UC Davis

Multirotor unmanned aerial vehicles (UAVs) have gained significant popularity in recent years due to their high maneuverability and vertical take-off and landing capability. The new roles require that the future multirotor UAVs will need to fly in a variety of challenging environments and the flight performance may significantly degrade due to the shift from nominal flight conditions. As their usage expands to increasingly challenging environments, the need for reliable and high-performance flight behavior becomes more pressing. The dissertation addresses these difficulties through a series of research efforts aimed at improving the overall flight performance of multirotor UAVs in challenging conditions.

First, an experimental study was conducted to identify a data-driven ground effect model for a small quadcopter, which takes into account the interference among the rotors and was validated through flight experiments. An adaptive control scheme was then developed to counter the model uncertainty resulting from the complex aerodynamics, leading to improved command tracking performance when the UAV is in the ground effect region. The effectiveness of the developed controller was demonstrated on a real quadcopter, with results showing superior performance compared to a traditional PID controller.

Second, the effect of wind on a hovering octocopter was investigated and modeled through field experiments. A data-driven approach was used to model the wind effects on the bare airframe by directly measuring the wind and including it as a control input. A state space model that explicitly considers the wind effect was identified from real flight data using a system identification approach. The validation results show that a significant error reduction can be achieved by considering wind effects and adding a correction term. The identified model can serve as a foundation for the future development of model-based controllers for outdoor multirotor aircraft, enhancing their flight performance in windy conditions.

Lastly, a vision-based control solution was developed in order to navigate the UAVs inside complex, unstructured, and GPS-denied environments. The proposed solution leverages imitation learning and a variational autoencoder neural network to enable the autonomous agent to learn reactive strategies from human experience effectively and efficiently. The learning frame- work and the developed controller were demonstrated in simulated riverine environments first and then validated in a real orchard on a custom-built quadcopter, with results outperforming existing baseline algorithms. The proposed vision-based control solution is expected to significantly enhance the performance of multirotor UAVs in complex and GPS-denied environments, where traditional navigation methods may not be applicable.