Niantic’s new artificial intelligence service promises to revolutionize how delivery robots navigate urban environments, leveraging a treasure trove of geospatial data collected over years by millions of Pokémon Go players.
Originally launched in 2016, Pokémon Go became the first global augmented reality phenomenon. Within just 60 days, 500 million people downloaded the app, taking to the streets of cities worldwide—from Chicago to Tokyo—with their phones held high, capturing virtual creatures superimposed onto real buildings and landmarks. Remarkably, eight years later, over 100 million players remained active in 2024.
This massive player base, pointing their cameras at countless corners of the planet, has generated an unprecedented visual archive. Niantic Spatial—the AI division the company created last May—has been using this flood of imagery to build what it calls a “world model”: technology that connects artificial intelligence with real physical environments with millimeter precision.
From Phone Screens to Sidewalks
Their visual positioning system can determine a device’s exact location—with a margin of error of just a few centimeters—by analyzing a few snapshots of the surroundings. And although this technology was initially conceived for future augmented reality glasses, Niantic has found an unexpected customer: autonomous delivery robots.
The startup Coco Robotics, which operates nearly a thousand suitcase-sized robots in cities like Los Angeles, Chicago, and Helsinki, will be the first to integrate this solution. Their machines, which travel along sidewalks at 5 mph making food deliveries, face a critical challenge: GPS fails spectacularly in urban “canyons” of skyscrapers, where signals bounce and can drift up to 50 meters, sending the robot to the wrong sidewalk.
The Living Map
Niantic’s secret lies in the quality of its data. For years, the company’s games—including Pokémon Go and the earlier Ingress—encouraged users to visit specific locations (gyms, stops, battle arenas), generating over a million “hot spots” worldwide. For each of these, the company has thousands of images taken from different angles, at different times of day and weather conditions, all tagged with extremely precise metadata about the phone’s position and orientation.
With this archive of 30 billion images, they have trained a model capable of orienting itself even in places that aren’t part of those hot spots. Coco’s robots, equipped with four cameras, will thus be able to stop exactly at a customer’s door or position themselves correctly in a restaurant pickup zone—something GPS couldn’t guarantee.
The Future: Maps Designed for Machines
For John Hanke, CEO of Niantic Spatial, we are witnessing the beginning of a “Cambrian explosion in robotics.” As more machines share spaces with humans—sidewalks, construction sites, shopping centers—they will need a level of spatial understanding similar to our own. The alliance with Coco is just the first step toward what Hanke calls a “living map”: a hyper-precise digital replica of the world that constantly updates as the robots themselves generate new mapping data while moving.
But the qualitative leap goes beyond precision. If traditional maps helped people orient themselves, maps for machines will need to function as semantic guides: not only indicating where things are, but labeling each object with descriptions that allow AI to understand its function in the world. “Building an understanding of how the connective tissue of the world works,” as Hanke puts it. While other companies develop models that generate fantasy virtual worlds on the fly, Niantic is betting on digitally recreating the real world in all its detail.
By: Nestor Castillo, ForAllTechNews Director