Drone Path Planning Based on Policy in the Latent Action Space with Perturbation and Conservative Q-Learning

This paper focus on the drone path plafnning problem in Drone-Assisted Vehicular Networks (DAVN). Although DAVN enhances the coverage of vehicular networks and improves road safety, it faces challenges related to high energy consumption and unstable communication latency between drones and vehicles. In response to these challenges, this paper addresses the drone path planning problem by leveraging the offline Reinforcement Learning (RL) algorithm. First, an extensive dataset is built with realistic modeling of Drone-to-Vehicle (D2V) communication latency and drone energy consumption. Then, an offline RL model based on the Policy in the Latent Action Space (PLAS) and Conservative Q-Learning (CQL) model is proposed with detailed reward function, and is trained with the pre-collected dataset. Experimental results demonstrate that compared to elliptical paths, the predicted paths from the trained model achieve improvements of 58.23%, 49.22% and 19.7% in flight path length, drone energy consumption and D2V communication latency, respectively.