Assessing the Readiness of Freeway Ramps for Automated Vehicles: A Software-in-the-Loop Simulation Study

Abstract

Current road infrastructure designed for human drivers may pose safety challenges for automated vehicles (AVs). In response to the higher safety risks on freeway ramps and the limited road-readiness studies on AV decision-making, this study examines the impacts of freeway-ramp design on AVs using simulation-in-the-loop tests. Thirty ramps across six interchanges along Shanghai’s first freeway for AV testing were modeled. Simulation tests were conducted in the Car Learning to Act (CARLA) simulator using the Traffic Manager (TM) module in autopilot mode. Rapid deceleration of the simulated AV served as the label. Six variables across three feature types were considered, including road geometry, road segment, and vehicle operation. The Extreme Gradient-Boosting (XGBoost) machine-learning algorithm and SHapley Additive exPlanations (SHAP) were used for modeling and interpretation. The results indicated that: 1) all features were correlated with rapid-deceleration events, especially the average 200 m downstream curvature and the target speed; 2) higher rapid-deceleration probability was associated with larger downstream curvature, smaller upstream curvature, higher target speeds, and ramp type with more complicated geometries; and 3) several interaction effects were observed, including curvatures of adjacent alignments, curvatures at various target speeds, and rates of change in combined horizontal and vertical curves. These findings contribute to AV safety by: 1) guiding the geometric design for improved freeway-ramp readiness; and 2) providing traffic departments and automobile manufacturers with AV operational design domain (ODD) restriction references.