Pokémon Go players helped train robots to navigate cities

Nearly a decade after its release, Pokémon Go is influencing artificial intelligence in an unexpected way.

According to a report by MIT Technology Review(nuova finestra), Niantic Spatial, an AI company spun out of the game’s original developer, is using more than 30 billion images(nuova finestra) of urban landmarks captured by Pokémon Go players to train AI systems that help robots understand where they are in the real world.

When Pokémon Go launched in 2016, millions of players walked through their cities pointing phones at buildings, parks, and landmarks while searching for digital creatures. In doing so, they created one of the largest datasets of real-world imagery ever collected. Niantic Spatial trained its model on 30 billion images captured across urban environments, many clustered around locations that players frequently visited in the game. Sometimes, digital creatures appeared in or near private spaces such as apartment buildings or residential courtyards, which means those moments may have become part of AI datasets.

That data is now being used to build what researchers call a “world model” — an AI system designed to help machines interpret and navigate physical environments.

It also highlights something easy to overlook about modern technology: everyday app activity can quietly become valuable training data for AI systems years later — something that Pokémon Go players couldn’t have possibly anticipated in 2016, let alone meaningfully consented to.

Why GPS struggles in cities

GPS signals often bounce off tall buildings, drift significantly, or disappear entirely in dense urban environments, a problem sometimes called the “urban canyon” effect. As MIT Technology Review explains in its report on the technology, even smartphone location indicators can drift dozens of meters in cities, often placing a device on the wrong block or the wrong side of the street.

Niantic Spatial’s solution relies on a visual positioning system(nuova finestra), which determines location by analyzing what a camera sees. By comparing snapshots of nearby buildings and landmarks with its massive dataset, the system can reportedly pinpoint a location within a few centimeters.

From augmented reality to delivery robots

One of the first real-world tests of Niantic Spatial’s technology is happening through a partnership with Coco Robotics, a startup operating sidewalk delivery robots across several cities.

The company’s robots carry groceries and restaurant orders in places like Los Angeles, Chicago, Miami, and Helsinki. According to the MIT Technology Review report, Coco’s robots have already completed more than half a million deliveries, covering millions of miles.

Navigating dense cities reliably remains a major challenge for autonomous machines. By combining cameras on the robots with Niantic Spatial’s visual positioning system, the machines can better determine exactly where they are, allowing them to stop precisely at pickup locations or outside a customer’s door.

The rise of “world models”

The project reflects a broader trend in AI development.

Large language models are trained on images and text from the internet. World models, by contrast, aim to help machines understand how the physical world itself is structured — where objects are located, how spaces connect, and how to move through them safely.

Niantic Spatial says its long-term goal is to build a constantly updated “living map” of the world that robots and other AI systems can use to navigate.

What this means for privacy

This story also illustrates a broader shift in how data created by everyday people using everyday apps is being reused.

Millions of people downloaded Pokémon Go to play a game. But along the way, they also generated billions of images and precise location signals tied to real-world places. This data is now helping train AI systems designed to map and navigate the physical world. That’s a long way off from the player’s intention to pick up the game in search of Pikachu. 

This is becoming a familiar pattern in the AI economy. Activities that seem unrelated to artificial intelligence — taking photos, browsing the web, using apps — often end up producing the datasets used to train powerful new systems. Modern AI models are frequently built using large collections of data scraped from the internet and other digital sources, which researchers say can sometimes include personal information or sensitive data.

As companies race to build world models and other forms of AI, the question becomes just where the training data comes from — and whether the people who generated it ever realized how that data would be used, or would still agree to use those apps if they had known the consequences. Policymakers and researchers have increasingly called for clearer consent standards(nuova finestra) for data used in AI training, arguing that creators and users should have more control over how their data is reused.

These AI privacy concerns are already playing out in court. Publishers, authors, and media companies have filed lawsuits claiming their work was used to train AI systems without permission, including a high-profile case in which The New York Times sued OpenAI and Microsoft(nuova finestra) over the alleged use of its journalism in AI training datasets.

For companies building AI that navigates the physical world, datasets created through games, apps, and digital platforms are quickly becoming some of the most valuable assets in the technology industry. As that happens, questions about transparency, consent, and how user-generated data is reused are becoming harder to ignore.

Not everyone is pursuing this data-hungry model. At Proton, we believe in private, transparent AI that gives you the benefits of an AI assistant(nuova finestra) without the privacy costs. Lumo never logs your data, trains on your sensitive conversations, or shares your information with anyone. It means using AI without wondering whether today’s interactions might quietly become tomorrow’s datasets.