🤖 AI Summary
Multi-agent path finding (MAPF) suffers from exponential growth in computational and memory overhead as agent count increases, severely limiting scalability in complex scenarios. To address this, we propose the first general-purpose hierarchical decomposition framework that automatically partitions large-scale MAPF instances into conflict-aware, approximately independent subproblems—compatible with seven mainstream solvers (e.g., A*, CBS, ICTS) without modifying their implementations. Our method combines graph partitioning with conflict-driven subproblem decomposition and employs a robust solution merging mechanism, achieving decomposition in ≈1 second. On standard benchmarks, it reduces memory consumption by >60% on average and significantly lowers runtime; serial solvers handle 2–3× more agents while maintaining <1% solution loss. The implementation is open-sourced.
📝 Abstract
Multi-agent pathfinding (MAPF) holds significant utility within autonomous systems, however, the calculation and memory space required for multi-agent path finding (MAPF) grows exponentially as the number of agents increases. This often results in some MAPF instances being unsolvable under limited computational resources and memory space, thereby limiting the application of MAPF in complex scenarios. Hence, we propose a decomposition approach for MAPF instances, which breaks down instances involving a large number of agents into multiple isolated subproblems involving fewer agents. Moreover, we present a framework to enable general MAPF algorithms to solve each subproblem independently and merge their solutions into one conflict-free final solution, and avoid loss of solvability as much as possible. Unlike existing works that propose isolated methods aimed at reducing the time cost of MAPF, our method is applicable to all MAPF methods. In our results, we apply decomposition to multiple state-of-the-art MAPF methods using a classic MAPF benchmarkfootnote{https://movingai.com/benchmarks/mapf.html}. The decomposition of MAPF instances is completed on average within 1s, and its application to seven MAPF methods reduces the memory usage or time cost significantly, particularly for serial methods. Based on massive experiments, we speculate the possibilty about loss of solvability caused by our method is $<$ 1%. To facilitate further research within the community, we have made the source code of the proposed algorithm publicly availablefootnote{https://github.com/JoeYao-bit/LayeredMAPF/tree/minimize_dependence}.