Feature-Matched Robust Control Enabled by Data-Driven Dynamics Online Learning

This study proposes a robust adaptive control framework incorporating Gaussian process (GP)-based dynamics learning method. In this approach, the aerodynamic model is captured from real-time angular rate measurements, and four control-oriented features are extracted from the learned model for controller adaptation. In this proposed control framework, three innovations are achieved: 1) The control commands are evaluated based on the posterior mean function of dynamics model. 2) Lyapunov stability condition are derived, where the uncertainty upper bound can be estimated by the GP prediction variance. 3) Singular value decomposition (SVD) of the control effect uncertainty is applied to analyze the worst-case scenarios. Simulations on a Winged-Cone aircraft with 40% parameters offset has been conducted, verifying the effectiveness of the proposed control framework, which achieves significant performance improvements compared to active disturbance rejection controller.