H2 Asset Preparation#

Prepare the Unitree H2 Plus simulation model, Pink kinematics, and AGILE policy for recording and evaluation in Isaac Lab-Arena.

Prerequisites#

Complete Installation. Run commands from the repository root on the GPU workstation. Recording later requires a PICO 4 Ultra headset on the workstation’s LAN and its paired handheld XR controllers.

The AGILE policy is available here.

Prepare the Simulation Assets#

  1. Sync the repository with the converter used by H2 validation:

    uv sync --frozen --extra converter
    
  2. Generate the content-pinned Pink sidecar in the Isaac runtime:

    Warning

    Read the NVIDIA Omniverse license terms before running the following setup block. The EULA exports record your acceptance and suppress the runtime license prompt. Run them only if you accept the terms.

    export ACCEPT_EULA=Y
    export OMNI_KIT_ACCEPT_EULA=YES
    export E2E_H2_ASSET_CACHE="$HOME/.cache/e2e-reference-app/h2plus"
    
    ./tools/arena_launch.sh python tools/materialize_h2_pink_sidecar.py \
      --cache-root "$E2E_H2_ASSET_CACHE" \
      --regenerate
    

    Successful completion verifies the model and generated sidecars against their pinned digests.

  3. In the command’s final JSON output, check that status is ok. Copy the value of pink_urdf, which ends in unitree_h2.urdf, into the export below:

    export E2E_H2_PINK_URDF="/absolute/path/to/unitree_h2.urdf"
    

Download the AGILE Policy#

  1. Download the repository, revision, and filename specified in configs/robots/h2.yaml, then verify the configured digest. Continue only after each command succeeds:

    read -r H2_AGILE_REPOSITORY H2_AGILE_REVISION H2_AGILE_FILENAME < <(
      uv run --frozen --no-sync python -c \
        'from e2e_robot_workflows.robots.h2 import load_h2_robot_config; c = load_h2_robot_config(); print(c.policy_repository, c.policy_revision, c.policy_filename)'
    )
    
    E2E_H2_AGILE_POLICY="$(
      uv run --frozen --no-sync hf download --quiet \
        "$H2_AGILE_REPOSITORY" "$H2_AGILE_FILENAME" \
        --revision "$H2_AGILE_REVISION"
    )"
    
    export E2E_H2_AGILE_POLICY
    
    uv run --frozen --no-sync python -c \
      'from e2e_robot_workflows.robots.h2 import load_h2_robot_config; print(load_h2_robot_config().resolve_policy())'
    

    Confirm the command prints the downloaded agile_velocity.onnx path. A wrong file fails its SHA-256 check.

Prepare the GR00T Runtime#

  1. Follow Prepare NVIDIA Isaac GR00T in Your Workspace to create a clean, standalone Isaac-GR00T checkout at the revision pinned in runtime.toml.

  2. Set its absolute path and verify the runtime:

    export ISAAC_GR00T_ROOT=/absolute/path/to/Isaac-GR00T
    
    (cd "$ISAAC_GR00T_ROOT" && uv sync --frozen)
    

    After synchronization succeeds, check the runtime:

    ./tools/gr00t_runtime.sh --gr00t-root "$ISAAC_GR00T_ROOT" --check
    

    Confirm the command prints TorchCodec 0.4.0: VideoDecoder. H2 validation uses GR00T’s training loader and TorchCodec to check converted video data.

For New Terminals#

Use the shared Environment Variable Reference to restore this setup’s asset paths, runtime settings, and run variables in each new terminal. It covers simulation, CloudXR, AutoData, training, and evaluation.

Expected Result#

The model and sidecars are cached outside the repository. The AGILE policy passes its digest check, and the pinned GR00T runtime passes its loader check.

Next Stage#

Continue to H2 Simulation Validation.