Reproduce “Learning Dynamic Weight Adjustment for Spatial-Temporal Trajectory Planning in Crowd Navigation” on OpenResearch

About this paper

Muqing Cao, Xinhang Xu, Yizhuo Yang, Jianping Li, Tongxing Jin, Pengfei Wang, Tzu-Yi Hung, Guosheng Lin and 1 more

A hybrid crowd-navigation system uses reinforcement learning to adjust the objective weights of a spatial-temporal trajectory optimizer rather than directly generating control commands. Across simulated environments, it achieved 95–98% completion rates and fewer collision events than fixed-weight planning, while real-robot trials showed context-dependent changes in speed, clearance, and turning priorities.

Key insights

• Learning planner priorities can provide adaptive behavior while retaining the interpretability and dynamic structure of optimization-based trajectory generation.
• The learned policy increases human-clearance emphasis in dense crowds, static-obstacle emphasis near fixed hazards, and yaw-rate emphasis around corners; objective scales are also learned rather than assumed equal.
• The approach adapts to local context without exposing explicit goal coordinates, using the previous planned path as directional information in the observation.

Summary by alphaXiv.