Data-Driven Edge-Case Validation Framework for E2E DCAS: Integrating ISO/TS 5083 and ISO/PAS 8800 for Regulatory-Compliant Safety Assurance

Abstract

This paper presents a standards-aligned verification and validation (V&V) framework for end-to-end (E2E) driver control assistance systems (DCAS), developed to ensure compliance with ISO/TS 5083 and ISO/PAS 8800. Conventional safety assessment methods, such as those based on simple mileage accumulation or random sampling, cannot effectively capture rare edge cases and are unable to identify functional insufficiencies within the black-box architectures of E2E AI. To overcome these challenges, we propose a data-driven scenario prioritization methodology that identifies safety-critical domains based solely on data characteristics prior to simulation. A composite metric ( $P_{i}$ ) integrating data scarcity, kinematic difficulty, and integrity is constructed to filter a massive dataset of 2.5 million scenarios to 100,000 high-priority cases. The prioritized scenarios are systematically mapped to five representative maneuver groups defined with reference to UN Regulation No. 171, Annex 4. Each maneuver group is independently evaluated using 3,131 closed-loop simulation runs, yielding 15,655 runs in total. The V&V environment employs a high-fidelity digital twin that co-simulates IPG CarMaker and MATLAB/Simulink to evaluate a recurrent proximal policy optimization agent under safety of the intended functionality (SOTIF) relevant triggering conditions, including stochastic sensor disturbances. Extensive closed-loop simulations reveal a strong Pearson correlation ( $r\gt 0.95$ ) between the $P_{i}$ index and actual collision rates. The results demonstrate that the proposed framework effectively resolves the black-box prioritization challenge by quantitatively uncovering “long-tail” blind spots that cannot be resolved by increasing training data volume alone. Overall, this research provides traceable, objective evidence to support a defensible safety assurance argument and offers a critical technical foundation for the regulatory approval of E2E autonomous driving systems.