Grand Theft Auto V replicates almost any situation a driver could face: city traffic, pedestrians, car accidents, clogged freeways. What better school for self-driving cars to learn how to drive?
MIT Technology Review reports that a team of researchers at Intel Labs and Darmstadt University in Germany is using GTA V as the basis for a machine learning program that could lead to safer, better self-driving cars.
Machine learning is the process by which computers learn, taking in and crunching huge amounts of data to learn lessons that can then be applied generally to a variety of situations. In other words, pretty much what human beings do through trial-and-error, only computers can achieve this much faster than we can.
In order for machine learning to work, however, you need to generate the data for computers to process. When it comes to all the variables that a self-driving car has to account for in order to navigate the road safely, generating that data can be a labor-intensive process that slows down the computer’s ability to learn.
Enter Grand Theft Auto V, a game that, when patched with the right mods, can look almost as good as film. Urban driving, suburban driving, dirt roads, bicyclers, Lear jets landing on four-lane highways, GTA V has it all. What the researchers at Intel Labs and Darmstadt University figured out is how to break the game images down into categories that artificial intelligence routines can recognize.
If an AI knows where the roads, buildings, and people are it can learn how to drive safely on those roads. And GTA V’s graphics are so advanced, even approaching photorealism without mods, that an AI learning on GTA V might be able to take those lessons and turn them into practical knowledge for real world driving.
This technique could speed up the process by which we finally all get our self-driving cars, just as long as we make sure said cars don’t learn the wrong lessons from playing GTA V. For instance, driving like this:
H/T Daily Mail