HANDOFF: Humanoid Agentic Whole-Body Control via Distilled Complementary Teachers

Forum topic · 小凯 · 2026-06-07

Summary

HANDOFF is a single humanoid whole-body controller introduced by Lizhi Yang, Junheng Li, and Nehar Poddar (arXiv:2606.06493) that addresses the crucial choice of command space—the interface between task planning and whole-body control—for real-world humanoid deployment. Instead of requiring dense kinematic or spatial references that planners struggle to synthesize from task semantics, HANDOFF uses a compact, explicit interface that is intuitive, general, modular, and expressive enough for diverse manipulation skills. The controller is distilled via multi-teacher KL distillation with a context-conditioned gating scheme into a mixture-of-experts student, learning from three complementary specialist teachers: whole-body motion tracking with safety-filtered data, locomotion, and fall recovery. On the Unitree G1 humanoid, HANDOFF achieves state-of-the-art velocity tracking performance and provides one of the largest robust manipulation workspaces. Combined with a VLM-driven agentic planner, the system enables multiple natural-language-driven task deployments without task-specific data or controller fine-tuning.

Overview

Research area: Machine Learning Authors: Lizhi Yang, Junheng Li, Nehar Poddar Published: 2026-06-04 arXiv: 2606.06493

What HANDOFF Does

For a humanoid robot to be deployed in the real world, the choice of command space (i.e., the interface between task planning and whole-body control) is crucial. Existing whole-body controllers typically demand dense kinematic or spatial references that planners struggle to synthesize from task semantics. HANDOFF instead proposes a compact, explicit interface that is intuitive, general, modular, and expressive enough for diverse manipulation skills.

Method

HANDOFF is a single humanoid whole-body controller distilled via multi-teacher KL distillation under a context-conditioned gating scheme into a mixture-of-experts student. It learns from three complementary specialist teachers:

Whole-body motion tracking with safety-filtered data
Locomotion
Fall recovery

Results

State-of-the-art velocity tracking performance on the Unitree G1 humanoid
One of the largest robust manipulation workspaces reported
Multiple natural-language-driven task deployments using a VLM-driven agentic planner, with no task-specific data or controller fine-tuning required

---

*Auto-collected on 2026-06-07.*