Researchers from the Georgian Institute of Technology (USA) have developed an Artificial Intelligence (AI) able to recreate a video game by analysing gameplay videos frame by frame. With no prior knowledge of the source code.
The research team presented their findings at the recent 26th International Joint Conference on Artificial Intelligence held in Melbourne in August with a paper entitled Gaming Engine Learning from Video.
They showed their new approach to AI can create a game engine based on source video footage, covering character movement and graphic rendering. This technique was developed for 2D platform games, and the group have showed success with titles such as Super Mario Bros, which was presented at the Melbourne conference showcasing the most advanced stage of the current AI design, as well as other titles including Sonic the Hedgehog and Mega Man.
The accuracy of this system was increased when multiple videos were analysed, and it must be noted that although the AI will copy the game based on what is analyses, the gameplay may differ. The developed game may not include concepts or goals contained outside of the analysed footage of the source game, or where action occur away from the main screen. Visually however, there is little difference between the two game versions.
To develop the Super Mario Bros clone, the AI was subjected to videos showing two different playing styles: rushing through the levels as fast as possible, and exploring each level in detail to find power-ups and hidden areas. This allowed the AI to recreate features such as jumping (including double jumping), how to defeat enemies, and the danger of pits and other hazards.
Speaking with Checkpoint, the lead author of the paper, PhD student Matthew Guzdial, indicated that this was only the start. He mentioned that:
“One of the reasons for choosing to test out this research with platformers was because their physics systems can be understood as simplified versions of real world physics.”
“Short term, we’re looking into how the learned game engines can be used to automatically create novel games. Imagine a future system where you can show it two videos from your favourite games and it gives you a new game that’s a mixture of your two favourites. We’re also looking into how these models can be used to improve automated game playing.”
With advancements like these, it looks as though the game development process is set to become more refined. And this system does not seem to be limited to just copying 2D platform games, with groups looking to expand these techniques to more complex genres and domains, and to applications outside of the gaming world. Previous work from this group also included generating interactive fiction games from stories.
Perhaps one day it would be possible to turn any video into a game?