Data-Efficient Gaussian Process Online Learning for Adaptive Control of Multi-DoF Robotic Arms

摘要

Adaptive control depending on parametric adaptive elements has been shown to be effective in handling system uncertainty. However, it is restricted by the fixed prior knowledge on parametric models or in the face of stochastic uncertainty. Gaussian process (GP) is a typical Bayesian nonparametric model that contains fewer adjustable parameters and can describe stochastic uncertainty owing to introducing a probabilistic framework. This paper employs an online GP-based adaptive control framework for trajectory tracking of robotic arms with multiple degrees of freedom (DoFs). Besides, four GP learning schemes are compared during robot control, including a baseline scheme called full-state learning, and three data-efficient schemes: inward cascaded learning, outward cascaded learning, and the proposed coplane learning that is inspired by the robot structure. Simulations are carried out based on a 7-DoF collaborative robot called Panda. It is demonstrated that the proposed control framework performs well for the Panda robot, and the proposed coplane learning outperforms the other three learning schemes in terms of computational efficiency and modeling accuracy.