[ooctipus]

Description

Integrate SuperMimic as a native G1 motion-tracking task and training backend. IsaacContrib-Motion-Tracking-G1 owns the scene, actuators, commands, observations, rewards, terminations and recorder term. The standard reinforcement-learning scripts select --rl_library supermimic for MLP or optional one-frame NeRD dynamics, differentiable policy learning, evaluation and video.

The learner owns replay and optimization. Native training/playback entrypoints own stepping, configuration, W&B uploads and resumable checkpoints. Robot/motion preparation, stationary endpoint goals, the upstream MIT notice and finite quaternion gradients at identity are included. No external SuperMimic runtime package or alternate task loop is required.

Dependency: #8822 provides shared capture and optional batched episode storage. This branch includes that unmerged commit and targets upstream main from the fork. Review only the SuperMimic changes. It does not depend on the separate offline NeRD/Cartpole/ANYmal PR #8825. Its optional NeRD architecture comes from PhysicsNeMo Robotics.

The SuperMimic guide covers pinned asset/motion preparation, training, logging, checkpoint recovery and full-corpus evaluation.

Validation

The full warnings-as-errors Sphinx HTML build passed for this branch.

102 focused tests passed in separate file processes: 7 learner/architecture, 3 G1 capture/task, 55 entrypoint, 33 task-discovery, 2 motion-preparation and 2 quaternion conversion/gradient tests. Full repository formatting/changelog gates and git diff --check passed.

The split preserves the runtime source from #8806. NeRD-specific ownership assertions moved to the NeRD PR; this branch's architecture checks cover G1, SuperMimic and native entrypoints without requiring the offline NeRD files. Imports are verified against this isolated checkout, with installed editable-package fallback disabled; the offline NeRD modules are unavailable in this environment.

The original native workflow's tensor-parity, train/save/resume and asset/video evidence remains recorded in #8806. This split does not rerun the full corpus or claim the paper's zero-failure result. Motion and robot data remain separately prepared inputs; evaluation reports failures over the actual prepared corpus. The default MLP path does not need PhysicsNeMo.

Type of change

• Native G1 task and SuperMimic learning backend
• Quaternion-gradient fix
• Motion preparation tools, task discovery and documentation

Checklist

• Read the contribution guidelines
• Ran repository formatting and changelog checks
• Updated documentation
• Validated changed behavior and applied the test-audit criteria
• Added package changelog fragments
• Contributor already listed in CONTRIBUTORS.md
• No runtime warnings (existing dependency/backend warnings remain)