After Intel researchers displayed a photorealistic version of Rockstar’s GTA V using machine learning last month. It turns out, taking inspiration from that a group of AI enthusiasts has built an AI tool capable of creating a playable world of GTA V on its own in real-time with impressive details like car shadow, a reflection of the sun, and much more.
As first spotted by Engadget, on Friday, Harrison Kinsley shared a video on Youtube showing GAN Theft Auto, a neural network that can generate a playable stretch of the GTA V world on its own. Harrison Kinsley and collaborator Daniel Kukiela made GAN Theft Auto – AI version of GTA V – with help of Nvidia’s GameGAN – a Generative Adversarial Network consisting of two competing neural networks: a generator and a discriminator. Kinsley told in his video that “Every pixel is generated from a neural network. There are no rules written here by us or the [RAGE] engine.”
As Engadget notes, “The generator is trained on a sample dataset and then produces the content based on what it saw. Meanwhile, the discriminator compares the output of the generator with the original dataset, in the process and “coaches” its counterpart to output content that is closer and closer to the source material.”
To make this project a reality Nvidia loaned Kinsley a DGX station A100 – which has four A100 GPUs and a 64-core AMD server CPU. And by seeing the end result it looks like Kinsley and Kukiela have used all computing power of the rig to run 12rules-based AI. They also explained that they have to develop a supersampling AI, which helps in cleaning the output, so that game won’t look pixelated.
If you want to check GAN Theft Auto (AI version of GTA V) by yourself, thankfully Kinsley and Kukieła have uploaded the project on GitHub and said most computers would able to run the demo. Last year using the same GameGAN the group recreated Pac-Man by watching another AI play through the game.