Massively Parallel Virtual Testing of Safety-Relevant Driving Systems
AI/Machine Learning/Deep Learning
TimeTuesday, June 18th1:45pm - 2:07pm
DescriptionTesting and validation are cost-intense parts in the automotive development process already today. On top of that, the emerging development of autonomous driving vehicles is demanding scenario and validation spaces being by orders of magnitude more complex than nowadays tasks and moreover not well-defined. E.g. the correct classification of objects depends on a vast variety of attributes, such as shape, color, texture, reflectivity, road infrastructure, ambient weather conditions are only one example. How to truly validate such complex systems in such complex variety of conditions? For many of such safety-critical systems this task has to be mastered by means of virtual validation and cover at least millions of virtual tests. Obviously, this tasks these have to be executed as fast as possible to enable a continuous development process for the development of autonomous systems.
Using high-performance computing (HPC) as well as artificial intelligence (AI) and machine learning (ML) algorithms we developed the framework VALICY, which is designed to master virtual validation tasks of high complexity. Systems under validation, e.g. decision functions in car control units or sensor functions of end-users, have been successfully integrated into the framework, and by that, validated on HPC systems on thousands of processors guaranteeing e.g. overnight feedback even for 15-20 dimensional variational test spaces. To exemplify the application of the framework, validation scenarios for camera-based object detection and classification functions will be shown and discussed in this talk.