Abstract
It is essential for the robot manipulator to adapt to unexpected events and dynamic environments while executing the physical contact-rich tasks. Although various methods have been proposed to improve the adaptability and generalization of robot manipulation, it is still difficult to deal with complex contact-rich tasks, e.g., rolling pizza dough and robot-assisted medical scanning, without the assistance from a human in the loop. In this paper, we proposed a novel framework combining learning from demonstration (LfD) and human experience to enhance the safety and adaptability of the robot manipulation. In this framework, dynamic ovement primitive (DMPs) is employed for learning from demonstration, and a method of human correction is applied to update the pre-trained DMPs model in real time. We conducted experiments on the Franka Emika Panda Robot with pizza dough rolling tasks. The results show that the proposed method could effectively improve the performance of the physical contact-rich tasks, and the online learning and human correction method through teleoperation provides a potential solution for advanced interaction tasks with complex and dynamic physical properties.

This publication has 33 references indexed in Scilit: