Nvidia’s AI creates game demo entirely on its own

Tom Bedford December 3, 2018

Nvidia is no stranger to machine learning, which it has now put to good use in a new game demo. The demo, an otherwise insignificant driving simulator, has one distinguishing feature — everything has been built by an AI.

Nvidia’s AI creates game demo entirely on its own

Visual data of a car’s surroundings from Nvidia’s experiments with self-driving cars, was broken down into separate categories depending on structure type (like trees, buildings, or other cars), to help the AI create different assets.

The environment of the demo was created in Unreal Engine 4, a popular graphics platform used in games like Spyro Reignited Trilogy, We Happy Few and SoulCalibur VI. A single engineer recreated the streets from Nvidia’s vehicle trials by placing objects, and the AI was designed to render each object’s graphics as the game is played.

The demo was plagued by several issues including object permanence, in which the AI forgot the particular graphics of an individual object when it was re-rendered as a driver returned to it. Another issue with the demo is its admittedly blurry graphics, which would look more at home in an early PS3 game than a modern, RTX-powered game. However these visual quirks a common symptom of AI-generated art,an AI-generated piece of art that sold at an auction in October also had this issue.

Interestingly, the entire demo was powered by a single computer — Nvidia’s Titan V. It’s a GPU created with AI researchers and data scientists in mind, which unfortunately means most amateur game designers can’t use it to easily generate their game’s graphics.

Nvidia’s AI-generated graphics game demo combines its two biggest areas of research: video games and vehicles. It hopes that in the future AI can mitigate some of the grunt work involved in video game development, although it admits it might be several decades before AI can generate blockbuster-style graphics.

Nvidia’s AI also has a lot in common with a current feature of games design known as procedural generation, a popular function for games with large worlds, in which an engine is taught particular rules and an algorithm extrapolates from these rules to automatically create a game world.

While AI-generated games may still be a long way off (especially given AI’s rather nonsensical writing style) it presents game designers, as well as AI researchers and development, a potential future tool for using AI to create content.