Multidimensional Assessment of Takeover Performance in Conditionally Automated Driving
URL: http://arxiv.org/abs/2507.22252v2
archive: archived pipeline: cataloged
Abstract
When automated driving systems encounter complex situations beyond their operational capabilities, they issue takeover requests, prompting drivers to resume vehicle control and return to the driving loop as a critical safety backup. However, this control transition places significant demands on drivers, requiring them to promptly respond to takeover requests while executing high-quality interventions. To ensure safe and comfortable control transitions, it is essential to develop a deep understanding of the key factors influencing various takeover performance aspects. This study evaluates drivers' takeover performance across three dimensions: response efficiency, user experience, and driving safety - using a driving simulator experiment. EXtreme Gradient Boosting (XGBoost) models are used to investigate the contributions of two critical factors, i.e., Situational Awareness (SA) and Spare Capacity (SC), in predicting various takeover performance metrics by comparing the predictive results to the baseline models that rely solely on basic Driver Characteristics (DC). The results reveal that (i) higher SA enables drivers to respond to takeover requests more quickly, particularly for reflexive responses; and (ii) SC shows a greater overall impact on takeover quality than SA, where higher SC generally leads to enhanced subjective rating scores and objective execution trajectories. These findings highlight the distinct yet complementary roles of SA and SC in shaping performance components, offering valuable insights for optimizing human-vehicle interactions and enhancing automated driving system design.
Summary
Driving simulator study of takeover performance in conditionally automated driving across nine scenarios (three traffic densities x three non-driving related tasks). XGBoost models with SHAP values predict ten takeover metrics spanning response efficiency, user experience, and driving safety. Compares baseline driver-characteristics (DC) models against DC+SA, DC+SC, and DC+SA+SC to isolate the contributions of Situational Awareness and Spare Capacity. Results cross-validated with Random Forest and LightGBM.
Key finding
Higher Situational Awareness shortens reaction times to takeover requests, especially reflexive responses, while Spare Capacity has a larger overall effect on takeover quality, raising both subjective ratings and objective trajectory measures. SA contributes minimally once DC and SC are already in the model.
Methodology
Fixed-base medium-fidelity driving simulator at TU Delft. 57 drivers, nine takeover scenarios per driver (3 traffic densities x 3 NDRTs); final analyzable dataset of 466 takeovers after exclusions. SA and SC measured via established questionnaires plus eye-tracking-derived cognition metrics. XGBoost regression/classification with SHAP attribution; cross-validated with Random Forest and LightGBM.
Sample size: 57 drivers, 466 takeovers analyzed
Quality score: 5 / 5