Koopman-Based Robust Learning Control With Extended State Observer

摘要

A key challenge in data-driven robot control is enabling robots to autonomously gather the most informative data during training while maintaining robust performance when deployed in new tasks or encountering unknown external disturbances. In this paper, we propose a robust active learning (RAL) control method designed to optimize data efficiency during model learning while ensuring robust and precise control during task execution. This approach integrates Koopman-based modeling with an active learning algorithm to enhance model learning efficiency, and an extended state observer (ESO)-assisted tracking control to ensure precise robot position control in the presence of unknown disturbances. The effectiveness of the proposed method is validated through various simulations and experiments, demonstrating significant improvements in data efficiency and robustness against unknown disturbances.