Advances in Multi-Agent Path Finding

Licentiatavhandling, 2025

Multi-Agent Path Finding (MAPF) addresses the collision-free coordination of multiple agents moving through a shared environment, finding applications in warehouses, airports, and video games, to name a few. Classical MAPF assumes discrete time and agents with pre-assigned goals, yet real-world scenarios demand continuous-time operations with dynamically arriving tasks. This thesis explores the techniques employed by scalable discrete-time MAPF and Lifelong MAPF (LMAPF) solvers; the challenges that arise from extending LMAPF to continuous time (LMAPF_R); and how MAPF in continuous-time (MAPF_R) can be solved for optimal solutions. We find that many scalable MAPF methods rely on prioritized planning, windowed planning, and dimensional simplifications. These insights culminate in a method that solves LMAPF for hundreds of agents with competitive throughput. Furthermore, addressing LMAPF_R presents challenges related to asynchronous movements, a dense search space, and volumetric agents complicating real-time collision detection and resolution. Our proposed algorithm is, to our knowledge, the first to address LMAPF_R, capable of planning up to a thousand agents in real time. Finally, Continuous-time Conflict-Based Search (CCBS) was thought to be the sole MAPF_R method for optimal solutions. Recent work suggests otherwise. We explain why CCBS could, and provide experimental evidence that CCBS does, return sub-optimal solutions. Consequently, we present a new MAPF_R solver - to our knowledge, the first to provably return optimal solutions in finite time. This work advances both the practical applicability and theoretical foundations of MAPF, bridging the gap between discrete-time abstractions and continuous-time realities.