H2 Data Generation With Isaac AutoData (Optional)#

Use Isaac AutoData to generate perturbed demonstrations from accepted episodes recorded with Unitree H2 in simulation. Skip this stage to convert your recorded episodes or download generated demonstrations for conversion.

Prerequisites#

Complete H2 Asset Preparation on the GPU workstation. In each new terminal, restore the simulation and H2 asset exports listed under For New Terminals. Run commands from the repository root.

conda is needed only if the provision command must create the isolated AutoData runtime. Generation uses one environment because the pinned H2 AGILE policy supports batch size one.

Prepare a Recording#

  1. Complete H2 Demonstration Recording, including strict state replay, using the current H2 workflow.

  2. Keep that run’s H2_RUN_DIR and H2_RECORDING values. If you open a new terminal, restore those values to the existing run before continuing.

The published H2 HDF5 downloads were recorded with different task settings and are provided for conversion. Generation requires a recording whose saved software versions and simulation settings match this checkout.

Prepare Isaac AutoData#

  1. On a fresh setup, provision the pinned AutoData checkout. Skip this command if it is already provisioned:

    ./tools/launch_autodata.sh provision
    
  2. Verify the Isaac Lab-Arena generation route:

    ./tools/launch_autodata.sh generate --lane arena --check
    

    The check must print the pinned AutoData revision and report the pinned Isaac Lab-Arena runtime as its interpreter.

Generate Demonstrations#

  1. Generate from the recorded HDF5 file:

    H2_GENERATED="$H2_RUN_DIR/generated.hdf5"
    
    H2_TARGET_DEMOS=50
    

    Choose a wall-clock budget before starting. It includes simulator startup, all successful and failed attempts, and dataset finalization. Measure a small run on the same workstation and size the budget from its time per attempt multiplied by --max_attempts, with extra time for startup, finalization, and variation. One hour is not a completion guarantee for 50 successes. The four-hour example below is a starting budget; increase it if your measured attempt rate requires more time.

    H2_GENERATION_TIMEOUT_SECONDS=14400
    
    ./tools/launch_autodata.sh generate --lane arena \
      --generation-run-timeout-seconds "$H2_GENERATION_TIMEOUT_SECONDS" \
      --input_file "$H2_RECORDING" \
      --output_file "$H2_GENERATED" \
      --result_file "$H2_RUN_DIR/generation.json" \
      -- --device cuda:0 \
         --requested_successes "$H2_TARGET_DEMOS" \
         --max_attempts 100
    

    Arena recordings already contain the boundary signals, so no annotation step is needed. Recordings from another checkout are accepted when their saved robot and task settings, software versions, action layout, and pickup/placement markers match. The source commit and uncommitted-change status are retained in the metadata.

  2. Confirm the launcher exits successfully and prints Generation complete: 50/50 successful demonstrations, followed by the attempt count and the dataset and result paths. This summary appears after the launcher validates attempt accounting, manifests, lineage, and matching artifact hashes, and publishes the finalized dataset. generation.json must report status: ok and num_success: 50.

    Reaching the attempt limit or timeout before 50 successes is a failed run. Files retained after a failed run may lack a finalized manifest and are not ready for conversion. Inspect generation.json, then rerun into a new output directory with a budget sized for the observed attempt rate.

  3. After successful generation, select its output for conversion:

    H2_TRAINING_HDF5="$H2_GENERATED"
    
    H2_TRAINING_DEMOS="$H2_TARGET_DEMOS"
    

Measured joint-limit faults consume an attempt and discard its incomplete recording before retrying. Other faults stop generation. For a separate run, choose a new output directory and pass --generation_seed after -- to select a seed.

Expected Result#

$H2_RUN_DIR/generated.hdf5 contains the requested accepted perturbations. generation.json records the result and hashes. Completed failed perturbations remain in the sibling generated_failed.hdf5 artifact for inspection.

Next Stage#

Continue to H2 Dataset Conversion to replay and validate the generated episodes before training.