首页|期刊导航|现代信息科技|基于强化学习的警用无人车路径规划研究

基于强化学习的警用无人车路径规划研究OA

Research on Path Planning for Police Unmanned Ground Vehicles Based on Reinforcement Learning

中文摘要英文摘要

为提升公安新质战斗力,针对警用无人车巡逻中的避障需求,提出一种基于Q-learning的路径规划方法.采用栅格法构建含圆形与矩形组合障碍的 20×20 栅格地图,设计 8 方向动作空间、ε-贪婪策略及差异化奖励函数,并引入基于目标距离的 Q 值表初始化方法以缓解地图扩大导致的计算量激增问题.在相同环境下与 A*、SARSA、人工势场法及标准Q-learning进行10次独立对比实验,结果表明,改进Q-learning平均路径长度为29.98±0.88,优于SARSA(31.76±4.49)、人工势场法(32.63)和标准 Q-learning(32.15±1.62),A* 算法路径最短(26.97),但其依赖完整先验地图信息,不具备在线学习能力;首次迭代步数仅 45±5 步,较标准 Q-learning 与 SARSA 降低约 96%;总训练步数较标准 Q-learning 减少约 28%,较 SARSA 减少约 70%.研究验证了该方法在警用无人车复杂巡逻环境路径规划中的可行性与有效性.

To enhance the new-quality public security combat effectiveness,this study proposes a Q-learning-based path planning method for obstacle avoidance in police Unmanned Ground Vehicle(UGV)patrols.A 20×20 grid map with circular and rectangular obstacles is constructed,featuring an 8-directional action space,the ε-greedy strategy,and differentiated reward functions.A target-distance-based Q-table initialization method is introduced to mitigate the computational surge caused by map expansion.Comparative experiments under the same environment(10 independent runs)against A*,SARSA,Artificial Potential Field(APF),and standard Q-learning show that the improved Q-learning achieves an average path length of 29.98±0.88,outperforming SARSA(31.76±4.49),APF(32.63),and standard Q-learning(32.15±1.62).The A* algorithm achieves the shortest path(26.97),but it relies on complete prior map information and lacks online learning capability.The first-iteration steps drop to 45±5,approximately a 96%reduction relative to standard Q-learning and SARSA.The total training steps decrease by approximately 28%and 70%relative to standard Q-learning and SARSA,respectively.These results validate the method's feasibility and effectiveness for complex police UGV patrol path planning.

俞正秋;李增

中国人民警察大学 研究生学院,河北 廊坊 065000中国人民警察大学 警务装备技术学院,河北 廊坊 065000

信息技术与安全科学

强化学习无人车路径规划警用Q值表仿真

reinforcement learningunmanned ground vehiclepath planningpolicingQ-tablesimulation

《现代信息科技》 2026 (18)

86-92,7

10.19850/j.cnki.2096-4706.2026.18.015

评论