Centralized Stochastic Multi-agent Pathfinding under Partial Observability

Multi-Agent Pathfinding (MAPF) is the problem of finding paths for multiple agents where each agent aims to reach a given goal location without conflicting with the other agents. In MAPF applications with physical robots, we can expect the agents to have stochastic behavior and imperfect localization. Planning for such centrally-controlled agents can be viewed as a special case of Partially Observable Markov Decision Process (POMDP), but off-the-shelf POMDP solvers cannot scale to plan for even a very small number of agents, due to the exponentially large size of the state and action spaces. Instead, we propose the Online Prioritized Planning (OPP) approach, where each agent computes and follows its individually-optimal policy until a potential conflict is detected. OPP resolves detected potential conflicts by replanning online for a subset of the agents so as to avoid positions that are potentially occupied by other agents. We describe how OPP can be implemented and propose two extensions that encourage the agents to leverage localization actions when needed. We evaluate OPP and its extensions empirically to highlight the pros and cons of our approach and show it can scale better than an offline baseline.