Yes, the original looks a lot better. (Screenshot: Rockstar Games)
GAN Theft Auto is a project that uses Nvidia’s GameGAN (Generative Adversarial Networks) technology to train an AI which then tries to create its own version of Build Grand Theft Auto V. The result does not seem to knock the inclined observer off the chair, but is in any case very clearly recognizable as a GTA.
The AI ​​can already build simple games like Pacman
It is also playable. The GameGAN fork shows how capable the technology is in handling games that are more complex than Pacman. On the occasion of its fortieth birthday, Nvidia had the cult game from the eighties recreated by an AI.
The Nvidia AI was there proceeded as a human developer would have done. She tried to understand the rules of the game by watching and then to imitate it programmatically. Then she trained with 50,000 Pacman runs. In the end there was a fully functional version of the game that works without any underlying game engine.
This is how the KI built GTA5
Github user Sentdex proceeded similarly. In the following video he shows that the AI ​​was trained on the basis of various game scenes by showing it examples of cars driving in GTA5, with the main aim, but not only, of differentiating the surroundings from the vehicle.
In the end, Sentdex succeeded in creating a playable version of an – admittedly very simple – section of the game. Visually, the result is not necessarily a pleasure. The environment looks as if you are looking at it with eleven diopters, where you don’t need glasses.
Nevertheless, for the fact that the result was developed by an AI and is fully playable, we can be clearly impressed. By the way, a few months ago an Intel AI dealt with GTA5.
