Learning Asynchronous Upper-body Task-space Trajectory Tracking Policy for Humanoid Robots

2026-06-24

AI Summary

This work addresses the temporal asynchrony and incomplete control structure between sparse, low-frequency task trajectories from high-level planners and high-frequency whole-body controllers by proposing an asynchronous upper-body task-space tracking framework. The approach leverages teacher-student distillation for policy initialization and introduces an asynchronous tracking mechanism that obviates explicit coordinate frame estimation. It further integrates model predictive control (MPC) post-processing with action-level and forward-kinematics-level self-guidance to effectively mitigate policy drift. By caching future trajectories, utilizing execution-time indexing, and employing sliding-window global reward training, inter-frame drift is significantly reduced. Experiments demonstrate superior low-frequency trajectory tracking performance in both simulation and on the Unitree G1 hardware, substantially outperforming synchronous and decoupled baselines while exhibiting strong out-of-distribution action generalization.

Abstract

High-level humanoid planners often output sparse task-space, low-rate trajectories, whereas whole-body controllers run at high frequency. This creates temporal asynchrony between the planning and execution, and structural incompleteness for full-body control. We propose an asynchronous upper body task-space tracking framework for humanoids. A student policy is initialized by teacher-student distillation, conditioned on the full cached future trajectory and an execution-time index, and trained with a sliding-window global reward to reduce frame drift without explicit frame estimation. For task-specific post-training, an MPC module completes sparse references into floating-base and upper-body guidance, while action- and FK level self-guidance constrain policy drift. Simulation and Unitree G1 hardware experiments show improved tracking under low update rates, stronger performance than synchronous and decoupled baselines, and safer adaptation to out-of-distribution motions.

Problem

• asynchronous control
• task-space trajectory tracking
• humanoid robots
• temporal asynchrony
• structural incompleteness

Innovation

• asynchronous trajectory tracking
• teacher-student distillation
• sliding-window reward
• MPC-guided policy refinement
• self-guidance constraint