Interactive Driving Turing Test with Think-Aloud Protocol: How Realistic Are Behavioural Driver Models?

Abstract

Driver models are essential for virtual safety assessments of automated vehicles. This study evaluates the realism of the i4Driving model, a behavioural driver model designed to emulate human-like driving, using an interactive driving Turing test combined with a think-aloud protocol in connected driving simulators. Thirty participants interacted with either a human-controlled vehicle or the i4Driving model across three motorway scenarios and rated realism, predictability, safety, and aggressiveness. Results showed that the i4Driving model was perceived as less realistic and predictable than human drivers (p < 0.001), yet participants could not reliably distinguish between the two (classification accuracy = 0.673). Think-aloud analysis revealed specific shortcomings, such as unrealistic merging and speed adaptation, alongside instances of naturalistic behaviour. These findings highlight the need for improvements in tactical decision-making and demonstrate the value of combining subjective ratings with qualitative insights for refining driver models.

Keywords: driving Turing test, driving behaviour, human-likeness, subjective rating, connected driving simulators, driver model

DOI: 10.54941/ahfe1007130