STF: Spatio-Temporal Fusion-Based Multi-Agent Path-Finding

Abstract

Multi-agent pathfinding (MAPF) plays a critical role in tasks involving collaborative multi-agent systems, particularly in complex environments, where it ensures the safe navigation of each agent while avoiding conflicts. However, existing methods still face numerous challenges, especially in high-density scenarios, where the risk of collisions between agents is higher, and myopic decision-making often leads to a decrease in both system efficiency and safety. To address these issues, we propose a spatio-temporal fusion-based (STF) MAPF method. In this method, we propose a spatio-temporal fusion module that effectively integrates spatial and temporal information, thereby mitigating myopic behavior and significantly reducing collision rates. Specifically, we design an attention-based spatial filter for the spatial observation encoder to process local spatial environment information and capture global spatial relationships within the environment. Additionally, we introduce a temporal future intention prediction module that provides the agents' short-term future intentions as temporal information. Experimental results demonstrate that STF significantly improves the safety and cooperation capabilities of agents in MAPF tasks, offering an effective solution to address the core challenges in MAPF.