Learning-based intelligent control and safety assurance of unmanned autonomous vehicles

Autonomous vehicles operating in unstructured, dynamic environments face safety-critical constraints under model mismatch, sensor noise, and real-time limits. This thesis develops NMPC-based control frameworks with embedded formal safety guarantees addressing: (i) feasibility—real-time solvable optimization; (ii) scalability—manageable computational demands for multi-agent systems; (iii) robustness—resilience to uncertainty including non-Gaussian disturbances. A data-driven approach identifies robot dynamics via Sparse Identification of Nonlinear Dynamics (SINDY), superior to NARX in noise robustness. Control Barrier Functions (CBFs) enforce collision avoidance as hard NMPC constraints, enabling computationally efficient short-horizon controllers with safety certification. Husky A200 experiments validate real-time navigation in cluttered environments. Relax-CBFs address the safety-feasibility trade-off through slack variables, improving trajectory smoothness near obstacles. This extends to multi-agent Autonomous Underwater Vehicle teams via distributed NMPC over bandwidth/latency-constrained networks, enabling coordinated formation control and obstacle avoidance. Distributional robustness against heavy-tailed uncertainty is achieved through Wasserstein-robust CVaR CBFs within discounted NMPC. Time-varying LQR feedback and Temporal-Difference learning enable online adaptation of NMPC weights and safety hyperparameters, demonstrating superior safety-performance trade-offs under parametric mismatch and actuator noise. This work provides a coherent framework from data-driven prediction with embedded safety through feasibility-aware scaling to learning-enhanced safety for autonomous vehicles. Thesis is embargoed until 31 July 2027.