An Introduction to Graph Neural Networks for Localization and Mapping on Fixed Graph Structures

Graph neural networks have a variety of applications in different fields of science where graph theory has its application. One of such fields is indoor localization and mapping problems. One of popular methods in this field is pose-graph construction to describe mobile agent poses as the nodes and environment restriction through the edges. Another application in the context of pose-graphs is the pose-graph optimization which goal is to fit a pre-created pose-graph in such a way to represent environment restrictions more accurately. What is not covered enough in modern research is localization and mapping on known and fixed graphs performed only with objects on their nodes or edges without altering graph configuration in the process. The goal of this paper is to illustrate the applicability of graph neural networks to this class of problems and to provide an introductory description of such problems. In the paper, several graph convolutional neural network models are described. Then, the localization and mapping problems statements are discussed. The models are evaluated on a synthetic dataset.