Superbrains for self-driving cars

Today’s modern driver assistance systems already use AI for decision-making and supporting vehicle drivers. They have to detect the traffic situation in a split second by means of sensor systems such as lidar, radar and cameras, and react appropriately. Yet there are many hurdles to overcome before a system can be used for autonomous driving. Whereas a human being recognizes a car after seeing a few dozen pictures, an autonomous driving algorithm has to learn to distinguish between pedestrians, trees and other vehicles. To this end, it is ‘fed’ with millions of pictures from real-life traffic situations.

The algorithm learns gradually, through trial-and-error, until the machine knows how the world surrounding the vehicle works, and how the vehicle should behave in it. So far most of the data for simulations has come from test drives taking place all over the globe. Every day, hundreds of vehicles are being driven to generate many terabytes of data. In the automotive industry deep learning and virtual simulations have become established for training the systems. Deep learning emulates the way a human brain learns. The machine’s artificial neural network learns through experience and links new information with existing knowledge.

All that is, however, very complex and time-consuming. New supercomputers can create the data artificially. In addition, they can play out any scenario and any eventuality in road traffic. For example, the car in front brakes suddenly, a child jumps onto the roadway, or the sun dazzles the on-board cameras. According to experts, supercomputers allow the normal product cycles in the automotive industry to be shortened – from an average of 60 months to 24. Continental’s supercomputer in a Frankfurt datacenter, for instance, uses more than 400 graphics processors from NVIDIA and carries out quadrillions of computing operations in one second (> 50 petaflops). In IT, that’s the equivalent of indicating a car’s horsepower – in this case, something like a Formula 1 racing car.

This computing power puts the supercomputer at the top of the pile in the vehicle industry. Without it, thousands of hours of training would be needed to handle the millions of images and huge volumes of data. The powerful computer, on the other hand, reduces the necessary time from several weeks to just a few hours. The manufacturer ŠKODA AUTO has also taken a high-performance computer into operation at its Czech headquarters in Mladá Boleslav. Divisions such as Production and Technical Development use the computer for visualizations, virtual reality presentations and simulations in the areas of aeroacoustics and aerodynamics, and in developing the engine portfolio. Drivers will be pleased because thanks to the supercomputer combined with advanced engineering, they can look forward to a new level of comfort and mobility.

(Stagephoto © Volkswagen AG)