Self-localization of autonomous robots by hidden representations

We present a framework for constructing representations of space in an autonomous agent which does not obtain any direct information about its location. Instead the algorithm relies exclusively on inputs from its sensors. Activations within a neural network are propagated in time depending on the input from receptors which signal the agent's own actions. The connections of the network to receptors for external stimuli are adapted according to a Hebbian learning rule derived from the prediction error on sensory inputs one time step ahead. During exploration of the environment the respective cells become selectively activated by particular locations and directions even when relying on highly ambiguous stimuli.