NVIDIA's AI learns the rules of Pac-Man and creates a 'fully functional' clone

The 40th-anniversary recreation could change how we teach autonomous machines.

For Pac-Man’s 40th anniversary on Friday, NVIDIA announced it had recreated the iconic arcade game without the support of a game engine. After feeding GameGAN, a generative adversarial network, 50,000 episodes of the game, the graphics company produced a faithful (and admittedly blurry) clone with full functionality. GameGAN is a new AI model that could eventually do more than create game levels it’s never seen before; it could train autonomous robots and vehicles.

A new way to create Pac-Man NVIDIA worked with Pac-Man’s creator Bandai Namco to develop the AI model. GameGAN uses competing neural networks to learn the rules of the game and keep track of the layout. This functionality combined with the model’s ability to separate different parts of the game from each other can help developers try out new levels, characters, and themes.

The training process was too much for humans, but by using AI, there were initially issues with teaching the model the concept of dying in-game, according to The Verge. To combat this, early versions had all the ghosts follow Pac-Man, but not quite reach him.

Non-gaming applications — The rule and environment-learning capabilities of GameGAN could help future machine learning programs. The AI can save time by creating simulations instead of requiring developers to do so, with applications like learning the layout of a warehouse and training self-driving cars.

“We could eventually have an AI that can learn to mimic the rules of driving, the laws of physics, just by watching videos and seeing agents take actions in an environment,” said Sanja Fidler, director of NVIDIA’s Toronto research lab, in a statement.

For now, it’s a happy birthday to Pac-Man, but soon this kind of AI could power a wide range of autonomous machines.