Opportunities for graph learning in robotics

In the last few years, robotics highly benefited from the use of machine and deep learning to process data stream captured by robots during their tasks. Yet, encoding data in grids (images) or vectors (time-series) significantly limits the type of data that can be processed to euclidean only. To unlock the potential of deep learning also to unstructured data, such as point clouds or functional relations, a rising - yet under-explored - approach lies on the use of graph neural networks (GNNs). With this manuscript, we intend to deliver a brief introduction to GNNs for robotics applications, together with a concise revision of notable applications in the field, with the aim of fostering the use of this learning strategy in a wider context and highlighting potential future research directions.