• Proposes FS-DRL, a dual reinforced learning algorithm combining CNN and DQN for fine-grained scheduling of autonomous vehicles at non-signalized intersections.
• Achieves 96.5% accuracy and improves traffic efficiency by 61.54%–85.37% compared to Dijkstra, RNN, and ant colony-based methods.
• Introduces an adaptive learning rate that ensures convergence with minimal training loops, enhancing practical applicability.
• Addresses the challenge of dynamic, non-signalized intersection environments, offering a scalable distributed scheduling solution.