Development of body-based spatial knowledge through mental imagery in an artificial agent

Distance perception for mobile agents is of great importance for safe navigation in unknown environments. Traditional methods make use of analytical solutions. Yet, according to some research hypothesis, distance perception is not the result of mathematical calculations, but an emergent consequence of an association process, where visual and tactile information acquire a central role. Designing models closer to natural cognition poses paramount challenges to artificial intelligence (AI), which call for a review of some of the foundations of current methods. Our work is framed in the embodied cognition paradigm, which highlights the importance of the body for the development of cognitive processes. We provide theoretical grounds and empirical evidence for an artificial account of distance perception through a multimodal association process. By learning multimodal sensorimotor schemes, an agent is capable of perceiving affordances related to distance perception without any non-body-based geometric knowledge. We let an agent interact with an environment cluttered with objects, while learning multimodal sensorimotor associations. The learned spatial relations are thoroughly characterized to show how the model depends on the agent’s specific sensorimotor capabilities. The system is tested in a passability experiment and a navigation task, showing the agent anticipates undesired situations using the learned model predictions.