CoR-Lab

Abstract—We present a novel approach to learn and combine multiple input to output mappings. Our system can employ the mappings to find solutions that satisfy multiple task constraints simultaneously. This is done by training a network for each map-ping independently and maintaining all solutions to multivalued mappings. Redundancies are resolved online through dynamic competitions in neural fields. The performance of the approach is demonstrated in the example application of inverse kinematics learning. We show simulation results for the humanoid robot iCub where we trained two networks: One to learn the kinematics of the robot’s arm and one to learn which postures are close to joint limits. We show how our approach can be used to easily integrate multiple mappings that have been learned separately from each other. When multiple goals are given to the system, such as reaching for a target location and avoiding joint limits, it dynamically selects a solution that satisfies as many goals as possible. I