Graph Neural Networks for Mobile Robots: A Systemic GNN Design Solution for Traffic in AGV Systems

Abstract

Automated Guided Vehicle (AGV) systems improve modern warehouse efficiency but require extensive effort in designing the virtual road networks (also known as layouts). A key evaluation metric in this process is waiting time. Traditional simulation-based methods for waiting time estimation are time-consuming, high lighting the need for faster predictive models. In this thesis, we explore Graph Neural Networks (GNNs) for predicting the waiting time on each road segment. We propose a hierarchical GNN framework that integrates a classifier to detect con gested segments and a regressor to estimate the waiting time, effectively addressing the wide-spreading data imbalance issues. Experimental results demonstrate that the framework captures meaningful patterns, providing a potential alternative to traditional simulations in layout design.