Optimizing Priority-Based Search for Lifelong Multi-Agent Path Finding

Abstract

The lifelong Multi-Agent Path Finding (MAPF) problem requires planning collision-free trajectories for agents operating continuously in dynamic environments. Traditional solvers such as Priority-Based Search (PBS) use fixed branching heuristics, which can be inefficient in high-congestion scenarios. This work explores how learning-based methods can improve PBS decision-making. We develop supervised learning (SL) policies trained from high-quality beam search trajectories and reinforcement learning (RL) policies learned directly through simulation, enabling adaptive branching strategies. Evaluations on warehouse-style and Kiva-style maps with varying agent densities show that learned policies can significantly boost throughput in congested warehouse layouts, while identifying scenarios where classical heuristics remain competitive. Our findings provide guidance on solver selection based on environment layout and congestion characteristics.