Geometric disturbance observer based nonlinear model predictive control and trajectory planning in safe convex corridors for UAVs

This thesis develops and validates an integrated framework for UAV autonomy in complex environments, spanning disturbance-resilient control, trajectory planning in obstacle-dense environments, and a hardware-in-the-loop (HITL) simulation platform for validating planning and control algorithms. The thesis consists of three main components. First, a geometric disturbance observer-based nonlinear model predictive control (NMPC) architecture is presented for quadrotor trajectory tracking. The framework couples an extended-state extended Kalman filter formulated on the SO(3) manifold with a disturbance-aware NMPC that embeds six-degree-of-freedom disturbance estimates as time-varying inputs within the prediction model, enabling proactive compensation within the receding-horizon optimization. Second, a time-efficient trajectory planning framework for fixed-wing UAVs combines greedy corridor generation along bidirectional rapidly-exploring random tree path with differential-flatness-based convex optimization. The approach constructs a safe corridor of overlapping convex regions significantly faster than workspace-wide convex decomposition methods and formulates the complete planning problem as a single convex quadratic program, guaranteeing collision-free and dynamically feasible solutions without nonlinear refinement. Third, a comprehensive hardware-in-the-loop ii testing framework is developed for the DJI M300 RTK platform, integrating the DJI Manifold 2-G onboard computer with the flight controller through the DJI Onboard SDK and Robot Operating System. Simulation results demonstrate that the geometric disturbance observer-based NMPC achieves 60% reduction in position root mean square error compared to baseline NMPC under periodic six-degree-of-freedom disturbances. The proposed trajectory planning framework achieves a 36% reduction in traversal time with a 17� computational speedup relative to the Graph of Convex Sets (GCS) planner refined with Kinodynamic Trajectory Optimization (KTO) as the baseline method, and demonstrating a 99.4% success rate across 1,001 randomized scenarios. The hardware-in-the-loop simulation framework for the DJI M300 RTK drone successfully validates the implementations of proportional�integral�derivative (PID), linear quadratic regulator (LQR), and model predictive control (MPC) algorithms. Together, these contributions form a complete pipeline for robust UAV control and planning: from disturbance-resilient NMPC and convex optimization-based trajectory planning to pre-flight validation on real hardware, providing a foundation for safer autonomous UAV operations in uncertain and obstacle-dense environments.