This article presents an approach to the mathematical model of a fixed wing unmanned aerial vehicle prototype. The model is split in two different parts, related to the longitudinal and lateral stability, respectively. For this, Newton-Euler formulation is used as well as basic aerodynamic theory. Aerodynamic coefficients, inertias and characteristic points of the aircraft are obtained through simulations with an open-source software called XFLR-5, and the physical parameters of the model match the prototype's. Then, a longitudinal control strategy describes the altitude control in a cascade architecture, whose inner loop conveniently manoeuvres the pitch angle by acting on the symmetric flag deflection. Frequency domain techniques are used to design PID controllers.

In this work, we apply artificial intelligence to guide a drone to a certain point autonomously. Unreal engine creates a virtual environment where the drone can fly, and the algorithm is trained simulating the drone dynamics thanks to Airsim plugin. The implemented algorithm is Asynchronous Actor-Critic Advantage (A3C), which trains a neural network with less computing resources than standard reinforcement learning algorithms that normally needs costly GPUs. To prove these advantages, several experiments are run using a different number of parallel simulations (threads). The drone should reach a point randomly generated each episode. The reward, the value and the advantage function are used to evaluate the performance. As expected, these experiments show that a higher number of threads helps the leaning process improve and become more stable. These learning results are of interest to optimize the computing resources in future applications.