Neural dynamics parametrically controlled by image correlations organize robot navigation

Abstract

Behavior-based robot designs confront the problem how different elementary behaviors can be integrated. We address two aspects of this problem, the stabilisation of decisions based on changing behavioral requirements and the fusion of multiple sources of qualitative sensory information. These issues are studied in the context of a vision-guided mobile robot that is endowed with the ability to reach a goal while it avoids obstacles. Behavior is organized from the “inside” of the robot. Even in the absense of external stimuli the internal dynamics generate behavior. By exploiting image correlations the visual sensors provide coarse qualitative estimates of spatial relations. These estimates are immediately coupled into neural dynamics realized by neural fields that generate obstacle avoidance and homing behaviors of an autonomous mobile robot. The background of the neural field approach is theoretical work on the function of cerebral cortex ([1]). Neural fields represent a state space approach to estimation and control. Convergent information cooperates and divergent information competes in shaping the stable attractor states. The states of the dynamical systems are on the one hand instrumental for the control of behavior on the other hand they provide concise hypotheses for the interpretation of sensory input. The concept of dynamic state deals effectively with contradictory information and assures that only behaviorally relevant information is extracted from the input. The navigation scheme works succesfully in real-time with image resolutions as poor as 322 pixels.