Hybrid A*와 강화학습을 이용한 다중 에이전트 경로 계획

Abstract

As autonomous driving technology has recently expanded and increased, the need for a path planning technology considering interaction with surrounding agents is increasing. In addition, Multi-Agent path planning is essential when a large number of autonomous driving mobility drives in a narrow environment. In particular, Multi-Agent path planning requires collision avoidance algorithms for dynamic obstacles. Therefore, in this paper, we propose a decentralized Multi-Agent path planning algorithm considering their interactions with surrounding agents. In global path planning, we use Hybrid A* to allow creating physically feasible paths by considering the kinematic characteristics of mobility. In local path planning, we used Proximal Policy Optimization(PPO), one of the reinforcement learning methods, to learn interactions with surrounding agents. The proposed algorithm is trained and evaluated through Unity ML-Agents. As a result of the simulation, we find that the success rate increases due to fewer collisions with surrounding agents than the traditionally frequently used RRT*-based path planning methods.