A World Model-Based Architecture for Reasoning About Vehicle Reactions to Achieve Proactive Rear-End Collision Avoidance

Autonomous Vehicles (AVs) currently face a significantly higher risk of being rear-ended by Human-Driven Vehicles (HDVs) compared to conventional vehicles—approximately 1.6 times higher. This vulnerability largely stems from the prevailing unilateral decision-making paradigm, which predicts surrounding traffic behaviors independently of the AV’s actions, thereby neglecting the reciprocal interplay between the AV and following HDVs. Consequently, AVs may execute maneuvers that are technically feasible but unexpected or hazardous for nearby human drivers, which can increase rear-end collision risk in high-interaction scenarios. To bridge this gap, this study proposes a World Model-based architecture designed to explicitly reason about these reactive behaviors. First, a World Model is established to learn environmental dynamics, enabling continuous inference of how HDVs react to specific AV actions. Second, a planner based on the Actor-Critic architecture is developed and trained entirely within this World Model, facilitating the learning of farsighted, closed-loop interaction policies. This architecture is implemented in a novel longitudinal control method, named ThinkACC, and tested in high-interaction platoon scenarios. Compared to baseline Adaptive Cruise Control (ACC) system, ThinkACC achieves an 86.8% reduction in rear-end collisions while maintaining high levels of comfort and efficiency. In particular, the agent emerges with human-like defensive driving behaviors, such as proactively accelerating to mitigate risks of rearward collision. These findings demonstrate the capability of World Models to reason about bidirectional interactions, offering a viable path for enhancing AVs safety and generalization in mixed traffic environments.