Using Self-Imitation to Direct Learning

Abstract

An evolutionary predecessor to observational imitation may have been self-imitation. Self-imitation is where an agent is able to learn and replicate actions it has experienced through the manipulation of its body by another. This form of imitative learning has the advantage of avoiding some of the complexities encountered in observational learning such as the correspondence problem. We investigate how a system using self-imitation can be constructed with reference to psychological models of motor control including ideomotor theory and ideas from social scaffolding seen in animals to allow us to construct a robotic control system. The system allows a human trainer to teach a robot new skills and modify existing skills. Additionally the system allows the robot to notify the trainer when it is being taught skills it already possesses. We argue that this mechanism may be the first step towards the transformation from self-imitation to observational imitation. We demonstrate the system on a physical Pioneer robot with a 5-DOF arm and pan/tilt camera which is taught using self-imitation to track and point to coloured objects