Generation of Virtual Test Scenarios for Training and Validation of AI-based Systems

Technologies in the field of artificial intelligence (AI) are increasingly used in a wide variety of industries. However, in addition to their considerable potential, it is important to note that the behavior of an AI-based system cannot be predicted based on its internal architecture, and thus its correct behavior cannot be guaranteed. Instead, both the performance and the reliability depend on the amount and quality of training data. The greater the complexity of a task required from the AI, the less likely it is that the necessary training and validation data can be collected through real measurement series in practice. A popular example is the situation awareness required for autonomous driving, which requires an ever-increasing amount of kilometers to be driven. Therefore, people are now gradually shifting to the inclusion of synthetic training data generated by simulation. This is where this paper comes in. We present a concept for a flexible generation of virtual test scenarios based on a systematic use of digital twins and virtual testbeds, allowing to generate training and validation data for AI-based systems in appropriate quantity, quality, and time.