Gran Turismo 4 has been used to train a new AI research project called Multiverse. It’s the first project from a new AI development company called Enigma Labs, comprising some young programmers (and Gran Turismo fans), to demonstrate a new technique for creating AI-generated worlds for multiplayer games.
Unlike traditional video games, which use a strictly-defined game engine, graphic models, and textures to generate the images you see on screen, a new breed of projects like Multiverse are pushing the limits of AI image generation to create playable video games in entirely different ways.
Multiverse was built by training its AI engine on video frames of Gran Turismo 4 gameplay around the Tsukuba circuit. Tsukuba’s relatively short length and uncomplicated layout made it an ideal environment for the AI to learn. These characteristics simplified the task of teaching the AI to understand not only the track’s geometry and appearance but also the complex interactions between two racing vehicles.
To gather the necessary gameplay footage, Enigma Labs employed a clever technique leveraging Gran Turismo 4’s own B-Spec mode. For those unfamiliar, B-Spec allows players to take on a directorial role, guiding an AI driver. By hacking the game to create 1-v-1 B-Spec races, the team could autonomously generate a dataset of 1-v-1 race footage without requiring countless hours of manual human input, significantly scaling up their data collection efforts.
What truly sets Multiverse apart — and positions it as a notable research project — is its successful integration of a multiplayer experience.
While AI-generated single-player worlds such as Oasis have seen some experimental success, creating a coherent and synchronized shared environment for two players — where each player’s actions meaningfully and consistently affect a shared reality viewed from independent perspectives — has been a significant hurdle.
Multiverse tackles this head-on, making it a novel exploration into AI-driven multiplayer game spaces. Enigma Labs states this is the first AI multiplayer world model of its kind.
To learn more about how Multiverse was built, Enigma Labs has published an extensive blog post documenting the team’s journey and how the simulation works.
For those with the technical inclination and capable hardware, Enigma Labs has made Multiverse an open-source project. The code, data, and even the pre-trained model are publicly available on GitHub and Hugging Face, allowing interested users to download and potentially run the simulation themselves. Remarkably, the team reports the initial development cost for this proof-of-concept was under $1,500 USD.
Projects like Multiverse, while still in the research phase, offer a fascinating glimpse into what could be the future of game development. The idea of game worlds being generated on-the-fly by AI, rather than meticulously pre-constructed, opens up a new paradigm for creating dynamic and responsive interactive experiences. It’s a revolutionary concept that could reshape how games are made and played in the years to come.
See more articles on Artificial Intelligence.
