Developmental Human-Robot Imitation Learning of Drawing with a Neuro Dynamical System

Abstract—This paper mainly deals with influences of teach-ing style and developmental processes in learning model to the acquired representations (primitives). We investigate these in-fluences by introducing a hierarchical recurrent neural network for robot model, and a form of motionese (a caregiver’s use of simpler and more exaggerated motions when showing a task to an infants). We modified a Multiple Timescales Recurrent Neural Network (MTRNN) for robot’s self-model. The number of layers in the MTRNN increases according to learn complex events. We investigate our approach with a humanoid robot “Actroid ” through conducting an imitation experiment in which a human caregiver gives the robot a task of pushing two buttons. Experiment results and analysis confirm that learning with phased teaching and structuring enables to acquire the clear motion primitives as the activities in the fast context layer of MTRNN and to the robot to handle unknown motions. I