人間の直接教示動作の統計的性質からのロボットへの作業スキル移植

—折り紙動作を題材として—

抄録

Although the demand to build a dexterous robot like a human, i.e., a robot that can execute a given task robustly against the variation of the environment, is increasing, it is difficult to implement an explicit sensory feedback law to adapt the variation of the environment. To solve the above problem, transferring human skills to robots has attracted attention in the robotics community and direct teaching is known as one of the powerful methods to transfer human skills to robots. We have already proposed a method to extract human skill automatically by using direct teaching. In this method, human skill is extracted in two aspects, i.e., appropriate nominal trajectories and sensory feedback laws. However, the proposed method didn't consider the direction of correlation between the force and the velocity and is limited to a specific DOF system with a specific number of sensory inputs. In this paper we extend the proposed method so that considered the direction of correlation between the force and the velocity by using canonical correlations and it can be applied to arbitrary DOF systems with arbitrary number of sensory inputs.