Nvidia’s AI recreated ‘Pac-Man’ from scratch by watching it

Nvidia's AI recreated Pac-Man from scratch.
Image: Nvidia

Nvidia is best known for its graphics cards and other computing hardware. However, it is starting to lean more heavily into the artificial intelligence (AI) space. Nvidia contributes research to the field on a regular basis and has some impressive accomplishments.

Its latest project is certainly noteworthy. A team of Nvidia researchers trained an AI system to recreate the game “Pac-Man” from scratch. The system was allowed to watch the game being played and then made a near-exact replica without any human coding. It comes at the perfect time as “Pac-Man” is celebrating its 40th anniversary on Friday.

Watch and Learn

Typically, AI learns through trial and error by building off of existing data. This experiment was unique because the AI model was only fed visual data of “Pac-Man” being played along with the accompanying controller inputs. From that information alone, the AI was able to recreate the game in its entirety. It didn’t have access to any pre-rendered images or human code to go on.

The result was a playable rendition of the classic arcade game. To be fair, it isn’t a perfect replica. Much of the game’s imagery is blurry. Moreover, the AI didn’t appear to replicate the unique personalities of the ghosts. In the real version of “Pac-Man,” each ghost acts differently and moves according to its programmed behaviors.

Nonetheless, Nvidia’s AI version has all of the game’s basic elements. Players eat pellets while avoiding ghosts and navigating a maze.

Nvidia’s vice president of simulation technology, Rev Lebaredian, says, “[It’s] similar to how a human programmer can watch many episodes of ‘Pac-Man’ on YouTube and infer what the rules of the games are and reconstruct them.”

The team used a program called GameGAN to recreate the iconic arcade title. It works in two halves—one attempting to replicate input data while the other compares it to the original source. Any data that doesn’t match is rejected and the process starts anew until a match is made.

Lebaredian notes that the team worked in collaboration with “Pac-Man’s” creator, Bandai Namco to bring the project to life for the game’s 40th birthday. Nvidia plans to release the AI-generated version online in the near future.

Future Implications

While recreating an arcade game is a neat experiment, it isn’t the sort of application that gets people excited about the future of AI. Still, the Nvidia team notes that it is building the foundation of AI-driven video game design.

For instance, developers could input their work into the AI algorithm and use it to create variations or even new levels. Sanja Fidler, director of Nvidia’s Toronto research lab, notes, “You could use this to mash different games together giving additional power to game developers by [letting them] blend together different games.”

Meanwhile, researchers envision a similar AI learning actions in the real world by watching them play out. The system could do things like observe automated trolleys in a warehouse and then create its own navigation software.

GameGAN is impressive and the early results of projects like this inspire plenty of hope for the future of AI.


Please enter your comment!
Please enter your name here