This version of the GTA V is made by AI only

Nvidia and AI are a long love affair. The company is constantly developing new systems, often based on the negative networks it creates (or GAN), And are often very enjoyable. Recently, for example, he gave us the news Canvas, His “landscape painting” system.

But with Camcon, Raises the green panel bar even higher: This completely crazy system aims to emulate a game visually by recording only the image and controls. Nothing else. No fate, no goal, nothing! At first glance, this system does not seem large .. Besides, games like Doom, or Backman can be reproduced with a reliability. Decide for yourself: left, original and right, AI-generated version.

The result is almost identical, and not just visually: the GAN version operates in a completely similar way. Very definite, and inevitably, some people quickly wondered how far this idea could go. You know Grand Theft Auto, now here … GAN Theft Auto. You will understand it; The training of an AI wizard (YouTuber Syntex) decided to avoid all activities GTA V from Back-Man. After we imagined the longest hours of training on a special machine (PC with 4 Nvidia A100 GPS and 64-core AMD Epic processor), he achieved a breathtaking result.

“Watch and learn“

Those who have already played the famous Rockstar title will immediately recognize the game, despite the numerous artifacts. The one in front of you is not reusable by AI. It really isRelease Directly from GAN: Syntex operates entirely on the visual representation of the neural network. Usually, we set up an AI in an environment; Here, AI’s just the environment!

It may seem trivial, but this is the most incredible feature of Camcon. In this system, there is no game machine and no pre-written code by the physics or manipulator. Images only, and player controls. This means that in the GTA example, the Camcon expands everything you see on the screen. This system “rediscovered” the effects of light, shadows and the physics of the vehicle. He also distinguishes between standard elements of decor and dynamic materials. The idea doesn’t stop there. If you provided enough game footage for AI, you could recreate the wholeness of this complex game.

This includes abstract concepts such as weapons, kidnapping, civilian but police chase. For example, during its demonstration, Syntex showed that it had learned to manage its structure His car collisions. Of course, we are still far from the real game; But in the end, it’s a matter of AI training time, and there is no ideological limit. Absolutely crazy, as we recall, GAN created all of this only from pictures without any physical machine.

What is GAN?

To accomplish this feat, in another era, a GAN would be based on two components, ending in a witch experiment. The first is a generator, which we will train with the original game. Its role is to produce the images you see. He would then show his plan discriminating against second parties. Its role is to try to distinguish truth from falsehood; So he compares the generator’s program to the actual game sequence. If the discriminator makes a credible (“true”) judgment, assuming it is consistent, the rendering confirms this point. Or, as Nvidia puts it succinctly, “The generator is trying to deceive the discriminator, trying not to fall into the trap ”.

AI used for video games is definitely a growing field. There is no doubt that this system is still set to progress; Seeing as Camcon has gone from back-man to GTA V in a year, the fantasies to come are allowed for many years to come. When was the first AAA title created by AI?

Everything becomes even more unbelievable when we imagine using such a system in real life. For example, can this system revolutionize simulation techniques? Is it possible to even model real world events using this system? The future will tell. But if that happens occasionally, one article may not be enough to discuss it!

For those who want to try GAN Theft Auto and have technical skills, the necessary equipment is available GitHub.