Practical validation of synthetic pre-crash scenarios

A team of researchers (Wu, Sander, Flannagan, Bärgman) published a paper on arXiv addressing a critical gap in autonomous vehicle safety validation: proving that synthetic pre-crash scenarios are practically equivalent to real-world crashes for simulation-based testing. Traditional statistical significance tests are ill-suited because they detect differences, not similarity. The team extends their previous Bayesian Region of Practical Equivalence (ROPE) framework by introducing a binning-based approach that defines appropriate statistics and equivalence criteria. This allows them to measure practically meaningful distributional differences between synthetic and real crash datasets.

The framework's applicability was demonstrated through a case study on rear-end pre-crash scenarios, testing two synthetic datasets against a reference dataset in the context of an Automatic Emergency Braking (AEB) system safety assessment. Results showed informative quantitative assessments of practical equivalence and diagnostic insights into where datasets diverge. Although focused on rear-end crashes, the method is generic and extensible to other validation contexts, providing an interpretable and principled basis for equivalence assessment across diverse synthetic data applications.