Control Design of a UAV using Deep Reinforcement Learning

This thesis presents the design and implementation of a UAV flight-control system using Deep Reinforcement Learning (DRL). Traditional PID-based and adaptive controllers, while effective, require labor-intensive tuning and are highly dependent on specific airframes. To address these limitations, this research explores the application of a Deep Deterministic Policy Gradient (DDPG) agent for low-level UAV control. A complete simulation environment was developed in MATLAB Simulink using the Aerospace Blockset and validated through hardware deployment on Parrot Mambo and Rolling Spider minidrones. The RL-based controller demonstrated stable autonomous flight, outperforming PID controllers under varying initial conditions and disturbances. The study establishes that DRL can serve as a generic, transferable control strategy for multirotor UAVs, enabling adaptive flight across platforms with minimal retuning.