usbmodem575E0031751",
+            motors={
+                # name: (index, model)
+                "shoulder_pan": (1, "xl330-m077"),
+                "shoulder_lift": (2, "xl330-m077"),
+                "elbow_flex": (3, "xl330-m077"),
+                "wrist_flex": (4, "xl330-m077"),
+                "wrist_roll": (5, "xl330-m077"),
+                "gripper": (6, "xl330-m077"),
+            },
+        ),
+    }
+    follower_arms = {
+        "main": DynamixelMotorsBus(
+            port="/dev/tty.usbmodem575E0032081",
+            motors={
+                # name: (index, model)
+                "shoulder_pan": (1, "xl430-w250"),
+                "shoulder_lift": (2, "xl430-w250"),
+                "elbow_flex": (3, "xl330-m288"),
+                "wrist_flex": (4, "xl330-m288"),
+                "wrist_roll": (5, "xl330-m288"),
+                "gripper": (6, "xl330-m288"),
+            },
+        ),
+    }
+    robot = ManipulatorRobot(
+        robot_type="koch",
+        calibration_dir=".cache/calibration/koch",
+        leader_arms=leader_arms,
+        follower_arms=follower_arms,
+    )
+
+    # Connect motors buses and cameras if any (Required)
+    robot.connect()
+
+    while True:
+        robot.teleop_step()
+    ```
+
+    Example of highest frequency data collection without camera:
+    ```python
+    # Assumes leader and follower arms have been instantiated already (see first example)
+    robot = ManipulatorRobot(
+        robot_type="koch",
+        calibration_dir=".cache/calibration/koch",
+        leader_arms=leader_arms,
+        follower_arms=follower_arms,
+    )
+    robot.connect()
+    while True:
+        observation, action = robot.teleop_step(record_data=True)
+    ```
+
+    Example of highest frequency data collection with cameras:
+    ```python
+    # Defines how to communicate with 2 cameras connected to the computer.
+    # Here, the webcam of the laptop and the phone (connected in USB to the laptop)
+    # can be reached respectively using the camera indices 0 and 1. These indices can be
+    # arbitrary. See the documentation of `OpenCVCamera` to find your own camera indices.
+    cameras = {
+        "laptop": OpenCVCamera(camera_index=0, fps=30, width=640, height=480),
+        "phone": OpenCVCamera(camera_index=1, fps=30, width=640, height=480),
+    }
+
+    # Assumes leader and follower arms have been instantiated already (see first example)
+    robot = ManipulatorRobot(
+        robot_type="koch",
+        calibration_dir=".cache/calibration/koch",
+        leader_arms=leader_arms,
+        follower_arms=follower_arms,
+        cameras=cameras,
+    )
+    robot.connect()
+    while True:
+        observation, action = robot.teleop_step(record_data=True)
+    ```
+
+    Example of controlling the robot with a policy (without running multiple policies in parallel to ensure highest frequency):
+    ```python
+    # Assumes leader and follower arms + cameras have been instantiated already (see previous example)
+    robot = ManipulatorRobot(
+        robot_type="koch",
+        calibration_dir=".cache/calibration/koch",
+        leader_arms=leader_arms,
+        follower_arms=follower_arms,
+        cameras=cameras,
+    )
+    robot.connect()
+    while True:
+        # Uses the follower arms and cameras to capture an observation
+        observation = robot.capture_observation()
+
+        # Assumes a policy has been instantiated
+        with torch.inference_mode():
+            action = policy.select_action(observation)
+
+        # Orders the robot to move
+        robot.send_action(action)
+    ```
+
+    Example of disconnecting which is not mandatory since we disconnect when the object is deleted:
+    ```python
+    robot.disconnect()
+    ```
+    """
+
+    def __init__(
+        self,
+        config: ManipulatorRobotConfig | None = None,
+        calibration_dir: Path = ".cache/calibration/koch",
+        **kwargs,
+    ):
+        if config is None:
+            config = ManipulatorRobotConfig()
+        # Overwrite config arguments using kwargs
+        self.config = replace(config, **kwargs)
+        self.calibration_dir = Path(calibration_dir)
+
+        self.robot_type = self.config.robot_type
+        self.leader_arms = self.config.leader_arms
+        self.follower_arms = self.config.follower_arms
+        self.cameras = self.config.cameras
+        self.is_connected = False
+        self.logs = {}
+
+        action_names = [f"{arm}_{motor}" for arm, bus in self.leader_arms.items() for motor in bus.motors]
+        state_names = [f"{arm}_{motor}" for arm, bus in self.follower_arms.items() for motor in bus.motors]
+        self.names = {
+            "action": action_names,
+            "observation.state": state_names,
+        }
+
+    @property
+    def has_camera(self):
+        return len(self.cameras) > 0
+
+    @property
+    def num_cameras(self):
+        return len(self.cameras)
+
+    @property
+    def available_arms(self):
+        available_arms = []
+        for name in self.follower_arms:
+            arm_id = get_arm_id(name, "follower")
+            available_arms.append(arm_id)
+        for name in self.leader_arms:
+            arm_id = get_arm_id(name, "leader")
+            available_arms.append(arm_id)
+        return available_arms
+
+    def connect(self):
+        if self.is_connected:
+            raise RobotDeviceAlreadyConnectedError(
+                "ManipulatorRobot is already connected. Do not run `robot.connect()` twice."
+            )
+
+        if not self.leader_arms and not self.follower_arms and not self.cameras:
+            raise ValueError(
+                "ManipulatorRobot doesn't have any device to connect. See example of usage in docstring of the class."
+            )
+
+        # Connect the arms
+        for name in self.follower_arms:
+            print(f"Connecting {name} follower arm.")
+            self.follower_arms[name].connect()
+        for name in self.leader_arms:
+            print(f"Connecting {name} leader arm.")
+            self.leader_arms[name].connect()
+
+        if self.robot_type in ["koch", "aloha"]:
+            from lerobot.common.robot_devices.motors.dynamixel import TorqueMode
+        elif self.robot_type in ["so100", "moss"]:
+            from lerobot.common.robot_devices.motors.feetech import TorqueMode
+
+        # We assume that at connection time, arms are in a rest position, and torque can
+        # be safely disabled to run calibration and/or set robot preset configurations.
+        for name in self.follower_arms:
+            self.follower_arms[name].write("Torque_Enable", TorqueMode.DISABLED.value)
+        for name in self.leader_arms:
+            self.leader_arms[name].write("Torque_Enable", TorqueMode.DISABLED.value)
+
+        self.activate_calibration()
+
+        # Set robot preset (e.g. torque in leader gripper for Koch v1.1)
+        if self.robot_type == "koch":
+            self.set_koch_robot_preset()
+        elif self.robot_type == "aloha":
+            self.set_aloha_robot_preset()
+        elif self.robot_type in ["so100", "moss"]:
+            self.set_so100_robot_preset()
+
+        # Enable torque on all motors of the follower arms
+        for name in self.follower_arms:
+            print(f"Activating torque on {name} follower arm.")
+            self.follower_arms[name].write("Torque_Enable", 1)
+
+        if self.config.gripper_open_degree is not None:
+            if self.robot_type in ["aloha", "so100", "moss"]:
+                raise NotImplementedError(
+                    f"{self.robot_type} does not support position AND current control in the handle, which is require to set the gripper open."
+                )
+            # Set the leader arm in torque mode with the gripper motor set to an angle. This makes it possible
+            # to squeeze the gripper and have it spring back to an open position on its own.
+            for name in self.leader_arms:
+                self.leader_arms[name].write("Torque_Enable", 1, "gripper")
+                self.leader_arms[name].write("Goal_Position", self.config.gripper_open_degree, "gripper")
+
+        # Check both arms can be read
+        for name in self.follower_arms:
+            self.follower_arms[name].read("Present_Position")
+        for name in self.leader_arms:
+            self.leader_arms[name].read("Present_Position")
+
+        # Connect the cameras
+        for name in self.cameras:
+            self.cameras[name].connect()
+
+        self.is_connected = True
+
+    def activate_calibration(self):
+        """After calibration all motors function in human interpretable ranges.
+        Rotations are expressed in degrees in nominal range of [-180, 180],
+        and linear motions (like gripper of Aloha) in nominal range of [0, 100].
+        """
+
+        def load_or_run_calibration_(name, arm, arm_type):
+            arm_id = get_arm_id(name, arm_type)
+            arm_calib_path = self.calibration_dir / f"{arm_id}.json"
+
+            if arm_calib_path.exists():
+                with open(arm_calib_path) as f:
+                    calibration = json.load(f)
+            else:
+                print(f"Missing calibration file '{arm_calib_path}'")
+
+                if self.robot_type in ["koch", "aloha"]:
+                    from lerobot.common.robot_devices.robots.dynamixel_calibration import run_arm_calibration
+
+                    calibration = run_arm_calibration(arm, self.robot_type, name, arm_type)
+
+                elif self.robot_type in ["so100", "moss"]:
+                    from lerobot.common.robot_devices.robots.feetech_calibration import (
+                        run_arm_auto_calibration,
+                        run_arm_manual_calibration,
+                    )
+
+                    # TODO(rcadene): better way to handle mocking + test run_arm_auto_calibration
+                    if arm_type == "leader" or arm.mock:
+                        calibration = run_arm_manual_calibration(arm, self.robot_type, name, arm_type)
+                    elif arm_type == "follower":
+                        calibration = run_arm_auto_calibration(arm, self.robot_type, name, arm_type)
+                    else:
+                        raise ValueError(arm_type)
+
+                print(f"Calibration is done! Saving calibration file '{arm_calib_path}'")
+                arm_calib_path.parent.mkdir(parents=True, exist_ok=True)
+                with open(arm_calib_path, "w") as f:
+                    json.dump(calibration, f)
+
+            return calibration
+
+        for name, arm in self.follower_arms.items():
+            calibration = load_or_run_calibration_(name, arm, "follower")
+            arm.set_calibration(calibration)
+        for name, arm in self.leader_arms.items():
+            calibration = load_or_run_calibration_(name, arm, "leader")
+            arm.set_calibration(calibration)
+
+    def set_koch_robot_preset(self):
+        def set_operating_mode_(arm):
+            from lerobot.common.robot_devices.motors.dynamixel import TorqueMode
+
+            if (arm.read("Torque_Enable") != TorqueMode.DISABLED.value).any():
+                raise ValueError("To run set robot preset, the torque must be disabled on all motors.")
+
+            # Use 'extended position mode' for all motors except gripper, because in joint mode the servos can't
+            # rotate more than 360 degrees (from 0 to 4095) And some mistake can happen while assembling the arm,
+            # you could end up with a servo with a position 0 or 4095 at a crucial point See [
+            # https://emanual.robotis.com/docs/en/dxl/x/x_series/#operating-mode11]
+            all_motors_except_gripper = [name for name in arm.motor_names if name != "gripper"]
+            if len(all_motors_except_gripper) > 0:
+                # 4 corresponds to Extended Position on Koch motors
+                arm.write("Operating_Mode", 4, all_motors_except_gripper)
+
+            # Use 'position control current based' for gripper to be limited by the limit of the current.
+            # For the follower gripper, it means it can grasp an object without forcing too much even tho,
+            # it's goal position is a complete grasp (both gripper fingers are ordered to join and reach a touch).
+            # For the leader gripper, it means we can use it as a physical trigger, since we can force with our finger
+            # to make it move, and it will move back to its original target position when we release the force.
+            # 5 corresponds to Current Controlled Position on Koch gripper motors "xl330-m077, xl330-m288"
+            arm.write("Operating_Mode", 5, "gripper")
+
+        for name in self.follower_arms:
+            set_operating_mode_(self.follower_arms[name])
+
+            # Set better PID values to close the gap between recorded states and actions
+            # TODO(rcadene): Implement an automatic procedure to set optimial PID values for each motor
+            self.follower_arms[name].write("Position_P_Gain", 1500, "elbow_flex")
+            self.follower_arms[name].write("Position_I_Gain", 0, "elbow_flex")
+            self.follower_arms[name].write("Position_D_Gain", 600, "elbow_flex")
+
+        if self.config.gripper_open_degree is not None:
+            for name in self.leader_arms:
+                set_operating_mode_(self.leader_arms[name])
+
+                # Enable torque on the gripper of the leader arms, and move it to 45 degrees,
+                # so that we can use it as a trigger to close the gripper of the follower arms.
+                self.leader_arms[name].write("Torque_Enable", 1, "gripper")
+                self.leader_arms[name].write("Goal_Position", self.config.gripper_open_degree, "gripper")
+
+    def set_aloha_robot_preset(self):
+        def set_shadow_(arm):
+            # Set secondary/shadow ID for shoulder and elbow. These joints have two motors.
+            # As a result, if only one of them is required to move to a certain position,
+            # the other will follow. This is to avoid breaking the motors.
+            if "shoulder_shadow" in arm.motor_names:
+                shoulder_idx = arm.read("ID", "shoulder")
+                arm.write("Secondary_ID", shoulder_idx, "shoulder_shadow")
+
+            if "elbow_shadow" in arm.motor_names:
+                elbow_idx = arm.read("ID", "elbow")
+                arm.write("Secondary_ID", elbow_idx, "elbow_shadow")
+
+        for name in self.follower_arms:
+            set_shadow_(self.follower_arms[name])
+
+        for name in self.leader_arms:
+            set_shadow_(self.leader_arms[name])
+
+        for name in self.follower_arms:
+            # Set a velocity limit of 131 as advised by Trossen Robotics
+            self.follower_arms[name].write("Velocity_Limit", 131)
+
+            # Use 'extended position mode' for all motors except gripper, because in joint mode the servos can't
+            # rotate more than 360 degrees (from 0 to 4095) And some mistake can happen while assembling the arm,
+            # you could end up with a servo with a position 0 or 4095 at a crucial point See [
+            # https://emanual.robotis.com/docs/en/dxl/x/x_series/#operating-mode11]
+            all_motors_except_gripper = [
+                name for name in self.follower_arms[name].motor_names if name != "gripper"
+            ]
+            if len(all_motors_except_gripper) > 0:
+                # 4 corresponds to Extended Position on Aloha motors
+                self.follower_arms[name].write("Operating_Mode", 4, all_motors_except_gripper)
+
+            # Use 'position control current based' for follower gripper to be limited by the limit of the current.
+            # It can grasp an object without forcing too much even tho,
+            # it's goal position is a complete grasp (both gripper fingers are ordered to join and reach a touch).
+            # 5 corresponds to Current Controlled Position on Aloha gripper follower "xm430-w350"
+            self.follower_arms[name].write("Operating_Mode", 5, "gripper")
+
+            # Note: We can't enable torque on the leader gripper since "xc430-w150" doesn't have
+            # a Current Controlled Position mode.
+
+        if self.config.gripper_open_degree is not None:
+            warnings.warn(
+                f"`gripper_open_degree` is set to {self.config.gripper_open_degree}, but None is expected for Aloha instead",
+                stacklevel=1,
+            )
+
+    def set_so100_robot_preset(self):
+        for name in self.follower_arms:
+            # Mode=0 for Position Control
+            self.follower_arms[name].write("Mode", 0)
+            # Set P_Coefficient to lower value to avoid shakiness (Default is 32)
+            self.follower_arms[name].write("P_Coefficient", 16)
+            # Set I_Coefficient and D_Coefficient to default value 0 and 32
+            self.follower_arms[name].write("I_Coefficient", 0)
+            self.follower_arms[name].write("D_Coefficient", 32)
+            # Close the write lock so that Maximum_Acceleration gets written to EPROM address,
+            # which is mandatory for Maximum_Acceleration to take effect after rebooting.
+            self.follower_arms[name].write("Lock", 0)
+            # Set Maximum_Acceleration to 254 to speedup acceleration and deceleration of
+            # the motors. Note: this configuration is not in the official STS3215 Memory Table
+            self.follower_arms[name].write("Maximum_Acceleration", 254)
+            self.follower_arms[name].write("Acceleration", 254)
+
+    def teleop_step(
+        self, record_data=False
+    ) -> None | tuple[dict[str, torch.Tensor], dict[str, torch.Tensor]]:
+        if not self.is_connected:
+            raise RobotDeviceNotConnectedError(
+                "ManipulatorRobot is not connected. You need to run `robot.connect()`."
+            )
+
+        # Prepare to assign the position of the leader to the follower
+        leader_pos = {}
+        for name in self.leader_arms:
+            before_lread_t = time.perf_counter()
+            leader_pos[name] = self.leader_arms[name].read("Present_Position")
+            leader_pos[name] = torch.from_numpy(leader_pos[name])
+            self.logs[f"read_leader_{name}_pos_dt_s"] = time.perf_counter() - before_lread_t
+
+        # Send goal position to the follower
+        follower_goal_pos = {}
+        for name in self.follower_arms:
+            before_fwrite_t = time.perf_counter()
+            goal_pos = leader_pos[name]
+
+            # Cap goal position when too far away from present position.
+            # Slower fps expected due to reading from the follower.
+            if self.config.max_relative_target is not None:
+                present_pos = self.follower_arms[name].read("Present_Position")
+                present_pos = torch.from_numpy(present_pos)
+                goal_pos = ensure_safe_goal_position(goal_pos, present_pos, self.config.max_relative_target)
+
+            # Used when record_data=True
+            follower_goal_pos[name] = goal_pos
+
+            goal_pos = goal_pos.numpy().astype(np.int32)
+            self.follower_arms[name].write("Goal_Position", goal_pos)
+            self.logs[f"write_follower_{name}_goal_pos_dt_s"] = time.perf_counter() - before_fwrite_t
+
+        # Early exit when recording data is not requested
+        if not record_data:
+            return
+
+        # TODO(rcadene): Add velocity and other info
+        # Read follower position
+        follower_pos = {}
+        for name in self.follower_arms:
+            before_fread_t = time.perf_counter()
+            follower_pos[name] = self.follower_arms[name].read("Present_Position")
+            follower_pos[name] = torch.from_numpy(follower_pos[name])
+            self.logs[f"read_follower_{name}_pos_dt_s"] = time.perf_counter() - before_fread_t
+
+        # Create state by concatenating follower current position
+        state = []
+        for name in self.follower_arms:
+            if name in follower_pos:
+                state.append(follower_pos[name])
+        state = torch.cat(state)
+
+        # Create action by concatenating follower goal position
+        action = []
+        for name in self.follower_arms:
+            if name in follower_goal_pos:
+                action.append(follower_goal_pos[name])
+        action = torch.cat(action)
+
+        # Capture images from cameras
+        images = {}
+        for name in self.cameras:
+            before_camread_t = time.perf_counter()
+            images[name] = self.cameras[name].async_read()
+            images[name] = torch.from_numpy(images[name])
+            self.logs[f"read_camera_{name}_dt_s"] = self.cameras[name].logs["delta_timestamp_s"]
+            self.logs[f"async_read_camera_{name}_dt_s"] = time.perf_counter() - before_camread_t
+
+        # Populate output dictionnaries
+        obs_dict, action_dict = {}, {}
+        obs_dict["observation.state"] = state
+        action_dict["action"] = action
+        for name in self.cameras:
+            obs_dict[f"observation.images.{name}"] = images[name]
+
+        return obs_dict, action_dict
+
+    def capture_observation(self):
+        """The returned observations do not have a batch dimension."""
+        if not self.is_connected:
+            raise RobotDeviceNotConnectedError(
+                "ManipulatorRobot is not connected. You need to run `robot.connect()`."
+            )
+
+        # Read follower position
+        follower_pos = {}
+        for name in self.follower_arms:
+            before_fread_t = time.perf_counter()
+            follower_pos[name] = self.follower_arms[name].read("Present_Position")
+            follower_pos[name] = torch.from_numpy(follower_pos[name])
+            self.logs[f"read_follower_{name}_pos_dt_s"] = time.perf_counter() - before_fread_t
+
+        # Create state by concatenating follower current position
+        state = []
+        for name in self.follower_arms:
+            if name in follower_pos:
+                state.append(follower_pos[name])
+        state = torch.cat(state)
+
+        # Capture images from cameras
+        images = {}
+        for name in self.cameras:
+            before_camread_t = time.perf_counter()
+            images[name] = self.cameras[name].async_read()
+            images[name] = torch.from_numpy(images[name])
+            self.logs[f"read_camera_{name}_dt_s"] = self.cameras[name].logs["delta_timestamp_s"]
+            self.logs[f"async_read_camera_{name}_dt_s"] = time.perf_counter() - before_camread_t
+
+        # Populate output dictionnaries and format to pytorch
+        obs_dict = {}
+        obs_dict["observation.state"] = state
+        for name in self.cameras:
+            obs_dict[f"observation.images.{name}"] = images[name]
+        return obs_dict
+
+    def send_action(self, action: torch.Tensor) -> torch.Tensor:
+        """Command the follower arms to move to a target joint configuration.
+
+        The relative action magnitude may be clipped depending on the configuration parameter
+        `max_relative_target`. In this case, the action sent differs from original action.
+        Thus, this function always returns the action actually sent.
+
+        Args:
+            action: tensor containing the concatenated goal positions for the follower arms.
+        """
+        if not self.is_connected:
+            raise RobotDeviceNotConnectedError(
+                "ManipulatorRobot is not connected. You need to run `robot.connect()`."
+            )
+
+        from_idx = 0
+        to_idx = 0
+        action_sent = []
+        for name in self.follower_arms:
+            # Get goal position of each follower arm by splitting the action vector
+            to_idx += len(self.follower_arms[name].motor_names)
+            goal_pos = action[from_idx:to_idx]
+            from_idx = to_idx
+
+            # Cap goal position when too far away from present position.
+            # Slower fps expected due to reading from the follower.
+            if self.config.max_relative_target is not None:
+                present_pos = self.follower_arms[name].read("Present_Position")
+                present_pos = torch.from_numpy(present_pos)
+                goal_pos = ensure_safe_goal_position(goal_pos, present_pos, self.config.max_relative_target)
+
+            # Save tensor to concat and return
+            action_sent.append(goal_pos)
+
+            # Send goal position to each follower
+            goal_pos = goal_pos.numpy().astype(np.int32)
+            self.follower_arms[name].write("Goal_Position", goal_pos)
+
+        return torch.cat(action_sent)
+
+    def print_logs(self):
+        pass
+        # TODO(aliberts): move robot-specific logs logic here
+
+    def disconnect(self):
+        if not self.is_connected:
+            raise RobotDeviceNotConnectedError(
+                "ManipulatorRobot is not connected. You need to run `robot.connect()` before disconnecting."
+            )
+
+        for name in self.follower_arms:
+            self.follower_arms[name].disconnect()
+
+        for name in self.leader_arms:
+            self.leader_arms[name].disconnect()
+
+        for name in self.cameras:
+            self.cameras[name].disconnect()
+
+        self.is_connected = False
+
+    def __del__(self):
+        if getattr(self, "is_connected", False):
+            self.disconnect()

lerobot/common/robot_devices/robots/stretch.py

@@ -0,0 +1,216 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import time
+from dataclasses import dataclass, field, replace
+
+import torch
+from stretch_body.gamepad_teleop import GamePadTeleop
+from stretch_body.robot import Robot as StretchAPI
+from stretch_body.robot_params import RobotParams
+
+from lerobot.common.robot_devices.cameras.utils import Camera
+
+
+@dataclass
+class StretchRobotConfig:
+    robot_type: str | None = "stretch"
+    cameras: dict[str, Camera] = field(default_factory=lambda: {})
+    # TODO(aliberts): add feature with max_relative target
+    # TODO(aliberts): add comment on max_relative target
+    max_relative_target: list[float] | float | None = None
+
+
+class StretchRobot(StretchAPI):
+    """Wrapper of stretch_body.robot.Robot"""
+
+    def __init__(self, config: StretchRobotConfig | None = None, **kwargs):
+        super().__init__()
+        if config is None:
+            config = StretchRobotConfig()
+        # Overwrite config arguments using kwargs
+        self.config = replace(config, **kwargs)
+
+        self.robot_type = self.config.robot_type
+        self.cameras = self.config.cameras
+        self.is_connected = False
+        self.teleop = None
+        self.logs = {}
+
+        # TODO(aliberts): test this
+        RobotParams.set_logging_level("WARNING")
+        RobotParams.set_logging_formatter("brief_console_formatter")
+
+        self.state_keys = None
+        self.action_keys = None
+
+    def connect(self) -> None:
+        self.is_connected = self.startup()
+        if not self.is_connected:
+            print("Another process is already using Stretch. Try running 'stretch_free_robot_process.py'")
+            raise ConnectionError()
+
+        for name in self.cameras:
+            self.cameras[name].connect()
+            self.is_connected = self.is_connected and self.cameras[name].is_connected
+
+        if not self.is_connected:
+            print("Could not connect to the cameras, check that all cameras are plugged-in.")
+            raise ConnectionError()
+
+        self.run_calibration()
+
+    def run_calibration(self) -> None:
+        if not self.is_homed():
+            self.home()
+
+    def teleop_step(
+        self, record_data=False
+    ) -> None | tuple[dict[str, torch.Tensor], dict[str, torch.Tensor]]:
+        # TODO(aliberts): return ndarrays instead of torch.Tensors
+        if not self.is_connected:
+            raise ConnectionError()
+
+        if self.teleop is None:
+            self.teleop = GamePadTeleop(robot_instance=False)
+            self.teleop.startup(robot=self)
+
+        before_read_t = time.perf_counter()
+        state = self.get_state()
+        action = self.teleop.gamepad_controller.get_state()
+        self.logs["read_pos_dt_s"] = time.perf_counter() - before_read_t
+
+        before_write_t = time.perf_counter()
+        self.teleop.do_motion(robot=self)
+        self.push_command()
+        self.logs["write_pos_dt_s"] = time.perf_counter() - before_write_t
+
+        if self.state_keys is None:
+            self.state_keys = list(state)
+
+        if not record_data:
+            return
+
+        state = torch.as_tensor(list(state.values()))
+        action = torch.as_tensor(list(action.values()))
+
+        # Capture images from cameras
+        images = {}
+        for name in self.cameras:
+            before_camread_t = time.perf_counter()
+            images[name] = self.cameras[name].async_read()
+            images[name] = torch.from_numpy(images[name])
+            self.logs[f"read_camera_{name}_dt_s"] = self.cameras[name].logs["delta_timestamp_s"]
+            self.logs[f"async_read_camera_{name}_dt_s"] = time.perf_counter() - before_camread_t
+
+        # Populate output dictionnaries
+        obs_dict, action_dict = {}, {}
+        obs_dict["observation.state"] = state
+        action_dict["action"] = action
+        for name in self.cameras:
+            obs_dict[f"observation.images.{name}"] = images[name]
+
+        return obs_dict, action_dict
+
+    def get_state(self) -> dict:
+        status = self.get_status()
+        return {
+            "head_pan.pos": status["head"]["head_pan"]["pos"],
+            "head_tilt.pos": status["head"]["head_tilt"]["pos"],
+            "lift.pos": status["lift"]["pos"],
+            "arm.pos": status["arm"]["pos"],
+            "wrist_pitch.pos": status["end_of_arm"]["wrist_pitch"]["pos"],
+            "wrist_roll.pos": status["end_of_arm"]["wrist_roll"]["pos"],
+            "wrist_yaw.pos": status["end_of_arm"]["wrist_yaw"]["pos"],
+            "gripper.pos": status["end_of_arm"]["stretch_gripper"]["pos"],
+            "base_x.vel": status["base"]["x_vel"],
+            "base_y.vel": status["base"]["y_vel"],
+            "base_theta.vel": status["base"]["theta_vel"],
+        }
+
+    def capture_observation(self) -> dict:
+        # TODO(aliberts): return ndarrays instead of torch.Tensors
+        before_read_t = time.perf_counter()
+        state = self.get_state()
+        self.logs["read_pos_dt_s"] = time.perf_counter() - before_read_t
+
+        if self.state_keys is None:
+            self.state_keys = list(state)
+
+        state = torch.as_tensor(list(state.values()))
+
+        # Capture images from cameras
+        images = {}
+        for name in self.cameras:
+            before_camread_t = time.perf_counter()
+            images[name] = self.cameras[name].async_read()
+            images[name] = torch.from_numpy(images[name])
+            self.logs[f"read_camera_{name}_dt_s"] = self.cameras[name].logs["delta_timestamp_s"]
+            self.logs[f"async_read_camera_{name}_dt_s"] = time.perf_counter() - before_camread_t
+
+        # Populate output dictionnaries
+        obs_dict = {}
+        obs_dict["observation.state"] = state
+        for name in self.cameras:
+            obs_dict[f"observation.images.{name}"] = images[name]
+
+        return obs_dict
+
+    def send_action(self, action: torch.Tensor) -> torch.Tensor:
+        # TODO(aliberts): return ndarrays instead of torch.Tensors
+        if not self.is_connected:
+            raise ConnectionError()
+
+        if self.teleop is None:
+            self.teleop = GamePadTeleop(robot_instance=False)
+            self.teleop.startup(robot=self)
+
+        if self.action_keys is None:
+            dummy_action = self.teleop.gamepad_controller.get_state()
+            self.action_keys = list(dummy_action.keys())
+
+        action_dict = dict(zip(self.action_keys, action.tolist(), strict=True))
+
+        before_write_t = time.perf_counter()
+        self.teleop.do_motion(state=action_dict, robot=self)
+        self.push_command()
+        self.logs["write_pos_dt_s"] = time.perf_counter() - before_write_t
+
+        # TODO(aliberts): return action_sent when motion is limited
+        return action
+
+    def print_logs(self) -> None:
+        pass
+        # TODO(aliberts): move robot-specific logs logic here
+
+    def teleop_safety_stop(self) -> None:
+        if self.teleop is not None:
+            self.teleop._safety_stop(robot=self)
+
+    def disconnect(self) -> None:
+        self.stop()
+        if self.teleop is not None:
+            self.teleop.gamepad_controller.stop()
+            self.teleop.stop()
+
+        if len(self.cameras) > 0:
+            for cam in self.cameras.values():
+                cam.disconnect()
+
+        self.is_connected = False
+
+    def __del__(self):
+        self.disconnect()

lerobot/common/robot_devices/robots/utils.py

@@ -1,9 +1,20 @@
 from typing import Protocol
 
 
+def get_arm_id(name, arm_type):
+    """Returns the string identifier of a robot arm. For instance, for a bimanual manipulator
+    like Aloha, it could be left_follower, right_follower, left_leader, or right_leader.
+    """
+    return f"{name}_{arm_type}"
+
+
 class Robot(Protocol):
-    def init_teleop(self): ...
+    # TODO(rcadene, aliberts): Add unit test checking the protocol is implemented in the corresponding classes
+    robot_type: str
+
+    def connect(self): ...
     def run_calibration(self): ...
     def teleop_step(self, record_data=False): ...
     def capture_observation(self): ...
     def send_action(self, action): ...
+    def disconnect(self): ...

lerobot/common/robot_devices/utils.py

@@ -1,3 +1,35 @@
+import platform
+import time
+
+
+def busy_wait(seconds):
+    if platform.system() == "Darwin":
+        # On Mac, `time.sleep` is not accurate and we need to use this while loop trick,
+        # but it consumes CPU cycles.
+        # TODO(rcadene): find an alternative: from python 11, time.sleep is precise
+        end_time = time.perf_counter() + seconds
+        while time.perf_counter() < end_time:
+            pass
+    else:
+        # On Linux time.sleep is accurate
+        if seconds > 0:
+            time.sleep(seconds)
+
+
+def safe_disconnect(func):
+    # TODO(aliberts): Allow to pass custom exceptions
+    # (e.g. ThreadServiceExit, KeyboardInterrupt, SystemExit, UnpluggedError, DynamixelCommError)
+    def wrapper(robot, *args, **kwargs):
+        try:
+            return func(robot, *args, **kwargs)
+        except Exception as e:
+            if robot.is_connected:
+                robot.disconnect()
+            raise e
+
+    return wrapper
+
+
 class RobotDeviceNotConnectedError(Exception):
     """Exception raised when the robot device is not connected."""
 

lerobot/common/utils/utils.py

@@ -14,7 +14,9 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import logging
+import os
 import os.path as osp
+import platform
 import random
 from contextlib import contextmanager
 from datetime import datetime, timezone
@@ -27,6 +29,18 @@ import torch
 from omegaconf import DictConfig
 
 
+def none_or_int(value):
+    if value == "None":
+        return None
+    return int(value)
+
+
+def inside_slurm():
+    """Check whether the python process was launched through slurm"""
+    # TODO(rcadene): return False for interactive mode `--pty bash`
+    return "SLURM_JOB_ID" in os.environ
+
+
 def get_safe_torch_device(cfg_device: str, log: bool = False) -> torch.device:
     """Given a string, return a torch.device with checks on whether the device is available."""
     match cfg_device:
@@ -158,7 +172,6 @@ def init_hydra_config(config_path: str, overrides: list[str] | None = None) -> D
         version_base="1.2",
     )
     cfg = hydra.compose(Path(config_path).stem, overrides)
-
     return cfg
 
 
@@ -177,3 +190,30 @@ def print_cuda_memory_usage():
 
 def capture_timestamp_utc():
     return datetime.now(timezone.utc)
+
+
+def say(text, blocking=False):
+    # Check if mac, linux, or windows.
+    if platform.system() == "Darwin":
+        cmd = f'say "{text}"'
+        if not blocking:
+            cmd += " &"
+    elif platform.system() == "Linux":
+        cmd = f'spd-say "{text}"'
+        if blocking:
+            cmd += "  --wait"
+    elif platform.system() == "Windows":
+        # TODO(rcadene): Make blocking option work for Windows
+        cmd = (
+            'PowerShell -Command "Add-Type -AssemblyName System.Speech; '
+            f"(New-Object System.Speech.Synthesis.SpeechSynthesizer).Speak('{text}')\""
+        )
+
+    os.system(cmd)
+
+
+def log_say(text, play_sounds, blocking=False):
+    logging.info(text)
+
+    if play_sounds:
+        say(text, blocking)

lerobot/configs/env/aloha_real.yaml

@@ -0,0 +1,10 @@
+# @package _global_
+
+fps: 30
+
+env:
+  name: real_world
+  task: null
+  state_dim: 18
+  action_dim: 18
+  fps: ${fps}

lerobot/configs/policy/act_aloha_real.yaml

@@ -1,16 +1,22 @@
 # @package _global_
 
-# Use `act_real.yaml` to train on real-world Aloha/Aloha2 datasets.
-# Compared to `act.yaml`, it contains 4 cameras (i.e. cam_right_wrist, cam_left_wrist, images,
-# cam_low) instead of 1 camera (i.e. top). Also, `training.eval_freq` is set to -1. This config is used
-# to evaluate checkpoints at a certain frequency of training steps. When it is set to -1, it deactivates evaluation.
-# This is because real-world evaluation is done through [dora-lerobot](https://github.com/dora-rs/dora-lerobot).
-# Look at its README for more information on how to evaluate a checkpoint in the real-world.
+# Use `act_aloha_real.yaml` to train on real-world datasets collected on Aloha or Aloha-2 robots.
+# Compared to `act.yaml`, it contains 4 cameras (i.e. cam_right_wrist, cam_left_wrist, cam_high, cam_low) instead of 1 camera (i.e. top).
+# Also, `training.eval_freq` is set to -1. This config is used to evaluate checkpoints at a certain frequency of training steps.
+# When it is set to -1, it deactivates evaluation. This is because real-world evaluation is done through our `control_robot.py` script.
+# Look at the documentation in header of `control_robot.py` for more information on how to collect data , train and evaluate a policy.
 #
-# Example of usage for training:
+# Example of usage for training and inference with `control_robot.py`:
 # ```bash
 # python lerobot/scripts/train.py \
-#   policy=act_real \
+#   policy=act_aloha_real \
+#   env=aloha_real
+# ```
+#
+# Example of usage for training and inference with [Dora-rs](https://github.com/dora-rs/dora-lerobot):
+# ```bash
+# python lerobot/scripts/train.py \
+#   policy=act_aloha_real \
 #   env=dora_aloha_real
 # ```
 
@@ -36,10 +42,11 @@ override_dataset_stats:
     std: [[[0.229]], [[0.224]], [[0.225]]]  # (c,1,1)
 
 training:
-  offline_steps: 100000
+  offline_steps: 80000
   online_steps: 0
   eval_freq: -1
-  save_freq: 20000
+  save_freq: 10000
+  log_freq: 100
   save_checkpoint: true
 
   batch_size: 8
@@ -62,7 +69,7 @@ policy:
 
   # Input / output structure.
   n_obs_steps: 1
-  chunk_size: 100 # chunk_size
+  chunk_size: 100
   n_action_steps: 100
 
   input_shapes:
@@ -107,7 +114,7 @@ policy:
   n_vae_encoder_layers: 4
 
   # Inference.
-  temporal_ensemble_coeff: null
+  temporal_ensemble_momentum: null
 
   # Training and loss computation.
   dropout: 0.1

lerobot/configs/policy/act_real_no_state.yaml

@@ -1,110 +0,0 @@
-# @package _global_
-
-# Use `act_real_no_state.yaml` to train on real-world Aloha/Aloha2 datasets when cameras are moving (e.g. wrist cameras)
-# Compared to `act_real.yaml`, it is camera only and does not use the state as input which is vector of robot joint positions.
-# We validated experimentaly that not using state reaches better success rate. Our hypothesis is that `act_real.yaml` might
-# overfits to the state, because the images are more complex to learn from since they are moving.
-#
-# Example of usage for training:
-# ```bash
-# python lerobot/scripts/train.py \
-#   policy=act_real_no_state \
-#   env=dora_aloha_real
-# ```
-
-seed: 1000
-dataset_repo_id: lerobot/aloha_static_vinh_cup
-
-override_dataset_stats:
-  observation.images.cam_right_wrist:
-    # stats from imagenet, since we use a pretrained vision model
-    mean: [[[0.485]], [[0.456]], [[0.406]]]  # (c,1,1)
-    std: [[[0.229]], [[0.224]], [[0.225]]]  # (c,1,1)
-  observation.images.cam_left_wrist:
-    # stats from imagenet, since we use a pretrained vision model
-    mean: [[[0.485]], [[0.456]], [[0.406]]]  # (c,1,1)
-    std: [[[0.229]], [[0.224]], [[0.225]]]  # (c,1,1)
-  observation.images.cam_high:
-    # stats from imagenet, since we use a pretrained vision model
-    mean: [[[0.485]], [[0.456]], [[0.406]]]  # (c,1,1)
-    std: [[[0.229]], [[0.224]], [[0.225]]]  # (c,1,1)
-  observation.images.cam_low:
-    # stats from imagenet, since we use a pretrained vision model
-    mean: [[[0.485]], [[0.456]], [[0.406]]]  # (c,1,1)
-    std: [[[0.229]], [[0.224]], [[0.225]]]  # (c,1,1)
-
-training:
-  offline_steps: 100000
-  online_steps: 0
-  eval_freq: -1
-  save_freq: 20000
-  save_checkpoint: true
-
-  batch_size: 8
-  lr: 1e-5
-  lr_backbone: 1e-5
-  weight_decay: 1e-4
-  grad_clip_norm: 10
-  online_steps_between_rollouts: 1
-
-  delta_timestamps:
-    action: "[i / ${fps} for i in range(${policy.chunk_size})]"
-
-eval:
-  n_episodes: 50
-  batch_size: 50
-
-# See `configuration_act.py` for more details.
-policy:
-  name: act
-
-  # Input / output structure.
-  n_obs_steps: 1
-  chunk_size: 100 # chunk_size
-  n_action_steps: 100
-
-  input_shapes:
-    # TODO(rcadene, alexander-soare): add variables for height and width from the dataset/env?
-    observation.images.cam_right_wrist: [3, 480, 640]
-    observation.images.cam_left_wrist: [3, 480, 640]
-    observation.images.cam_high: [3, 480, 640]
-    observation.images.cam_low: [3, 480, 640]
-  output_shapes:
-    action: ["${env.action_dim}"]
-
-  # Normalization / Unnormalization
-  input_normalization_modes:
-    observation.images.cam_right_wrist: mean_std
-    observation.images.cam_left_wrist: mean_std
-    observation.images.cam_high: mean_std
-    observation.images.cam_low: mean_std
-  output_normalization_modes:
-    action: mean_std
-
-  # Architecture.
-  # Vision backbone.
-  vision_backbone: resnet18
-  pretrained_backbone_weights: ResNet18_Weights.IMAGENET1K_V1
-  replace_final_stride_with_dilation: false
-  # Transformer layers.
-  pre_norm: false
-  dim_model: 512
-  n_heads: 8
-  dim_feedforward: 3200
-  feedforward_activation: relu
-  n_encoder_layers: 4
-  # Note: Although the original ACT implementation has 7 for `n_decoder_layers`, there is a bug in the code
-  # that means only the first layer is used. Here we match the original implementation by setting this to 1.
-  # See this issue https://github.com/tonyzhaozh/act/issues/25#issue-2258740521.
-  n_decoder_layers: 1
-  # VAE.
-  use_vae: true
-  latent_dim: 32
-  n_vae_encoder_layers: 4
-
-  # Inference.
-  temporal_ensemble_coeff: null
-
-  # Training and loss computation.
-  dropout: 0.1
-  kl_weight: 10.0

lerobot/configs/policy/tdmpc.yaml

@@ -12,6 +12,7 @@ training:
   grad_clip_norm: 10.0
   lr: 3e-4
 
+  save_freq: 10000
   eval_freq: 5000
   log_freq: 100
 

lerobot/configs/robot/aloha.yaml

@@ -0,0 +1,117 @@
+# [Aloha: A Low-Cost Hardware for Bimanual Teleoperation](https://www.trossenrobotics.com/aloha-stationary)
+# https://aloha-2.github.io
+
+# Requires installing extras packages
+# With pip: `pip install -e ".[dynamixel intelrealsense]"`
+# With poetry: `poetry install --sync --extras "dynamixel intelrealsense"`
+
+# See [tutorial](https://github.com/huggingface/lerobot/blob/main/examples/9_use_aloha.md)
+
+
+_target_: lerobot.common.robot_devices.robots.manipulator.ManipulatorRobot
+robot_type: aloha
+# Specific to Aloha, LeRobot comes with default calibration files. Assuming the motors have been
+# properly assembled, no manual calibration step is expected. If you need to run manual calibration,
+# simply update this path to ".cache/calibration/aloha"
+calibration_dir: .cache/calibration/aloha_default
+
+# /!\ FOR SAFETY, READ THIS /!\
+# `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
+# Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
+# the number of motors in your follower arms.
+# For Aloha, for every goal position request, motor rotations are capped at 5 degrees by default.
+# When you feel more confident with teleoperation or running the policy, you can extend
+# this safety limit and even removing it by setting it to `null`.
+# Also, everything is expected to work safely out-of-the-box, but we highly advise to
+# first try to teleoperate the grippers only (by commenting out the rest of the motors in this yaml),
+# then to gradually add more motors (by uncommenting), until you can teleoperate both arms fully
+max_relative_target: 5
+
+leader_arms:
+  left:
+    _target_: lerobot.common.robot_devices.motors.dynamixel.DynamixelMotorsBus
+    port: /dev/ttyDXL_leader_left
+    motors:  # window_x
+      # name: (index, model)
+      waist: [1, xm430-w350]
+      shoulder: [2, xm430-w350]
+      shoulder_shadow: [3, xm430-w350]
+      elbow: [4, xm430-w350]
+      elbow_shadow: [5, xm430-w350]
+      forearm_roll: [6, xm430-w350]
+      wrist_angle: [7, xm430-w350]
+      wrist_rotate: [8, xl430-w250]
+      gripper: [9, xc430-w150]
+  right:
+    _target_: lerobot.common.robot_devices.motors.dynamixel.DynamixelMotorsBus
+    port: /dev/ttyDXL_leader_right
+    motors:  # window_x
+      # name: (index, model)
+      waist: [1, xm430-w350]
+      shoulder: [2, xm430-w350]
+      shoulder_shadow: [3, xm430-w350]
+      elbow: [4, xm430-w350]
+      elbow_shadow: [5, xm430-w350]
+      forearm_roll: [6, xm430-w350]
+      wrist_angle: [7, xm430-w350]
+      wrist_rotate: [8, xl430-w250]
+      gripper: [9, xc430-w150]
+
+follower_arms:
+  left:
+    _target_: lerobot.common.robot_devices.motors.dynamixel.DynamixelMotorsBus
+    port: /dev/ttyDXL_follower_left
+    motors:
+      # name: [index, model]
+      waist: [1, xm540-w270]
+      shoulder: [2, xm540-w270]
+      shoulder_shadow: [3, xm540-w270]
+      elbow: [4, xm540-w270]
+      elbow_shadow: [5, xm540-w270]
+      forearm_roll: [6, xm540-w270]
+      wrist_angle: [7, xm540-w270]
+      wrist_rotate: [8, xm430-w350]
+      gripper: [9, xm430-w350]
+  right:
+    _target_: lerobot.common.robot_devices.motors.dynamixel.DynamixelMotorsBus
+    port: /dev/ttyDXL_follower_right
+    motors:
+      # name: [index, model]
+      waist: [1, xm540-w270]
+      shoulder: [2, xm540-w270]
+      shoulder_shadow: [3, xm540-w270]
+      elbow: [4, xm540-w270]
+      elbow_shadow: [5, xm540-w270]
+      forearm_roll: [6, xm540-w270]
+      wrist_angle: [7, xm540-w270]
+      wrist_rotate: [8, xm430-w350]
+      gripper: [9, xm430-w350]
+
+# Troubleshooting: If one of your IntelRealSense cameras freeze during
+# data recording due to bandwidth limit, you might need to plug the camera
+# on another USB hub or PCIe card.
+cameras:
+  cam_high:
+    _target_: lerobot.common.robot_devices.cameras.intelrealsense.IntelRealSenseCamera
+    serial_number: 128422271347
+    fps: 30
+    width: 640
+    height: 480
+  cam_low:
+    _target_: lerobot.common.robot_devices.cameras.intelrealsense.IntelRealSenseCamera
+    serial_number: 130322270656
+    fps: 30
+    width: 640
+    height: 480
+  cam_left_wrist:
+    _target_: lerobot.common.robot_devices.cameras.intelrealsense.IntelRealSenseCamera
+    serial_number: 218622272670
+    fps: 30
+    width: 640
+    height: 480
+  cam_right_wrist:
+    _target_: lerobot.common.robot_devices.cameras.intelrealsense.IntelRealSenseCamera
+    serial_number: 130322272300
+    fps: 30
+    width: 640
+    height: 480

lerobot/configs/robot/koch.yaml

@@ -1,5 +1,12 @@
-_target_: lerobot.common.robot_devices.robots.koch.KochRobot
-calibration_path: .cache/calibration/koch.pkl
+_target_: lerobot.common.robot_devices.robots.manipulator.ManipulatorRobot
+robot_type: koch
+calibration_dir: .cache/calibration/koch
+
+# `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
+# Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
+# the number of motors in your follower arms.
+max_relative_target: null
+
 leader_arms:
   main:
     _target_: lerobot.common.robot_devices.motors.dynamixel.DynamixelMotorsBus
@@ -12,6 +19,7 @@ leader_arms:
       wrist_flex: [4, "xl330-m077"]
       wrist_roll: [5, "xl330-m077"]
       gripper: [6, "xl330-m077"]
+
 follower_arms:
   main:
     _target_: lerobot.common.robot_devices.motors.dynamixel.DynamixelMotorsBus
@@ -24,6 +32,7 @@ follower_arms:
       wrist_flex: [4, "xl330-m288"]
       wrist_roll: [5, "xl330-m288"]
       gripper: [6, "xl330-m288"]
+
 cameras:
   laptop:
     _target_: lerobot.common.robot_devices.cameras.opencv.OpenCVCamera
@@ -37,3 +46,8 @@ cameras:
     fps: 30
     width: 640
     height: 480
+
+# ~ Koch specific settings ~
+# Sets the leader arm in torque mode with the gripper motor set to this angle. This makes it possible
+# to squeeze the gripper and have it spring back to an open position on its own.
+gripper_open_degree: 35.156

lerobot/configs/robot/koch_bimanual.yaml

@@ -0,0 +1,75 @@
+_target_: lerobot.common.robot_devices.robots.manipulator.ManipulatorRobot
+robot_type: koch
+calibration_dir: .cache/calibration/koch_bimanual
+
+# `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
+# Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
+# the number of motors in your follower arms.
+max_relative_target: null
+
+leader_arms:
+  left:
+    _target_: lerobot.common.robot_devices.motors.dynamixel.DynamixelMotorsBus
+    port: /dev/tty.usbmodem585A0085511
+    motors:
+      # name: (index, model)
+      shoulder_pan: [1, "xl330-m077"]
+      shoulder_lift: [2, "xl330-m077"]
+      elbow_flex: [3, "xl330-m077"]
+      wrist_flex: [4, "xl330-m077"]
+      wrist_roll: [5, "xl330-m077"]
+      gripper: [6, "xl330-m077"]
+  right:
+    _target_: lerobot.common.robot_devices.motors.dynamixel.DynamixelMotorsBus
+    port: /dev/tty.usbmodem575E0031751
+    motors:
+      # name: (index, model)
+      shoulder_pan: [1, "xl330-m077"]
+      shoulder_lift: [2, "xl330-m077"]
+      elbow_flex: [3, "xl330-m077"]
+      wrist_flex: [4, "xl330-m077"]
+      wrist_roll: [5, "xl330-m077"]
+      gripper: [6, "xl330-m077"]
+
+follower_arms:
+  left:
+    _target_: lerobot.common.robot_devices.motors.dynamixel.DynamixelMotorsBus
+    port: /dev/tty.usbmodem585A0076891
+    motors:
+      # name: (index, model)
+      shoulder_pan: [1, "xl430-w250"]
+      shoulder_lift: [2, "xl430-w250"]
+      elbow_flex: [3, "xl330-m288"]
+      wrist_flex: [4, "xl330-m288"]
+      wrist_roll: [5, "xl330-m288"]
+      gripper: [6, "xl330-m288"]
+  right:
+    _target_: lerobot.common.robot_devices.motors.dynamixel.DynamixelMotorsBus
+    port: /dev/tty.usbmodem575E0032081
+    motors:
+      # name: (index, model)
+      shoulder_pan: [1, "xl430-w250"]
+      shoulder_lift: [2, "xl430-w250"]
+      elbow_flex: [3, "xl330-m288"]
+      wrist_flex: [4, "xl330-m288"]
+      wrist_roll: [5, "xl330-m288"]
+      gripper: [6, "xl330-m288"]
+
+cameras:
+  laptop:
+    _target_: lerobot.common.robot_devices.cameras.opencv.OpenCVCamera
+    camera_index: 0
+    fps: 30
+    width: 640
+    height: 480
+  phone:
+    _target_: lerobot.common.robot_devices.cameras.opencv.OpenCVCamera
+    camera_index: 1
+    fps: 30
+    width: 640
+    height: 480
+
+# ~ Koch specific settings ~
+# Sets the leader arm in torque mode with the gripper motor set to this angle. This makes it possible
+# to squeeze the gripper and have it spring back to an open position on its own.
+gripper_open_degree: 35.156

lerobot/configs/robot/moss.yaml

@@ -0,0 +1,56 @@
+# [Moss v1 robot arm](https://github.com/jess-moss/moss-robot-arms)
+
+# Requires installing extras packages
+# With pip: `pip install -e ".[feetech]"`
+# With poetry: `poetry install --sync --extras "feetech"`
+
+# See [tutorial](https://github.com/huggingface/lerobot/blob/main/examples/11_use_moss.md)
+
+_target_: lerobot.common.robot_devices.robots.manipulator.ManipulatorRobot
+robot_type: moss
+calibration_dir: .cache/calibration/moss
+
+# `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
+# Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
+# the number of motors in your follower arms.
+max_relative_target: null
+
+leader_arms:
+  main:
+    _target_: lerobot.common.robot_devices.motors.feetech.FeetechMotorsBus
+    port: /dev/tty.usbmodem58760431091
+    motors:
+      # name: (index, model)
+      shoulder_pan: [1, "sts3215"]
+      shoulder_lift: [2, "sts3215"]
+      elbow_flex: [3, "sts3215"]
+      wrist_flex: [4, "sts3215"]
+      wrist_roll: [5, "sts3215"]
+      gripper: [6, "sts3215"]
+
+follower_arms:
+  main:
+    _target_: lerobot.common.robot_devices.motors.feetech.FeetechMotorsBus
+    port: /dev/tty.usbmodem58760431191
+    motors:
+      # name: (index, model)
+      shoulder_pan: [1, "sts3215"]
+      shoulder_lift: [2, "sts3215"]
+      elbow_flex: [3, "sts3215"]
+      wrist_flex: [4, "sts3215"]
+      wrist_roll: [5, "sts3215"]
+      gripper: [6, "sts3215"]
+
+cameras:
+  laptop:
+    _target_: lerobot.common.robot_devices.cameras.opencv.OpenCVCamera
+    camera_index: 0
+    fps: 30
+    width: 640
+    height: 480
+  phone:
+    _target_: lerobot.common.robot_devices.cameras.opencv.OpenCVCamera
+    camera_index: 1
+    fps: 30
+    width: 640
+    height: 480

lerobot/configs/robot/so100.yaml

@@ -0,0 +1,56 @@
+# [SO-100 robot arm](https://github.com/TheRobotStudio/SO-ARM100)
+
+# Requires installing extras packages
+# With pip: `pip install -e ".[feetech]"`
+# With poetry: `poetry install --sync --extras "feetech"`
+
+# See [tutorial](https://github.com/huggingface/lerobot/blob/main/examples/10_use_so100.md)
+
+_target_: lerobot.common.robot_devices.robots.manipulator.ManipulatorRobot
+robot_type: so100
+calibration_dir: .cache/calibration/so100
+
+# `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
+# Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
+# the number of motors in your follower arms.
+max_relative_target: null
+
+leader_arms:
+  main:
+    _target_: lerobot.common.robot_devices.motors.feetech.FeetechMotorsBus
+    port: /dev/tty.usbmodem585A0077581
+    motors:
+      # name: (index, model)
+      shoulder_pan: [1, "sts3215"]
+      shoulder_lift: [2, "sts3215"]
+      elbow_flex: [3, "sts3215"]
+      wrist_flex: [4, "sts3215"]
+      wrist_roll: [5, "sts3215"]
+      gripper: [6, "sts3215"]
+
+follower_arms:
+  main:
+    _target_: lerobot.common.robot_devices.motors.feetech.FeetechMotorsBus
+    port: /dev/tty.usbmodem585A0080971
+    motors:
+      # name: (index, model)
+      shoulder_pan: [1, "sts3215"]
+      shoulder_lift: [2, "sts3215"]
+      elbow_flex: [3, "sts3215"]
+      wrist_flex: [4, "sts3215"]
+      wrist_roll: [5, "sts3215"]
+      gripper: [6, "sts3215"]
+
+cameras:
+  laptop:
+    _target_: lerobot.common.robot_devices.cameras.opencv.OpenCVCamera
+    camera_index: 0
+    fps: 30
+    width: 640
+    height: 480
+  phone:
+    _target_: lerobot.common.robot_devices.cameras.opencv.OpenCVCamera
+    camera_index: 1
+    fps: 30
+    width: 640
+    height: 480

lerobot/configs/robot/stretch.yaml

@@ -0,0 +1,33 @@
+# [Stretch3 from Hello Robot](https://hello-robot.com/stretch-3-product)
+
+# Requires installing extras packages
+# With pip: `pip install -e ".[stretch]"`
+# With poetry: `poetry install --sync --extras "stretch"`
+
+# See [tutorial](https://github.com/huggingface/lerobot/blob/main/examples/8_use_stretch.md)
+
+
+_target_: lerobot.common.robot_devices.robots.stretch.StretchRobot
+robot_type: stretch3
+
+cameras:
+  navigation:
+    _target_: lerobot.common.robot_devices.cameras.opencv.OpenCVCamera
+    camera_index: /dev/hello-nav-head-camera
+    fps: 10
+    width: 1280
+    height: 720
+    rotation: -90
+  head:
+    _target_: lerobot.common.robot_devices.cameras.intelrealsense.IntelRealSenseCamera.init_from_name
+    name: Intel RealSense D435I
+    fps: 30
+    width: 640
+    height: 480
+    rotation: 90
+  wrist:
+    _target_: lerobot.common.robot_devices.cameras.intelrealsense.IntelRealSenseCamera.init_from_name
+    name: Intel RealSense D405
+    fps: 30
+    width: 640
+    height: 480

lerobot/scripts/configure_motor.py

@@ -0,0 +1,145 @@
+"""
+This script configure a single motor at a time to a given ID and baudrate.
+
+Example of usage:
+```bash
+python lerobot/scripts/configure_motor.py \
+  --port /dev/tty.usbmodem585A0080521 \
+  --brand feetech \
+  --model sts3215 \
+  --baudrate 1000000 \
+  --ID 1
+```
+"""
+
+import argparse
+import time
+
+
+def configure_motor(port, brand, model, motor_idx_des, baudrate_des):
+    if brand == "feetech":
+        from lerobot.common.robot_devices.motors.feetech import MODEL_BAUDRATE_TABLE
+        from lerobot.common.robot_devices.motors.feetech import (
+            SCS_SERIES_BAUDRATE_TABLE as SERIES_BAUDRATE_TABLE,
+        )
+        from lerobot.common.robot_devices.motors.feetech import FeetechMotorsBus as MotorsBusClass
+    elif brand == "dynamixel":
+        from lerobot.common.robot_devices.motors.dynamixel import MODEL_BAUDRATE_TABLE
+        from lerobot.common.robot_devices.motors.dynamixel import (
+            X_SERIES_BAUDRATE_TABLE as SERIES_BAUDRATE_TABLE,
+        )
+        from lerobot.common.robot_devices.motors.dynamixel import DynamixelMotorsBus as MotorsBusClass
+    else:
+        raise ValueError(
+            f"Currently we do not support this motor brand: {brand}. We currently support feetech and dynamixel motors."
+        )
+
+    # Check if the provided model exists in the model_baud_rate_table
+    if model not in MODEL_BAUDRATE_TABLE:
+        raise ValueError(
+            f"Invalid model '{model}' for brand '{brand}'. Supported models: {list(MODEL_BAUDRATE_TABLE.keys())}"
+        )
+
+    # Setup motor names, indices, and models
+    motor_name = "motor"
+    motor_index_arbitrary = motor_idx_des  # Use the motor ID passed via argument
+    motor_model = model  # Use the motor model passed via argument
+
+    # Initialize the MotorBus with the correct port and motor configurations
+    motor_bus = MotorsBusClass(port=port, motors={motor_name: (motor_index_arbitrary, motor_model)})
+
+    # Try to connect to the motor bus and handle any connection-specific errors
+    try:
+        motor_bus.connect()
+        print(f"Connected on port {motor_bus.port}")
+    except OSError as e:
+        print(f"Error occurred when connecting to the motor bus: {e}")
+        return
+
+    # Motor bus is connected, proceed with the rest of the operations
+    try:
+        print("Scanning all baudrates and motor indices")
+        all_baudrates = set(SERIES_BAUDRATE_TABLE.values())
+        motor_index = -1  # Set the motor index to an out-of-range value.
+
+        for baudrate in all_baudrates:
+            motor_bus.set_bus_baudrate(baudrate)
+            present_ids = motor_bus.find_motor_indices(list(range(1, 10)))
+            if len(present_ids) > 1:
+                raise ValueError(
+                    "Error: More than one motor ID detected. This script is designed to only handle one motor at a time. Please disconnect all but one motor."
+                )
+
+            if len(present_ids) == 1:
+                if motor_index != -1:
+                    raise ValueError(
+                        "Error: More than one motor ID detected. This script is designed to only handle one motor at a time. Please disconnect all but one motor."
+                    )
+                motor_index = present_ids[0]
+
+        if motor_index == -1:
+            raise ValueError("No motors detected. Please ensure you have one motor connected.")
+
+        print(f"Motor index found at: {motor_index}")
+
+        if brand == "feetech":
+            # Allows ID and BAUDRATE to be written in memory
+            motor_bus.write_with_motor_ids(motor_bus.motor_models, motor_index, "Lock", 0)
+
+        if baudrate != baudrate_des:
+            print(f"Setting its baudrate to {baudrate_des}")
+            baudrate_idx = list(SERIES_BAUDRATE_TABLE.values()).index(baudrate_des)
+
+            # The write can fail, so we allow retries
+            motor_bus.write_with_motor_ids(motor_bus.motor_models, motor_index, "Baud_Rate", baudrate_idx)
+            time.sleep(0.5)
+            motor_bus.set_bus_baudrate(baudrate_des)
+            present_baudrate_idx = motor_bus.read_with_motor_ids(
+                motor_bus.motor_models, motor_index, "Baud_Rate", num_retry=2
+            )
+
+            if present_baudrate_idx != baudrate_idx:
+                raise OSError("Failed to write baudrate.")
+
+        print(f"Setting its index to desired index {motor_idx_des}")
+        motor_bus.write_with_motor_ids(motor_bus.motor_models, motor_index, "Lock", 0)
+        motor_bus.write_with_motor_ids(motor_bus.motor_models, motor_index, "ID", motor_idx_des)
+
+        present_idx = motor_bus.read_with_motor_ids(motor_bus.motor_models, motor_idx_des, "ID", num_retry=2)
+        if present_idx != motor_idx_des:
+            raise OSError("Failed to write index.")
+
+        if brand == "feetech":
+            # Set Maximum_Acceleration to 254 to speedup acceleration and deceleration of
+            # the motors. Note: this configuration is not in the official STS3215 Memory Table
+            motor_bus.write("Lock", 0)
+            motor_bus.write("Maximum_Acceleration", 254)
+
+            motor_bus.write("Goal_Position", 2048)
+            time.sleep(4)
+            print("Present Position", motor_bus.read("Present_Position"))
+
+            motor_bus.write("Offset", 0)
+            time.sleep(4)
+            print("Offset", motor_bus.read("Offset"))
+
+    except Exception as e:
+        print(f"Error occurred during motor configuration: {e}")
+
+    finally:
+        motor_bus.disconnect()
+        print("Disconnected from motor bus.")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--port", type=str, required=True, help="Motors bus port (e.g. dynamixel,feetech)")
+    parser.add_argument("--brand", type=str, required=True, help="Motor brand (e.g. dynamixel,feetech)")
+    parser.add_argument("--model", type=str, required=True, help="Motor model (e.g. xl330-m077,sts3215)")
+    parser.add_argument("--ID", type=int, required=True, help="Desired ID of the current motor (e.g. 1,2,3)")
+    parser.add_argument(
+        "--baudrate", type=int, default=1000000, help="Desired baudrate for the motor (default: 1000000)"
+    )
+    args = parser.parse_args()
+
+    configure_motor(args.port, args.brand, args.model, args.ID, args.baudrate)

lerobot/scripts/control_robot.py

@@ -99,160 +99,70 @@ python lerobot/scripts/control_robot.py record \
 """
 
 import argparse
-import concurrent.futures
-import json
 import logging
-import os
-import platform
-import shutil
 import time
-import traceback
-from contextlib import nullcontext
-from functools import cache
 from pathlib import Path
-
-import cv2
-import torch
-import tqdm
-from omegaconf import DictConfig
-from PIL import Image
-from termcolor import colored
+from typing import List
 
 # from safetensors.torch import load_file, save_file
-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.policies.factory import make_policy
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.common.robot_devices.control_utils import (
+    control_loop,
+    has_method,
+    init_keyboard_listener,
+    init_policy,
+    log_control_info,
+    record_episode,
+    reset_environment,
+    sanity_check_dataset_name,
+    stop_recording,
+    warmup_record,
+)
 from lerobot.common.robot_devices.robots.factory import make_robot
 from lerobot.common.robot_devices.robots.utils import Robot
-from lerobot.common.utils.utils import get_safe_torch_device, init_hydra_config, init_logging, set_global_seed
-from lerobot.scripts.eval import get_pretrained_policy_path
-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,
-)
-
-########################################################################################
-# Utilities
-########################################################################################
-
-
-def say(text, blocking=False):
-    # Check if mac, linux, or windows.
-    if platform.system() == "Darwin":
-        cmd = f'say "{text}"'
-    elif platform.system() == "Linux":
-        cmd = f'spd-say "{text}"'
-    elif platform.system() == "Windows":
-        cmd = (
-            'PowerShell -Command "Add-Type -AssemblyName System.Speech; '
-            f"(New-Object System.Speech.Synthesis.SpeechSynthesizer).Speak('{text}')\""
-        )
-
-    if not blocking and platform.system() in ["Darwin", "Linux"]:
-        # TODO(rcadene): Make it work for Windows
-        # Use the ampersand to run command in the background
-        cmd += " &"
-
-    os.system(cmd)
-
-
-def save_image(img_tensor, key, frame_index, episode_index, videos_dir):
-    img = Image.fromarray(img_tensor.numpy())
-    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 busy_wait(seconds):
-    # Significantly more accurate than `time.sleep`, and mendatory for our use case,
-    # but it consumes CPU cycles.
-    # TODO(rcadene): find an alternative: from python 11, time.sleep is precise
-    end_time = time.perf_counter() + seconds
-    while time.perf_counter() < end_time:
-        pass
-
-
-def none_or_int(value):
-    if value == "None":
-        return None
-    return int(value)
-
-
-def log_control_info(robot, dt_s, episode_index=None, frame_index=None, fps=None):
-    log_items = []
-    if episode_index is not None:
-        log_items += [f"ep:{episode_index}"]
-    if frame_index is not None:
-        log_items += [f"frame:{frame_index}"]
-
-    def log_dt(shortname, dt_val_s):
-        nonlocal log_items
-        log_items += [f"{shortname}:{dt_val_s * 1000:5.2f} ({1/ dt_val_s:3.1f}hz)"]
-
-    # total step time displayed in milliseconds and its frequency
-    log_dt("dt", dt_s)
-
-    for name in robot.leader_arms:
-        key = f"read_leader_{name}_pos_dt_s"
-        if key in robot.logs:
-            log_dt("dtRlead", robot.logs[key])
-
-    for name in robot.follower_arms:
-        key = f"write_follower_{name}_goal_pos_dt_s"
-        if key in robot.logs:
-            log_dt("dtWfoll", robot.logs[key])
-
-        key = f"read_follower_{name}_pos_dt_s"
-        if key in robot.logs:
-            log_dt("dtRfoll", robot.logs[key])
-
-    for name in robot.cameras:
-        key = f"read_camera_{name}_dt_s"
-        if key in robot.logs:
-            log_dt(f"dtR{name}", robot.logs[key])
-
-    info_str = " ".join(log_items)
-    if fps is not None:
-        actual_fps = 1 / dt_s
-        if actual_fps < fps - 1:
-            info_str = colored(info_str, "yellow")
-    logging.info(info_str)
-
-
-@cache
-def is_headless():
-    """Detects if python is running without a monitor."""
-    try:
-        import pynput  # noqa
-
-        return False
-    except Exception:
-        print(
-            "Error trying to import pynput. Switching to headless mode. "
-            "As a result, the video stream from the cameras won't be shown, "
-            "and you won't be able to change the control flow with keyboards. "
-            "For more info, see traceback below.\n"
-        )
-        traceback.print_exc()
-        print()
-        return True
-
+from lerobot.common.robot_devices.utils import busy_wait, safe_disconnect
+from lerobot.common.utils.utils import init_hydra_config, init_logging, log_say, none_or_int
 
 ########################################################################################
 # Control modes
 ########################################################################################
 
 
-def calibrate(robot: Robot):
-    if robot.calibration_path.exists():
-        print(f"Removing '{robot.calibration_path}'")
-        robot.calibration_path.unlink()
+@safe_disconnect
+def calibrate(robot: Robot, arms: list[str] | None):
+    # TODO(aliberts): move this code in robots' classes
+    if robot.robot_type.startswith("stretch"):
+        if not robot.is_connected:
+            robot.connect()
+        if not robot.is_homed():
+            robot.home()
+        return
+
+    if arms is None:
+        arms = robot.available_arms
+
+    unknown_arms = [arm_id for arm_id in arms if arm_id not in robot.available_arms]
+    available_arms_str = " ".join(robot.available_arms)
+    unknown_arms_str = " ".join(unknown_arms)
+
+    if arms is None or len(arms) == 0:
+        raise ValueError(
+            "No arm provided. Use `--arms` as argument with one or more available arms.\n"
+            f"For instance, to recalibrate all arms add: `--arms {available_arms_str}`"
+        )
+
+    if len(unknown_arms) > 0:
+        raise ValueError(
+            f"Unknown arms provided ('{unknown_arms_str}'). Available arms are `{available_arms_str}`."
+        )
+
+    for arm_id in arms:
+        arm_calib_path = robot.calibration_dir / f"{arm_id}.json"
+        if arm_calib_path.exists():
+            print(f"Removing '{arm_calib_path}'")
+            arm_calib_path.unlink()
+        else:
+            print(f"Calibration file not found '{arm_calib_path}'")
 
     if robot.is_connected:
         robot.disconnect()
@@ -260,429 +170,189 @@ def calibrate(robot: Robot):
     # Calling `connect` automatically runs calibration
     # when the calibration file is missing
     robot.connect()
+    robot.disconnect()
+    print("Calibration is done! You can now teleoperate and record datasets!")
 
 
-def teleoperate(robot: Robot, fps: int | None = None, teleop_time_s: float | None = None):
-    # TODO(rcadene): Add option to record logs
-    if not robot.is_connected:
-        robot.connect()
-
-    start_teleop_t = time.perf_counter()
-    while True:
-        start_loop_t = time.perf_counter()
-        robot.teleop_step()
-
-        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)
-
-        if teleop_time_s is not None and time.perf_counter() - start_teleop_t > teleop_time_s:
-            break
+@safe_disconnect
+def teleoperate(
+    robot: Robot, fps: int | None = None, teleop_time_s: float | None = None, display_cameras: bool = False
+):
+    control_loop(
+        robot,
+        control_time_s=teleop_time_s,
+        fps=fps,
+        teleoperate=True,
+        display_cameras=display_cameras,
+    )
 
 
+@safe_disconnect
 def record(
     robot: Robot,
-    policy: torch.nn.Module | None = None,
-    hydra_cfg: DictConfig | None = None,
+    root: str,
+    repo_id: str,
+    single_task: str,
+    pretrained_policy_name_or_path: str | None = None,
+    policy_overrides: List[str] | None = None,
     fps: int | None = None,
-    root="data",
-    repo_id="lerobot/debug",
-    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_writers=8,
-    force_override=False,
-):
+    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,
+    num_image_writer_processes: int = 0,
+    num_image_writer_threads_per_camera: int = 4,
+    display_cameras: bool = True,
+    play_sounds: bool = True,
+    tags: str = None,
+    force_override: bool = False,
+) -> LeRobotDataset:
     # TODO(rcadene): Add option to record logs
-    # TODO(rcadene): Clean this function via decomposition in higher level functions
+    listener = None
+    events = None
+    policy = None
+    device = None
+    use_amp = None
 
-    _, dataset_name = repo_id.split("/")
-    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})."
-        )
+    if single_task:
+        task = single_task
+    else:
+        raise NotImplementedError("Only single-task recording is supported for now")
 
-    if not video:
-        raise NotImplementedError()
+    # Load pretrained policy
+    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}).")
+        elif fps != policy_fps:
+            logging.warning(
+                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 = LeRobotDataset.create(
+        repo_id,
+        fps,
+        root=root,
+        robot=robot,
+        image_writer_processes=num_image_writer_processes,
+        image_writer_threads_per_camera=num_image_writer_threads_per_camera,
+        use_videos=video,
+    )
 
     if not robot.is_connected:
         robot.connect()
 
-    local_dir = Path(root) / repo_id
-    if local_dir.exists() and force_override:
-        shutil.rmtree(local_dir)
+    listener, events = init_keyboard_listener()
 
-    episodes_dir = local_dir / "episodes"
-    episodes_dir.mkdir(parents=True, exist_ok=True)
+    # 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,
+    # 3. place the cameras windows on screen
+    enable_teleoperation = policy is None
+    log_say("Warmup record", play_sounds)
+    warmup_record(robot, events, enable_teleoperation, warmup_time_s, display_cameras, fps)
 
-    videos_dir = local_dir / "videos"
-    videos_dir.mkdir(parents=True, exist_ok=True)
+    if has_method(robot, "teleop_safety_stop"):
+        robot.teleop_safety_stop()
 
-    # 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)
-        episode_index = rec_info["last_episode_index"] + 1
-    else:
-        episode_index = 0
+    recorded_episodes = 0
+    while True:
+        if recorded_episodes >= num_episodes:
+            break
 
-    if is_headless():
-        logging.info(
-            "Headless environment detected. On-screen cameras display and keyboard inputs will not be available."
+        # 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: ")
+
+        episode_index = dataset.episode_buffer["episode_index"]
+        log_say(f"Recording episode {episode_index}", play_sounds)
+        record_episode(
+            dataset=dataset,
+            robot=robot,
+            events=events,
+            episode_time_s=episode_time_s,
+            display_cameras=display_cameras,
+            policy=policy,
+            device=device,
+            use_amp=use_amp,
+            fps=fps,
         )
 
-    # 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.
-    exit_early = False
-    rerecord_episode = False
-    stop_recording = False
+        # Execute a few seconds without recording to give time to manually reset the environment
+        # Current code logic doesn't allow to teleoperate during this time.
+        # 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"]
+        ):
+            log_say("Reset the environment", play_sounds)
+            reset_environment(robot, events, reset_time_s)
 
-    # Only import pynput if not in a headless environment
-    if not is_headless():
-        from pynput import keyboard
+        if events["rerecord_episode"]:
+            log_say("Re-record episode", play_sounds)
+            events["rerecord_episode"] = False
+            events["exit_early"] = False
+            dataset.clear_episode_buffer()
+            continue
 
-        def on_press(key):
-            nonlocal exit_early, rerecord_episode, stop_recording
-            try:
-                if key == keyboard.Key.right:
-                    print("Right arrow key pressed. Exiting loop...")
-                    exit_early = True
-                elif key == keyboard.Key.left:
-                    print("Left arrow key pressed. Exiting loop and rerecord the last episode...")
-                    rerecord_episode = True
-                    exit_early = True
-                elif key == keyboard.Key.esc:
-                    print("Escape key pressed. Stopping data recording...")
-                    stop_recording = True
-                    exit_early = True
-            except Exception as e:
-                print(f"Error handling key press: {e}")
+        dataset.add_episode(task)
+        recorded_episodes += 1
 
-        listener = keyboard.Listener(on_press=on_press)
-        listener.start()
+        if events["stop_recording"]:
+            break
 
-    # Load policy if any
-    if policy is not None:
-        # Check device is available
-        device = get_safe_torch_device(hydra_cfg.device, log=True)
+    log_say("Stop recording", play_sounds, blocking=True)
+    stop_recording(robot, listener, display_cameras)
 
-        policy.eval()
-        policy.to(device)
+    if dataset.image_writer is not None:
+        logging.info("Waiting for image writer to terminate...")
+        dataset.image_writer.stop()
 
-        torch.backends.cudnn.benchmark = True
-        torch.backends.cuda.matmul.allow_tf32 = True
-        set_global_seed(hydra_cfg.seed)
-
-        # override fps using policy fps
-        fps = hydra_cfg.env.fps
-
-    # Execute a few seconds without recording data, to give times
-    # to the robot devices to connect and start synchronizing.
-    timestamp = 0
-    start_warmup_t = time.perf_counter()
-    is_warmup_print = False
-    while timestamp < warmup_time_s:
-        if not is_warmup_print:
-            logging.info("Warming up (no data recording)")
-            say("Warming up")
-            is_warmup_print = True
-
-        start_loop_t = time.perf_counter()
-
-        if policy is None:
-            observation, action = robot.teleop_step(record_data=True)
-        else:
-            observation = robot.capture_observation()
-
-        if not is_headless():
-            image_keys = [key for key in observation if "image" in key]
-            for key in image_keys:
-                cv2.imshow(key, cv2.cvtColor(observation[key].numpy(), cv2.COLOR_RGB2BGR))
-            cv2.waitKey(1)
-
-        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_warmup_t
-
-    # Save images using threads to reach high fps (30 and more)
-    # Using `with` to exist smoothly if an execption is raised.
-    # Using only 4 worker threads to avoid blocking the main thread.
-    futures = []
-    with concurrent.futures.ThreadPoolExecutor(max_workers=num_image_writers) as executor:
-        # Start recording all episodes
-        while episode_index < num_episodes:
-            logging.info(f"Recording episode {episode_index}")
-            say(f"Recording episode {episode_index}")
-            ep_dict = {}
-            frame_index = 0
-            timestamp = 0
-            start_episode_t = time.perf_counter()
-            while timestamp < episode_time_s:
-                start_loop_t = time.perf_counter()
-
-                if policy is None:
-                    observation, action = robot.teleop_step(record_data=True)
-                else:
-                    observation = robot.capture_observation()
-
-                image_keys = [key for key in observation if "image" in key]
-                not_image_keys = [key for key in observation if "image" not in key]
-
-                for key in image_keys:
-                    futures += [
-                        executor.submit(
-                            save_image, observation[key], key, frame_index, episode_index, videos_dir
-                        )
-                    ]
-
-                if not is_headless():
-                    image_keys = [key for key in observation if "image" in key]
-                    for key in image_keys:
-                        cv2.imshow(key, cv2.cvtColor(observation[key].numpy(), cv2.COLOR_RGB2BGR))
-                    cv2.waitKey(1)
-
-                for key in not_image_keys:
-                    if key not in ep_dict:
-                        ep_dict[key] = []
-                    ep_dict[key].append(observation[key])
-
-                if policy is not None:
-                    with (
-                        torch.inference_mode(),
-                        torch.autocast(device_type=device.type)
-                        if device.type == "cuda" and hydra_cfg.use_amp
-                        else nullcontext(),
-                    ):
-                        # Convert to pytorch format: channel first and float32 in [0,1] with batch dimension
-                        for name in observation:
-                            if "image" in name:
-                                observation[name] = observation[name].type(torch.float32) / 255
-                                observation[name] = observation[name].permute(2, 0, 1).contiguous()
-                            observation[name] = observation[name].unsqueeze(0)
-                            observation[name] = observation[name].to(device)
-
-                        # Compute the next action with the policy
-                        # based on the current observation
-                        action = policy.select_action(observation)
-
-                        # Remove batch dimension
-                        action = action.squeeze(0)
-
-                        # Move to cpu, if not already the case
-                        action = action.to("cpu")
-
-                    # Order the robot to move
-                    robot.send_action(action)
-                    action = {"action": action}
-
-                for key in action:
-                    if key not in ep_dict:
-                        ep_dict[key] = []
-                    ep_dict[key].append(action[key])
-
-                frame_index += 1
-
-                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 exit_early:
-                    exit_early = False
-                    break
-
-            if not stop_recording:
-                # Start resetting env while the executor are finishing
-                logging.info("Reset the environment")
-                say("Reset the environment")
-
-            timestamp = 0
-            start_vencod_t = time.perf_counter()
-
-            # During env reset we save the data and encode the videos
-            num_frames = frame_index
-
-            for key in image_keys:
-                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():
-                    video_path.unlink()
-                # Store the reference to the video frame, even tho the videos are not yet encoded
-                ep_dict[key] = []
-                for i in range(num_frames):
-                    ep_dict[key].append({"path": f"videos/{fname}", "timestamp": i / fps})
-
-            for key in not_image_keys:
-                ep_dict[key] = torch.stack(ep_dict[key])
-
-            for key in action:
-                ep_dict[key] = torch.stack(ep_dict[key])
-
-            ep_dict["episode_index"] = torch.tensor([episode_index] * num_frames)
-            ep_dict["frame_index"] = torch.arange(0, num_frames, 1)
-            ep_dict["timestamp"] = torch.arange(0, num_frames, 1) / fps
-
-            done = torch.zeros(num_frames, dtype=torch.bool)
-            done[-1] = True
-            ep_dict["next.done"] = done
-
-            ep_path = episodes_dir / f"episode_{episode_index}.pth"
-            print("Saving episode dictionary...")
-            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)
-
-            is_last_episode = stop_recording or (episode_index == (num_episodes - 1))
-
-            # Wait if necessary
-            with tqdm.tqdm(total=reset_time_s, desc="Waiting") as pbar:
-                while timestamp < reset_time_s and not is_last_episode:
-                    time.sleep(1)
-                    timestamp = time.perf_counter() - start_vencod_t
-                    pbar.update(1)
-                    if exit_early:
-                        exit_early = False
-                        break
-
-            # Skip updating episode index which forces re-recording episode
-            if rerecord_episode:
-                rerecord_episode = False
-                continue
-
-            episode_index += 1
-
-            if is_last_episode:
-                logging.info("Done recording")
-                say("Done recording", blocking=True)
-                if not is_headless():
-                    listener.stop()
-
-                logging.info("Waiting for threads writing the images on disk to terminate...")
-                for _ in tqdm.tqdm(
-                    concurrent.futures.as_completed(futures), total=len(futures), desc="Writting images"
-                ):
-                    pass
-                break
-
-    robot.disconnect()
-    if not is_headless():
-        cv2.destroyAllWindows()
-
-    num_episodes = episode_index
-
-    logging.info("Encoding videos")
-    say("Encoding videos")
-    # 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:
-            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)
-
-    logging.info("Concatenating episodes")
-    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)
-
-    total_frames = data_dict["frame_index"].shape[0]
-    data_dict["index"] = torch.arange(0, total_frames, 1)
-
-    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,
-    )
     if run_compute_stats:
         logging.info("Computing dataset statistics")
-        say("Computing dataset statistics")
-        stats = compute_stats(lerobot_dataset)
-        lerobot_dataset.stats = stats
-    else:
-        stats = {}
-        logging.info("Skipping computation of the dataset statistics")
 
-    hf_dataset = hf_dataset.with_format(None)  # to remove transforms that cant be saved
-    hf_dataset.save_to_disk(str(local_dir / "train"))
-
-    meta_data_dir = local_dir / "meta_data"
-    save_meta_data(info, stats, episode_data_index, meta_data_dir)
+    dataset.consolidate(run_compute_stats)
 
+    # lerobot_dataset = create_lerobot_dataset(dataset, run_compute_stats, push_to_hub, tags, play_sounds)
     if push_to_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)
+        dataset.push_to_hub()
 
-    logging.info("Exiting")
-    say("Exiting")
-    return lerobot_dataset
+    log_say("Exiting", play_sounds)
+    return dataset
 
 
-def replay(robot: Robot, episode: int, fps: int | None = None, root="data", repo_id="lerobot/debug"):
+@safe_disconnect
+def replay(
+    robot: Robot,
+    root: Path,
+    repo_id: str,
+    episode: int,
+    fps: int | None = None,
+    play_sounds: bool = True,
+    local_files_only: bool = True,
+):
+    # 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()
 
-    logging.info("Replaying episode")
-    say("Replaying episode", blocking=True)
-    for idx in range(from_idx, to_idx):
+    log_say("Replaying episode", play_sounds, blocking=True)
+    for idx in range(dataset.num_samples):
         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
@@ -712,16 +382,40 @@ if __name__ == "__main__":
     )
 
     parser_calib = subparsers.add_parser("calibrate", parents=[base_parser])
+    parser_calib.add_argument(
+        "--arms",
+        type=str,
+        nargs="*",
+        help="List of arms to calibrate (e.g. `--arms left_follower right_follower left_leader`)",
+    )
 
     parser_teleop = subparsers.add_parser("teleoperate", parents=[base_parser])
     parser_teleop.add_argument(
         "--fps", type=none_or_int, default=None, help="Frames per second (set to None to disable)"
     )
+    parser_teleop.add_argument(
+        "--display-cameras",
+        type=int,
+        default=1,
+        help="Display all cameras on screen (set to 1 to display or 0).",
+    )
 
     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,
@@ -772,10 +466,25 @@ if __name__ == "__main__":
         help="Add tags to your dataset on the hub.",
     )
     parser_record.add_argument(
-        "--num-image-writers",
+        "--num-image-writer-processes",
         type=int,
-        default=8,
-        help="Number of threads writing the frames as png images on disk. Don't set too much as you might get unstable fps due to main thread being blocked.",
+        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 many threads might cause unstable teleoperation fps due to main thread being blocked. "
+            "Not enough threads might cause low camera fps."
+        ),
     )
     parser_record.add_argument(
         "--force-override",
@@ -839,19 +548,7 @@ if __name__ == "__main__":
         teleoperate(robot, **kwargs)
 
     elif control_mode == "record":
-        pretrained_policy_name_or_path = args.pretrained_policy_name_or_path
-        policy_overrides = args.policy_overrides
-        del kwargs["pretrained_policy_name_or_path"]
-        del kwargs["policy_overrides"]
-
-        policy_cfg = None
-        if pretrained_policy_name_or_path is not None:
-            pretrained_policy_path = get_pretrained_policy_path(pretrained_policy_name_or_path)
-            policy_cfg = init_hydra_config(pretrained_policy_path / "config.yaml", policy_overrides)
-            policy = make_policy(hydra_cfg=policy_cfg, pretrained_policy_name_or_path=pretrained_policy_path)
-            record(robot, policy, policy_cfg, **kwargs)
-        else:
-            record(robot, **kwargs)
+        record(robot, **kwargs)
 
     elif control_mode == "replay":
         replay(robot, **kwargs)

lerobot/scripts/eval.py

@@ -57,7 +57,7 @@ import gymnasium as gym
 import numpy as np
 import torch
 from huggingface_hub import snapshot_download
-from huggingface_hub.utils._errors import RepositoryNotFoundError
+from huggingface_hub.errors import RepositoryNotFoundError
 from huggingface_hub.utils._validators import HFValidationError
 from torch import Tensor, nn
 from tqdm import trange
@@ -70,7 +70,13 @@ from lerobot.common.policies.factory import make_policy
 from lerobot.common.policies.policy_protocol import Policy
 from lerobot.common.policies.utils import get_device_from_parameters
 from lerobot.common.utils.io_utils import write_video
-from lerobot.common.utils.utils import get_safe_torch_device, init_hydra_config, init_logging, set_global_seed
+from lerobot.common.utils.utils import (
+    get_safe_torch_device,
+    init_hydra_config,
+    init_logging,
+    inside_slurm,
+    set_global_seed,
+)
 
 
 def rollout(
@@ -79,7 +85,6 @@ def rollout(
     seeds: list[int] | None = None,
     return_observations: bool = False,
     render_callback: Callable[[gym.vector.VectorEnv], None] | None = None,
-    enable_progbar: bool = False,
 ) -> dict:
     """Run a batched policy rollout once through a batch of environments.
 
@@ -109,7 +114,6 @@ def rollout(
             are returned optionally because they typically take more memory to cache. Defaults to False.
         render_callback: Optional rendering callback to be used after the environments are reset, and after
             every step.
-        enable_progbar: Enable a progress bar over rollout steps.
     Returns:
         The dictionary described above.
     """
@@ -136,7 +140,7 @@ def rollout(
     progbar = trange(
         max_steps,
         desc=f"Running rollout with at most {max_steps} steps",
-        disable=not enable_progbar,
+        disable=inside_slurm(),  # we dont want progress bar when we use slurm, since it clutters the logs
         leave=False,
     )
     while not np.all(done):
@@ -210,8 +214,6 @@ def eval_policy(
     videos_dir: Path | None = None,
     return_episode_data: bool = False,
     start_seed: int | None = None,
-    enable_progbar: bool = False,
-    enable_inner_progbar: bool = False,
 ) -> dict:
     """
     Args:
@@ -224,8 +226,6 @@ def eval_policy(
             the "episodes" key of the returned dictionary.
         start_seed: The first seed to use for the first individual rollout. For all subsequent rollouts the
             seed is incremented by 1. If not provided, the environments are not manually seeded.
-        enable_progbar: Enable progress bar over batches.
-        enable_inner_progbar: Enable progress bar over steps in each batch.
     Returns:
         Dictionary with metrics and data regarding the rollouts.
     """
@@ -266,7 +266,8 @@ def eval_policy(
     if return_episode_data:
         episode_data: dict | None = None
 
-    progbar = trange(n_batches, desc="Stepping through eval batches", disable=not enable_progbar)
+    # we dont want progress bar when we use slurm, since it clutters the logs
+    progbar = trange(n_batches, desc="Stepping through eval batches", disable=inside_slurm())
     for batch_ix in progbar:
         # Cache frames for rendering videos. Each item will be (b, h, w, c), and the list indexes the rollout
         # step.
@@ -285,7 +286,6 @@ def eval_policy(
             seeds=list(seeds) if seeds else None,
             return_observations=return_episode_data,
             render_callback=render_frame if max_episodes_rendered > 0 else None,
-            enable_progbar=enable_inner_progbar,
         )
 
         # Figure out where in each rollout sequence the first done condition was encountered (results after
@@ -454,6 +454,16 @@ def main(
     else:
         hydra_cfg = init_hydra_config(hydra_cfg_path, config_overrides)
 
+    if hydra_cfg.eval.batch_size > hydra_cfg.eval.n_episodes:
+        raise ValueError(
+            "The eval batch size is greater than the number of eval episodes "
+            f"({hydra_cfg.eval.batch_size} > {hydra_cfg.eval.n_episodes}). As a result, {hydra_cfg.eval.batch_size} "
+            f"eval environments will be instantiated, but only {hydra_cfg.eval.n_episodes} will be used. "
+            "This might significantly slow down evaluation. To fix this, you should update your command "
+            f"to increase the number of episodes to match the batch size (e.g. `eval.n_episodes={hydra_cfg.eval.batch_size}`), "
+            f"or lower the batch size (e.g. `eval.batch_size={hydra_cfg.eval.n_episodes}`)."
+        )
+
     if out_dir is None:
         out_dir = f"outputs/eval/{dt.now().strftime('%Y-%m-%d/%H-%M-%S')}_{hydra_cfg.env.name}_{hydra_cfg.policy.name}"
 
@@ -487,8 +497,6 @@ def main(
             max_episodes_rendered=10,
             videos_dir=Path(out_dir) / "videos",
             start_seed=hydra_cfg.seed,
-            enable_progbar=True,
-            enable_inner_progbar=True,
         )
     print(info["aggregated"])
 

lerobot/scripts/find_motors_bus_port.py

@@ -0,0 +1,36 @@
+import time
+from pathlib import Path
+
+
+def find_available_ports():
+    ports = []
+    for path in Path("/dev").glob("tty*"):
+        ports.append(str(path))
+    return ports
+
+
+def find_port():
+    print("Finding all available ports for the MotorsBus.")
+    ports_before = find_available_ports()
+    print(ports_before)
+
+    print("Remove the usb cable from your MotorsBus and press Enter when done.")
+    input()
+
+    time.sleep(0.5)
+    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.")
+    elif len(ports_diff) == 0:
+        raise OSError(f"Could not detect the port. No difference was found ({ports_diff}).")
+    else:
+        raise OSError(f"Could not detect the port. More than one port was found ({ports_diff}).")
+
+
+if __name__ == "__main__":
+    # Helper to find the usb port associated to all your MotorsBus.
+    find_port()

lerobot/scripts/push_dataset_to_hub.py

@@ -66,6 +66,8 @@ 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:
+        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
     elif raw_format == "xarm_pkl":
@@ -197,9 +199,25 @@ 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)
-    hf_dataset, episode_data_index, info = from_raw_to_lerobot_format(
-        raw_dir, videos_dir, fps, video, episodes, encoding
-    )
+
+    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)
 
     lerobot_dataset = LeRobotDataset.from_preloaded(
         repo_id=repo_id,
@@ -242,7 +260,7 @@ 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)
 
@@ -268,7 +286,7 @@ def main():
         "--raw-format",
         type=str,
         required=True,
-        help="Dataset type (e.g. `pusht_zarr`, `umi_zarr`, `aloha_hdf5`, `xarm_pkl`, `dora_parquet`).",
+        help="Dataset type (e.g. `pusht_zarr`, `umi_zarr`, `aloha_hdf5`, `xarm_pkl`, `dora_parquet`, `openx_rlds`).",
     )
     parser.add_argument(
         "--repo-id",
@@ -328,6 +346,13 @@ def main():
         default=0,
         help="When set to 1, resumes a previous run.",
     )
+    parser.add_argument(
+        "--cache-dir",
+        type=Path,
+        required=False,
+        default="/tmp",
+        help="Directory to store the temporary videos and images generated while creating the dataset.",
+    )
     parser.add_argument(
         "--tests-data-dir",
         type=Path,

lerobot/scripts/train.py

@@ -241,6 +241,7 @@ def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = No
         raise NotImplementedError()
 
     init_logging()
+    logging.info(pformat(OmegaConf.to_container(cfg)))
 
     if cfg.training.online_steps > 0 and isinstance(cfg.dataset_repo_id, ListConfig):
         raise NotImplementedError("Online training with LeRobotMultiDataset is not implemented.")
@@ -287,6 +288,16 @@ def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = No
             "you meant to resume training, please use `resume=true` in your command or yaml configuration."
         )
 
+    if cfg.eval.batch_size > cfg.eval.n_episodes:
+        raise ValueError(
+            "The eval batch size is greater than the number of eval episodes "
+            f"({cfg.eval.batch_size} > {cfg.eval.n_episodes}). As a result, {cfg.eval.batch_size} "
+            f"eval environments will be instantiated, but only {cfg.eval.n_episodes} will be used. "
+            "This might significantly slow down evaluation. To fix this, you should update your command "
+            f"to increase the number of episodes to match the batch size (e.g. `eval.n_episodes={cfg.eval.batch_size}`), "
+            f"or lower the batch size (e.g. `eval.batch_size={cfg.eval.n_episodes}`)."
+        )
+
     # log metrics to terminal and wandb
     logger = Logger(cfg, out_dir, wandb_job_name=job_name)
 
@@ -372,7 +383,7 @@ def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = No
             logging.info(f"Checkpoint policy after step {step}")
             # Note: Save with step as the identifier, and format it to have at least 6 digits but more if
             # needed (choose 6 as a minimum for consistency without being overkill).
-            logger.save_checkpont(
+            logger.save_checkpoint(
                 step,
                 policy,
                 optimizer,

lerobot/scripts/visualize_dataset_html.py

@@ -57,7 +57,6 @@ import logging
 import shutil
 from pathlib import Path
 
-import torch
 import tqdm
 from flask import Flask, redirect, render_template, url_for
 
@@ -65,19 +64,6 @@ from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.common.utils.utils import init_logging
 
 
-class EpisodeSampler(torch.utils.data.Sampler):
-    def __init__(self, dataset, episode_index):
-        from_idx = dataset.episode_data_index["from"][episode_index].item()
-        to_idx = dataset.episode_data_index["to"][episode_index].item()
-        self.frame_ids = range(from_idx, to_idx)
-
-    def __iter__(self):
-        return iter(self.frame_ids)
-
-    def __len__(self):
-        return len(self.frame_ids)
-
-
 def run_server(
     dataset: LeRobotDataset,
     episodes: list[int],
@@ -111,11 +97,14 @@ def run_server(
             "num_episodes": dataset.num_episodes,
             "fps": dataset.fps,
         }
-        video_paths = get_episode_video_paths(dataset, episode_id)
+        video_paths = [dataset.get_video_file_path(episode_id, key) for key in dataset.video_keys]
+        tasks = dataset.episode_dicts[episode_id]["tasks"]
         videos_info = [
-            {"url": url_for("static", filename=video_path), "filename": Path(video_path).name}
+            {"url": url_for("static", filename=video_path), "filename": video_path.name}
             for video_path in video_paths
         ]
+        videos_info[0]["language_instruction"] = tasks
+
         ep_csv_url = url_for("static", filename=get_ep_csv_fname(episode_id))
         return render_template(
             "visualize_dataset_template.html",
@@ -147,10 +136,10 @@ def write_episode_data_csv(output_dir, file_name, episode_index, dataset):
     # init header of csv with state and action names
     header = ["timestamp"]
     if has_state:
-        dim_state = len(dataset.hf_dataset["observation.state"][0])
+        dim_state = dataset.shapes["observation.state"]
         header += [f"state_{i}" for i in range(dim_state)]
     if has_action:
-        dim_action = len(dataset.hf_dataset["action"][0])
+        dim_action = dataset.shapes["action"]
         header += [f"action_{i}" for i in range(dim_action)]
 
     columns = ["timestamp"]
@@ -181,11 +170,24 @@ def get_episode_video_paths(dataset: LeRobotDataset, ep_index: int) -> list[str]
     # get first frame of episode (hack to get video_path of the episode)
     first_frame_idx = dataset.episode_data_index["from"][ep_index].item()
     return [
-        dataset.hf_dataset.select_columns(key)[first_frame_idx][key]["path"]
-        for key in dataset.video_frame_keys
+        dataset.hf_dataset.select_columns(key)[first_frame_idx][key]["path"] for key in dataset.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:
+        return None
+
+    # get first frame index
+    first_frame_idx = dataset.episode_data_index["from"][ep_index].item()
+
+    language_instruction = dataset.hf_dataset[first_frame_idx]["language_instruction"]
+    # TODO (michel-aractingi) hack to get the sentence, some strings in openx are badly stored
+    # with the tf.tensor appearing in the string
+    return language_instruction.removeprefix("tf.Tensor(b'").removesuffix("', shape=(), dtype=string)")
+
+
 def visualize_dataset_html(
     repo_id: str,
     root: Path | None = None,
@@ -200,8 +202,8 @@ def visualize_dataset_html(
 
     dataset = LeRobotDataset(repo_id, root=root)
 
-    if not dataset.video:
-        raise NotImplementedError(f"Image datasets ({dataset.video=}) are currently not supported.")
+    if len(dataset.image_keys) > 0:
+        raise NotImplementedError(f"Image keys ({dataset.image_keys=}) are currently not supported.")
 
     if output_dir is None:
         output_dir = f"outputs/visualize_dataset_html/{repo_id}"
@@ -221,7 +223,7 @@ def visualize_dataset_html(
     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())
+        ln_videos_dir.symlink_to((dataset.root / "videos").resolve())
 
     template_dir = Path(__file__).resolve().parent.parent / "templates"
 

lerobot/templates/visualize_dataset_template.html

@@ -14,7 +14,7 @@
 <!-- Use [Alpin.js](https://alpinejs.dev), a lightweight and easy to learn JS framework -->
 <!-- Use [tailwindcss](https://tailwindcss.com/), CSS classes for styling html -->
 <!-- Use [dygraphs](https://dygraphs.com/), a lightweight JS charting library -->
-<body class="flex h-screen max-h-screen bg-slate-950 text-gray-200" x-data="createAlpineData()" @keydown.window="(e) => {
+<body class="flex flex-col md:flex-row h-screen max-h-screen bg-slate-950 text-gray-200" x-data="createAlpineData()" @keydown.window="(e) => {
     // Use the space bar to play and pause, instead of default action (e.g. scrolling)
     const { keyCode, key } = e;
     if (keyCode === 32 || key === ' ') {
@@ -30,7 +30,7 @@
     }
 }">
     <!-- Sidebar -->
-    <div x-ref="sidebar" class="w-60 bg-slate-900 p-5 break-words max-h-screen overflow-y-auto">
+    <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>
 
         <ul>
@@ -46,7 +46,8 @@
         </ul>
 
         <p>Episodes:</p>
-        <ul class="ml-2">
+        <!-- episodes menu for medium & large screens -->
+        <ul class="ml-2 hidden md:block">
             {% for episode in episodes %}
             <li class="font-mono text-sm mt-0.5">
                 <a href="episode_{{ episode }}" class="underline {% if episode_id == episode %}font-bold -ml-1{% endif %}">
@@ -56,26 +57,47 @@
             {% endfor %}
         </ul>
 
+        <!-- episodes menu for small screens -->
+        <div class="flex overflow-x-auto md:hidden">
+            {% for episode in episodes %}
+            <p class="font-mono text-sm mt-0.5 border-r last:border-r-0 px-2 {% if episode_id == episode %}font-bold{% endif %}">
+                <a href="episode_{{ episode }}" class="">
+                    {{ episode }}
+                </a>
+            </p>
+            {% endfor %}
+        </div>
+
     </div>
 
     <!-- Toggle sidebar button -->
-    <button class="flex items-center opacity-50 hover:opacity-100 mx-1"
+    <button class="flex items-center opacity-50 hover:opacity-100 mx-1 hidden md:block"
         @click="() => ($refs.sidebar.classList.toggle('hidden'))" title="Toggle sidebar">
         <div class="bg-slate-500 w-2 h-10 rounded-full"></div>
     </button>
 
     <!-- Content -->
-    <div class="flex-1 max-h-screen flex flex-col gap-4 overflow-y-auto">
+    <div class="max-h-screen flex flex-col gap-4 overflow-y-auto md:flex-1">
         <h1 class="text-xl font-bold mt-4 font-mono">
             Episode {{ episode_id }}
         </h1>
 
+        <!-- Error message -->
+        <div class="font-medium text-orange-700 hidden" :class="{ 'hidden': !videoCodecError }">
+            <p>Videos could NOT play because <a href="https://en.wikipedia.org/wiki/AV1" target="_blank" class="underline">AV1</a> decoding is not available on your browser.</p>
+            <ul class="list-decimal list-inside">
+                <li>If iPhone: <span class="italic">It is supported with A17 chip or higher.</span></li>
+                <li>If Mac with Safari: <span class="italic">It is supported on most browsers except Safari with M1 chip or higher and on Safari with M3 chip or higher.</span></li>
+                <li>Other: <span class="italic">Contact the maintainers on LeRobot discord channel:</span> <a href="https://discord.com/invite/s3KuuzsPFb" target="_blank" class="underline">https://discord.com/invite/s3KuuzsPFb</a></li>
+            </ul>
+        </div>
+
         <!-- Videos -->
         <div class="flex flex-wrap gap-1">
             {% for video_info in videos_info %}
-            <div class="max-w-96">
+            <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>
-                <video autoplay muted loop type="video/mp4" class="min-w-64" @timeupdate="() => {
+                <video muted loop type="video/mp4" class="object-contain w-full h-full" @canplaythrough="videoCanPlay" @timeupdate="() => {
                     if (video.duration) {
                       const time = video.currentTime;
                       const pc = (100 / video.duration) * time;
@@ -100,6 +122,13 @@
             {% endfor %}
         </div>
 
+        <!-- Language instruction -->
+        {% if videos_info[0].language_instruction %}
+        <p class="font-medium mt-2">
+            Language Instruction: <span class="italic">{{ videos_info[0].language_instruction }}</span>
+        </p>
+        {% endif %}
+
         <!-- Shortcuts info -->
         <div class="text-sm hidden md:block">
             Hotkeys: <span class="font-mono">Space</span> to pause/unpause, <span class="font-mono">Arrow Down</span> to go to next episode, <span class="font-mono">Arrow Up</span> to go to previous episode.
@@ -107,12 +136,12 @@
 
         <!-- Controllers -->
         <div class="flex gap-1 text-3xl items-center">
-            <button x-ref="btnPlay" class="-rotate-90 hidden" class="-rotate-90" title="Play. Toggle with Space" @click="() => {
+            <button x-ref="btnPlay" class="-rotate-90" class="-rotate-90" title="Play. Toggle with Space" @click="() => {
                 videos.forEach(video => video.play());
                 $refs.btnPlay.classList.toggle('hidden');
                 $refs.btnPause.classList.toggle('hidden');
             }">🔽</button>
-            <button x-ref="btnPause" title="Pause. Toggle with Space" @click="() => {
+            <button x-ref="btnPause" class="hidden" title="Pause. Toggle with Space" @click="() => {
                 videos.forEach(video => video.pause());
                 $refs.btnPlay.classList.toggle('hidden');
                 $refs.btnPause.classList.toggle('hidden');
@@ -125,7 +154,6 @@
                 @click="() => (videos.forEach(video => (video.currentTime = 0.0)))">↩️</button>
             <input x-ref="slider" max="100" min="0" step="1" type="range" value="0" class="w-80 mx-2" @input="() => {
                 const sliderValue = $refs.slider.value;
-                $refs.btnPause.click();
                 videos.forEach(video => {
                     const time = (video.duration * sliderValue) / 100;
                     video.currentTime = time;
@@ -177,9 +205,9 @@
                             </td>
                             <template x-for="(cell, colIndex) in row">
                                 <td x-show="cell" class="border border-slate-700">
-                                    <div class="flex gap-x-2 w-24 justify-between px-2">
+                                    <div class="flex gap-x-2 w-24 justify-between px-2" :class="{ 'hidden': cell.isNull }">
                                         <input type="checkbox" x-model="cell.checked" @change="updateTableValues()">
-                                        <span x-text="`${cell.value.toFixed(2)}`"
+                                        <span x-text="`${!cell.isNull ? cell.value.toFixed(2) : null}`"
                                             :style="`color: ${cell.color}`"></span>
                                     </div>
                                 </td>
@@ -201,16 +229,29 @@
                 dygraph: null,
                 currentFrameData: null,
                 columnNames: ["state", "action", "pred action"],
-                nColumns: {% if has_policy %}3{% else %}2{% endif %},
+                nColumns: 2,
+                nStates: 0,
+                nActions: 0,
                 checked: [],
                 dygraphTime: 0.0,
                 dygraphIndex: 0,
                 videos: null,
                 video: null,
                 colors: null,
+                nVideos: {{ videos_info | length }},
+                nVideoReadyToPlay: 0,
+                videoCodecError: false,
 
                 // alpine initialization
                 init() {
+                    // check if videos can play
+                    const dummyVideo = document.createElement('video');
+                    const canPlayVideos = dummyVideo.canPlayType('video/mp4; codecs="av01.0.05M.08"'); // codec source: https://huggingface.co/blog/video-encoding#results
+                    if(!canPlayVideos){
+                        this.videoCodecError = true;
+                    }
+
+                    // process CSV data
                     this.videos = document.querySelectorAll('video');
                     this.video = this.videos[0];
                     this.dygraph = new Dygraph(document.getElementById("graph"), '{{ ep_csv_url }}', {
@@ -235,17 +276,19 @@
                                 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
-                                let lightnessIdx = 0;
-                                const chunkSize = Math.ceil(seriesNames.length / this.nColumns);
-                                for (let i = 0; i < seriesNames.length; i += chunkSize) {
-                                    const lightness = LIGHTNESS[lightnessIdx];
-                                    for (let hue = 0; hue < 360; hue += parseInt(360/chunkSize)) {
-                                        const color = `hsl(${hue}, 100%, ${lightness}%)`;
-                                        colors.push(color);
-                                    }
-                                    lightnessIdx += 1;
+                                // colors for "state" lines
+                                for (let hue = 0; hue < 360; hue += parseInt(360/this.nStates)) {
+                                    const color = `hsl(${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);
                                 }
                                 this.dygraph.updateOptions({ colors });
                                 this.colors = colors;
@@ -272,37 +315,40 @@
                     if (!this.currentFrameData) {
                         return [];
                     }
-                    const columnSize = Math.ceil(this.currentFrameData.length / this.nColumns);
-                    return Array.from({
-                        length: columnSize
-                    }, (_, rowIndex) => {
-                        const row = [
-                            this.currentFrameData[rowIndex] || null,
-                            this.currentFrameData[rowIndex + columnSize] || null,
-                        ];
-                        if (this.nColumns === 3) {
-                            row.push(this.currentFrameData[rowIndex + 2 * columnSize] || null)
-                        }
-                        return row;
-                    });
+                    const rows = [];
+                    const nRows = Math.max(this.nStates, this.nActions);
+                    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
+                        rowIndex += 1;
+                        rows.push(row);
+                    }
+                    return rows;
                 },
                 isRowChecked(rowIndex) {
-                    return this.rows[rowIndex].every(cell => cell && cell.checked);
+                    return this.rows[rowIndex].every(cell => cell && (cell.isNull || cell.checked));
                 },
                 isColumnChecked(colIndex) {
-                    return this.rows.every(row => row[colIndex] && row[colIndex].checked);
+                    return this.rows.every(row => row[colIndex] && (row[colIndex].isNull || row[colIndex].checked));
                 },
                 toggleRow(rowIndex) {
                     const newState = !this.isRowChecked(rowIndex);
                     this.rows[rowIndex].forEach(cell => {
-                        if (cell) cell.checked = newState;
+                        if (cell && !cell.isNull) cell.checked = newState;
                     });
                     this.updateTableValues();
                 },
                 toggleColumn(colIndex) {
                     const newState = !this.isColumnChecked(colIndex);
                     this.rows.forEach(row => {
-                        if (row[colIndex]) row[colIndex].checked = newState;
+                        if (row[colIndex] && !row[colIndex].isNull) row[colIndex].checked = newState;
                     });
                     this.updateTableValues();
                 },
@@ -343,7 +389,6 @@
                     window.history.replaceState({}, '', url.toString());
                 },
 
-
                 formatTime(time) {
                     var hours = Math.floor(time / 3600);
                     var minutes = Math.floor((time % 3600) / 60);
@@ -351,6 +396,14 @@
                     return (hours > 0 ? hours + ':' : '') + (minutes < 10 ? '0' + minutes : minutes) + ':' + (seconds <
                         10 ?
                         '0' + seconds : seconds);
+                },
+
+                videoCanPlay() {
+                    this.nVideoReadyToPlay += 1;
+                    if(this.nVideoReadyToPlay == this.nVideos) {
+                        // start autoplay all videos in sync
+                        this.$refs.btnPlay.click();
+                    }
                 }
             };
         }

pyproject.toml

@@ -43,8 +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 = ">=0.23.0"}
-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}
@@ -64,9 +64,13 @@ pandas = {version = ">=2.2.2", optional = true}
 scikit-image = {version = ">=0.23.2", optional = true}
 dynamixel-sdk = {version = ">=3.7.31", optional = true}
 pynput = {version = ">=1.7.7", optional = true}
-# TODO(rcadene, salibert): 71.0.1 has a bug
-setuptools = {version = "!=71.0.1", optional = true}
-
+feetech-servo-sdk = {version = ">=1.0.0", optional = true}
+setuptools = {version = "!=71.0.1", optional = true}  # TODO(rcadene, aliberts): 71.0.1 has a bug
+pyrealsense2 = {version = ">=2.55.1.6486", markers = "sys_platform != 'darwin'", optional = true}  # TODO(rcadene, aliberts): Fix on Mac
+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"
 
 
 [tool.poetry.extras]
@@ -75,10 +79,13 @@ pusht = ["gym-pusht"]
 xarm = ["gym-xarm"]
 aloha = ["gym-aloha"]
 dev = ["pre-commit", "debugpy"]
-test = ["pytest", "pytest-cov"]
+test = ["pytest", "pytest-cov", "pyserial"]
 umi = ["imagecodecs"]
 video_benchmark = ["scikit-image", "pandas"]
-koch = ["dynamixel-sdk", "pynput"]
+dynamixel = ["dynamixel-sdk", "pynput"]
+feetech = ["feetech-servo-sdk", "pynput"]
+intelrealsense = ["pyrealsense2"]
+stretch = ["hello-robot-stretch-body", "pyrender", "pyrealsense2", "pynput"]
 
 [tool.ruff]
 line-length = 110

tests/conftest.py

@@ -13,28 +13,107 @@
 # 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 traceback
+
 import pytest
+from serial import SerialException
 
+from lerobot import available_cameras, available_motors, available_robots
 from lerobot.common.utils.utils import init_hydra_config
-
-from .utils import DEVICE, KOCH_ROBOT_CONFIG_PATH
+from tests.utils import DEVICE, ROBOT_CONFIG_PATH_TEMPLATE, make_camera, make_motors_bus
 
 
 def pytest_collection_finish():
     print(f"\nTesting with {DEVICE=}")
 
 
-@pytest.fixture(scope="session")
-def is_koch_available():
+@pytest.fixture
+def is_robot_available(robot_type):
+    if robot_type not in available_robots:
+        raise ValueError(
+            f"The robot type '{robot_type}' is not valid. Expected one of these '{available_robots}"
+        )
+
     try:
         from lerobot.common.robot_devices.robots.factory import make_robot
 
-        robot_cfg = init_hydra_config(KOCH_ROBOT_CONFIG_PATH)
+        config_path = ROBOT_CONFIG_PATH_TEMPLATE.format(robot=robot_type)
+        robot_cfg = init_hydra_config(config_path)
         robot = make_robot(robot_cfg)
         robot.connect()
         del robot
         return True
+
     except Exception as e:
-        print("A koch robot is not available.")
-        print(e)
+        print(f"\nA {robot_type} robot is not available.")
+
+        if isinstance(e, ModuleNotFoundError):
+            print(f"\nInstall module '{e.name}'")
+        elif isinstance(e, SerialException):
+            print("\nNo physical motors bus detected.")
+        else:
+            traceback.print_exc()
+
         return False
+
+
+@pytest.fixture
+def is_camera_available(camera_type):
+    if camera_type not in available_cameras:
+        raise ValueError(
+            f"The camera type '{camera_type}' is not valid. Expected one of these '{available_cameras}"
+        )
+
+    try:
+        camera = make_camera(camera_type)
+        camera.connect()
+        del camera
+        return True
+
+    except Exception as e:
+        print(f"\nA {camera_type} camera is not available.")
+
+        if isinstance(e, ModuleNotFoundError):
+            print(f"\nInstall module '{e.name}'")
+        elif isinstance(e, ValueError) and "camera_index" in e.args[0]:
+            print("\nNo physical camera detected.")
+        else:
+            traceback.print_exc()
+
+        return False
+
+
+@pytest.fixture
+def is_motor_available(motor_type):
+    if motor_type not in available_motors:
+        raise ValueError(
+            f"The motor type '{motor_type}' is not valid. Expected one of these '{available_motors}"
+        )
+
+    try:
+        motors_bus = make_motors_bus(motor_type)
+        motors_bus.connect()
+        del motors_bus
+        return True
+
+    except Exception as e:
+        print(f"\nA {motor_type} motor is not available.")
+
+        if isinstance(e, ModuleNotFoundError):
+            print(f"\nInstall module '{e.name}'")
+        elif isinstance(e, SerialException):
+            print("\nNo physical motors bus detected.")
+        else:
+            traceback.print_exc()
+
+        return False
+
+
+@pytest.fixture
+def patch_builtins_input(monkeypatch):
+    def print_text(text=None):
+        if text is not None:
+            print(text)
+
+    monkeypatch.setattr("builtins.input", print_text)

tests/data/lerobot/aloha_mobile_cabinet/meta_data/info.json

@@ -1,3 +1,11 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:50e40e4c2bb523fca0b54e9a9635281312e9c6f9d757db03c06a0865c5508f29
-size 188
+{
+    "codebase_version": "v1.6",
+    "fps": 50,
+    "video": true,
+    "encoding": {
+        "vcodec": "libsvtav1",
+        "pix_fmt": "yuv420p",
+        "g": 2,
+        "crf": 30
+    }
+}

tests/data/lerobot/aloha_mobile_cabinet/train/dataset_info.json

@@ -1,3 +1,61 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:8b7fbedfdb3d536847bc6fadf2cbabb9f2b5492edf3e2c274a3e8ffb447105e8
-size 1166
+{
+  "citation": "",
+  "description": "",
+  "features": {
+    "observation.images.cam_high": {
+      "_type": "VideoFrame"
+    },
+    "observation.images.cam_left_wrist": {
+      "_type": "VideoFrame"
+    },
+    "observation.images.cam_right_wrist": {
+      "_type": "VideoFrame"
+    },
+    "observation.state": {
+      "feature": {
+        "dtype": "float32",
+        "_type": "Value"
+      },
+      "length": 14,
+      "_type": "Sequence"
+    },
+    "observation.effort": {
+      "feature": {
+        "dtype": "float32",
+        "_type": "Value"
+      },
+      "length": 14,
+      "_type": "Sequence"
+    },
+    "action": {
+      "feature": {
+        "dtype": "float32",
+        "_type": "Value"
+      },
+      "length": 14,
+      "_type": "Sequence"
+    },
+    "episode_index": {
+      "dtype": "int64",
+      "_type": "Value"
+    },
+    "frame_index": {
+      "dtype": "int64",
+      "_type": "Value"
+    },
+    "timestamp": {
+      "dtype": "float32",
+      "_type": "Value"
+    },
+    "next.done": {
+      "dtype": "bool",
+      "_type": "Value"
+    },
+    "index": {
+      "dtype": "int64",
+      "_type": "Value"
+    }
+  },
+  "homepage": "",
+  "license": ""
+}

tests/data/lerobot/aloha_mobile_cabinet/train/state.json

@@ -1,3 +1,13 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:34ece24fb6b302db0b68987858509f31713fb299faa9a9d34b8fd68f10bc3100
-size 247
+{
+  "_data_files": [
+    {
+      "filename": "data-00000-of-00001.arrow"
+    }
+  ],
+  "_fingerprint": "3e76021c95d21c4d",
+  "_format_columns": null,
+  "_format_kwargs": {},
+  "_format_type": null,
+  "_output_all_columns": false,
+  "_split": null
+}

tests/data/lerobot/aloha_mobile_chair/meta_data/info.json

@@ -1,3 +1,11 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:50e40e4c2bb523fca0b54e9a9635281312e9c6f9d757db03c06a0865c5508f29
-size 188
+{
+    "codebase_version": "v1.6",
+    "fps": 50,
+    "video": true,
+    "encoding": {
+        "vcodec": "libsvtav1",
+        "pix_fmt": "yuv420p",
+        "g": 2,
+        "crf": 30
+    }
+}

tests/data/lerobot/aloha_mobile_chair/train/dataset_info.json

@@ -1,3 +1,61 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:8b7fbedfdb3d536847bc6fadf2cbabb9f2b5492edf3e2c274a3e8ffb447105e8
-size 1166
+{
+  "citation": "",
+  "description": "",
+  "features": {
+    "observation.images.cam_high": {
+      "_type": "VideoFrame"
+    },
+    "observation.images.cam_left_wrist": {
+      "_type": "VideoFrame"
+    },
+    "observation.images.cam_right_wrist": {
+      "_type": "VideoFrame"
+    },
+    "observation.state": {
+      "feature": {
+        "dtype": "float32",
+        "_type": "Value"
+      },
+      "length": 14,
+      "_type": "Sequence"
+    },
+    "observation.effort": {
+      "feature": {
+        "dtype": "float32",
+        "_type": "Value"
+      },
+      "length": 14,
+      "_type": "Sequence"
+    },
+    "action": {
+      "feature": {
+        "dtype": "float32",
+        "_type": "Value"
+      },
+      "length": 14,
+      "_type": "Sequence"
+    },
+    "episode_index": {
+      "dtype": "int64",
+      "_type": "Value"
+    },
+    "frame_index": {
+      "dtype": "int64",
+      "_type": "Value"
+    },
+    "timestamp": {
+      "dtype": "float32",
+      "_type": "Value"
+    },
+    "next.done": {
+      "dtype": "bool",
+      "_type": "Value"
+    },
+    "index": {
+      "dtype": "int64",
+      "_type": "Value"
+    }
+  },
+  "homepage": "",
+  "license": ""
+}

tests/data/lerobot/aloha_mobile_chair/train/state.json

@@ -1,3 +1,13 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:38cf4116a65cb92a5c43f9b9da7a7b81cfa9168b17605c8c456f7d3a3a23b77a
-size 247
+{
+  "_data_files": [
+    {
+      "filename": "data-00000-of-00001.arrow"
+    }
+  ],
+  "_fingerprint": "872117944c4ecdff",
+  "_format_columns": null,
+  "_format_kwargs": {},
+  "_format_type": null,
+  "_output_all_columns": false,
+  "_split": null
+}

tests/data/lerobot/aloha_mobile_elevator/meta_data/info.json

@@ -1,3 +1,11 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:50e40e4c2bb523fca0b54e9a9635281312e9c6f9d757db03c06a0865c5508f29
-size 188
+{
+    "codebase_version": "v1.6",
+    "fps": 50,
+    "video": true,
+    "encoding": {
+        "vcodec": "libsvtav1",
+        "pix_fmt": "yuv420p",
+        "g": 2,
+        "crf": 30
+    }
+}

tests/data/lerobot/aloha_mobile_elevator/train/dataset_info.json

@@ -1,3 +1,61 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:8b7fbedfdb3d536847bc6fadf2cbabb9f2b5492edf3e2c274a3e8ffb447105e8
-size 1166
+{
+  "citation": "",
+  "description": "",
+  "features": {
+    "observation.images.cam_high": {
+      "_type": "VideoFrame"
+    },
+    "observation.images.cam_left_wrist": {
+      "_type": "VideoFrame"
+    },
+    "observation.images.cam_right_wrist": {
+      "_type": "VideoFrame"
+    },
+    "observation.state": {
+      "feature": {
+        "dtype": "float32",
+        "_type": "Value"
+      },
+      "length": 14,
+      "_type": "Sequence"
+    },
+    "observation.effort": {
+      "feature": {
+        "dtype": "float32",
+        "_type": "Value"
+      },
+      "length": 14,
+      "_type": "Sequence"
+    },
+    "action": {
+      "feature": {
+        "dtype": "float32",
+        "_type": "Value"
+      },
+      "length": 14,
+      "_type": "Sequence"
+    },
+    "episode_index": {
+      "dtype": "int64",
+      "_type": "Value"
+    },
+    "frame_index": {
+      "dtype": "int64",
+      "_type": "Value"
+    },
+    "timestamp": {
+      "dtype": "float32",
+      "_type": "Value"
+    },
+    "next.done": {
+      "dtype": "bool",
+      "_type": "Value"
+    },
+    "index": {
+      "dtype": "int64",
+      "_type": "Value"
+    }
+  },
+  "homepage": "",
+  "license": ""
+}

tests/data/lerobot/aloha_mobile_elevator/train/state.json

@@ -1,3 +1,13 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:d394d451929b805f2d94f9fc5b12d15c31cfc494df76d7d642b63378b8ba0131
-size 247
+{
+  "_data_files": [
+    {
+      "filename": "data-00000-of-00001.arrow"
+    }
+  ],
+  "_fingerprint": "ba1d9dc6ea5a9717",
+  "_format_columns": null,
+  "_format_kwargs": {},
+  "_format_type": null,
+  "_output_all_columns": false,
+  "_split": null
+}

tests/data/lerobot/aloha_mobile_shrimp/meta_data/info.json

@@ -1,3 +1,11 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:50e40e4c2bb523fca0b54e9a9635281312e9c6f9d757db03c06a0865c5508f29
-size 188
+{
+    "codebase_version": "v1.6",
+    "fps": 50,
+    "video": true,
+    "encoding": {
+        "vcodec": "libsvtav1",
+        "pix_fmt": "yuv420p",
+        "g": 2,
+        "crf": 30
+    }
+}

tests/data/lerobot/aloha_mobile_shrimp/train/dataset_info.json

@@ -1,3 +1,61 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:8b7fbedfdb3d536847bc6fadf2cbabb9f2b5492edf3e2c274a3e8ffb447105e8
-size 1166
+{
+  "citation": "",
+  "description": "",
+  "features": {
+    "observation.images.cam_high": {
+      "_type": "VideoFrame"
+    },
+    "observation.images.cam_left_wrist": {
+      "_type": "VideoFrame"
+    },
+    "observation.images.cam_right_wrist": {
+      "_type": "VideoFrame"
+    },
+    "observation.state": {
+      "feature": {
+        "dtype": "float32",
+        "_type": "Value"
+      },
+      "length": 14,
+      "_type": "Sequence"
+    },
+    "observation.effort": {
+      "feature": {
+        "dtype": "float32",
+        "_type": "Value"
+      },
+      "length": 14,
+      "_type": "Sequence"
+    },
+    "action": {
+      "feature": {
+        "dtype": "float32",
+        "_type": "Value"
+      },
+      "length": 14,
+      "_type": "Sequence"
+    },
+    "episode_index": {
+      "dtype": "int64",
+      "_type": "Value"
+    },
+    "frame_index": {
+      "dtype": "int64",
+      "_type": "Value"
+    },
+    "timestamp": {
+      "dtype": "float32",
+      "_type": "Value"
+    },
+    "next.done": {
+      "dtype": "bool",
+      "_type": "Value"
+    },
+    "index": {
+      "dtype": "int64",
+      "_type": "Value"
+    }
+  },
+  "homepage": "",
+  "license": ""
+}

tests/data/lerobot/aloha_mobile_shrimp/train/state.json

@@ -1,3 +1,13 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:69435f30146a309c8d7d0eb01216555bf0547095db1fc9c20218d481d6fe62c8
-size 247
+{
+  "_data_files": [
+    {
+      "filename": "data-00000-of-00001.arrow"
+    }
+  ],
+  "_fingerprint": "d95a3a7eae59566d",
+  "_format_columns": null,
+  "_format_kwargs": {},
+  "_format_type": null,
+  "_output_all_columns": false,
+  "_split": null
+}

tests/data/lerobot/aloha_mobile_wash_pan/meta_data/info.json

@@ -1,3 +1,11 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:50e40e4c2bb523fca0b54e9a9635281312e9c6f9d757db03c06a0865c5508f29
-size 188
+{
+    "codebase_version": "v1.6",
+    "fps": 50,
+    "video": true,
+    "encoding": {
+        "vcodec": "libsvtav1",
+        "pix_fmt": "yuv420p",
+        "g": 2,
+        "crf": 30
+    }
+}