Replicating centralized MAPF solvers with MAPF-GPT

Determine how effectively MAPF-GPT, a decentralized imitation-learning policy for multi-agent pathfinding, can replicate the behavior of centralized MAPF solvers other than LaCAM, specifically the optimal Conflict-Based Search (CBS) algorithm.

Background

MAPF-GPT is trained via supervised imitation learning on a large dataset of expert demonstrations generated by the centralized LaCAM solver, producing a decentralized policy that uses only local observations without communication. The model achieves strong zero-shot performance across multiple benchmarks.

The authors note that, while effective when trained on LaCAM trajectories, it remains unclear whether MAPF-GPT can similarly imitate the behavior of other centralized MAPF solvers, such as Conflict-Based Search (CBS), which provides optimal solutions. Clarifying this would assess the generality of MAPF-GPT with respect to different expert policies and would guide data collection and training strategies.