Zoom link: https://upenn.zoom.us/j/93917708817

Recent advances in learning-enabled control have dramatically expanded the capabilities of autonomous systems, with applications ranging from robots and self-driving cars to industrial automation systems. However, reliable real-world deployment remains fundamentally constrained by the uncertainty in learned components, the complexity of large-scale systems, and the high cost of collecting system data. Toward addressing these limitations, this thesis advances the statistical foundations and control architectures underlying learning-enabled autonomy.

We begin with the problem of learning nonlinear systems, deriving finite-sample guarantees for the identification of time-varying systems and the selection of the model class from which the system is learned. We then establish statistical safety guarantees and certified mission completion for learning-enabled model predictive controllers under unknown system noise and stochastic external agents. Next, we broaden the scope from single-layer controllers to layered control architectures, co-designing layers operating at different timescales with progressively more detailed system models, while accounting for constraints, stochastic noise, and partial observations. Finally, we shift from a single control task to multitask architectures, where a common controller is learned over a distribution of related systems and control missions. For linear quadratic control, we establish performance and generalization guarantees for multitask policy optimization via novel system-theoretic measures of task heterogeneity.