Spatial relational reasoning using graph neural networks: Learning representation, planning and search strategies

Recent advances in computer vision techniques have greatly extended the capabilities of robots to perceive objects in their environment. Nonetheless, robots still cannot match the ability of humans to make decisions and act in unstructured physical environments. This is particularly so where these environments are subject to unplanned changes over time. This thesis explores and proposes methods based on Graph Neural Networks, that an agent may use to act in an environment with dynamic spatial relationships that are subject to unexpected changes