GNN-Based Scout Planning for Heterogeneous Robot Teams

A new framework known as Scout-Assisted Planning (SAP) has been introduced by researchers for diverse teams of robots. In this system, unmanned aerial vehicles (UAVs) actively collect environmental data to aid unmanned ground vehicles (UGVs) as they navigate through partially familiar terrains. This strategy mitigates the expensive backtracking that UGVs face when they encounter blocked paths that are only discovered through direct exploration. To enhance scouting efficiency, the team presents Information Gain-based Action Pruning, which evaluates potential scouting actions based on their anticipated effects on UGV operations. Due to the high cost of precise calculations, a Graph Neural Network (GNN) model is created to forecast information gain from the graph structure and belief state, allowing for real-time planning without compromising quality. Tests conducted in three different environments validate the effectiveness of SAP with GNN predictions. The research can be found on arXiv under reference 2605.22693.