Researchers are turning to an ironic source for help in teaching machine learning programs about how to drive: the chaotic vigilante driving game Grand Theft Auto.
Bear with us here, because this actually makes a ton of sense. Researchers at Intel Labs and Darmstadt University in Germany could map their own simulators to create real-world scenarios, but there’s already a program out there with a realistic driving experience complete with pedestrians and other vehicles: Grand Theft Auto.
So they’ve essentially annotated a version of the game to give cars real-world driving experience. No, they don’t get guns or any of the “other” parts of the game, but anyone who has played GTA knows that the driving conditions are pretty true to real life. So assuming you’re not trying to terrorize a fictional city, having an AI use it to learn about driving makes sense.
Here’s a look at how it maps:
Of course if the machine learning program does hit a pedestrian, and manages to get stars, it might learn a lot of lessons we don’t want it learning. Grand Theft Auto’s police force has more of a shoot first approach to law enforcement.
It would be pretty disheartening for an AI system to learn that pulling over for the cops is dangerous. This is how you get Skynet.
[H/T MIT Technology Review]
Correction: the original version misidentified the name of the university using Grand Theft Auto for machine learning
