Experts at TÜV Süd have estimated that autonomous vehicles’ AI systems need to be able to fully master around 100 million situations for each automated driving feature. The algorithms these situations are based on use deep learning to make (the right) decisions autonomously. TÜV Süd now intends to have them validated and certified for road safety. The crux of the undertaking is a platform called Genesis on which users will be able to upload their data and modules.
According to Dr. Christian Müller of the German Research Center for Artificial Intelligence (DFKI), the industry is showing keen interest. TÜV SÜD’s Dr. Houssem Abdellatif adds that the results achieved to date using deep-learning methods are “astonishingly good,” although no one knows yet exactly “what’s actually happening.” For this reason, one of the goals of the collaboration is to research this using virtual traffic situations. A practical test follows the theoretical one – just like the normal requirements for new drivers. The results will be used to develop a certificate – a kind of “driver’s license” – that verifies the safety of an algorithm’s driving skills.
Another strategy to increase the functional safety of AI systems in autonomous vehicles is described by well-versed software tester Dr. Peter Liggesmeyer , managing director of the Fraunhofer Institute for Experimental Software Engineering IESE , in Capital Magazine: monitoring the systems using conventional software. The AI solution could be used to help deal with complex actual circumstances, while the conventional software would intervene when artificial intelligence fails – not unlike a safety net for a trapeze artist, Liggesmeyer argues. It would then be able to bring the vehicle to a halt at the edge of the road, for example.