An LLM-driven Scenario Generation Pipeline Using an Extended Scenic DSL for Autonomous Driving Safety Validation

TL;DR

This paper introduces a scalable pipeline that leverages GPT-4o and an extended Scenic DSL to automatically convert crash reports into executable simulation scenarios for autonomous driving safety testing, improving accuracy and variability capture.

Contribution

It presents a novel intermediate Scenic DSL layer and a pipeline that enhances scenario generation accuracy and scalability compared to prior direct text-to-scenario methods.

Findings

100% correctness in environmental and road attributes extraction

97-98% accuracy in trajectory extraction

Successfully triggered traffic violations in 2,000 simulated scenarios

Abstract

Real-world crash reports, which combine textual summaries and sketches, are valuable for scenario-based testing of autonomous driving systems (ADS). However, current methods cannot effectively translate this multimodal data into precise, executable simulation scenarios, hindering the scalability of ADS safety validation. In this work, we propose a scalable and verifiable pipeline that uses a large language model (GPT-4o mini) and a probabilistic intermediate representation (an Extended Scenic domain-specific language) to automatically extract semantic scenario configurations from crash reports and generate corresponding simulation-ready scenarios. Unlike earlier approaches such as ScenicNL and LCTGen (which generate scenarios directly from text) or TARGET (which uses deterministic mappings from traffic rules), our method introduces an intermediate Scenic DSL layer to separate high-level semantic understanding from low-level scenario rendering, reducing errors and capturing real-world variability. We evaluated the pipeline on cases from the NHTSA CIREN database. The results show high accuracy in knowledge extraction: 100% correctness for environmental and road network attributes, and 97% and 98% for oracle and actor trajectories, respectively, compared to human-derived ground truth. We executed the generated scenarios in the CARLA simulator using the Autoware driving stack, and they consistently triggered the intended traffic-rule violations (such as opposite-lane crossing and red-light running) across 2,000 scenario variations. These findings demonstrate that the proposed pipeline provides a legally grounded, scalable, and verifiable approach to ADS safety validation.