UAV Flight Control via Quaternion Neural Networks and Closed Loop Lyapunov Function

This work investigates the use of Quaternion Neu-ral Networks (QNNs) to control a hovering manoeuvre in a realistically simulated quadrotor model. The use of QNNs is well justified due to the nature of the signals to be controlled, which contain quantities that are directly accessible to quaternion computation. Using imitation learning with stability constraints, a neural network based controller is trained to mimic a linear quadratic regulator. A comparative analysis between traditional neural networks and QNNs is performed to develop a more robust controller. The quaternion-based representation can effectively handle multidimensional data and enables more accurate and robust attitude control of a quadrotor certified, via a numerical approach, through a neural-based Lyapunov function.