From Echoes to Maps: Collaborative Learning of Unstructured Spaces with Multipaths and Graph Neural Network

Abstract

This paper proposes a multi-agent decentralized mapping algorithm for unstructured environments that utilizes distance measurements from a single antenna. This constraint necessitates robust data association to track consecutive non-line-of-sight (NLOS) paths. We address this with a framework that first employs graph matching to unambiguously identify consecutive NLOS paths, followed by a graph neural network (GNN) that estimates reflection points using the spatial-temporal graph structure. Evaluated in a complex curved wall scenario, our approach demonstrates robust mapping accuracy compared to benchmarks, including a state-of-the-art approach, and achieves real-time performance with a computational time of 0.014 second.