Google researchers announced a new artificial intelligence (AI) game engine last week. Dubbed GameNGen, it is entirely powered by a neural model and is capable of real-time generation over a long trajectory. The researchers claim that the game engine can generate complex environments at a high number of frames. Notably, the company claims that the game engine was able to interactively simulate the classic game Doom at more than 20 frames per second. The GameNGen can run the game with a single Tensor Processor Unit (TPU).
Google Unveils GameNGen
The tech giant published a paper on the neural model-powered game engine in the online pre-print journal arXiv. It also detailed the model in a GitHub listing. Creation of a game engine is a significantly complex task, as the system requires to not only generate complex 2D and 3D environments at a high speed consistently, but it also needs to do so with logical sequencing keeping level progression in mind.
Highlighting the achievement of the GameNGen, the paper highlighted that the game engine was able to interactively simulate the 1993 video game Doom at more than 20 frames per second. Interactive simulation means these generations were not static videos or images, but players can interact with these generated elements as well.
The paper claims that two processes were followed to train the AI-powered game engine. First was training using Stable Diffusion v1.4. The researchers also used a novel method to mitigate auto-regression (when the AI model generates the next sequence based on the information of the past sequence) drift, where it added Gaussian noise to encode the frames.
The second part involved the usage of automatic reinforcement learning (RL) agents. The paper stated that the collection of data at scale using human players would not have been possible. As a result, the researchers used automated AI-powered agents that played the game, allowing the collection of a large sample of data.
At present, the AI game engine is not available for people to download or test out. The model is still kept under wraps and only the research paper is available. Notably, publishing a paper on arXiv does not require peer review, so a full evaluation of the claims and methodology is yet to be done.
