Date of Award
8-19-2024
Degree Name
Doctor of Philosophy (PhD)
Degree Type
Dissertation
Abstract
Unmanned Aerial Vehicles (UAVs), initially developed for military applications, have now become indispensable tools across various civilian sectors. To accomplish flight tasks efficiently and effectively, two aspects of design are necessarily involved. At the low level, the intricate controller governs the UAV's immediate response to dynamic inputs and disturbances, ensuring stability, maneuverability, and precise trajectory tracking. Meanwhile, at the high level, strategic motion planning algorithms coordinate complex flight trajectories, taking environmental constraints, mission objectives, and dynamic obstacles into account. Recognizing the challenges posed by UAV dynamics and environment, this thesis deals with controller design and motion planning by harnessing available data resources. The first part of the thesis emphasizes the design of data-driven controllers, leveraging abundant data and computational power to synthesize controllers efficiently. While model-based learning techniques have shown robustness and scalability, they still face challenges such as extensive training process and heavy customization of flight stacks during implementation and model exploitation at the outset of learning. To overcome these issues, two novel approaches are proposed: rapid attitude controller design enabled by flight data and sequential model learning and control synthesis. Both approaches are demonstrated to be data-efficient and computationally lightweight. The second part of the thesis targets specific flight scenarios: navigating through windy environments and coordinating multiple UAVs in highly constrained spaces, commonly encountered in outdoor operations. Existing motion planners suffer from suboptimal performance and limited scalability. To address these challenges, we introduce two innovative data-driven approaches tailored to each scenario. Firstly, high-fidelity wind simulation results, derived from information sampled at key locations, are seamlessly integrated into the drone's motion planning process to optimize trajectories in windy conditions. Secondly, by centralizing all drones' original motion information, our planner efficiently handles prioritized multi-agent planning, treating temporal collisions as spatial obstacles. All proposed approaches have been deployed on physical quadcopter platforms, demonstrating their practicality and efficacy in real world settings.
Language
English (en)
Chair
Shen Zeng