Integrating RAG and Reasoning for the Realization of a Concrete Driving Scenario within an LLM-Based Framework

作者：Majid Jegarian, Shahrzad Hanifehbakkheyrabadi, Jonas Freyer, Katharina Bause, Tobias Düser · 年份：2025 · DOI：10.1109/iavvc61942.2025.11219551 · 被引用次数：1 · 研究领域：Autonomous Vehicle Technology and Safety、Human-Automation Interaction and Safety、Modeling and Simulation Systems

This paper presents an end-to-end framework that leverages Large Language Models (LLMs) to generate simulation-ready driving scenarios from natural language input, addressing key limitations in automated vehicle (AV) validation processes. Existing tools for scenario generation often lack scalability, semantic rigor, and the capacity for hierarchical, multi-step reasoning required for scenario concretization. The proposed approach translates high-level natural language prompts into fully specified, executable scenarios through a reasoning-enhanced LLM pipeline that integrates Retrieval-Augmented Generation (RAG), Chain-of-Thought (CoT) prompting, and structured selfevaluation. Abstract scenarios are systematically refined into parameter-complete, logically coherent representations suitable for direct execution in industry-standard simulators. The framework generalizes across scenario specification formats by dynamically extracting and applying structural constraints from formal documentation. Experimental results demonstrate that the system produces syntactically valid and semantically consistent scenarios without manual intervention. This work contributes a scalable, automation-driven approach to scenario generation that bridges high-level user intent and fully executable simulation files for AV testing.