Skill Transfer between Industrial Robots by Sparse learning

Abstract

Recently, by increasing the productivity of industrial manufacture, industrial robots have played a key role in many fields of industry (e.g. automobile production, food production, etc.) However, there are two problems rarely mentioned in this field. First, compared with automatization in other fields, industrial robots are programmed manually by a human operator. Second, because of the physical difference between robots and difference of operating platform, there doesn’t exist a general method to define the skill (motion records) of robots and make it possible to reuse the skills between robots. In this work, we are trying to propose a skill definition of transfer system which combine the strengths of traditional DMP algorithm and deep learning method. Specifically, in our method, a set of motion primitive bases are generated from motion records in different robots. Skills are re-defined by the linear coefficient of the primitive bases and transferred based on motion primitive bases translation between different platforms. Experiment shows that our method can successfully transfer skills between different models with less space requirement.