Unitree G1 Dataset Conversion#

Dataset conversion turns recorded MCAP sessions into one LeRobot dataset for GR00T N1.7 fine-tuning.

Prerequisites#

  • Retain the complete session directories from Data Recording.

  • Run on an x86_64 workstation or Thor. The converter requires neither a GPU nor ROS, and can run on the host or inside the Isaac ROS container.

  • Set ISAAC_ROS_WS to the workspace on this machine. On a workstation, copy the recordings into its recordings/ directory and clone the pinned isaac_ros_data_tools source from Isaac ROS Setup.

Choose the Dataset Directory#

Set the output path once in the terminal used below. Choose a directory that does not exist; change lerobot_output if that name is already in use.

DATASET_DIR="${ISAAC_ROS_WS}/recordings/lerobot_output"

Use the Published Dataset (Optional)#

To skip recording and conversion, download the published nvidia/GR00T-N1.7-AppleToPlate LeRobot dataset. Install the Hugging Face CLI and authenticate with an account that has access to the dataset before downloading:

hf download \
    nvidia/GR00T-N1.7-AppleToPlate \
    --repo-type dataset \
    --local-dir "${DATASET_DIR}"

Then continue to Validate Output for Fine-Tuning.

Install the Converter With uv#

Install uv using the host tools instructions before continuing.

  1. Enter the converter source directory, then create and activate its environment in the same terminal:

    cd ${ISAAC_ROS_WS}/src/isaac_ros_data_tools/isaac_ros_mcap_lerobot_converter
    
    uv venv --python 3.12 .venv
    
    source .venv/bin/activate
    

    Install the converter in that environment:

    uv pip install -e .
    

Convert a Recording Session#

  1. Run the following script to convert the rosbags of your recording sessions. Replace <session_name> with the name of your recording session.

    To combine multiple recording sessions, pass multiple --rosbags-dir arguments to convert all sessions into a single LeRobot dataset. Episodes are numbered sequentially across sessions in the order listed.

    mcap-to-lerobot \
        --rosbags-dir ${ISAAC_ROS_WS}/recordings/<session_name> \
        --output-dir "${DATASET_DIR}" \
        --task "move the apple to the plate" \
        --fps 30 \
        --robot-type unitree_g1
    

    See also

    Refer to the isaac_ros_mcap_lerobot_converter reference for full CLI details and --fps/sync_rate guidance.

Validate Output for Fine-Tuning#

  1. Check the dataset directory and metadata before training:

    ls "${DATASET_DIR}"
    
  2. Confirm the output is ready:

    • [ ] Output directory contains four folders: data/, meta/, videos/, and images/.

    • [ ] Output directory contains new_embodiment_config_defaults.py for GR00T fine-tuning.

    • [ ] data/, meta/, and videos/ contain data.

    • [ ] images/ exists but is empty.

    • [ ] Episode count aligns with source sessions.

    • [ ] fps in metadata matches the recorder sync_rate (30 Hz by default).

Next Stage#

The verified LeRobot dataset and new_embodiment_config_defaults.py are the inputs to Unitree G1 Policy Training. Use the path stored in DATASET_DIR as <dataset-dir> in training and LEAPP export. If training runs on another machine, transfer the complete dataset and use its destination path instead.