Improving Robustness and Safety of Planning and Control Algorithms for Small Uncrewed Aircraft

Abstract

Autonomous systems operate through multiple interacting components to perform tasks such as perception, planning, and control. This dissertation develops and applies methods to quantify performance and robustness across these components, focusing on distributed optimization algorithms, motion planners, and learning-based controllers for small fixed-wing uncrewed aircraft systems. First, distributed optimization algorithms are analyzed as foundational elements of networked systems. Extending consensus and subspace-constrained methods, this work applies the integral quadratic constraint framework to characterize worst-case convergence rates and robustness under stochastic disturbances. Next, the dissertation addresses safety-critical motion planning. Using motion primitives derived from kinematic and experimentally identified nonlinear models, planners are designed with provable guarantees of obstacle avoidance under certain assumptions. Formal verification tools, including branch-and-bound solvers and satisfiability modulo theories frameworks, are used to prove safety and guide design choices, with results validated in simulation and flight experiments. Finally, learning-based controllers are developed for robust path-following under model uncertainty. Adversarial reinforcement learning is applied to train controllers against a learnable adversary that perturbs aerodynamic coefficients, producing policies that generalize effectively to aerodynamic modeling errors within the considered bounds. Additionally, a hypernetwork-conditioned reinforcement learning approach is introduced to enable adaptation to parameterized actuator faults, improving robustness and generalization to unforeseen time-varying failure modes. Collectively, these contributions provide methods to characterize and improve the robustness of individual components, offering a foundation for future efforts to achieve system-level guarantees.