Last month, you could have seen {that a} group of researchers created a machine studying system that might remodel the presentation of Grand Theft Auto V into one thing that appears virtually photorealistic. It seems, at about the identical time, one other group of AI lovers have been engaged on one thing much more spectacular involving Rockstar’s open world title. On Friday, YouTuber Harrison Kinsley shared a video showing off GAN Theft Auto, a neural community that may generate a playable stretch of Grand Theft Auto V’s sport world by itself.

Kinsley and collaborator Daniel Kukieła made GAN Theft Auto with GameGAN, which final 12 months recreated Pac-Man by watching one other AI play by way of the sport. GameGAN, because the title suggests, is a generative adversarial community. Every GAN consists of two competing neural networks: a generator and a discriminator. The generator is skilled on a pattern dataset after which informed to supply content material based mostly on what it noticed. The discriminator, in the meantime, will examine the output of the generator with the unique dataset, within the course of “coaching” its counterpart to output content material that’s nearer and nearer to the supply materials.

“Every pixel you see here is generated from a neural network while I play,” Kinsley says within the video. “The neural network is the entire game. There are no rules written here by us or the [RAGE] engine.”

Training a GAN is a really GPU intensive process. NVIDIA loaned Kinsley a DGX Station A100 laptop to make the undertaking a actuality. The system comes with 4 of the corporate’s A100 GPUs and a 64-core AMD server CPU. Kinsley and Kukieła used all of that computing energy to run 12 rules-based AI concurrently. Those packages would drive the identical stretch of freeway, amassing the information the neural community wanted to start out producing its personal sport world. The two additionally developed a supersampling AI to wash up the output of the neural community so it wouldn’t look so pixelated.

As you’ll be able to see from the video, the community fashions a shocking variety of techniques from the sport. As the automobile strikes, so does the shadow beneath it and the reflection of the solar on its again windshield. The mountains within the distance even get nearer as effectively. That’s not one thing Kinsley essentially anticipated the AI would do when he and Kukieła first began coaching the AI.

There’s additionally a dream-like high quality to the gameplay and that is thanks partly to the very fact the neural community doesn’t replicate each facet of GTA V completely. For one, collisions give it hassle. Kinsley says there was one occasion the place he noticed an oncoming police cruiser cut up into two simply because it was about to crash together with his automobile.

If you need to strive GAN Theft Auto for your self, Kinsley and Kukieła have uploaded the undertaking to GitHub. They say most computer systems ought to have the ability to run the demo.