Data-driven Tuning of Autopilots with Fictitious Reference Iterative Tuning

Control parameter tuning is a well recognized challenge in the control field, especially in the absence of a priori knowledge on the system dynamics. Autopilots are a natural representation of this challenge, as they involve both parameter tuning of the autopilot laws and uncertainty of the vehicle dynamics. Building upon the Fictitious Reference Iterative Tuning (FRIT) idea, this work proposes an improved FRIT method and deploys it in autopilots of Unmanned Aerial Vehicles (UAVs). By introducing a novel data-driven cost based on mixed sensitivity design, the proposed FRIT method can handle multivariable loops with possible couplings among the various loops, which is a key aspect of autopilot architectures. The proposed FRIT does not require any a priori knowledge about the vehicle dynamics: it suffices to collect data during vehicle operation and exploit the available knowledge of autopilot control loops to perform a oneshot experiment to tune the control gains. Notably, our tests show that, even in the case that the control parameters are obtained using data collected in nominal operation, the corresponding controllers perform good also in non-nominal uncertain scenarios involving load changes or wind disturbances. The feasibility and robustness of the proposed approach is further validated via comparisons with standard FRIT methods and with a set of state-of-the-art control parameter tuning methods.