A Proximal Policy Optimization Reinforcement Learning Approach to Unmanned Aerial Vehicles Attitude Control

The latest developments in the field of Machine Learning (ML), especially Reinforcement Learning (RL) techniques, reduce the need of having pre-existing data available. In this paper, we are presenting a Reinforcement Learning approach to Unmanned Aerial Vehicles (UAV) trajectory tracking and attitude control for an X configuration quadcopter. The proposed solution aims to tackle different maneuvers and to be able to withstand a wide variety of environmental disturbances, both while ensuring the success of the mission for which the Unmanned Aerial Vehicle has been designed. The Proximal Policy Optimization (PPO) solution has first been trained in a simulation environment. The model of the vehicle is designed to take into account various configurations, including changes of mass, while the model of the environment contains various disturbances sources.