Kravbaserad generering och validering av AI-skapade scenarios for AV-träning och utvärdering

Research Project, 2027 – 2028

Automated vehicles (AVs) rely heavily on virtual testing, which requires large numbers of realistic driving scenarios covering both everyday situations and rare, safety‑critical events. Recent advances in AI, Large Language Models (LLMs), and Vision‑Language Models (VLMs) have accelerated situational understanding and synthetic scenario generation, but important gaps remain. Requirements for scenario generation are not well defined, AI‑based methods for producing meaningful scenario variants are still limited, and robust ways to verify whether AI‑generated scenarios reflect real‑world behavior are largely missing. These gaps risk AVs being trained and evaluated on unrealistic or insufficiently broad scenario sets, affecting safety and user acceptance.

This project aims to strengthen AI‑ and LLM‑based scenario generation for AV development in three areas. WP1 will develop methods for extracting scenario‑generation requirements from legislation, literature, reports, and data, followed by expert validation. WP2 will build pipelines that convert textual crash descriptions into 3D motion scenarios and generate LLM‑based scenario variants to broaden coverage, including edge cases. WP3 will establish methods for assessing the realism and representativeness of AI‑generated scenarios, integrating WP1 requirements and benchmarking both project‑developed and commercial scenario-generation tools.

Expected results include a requirements framework, AI-based scenario variant generation, V&V methods, and an end‑to‑end process linking requirements, generation, and validation. With industry partners involved, uptake should be rapid. Outcomes will be disseminated through conference submissions, a potential journal article, and further collaboration, supporting the automotive industry in advancing reliable AI‑based scenario generation for AV training and safety assessment.