How Pokémon Go is building a world model in the age of AI

30 billion images power robot navigation

What would you do with 30 billion images and scans captured by Pokémon Go players as they roam cities around the world? Niantic, the publisher behind the hit mobile game, has been tapping a colossal corpus amassed since 2016. As they played, millions of users—often without realising—fed an immense cache of urban imagery paired with metadata such as orientation, speed and environmental conditions. According to MIT Technology Review, that trove now trains a Visual Positioning System (VPS) and a large geospatial model that can pinpoint a device to within a few centimetres and interpret the built environment from simple landmark views.

Niantic first retained player photos and data to make augmented reality work and to anchor virtual objects credibly and precisely in the real world. It needed to know not just where a user was, but which way they were facing and what their surroundings looked like. In the AI era, those signals have only grown in value, and Niantic now sees the data as good for much more than placing 3D Pokémon convincingly on a phone screen.

From chasing Pikachu to moving robots

This technical layer turns out to fix a familiar city problem: GPS unreliability in urban “canyons”, where signals drift amid towers and flyovers. By combining cameras with VPS, Niantic Spatial promises a far steadier lock than GPS alone, whose error can balloon to tens of metres. Such granularity is becoming a prerequisite for positioning‑sensitive tasks, from delivery to guiding mobile assistants.

The partnership with Coco Robotics shows what that looks like in practice. The company runs about 1,000 robots across several US and European cities and claims more than half a million automated deliveries. Blending GPS with visual localisation trained on Niantic’s data helps those robots stop at the right restaurant entrance, avoid obstructing pedestrians and line up at the correct threshold rather than a few metres from the door. “It turns out that getting Pikachu to realistically run around and getting Coco’s robot to safely and accurately move through the world is actually the same problem,” said John Hanke, chief executive of Niantic Spatial, as quoted in a BFMTV article.

World model and a living map

By developing this strand of “geospatial intelligence”, Niantic Spatial is laying groundwork for what it calls a living map: a hyper‑detailed world model, continuously refreshed by robot sensors and other sources. The ambition is to move beyond simple location to describe what is happening in a place and how it is changing. That spatial understanding could feed AI perception, planning and simulation across logistics, mobility and urban planning—and, in time, more contextual travel and city‑tour experiences.

Of course, Pokémon hunters may feel blindsided to learn their kilometres on foot have, unbeknown to many, created valuable assets—at least for those who did not read Niantic’s terms of service closely. The use of user‑generated content by contributors who are not always aware of how it will be exploited feeds into wider debates over ownership of data used to train AI models. As these “world models” grow in scope, contributor information and consent will be critical. Transparency on data use, workable opt‑out mechanisms and technical safeguards will likely need to go well beyond a few lines buried in sprawling T&Cs.