融合课程学习与自适应奖励塑形的移动机器人导航方法OA

Sparse rewards in end-to-end mobile robot navigation often lead to low learning efficiency and unstable convergence in conventional reinforcement learning methods.This study developed a deep reinforcement learning approach that integrated curriculum learning and adaptive collision entropy reward shaping.It introduced collision entropy derived from 2D LiDAR as a dynamic shaping term to quantify environmental uncertainty and guide obstacle-aware exploration.An adaptive scheduling mechanism adjusted the shaping coefficient β according to the slope of success rate variation,which balanced exploration and convergence during training.A staged curriculum progressively increased environmental difficulty to stabilize learning.Simula-tion results show that the proposed method achieves an 84.5％success rate and reduces the collision rate to 14.5％in dense environments.The method effectively alleviates sparse reward issues and improves navigation performance.