Bridging demographics,perception,and performance: A human-AI approach to vehicle evaluation in a dynamic simulator

Abstract

The evaluation of virtual vehicle models in high-fidelity driving simulators currently depends strongly on the experience and subjective judgment of a limited group of human drivers. In this context, this work proposes a data-driven Human–AI methodology that integrates driver characteristics (age and driving experience), drivers’ subjective assessments (five SAE-J1441-based questions), and virtual vehicle model telemetry. Virtual driver and driver-in-the-loop simulations were conducted on a closed circuit with N = 25 volunteers (12 females, 13 males; mean age 22 ± 7 years; driving experience 4 ± 6 years). In addition, data augmentation via standard-deviation-based synthesis, followed by Kolmogorov-Smirnov acceptance, increased sample representativeness and expanded the original dataset from 75 records to 225. At the end of the work, models yielded high classification performance on test sets (accuracies concentrated on the main diagonal: 86%, 90%, 90%, 93%, and 90% for Q0–Q4, respectively). Consequently, these results demonstrate that combining demographic and objective telemetry inputs enables reliable prediction of subjective drivability evaluations, thereby reducing dependence on individual qualified drivers and potentially reducing simulator operational costs.