Multi-CrowdSurfer: A Position Paper on Applying Single-Agent Generative Priors to Decentralized MAPF
Accepted at AAAI Workshop on Multi-Agent Path Finding, 2026
Decentralized MAPF planners often suffer from myopic failures. We propose a new direction: equipping agents with a sophisticated, generative local planner, CrowdSurfer(IEEE ICRA 2025), validated in the single-agent crowd-navigation domain. This paper investigates the viability of this domain transfer. We hypothesize that its multi-modal, long-horizon nature could reduce non-symmetric conflicts. However, we explicitly acknowledge the domain gap: navigating stochastic crowds is not the same as coordinating with rational agents. We also identify a critical open question: can such a planner, based on implicit coordination, resolve the symmetric deadlocks that plague MAPF? This position paper explores this question and welcomes discussion from the community on this conceptual proposal.
Recommended citation: Antareep Singha, Laksh Nanwani, Arun Kumar Singh, K. Madhava Krishna.
Download Paper | Project Website