Learning Compositional Hierarchies Of A Sensorimotor System

We address the problem of learning static spatial representation of a robot motor system and the environment to solve a general forward/inverse kinematics problem. The latter proves complex for high degree-of-freedom systems. The proposed architecture relates to a recent research in cognitive science, which provides a solid evidence that perception and action share common neural architectures. We propose to model both a motor system and an environment with compositional hierarchies and develop an algorithm for learning them together with a mapping between the two. We show that such a representation enables efficient learning and inference of robot states. We present our experiments in a simulated environment and with a humanoid robot Nao.