GAN Theft Auto is a project where Nvidia’s GameGAN (Generative Adversarial Networks) technology is used to train an AI to try and build its own version of Grand Theft Auto V. And while of course it’s far from perfect, it’s also pretty instantly recognisable as GTAV.
sentdex explains the project in the video below, where the AI was trained by showing it loads of different examples of cars driving in the game, and by extension shown stuff like what cars and the environment looked like and then set to work.
What’s really important to note here is that this isn’t just simulated footage, this is playable. And while what’s available to the player is pretty basic — you can only drive a car around — this is still pretty damn incredible to see, considering this is something that was just conjured by an AI.
Even the visuals are impressive. Yes, there’s the trademark, watery “dreamscape” look to everything given the fact this was made by computers and not humans, but squint past that and you can see that the AI hasn’t just replicated the overall aesthetic of the game, but even some of its visual effects as well, like the way the car rolls slightly when turning.
To show you how the AI was trained, here’s a look at GameGAN being fed the footage, then chomping it all up and learning stuff like which parts are the road.
This is the second time in two months I’ve got to write about AI and Grand Theft Auto V. Back in May, we saw a demonstration of how AI could be trained to make the game’s visuals almost photorealistic, and the results — albeit from a short demo — were pretty damn close to achieving that.