Autonomous Navigation of Drones Using Explainable Deep Reinforcement Learning in Foggy Environment

Autonomous navigation of Unmanned Aerial Vehicles (UAV) in complex environments is still a challenging field. Recognizing UAV real-time perception as a sequential decision-making challenge, researchers increasingly adopt learning-based methods, leveraging machine learning to enhance navigation in complex environments. In this paper, a novel deep reinforcement learning (DRL) model has been proposed for the smooth navigation of the UAV. The paper provides an overview of existing techniques, laying the foundation for our proposed work, which not only addresses certain limitations but also demonstrates superior performance in complex environments. The simulation environment is built using Unreal Engine and connections are established using AirSim APIs. The implementation of the TD3 and PPO algorithm is chosen for its exceptional adaptability in continuous action spaces. A novel integration of Image adaptation and the Autonomous navigation model has been carried which increases the robustness of the drone in foggy conditions. Our model undergoes training in a customized foggy landscape mountainous environment, and the results, obtained after rigorous training, are thoroughly analyzed. The state-action pair of our trained TD3 agent is explained using LIME and SHAP techniques. The paper concludes by presenting promising directions for further exploration and advancement in this evolving field.