Bayesian Optimization-Based Efficient Impedance Controller Tuning for Robotic Interaction With Force Feedback

Abstract

Impedance control is widely used in manipulation tasks of robots for the purpose of modulating the contact force between the robot and the interaction environment. Currently, it is achievable to adapt the impedance control parameters to various interactive objects via optimization or learning approaches. However, the adapt efficiency is still one of the most important unsolved challenges. In this regard, this paper presents a Bayesian optimization-based efficient tuning strategy for the impedance controller. The interaction process from impedance control parameters to the produced contact force is modeled as a Gaussian Process and the expected improvement plus (EIP) acquisition function is employed as a sampling criterion to tune the impedance control parameters. For interacting with objects, the efficiency and performance of the proposed method are numerically verified and compared with two classical adaptive impedance control schemes. With a real-world robotic manipulator, it is demonstrated that thanks to the inherent data efficiency of the Bayesian optimization, our method is able to find suitable impedance parameters via as few as a dozen of interaction attempts. Finally, the robustness of the proposed method against prior data is discussed.