End-to-End 자율주행 AI의 안전성 확보를 위한 기술 동향과 국제 안전 표준 관점의 분석

Abstract

End-to-end (E2E) learning-based autonomous driving has emerged as a promising paradigm that directly maps sensor inputs to vehicle control commands using data-driven models. Compared to conventional modular architectures, E2E approaches offer advantages in terms of architectural simplification and holistic learning of complex driving contexts from large-scale data. However, since E2E systems rely on probabilistic decision-making, exhibit limited explainability, and remain vulnerable to distribution shifts, edge cases, and long-tail scenarios, those benefits actually introduce fundamental challenges in safety assurance. This paper reviews recent technological developments in E2E autonomous driving and systematically analyzes key AI safety issues from a system-level perspective. Core challenges including rare operational scenarios, uncertainty under out-of-distribution conditions, and limitations in traceability and accountability are examined in relation to existing automotive safety frameworks. In particular, the paper investigates how ISO 21448 (Safety of the Intended Functionality, SOTIF) and UL 4600 can be reinterpreted and applied to learning-based autonomous driving systems to complement traditional failure-based functional safety standards. To address the structural mismatch between E2E architectures and existing safety standards, this paper discusses rule-based safety shields and scenario-based validation as practical and standard-compatible mechanisms for mitigating non-failure-based risks and constructing evidence-driven safety arguments. The analysis demonstrates that instead of E2E autonomous driving invalidating existing safety frameworks, it actually necessitates their complementary and systematic integration to achieve robust safety assurance in learning-based autonomous driving systems.