Google Unveils AI Model Called Genie That Can Generate 2D Video Games

Google has launched one other generative artificial intelligence (AI) mannequin that may create countless numbers of 2D platformer video video games. Genie is being touted as an action-controllable world mannequin that was skilled on unsupervised online game information. It makes use of predictive evaluation to generate online game ranges and may also management a playable character and decide its actions. Interestingly, OpenAI additionally introduced a world mannequin earlier this month known as Sora, which might generate hyperrealistic movies of as much as one minute in size.

The announcement was made by Tim Rocktäschel, Open-Endedness Team Lead, Google DeepMind, by way of a sequence of posts on X (previously often known as Twitter). He stated, “We introduce Genie, a foundation world model trained exclusively from Internet videos that can generate an endless variety of action-controllable 2D worlds given image prompts.” Genie is exclusive within the side that it might probably solely generate one particular factor, and additionally it is the one video game-generating mannequin that has been publicly introduced thus far.

Google’s Genie AI mannequin isn’t open to the general public but and solely exists as a analysis mannequin for now. This is why its user-centric functionalities are usually not recognized but. It can generate online game ranges utilizing photographs, however whether or not it might probably take textual content prompts and even video prompts isn’t recognized. A preprint model of the paper was posted on-line which highlights its technical points. The AI mannequin was skilled unsupervised on 2,00,000 hours of online game footage and comprises 11 billion parameters. The structure of the mannequin makes use of three completely different components — a spatiotemporal video tokenizer, an autoregressive dynamics mannequin, and a easy and scalable latent motion mannequin.

How Google Genie Works

To simplify, the spatiotemporal video tokenizer takes online game footage, breaks it down into smaller chunks of datasets, often known as tokens, that may be consumed by the muse mannequin. Spatiotemporal explains that the info is damaged down each in time and house (For instance, a video was damaged down into 2-second clips, however every body was additionally damaged down into a number of items).

The autoregressive dynamic mannequin comes subsequent. Autoregressive fashions primarily predict the long run primarily based on how one thing has carried out up to now, and a dynamic mannequin is accountable for understanding how issues change and transfer over time. So this half is the place the predictive evaluation begins. The last element is the latent motion mannequin. This is the place the AI understands how the playable character strikes and traverses within the online game world.

“Genie’s learned latent action space is not just diverse and consistent, but also interpretable. After a few turns, humans generally figure out a mapping to semantically meaningful actions (like going left, right, jumping etc.),” stated Rocktäschel. This half is essential as a result of it highlights that the primary downside this AI mannequin solves is not only producing 2D online game ranges, but additionally understanding how primary actions happen, and the way that info can be utilized to navigate real-world terrains.

Highlighting this, he added, “Genie’s model is general and not constrained to 2D. We also train a Genie on robotics data (RT-1) without actions, and demonstrate that we can learn an action controllable simulator there too. We think this is a promising step towards general world models for AGI.”