Safety evidence for automated vehicles requires extensive field tests, which are extremely monetary and time-consuming. To reduce the testing efforts, validation can be conducted in the virtual environment. This has been studied for several years in the field of AD/ADAS. However, a generic method for virtual validation as well as a standardized workflow and tool chain, which can be easily adopted and implemented by industrial partners, is still necessary. Furthermore, the transferability of this method to powertrain is of interest as well, since it can potentially save great cost for testing.
In a first step, real world and simulated data describing relevant vehicle and environmental parameters for virtual validation scenarios are aggregated. The data is then analyzed and the relevant vehicle and environmental features (e.g., speed, acceleration, traffic density, number of lanes etc.) are extracted. A statistical analysis of the extracted features and description of their correlations is conducted. This statistical parameter representation is the basis for the generation of a realistic baseline scenario set, represented in the OpenX format.
Based on the extracted baseline scenario, AI-based scenario generation and variation is conducted to explore the scenario space regarding both coverage and critical cases. Here importance sampling is used to limit the effort for testing and reinforcement learning is responsible for generating rare and critical cases. Scenario variation interacts with a multi-agent simulation encapsulated in an industrial simulation environment. In this way, virtual validation can be carried out and its plausibility can be checked by comparing selected parameter distributions of virtual validation against real word dataset.
