RegulaRAG: AI Pipeline for Regulation-Compliant Scenario Generation in Automotive Safety

A recent study presents RegulaRAG, a Retrieval-Augmented Generation (RAG) framework aimed at producing test scenarios that comply with regulations for safety-critical automotive systems. This system tackles the issue of grounding Large Language Models (LLMs) within extensive hierarchical standards, specifically UN Regulation No. 152 (AEBS). RegulaRAG integrates SmartChunking, which enhances paragraphs and tables through graph traversal, alongside a Smart Retrieve & Rerank approach. The research assesses the system using a curated dataset that encompasses all scenarios in UN Regulation No. 152, employing a three-step progressive search to determine optimal retrieval parameters, direct comparisons with five baseline RAG systems, and a robustness stress test featuring distractor content. Outputs are evaluated with a tailored penalized metric. This paper can be found on arXiv (2608.16394) and is a new submission. Its implications are crucial for the automotive sector, as it seeks to enhance the validation of advanced driver-assistance systems by ensuring scenario generation aligns with regulatory requirements.