Challenge 4

Generative AI for Automatic Test Case Generation from Crash Databases & Standards_Siemens Industry Software NV 

Team: 

Crash2Test — Generative AI for Crash-Derived Validation Scenarios

Proposal title: 

University of West Bohemia (UWB), Czechia (lead), with ARRK Engineering GmbH, Germany

Organisation

The University of West Bohemia in Pilsen, Czechia, is a public university founded in 1991; its team has published research on crash-image analysis and vision-language crash reasoning. Technical partner ARRK Engineering GmbH, a Munich-based engineering services company founded in 1967, contributes ADAS validation and OEM engineering experience.

Project info

Crash2Test turns fragmented crash evidence into validation-ready tests through one uncertainty-aware pipeline. Accident records, crash reports, structured data and multi-view post-crash imagery are transformed into Functional Scenarios, enriched into Logical Scenarios and linked to Euro NCAP and other validation references to reveal coverage gaps and generate simulation-ready tests for CARLA or OpenSCENARIO.

It builds on UWB’s research prototypes. Domain-adapted vision-language models read multi-view crash imagery with text and metadata; the team’s Crash-VQA work raised an open-weight model’s accuracy from 45.4% to 67.3%. A multi-view encoder with custom attention estimates crash mechanics such as CDC attributes and ΔV. An LLM with retrieval over crash taxonomies, standards and Euro NCAP material grounds interpretation. A structured scenario layer stores actors, road context, manoeuvres and parameter ranges with evidence and confidence per attribute, rather than inventing missing parameters. Schema-constrained generation converts Functional into Logical Scenarios, hybrid semantic and rule-based matching detects gaps, and outputs are checked for traffic, kinematic and multi-agent consistency before simulation.

ARRK validates the workflow in a representative ADAS validation environment. Targets include at least 95% of attributes linked to evidence, matching F1 of at least 0.85 and at least 50% less engineer time per accepted test.

Details

Automotive validation depends on realistic scenarios, but crash evidence is spread across reports, databases, images and standards. Deriving test scenarios from it by hand is slow and hard to trace, and existing tools address only isolated steps such as crash-image classification or text-to-scenario generation.

Crash2Test, developed by the University of West Bohemia with ARRK Engineering, links these steps into one evidence-grounded workflow from real crash data to executable test cases. Vision-language models adapted to the crash domain read post-crash photographs together with reports and records, and dedicated models recover crash mechanics such as impact configuration and change in velocity. Retrieval over standards and Euro NCAP material grounds the interpretation, and every generated attribute keeps its evidence and a confidence value. Accident-derived scenarios are then matched to validation catalogues to show where testing falls short, and converted into simulation-ready tests that are checked for physical consistency.

The components already exist as research prototypes on real crash data, giving a TRL 3 starting point. During the competition the partners will integrate them, validate the pipeline on heterogeneous crash sources and reference standards, and demonstrate TRL 5 through containerised end-to-end runs and integration with an industrial validation workflow at ARRK. Targets include Functional Scenario macro-F1 of at least 0.80, at least 80% correct Logical Scenario attributes, at least 95% of attributes linked to evidence, catalogue-matching F1 of at least 0.85, at least 90% physics-consistent outputs, and reproducible batch processing of at least 10,000 accident records. ARRK provides a direct route to industrial pilots with OEMs and Tier-1 suppliers.