Physics-Guided neural networks for the robust control of drones

Throughout recent History Model-Based controllers dominated the scene of flight controller design and their effectiveness has been proven several times. Despite that, their reliance on the accuracy of the mathematical model used in order to represent real plant might lead to an explosion of the complexity of the problem in case of strongly non-linear systems such as UAVs. In the following work, we are going to propose an alternative approach for the design of flight controllers based on the use of ANN and their capabilities of being universal approximators in order to overcome some of the flaws of standard Model-Based controllers. We are going to investigate the effectiveness of a mixed approach of both Data and Information Driven techniques using Physics-Guided Neural Networks for the approximation of the real plant dynamics. Moreover, we are going to implement the Dynamics inversion of the plant using such techniques in order to design a flight controller based on feedback linearization. For the simulation of the plant and the data collection needed in order to train and validate the ML models we made use of Microsoft Arisim flight simulator while the ANN Development has been made through the Pytorch Framework. The Flight controller connection with the Quadrotor in the Airsim Environment uses the PX4 firmware and Mavlink communication protocol. The model we implemented is able to invert the Dynamics of the drone starting from a set of measurements and targets at given time instance (such as RPY angles, Velocities and accelerations in the body frame) and computes with good accuracy the set of forces applied on each vertex on the quad-rotor. Different Architectures has been tested, mainly LSTM and Feedforward, using both the standard Data-Driven procedure and Data-Information mixed procedure, and in each case the latter always improved the RMSE loss with respect to the counter-part of about 10%-20% with faster convergence.