This study investigates the evolution of crash risk in urban road networks as the market penetration of Autonomous Vehicles (AVs) increases across different levels of automation. Using microscopic traffic simulations in a dense metropolitan area (Athens, Greece), this study integrates different vehicle automation levels, including conventional vehicles, partially automated (SAE Level 2–3), fully automated (SAE Level 4–5), and aggressively behaving AVs. An XGBoost machine learning model, combined with SHAP explainability, is employed to identify the most influential traffic and behavioral features associated with crash risk. Findings revealed that automation influences crash risk, while its impact is neither uniform nor isolated and it is shaped by traffic composition, infrastructure, and behavioral diversity. Fully automated vehicles reduce crash risk as their share grows, while partially automated vehicles slightly increase it in mixed-traffic conditions. Aggressively behaving AVs have minimal impact, mostly at high penetration rates. These findings underscore the need for network-specific planning and offer guidance for AV deployment, infrastructure upgrades and regulatory frameworks to support a safer transition to automated mobility.
Oikonomou et al. (Fri,) studied this question.