[Au Jus Demon]

Pokémon Go Players Accidentally Built One of the Largest AI Datasets Ever

You may have helped train robots without realizing it.

Over the past eight years, Pokémon Go players have captured tens of billions of images of the real world while playing the game.

Those images are now being used to help autonomous robots move through real city streets.

What started as a mobile AR game slowly became one of the largest crowdsourced mapping datasets ever created.

30 billion images from over 1 million mapped locations.

What Just Happened

On March 10, 2026, Niantic Spatial announced a partnership with Coco Robotics.

Niantic Spatial is an AI company created from the technology behind Pokémon Go.

Coco Robotics operates autonomous delivery robots in several cities including:

• Los Angeles
• Chicago
• Miami
• Jersey City
• Helsinki

These robots will use Niantic's Visual Positioning System (VPS).

Instead of relying only on GPS, the robots determine their position using computer vision.

They capture images of nearby buildings and landmarks and compare those images against Niantic’s visual database.

This allows the robot to calculate its position much more precisely.

GPS in dense cities can drift 20 to 50 meters.

Visual positioning systems can locate something within a few centimeters.

For a robot navigating sidewalks and building entrances, that difference is extremely important.

Where The Dataset Came From

Pokémon Go launched in 2016 and quickly became one of the most popular mobile games ever released.

Within about 60 days it reached more than 500 million installs.

Millions of people were suddenly walking around cities pointing their phones at parks, buildings, statues, and landmarks.

But the real data collection began later.

In 2020 Niantic introduced AR Mapping tasks.

Players were asked to scan real world objects by walking around them with their camera active.

Inside the game this appeared as a simple research task that rewarded players with items.

Each scan uploaded spatial information such as:

• camera images
• GPS coordinates
• device orientation
• direction of movement
• sensor data

Individually these scans looked like normal gameplay actions.

But across millions of players they created a massive visual map of the real world.

Niantic reportedly now has around 30 billion images covering more than one million locations.

Why This Data Is Valuable

Most mapping companies rely on dedicated camera vehicles.

Google Street View is a well known example. Camera cars drive through cities capturing images every so often.

Those images are consistent but limited.

Pokémon Go players produced something very different.

Images were captured under many different conditions such as:

• different lighting conditions
• different weather
• different times of day
• different viewing angles
• crowded environments

A single Pokéstop might be scanned hundreds or even thousands of times under completely different conditions.

This variety is extremely useful for training computer vision systems used in robotics and navigation.

No traditional mapping company could realistically recreate this diversity at the same scale.

The Bigger Goal

Niantic is not just building maps.

Their long term goal is something called a Large Geospatial Model.

Large language models learn patterns in text.

Large geospatial models learn patterns in the physical world instead of language.

Instead of understanding language, the system learns to understand physical spaces and environments.

Technology like this could power systems such as:

• delivery robots
• AR glasses
• autonomous vehicles
• robotics navigation systems
• digital twins of cities

Every robot or device using the system can also feed new images back into the dataset.

Over time the map becomes more accurate and more detailed.

Developer Uses (Unity / AR)

This technology is also available to developers.

Niantic's Lightship VPS can be used inside Unity to build location based AR experiences.

Developers can anchor digital objects to real world locations with much higher precision than GPS alone.

Examples include:

• AR games tied to real world landmarks
• interactive public installations
• location based scavenger hunts
• persistent AR objects that stay in the same place
• multiplayer AR experiences in shared locations

Instead of relying only on GPS, the system uses visual localization to anchor content to the environment.

This keeps AR content stable even in dense urban areas.

Something Interesting To Think About

Pokémon Go launched in 2016.

Years later, AR scanning tasks helped build this dataset.

Every app collects data.

But how many of them are accidentally building something as massive as a global spatial map?

[MaîtreCornichon]

It's fascinating to see how a game can contribute to such advanced AI technology in unexpected ways.

[Wake]

I always had ideas of what the data they collect would be used for, delivery robots or self driving cars was always my guess. Very informative thread nicely done.