Abstract Addressing the three major bottlenecks in humanoid robot motion generation: motion discretization, strategy homogeneity, and dynamic response delay, this paper pioneers a deep reinforcement learning framework integrated with a sparse attention mechanism. Through curvature-adaptive hyperbolic space modeling, the range of human joint activities is transformed into a spatial metric tensor, breaking through the fundamental conflict between biological motion continuity and algorithmic discretization, with a 52% reduction in the standard deviation of trajectory curvature. Based on Mobius algebra, a solution space entropy optimization mechanism is constructed to maximize strategy entropy coverage in unstructured task spaces, resulting in an 85.7% increase in strategy diversity and an increase in the number of effective strategies to 28. A dynamic hybrid replanning architecture is designed, combining node self-motion optimization and greedy path reconnection dual-mode mechanisms, which compresses the trajectory replanning delay to 18.3 ms in a 7—Degree of Freedom (DOF) system, breaking through the industrial safety threshold of 25 ms. Experimental results demonstrate substantial improvements in trajectory similarity, dynamic obstacle avoidance success rates, and energy efficiency in both industrial assembly and household service scenarios. This technology reduces the accident rate in human–robot collaboration by two orders of magnitude, providing an intelligent control paradigm for the multi-billion-unit collaborative robot market that combines biological motion fidelity with millisecond-level safety response.
Chen et al. (2025) studied this question.