Mesbahi, MehranVasisht, Soumya2019-10-152019-10-152019-10-152019Vasisht_washington_0250E_20816.pdfhttp://hdl.handle.net/1773/44696Thesis (Ph.D.)--University of Washington, 2019Autonomous aerial robots provide new possibilities to study interesting phenomena and offer a unique vantage point for many surveillance and tracking tasks. Tracking a rogue or an unknown target is an important task in which an agent typically adopts a reactive strategy to the changes reflected in the target observations. As these aerial vehicles increasingly share airspace with fixed wing commercial airplanes, it has become critical to establish reliable, high quality tracking strategies. This work seeks to leverage the concepts of modern control theory, statistics and reinforcement learning to enhance traditional tracking control design strategies to achieve improved tracking performance. A data-guided approach is proposed which shows that embedding observation data in to the control loop improves tracking performance for certain classes of target systems. A comparative study of model-based and model-free approaches for tracking is presented in which an agent, guided by vision-based sensors, directly learns an optimal policy to track the unknown reference trajectory. In addition, a distributed framework is developed in which multiple agents perform consensus on the learned parameters to improve tracking accuracy. Numerical simulations are presented to validate this data-guided tracking scheme for a single agent and a network of agents.application/pdfen-USnoneControl SystemsLinear Quadratic RegulatorMachine LearningRoboticsUnmanned Aerial VehiclesAerospace engineeringRoboticsComputer scienceAeronautics and astronauticsThesis