Title: Safe2Drive: Assessing the Safety Proficiency of End-to-End Autonomous Driving Systems

Abstract:

While recent end-to-end (E2E) autonomous driving policies have demonstrated impressive performance metrics in closed-loop simulations, their ability to navigate common safety-critical situations remains uncertain. To address this gap, we introduce Safe2Drive (S2D), a collection of scenario extensions aligned with Bench2Drive that target three prevalent categories of road hazards: construction zones, pedestrians jaywalking, and vulnerable road users (VRUs) hidden from view. Safe2Drive incorporates 100 complex, everyday scenarios and establishes the SafeDriving Score (SDS), a safety-oriented evaluation metric. This metric enhances previous standards by assessing pre-crash braking capabilities, contact with objects in work zones, lane centering precision, and driving smoothness.

Our evaluation of two leading state-of-the-art policies, LEAD and SimLingo, on the S2D platform reveals a significant decline in performance compared to their Bench2Drive baseline scores. Specifically, LEAD’s score fell from 94.70 on Bench2Drive to 39.95 on S2D, while SimLingo’s dropped from 85.07 to 41.00. Furthermore, the SDS results were notably low, recording 11.85 for LEAD and 15.27 for SimLingo. These findings underscore fragile safety behaviors within the models, including inadequate handling of work zones, violations of traffic signals, and delayed or missing braking responses when encountering pedestrians. This study emphasizes a critical deficiency in safe behavioral reasoning among E2E models, even when evaluated on CARLA environments included in their training data. We intend to make the code and video recordings for all 100 S2D scenarios publicly available.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC