G1 Simulation#

Launch the simulated Unitree G1 apple-to-plate scene and practice AGILE locomotion before recording demonstrations.

Requirements#

Use the workstation, repository checkout, and locked environment prepared by Installation. Run the commands below from the repository root.

Complete Create and Check the Simulation Environment before launching the viewer. Both G1 compositions use that environment. On first use, Isaac Sim may request NVIDIA EULA acceptance and download robot, scene, and policy assets.

Drive the Unitree G1#

  1. Launch the native AGILE viewer:

    uv run scripts/locomotion_with_agile.py --device cuda:0 --viz kit
    
  2. Click the viewport to give it keyboard focus. Use the controls below to move the robot, then release the movement keys to stop.

    Input

    Command

    Up / Down

    Walk forward / backward

    Left / Right

    Turn

    A / D

    Strafe

    Release Up, Down, Left, Right, A, and D

    Stop and balance

  3. Confirm that the robot stands upright, responds to the controls, and balances when you release them. Stop the viewer with Ctrl+C in its terminal before starting another simulation.

Use --help to list options such as --speed, --turn, --hip_height, and constant velocity commands such as --vx 0.3.

Note

This viewer enables AGILE for its own process. Select the recording mode separately using the commands in G1 demonstration recording.

Optional Isaac Lab-Arena Simulation#

Use this alternative to preview the composition used for checkpoint evaluation. It has the same keyboard controls and uses Arena’s own locomotion policy.

  1. Launch it from the repository root:

    ./tools/launch_arena_locomotion.sh --viz kit
    

    The launcher fetches and validates the pinned Arena checkout, then uses the repository’s locked environment. See How Arena Is Obtained for checkout details.

  2. Repeat the movement and balance check above, then stop the viewer with Ctrl+C in its terminal.

Next Stage#

Continue to G1 demonstration recording to capture apple-to-plate demonstrations with a headset. This viewer stage produces no training dataset.