G1 Data Generation With Isaac AutoData#

Isaac AutoData expands recorded Unitree G1 apple-to-plate demonstrations by replaying adapted trajectories in randomized scenes. This stage is optional. Recordings made with Pink IK and AGILE in either Isaac Lab-Arena or native Isaac Lab can also go directly to dataset conversion. For Unitree H2 Plus, use H2 data generation.

Requirements#

  • Complete Installation on the GPU workstation.

  • Use replayed and inspected demonstrations recorded with G1 compatible with this checkout. Keep the original HDF5 file unchanged.

  • Use a Linux GPU workstation and run commands from this repository’s root.

  • Prepare a clean AutoData checkout at the revision in runtime.toml. Native commands and validate-dataset also need AutoData’s interpreter. Follow First-Time Provisioning if either is missing.

Select an Isaac Lab-Arena Recording#

  1. Select your accepted recording in the terminal used for generation:

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

For native Isaac Lab recordings, use the native recording paths shown below instead.

Published HDF5 files were recorded with different simulator and task settings and are provided for conversion. To skip recording and generation, use the generated HDF5 download.

Choose the Steps for Your Recording Method#

Follow the steps for the method you used to record your demonstrations:

Recording

Source HDF5

Sequence

Isaac Lab-Arena with Pink IK and AGILE

outputs/g1/recording-arena.hdf5

Generate → validate

Native Isaac Lab with a fixed base or AGILE

outputs/g1/recording-native.hdf5

Annotate → validate → generate → validate

The commands call this recording method a lane: --lane arena selects Isaac Lab-Arena, and --lane native selects native Isaac Lab.

The metadata embedded in the recording identifies its action layout and manipulating hand. The command fails before simulation starts if the selected recording method or robot-control configuration conflicts with that metadata. Isaac Lab-Arena recordings made with Pink IK already contain subtask boundaries and must not be annotated. Arena joint-control recordings are for recording and replay, not AutoData.

Both recording methods support the supplied policy training and evaluation configuration when they use AGILE. Native fixed-base data omits its required navigation and base-height fields and needs custom training and evaluation integration.

First-Time Provisioning#

  1. With Git and Conda on PATH, provision the pinned source and interpreter:

    ./tools/launch_autodata.sh provision
    

    The launcher sets PYTHONNOUSERSITE=1 for all commands to prevent user-site packages from shadowing the environment.

    The default checkout is the sibling Isaac-AutoData-e2e and the Conda environment is isaac_autodata. Use --autodata-root or ISAAC_AUTODATA_ROOT for another checkout; use --autodata-python or AUTODATA_PYTHON for another interpreter. Provisioning includes the pinned extras required by native operations.

  2. Check the prepared runtime:

    ./tools/launch_autodata.sh --check
    

    For Arena generation, also check its runtime:

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

    The checks print the pinned checkout and interpreter. Arena generation uses this repository’s locked Arena runtime. It does not use the AutoData interpreter for simulation.

Tip

For an existing pinned checkout and interpreter, run ./tools/launch_autodata.sh setup to refresh the extras. Use the same checkout and interpreter overrides as the existing installation.

Isaac Lab-Arena Recording Generation#

  1. Generate directly from the selected Isaac Lab-Arena recording made with Pink IK. This requests one success within three attempts and a seven-minute process deadline:

    ./tools/launch_autodata.sh generate --lane arena \
      --generation-run-timeout-seconds 420 \
      --input_file "$G1_RECORDING" \
      --output_file outputs/g1/generated-arena.hdf5 \
      --result_file outputs/g1/generation-arena.json \
      -- --device cuda:0 --requested_successes 1 --max_attempts 3
    

    The input must contain successful episodes, subtask boundaries, and saved software versions and task settings that match this checkout. The recording metadata selects the left- or right-hand task description and retains the source commit.

  2. Validate the generated file:

    ./tools/launch_autodata.sh validate-dataset outputs/g1/generated-arena.hdf5
    

    A completed run writes outputs/g1/generated-arena.hdf5 and reports status: ok in outputs/g1/generation-arena.json, with attempt counts and artifact hashes. An exhausted attempt budget or deadline is a failed run, even if partial artifacts remain.

Native Recording Generation#

  1. Annotate your native recording to find where pickup ends and transport begins. --wbc-mode auto reads fixed or AGILE mode from the manifest:

    ./tools/launch_autodata.sh annotate --lane native --wbc-mode auto \
      --device cpu \
      --input_file outputs/g1/recording-native.hdf5 \
      --output_file outputs/g1/annotated-native.hdf5
    

    Only episodes satisfying the task’s success condition are exported. An empty annotated file means no source episode passed that condition. For manual boundary annotation, add --manual: N plays, B pauses, S marks the boundary, and Q skips an episode.

  2. Validate the annotated pool:

    ./tools/launch_autodata.sh validate-dataset outputs/g1/annotated-native.hdf5
    
  3. Generate ten successful episodes, with a seven-minute process deadline:

    ./tools/launch_autodata.sh generate --lane native --wbc-mode auto \
      --generation-run-timeout-seconds 420 \
      --viz none --device cuda:0 --num_envs 5 --generation_num_trials 10 \
      --input_file outputs/g1/annotated-native.hdf5 \
      --output_file outputs/g1/generated-native.hdf5 \
      --result_file outputs/g1/generation-native.json
    

    This task retries failed attempts until ten successes or the deadline. Failed attempts are not exported. CUDA is recommended for AGILE generation; CPU supports small checks but runs substantially slower.

  4. Validate the generated file:

    ./tools/launch_autodata.sh validate-dataset outputs/g1/generated-native.hdf5
    

    A completed run writes outputs/g1/generated-native.hdf5 and reports status: ok in outputs/g1/generation-native.json. A timeout or failed run retains an error result. Investigate it before using the output for training. Continue to G1 Dataset Conversion.

Tip

For native trajectory inspection only, generate --no-cameras skips video. That output cannot be converted for training. Arena generation requires cameras and rejects --no-cameras.

Next Stage#

For data generated from either AGILE recording method above, continue to G1 Dataset Conversion with the validated generated HDF5 file. Select the command for the method you used to generate the data, so its source details are retained in the training dataset.

Supporting Guidance#

Native Isaac Lab integration requirements, simulator validation commands, and the upstream example are in Native AutoData Integration.

Only older recordings without e2e_workflow_manifest need Legacy Recording Migration. Do not migrate current recordings or newly merged files to work around a missing or incompatible manifest.