From ambiguity to precision: Digitized traffic rules for autonomous driving

In the mixed traffic streams with interacting human-driven and automated vehicles (AVs), AVs must accurately infer and comply with the traffic rules followed by human drivers, which is a prerequisite for safety. However, current traffic rules often use ambiguous expressions such as "not impede," which are easily interpreted by humans but difficult for AV implementation. To address this challenge, this study proposes a rule optimization framework that combines knowledge-based reasoning with data-driven optimization. The methodology is built on safety‑critical event (SCE) data collected from six intersections across three Chinese cities. The methodology involves three stages: (1) Semantic classification is conducted to categorize traffic rules; (2) For each category, Metric Temporal Logic is used to formalize natural language rules. The underspecified parameters are supplemented using domain knowledge extracted from 26 regulatory documents across ten countries/organizations. Initial parameter ranges are then estimated by incorporating behavioral patterns observed in SCEs; (3) A genetic algorithm is employed to calibrate the optimal parameters within simulation environments reconstructed from real-world SCEs. Experimental evaluations show that: (1) Relative to human drivers, the rule-integrated planner enhances safety performance by over 90% while maintaining both operational efficiency and ride comfort; (2) When embedded into existing planners on the CommonRoad and INTERACTION benchmarks, the optimized rules consistently reduce collision risks, confirming strong generalizability; (3) The final performance is primarily driven by parameters that control turn ranges, temporal safety thresholds, and longitudinal distances. Overall, this study introduces an interpretable and transferable rule optimization framework that improves safety and strengthens the alignment of AV behavior across different driving environments.