Abstract
Wind disturbances and faults in the construction of a quadrotor unmanned aerial vehicle (UAV) can cause errors in its steady state value and cause the UAV to drift over time. Civilian GPS has a low positional accuracy, meaning that a UAV can drift away from its target position without the GPS registering it, causing significant difficulties in high accuracy positional control. The work presented in this thesis outlines the design of a quadrotor UAV with a novel reinforcement learning based neural network controller that reads accelerometer data from several different sensors mounted on the UAV and trains it in a live environment using feedback from a camera imaging system. Results show that the neural network was able to train with the camera imaging system and improved the accuracy of its decision making by 43% after only ten training trials for the first trial, and by 14% for the second trial, after which damage to the quadrotor during the training process ended the experiment. This neural network shows promising applications for learning nonlinear dynamics and assisting in the control of robotics and mechatronic systems.