For no other field is the fact that innovations happens at the crossroads of disciplines more valid than for autonomous driving. AI, neuronal networks, machine learning, sensor technologies, processing power, behavior science, virtual reality and simulations push the development of autonomous cars forward.
When recently Waymo-CEO John Krafcik reported that the Google cars drove in addition to the 3 million miles in real world city traffic also 1 billion simulated miles, and that those simulated miles are the dominant form of improving the technology, even experts looked astonished. Not surprisingly, as simulations allow re-experiencing scenarios that in reality the cars would rarely encounter. In a simulator those scenarios can be repeated as often es required, and under as many conditions and variations as wanted.
Other startups bet on simulations as well to advance their technologies and algorithms. The Hungarian company AImotive developed a simulator that helps to identify objects better. The objects are being annotated to categorize them properly.
The German startup understand.ai works on automatic annotation with its AI-system. In every traffic situations autonomous cars must recognize, which objects are cars, pedestrians, bicyclists and other objects. With videos from real traffic or simulations a trained AI-system can recognize them in real-time. This approach is not only suitable for autonomous cars, but also for cancer diagnosis or the analysis of drone images.
It’s fitting that the students in Udacity’s nanodegree for self-driving use video games and simulations as well to improve their algorithms. Here is a more detailed blog of how the following video was created.
This article is also available in German.