Liu, Yongzhi (Southeast University), Zhang, Sunan (Southeast University), shen, changfeng (Southeast University), zhang, xiangwei (Southeast University), Zhuang, Weichao (Southeast University)

Risk-Aware Dual-Policy Coordination with World Model for Safe and Adaptive Highway Autonomous Driving

Scheduled for presentation during the Regular Session "S11c-Motion Planning, Trajectory Optimization, and Control for Autonomous Vehicles" (WE-LA-T11), Wednesday, November 19, 2025, 16:40−17:00, Broadbeach 1&2

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 19, 2025

Abstract

The inherent tension between safety and operational efficiency poses a critical challenge for reinforcement learning in autonomous driving, particularly in complex, unpredictable traffic environments. We propose a risk-aware dual-policy framework that combines trajectory imagination with hierarchical policy coordination. First, our method evaluates candidate actions via model-based trajectory rollouts, dynamically filtering unsafe behaviors while balancing safety and efficiency. Second, a scenarioadaptive dual-policy architecture is introduced. Short-term Gaussian policy extracts executable actions from low-risk imagined trajectories for responsive control, while a longterm Lagrangian SAC policy optimizes global decisions using real and imagined data. These policies co-evolve through closed-loop interaction, continuously improving safety and performance. Experiments in highway scenarios (lane changes, dense traffic, and sudden risks) show our method suppresses risky behaviors in high-risk settings while maintaining efficiency in low-risk conditions. Compared to baselines, the proposed approach achieves superior safety compliance, adaptability, and success rates, advancing reliable autonomous driving.