Pokémon GO Data Powers AI & Maps – Better Than Street View?

A decade after its release, the augmented reality game Pokémon Proceed is having a second life – not as a cultural phenomenon, but as a training ground for artificial intelligence. Niantic Spatial, a spin-off from the game’s creator Niantic, is leveraging the 30 billion images captured by players to build a highly detailed world model for robots, starting with Coco Robotics’ delivery fleet.

The premise of Pokémon Go, launched in 2016, was simple: players used their smartphones to locate and capture virtual Pokémon characters that appeared in the real world. This required millions of users to photograph landmarks, street corners, and buildings, effectively crowdsourcing a massive dataset of visual information tagged with precise location data. Whereas the game’s popularity waned after its initial surge – though Scopely, which acquired Pokémon Go from Niantic, reports over 100 million players remained active in 2024 – the data it generated proved remarkably valuable.

Niantic Spatial’s technology can now pinpoint a location to within centimeters using just a few snapshots of surrounding buildings, a level of accuracy that surpasses traditional GPS, particularly in urban canyons and other areas where satellite signals are unreliable. “We look at the player data as very high-quality ground training data for other lower-quality datasets,” explained Brian McClendon, Niantic Spatial’s chief technology officer, in a statement. McClendon, one of the original creators of Google Earth, emphasized the strategy of using concentrated data to train models and then applying them more broadly.

The first major deployment of this technology is a partnership with Coco Robotics, which operates approximately 1,000 delivery robots across cities in the US and Europe, including Los Angeles, Chicago, Miami, Jersey City, and Helsinki. The robots will use the Niantic Spatial model to navigate complex urban environments with greater precision, improving delivery times and reliability. The initial focus is on “last-mile” delivery, the final leg of the journey from a distribution center to the customer’s door.

The success of Pokémon Go in gathering this data was, in part, accidental. The game’s design inherently encouraged players to explore and document their surroundings in a way that traditional mapping efforts, like Google Street View, could not replicate. Players ventured into more inaccessible locations and captured images in varying seasons and lighting conditions, creating a more robust and comprehensive dataset. As one observer noted, players were effectively performing unpaid labor, creating a valuable asset for the robotics industry.

While augmented reality was initially touted as the primary application for this type of data, McClendon noted a shift in focus. “Everybody thought that AR was the future, that AR glasses were coming,” he said. “And then robots became the audience.”