Paper Overview
Field: ML Authors: Lizhi Yang, Junheng Li, Nehar Poddar Published: 2025-06-11 arXiv: 2506.08300
Abstract (Original)
For a humanoid robot to be deployed in the real world, the choice of command space is crucial. We introduce a compact, explicit interface that is intuitive, general, modular, and expressive enough for diverse manipulation skills. We introduce HANDOFF, a single humanoid whole-body controller distilled via multi-teacher KL distillation under a context-conditioned gating scheme into a mixture-of-experts student from three complementary specialists. On the Unitree G1, HANDOFF matches state-of-the-art velocity tracking and offers one of the largest robust manipulation workspaces.
Key Points
- Problem: Existing whole-body controllers typically require dense kinematic or spatial references, which task planners struggle to synthesize from task semantics. The command space (interface between task planning and whole-body control) is critical for real-world humanoid deployment.
- Solution: A compact, explicit interface that is intuitive, general, modular, and expressive enough to encode diverse manipulation skills.
- Method: HANDOFF — a single whole-body controller trained via multi-teacher KL distillation under a context-conditioned gating scheme, producing a mixture-of-experts student from three complementary specialist teachers:
- Whole-body motion tracking
- Safety-filtered data
- Locomotion and fall recovery
- Results: On the Unitree G1 humanoid, HANDOFF matches state-of-the-art velocity tracking and provides one of the largest robust manipulation workspaces.
*Auto-collected on 2025-06-11*