Improving Multi-Agent Path Finding via Multi Visual Model and Channel

Multi-Agent Path Finding (MAPF) aims to plan conflict-free joint optimal paths effectively for a group of agents from their respective starting points to destinations. This task is challenging due to its computational complexity. Current works focus on reducing the number of potential conflicts to further decrease the time required for path planning. In this study, we integrate deep learning-based architecture with the sub-dimension expansion method, aiming to transform the multi-agent pathfinding task into an image classification task. We employ visual models as single-agent path planner to predict optimal path based on the observation space of each agent, which reduces the number of conflicts and improves the performance of the planner significantly. The experimental results show that, in comparison to Learning-Assisted M*, our method achieves a reduction in potential conflicts and implements MAPF task effectively only with a minor increase in time consumption.