Explainable AI analysis of brake control in CARLA through reference and distilled policies.

Abstract

This paper presents an explainable AI analysis of brake control in CARLA closed-loop driving. Longitudinal braking is studied through a threshold policy, a risk-aware reference policy, and a distilled policy learned from reference rollouts. The framework combines structured traffic-state features, high-fidelity XGBoost surrogates, and SHAP to analyze deployed brake behavior across Town10HD and Town05. Results show that brake generation is consistently dominated by front-vehicle distance, relative speed, and time to collision, while lateral variables contribute weakly. The distilled controller retains a forward-risk-oriented explanation structure rather than behaving as an arbitrary black box, but the degree of apparent semantic alignment with the reference policy is environment-dependent. Fallback-aware analysis shows that this alignment is substantially reinforced by the deployed safety fallback in Town10HD, while Town05 provides a cleaner view of the learned component itself. These findings show that learned brake control can remain interpretable when grounded in semantically structured state representations.