Reference#

Reference pages collect the exact sources of runtime, configuration, and validation behavior.

Use the Environment Variable Reference to restore runtime, asset, and run settings when opening a new terminal in any workflow.

Configuration and Version References#

Concern

Authoritative source

Python dependencies

uv.lock

Tested external revisions and released checkpoint

runtime.toml

Robot model and robot-control defaults

configs/robots/

Task discovery

configs/tasks/

Dataset and policy stages

configs/stages/

Validation entry points

tools/validate_repository.sh

Supported Configurations and Limitations#

  • Simulator, Isaac Lab-Arena, and XR behavior has been validated on Linux x86_64.

  • The application has Unitree G1 and Unitree H2 Plus adapters for the static-apple task. Unitree G1 retains built-in and Isaac Lab-Arena paths; Unitree H2 is Isaac Lab-Arena-only.

  • Recording and checkpoint evaluation share the settled-placement predicate. Recording demonstrations with H2 additionally requires pickup and receiving-hand boundaries and follow-through; its evaluation placement metric does not establish those handover events. With automatic completion enabled, manual finish/reset ends an episode but does not override the policy label. On Unitree G1, passing --no-auto_success switches to operator-owned acceptance. The Isaac AutoData variant of the task adds a success termination and a subtask boundary signal, which score and annotate generated attempts rather than recordings.

  • Isaac Lab-Arena’s code is an external source checkout rather than a locked dependency, because its wheel ships no data files. It is pinned by revision in runtime.toml, enforced by tools/check_arena_pin.py, fetched by tools/fetch_arena_checkout.sh when E2E_ARENA_ROOT is unset, and overlaid on this environment by tools/arena_launch.sh. Its dependencies are locked here, in the arena dependency group, so Isaac Lab-Arena runs in one interpreter with the simulator versions this repository pins.

  • Isaac AutoData uses its own checkout, pinned by revision in runtime.toml, and is reached only through tools/launch_autodata.sh. Native operations use AutoData’s isolated interpreter; Isaac Lab-Arena generation overlays its source on the repository’s Isaac Lab-Arena runtime.

  • The HDF5-to-LeRobot command writes a GR00T LeRobot v2.1 dataset by default, including each workflow’s recorded command fields: demonstrations recorded with Unitree G1 using Pink IK and AGILE in Isaac Lab-Arena, or AGILE in native Isaac Lab, include navigation and base height, while demonstrations with Unitree H2 include navigation only. --inspect opts out and only reports whether the configured fields resolve. Two omissions are deliberate: meta/stats.json is generated downstream in the GR00T environment, and the writer drops modality groups a run does not produce. The supplied modality for policy training and evaluation with Unitree G1 expects its locomotion commands. Both recording methods described above, including generated data, provide them. Demonstrations recorded with a fixed-base G1 omit them and need custom training and inference integration.

  • The static_apple_v0 environment preserves the workflow’s recording scene. It does not claim equivalence with Isaac Lab-Arena’s first-party Galileo scene or released-checkpoint performance. The Arena path uses this same scene as e2e_static_apple rather than reusing Isaac Lab-Arena’s scene.

  • AGILE inputs are pinned at their ownership seams. The built-in Unitree G1 task uses the TorchScript digest and Isaac Lab-Arena Unitree G1 uses the Isaac Lab-Arena policy revision recorded in runtime.toml. Unitree H2 uses the repository, revision, filename, and digest in configs/robots/h2.yaml.

  • Simulator assets, model checkpoints, and AGILE policies remain external. Initial setup requires network access and sufficient local storage.

For setup, local builds, validation, and integration commands, use the Developer Guide.

For hardware, camera, streaming, and compute-unit issues on the physical Unitree G1 path, use Troubleshooting.