Mobility-Aware Assisted Deep Reinforcement Learning for Collaborative Task Migration and Resource Allocation in Vehicular Edge Computing

Driven by advances in artificial intelligence and wireless communication technologies, vehicular edge computing (VEC) has become a vital component of intelligent transportation systems. However, high vehicle mobility in VEC may cause task execution interruptions as vehicles move beyond the coverage of the roadside units (RSUs). Additionally, uneven vehicle distribution and static RSUs deployment can lead to workload imbalance and inefficient resource allocation. Although collaborative task offloading and resource allocation among RSUs can mitigate some existing challenges, vehicle mobility-aware proactive decision-making is the key to achieving lower delay and better workload balancing. Therefore, we propose a novel mobility-aware task migration (MATM) framework, which integrates a Mamba-based trajectory prediction module and a soft actor–critic (SAC)-based intelligent decision-making module. MATM leverages a multi-level residual connection architecture to accurately capture the future spatiotemporal information of vehicles. This information serves as environmental state guidance for the SAC decision module to perform dynamic task migration and resource allocation. By integrating spatiotemporal information mining with deep reinforcement learning (DRL), MATM enables real-time optimal forward-looking service decisions. Extensive experiments on real-world vehicle trajectory datasets from Cologne demonstrate that MATM outperforms both heuristic and DRL-based methods, reducing the average task completion delay by 8.3%-54.5%.