AI Recreates Pac-Man From Scratch, Without Any Human Input
Researchers at Nvidia have managed to train an AI to develop and manage its own version of Pac-Man.
Today, one of gaming's most recognizable icons turns 40. And to celebrate, Nvidia and Bandai Namco have partnered up to recreate Pac-Man, only this time without any actual human input.
Using Nvidia's GameGAN, the company has successfully trained an AI model to recreate and play Pac-Man without the use of a game engine. Instead, the model recreates each frame of the game manually on screen, and tracking not only the movement of the pie-chart protagonist but the pellets it eats, the ghosts it gobbles, and the score it racks up.
The AI was trained using millions of frames from the original Pac-Man, supplied by Bandai Namco, along with recorded player inputs. After chewing through 50,000 games, the model was able to replicate static elements of a simple Pac-Man game, while also understanding how to operate its ruleset, without a single line of game code supplied by the developers.
This is a neat homage to one of the foundations of gaming as we know it, but it's also a glimpse into the future of AI and its role in gaming. Nvidia and its researchers in this field believe this could eventually be used to help game developers iterate on level designs by feeding an AI model existing level layouts, letting it generate additional content that fits with what already exists. It could also be used in both virtual and real-world applications of teaching AI the simplicities of human behaviors, to better recreate them either in a game or outside (like self-driving cars, for example).
For now, though it's an intriguing glimpse into the growing power of AI and Nvidia's role in it, which should expand now that the company has revealed its next-generation Ampere processing architecture. You can read the full research paper on the GameGAN AI model here.
GameSpot may get a commission from retail offers.
Got a news tip or want to contact us directly? Email firstname.lastname@example.org