Learning-based robot control with localized sparse online Gaussian process

In recent years, robots have been increasingly utilized in applications with complex unknown environments, which makes system modeling challenging. In order to meet the demand from such applications, an experience-based learning approach can be used. In this paper, a novel learning algorithm is proposed, which can learn an unknown system model from given data iteratively using a localization approach to manage the computational costs for real time applications. The algorithm segments the data domain by measuring significance of data. As case studies, the proposed algorithm is tested on the control of the mecanum-wheeled robot and in learning the inverse kinematics of a kinematically-redundant manipulator. As the result, the algorithm achieves the on-line system model learning for real time robotics applications.