Region-specific highway driving scenarios generation for accelerating automated driving systems validation: A large-language-model assisted framework

Abstract

The safety and reliability of Automated Driving Systems (ADS) must be validated before large-scale deployment, and scenario-based testing is a promising way to improve validation efficiency and reduce cost. However, unidentified cross-regional differences in driving scenarios force manufacturers to repeat extensive validation when deploying ADS-equipped vehicles in new regions. Quantifying these differences, so that already-validated common scenarios need not be retested, remains insufficiently addressed. This work proposes a 14-dimension highway driving behavioral taxonomy and builds an automated framework to compare naturalistic driving datasets from China and Germany, with every contrast computed within matched traffic states. For two context-dependent dimensions whose closed-form definitions are especially fragile, lane-change aggressiveness and interaction danger, a dual-track evaluation combining a deterministic formula with a Large Language Model (LLM) agent is introduced. Using Cliff's delta, significant and practically meaningful cross-regional differences are identified across multiple dimensions and traffic states, and the dual-track method localizes where the two label sets disagree. A regional differential filter then converts the identified divergences into a library of OpenSCENARIO test fragments that are batch-executable on a Hardware-in-the-Loop (HiL) bench, offering a practical way to cut adaptation-validation effort when transferring an ADS between regions.