Abstract: The current methods for human-robot skill transfer in robotics mainly focus on discrete point-to-point or kinematic aspects, while the learning of periodic dynamic skills has not been adequately considered. This article proposes a multi-space fusion skill learning framework based on rhythmic dynamic movement primitive (rDMP) for learning periodic impedance skills (including motion trajectories and variable stiffness features) from human demonstrations. The force perception information of the robot’s workspace is used to estimate the endpoint stiffness matrix. However, due to its symmetric positive definite (SPD) characteristics and the dependence of traditional rDMP on Euclidean space data, this research designs a Riemannian metric-based rhythmic dynamic movement primitive (RM-rDMP) skill learning method to adapt to the stiffness matrix information on the SPD manifold and adopts an adaptive oscillator estimation system for frequency and phase. By using an impedance control strategy based on the robot’s dynamics model, both motion trajectories and variable stiffness features are assigned to the robot to achieve variable impedance skill encoding. Simulation and experimental studies demonstrate that the proposed method successfully transfers human-like kinematic and dynamic skills to the robot and enables it to have reproduction and generalization capabilities. In summary, the proposed method includes an endpoint stiffness estimation model and a multi-space fusion framework for periodic impedance skill learning, which is applicable to human-robot skill transfer control methods that need to consider both Euclidean space position and Riemannian space stiffness matrix.

Key words: robot dynamics, periodic variable impedance skill, rhythmic dynamic movement primitive, Riemannian manifold, human-robot skill transfer