POPM: A Two-Stage Multi-Agent Deep Reinforcement Learning Framework for Joint Vehicle-Application-Infrastructure Task Offloading in Vehicular Edge Computing

The rapid development of Internet of Vehicles technologies has generated a large number of latency-sensitive deep neural network task requests. Vehicular edge computing systems constitute a complex environment involving vehicles, computation application, and infrastructure components such as roadside units (RSUs). Existing studies often focus on optimizing only a subset of these elements, thereby overlooking the potential benefits of joint optimization across multiple layers. To address this limitation, this paper proposes a two-stage distributed scheduling framework, termed POPM, which enables prediction-assisted offloading optimization through joint modeling of the vehicle, task, and infrastructure layers. In the proposed framework, the first stage focuses on the vehicle and task layers, where latency-constrained visual perception tasks are modeled as directed acyclic graphs, and a multi-dimensional dynamic priority mechanism incorporating vehicle residence time is integrated with a MAPPO-based algorithm to enable fine-grained distributed local offloading. The second stage emphasizes the infrastructure layer, where RSU load prediction is performed using a TCN–FiLM model, and a MADDPG-based policy is adopted to formulate proactive cross-RSU migration strategies, thereby achieving dynamic global resource balancing and hotspot avoidance. Simulation results show that POPM improves the task success rate by approximately 18% on average compared with the baselines, while keeping the task completion latency within a competitive range. In addition, the proposed load-balancing mechanism improves fairness by approximately 14%.