Gaussian Process Inverse Dynamics Learning for Variable Stiffness Actuator Control

The control of variable stiffness actuators (VSAs) is challenging because they exhibit highly nonlinear characteristics and are difficult to model accurately. In this study, we propose a machine learning-based tracking control approach combining Gaussian process (GP) learning and low-gain feedback control for VSAs subjected to unknown dynamics, where the GP model learns the inverse dynamics of agonistic-antagonistic (AA)-VSAs to feedforward control and provides the model fidelity by the predicted variance for the online adjustment of feedback control gains. It is shown that the tracking error is uniformly ultimately bounded and exponentially converges to a small ball under a given probability. Experiments on an AA-VSA named qbmove Advanced have validated the superiority of the proposed method with respect to tracking accuracy and generalization.