Abstract:Morphing aircraft have attracted attention due to the uncertainties faced in attitude control. Probabilistic Inference for Learning Control (PILCO) methods, which do not rely on accurate models and represent uncertainties probabilistically while optimizing strategies through data-driven approaches, have achieved preliminary progress in the control problems with systematic uncertainties. However, existing studies still face limitations when dealing with high-order servo models and complex operating conditions. This paper proposes two improved PILCO methods that address the degradation of control performance caused by high-order servo models by using a multi-step history policy based on non-Markov processes. Additionally, it enhances the robustness of the control system under wind disturbances through multi-task learning based cost inference of multi-surrogate model and gradient weighting approaches. Using a low-speed morphing wing aircraft as the test object, the effectiveness of the improved PILCO control method was validated through mathematical simulations when used independently and in combination.