Mar 17, 2026
3 min read
Over the years, millions of users have taken pictures of streets, statues and monuments to catch Pokemon.That data now guides delivery robots through the streets of Los Angeles, Chicago and Helsinki with an accuracy of a few centimeters.
For years, millions of users around the world have spent millions of miles tracking Pikachu and Bulbasaur.What they didn't know is that they were creating one of the densest and most diverse collections of visual data ever collected, a database worth billions in the world of physical AI and robotics.
It is managed by Niantic Spatial, a company founded in May 2025 as a spin-off from Niantic after the sale of the gaming branch that includes Pokémon Go, Pikmin Bloom and Monster Hunter Now to Scopely for 3.5 billion dollars.Scopely may not realize that the team that built the spatial positioning technology behind the game and the entire database is worth far more than the Pokémon rights and games themselves.Because what Pokémon Go built is something that cannot be replicated today, even with the most advanced techniques.
A problem that GPS cannot solve
In fact, GPS works by measuring the travel time of radio signals from a constellation of orbiting satellites, and when the accuracy drops to a few meters in an open area in a dense urban area, the situation becomes more difficult.In fact, the signals bounce off the glass and concrete facades of the buildings, creating multiple displays that the receiver interprets as direct paths, this is called the "urban canyon."The result is that using the GPS signal to measure the location of the robot must have an error of tens of meters, which means that the robot will be sent to the wrong side of the road or even to the wrong place.The robots we're talking about, at least in this first phase, are delivery robots, and a delivery robot that goes about five kilometers an hour and carries very large pizzas or shopping bags, this level of error is unacceptable.If it wants to compete with human messengers, it must meet accurate and specific delivery times to pick-up and drop-off locations, then it will not make a mistake.
The most common solutions used so far to correct the low accuracy when "inside the cities" involve the use of additional sensors or additional data, such as tire distance measurements, LiDAR, or even the creation of local 3D maps through SLAM (Synonymous Localization and Mapping). Some companies that build these robots use robot sensors to create local maps in real time to monitor the location.edges, columns, and building profiles and then create subsequent paths using this information.However, there are structural limitations to this approach: Each robot creates and maintains its own map, private, and not shared.
Niantic Spatial has built over the years to "pick Pokemon" visual positioning systems (VPS) that provide centimeter-precision with six degrees of freedom, working even in areas where GPS cannot reach.The problem is how to get people to use the system so that they can create a detailed and detailed global model to cover millions of streets in dozens of cities.Pokemon Go. Over the years, Pokemon Go games have taken users to specific locations, transforming them as gyms, battlefields, and parking spaces, creating a global dense database of urban images, each of which corresponds to comprehensive metadata: latitude, longitude, camera direction, device location, and motion data.And this greatly expands the collection of high-quality three-dimensional scenes, often taken from several angles and under different lighting conditions.
According to Niantic Spatial, this resulted in more than 30 billion photos taken for free, centered around more than a million "hotspots" photographed from multiple angles at different times of the day and in variable weather conditions.All done by people who thought they were playing.
Today the project is called VPS, Visual Positioning System, but what the company actually builds with the collected data is an LGM, Large Geospatial Model.In short, they create an AI model that does for physical space what large language models do for text, and a series of local networks already trained to create a model capable of understanding the space and morphology of places that have never been visited before.They do this by combining.Today, we can ask AI to create an image that doesn't exist, and a geospatial model can understand the location and proportions of a given place.Niantic explains that it has already trained more than 50 million neural networks with more than 150 trillion parameters covering more than a million locations, and now they need to merge them.
What Niantic wants to do is ambitious: for example, if Pokémon Go players can take a picture of the entrance of a church in Milan and the background of the same church in Lyon, for example, it can understand, or imagine what the surface of a church in Krakow would look like based on the structure of the architecture it learned from all the samples it learned.
In Niantic Spatial's vision, this model will become a shared infrastructure between machines, AI agents, and human users, a semantically fused 3D model of the world, updated in real time as new sensors observe it.A digital twin of the Earth.
Coco Robotics will be the first testbed
The first company to try a VPS system rich with game data from Pokemon Go players is Coco Robotics, which makes deliveries using robots and operates about a thousand small carts on the sidewalks of American and European cities.Coco, surely someone has seen the videos, has recorded hundreds of thousands of deliveries covering more than one and a half million kilometers between Los Angeles, Chicago, Miami, Jersey City and Helsinki.Each robot stands four hips tall.It has a camera, and the addition of Niantic's system fuses the GPS signal with visual localization, which achieves the necessary accuracy not only for street navigation, but also for identifying the exact reception point.If it used to be a bit "random", today it can detect if there is a slide to reach the pavement and also calculate the best route for the last section.
There was just one small problem: Pokémon Go users didn't know they were building a training database for delivery robots and were collecting free miles for a company that would use the data they collected to create a basic model for future robots and automation.We imagine a delivery driver who has been playing Pokemon Go for years and whose contribution to the work has somehow cost him his job.
Niantic is specific on one point: the scanning function is completely optional, users must visit a specific publicly accessible location and deliberately click to start scanning, walking around with the app open does not train the AI model.It is true, however, that there are different nuances: it would not be the first time that data that users freely collect with a purpose ends up as fodder for something completely different.Recall, for example, that the police recently bought or obtained data from the navigation app Waze to investigate whether Google's captcha, which asks users to identify traffic lights and bicycles, are not bots, but tools for free cataloging of data sets that will later be used to train artificial intelligence.
