Researchers upgrade Super Mario Bros with AI

Computer programmers at IT University of Copenhagen have created a revolutionary new way to generate Super Mario Bros. The Danish programmers are reinventing the gaming landscape—MarioGPT, based on the GPT-2 language model. The team has released their findings in an open-access paper on the arXiv pre-print server.

AI is reinventing the classic arcade game

Mario Brothers debuted in 1983. It features two Italian plumbers coming from a sewer to rescue Princess Peach from a captor, Bowser. They must circumvent obstacles crafted from pipes and bricks to save the princess. The Danish researchers have taken the game to a new level by introducing simulated terrains that evolve depending on the level a player reaches.

Researchers have reinvented the rules of level-building for Super Mario Bros, using OpenAI’s GPT-2 open-source language model. By creating a small bit of Python code and training it with samples from the classic Super Mario Bros and another sequel, “Super Mario Brod: The Lost Levels,” they could translate user requests into exciting graphical representations that capture the essence of the thrilling video game franchise.

MarioGPT generates responses to user prompts by creating custom levels for their gaming adventure. The AI tool responds to phrases such as “many enemies, many pipes, low elevation, little blocks” Users can find the instructions and code for encoding and creating levels on GitHub. MarioGPT allows users to explore as many levels as they desire.

In other news, researchers at New York University Tandon collaborated with a colleague at the University of the Witwatersrand in a project similar to MarioGPT. The team leveraged large language models in generating and playing versions of Sokoban. The game involves players pushing boxes in a warehouse. They also published their project in a paper on the arXiv preprint server.

The featured image was taken from Super Mario