143 million players exploring streets in Pokémon GO, but helping Niantic build the largest real-world visual dataset for AI

Pokémon Go players have been scanning landmarks for years to “improve the gaming experience.” Now, this data is driving autonomous delivery robots through the streets of Los Angeles. When gamified incentive designs make unpaid data collection addictive, how much value does the word “voluntary” really carry?
(Background: Pokémon GO developer secures $300 million in funding! To develop the metaverse, valuation hits $9 billion)
(Additional context: All eyes on Meta! Using user FB and IG posts to “train AI”—how to oppose Facebook’s data collection?)

Table of Contents

Toggle

• From “Improving Game Maps” to Robot Navigation Engines
• Niantic’s Transformation: From Gaming Company to Spatial Data Firm
• Where Are the Boundaries of “Voluntary”?
• Data Metaphors in the Gaming Industry

While millions of Pokémon Go players step into the real world to catch digital Pokémon, they unknowingly contribute every scan image to the world’s largest real-world visual AI dataset. According to MIT Technology Review, this data is now used by Niantic Spatial to guide autonomous delivery robots through city streets with precise navigation.

From “Improving Game Maps” to Robot Navigation Engines

Since its launch in 2016, Pokémon Go has had an unobtrusive feature: players can choose to anonymously scan nearby public landmarks (statues, buildings, monuments). Officially, this is to “help improve the game’s spatial map.” Over the years, this system has trained a visual positioning system (VPS) with over 30 billion images.

In February this year, Niantic’s spatial tech subsidiary, Niantic Spatial, announced a partnership with Coco Robotics, based in Santa Monica, Los Angeles. Coco currently operates about 1,000 autonomous delivery robots in Los Angeles, Chicago, Miami, and Helsinki. The robots are roughly the size of a suitcase and can carry up to 8 large pizzas or 4 grocery bags at once.

VPS addresses a core challenge for city robots: GPS signals weaken or fail entirely in dense urban environments due to reflection and obstruction. VPS improves positioning accuracy to centimeter-level by matching real-time camera images with pre-built visual maps—allowing robots to stop precisely at restaurant entrances instead of crashing into parking barriers.

Niantic’s Transformation: From Gaming Company to Spatial Data Firm

In May 2025, Niantic Spatial officially spun off from Niantic Inc., becoming an independent company focused on spatial AI technology. This move was hardly surprising—by 2021, Niantic had already raised $300 million with a vision to build a real-world metaverse, reaching a valuation of $9 billion at the time.

For Niantic, Pokémon Go has always been more than just a game; it’s a global network of ground sensors. Niantic Spatial’s spokesperson told Decrypt:

“Our initial VPS was built using data voluntarily scanned by players in the game—but no single source defined this model. Our approach’s uniqueness lies in combining scale with ground-level detail, and increasingly, user-generated data is key to driving accuracy in the most critical environments.”

Where Are the Boundaries of “Voluntary”?

Criticism soon followed. On X (formerly Twitter), a user pointed out: “1.43 billion people think they’re catching Pokémon, but they’re actually building one of the largest real-world visual datasets in AI history.”

Another insightful critique came from a different user: “The killer feature isn’t the map, but the incentive design. Pokémon Go turns millions of players into unpaid edge case hunters, making data collection feel like a game.”

This is the core paradox of gamification—when something is packaged as a game, people no longer judge it by work standards. Players scan landmarks not for Niantic’s commercial benefit but to “enhance the gaming experience.” Yet, after a decade, the biggest beneficiaries of this data are autonomous delivery robots running in cities.

Niantic’s official response emphasizes voluntariness: “Players can choose to submit anonymous scans of public places to help improve VPS. These scans have always been and remain completely voluntary, and scan data is not linked to player accounts.”

However, the premise of “voluntary” depends on full awareness of data use. How many players who submitted scans foresaw that their contributions might one day be used to train robot navigation models?

Data Metaphors in the Gaming Industry

The Pokémon Go case offers a textbook example of how game mechanics can systematically convert user behavior into commercial data assets. This isn’t unique to Niantic—many games quietly collect behavioral data from every click, route, and decision.

The difference is that Pokémon Go’s data collection extends from the virtual game world into the physical space of the real world—players contribute not only gameplay actions but also perceptions and scans of the physical environment. This scale far surpasses any purely digital game dataset.

300 billion images—this number is the result of 143 million players’ footsteps.