Roughly 30 billion player-captured scans built a Visual Positioning System that locates a camera without GPS. Niantic Spatial's December 2025 deal with defense contractor Vantor would push that capability into drones, and nobody will say whose footage is inside the model.
Hundreds of millions of Pokémon Go players spent years filming streets, parks, and storefronts to earn in-game items. Those scans, roughly 30 billion of them, now sit inside a camera-based navigation model owned by Niantic Spatial, and a U.S. defense contractor is preparing to put that model into drones and ground robots. The interesting question here is not whether the technology works. It is what kind of training data a consumer game can quietly turn into, and how little of that pipeline is traceable once the data is folded into a model.
What is actually being claimed
The core technology is a Visual Positioning System, or VPS. GPS works by trilaterating a signal from satellites. VPS does something different: it figures out where a camera is by matching what the camera sees against a pre-built 3D model of a place. If the model contains enough recognizable geometry, a handful of distinctive reference points a few pixels wide can be enough to fix a pose. The output is the same as GPS, a location, but the input is pixels rather than radio.
This is not new as a concept. Structure-from-motion, visual SLAM, and place recognition have been research staples for over a decade, and they ship in Niantic's Lightship VPS developer platform already. What changes the calculus is the size and density of the reference map. A VPS is only as good as its prior model of the world, and building that model is the hard, expensive part. You need huge volumes of ground-level imagery from many angles, many times of day, across many seasons. Crowdsourcing that from a game is an efficient way to get it.
Since 2021, Pokémon Go asked players to record short 360-degree videos of real-world Pokéstops in exchange for in-game rewards. The scanning was optional, and Niantic asked separately for permission to retain the footage. Saying yes meant agreeing to extra terms that granted a transferable, sublicensable license, the legal language that lets a company resell imagery to third parties. Niantic Spatial has acknowledged that the scans trained an early version of its navigation model.
What is genuinely new
The news event is a partnership announced on December 16, 2025, between Niantic Spatial and Vantor, the defense and intelligence firm previously branded Maxar Intelligence. The plan fuses two positioning systems that operate in different layers. Niantic Spatial handles ground-level localization by aligning a camera feed against its model. Vantor's Raptor software, released in February 2025, does the equivalent job from the air using a drone camera matched against proprietary 3D terrain data.
The combined pitch is a shared coordinate frame with no satellite link. A drone overhead and an operator or vehicle on the ground would resolve to the same coordinates in real time, even under jamming. Vantor's own release names the threat plainly: GPS unavailability, spoofing, interference, and jamming. The target platforms listed include autonomous drones, vehicles, and augmented reality glasses. Field testing of the integrated system is planned for early 2026.
This matters because GPS denial is one of the most consequential capability gaps in unmanned systems right now. Electronic warfare units can switch on a jammer and strip a fleet of its positioning in seconds. Terrain-matching and visual navigation are among the few approaches that degrade gracefully under that pressure, because a camera does not depend on a signal an adversary can saturate. Programs like Shield AI's V-BAT lean on similar visual and inertial methods precisely so they keep flying when radio links die.
The training-data problem, in technical terms
The sharper issue is what happens to data once it enters a model, and why provenance becomes almost impossible to establish afterward. When you train a model on a dataset, you do not store the dataset inside the weights in any retrievable form. The images shape parameters through gradient updates, and individual contributions diffuse into distributed representations. There is no index mapping a given Pokéstop scan to a set of activations you could later point to and say, this came from that player's apartment.
That property is what makes the central denial in this story unfalsifiable. Asked whether the military-bound system relies on Pokémon Go imagery, Vantor said it would not use the game's data. It then declined to say whether the model it plans to deploy was already trained on those scans. Those are two different statements. "We will not feed it new game data" and "the model we are fielding was never trained on game data" do not overlap, and the gap between them is the entire dispute. Jeroen van den Hoven, an ethics and technology professor at TU Delft, made the point that once scans are baked into a model, tracing them back is close to impossible. That is correct as a matter of how these systems work, and it conveniently insulates any vendor from having to prove a negative.
The consent angle compounds it. A player who recorded a 360 sweep of a park, or in one documented case the inside of his own apartment, agreed to terms written for a game. Consent obtained for entertainment is a poor proxy for consent to a defense application, and the legal license language bridges that gap without the user ever revisiting it. Iris Muis, a data-ethics researcher at Utrecht University, framed the structural trap: a user at scan time cannot picture an application five years out that they might fundamentally object to.
How the pieces fit together commercially
Niantic's history makes the defense turn less surprising. The company traces back to Keyhole, a geographic mapping firm that took In-Q-Tel funding, the CIA's venture arm, in 2003. Google acquired Keyhole in 2004, and Keyhole's CEO John Hanke went on to build Google Maps, Google Earth, and Street View before founding Niantic inside Google in 2010 and spinning it out in 2015. Niantic's current CTO, Brian McClendon, also came from that Google geo lineage.
In 2025 the company split. Scopely, owned by Saudi Arabia's Savvy Games Group under the kingdom's Public Investment Fund, bought Niantic's games business for $3.5 billion in a deal that closed in late May. The technology platform spun off as a standalone Niantic Spatial under Hanke. Vantor, meanwhile, rebranded from Maxar Intelligence on October 1, 2025, and holds a $70 million follow-on award under the National Geospatial-Intelligence Agency's Global Enhanced GEOINT Delivery program. The map data and the games went in opposite directions, and the map ended up next to a national-security prime.
What this changes
For anyone building or evaluating visual navigation, the lesson is less about a single game and more about the supply chain of reference maps. The same data hunger applies to Meta's smart glasses, Apple's spatial hardware, and Waymo's street reconstructions. Any continuously scanning consumer device is, in effect, a distributed mapping rig, and the terms governing that footage rarely constrain downstream model training in ways a user would recognize.
The capability is real and arguably necessary. GPS-denied navigation is a problem worth solving, and visual positioning is a sound technical answer to it. The unresolved part is narrower: the field tests in early 2026 will demonstrate whether the air-to-ground system performs, but they will not reveal whose imagery sits inside the model, and so far neither company will. For a technology whose entire premise is knowing exactly where something is, the data's own location has been left conveniently unmappable.
Comments
Please log in or register to join the discussion