SA-DRL: Semantic-Augmented Deep Reinforcement Learning for UAV-Aided VANETs

A new framework, Semantic-Augmented Deep Reinforcement Learning (SA-DRL), has been proposed to address network fragmentation in Urban Vehicular Ad-Hoc Networks (VANETs). The framework, detailed in arXiv paper 2603.18871, leverages a pretrained Large Language Model (LLM) to generate topology-dependent action priors, guiding UAV deployment. The resulting algorithm, Semantic-Augmented PPO (SA-PPO), combines these priors with Proximal Policy Optimization (PPO) through Logit Fusion, improving exploration efficiency. Simulations using real-world urban trajectories show SA-PPO achieves faster convergence and better performance compared to baseline methods. The work addresses the challenge of inefficient exploration in DRL-based UAV deployment by incorporating road-topology guidance, offering a novel approach to maintaining connectivity in dynamic vehicular networks.