Multi Robot Path Planning based on Reinforcement Learning

摘要

In order to improve the success rate of path finding in the MAPF(Multi Agent Path Planning) system and solving quality of large scale MAPF problems, this paper designs a multi robot deep reinforcement learning algorithm based on the ISAC(Independent Soft Actor Critic) mechanism to solve the multi robot path planning problem. In order to overcome the problem of difficult training of multi robot models, a deep neural network with a communication layer was designed based on the parameter sharing ISAC algorithm. ISAC was combined with IL(imitation learning) to construct the ISAC-IL algorithm. The simulation results show that compared with precise algorithms and heuristic algorithms, the ISAC-IL algorithm has a higher success rate, shorter completion time, and can calculate higher quality optimization solutions.