Providing Evidence for the Validity of the Virtual Verification of Automated Driving Systems

Abstract

With the increasing complexity of automated driving systems, formal verification as well as statistical verification that solely relies on real-world testing methods, become infeasible. Virtual testing seems like a promising alternative to traditional methods, especially as part of a scenario-based verification and validation methodology. But in order to transfer the test results of a system from a simulation to the real world, we need to argue the validity of the virtual tests. Our proposed method enables this validity argumentation by comparing the virtual test traces against traces that have sufficiently similar recorded real-world traces. To reduce the amount of required real-world data, the method involves two mechanisms to generalize the validity statement of a single real-world trace to a set of virtual traces. The reduction of required data is showcased in a proof of concept that compares the needed amounts of data with a "naive" validation method and here presented enhancements in an ablation study.