基于改进MADDPG的多机器人路径规划方法研究  

Research on Path Planning Based on Improved MADDPG for Multi-Robot System

作　　者：贾思雨 毕凌滔 曹扬 吕乃冰 JIA Si-yu;BI Ling-tao;CAO Yang;LV Nai-bing(Beijing Aerocim Technology Co.,Ltd,Beijing 102308,China;Beijing University of Technology,Beijing 100124,China)

机构地区：[1]北京航天晨信科技有限责任公司,北京102308 [2]北京工业大学,北京市100124

出　　处：《计算机仿真》2024年第8期458-465,共8页Computer Simulation

摘　　要：为完成未知环境下救援物资的运输任务,研究了基于深度强化学习的多机器人路径规划方法。首先选用多智能体深度确定性策略梯度算法MADDPG算法为基础算法,然后针对算法存在的收敛速度慢甚至不收敛问题,引入了碰撞发生区域重点训练、经验池分离机制和优先经验回放等改进措施。最后基于Gazebo三维仿真平台搭建了两种仿真环境,并从训练结果和测试结果两个方面对改进算法与原始算法进行了对比分析。实验结果显示:改进算法相比于原始算法在各仿真环境中的任务成功率分别提高了21%和32%,平均路径长度分别缩短了12%和17%,这证明了改进算法可以有效提高算法的收敛速度以及机器人的避障能力,从而更好地应用于多台物资配送车辆的路径规划。In order to complete the transportation task of rescue materials under an unknown environment,a multi-robot path planning method based on deep reinforcement learning is studied.Firstly,the Multi-Agent Deep Deterministic Policy Gradient(MADDPG)is selected as the research algorithm.At the same time,in view of the slow convergence rate or even non-convergence problem of the algorithm,some improvement measures are introduced,such as focus training in collision occurrence area,experience pool separation mechanism and prioritized experience replay mechanism.Finally,two simulation environments are built based on the Gazebo 3D simulation platform,and the improved algorithm is compared with the original algorithm from two aspects of training results and test results.The experimental results show that compared with the original algorithm,the task success rate of the improved algorithm in the two simulation environments is increased by 21% and 32% respectively,and the average path length is shortened by 12% and 17% respectively,which proves that the improved algorithm can effectively improve the convergence speed of the algorithm and the obstacle avoidance ability of the robot,so as to better realize the path planning of multiple rescue vehicles.

关 键 词：多机器人路径规划 深度强化学习 多智能体深度确定性策略梯度 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]