Modelling riders’ intervention behavior during high-level autonomous driving under extreme conditions

The development of autonomous driving systems (ADS) has primarily focused on technical advancements to prevent accidents and enhance overall transport performance. While significant strides have been made in improving the performance of autonomous vehicles (AVs), there remains a substantial disconnect between the intelligence of AVs and user acceptance. This study aims to illuminate the differences in decision-making between the ADS and riders in AVs during extreme crash scenarios, identifying the specific factors that prompt rider interventions. We recreated three typical fatal road accidents from Australian roads, within a high-fidelity virtual reality (VR) environment. In these simulations, vehicles involved in the original crashes were replaced with fine-tuned level 4 AVs to evaluate whether the accidents would occur in a similar manner. We engaged human participants from diverse demographic backgrounds in a human-in-the-loop analysis, immersing them in these scenarios by sitting in the simulated AVs to gather insights into their perceptions and reactions. Our study investigated the factors influencing riders’ intervention behaviors and highlighted the decision-making disparities between ADS and human riders. We quantified the nature of these interventions during autonomous driving by defining turning and accelerating indices. The results revealed a strong correlation between interventions and vehicle movement, with intervention probabilities exceeding 80% when the AV’s acceleration index surpasses 0.7. Most importantly, our findings distinguished between unnecessary and necessary interventions during rider interactions with ADS under extreme conditions. We showed that necessary interventions can help refine ADS maneuvers at intersections by tempering aggressive responses, offering valuable guidance for system development. These insights not only inform current ADS enhancement strategies but pave the way for future research aimed at reducing unnecessary interventions while recognizing the value of necessary ones, ultimately supporting the broader adoption of high-level ADS.