Development of Spatial Mental Representations in an Embodied Artificial Neural Network Using Sequential and Egocentric Data

The current study investigated the development of spatial relations in an artificial neural network. Design constraints and requirements for the artificial neural network were formulated from literature, in an attempt to make the network psychologically and neurobiologically plausible. Egocentric route information was taught to a network using back propagation. The network was embodied in a Lego Mindstorms robot. The (embodied) network successfully managed to navigate a learned maze. Using principal component analysis to investigate the representations the network built, components for direction and location were found. They hinted at preparation effects and the basis for an emerging allocentric representation. No evidence for the ability to find novel routes was found.