Pokemon Go Built a 30 Billion Image AI Dataset

Players thought they were catching Pokemon. They were building one of the largest real-world visual datasets in AI history.

Niantic has disclosed that photos and AR scans collected through Pokemon Go produced a dataset of over 30 billion real-world images. The company's spin-off, Niantic Spatial, is now using that data to power a Visual Positioning System (VPS) that guides delivery robots with centimeter-level accuracy - far beyond what GPS can achieve.

The first commercial deployment: a partnership with Coco Robotics, announced March 10, to equip sidewalk delivery robots with visual navigation trained on eight years of Pokemon Go player scans.

Metric	Number
Total images collected	30+ billion
Scanned locations	10 million (1 million activated for VPS)
Fresh scans per week	~1 million
Peak monthly active players (2016)	~230 million
Data collection period	2016-2026 (8 years)
Positioning accuracy	Centimeter-level

"Whether players knew it or not, those scans were creating 3D models of the real world that would eventually power the Niantic model."
Popular Science

How They Built It

Starting in fall 2020, Niantic introduced "Field Research" tasks that asked players to walk around real-world landmarks - Pokestops, gyms, battle arenas - while their smartphone cameras captured images. Players received in-game rewards for scanning locations. Niantic used photogrammetry to convert these scans into detailed 3D models.

The genius of the approach: each location was scanned by thousands of different players, from different angles, at different times of day, in different weather and lighting conditions. A single Pokestop might have morning scans in sunshine, evening scans in rain, winter scans with snow cover, and weekend scans with crowds. No mapping company launching camera cars could have replicated this diversity on the same timeline or budget.

The images cluster around hot spots - locations that served as important in-game landmarks. For each of those million activated locations, Niantic Spatial has thousands of images taken in roughly the same place but with the natural variation that comes from different humans approaching from different directions at different times.

What It Powers Now

The Visual Positioning System

Niantic Spatial's VPS takes a handful of camera snapshots and matches them against its database to determine your exact position - down to a few centimeters. This works in environments where GPS fails: urban canyons between tall buildings, indoor spaces, and narrow sidewalks where a 5-meter GPS error is the difference between the sidewalk and the middle of the road.

Coco Robotics Partnership

Coco launches last-mile delivery robots for food and groceries in US and European cities. The integration allows robots to:

Position themselves at precise pickup spots outside restaurants
Navigate without blocking pedestrian traffic
Stop at the customer's exact door rather than "somewhere nearby"

Images captured by the robots' onboard cameras feed back into the VPS, creating a continuous improvement loop. The robots are both consumers and contributors to the dataset.

The Large Geospatial Model

Beyond VPS, Niantic is building a Large Geospatial Model (LGM) - a spatial AI framework trained on scans, 3D splats, LiDAR data, drone imagery, and VPS anchors. The vision: a model that understands physical spaces the way LLMs understand language, enabling AR glasses, autonomous robots, and spatial computing applications.

Section 5.2 of Niantic's Terms of Service grants the company broad rights over AR content uploads. The terms state Niantic can use submitted data however it wishes and "pass that freedom on to other entities." Players agreed to this upon installation.

But agreeing to a Terms of Service and understanding what you are consenting to are different things. Hackaday's coverage noted that "it's unlikely that many players will lose any sleep over the fact that they have unwittingly been collecting training data" - but also flagged the risk of "a scenario in which that data ends up getting licensed out for some purpose they aren't comfortable with."

Key privacy details:

Uploaded AR imagery is anonymized during processing
Players can opt out of uploading additional data from now on
Players can't remove data already pushed into the system
The VPS model is a derivative work - individual images aren't retrievable

The comparison to Google's reCAPTCHA is apt. For years, users solved CAPTCHAs thinking they were proving they were human. They were actually labeling training data for Google's self-driving car project. Pokemon Go's scanning feature followed the same pattern: a useful-seeming user interaction that doubled as a data collection pipeline.

Should You Care?

The 30 billion image number is striking, but the real story is the data quality. Staged photography and camera-car fleets produce uniform, controlled images. Pokemon Go players produced messy, diverse, real-world data - exactly what robotics AI needs. A delivery robot has to navigate in rain, at night, around construction, past parked cars. Pokemon Go players already captured all of those conditions, millions of times, across thousands of cities.

Niantic built the most complete street-level visual dataset in the world, and it didn't deploy a single camera car. It gave people a game and let them do the work. The question from now on isn't whether the data is valuable - Coco's partnership proves that. The question is whether "you agreed to the Terms of Service" is sufficient consent when the commercial application wasn't foreseeable to the people generating the data.

Thirty billion images. Eight years. Millions of players scanning every landmark, storefront, and sidewalk they walked past while chasing virtual creatures. Niantic turned a mobile game into the largest crowdsourced mapping operation in history, and now that map guides delivery robots. The players built the dataset. Niantic owns it. The robots use it. The only party that didn't benefit from the arrangement is the one that did all the work.

Sources: