Attention-based Priority Learning for Limited Time Multi-Agent Path Finding.

Solving large-scale Multi-Agent Path Finding (MAPF) within a limited time remains an open challenge, despite its importance for many robotic applications. Recent learning-based methods scale better than conventional ones, but remain suboptimal and often exhibit low success rates within a limited time on large-scale instances. These limitations often stem from their black-box nature. In this study, we propose a hybrid approach that incorporates prioritized planning with learning-based methods to explicitly address these challenges. We formulate prioritized planning as a Markov Decision Process and introduce a reinforcement learning-based prioritized planning paradigm. In doing so, we develop a novel Synthetic Score-based Attention Network (S2AN) to learn conflict/blocking relationships among agents, and deliver blocking-free priorities. By integrating priority mechanisms and leveraging a new attention-based neural network for enhanced multi-agent cooperative strategies, our method enhances solution completeness while trading off scalability and maintains linear time complexity, thus offering a robust avenue for large-scale MAPF tasks. Comparisons demonstrate its superiority over current learning-based methods in terms of solution quality, completeness, and reachability within limited time constraints, especially in large-scale scenarios. Moreover, an extensive set of numerical results reveals superior completeness compared to restricted-time Priority-Based Search (PBS) and Priority Inheritance with Backtracking (PIBT) in medium to large-scale obstacle-dense scenarios.