PCA-based genetic operator for evolving movements of humanoid robot

This paper proposes a new genetic operator in order to evolve the humanoid movements, which is composed of principal component analysis (PCA) and descent-based local optimization with respect to robot dynamics. The aim of the evolution is to let humanoid robots generate human-like and energy-efficient motions in real-time. We first capture human motions and build a set of movement primitives. The set is then evolved to the optimal movement primitives for the specific robot, which contain its dynamic characteristics, by using an evolutionary algorithm with the proposed genetic operator. Finally, the humanoid robot can generate arbitrary motions in real-time through the mathematical interpolation of the movement primitives in the evolved set. The evolved set of movement primitives endows the humanoid robot with natural motions which require minimal torque. This technique gives a systematic methodology for a humanoid robot to learn natural motions from human considering dynamics of the robot. The feasibility of our genetic operator is investigated by simulation experiments in regard to catching a ball that a man throws of the humanoid robot.