A artificial intelligence has recreated a short playable section of Grand Theft Auto V, without using a graphics engine or a single line of code. The experiment, called GAN Theft Auto, has been carried out with the red neuronal GameGAN created by Nvidia, which has been taught multiple fragments of gameplay to do the recreation, according to PC Gamer.
Before going into details a bit of context is necessary. In May 2020 Nvidia released a version of Pac-Man (which last year celebrated its 40th anniversary) created by this artificial intelligence. GameGAN is, in the words of the graphics card manufacturer, “an adversarial generative network”.
“The [redes] GAN can create images that look like photographs of human faceseven if the faces do not belong to a real person. GANs can achieve this level of realism by teaming up with a generator, which learns to produce the objective result, with a discriminator, which learns to distinguish real data from the generator’s results, “explained Nvidia in May of last year.” The generator tries to fool the discriminator, and the discriminator tries not to be fooled. “
Without code and without graphics engine, GameGAN created by itself the game
The video of GAN Theft Auto, from a playable demo downloadable on GitHub, shows a game blur in which the environment blends and blurs with the movement of the car. It is not a current game in terms of graphic quality, but there are very curious details such as the realistic shadows under the car and the sun’s reflections on the vehicle windows.
The demo is the result of teaching the neural network dozens of gameplay of researchers Harrison Kinsley and Daniel Kukiela driving down a highway GTA 5. GameGAN, run on a powerful workstation, I recreated the aforementioned demo, incorporating by itself vehicle physics, how the environment moves and even the controls. The environment itself is the neural network; the landscape is generated while the user plays.
At the moment it is unknown if there is any development of a vcommercial ideo game that is using this technology to recreate sections of a game or entire titles, but it is, to say the least, interesting. A decade ago it would be unthinkable to think of image reconstruction techniques such as Nvidia DLSS and today it is increasingly common to see this technology or others like it in computer games and consoles.