Hierarchical Planning for Heterogeneous Multi-Robot Routing Problems via Learned Subteam Performance

This letter considersa particular class of multi-robot task allocation problems, where tasks correspond to heterogeneous routing problems defined on different areas a given environment. We present hierarchical planner that breaks down the complexity this problem into two subproblems: high-level allocating robots tasks, and low-level computing actual paths for each subteam. The uses Graph Neural Network (GNN) as heuristic estimate subteam performance specific coalitions tasks. It then iteratively refines estimates real performances solutions become availableon testbed having area inspection base task, we empirically show our is able compute optimal or near-optimal (within 7%) approximately 16 times faster (on average) than an baseline computes plans all possible allocations in advance obtain precise times. Furthermore, GNN-based estimator can provide excellent trade-off between solution quality computation time compared other (non-learned) estimators.