Learning of robotic assembly based on force information

Abstract

This paper treats a practical method to generate assembly strategies applicable to part-mating tasks that are of particular interest. The difficulties in devising reliable assembly strategies results from various forms of uncertainty such as an imperfect knowledge of the parts being assembled and limitations of the devices performing the assembly. Our approach to cope with this problem is to have the robot learn the appropriate control response to measured force vectors, that is, the mapping between sensing data and corrective motion of robot, during task execution. In this paper, the mapping is acquired by using a learning algorithm and represented with a binary tree type database. Remarkable features of the proposed method are the use of a priori knowledge and accomplishment of the task with little human trouble. Experiments are carried out by taking account of practical production facilities. It is shown by experimental results that an ideal mapping is acquired effectively by using the proposed method and the assembly task is carried out smoothly.