APF-SAC: Real-Time Autonomous Navigation System for Railway Inspection UAVs Based on Improved SAC Algorithm

With the rapid expansion of railway networks, the safety of infrastructure has increasingly become a critical concern. In recent years, drones have gained widespread attention and pilot applications in the field of railway inspection, but given the complex environment along the railway, how to ensure the safety of flights is still a major challenge. This paper proposes a real-time autonomous navigation system for railway inspection UAVs based on an improved SAC algorithm. Compared with traditional deep reinforcement learning methods that face bottlenecks such as low sample efficiency and slow policy convergence in high-dimensional dynamic environments, the proposed UAV system integrates a lightweight monocular depth camera as the perception device. By estimating image depth to obtain 3D environmental structure information, the system effectively enhances perception range and adaptability. Furthermore, considering the elongated and ribbon-like characteristics of railway viaducts, a reward-punishment function is designed by incorporating the artificial potential field method. Finally, comparative simulation experiments in real railway scenarios demonstrate that the proposed method enables UAVs to achieve real-time autonomous navigation in complex railway environments.