Autonomous Control of an UAV in a Virtual Environment Using Machine Learning and Pathfinding

The paper aims to conduct a comparative analysis between machine learning and pathfinding methodologies for the autonomous control of Unmanned Aerial Vehicles (UAVs) within a simulated environment. The machine learning approach is delineated into two distinct phases. Firstly, the training of an agent utilizing deep reinforcement learning algorithms to govern the manipulation of a quadrocopter’s rotors. Subsequently, a dual-mode training strategy incorporating deep reinforcement learning and imitative learning algorithms is employed to enable the autonomous flight of a UAV within a virtual environment replete with obstacles. For autonomous UAV control, the A* search algorithm is implemented. The UAV’s navigation is facilitated through the utilization of LiDAR sensors to map the environment, with subsequent route adjustments. Furthermore, infrared sensors are deployed to detect and circumvent obstacles. The evaluation methodology encompasses a comparative analysis of performance metrics, specifically focusing on two distinct environmental settings with varying obstacle complexities. Parameters such as flight time and collision occurrences with the environment are measured and juxtaposed between the machine learning and pathfinding approaches. The empirical findings underscore the superior stabilization and maneuverability exhibited by the machine learning-based solution, resulting in notably improved performance metrics owing to its capacity for real-time adaptation to changes in velocity vectors, in contrast to pre-programmed stabilization techniques.