Niantic uses Pokemon Go data to build AI navigation system (19459000)
Niantic announced last week that it would create an AI model to navigate the physical world based on scans collected by players of its mobile game, such as Pokemon Go404 Media reports that it has received feedback from its Scaniverse users.
All AI models need training data. Companies have collected data through websites, YouTube videos and audio sources. This is the first time we’ve heard that AI training data was collected via a mobile game app.
“Over the past five years, Niantic has focused on building our Visual Positioning System (VPS), which uses a single image from a phone to determine its position and orientation using a 3D map built from people scanning interesting locations in our games and Scaniverse,” Niantic posted a blog entry on the company’s website.
The company’s creation is called a “large geospatial model” (LGM), drawing comparisons to large-language models (LLMs), like the ones that power ChatGPT. Niantic’s new model will process spaces by using geolocated images from its apps. Language models process text.
The size of Niantic’s data collection shows the company’s presence in the AR industry. The model is based on over 10 million scans of locations around the world, with users taking approximately 1 million new scans each week through Pokemon Go Scaniverse. These scans are from a pedestrian’s perspective, and capture areas that are not accessible to cars or street-view cameras.
First-person scans
The company claims to have trained more than fifty million neural networks, with each one representing a particular location or viewing angle. These networks compress thousands mapping images into digital representations. Together, they contain over 150 trillion parameters–adjustable values that help the networks recognize and understand locations. Niantic intends to combine the knowledge of multiple networks to create a comprehensive model that will be able to understand any location from unfamiliar angles.
