Novel Autonomous Drone Control System Utilizing Deep Reinforcement Learning

As drones become increasingly vital in disaster response, logistics, and environmental monitoring, there is a growing need for robust, decentralized autonomous control systems. This research develops a novel autonomous control system that enables multirotor drones to navigate autonomously and make real-time decisions using only onboard sensors, such as IMU and LiDAR, without reliance on GPS or other centralized systems. The control system is built using an advanced Deep Reinforcement Learning (DRL) model, the Twin-Delayed Deep Deterministic Policy Gradient (TD3), which outperforms other control models or algorithms in achieving real-time, continuous control. Inputs to the DRL algorithm, such as position, velocity, and sensor data, are calculated from the onboard flight controller. The control system outputs real-time velocity commands to enable adaptive flight. The TD3 algorithm is first trained within the Cosys-AirSim simulation environment built on Unreal Engine. The training process follows a curriculum learning approach to ensure the model adapts to increasingly complex environments incrementally. The trained control system was evaluated in simulation, achieving a success rate between 90% to 95%. Following simulation, the control system was deployed on a physical drone equipped with a Raspberry Pi companion computer, LiDAR, and an optical flow sensor. A pre-trained and fine-tuned MobileNetV3 convolutional neural network (CNN) model and a thermal sensor were also integrated to detect mission-specific objects. The drone successfully navigated without using GPS, located and approached the objects, which validates the effectiveness of the autonomous drone control system in real-world scenarios. This control system enhances real-time responsiveness, scalability, and flexibility, allowing drones to autonomously navigate to any designated point while adjusting its flight path based on real-time conditions.