Machine Learning Control of an Aerial Robot Based on a Tuned Deep Deterministic Policy Gradient Method

This paper is devoted to the development of a Machin Learning (ML) control for a quadcopter via a Tuned version of the Deep Deterministic Policy Gradient (DDPG) method. In this regard, the aerial robot explores the environment and acquires the appropriate performance by the reception of a reward in a simulated area. Normally quadcopters exploit two distinct controllers for direction and position control tasks. Based on Reinforcement Learning (RL), quadcopter control can be implemented by an integrated control system. Initially, a 6DOF simulation toolset of the aerial robot is developed for the generation of the required learning dataset. In this work, for take-off and hovering tasks, the performance of a tuned version of the DDPG method in controlling the vehicle is examined. The developed toolsets give a capacity for employment of this control method in other flight phases and more complicated missions.