A competitive mechanism for self-organized learning of sensorimotor mappings

How can a robot learn sensorimotor knowledge in a developmental way based on its own experiences solely? An important step is the acquisition of a body-schema-learning about the physical behavior of its own body, and how incoming sensory stimuli can be put in relation to the own body. In this work, we study how a competitive learning mechanism, which is related to the EM algorithm, can help to simplify the learning problem. We demonstrate how a robot can learn the way visual stimuli move as a consequence of the robots own actions of moving its camera or moving its end-effector in front of its camera. We show how the robot can discriminate stimuli originating from these two kinds of actions and learn the position of the end-effector in its visual input. Previous approaches have relied on a preprocessing step to “self-detect”, which we find is not necessary. The robot acquires a set of sensorimotor estimates, which could later be used, e.g. in visually guided reaching.