G1 Dataset Conversion#

Convert recorded or generated Unitree G1 apple-to-plate demonstrations into LeRobot format for GR00T N1.7 training.

Tip

To use the provided training-ready dataset, skip this module and continue to G1 policy training. The next module includes the download command and training instructions.

Prerequisites#

  • Complete Installation.

  • Prepare replayed and inspected demonstrations recorded with G1 or validated output from G1 data generation.

  • Use data recorded with Pink IK and AGILE in either Isaac Lab-Arena or native Isaac Lab, including camera video, navigation commands, and base-height commands. The supplied policy training recipe requires these fields. Native fixed-base data omits the navigation and base-height commands and needs custom training and evaluation integration.

Run commands from this repository’s root on the workstation. Choose the conversion section matching your input and an output directory that does not already exist. If the dataset is already converted, continue to Validate the Training Input.

Convert Recorded Demonstrations#

Use this section to convert your own recording from the earlier recording module.

Tip

To train with provided data, use the published LeRobot dataset instead.

  1. Set G1_RECORDING in the terminal used for conversion. This example selects your recording from Isaac Lab-Arena with Pink IK and AGILE. For native Isaac Lab with AGILE, use outputs/g1/recording-native.hdf5 instead.

    G1_RECORDING="$PWD/outputs/g1/recording-arena.hdf5"
    

Convert the Selected Recording#

  1. Convert your recording selected above:

    uv run --extra converter python scripts/convert_hdf5_to_lerobot.py \
      --input "$G1_RECORDING" \
      --output outputs/g1/lerobot
    

Convert Generated Demonstrations#

Choose either your own generated dataset below or published generated data. For native output, continue directly to Convert the Selected File.

  1. For your own Arena output, set G1_GENERATED in the terminal used for conversion, then continue to Convert the Selected File:

    G1_GENERATED="$PWD/outputs/g1/generated-arena.hdf5"
    

Use Published Generated Data#

Alternatively, download the published dataset of 2,500 generated Arena episodes. The HDF5 file is 79.2 GB. Allow additional space for the converted dataset.

  1. Complete Hugging Face access setup, then download the generated dataset:

    uv run --frozen --no-sync hf download nvidia/GR00T-E2E-Reference-Workflow-SIM-G1-N1.7-AppletoPlate-Dataset \
      e2e_ra_g1_apple_to_plate_2500_generated_demos.hdf5 \
      --repo-type dataset --revision 36234f4b48dad27f218d8e29ec7710ad5a90b48c \
      --local-dir outputs/g1/downloads
    
  2. Select the downloaded dataset for conversion:

    G1_GENERATED="$PWD/outputs/g1/downloads/e2e_ra_g1_apple_to_plate_2500_generated_demos.hdf5"
    

The HDF5 file includes camera frames and metadata describing its generation settings and source recording. Conversion preserves this metadata. No separate recording is needed for conversion.

Convert the Selected File#

The --lane option checks that the input is marked as generated data from the selected workflow. The converter saves the original metadata and the HDF5 file’s checksum in meta/e2e_workflow_lineage.json.

  1. Run one of the following commands, matching how the data was generated.

    For Isaac Lab-Arena output:

    uv run --extra converter python scripts/convert_hdf5_to_lerobot.py \
      --lane arena --input "$G1_GENERATED" \
      --output outputs/g1/lerobot-generated
    

    For native output:

    uv run --extra converter python scripts/convert_hdf5_to_lerobot.py \
      --lane native --input outputs/g1/generated-native.hdf5 \
      --output outputs/g1/lerobot-generated
    

Native output generated with --no-cameras cannot be converted for training. Arena generation requires cameras.

Validate the Training Input#

  1. Check the output directory for data/, videos/, and meta/, including meta/info.json and meta/modality.json. The converter writes the episode data and camera video; GR00T generates meta/stats.json on its first training run.

  2. Confirm that navigate_command and base_height_command appear under action in meta/modality.json. Conversion can succeed without those groups, but the supplied policy training recipe requires them.

To inspect another input before writing, use its conversion command with --inspect --strict and omit --output. It prints the resolved fields and fails when a required field or mapping is missing.

Expected Result#

outputs/g1/lerobot/ contains the converted recording, or outputs/g1/lerobot-generated/ contains the generated dataset. Keep the complete directory for training.

Next Stage#

Continue to G1 Policy Training. Set TRAIN_DATASET to the absolute path of the converted directory when preparing the training terminal.