This paper views the Multi-Agent Pickup and Delivery (MAPD) problem as an extension of Multi-Agent Path Finding (MAPF) and aims to address the performance degradation of learning-based methods in warehouse environments characterized by narrow aisles and long corridors. To achieve this, we define MAPF as a sequence modeling problem and demonstrate that sequence modeling-based pathfinding policies exhibit order-invariant optimality. We propose a Transformer-based Sequential Pathfinder (SePar) that reduces decision complexity through implicit information exchange while maintaining efficiency and global awareness. Experiments demonstrate that SePar outperforms existing learning-based methods on various MAPF tasks and variations, demonstrating excellent generalization to unseen environments. Furthermore, we highlight the importance of imitation learning in complex maps such as warehouses.
