From One to Many: Adaptive Multi-Agent Pathfinding in Heterogeneous Environments

Multi-agent pathfinding (MAPF) is a common abstraction of multi-robot trajectory planning problem, where multiple homogeneous robots simultaneously move in the shared environment. This paper addresses the heterogeneous MAPF problem, where a group of adaptive agents interacts with other agents (called impostors) that behave differently. The task remains cooperative, all agents should have the opportunity to reach their goals. We investigate how homogeneous methods can be enhanced for heterogeneous settings through three distinct approaches: planning-based, sampling-based, and learning-based methods. Our experimental framework employs the POGEMA benchmark to evaluate adaptive agents interacting with impostors following different policies (A* and PIBT). Our results demonstrate that all methods show significant performance improvements primarily with large agent populations, where frequent encounters with impostors necessitate conflict resolution. These findings indicate that while predictive modeling can enhance non-specialized algorithms when online training is impractical, learning-based methods offer superior adaptability to novel agent types in dynamic heterogeneous environments.