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

Researchers introduce HANDOFF, a humanoid robot whole-body controller that uses distilled multi-teacher learning to enable intuitive task planning and robust manipulation. The system demonstrates real-world feasibility on Unitree G1 robots with natural language task execution, advancing practical deployment of humanoid robots in complex environments.

HANDOFF represents a significant engineering advancement in humanoid robotics by solving a critical interface problem between high-level task planning and low-level motor control. Traditional whole-body controllers require dense kinematic references that are difficult for planners to generate from semantic task descriptions. This research bridges that gap through a compact, explicit command space interface that simplifies the integration of vision-language models with robotic control systems.

The technical approach leverages knowledge distillation from three specialized teacher networks—whole-body motion tracking, locomotion, and fall-recovery—into a single mixture-of-experts student model. This design philosophy echoes broader trends in AI where specialized expert systems are combined through gating mechanisms for improved generalization. The safety-filtered training data addresses a persistent challenge in robotics: ensuring controllers learn robust behaviors without catastrophic failure modes during real-world deployment.

The hardware validation on Unitree G1 demonstrates that the approach translates from simulation to physical systems, a notoriously difficult step in robotics research. Integration with VLM-driven agentic planners without task-specific fine-tuning suggests the controller generalizes across diverse manipulation scenarios. This modularity and expressiveness could accelerate the development of practical humanoid applications beyond controlled laboratory settings.

Looking forward, the key question is whether this command-space design becomes an industry standard interface. If HANDOFF's approach gains adoption, it could significantly reduce development cycles for new manipulation skills and enable faster scaling of humanoid deployment in real-world settings like warehousing, manufacturing, and service robotics.

• →HANDOFF solves the command-space interface problem between task planning and whole-body control through an intuitive, modular design.
• →Multi-teacher knowledge distillation with context-conditioned gating creates a robust single controller handling motion tracking, locomotion, and fall recovery.
• →Hardware validation on Unitree G1 demonstrates state-of-the-art velocity tracking and expanded manipulation workspace capabilities.
• →Integration with VLM-driven planners enables natural-language task execution without controller fine-tuning or task-specific training data.
• →The approach prioritizes safety through filtered data and generalizes across diverse manipulation skills, advancing practical humanoid deployment.