Researchers introduce HANDOFF, a unified humanoid whole-body controller that uses distilled knowledge from complementary specialists (motion tracking, locomotion, fall-recovery) to achieve robust task-space control. It demonstrates state-of-the-art performance and hardware feasibility
The development of effective command spaces for deploying humanoid robots in the real world is crucial. This paper proposes HANDOFF, a compact and expressive interface for whole-body control that bridges task planning and execution more effectively than existing methods. HANDOFF is achieved by distilling knowledge from three complementary specialists—whole-body motion tracking, locomotion, and fall-recovery—using multi-teacher KL distillation under a context-conditioned gating scheme. This results in a single, robust controller. Experimental results show that HANDOFF matches state-of-the-art velocity tracking on the Unitree G1. Furthermore, the method demonstrates hardware feasibility through natural-language-driven task roll-outs powered by a VLM-driven agentic planner, eliminating the need for task-specific controller fine-tuning.