Simultaneously Learning of Motion, Stiffness, and Force From Human Demonstration Based on Riemannian DMP and QP Optimization

Abstract

In this paper, we propose a motion, stiffness, and force learning framework based on an extended dynamic movement primitive (DMP) and quadratic programming (QP) optimization. The objective is to learn kinematic and dynamic operational parameters from a one-shot human demonstration, through measurement and estimation of the motion, 3-dimensional (3-D) endpoint stiffness, and applied forces of the human arm during manipulation tasks. To this end, first, the framework features an extended DMP to model the motion, stiffness, and force variations in Cartesian space and 2-D sphere manifold. Second, to account for collected errors and human-robot operation gaps, a QP optimization is applied to fine-tune the desired position of the controller. Finally, we validate the framework through two experiments in real scenarios on the Franka Emika Panda robot. Experimental results show that the robot can not only inherit the variation laws of motion, stiffness, and force in the human demonstration, but also exhibit certain generalization capabilities to other situations. The framework provides a reference for robots learning multiple skills via a one-shot human demonstration, which finds great potential application in human-robot cooperation, contact-rich scenarios, and skillful operations, where the motion, stiffness, and applied forces need to be considered simultaneously. Note to Practitioners—Fast programming in robotics through skill transfer plays a critical role in next-generation robots entering ordinary people’s lives. Existing research focuses more on skill learning at the kinematic level and lacks on the dynamic level, such as stiffness and contact force. The goal of this paper is to propose a novel framework for robots learning of motion, stiffness, and force variations from a one-shot human demonstration, simultaneously. To this end, a Riemannian-based DMP method is employed to model the variation laws of motion, stiffness, and force in Cartesian space and 2-D sphere manifold, respectively. In this way, the learning module needs to be run only once, and the patterns can also be generalized to other targets without repeated robot teaching and additional time-consuming processes. To accurately reproduce the learned skills, a human-like motion/stiffness/force controller combined with QP optimization is investigated. In this paper, rather than identifying real environmental parameters, we directly use interacted forces during the human demonstration to represent environmental effects and employ QP to update the desired position in a limited range to account for collected errors and human-robot operation gaps. Experiments on button pressing and polishing tasks by the Panda robot have achieved very good results. The work of this paper lays a foundation for multiple skills learning from human demonstration (LfHD).