Memory-based Simultaneous Learning of Motor, Perceptual, and Navigation Skills

A learning approach and associated experiments are described that allow robotic devices to learn the relationship between control actions and sensory feedback, to acquire skills, and to adapt to a changing environment. The approach frames the motor skill learning problem as one involving the co-evolution of perceptual and actuation structures. A memorybased learning approach is taken and a Cached Event String Automata (CESA) is introduced. The approach supports the learning of complex sensory-motor skills from a very small number of trials with incremental skill refinement. A key aspect of the learning approach advocated in this paper is the use of self-motivated exploration to refine the learning of sensory-motor skills.