AI-Enabled Vehicular Data Offloading for Sustainable Smart Cities: Taxonomy, KPI Models, and Open Challenges

Vehicular edge computing (VEC) is emerging as a key enabler for intelligent transportation systems that are both latency- and energy-sensitive. This survey is motivated by the need for a unified, KPI-driven view of AI-based vehicular computation offloading that explicitly links performance gains to sustainability objectives in smart cities. We synthesize recent advances in AI-powered offloading for vehicular networks, with emphasis on deep reinforcement learning (DRL) and multi-agent variants that learn adaptive, sequential policies under dynamic topology, fluctuating wireless capacity, and heterogeneous workloads. We propose a unified taxonomy that spans infrastructure-based, vehicle-assisted, and hybrid architectures; map offloading decisions to key performance dimensions (end-to-end latency, energy efficiency, reliability, throughput, and task-success rate); and formalize a minimal KPI model that links radio, compute, and caching components. The review compares algorithmic designs (DQN/DDPG/A3C/SAC, prioritized and federated variants, DRL+optimizer hybrids), scheduling granularities and baseline choices, while examining reproducibility factors such as simulators, mobility models, and dataset availability. We further discuss integration with enabling technologies (cellular vehicle-to-everything (C-V2X)/NR-V2X, reconfigurable intelligent surfaces (RIS), UAV relays, edge caching), security and privacy considerations, and the sustainability implications of AI-driven offloading for intelligent urban environments. The paper concludes with open challenges including non-stationarity, sim-to-real transfer, safety constraints, and explainability and outlines a research agenda toward robust, accountable, and resource-efficient offloading policies deployable in real world VEC systems.