Model-Centric Quantitative Methodology for Evaluating Simulation Reliability in Autonomous-Vehicle Testing

As autonomous vehicles continue to advance, autonomous driving systems are becoming increasingly complex. This growing complexity necessitates higher levels of verification, making simulation an essential tool for system validation. In this context, simulation models are being structured by integrating various submodels, such as vehicle dynamics, control, environment, and perception sensors. As these simulation models become more complex and diverse, quantitative verification methods are required to ensure that each model accurately reflects the real-world environment. Thus, this paper proposes a structured and quantitative framework for evaluating the reliability of each model used in integrated simulations. The proposed method decomposes the simulation architecture into four core model components, with the parameters of each model mapped to specific functional roles. Parameter-level fidelity is assessed by matching scenario-specific model activities across seven evaluation scenarios classified into lateral, longitudinal, and complex behaviors. A composite reliability index is constructed, which captures the morphological similarity, numerical accuracy, and phase-amplitude dynamics of time-series using statistical indices. The proposed method is applied to a vehicle-in-the-loop simulation platform to validate its applicability to real-world autonomous driving test data. This approach enables scenario-dependent, component-level simulation model evaluation and contributes to the establishment of a standardized simulation-based verification framework for autonomous vehicles.