A team of researchers from the University of Copenhagen has developed an AI to generate Super Mario Bros levels based on a series of words entered by the user. A tool more scientific than video game.
Even Mario video games can be generated indefinitely! After ChatGPT popularized the use of artificial intelligence to generate complex responses to user requests, MarioGPT proves that the recipe can also be applied to video games.
This is demonstrated by six researchers from the University of Information Technology in Copenhagen in an academic article pre-published online on February 12. Based on the first two Mario games (Super Mario Bros. And Super Mario Bros: The Lost Levelsreleased in 1985 and 1986), their tool makes it possible to generate random, but playable 2D levels of Nintendo’s hero, while adapting to the desires of users.
“Many pipes, many enemies, little blocks”
If creating new levels for the mustachioed plumber was already possible with the series of games Super Mario Maker or through other unofficial level generators, MarioGPT is the first to “generate various levels, but can also be based on textual proposals to control the generation of levels“, explain the researchers in the introduction of their article.
Indeed, the content produced by MarioGPT is linked to ” quick“, these more or less thoughtful combinations of words that we type in the search bar of an image-generating AI like Dall-E in order to obtain more or less precise results.
In the case of MarioGPT, the AI allows you to play on the frequency of certain game elements, enemies, or even on the general height of the level. The researchers’ article gives some examples: lots of pipes, lots of enemies, few blocks, low altitude“.
A different method of ChatGPT
To achieve this, the Danish researchers relied on the artificial intelligence model GPT-2, specifies the online media TechCrunch: a version older than the one used by ChatGPT and less good at mimicking human language.
But this is not a problem in the case of MarioGPT, since the group of academics used it instead to generate high-speed computer code: MarioGPT gradually learned to identify similar code patterns (gameplay elements, enemies, etc.) before repeating them according to the logic of the game.
Available for free… for insiders
If you want to test the program, it is available for free on Github… but it will be a bit more complex than using ChatGPT. MarioGPT requires some basics in video game development and knowing how to run code in Python 3.8, details the information siteSlash Gear.
Similarly, generated game levels are limited to one type of enemy generated at a time, rather than varying goombas and koopas. Because the objective of MarioGPT remains more scientific than playful: to prove that the procedural generation of video games by prompts is possible.
But Danish academics are not closing the door to improving their tool. As researcher Shyam Sudhakaran explains in his interview forTechCrunch“in the future, with larger datasets and more complicated prompts, we may need to use a more sophisticated model like GPT-3.»
Do you use Google News (News in France)? You can follow your favorite media. Follow Frandroid on Google News (and Numerama).