Last month, you may have seen that a group of researchers created a machine learning system that could transform the presentation of Grand Theft Auto V in something that . It turns out that around the same time, another group of AI enthusiasts were working on something even more impressive involving Rockstar’s open world title. YouTuber Harrison Kinsley shared a video on Friday showing off GAN Theft Auto, a neural network that can generate a playable stretch of Grand Theft Auto Vthe gaming world alone.
Kinsley and his collaborator Daniel Kukieła created GAN Theft Auto with GameGAN, which recreated last year watching another AI play in the game. GameGAN, as the name suggests, is a generative adversarial network. Each GAN consists of two competing neural networks: a generator and a discriminator. The builder is trained on a sample dataset and then prompted to produce content based on what they saw. The discriminator, on the other hand, will compare the output of the generator with the original dataset, in the process of “training” its counterpart to produce content that is closer and closer to the source material.
“Every pixel you see here is generated from a neural network as I play,” Kinsley explains in the video. “The neural network is the whole game. There are no rules written here by us or the [RAGE] engine.”
Training a GAN is a very GPU-intensive task. NVIDIA loaned Kinsley a DGX A100 data center computer to make the project a reality. The $ 200,000 system comes with eight of the and two 64-core AMD server processors. Kinsley and Kukieła used all this computing power to simultaneously run 12 rule-based AIs. These programs would drive on the same stretch of highway, collecting the data the neural network needed to start generating its own game world. The two also developed an oversampling AI to clean the output of the neural network so that she doesn’t look so pixelated.
As you can see from the video, the network models a surprising number of systems in the game. As the car moves, the shadow below and the reflection of the sun on its rear windshield do the same. . The mountains in the distance are also approaching. It’s not something Kinsley necessarily expected AI to do when he and Kukieła first started training AI.
There is also a dreamlike quality to the gameplay and this is in part due to the fact that the neural network does not reproduce all aspects of GTA V Perfectly. On the one hand, collisions cause him problems. Kinsley says there was one case where he saw an oncoming police car split in two as it was about to crash with his car.
If you want to try GAN Theft Auto for yourself, Kinsley and Kukieła uploaded the project to GitHub. They say most computers should be able to run the demo.
All products recommended by Engadget are selected by our editorial team, independent of our parent company. Some of our stories include affiliate links. If you buy something through any of these links, we may earn an affiliate commission.