End-To-End Workflows

In Memory

examples/in_memory_capture.py creates a typed SceneObservation and runs it through RetargetingExperiment without any intermediate file.

Skateboarding

experiment = RetargetingExperiment(
    observation=SkateboardingObservationRecipe(
        mocap=ViconRecordingSource(vicon_path, VICON_SCHEMA, name=demo),
        human_pose=GvhmrOutputSource(
            gvhmr_path,
            GVHMR_SCHEMA,
            fps=video_fps,
            name=demo,
        ),
    ),
    recipe=SkateboardingRetargetingRecipe(),
    robot=robot,
    retargeter=retargeter,
)
result = experiment.run()

The observation recipe estimates clock alignment from typed foot cues, selects the shared timeline, resamples each track by its own interpolation policy, converts frames, performs rigid point registration, reconstructs the board, and derives semantic contacts. The adaptation recipe resolves those semantics through robot roles.

Raw video can replace an existing pose output with VideoPoseSource and a HumanPoseEstimator. GvhmrEstimator is an optional implementation that uses a temporary workspace.

Holosoma

Holosoma climbing uses the same hierarchy with one mocap source and no temporal fusion:

experiment = RetargetingExperiment(
    observation=HolosomaClimbObservationRecipe.from_fixture(root),
    recipe=HolosomaClimbRetargetingRecipe(holosoma_root=root),
    robot=g1_spherehand_robot(root),
    retargeter_factory=mujoco_retargeter,
)

It is a specialization of the same observation, adaptation, compilation, and execution interfaces.