This repository contains MAPF-GPT, a deep learning-based model for solving MAPF problems. Trained with imitation learning on trajectories produced by LaCAM, it generates collision-free paths under partial observability without heuristics or agent communication. MAPF-GPT excels on unseen instances and outperforms state-of-the-art solvers.

[AAAI-2024] Follower: This study addresses the challenging problem of decentralized lifelong multi-agent pathfinding. The proposed Follower approach utilizes a combination of a planning algorithm for constructing a long-term plan and reinforcement learning for resolving local conflicts.

The POGEMA Toolbox is a comprehensive framework designed to facilitate the testing of learning-based approaches within the POGEMA environment. This toolbox offers a unified interface that enables the seamless execution of any learnable MAPF algorithm in POGEMA.

[AAAI-2024] MATS-LP addresses the challenging problem of decentralized lifelong multi-agent pathfinding. The proposed approach utilizes a combination of Monte Carlo Tree Search and reinforcement learning for resolving conflicts.