A Survey on Reinforcement Learning-Based Multi-Agent Path Planning

With the advancement of artificial intelligence (AI) and Internet of Things (IoT) technologies, industrial systems have witnessed a growing demand for intelligent management. Multi-agent systems (MAS), being applied across various domains, not only significantly enhance the autonomy and efficiency of intelligent systems but also demonstrate irreplaceable potential in hazardous and complex environments where mobile robots may even prove indispensable. Deep reinforcement learning (DRL) has emerged as one of the most effective solutions for multi-agent path finding (MAPF) problems, having successfully achieved global path planning in known environments. However, its performance remains unstable in map-free scenarios, a persistent challenge that has become a focal point in recent MAPF research. This paper synthesizes contemporary studies to provide an overview of current research progress in MAPF. We systematically categorize existing algorithms into centralized and decentralized frameworks, tracing their evolutionary trajectories. Through established evaluation metrics, the paper analyze and compares mainstream algorithms while summarizing their developmental processes. Furthermore, this work examines critical challenges in DRL-based MAPF implementations and proposes potential directions for future research, which may focus on enhancing environmental adaptability, optimizing multi-agent coordination mechanisms, and improving real-time decision-making capabilities in dynamic scenarios. The discussion aims to provide theoretical references for advancing intelligent path planning in complex industrial applications.