A Variable Impedance Skill Learning Algorithm Based on Kernelized Movement Primitives

This article proposes a novel learning from demonstrations (LfD) method based on kernelized movement primitives (KMP). The original KMP algorithm is excellent at generalizing and handling high-dimensional inputs, but it is slightly inadequate in reproduction accuracy. To address this issue, we make two improvements to the KMP algorithm. First, a multivariate Gaussian process is employed to model the reference trajectory, which preliminarily improves the reproduction accuracy of KMP. Second, an optimization problem is formulated to learn the hyperparameters of the kernel function in KMP, which reduces the dependence on experience. We also propose a novel variable impedance control approach to tradeoff contact compliance against tracking accuracy by utilizing the probabilistic properties of the KMP. Comparative simulations and experiments are conducted to validate the proposed LfD algorithm.