Abstract: Aiming at the issue of low path planning efficiency for Automated Guided Vehicle (AGV) facing mixed U-shaped and dynamic obstacles,an adaptive Proximal Distillation Evolutionary Reinforcement Learning algorithm guided by Intrinsic Reward Policy (IRPA-PDERL).Initially,Random Network Distillation (RND) was introduced as intrinsic rewards in the fitness function to enhance the policy of diversity of elite policy evaluation.An adaptive intrinsic reward weight factor  was designed to balance the algorithm on the capacity to exploration and exploitation,aiding AGV in selecting the optimal strategy in highly uncertain dynamic obstacle environments.The covariance parameters of the mutation operator were optimized by covariance matrix adaption evolution strategy to decrease parameter sensitivity.Comparing with multiple algorithms on three different environments,the experimental results showed that the improved algorithm reduced paths by 12.65%,13.44% and 12.87% in U-shaped sub-assembly line,dynamic obstacle and transmission assembly line environments respectively,with faster convergence speeds than the original algorithms,and had strong robustness in complex environments.

Key words: U-shaped obstacles, evolutionary reinforcement learning, intrinsic reward, adaptive reward weights, path planning

摘要： 针对自动引导运输车(AGV)面对混合U型障碍与动态障碍时路径规划效率低的问题,提出一种基于内在奖励策略引导的自适应近端蒸馏进化强化学习算法。在适应度函数中引入随机网络蒸馏模型作为内在奖励,以提升算法评估精英策略的多样性;设计自适应内在奖励权重因子α平衡算法探索与利用的能力,帮助AGV在不确定性强的动态障碍环境中选择最优策略。将变异算子的协方差参数通过协方差矩阵自适应进化策略进行优化,从而降低参数的敏感程度。通过与多种算法分别在3种不同的环境中对比表明,改进后的算法在U型分装线、动态障碍、变速箱装配线环境中的路径分别减少12.65%,13.44%,12.87%,相比原始算法的收敛速度更快,且在复杂环境中具有较强的鲁棒性。

关键词: U型障碍, 进化强化学习, 内在奖励, 自适应奖励权重, 路径规划