Final week, Niantic introduced plans to create an AI mannequin for navigating the bodily world utilizing scans collected from gamers of its cellular video games, similar to Pokémon Go, and from customers of its Scaniverse app, studies 404 Media.
All AI fashions require coaching information. To date, corporations have collected information from web sites, YouTube movies, books, audio sources, and extra, however that is maybe the primary we have heard of AI coaching information collected by way of a cellular gaming app.
“Over the previous 5 years, Niantic has centered on constructing our Visible Positioning System (VPS), which makes use of a single picture from a cellphone to find out its place and orientation utilizing a 3D map constructed from folks scanning fascinating areas in our video games and Scaniverse,” Niantic wrote in an organization weblog publish.
The corporate calls its creation a “massive geospatial mannequin” (LGM), drawing parallels to massive language fashions (LLMs) like the sort that energy ChatGPT. Whereas language fashions course of textual content, Niantic’s mannequin will course of bodily areas utilizing geolocated pictures collected by way of its apps.
The dimensions of Niantic’s information assortment reveals the corporate’s sizable presence within the AR area. The mannequin attracts from over 10 million scanned areas worldwide, with customers capturing roughly 1 million new scans weekly by way of Pokémon Go and Scaniverse. These scans come from a pedestrian perspective, capturing areas inaccessible to automobiles and street-view cameras.
First-person scans
The corporate studies it has educated greater than 50 million neural networks, every representing a particular location or viewing angle. These networks compress 1000’s of mapping pictures into digital representations of bodily areas. Collectively, they comprise over 150 trillion parameters—adjustable values that assist the networks acknowledge and perceive areas. A number of networks can contribute to mapping a single location, and Niantic plans to mix its information into one complete mannequin that may perceive any location, even from unfamiliar angles.
