Hierarchical Cloud-Edge Cooperative Control Framework for Multi-Vehicle Autonomous Driving Using Reinforcement Learning and Model Predictive Control

This study presents a hierarchical cooperative control framework for hybrid electric vehicles in intelligent transportation systems, termed “cloud-coordinated cooperative driving with deep reinforcement learning.” The framework integrates reinforcement learning with model predictive control (MPC) to balance adaptability and stability. A cloud-level coordinator utilizes the multi-agent deep deterministic policy gradient (MADDPG) algorithm for global policy optimization, while each vehicle executes a local Soft Actor–Critic (SAC) controller for real-time adaptive control. A predecessor–leader–follower-based cooperative economic MPC module is incorporated to ensure stable spacing and energy-efficient trajectory tracking. Furthermore, a hybrid offline–online learning mechanism with state synchronization and experience replay enhances convergence speed and generalization capability. Under three representative driving cycles: highway fuel economy test, US06 (a high-speed, high-acceleration cycle), and China light-duty vehicle test cycle with an initial state-of-charge (SOC) of 0.60, the proposed MADDPG-SAC algorithm consistently outperforms deep deterministic policy gradient (DDPG) and MADDPG. This method attains lower fuel consumption (17.93–20.62 L/100 km), higher minimum SOC (0.33–0.35), reduced speed tracking error (0.22–0.26 m/s) and spacing error (0.29–0.35 m), smoother control (mean jerk of 0.041–0.049 m/s 3 ), and faster convergence (11–13 episodes compared with 12–29 episodes for DDPG and MADDPG). Overall, the framework exhibits strong cooperative capability, energy efficiency, and driving stability, demonstrating its potential for large-scale energy-aware autonomous driving applications.