The Georgia Institute of Technology Research is developing an "artificially intelligent system" that can build Mario stages by watching YouTube or Twitch videos.
Nintendo is making significant efforts to allow players to make their own Mario levels, but researchers at the Georgia Institute of Technology have an intriguing take on the matter. A new computing system is designed to watch gameplay video on YouTube, Twitch or other such streaming services, analyze the footage, and create original new stages for the game.
This is the first discovery of its kind, and researchers tested this artificially intelligent system with the popular Super Mario Brothers, which will also allow the system to replicate results across other similar games as well.
As the Georgia Institute of Technology explains, the automatic level designer focuses on the gameplay, not the playable character itself, while also taking into account the locations of on-screen elements such as pipes, coins, blocks or other. Through careful analysts, the system determines the level design rule or other requirements.
Mario games, for instance, feature pipes that come out of the ground, so the system take it into account and creates its stage levels accordingly, avoiding the combination of pipes with grassy surfaces. The system also uses spatial analysis to ensure the jumps the hero must make, for instance, are not impossibly long.
According to Matthew Guzdial, lead researcher and Computer Science PhD student at Georgia Tech, the initial evaluation of this system suggests that it is able to develop level sections that are not only playable, but are also close to the original without the need to hand code design criteria.
By observing gameplay video to determine where players spend most of their game time, the system identifies the high-interaction areas and focuses on them to draw design information.
"Our system creates a model or template, and it's able to produce level sections that have never been seen before, do not appear random and can be traversed by the player," further explains Mark Riedl, associate professor of Interactive Computing. "One could say that the system 'studies' the design of Mario levels until it is able to create new playable areas."
The automatic level maker from Georgia Tech tool 17 samples of the original game and created 151 distinct level sections, keeping tabs on style variables and overall playability. The system then increased its output to 334 level sections as it reduced the limitations. According to Georgia Tech, all levels can be ported into a game engine to be played.
The researchers plan to create full-scale levels in the future, as well as evaluate how players interact in such levels and determine whether it's different than the original gameplay videos.
