Dynamic Skill Learning From Human Demonstration Based on the Human Arm Stiffness Estimation Model and Riemannian DMP

Abstract

Traditional skill transfer frameworks mainly focus on kinematics, while there has been a lack of studies on dynamics. This article develops a Riemannian-based dynamic movement primitive (DMP) framework for learning and generalizing multispace data, including position, orientation, and stiffness from human demonstration. A simplified geometric configuration of the human arm skeleton is adopted to extract its endpoint stiffness. The dynamic skills of a variable stiffness are obtained in real time and then transferred to robots. The effectiveness of the presented approach is verified by two experiments in real scenarios on the Franka Emika Panda robot. The experimental results indicate that both kinematic and dynamic skills can be learned and generalized by the extended DMP framework with high accuracy and strong correlation. The human-like variable impedance control of a robot can be successfully realized by using the proposed approach. Thus, the proposed approach, including the stiffness estimation model and the skill learning and generalization framework, is suitable for applications of human–robot collaboration, contact operation, and teleoperation, where the position, orientation, and stiffness need to be considered simultaneously.