ESE Ph.D. Thesis Defense: “Learning-based Model Predictive Control for Aerial Vehicles”

Learning-based model predictive control (MPC) is an increasingly prominent control paradigm in recent years. One primary approach in learning-based MPC is to leverage machine or deep learning tools to construct accurate dynamics models. These models are then deployed into optimization-based control schemes. Although these frameworks can potentially provide significant performance improvements, a number of key challenges need to be addressed before they can realize their full potential and prove reliable for real-world applications. In this defense, I highlight a set of frameworks that alleviate these challenges in a systematic and holistic manner.

The data-driven models within learning-based MPC frameworks aim to strike a balance between architectural complexity and expressiveness. However, many of these models either have elaborate architectures, or lack sufficient expressiveness to capture the intricate interactions and nonlinearities within the system dynamics. As the first contribution of this defense, I outline a novel learning-based MPC framework that addresses this balance. In particular, I show how an accurate and architecturally simple dynamics model can be constructed by combining prior knowledge of the system with a data-driven component. This combined model is lightweight, making it highly compatible with nonlinear MPC. Another challenge in learning-based MPC frameworks is that the dynamics models are typically learned offline. This implies that the models often fail to account for the dynamic uncertainties experienced by the system during deployment. Learning the models online is possible, but that typically requires substantial training data and computational resources. As the second contribution of this defense, I alleviate these shortcomings by developing an adaptive framework that boosts both sample and computational efficiency, along with dynamic uncertainty compensation in learning-based MPC. A third challenge of learning-based MPC frameworks, is that there are often little or no guarantees on the accuracy of the data-driven models. As the third contribution of this defense, I leverage recent uncertainty quantification techniques to extract uncertainty estimates for the model predictions, which are equipped with probabilistic guarantees. These estimates are shown to be not overly conservative and are incorporated within a robustification framework to enhance the robustness of the system.