Teaching by demonstration on dual-arm robot using variable stiffness transferring

Teaching by demonstration (TbD) techniques have been extensively investigated in the recent decades to enable transferring various task skills from human to robots. The traditional TbD techniques focus on teaching motion trajectories that may be sufficient for routine tasks with fixed objects. While for interactive tasks in contact with dynamic environment and objects, e.g., the payload of a robot manipulator may change from one to another, teaching robot only by motion demonstration may cause undesired contact force and inefficiency in the task execution. In this paper, we present a novel TbD method enhanced by transferring the stiffness profile during human robot interaction (HRI). The method is developed on a bimanual robot, whereas one slave arm plays the role of the tutee, and the other master arm coupled with human demonstrator plays the role of tutor. A rendering algorithm is employed to provide demonstrator with force feedback via a purposely built coupling device according to the motion disparity between the two arms. The muscle surface electromyography (sEMG) signals collected during HRI is processed to extract the demonstrator's variable stiffness as well as hand grasping patterns. Comparative tests have been carried out on a bimanual Baxter robot for a lifting task with three different set-ups: i) TbD with predefined fixed stiffness; ii) TbD with demonstrator transferred variable stiffness without force feedback; and iii) TbD with demonstrator transferred variable stiffness with force feedback. Results show that the proposed TbD method performs best by transferring the demonstrator's physical interactive skill to the robot in a natural and efficient manner.