Effective Learning and Online Modulation for Robotic Variable Impedance Skills

Learning robot manipulation skills from human demonstration is one of the key technologies for the cooperative robot. However, the traditional learning methods are difficult to be rapidly applied to complex force interaction tasks such as button-pressing, assembly, polishing, etc, which mainly focus on acquiring the movement skill by encoding the kinematics features and lacking consideration of the interactive force, stiffness, and other compliant profiles. Meanwhile, due to a large amount of high-dimensional demonstrations and task constraints, the existing methods are faced with the problems of low learning efficiency and difficult online modulation. To address those problems, a novel framework of robot skill learning is presented in this paper, which not only can simultaneously consider the movement and compliant profiles but also the implementation of effective learning and online modulation. Specifically, a Gaussian Mixture Model-Gaussian Mixture Regression (GMM-GMR) is first proposed for learning underlying dynamics from demonstrated trajectories, such that the skill can be formulated with two parameters at each time step, i.e. mean and variance. And then, a GMMGMR-based vector-valued Gaussian Process is proposed to autonomously modulate the trajectory according to the unexpected environmental changes and perturbations, such as unseen start-, via- or end-points. Those points are introduced by an external-force aware variable impedance controller by human intervention during skill reproduction. Consequently, a nonlinear mapping relationship between the variance and robot stiffness is investigated for encapsulating the compliant profiles. Experimental verification is performed on a button-pressing task with a Franka-Emika panda robot, results indicate that the presented framework can learn humanlike manipulation skills and reduce the complexity of robot application in complex force interaction tasks.