Presentation
Explaining Trust and Proactive Takeovers in Automated Driving: A Machine Learning Analysis with Neural, and Gaze Metrics
DescriptionDespite recent advancements towards fully automated vehicles (AVs), SAE L4/5 systems would likely remain limited by their operational design domains. This study attempts to explain drivers’ trust towards partially automated vehicles (L2/3) and takeover decisions using gaze and neural data. Fifty-one drivers (24 females) experienced an 85-minute simulation with four traffic events manipulating AV crash avoidance and silent failures. Participants were allowed to proactively take over the AV whenever they thought it was unsafe. XGBoost inference models were constructed with brain cortical activation and gaze metrics to predict drivers’ binary proactive takeover decisions (takeover; no-takeover) and subjective trust levels (high; low). The best model achieved an area-under-the-curve (AUC) of 0.905 and predicted proactive takeover with lower medial dorsolateral prefrontal cortex (DLPFC) and higher right DLPFC activations. This trend shows that proactive takeover reduces cognitive load as a risk-avoidance strategy. Meanwhile, models predicting subjective trust had an AUC of 0.631, with fixation duration on the steering peak and left DLPFC being the strongest predictors. These are reflections of trusting drivers’ goal hierarchy approach and offloading cognitive demands on the AV, without gaze attention on disengagement control for takeovers. Our findings provide design insights for adaptive AV using neurophysiological measures.
Event Type
Lecture
TimeTuesday, October 14th1:50pm - 2:10pm CDT
LocationGrand Hall J
Surface Transportation
Similar Presentations