GAN Theft Auto Is A Neural Network Based On NVIDIA GameGAN

Thanks to Artificial Intelligence, Researchers have created a new AI-trained game engine called GAN Theft Auto which imitates the driving experience in GTA V. The engine was created by developing and training a NVIDIA GameGAN based neural network.

Don’t get confused, the GAN Theft Auto isn’t a neural network trained to play GTA5, rather a simple engine that is trained to copy GTA5’s gameplay. In 2020, GameGAN (Generative Adversarial Networks) was used to recreate a PAC-MAN game. The engine required 50,000 episodes to be fully trained. The episodes were created by the neural network.

Two competing neural networks, a generator, and a discriminator to create a convincing copy of the game engine are required to create any GAN-based game. It was easy to create the PAC-MAN since it is a simple game that was fairly easy to train, however, GTA V is a different story.

The researchers also borrowed the DGX100 Station from NVIDIA. Since neural networks require a long time to train, the researchers came up with an interesting method. To feed the data they did not launch the game multiple times instead they spawned multiple instances of cars driving in the same sandbox, except the cars were clipping through each other and were made invisible.

The simple version of GAN Theft Auto gameplay is capable of only providing a very basic highway driving experience and that too with substandard graphics but they were improved using upscaling. Luckily the network was able to detect collisions of other cars and obstacles.

This is just a hint of what neural networks can do. Soon, game developers may use this to create an alternate storyline or game physics.

Via Sentdex