When you buy through affiliate links in our content, we may earn a commission at no extra cost to you. Learn 
Google’s GameNGen AI engine simulates a classic shooter title, Doom, achieving gameplay nearly indistinguishable from the original.
Google researchers have revealed an AI version of Doom based on gameplay clips.
A paper was published detailing how a small team could “interactively simulate” the game, with only a “slightly better than random chance” of humans being able to differentiate between the AI-powered version and the original content.
The scientists responsible, Dani Valevski, Yaniv Leviathan, Moab Arar, and Shlomi Fruchter, created GameNGen using Stability AI’s Stable Diffusion image generator.
The game engine is based on a neural network, allowing gameplay like the actual title: navigating, firing weapons, and evading damage from the enemy. The AI construct works to build a level around you in real-time as you survey the environment.
The simulation’s advances are limited, as reflected in the generator’s three-second memory, which can’t track what you did just a moment ago. It isn’t a finished concept, but GameNGen will contribute to the evolution of AI engines and games.
How GameNGen creates the AI Gameplay
Vast data are required to generate realistic gameplay, but the researchers were aware of the difficulty of obtaining the relevant data from human gamers to train the model. To alleviate this, a crucial first step was for an AI bot to be clued up to play Doom across all levels.
This allowed the generator to pick up on the range of player skill levels and capture a broad picture of Doom’s overall gameplay.
A detailed process included attention to flaws in image generation, meaning the researchers used previous frames to train new ones for improvement. The reworked frames were corrupted with Gaussian noise inputs on a separate network, which means an effective self-correcting loop was established to create longer periods of gameplay.
