Effective robot skill synthesis via divided control

Abstract

Learning from demonstration is a powerful method for obtaining task skills, which aim to eliminate the need for explicit robot programming. Classically, the tasks are demonstrated to the robot by means of either recorded human motion, direct kinesthetic teaching or through manual interfaces, which may not be applicable for task that involve dynamics. In such cases, human-in-the-Ioop robot learning with anthropomorphic and intuitive tele-operation may be more suitable. In this paper, we propose a divide-and-conquer approach for human-in-the-Ioop robot learning framework to improve the efficacy of skill synthesis. Usually a straightforward division of control between the human and the robot for skill transfer can be designed for effective skill transfer. With such a division, not only the human learning is sped up, but also the design of the autonomous part of the control policy is simplified by exploiting the human capability to learn to adapt to robot operation. In this study, the proposed approach is realized by using the `ball swapping task' on an anthropomorphic robotic arm-hand setup, where the balls must be swapped over the fingers without being dropped. In the current implementation, the control is divided over the control of the arm and the hand, where the human learns to control the position and the orientation of the hand to swap the balls, while the hand runs a periodic finger movement autonomously. Our results indicate that complex autonomous policies can be easily obtained by distributing control over the human operator and the robot in a human-in-the-loop control setup. In particular, we show that the human operator quickly learns to control the arm in such a way that the simple finger movements of the hand become effective ball swapping actions. The combination of human and robot control then yields an autonomous ball swapping skill, which can be further improved for speed.