When the developers of artificial intelligence for the autopilot are convinced that it has learned how to perform standard tasks, they are faced with the task of testing their brainchild in obviously dangerous situations. But where to get so many unpredictable drivers, how to create unforeseen situations on the road and involve AI in resolving conflict issues, without damaging anyone or anything? Intel figured out how to adapt the most reckless game for these purposes, where permanent chaos on the road is the norm.
At the Technological University of Darmstadt, together with American IT specialists, they developed a system for transmitting visual information from the game world to the sensors of a real autopilot. The artificial driver does not play the game, but he reacts to the events that unfold on the streets of Los Santos as completely authentic phenomena.
This is convenient - those who, at least sometimes, just traveled around the virtual world of Grand Theft Auto, without chases and shootouts, certainly paid attention to the high detail of aspects of traffic. There are signs, traffic lights, pedestrians of varying degrees of impudence, and both obedient and inadequate drivers. And they sit behind the wheel of a variety of vehicles, so that the desired variety of traffic is provided.
The drone is not faced with the task of going through this or that adventurous mission, on the contrary, they learn to drive safely in a constantly changing environment. This includes the appearance of animals on the road, and the change of weather, and contact with the police, and driving through very difficult junctions and intersections, as well as training in building the optimal route.
An additional argument in favor of learning the autopilot on the streets of the game world - engineers can simulate any situation and run the test countless times. Up to drive 3 million km per day, meticulously noting how the AI reacts to minor changes in the script. However, there are doubts that the authorities will count these virtual trips as real experience in training autopilot systems. Therefore, while everything is theoretical in nature and the "brought up GTA" robot will not soon go on the real road.