Ever since its groundbreaking release in 2016, Niantic’s Pokémon GO has made history in the world of AR games, putting Niantic on the map. At its highest player count, Pokémon GO had approximately 232 million active users. As of November 2024, the estimated player count is 5.9 million, which is still considerable given the game’s expected decline. Regardless, Pokémon GO’s ability to retain millions of users over the years has provided Niantic with a substantial wealth of player data for various purposes, along with their additional comparably successful AR-based mobile games.
Niantic announced their current attempts at building a Large Geospatial Model (LGM) in a blog post earlier this November, explaining that their user data would play an important role in this project. The blog post begins by highlighting the difference between human perception and machine perception. Human beings have spatial understanding, allowing us to fill in the details of our surrounding environment based on the countless spaces we’ve seen before. Machines, on the other hand, find this task exceedingly difficult. Even the most advanced forms of AI struggle with spatial awareness, which has led Niantic to work on their ideal Large Geospatial Model.
LGMs are based largely on the concept of Large Language Models (LLMs), which are trained on internet-scale collections of text and can interpret language in ways that go beyond human understanding. LGMs would be able to navigate the physical world in a comparably complex way. Geospatial models are generally built on large amounts of raw user data, which can later be used to develop large models capable of more complex location-based interactions.
The post explains that Niantic has focused on building their Visual Positioning System (VPS) over the past five years. Their VPS is able to determine position and orientation of a single image using a 3D map built from the company’s “Scaniverse.” This advanced system allows players to see digital content placed into their physical environment in a more precise way. Niantic’s VPS is built from billions of user scans from locations around the globe, which has created “a highly detailed understanding of the world.” These scans are especially helpful to their VPS, because they are “taken from a pedestrian perspective and include places inaccessible to cars.”
As of November 12, 2024, Niantic has scanned over 10 million locations around the world through games like Pokémon GO, Pikmin Bloom, Ingress, and other Niantic properties. 1 million of those scans are currently active for use via Niantic’s VPS service. They have trained “more than 50 million neural nets to date, where multiple networks can contribute to a single location.” When combined, these networks cover more than 150 trillion parameters optimized using machine learning.
Since the blog post’s release earlier this November, there has been a misunderstanding among the press Many news sources expressed the idea that Niantic was using all of their user data to train AI map models, entirely regardless of user consent. This has since been clarified by Niantic in an editor’s note at the head of the original blog post, which explains that simply walking around and playing their game does not train any AI models.
We use player-contributed scans of public real-world locations to help build our Large Geospatial Model. This scanning feature is completely optional – people have to visit a specific publicly-accessible location and click to scan. This allows Niantic to deliver new types of AR experiences for people to enjoy. Merely walking around playing our games does not train an AI model.
For anyone interested in trying out Niantic’s AR-based mobile games, all of them are free-to-play and currently available on both iOS and Android devices. Readers specifically interested in Niantic’s advancements in AI map models can monitor the company’s progress via blog posts on the official Niantic website.