Deep Learning-Based Autonomous Navigation for PAVs in Urban Airspaces via Synthetic Dataset Generation Framework

Personal Aerial Vehicles (PAVs) are emerging as a critical component of Urban Air Mobility (UAM), enabling short-range passenger transport in dense urban areas. As manual piloting becomes impractical under high traffic density and dynamic obstacles, safe operation in such environments necessitates reliable autonomous navigation. This study presents an end-to-end Autonomous Navigation and Control System (ANCS) for PAVs, developed and rigorously evaluated within a high-fidelity Unreal Engine-AirSim simulation environment. A comprehensive synthetic dataset comprising 29,320 high-resolution RGB images was generated using controlled PX4-SITL flights in complex urban layouts, with each frame labeled for three-dimensional waypoint regression. The proposed framework integrates a Convolutional Neural Network (CNN) for spatial feature extraction with a Gated Recurrent Unit (GRU) for temporal sequence modeling. To ensure passenger safety and flight stability, the architecture incorporates an active depth-based obstacle intervention mechanism (utilizing an empirically derived 15-meter threshold) alongside a post-prediction trajectory smoothing module. Comprehensive experiments, including robust ablation studies and statistical trials across multiple urban density scenarios and low-visibility conditions, demonstrate the system’s efficacy. The ANCS achieved a 100% navigation success rate and maintained a highly controlled simulated collision rate of 4.0%. The results confirm that conservative, safety-focused displacement prediction effectively supports short-term path planning for PAVs in complex urban airspace. The source code is publicly available at: https://github.com/Prajjval78/Pav_code_files