An empirical evaluation of learning-based multi-agent path finding algorithms in warehouse environments

In recent years, Multi-Agent Path Finding (MAPF) has become one of the most challenging and interesting fields in autonomous robotics and artificial intelligence. MAPF consists in computing collision-free paths for a group of agents that move from their initial locations to their goal locations in a shared environment. Many algorithms have been proposed to solve this problem using traditional search and planning approaches. The scarce scalability to hundreds or thousands of agents of some of these algorithms has recently pushed the community to investigate the use of Multi-Agent Reinforcement Learning (MARL) techniques for MAPF. Despite requiring extensive training, these learning-based approaches promise to scale better than traditional search and planning algorithms in complex environments, thanks to their decentralized execution. In this paper, we empirically evaluate and compare a representative sample of learning-based algorithms for MAPF, highlighting their strengths and weaknesses, also comparing them with traditional search and planning algorithms. Interestingly, while learning-based algorithms are usually trained and tested in randomly-generated environments, we test them in warehouse environments, to evaluate their practical applicability in realistic MAPF settings. Our results show that some learning-based algorithms nearly match the performance of search and planning algorithms in terms of path quality and show limited computing effort, proving their potential as a viable option for practical applications.