Motion Repertory for a Legged Robot from a Reflective Walk

Abstract

In this paper, we propose a method of action based perception categorization for a legged robot to move to any destination in the environment. Once the robot acquires the relationship between actions and the changes of the view, it can generate the motion with respect to the observed scene. By utilizing this relationship, it reaches the desired position by feeding back the difference between the goal view and the current one. We call a group of motions based on the relationship a motion repertory. The motion repertory consists of sets of data of robot motion, which are compressed by the wavelet transform, and motion pattern of visual features. The validity of the method is shown by a preliminary experiment.