Implementation of partially tuned PD controllers of a multirotor UAV using deep deterministic policy gradient

The present methodology employed in classical control systems is characterized by high costs, significant processing requirements, and inflexibility. In conventional practice, when the controller exhibits instability after being implemented on the hardware, it is often adjusted to achieve stability. However, this approach is not suitable for mass-produced systems like drones, which possess diverse manufacturing tolerances and delicate stability thresholds. The aim of this study is to design and evaluate a controller for a multirotor unmanned aerial vehicle (UAV) system that is capable of adapting its gains in accordance with changes in the system dynamics. The controller utilized in this research employs a Simulink-constructed model that has been taught by reinforcement learning techniques, specifically employing a deep deterministic policy gradient (DDPG) network. The Simulink model of the UAV establishes the framework within which the agent engages in learning through interaction with its surroundings. The DDPG algorithm is an off-policy reinforcement learning technique that operates in continuous action spaces and does not require a model. The efficacy of the cascaded PD controllers and neural network tuner is evaluated. The results revealed that the controller exhibited stability during several flight phases, including take-off, hovering, path tracking, and landing manoeuvres.