A discrete computational model of sensorimotor contingencies for object perception and control of behavior

According to Sensorimotor Contingency Theory (SCT), visual awareness in humans emerges from the specific properties of the relation between the actions an agent performs on an object and the resulting changes in sensory stimulation. The main consequence of this approach is that perception is based not only on information coming from the sensory system, but requires knowledge about the actions that caused this input. For the development of autonomous artificial agents the conclusion is that consideration of the actions, that cause changes in sensory measurements, could result in a more human-like performance in object recognition and manipulation than ever more sophisticated analyses of the sensory signal in isolation, an approach that has not been fully explored yet. We present the first results of a modeling study elucidating computational mechanisms implied by adopting SCT for robot control, and demonstrate the concept in two artificial agents. The model is given in abstract, probabilistic terms that lead to straightforward implementations on a computer, but also allow for a neurophysiological grounding. After demonstrating the emergence of object-following behavior in a computer simulation of the model, we present results on object perception in a real robot controlled by the model. We show how the model accounts for aspects of the robot's embodiment, and discuss the role of memory, behavior, and value systems with respect to SCT as a cognitive control architecture.