A visual attention mechanism for autonomous robots controlled by sensorimotor contingencies

Robot control architectures that are based on learning the dependencies between robot's actions and the resulting change in sensory input face the fundamental problem that for high-dimensional action and/or sensor spaces, the number of these sensorimotor dependencies can become huge. In this article we present a scenario of a robot that learns to avoid collisions with stationary objects from image-based motion flow and a collision detector. Following an information-theoretic approach, we demonstrate that the robot can infer image regions that facilitate the prediction of imminent collisions. This allows restricting the computation to the domain in the input space that is relevant for the given task, which enables learning sensorimotor contingencies in robots with high-dimensional sensor spaces.