Learning by Observation without Three-Dimensional Reconstruction

Abstract. This paper presents a basic skill of imitation learning from observation towards developmental cognitive robotics which aims at providing a new way to understand the cognitive process of human beings by realizing artificial systems based on the internal observer’s viewpoint. Unlike the conventional ones, the proposed method does not assume the global coordinate system in order to transform the observed motion in the demonstrator’s coordinate system to the learner’s one. Instead, the optic-geometrical constraint called “epipolar constraint” is used to reconstruct the view of the other agent, on which adaptive visual servoing is applied to imitate the observed motion. The experimental results and future issues are given.