Hierarchical Impedance, Force, and Manipulability Control for Robot Learning of Skills
Chao Zeng, Chenguang Yang, Zhehao Jin, Jianwei Zhang
Abstract
Learning from demonstration (LfD) has been considered an efficient way for skill transfer from a human user to a robot, to quickly program the robot to perform tasks. Current methods mostly focus on the learning and reproduction of one or several important manipulation features demonstrated by the human user in a specific task. Naturally, it is expected that robots could learn multiple manipulation features simultaneously from human demonstration toward a higher level of human-like dexterity. This work seeks to develop an LfD framework for learning human manipulation skills with multifeatured comprehensive features, mainly represented by motion trajectories, force profiles, and posture-dependent manipulability ellipsoids. To reproduce the learned features seamlessly, we first propose a hierarchical torque controller that assembles impedance control and manipulability control to achieve compliant trajectory-tracking and manipulability-tracking, respectively. We then compare the combinations of different impedance controllers and manipulability controllers concerning the tracking accuracies. Finally, within the hierarchical controller, we propose an optimization-based force control mechanism to reproduce the desired force profile, by adapting the desired stiffness online using quadratic programming, augmented with stiffness constraints for safety guarantee. For the evaluation part, the proposed framework is compared with the state-of-the-art in simulated tasks, demonstrating comparable performances and merits, and it is further verified in two real-world scenarios.