Mixture of experts improves multi-agent pathfinding on diverse maps
Map-Distribution Mixture-of-Experts for Multi-Agent Pathfinding
Artificial Intelligence
Summary
Figuring out how groups of robots can move efficiently without bumping into each other is a tricky problem. This paper looks at combining several specialized robot 'brains,' each good at different types of maps, and decides which to listen to based on where a robot is on the map. The researchers show this mixed approach can do better than many existing methods, though sometimes a single well-trained brain still performs best. This insight helps design smarter robot teams for places like warehouses.
What this means in practice
- •For warehouse logistics engineers: Improve robot fleet navigation by selectively combining specialized route planning models depending on layout types encountered.
- •For autonomous delivery fleet managers: Enhance path planning reliability in complex environments by integrating multiple expert models tuned for varied map distributions.
Authors
Fumimaro Suzuki, Keisuke Okumura
Abstract
Designing capable neural policies for multi-agent pathfinding (MAPF), which map local observations to action distributions, represents an important challenge in multi-agent learning, with practical relevance to scalable collective automation problems such as warehouse logistics. In contrast to prior studies focusing primarily on individual policy design, this study explores a meta-level Mixture-of-Experts (MoE) approach to learning-based MAPF. Specifically, given a collection of compact yet high-performing experts specialised for different map distributions, our architecture learns a router that selectively fuses their outputs according to each agent's local situation. We instantiate and validate this MoE approach with a commonly used graph neural network-based policy as well as a newly developed recursive action-token refinement model. Combined with supervised and reinforcement learning fine-tuning of the experts, our experiments show that MoE generally improves upon the pretrained policy and outperforms existing leading methods. At the same time, our findings reveal that appropriately trained, parameter-matched monolithic models still outperform MoE, highlighting a nuanced role for map-distribution MoE for MAPF.