Generation of human-aware navigation maps using graph neural networks

Minimising the discomfort caused by robots when navigating in social situations is crucial for them to be accepted. Graph Neural Net-works can process representations including arbitrarily complex relation-ships between entities such as human interactions. This is particularly interesting in the context of social navigation, where relational information should be considered. This paper presents a model combining Graph Neural Network (GNN) and Convolutional Neural Network (CNN) layers to produce cost maps for human-aware navigation in real-time. The model leverages the relational inductive bias of GNNs to generate scenario representations that can be efficiently exploited using CNNs. In addition, a framework to bootstrap existing zero-dimensional models to generate cost map datasets is proposed. The model is evaluated against the original zero-dimensional dataset and in simulated navigation tasks.The results outperform similar state-of-the-art-methods considering the accuracy for the dataset and the navigation metrics used. The applications of the proposed framework are not limited to human-aware navigation, it could be applied to other fields where cost map generation is needed.