Over the past decade, Unmanned Aerial Vehicles (UAVs) have attracted great interest from the robotics and control communities, whose wide variety of applications ranges from prospecting and rescuing to transportation and delivery. These tasks are normally challenging to quadcopters regarding the controller design, the adaptation of unmodeled dynamics stemming from cluttered and unpredictable environments, and varying system dynamics. System identification via the first principles, as a part of the controller design procedure, often has to face the tradeoff between fidelity and simplicity. The such tradeoff can be easily observed on quadcopters which is a highly agile and nonlinear under-actuated system. To mitigate these challenges, my research aims to solve attitude controller design and motion planning problems by purely leveraging flight data.

A well-designed attitude controller is critical, as it is normally used to stabilize the inner loop of the UAVs. However, related prior works often ask for extra efforts on either tuning parameters, identifying models, or developing unique flight stacks. I hereby propose a data-driven framework for learning dynamics and designing attitude controllers that can be rapidly deployed on an 'empty' quadcopter, so as to expedite the entire control design process. Such approaches are also investigated on a quadcopter with varying dynamics, which can be commonly seen when the delivery quadcopter loads/offloads packages. 

The richness of the collected data determines the quality of the learned dynamics and thus the designed controller. However, it is extremely demanding to fully explore the space for high-dimensional systems. Therefore, I proposed a target-oriented control synthesis approach that enables efficient point-wise steering by learning from the data. When environmental disturbances are involved, a neural network-based external governor is proposed to improve the tracking performance, where the inner loop system is stabilized by a baseline controller that is designed based on my first topic.