Visual GANs for End-to-End UAV Trajectory Generation in RIS-Assisted Energy-Efficient Wireless Vehicular Networks

Designing collision-free unmanned aerial vehicle (UAV) trajectories for line-of-sight (LoS) wireless communication in dense urban environments is crucial for enhancing the connectivity of intelligent ground vehicles. However, conventional path-planning techniques struggle with kinodynamic constraints, obstacle avoidance, and real-time motion feasibility, making the problem intractable for standard convex optimization. Additionally, beamforming, a key requirement for future intelligent transportation networks, is energy-intensive, necessitating energy-efficient alternatives like reconfigurable intelligent surfaces (RIS). However, deploying RIS on urban structures further complicates UAV path planning, as a direct LoS link must be established between the UAV, RIS, and ground vehicles. To address these challenges, we propose an AI-driven trajectory planning framework leveraging generative adversarial networks (GANs). Our end-to-end visual GANs (vGANs) framework processes raw bird’s eye view (BEV) images to generate motion primitives and control signals, dynamically optimizing UAV trajectories for enhanced channel performance and navigation feasibility. Simulation results demonstrate the effectiveness of vGANs in real-time UAV navigation, outperforming traditional methods in collision-free LoS communication. • First vGANs-based 3D UAV trajectory planning with kinodynamic constraints and valid LoS links. • Enables intelligent, collision-free UAV navigation for enhanced LoS wireless communication for intelligent autonomous vehicles in dense urban areas. • Transforms raw sensory data into UAV control signals via a unique two-player GAN architecture. • Real-time path generation where traditional methods like RRT fail in complex urban environments. • Outperforms deep reinforcement learning by eliminating reliance on Markov decision processes.