Bootstrapping inverse kinematics with Goal Babbling

We present an approach to learn inverse kinematics of redundant systems without prior- or expert-knowledge. The method allows for an iterative bootstrapping and refinement of the inverse kinematics estimate. We show that the information structure induced by goal-directed exploration enables an efficient resolution of inconsistent samples solely from observable data. The bootstrapped solutions are aligned for a maximum of movement efficiency, i.e. realizing an effector movement with a minimum of joint motion. We derive and illustrate the exploration and learning process with a low-dimensional kinematic example and show that the same procedure scales for high dimensional problems, such as hyperredundant planar arms with up to 50 degrees of freedom.