Latency-Aware Edge-AI Scheduling in Vehicular Ad-Hoc Networks

Abstract

Vehicular Ad‑Hoc Networks (VANETs) have emerged as a critical component of next‑generation intelligent transportation systems, providing real‑time data exchange among vehicles, infrastructure, and cloud services. In these environments, latency is often the most stringent Quality of Service (QoS) requirement, particularly for safety‑critical applications such as collision avoidance, emergency braking alerts, and cooperative driving maneuvers. Traditional centralized cloud processing introduces unacceptable delays due to backhaul transmission times and unpredictable network congestion. Mobile Edge Computing (MEC) mitigates some of these concerns by relocating computation closer to the data source—either on On‑Board Units (OBUs) within vehicles or at strategically deployed Roadside Units (RSUs). However, these edge nodes have heterogeneous capabilities and limited resources, making optimal scheduling of computational tasks both challenging and essential. This manuscript proposes a novel latency‑aware Edge‑AI scheduling framework tailored specifically for VANET scenarios. Our framework dynamically assesses network conditions, task urgency, data dependencies, and node processing capabilities to make real‑time scheduling decisions. At its core lies a hybrid heuristic‑reinforcement learning (RL) scheduler: initially seeded with a latency‑minimization heuristic to provide a strong starting policy, and subsequently refined through online RL to adapt to evolving network topologies and workload patterns. We define a composite latency metric that incorporates transmission delay, queuing delay, and processing time, alongside deadline adherence penalties. By modeling the scheduling problem as a Markov Decision Process (MDP), our RL agent learns to balance the trade‑off between minimizing total latency and avoiding deadline violations. We validate our approach using the Veins simulation platform integrated with SUMO for realistic vehicular mobility traces under varied urban density and speed profiles. Compared to state‑of‑the‑art baselines—including static round‑robin, dependency‑aware heuristics, and model‑free DDPG schedulers—our method reduces median end‑to‑end task latency by up to 35%, cuts deadline miss rates by over 60%, and maintains 99% compliance for high‑priority safety tasks. Furthermore, the online learning capability ensures robust adaptability to sudden traffic fluctuations and node failures. These results demonstrate the potential of latency‑aware Edge‑AI scheduling in enhancing the reliability, responsiveness, and safety of VANET applications, paving the way for truly real‑time cooperative driving systems.