Challenge 4
Generative AI for Automatic Test Case Generation from Crash Databases & Standards_Siemens Industry Software NV
Team:
Pleias, France
Proposal title:
CASCADE
Organisation
Pleias is a Paris-based AI company of 20 people that builds open training data, synthetic-data pipelines and small specialised language models. Its work includes Common Corpus, an open training dataset of 2.27 trillion tokens, and Synth, a synthetic data generation pipeline delivering state-of-the-art efficiency for foundation model training.
Project info
CASCADE is a hybrid generative-Bayesian environment covering the whole challenge. A discrete Bayesian network fitted to the open UK, French and Belgian crash datasets (STATS19, BAAC, Statbel) fixes what is statistically true: which crashes occur, how often and with which parameters. Generative models do the reading and writing but are never the source of a fact, which makes them admissible in safety validation.
Because Europe has no lawful corpus of crash evidence paired with scenarios, the environment manufactures one. A large open model renders sampled crashes into police-style narratives, records and claim summaries in four languages. Specialised Pleias models of 0.3B to 8B parameters, trained in the environment with inline citations, calibrated uncertainty and abstention, then extract harmonised Functional Scenarios and enrich them into Logical Scenarios; vision-language models read visual datasets and OpenStreetMap supplies road-geometry priors. A matcher reads Euro NCAP 2026 and UN R157 references into the same grammar and quantifies coverage gaps. Outputs are ISO 34505 test cases in OpenSCENARIO 1.3, executed in esmini and Simcenter Prescan.
Since every item’s truth is known, the environment is also the benchmark. Targets include at least 90% coverage of injury-crash probability mass per country and calibration error below 0.05.
Details
Europe’s open crash data consists of coded fields without narratives, in-depth data is restricted, and police reports contain personal data. As a result, there is no lawful corpus of crash evidence paired with validation scenarios on which to train or fairly test AI systems that turn accidents into test cases.
CASCADE, developed by Pleias, solves this with a hybrid environment. A Bayesian network fitted to national crash statistics from the UK, France and Belgium defines which crashes occur and how often; generative models then write realistic multilingual evidence from those known facts and read it back into structured scenarios. Because the true answer is known for every generated item, the same environment produces training data, serves as a benchmark and scores any extractor, enricher or matcher on the challenge KPIs. Small specialised models of under 8 billion parameters run on a single on-premise GPU, so manufacturers can keep accident data in-house.
The project will move from TRL 3 to TRL 5 by fitting the network to real national data, building a multilingual corpus of at least one million evidence-scenario pairs, and validating it on real crash narratives and expert review, with test cases run in esmini and Simcenter Prescan. Targets include a catalogue covering at least 90% of injury-crash probability mass per country, match precision and recall above 0.9, calibration error below 0.05, and more than 10,000 records processed per GPU-hour. The headline result will be the first quantified share of European injury crashes that fall outside the Euro NCAP 2026 catalogue, with executable test cases for the ten largest gaps.
