Hundreds of millions of people spent five years filming the world for a phone game, and the resulting dataset is now being fused into navigation software for military platforms that have to fly when GPS is jammed. The number that should stop you is 30 billion: that is how many geolocated images Niantic Spatial says it has gathered since 2021, when Pokémon Go started asking players to record short 360-degree sweeps of streets, statues, and storefronts for in-game rewards. On December 16, 2025, Niantic Spatial announced a partnership with Vantor to pair that ground-level model with aerial drone navigation for the defense market, with field testing slated for early 2026.

The easy headline is "your Pokémon data is steering missiles." That framing is the least useful thing you can take from this. If you build products that collect environmental data, or you plan to deploy anything that navigates on its own, the lesson is about provenance, the licensing chain that carried a games dataset out to a National Geospatial-Intelligence Agency prime contractor, and what that chain says about every broad terms-of-service grant your own users have signed.

The data outlived the game that collected it

Watch the corporate plumbing, because it is the whole mechanism. On March 12, 2025, Scopely (backed by Saudi Arabia's Savvy Games Group) agreed to buy Niantic's games business, including Pokémon Go, for $3.5 billion. Two months later, in May 2025, Niantic spun its geospatial AI division into a separate company, Niantic Spatial Inc. The game went to one owner. The map and the model stayed with another.

That split is the point. The scans were collected under a Pokémon Go feature, but the asset they created was deliberately kept out of the sale. A consumer game was, in retrospect, an enormous data-collection program whose output had a second life its players never agreed to in any meaningful sense. The thing being monetized at the end was never the game. It was the corpus.

Why a camera beats a satellite when someone is jamming you

Satellite positioning is the first thing a competent adversary degrades. Spoofing and jamming are ordinary features of contested airspace now, not exotic threats, and a drone that loses GPS loses the ability to know where it is. That is the problem a visual positioning system solves. Instead of trusting a satellite signal, a VPS matches a live camera feed against a pre-built 3D model of the world and works out the device's position from what it sees.

The threshold for this is lower than most people assume. Per DroneXL's reporting, two recognizable reference points a few pixels wide can be enough to localize a device against the model. That is why a camera-based approach survives in degraded environments where other sensors give up: it does not need a clean signal, it needs a couple of features it recognizes. And the recognition only works if the model already contains a rich, ground-level picture of the area. Which is exactly what 30 billion first-person scans buy you.

This is the part that makes the dataset strategic rather than merely large. Satellite imagery sees rooftops. Player scans captured the messy ground-level view of cities, the interiors of parks and storefronts and the spaces between buildings, the perspective a low-flying drone or a ground robot actually operates in. No competitor can buy that, because it took years of distributed human labor across the planet to assemble. The moat is not the algorithm. The moat is the corpus, and the corpus is not for sale.

Vantor is not running an experiment

It would be easy to dismiss this as a press-release stunt. The partner says otherwise. Vantor rebranded from Maxar Intelligence on October 1, 2025, and it is a prime contractor to the National Geospatial-Intelligence Agency. DroneXL, citing Vantor, reports it holds a roughly $70 million follow-on award under the Global Enhanced GEOINT Delivery program, serving more than 400,000 U.S. government users. The December 16 deal pairs Niantic Spatial's ground-level VPS with Vantor's Raptor aerial drone navigation software, which launched in February 2025 per SpaceNews. Ground truth from below, aerial perspective from above: that combination is what makes terminal navigation work without a satellite fix.

The people involved sharpen the picture further. Niantic Spatial's CTO is Brian McClendon, who previously led Google Maps, Earth, and Street View. Founder John Hanke built the teams behind Google's mapping products. This is the same group that mapped the world once, now applying the playbook to a first-person 3D layer and pointing it at defense. Treat it as a serious, well-resourced effort, not a side project.

The denial that doesn't deny what you think it denies

Here is the honest counterpoint to the outrage, and it cuts the other way from how it first reads. Vantor has stated it will not use Pokémon Go data. Good. But it has declined to say whether the model it is fielding was already trained on that data. Those are different claims, and the gap between them is the entire story for anyone who evaluates provenance for a living.

A model trained on a dataset does not become untrained when you stop feeding it new scans. The learned representation is baked in. So "we will not use the data going forward" is fully compatible with "the model we are deploying learned the world from that data." If you read the denial as a clean bill of health, you have been told something narrower than you heard. When a vendor answers the easy question and skips the hard one, the missing answer is the answer.

This is the LLM provenance problem, with worse inputs

Security and governance teams spent the last year learning to ask where an LLM's training data came from. Spatial models need the same scrutiny, and the inputs here are messier than a scraped text corpus. This data was collected from civilians, in public and private spaces, under game incentives, then relicensed across a corporate split and out to a defense contractor.

The lever that made all of it legal is two words in a terms-of-service grant. Niantic's 2021 scanning terms reportedly gave the company a transferable, sublicensable license to the imagery. "Transferable" is what let the data survive the sale to Scopely and stay with Niantic Spatial. "Sublicensable" is what lets it flow to Vantor. If your own privacy policy contains those two words next to user-generated sensor data, your users' data can legally end up somewhere you never named, and you will have signed off on it without noticing. The provenance trail is the asset and the liability at the same time.

What to audit before the early-2026 field test ships

Treat the early-2026 Vantor field test as your deadline. If integrated VPS plus Raptor testing goes well and moves into procurement, every defense and logistics integrator will start hunting for proprietary 3D world models, and consumer-collected spatial datasets become a contested asset class overnight. Get ahead of it now.

  1. Grep your own legal text this week. Search your privacy policy and terms of service for "transferable" and "sublicensable." If either appears next to user-generated sensor data (camera, location, audio, depth), assume that data can be sold or relicensed to a party you have never disclosed. Scope the grant to your own services before the dataset gets valuable enough that your own legal team resists narrowing it. After the corpus has a buyer, you have lost the argument.
  1. Demand training-data lineage when you procure autonomous navigation, the same as you would for an LLM. Put it in the contract. And when a vendor says it "will not use" a controversial dataset, ask the exact second question Vantor declined to answer: was the fielded model already trained on it? If the answer is no answer, score that as unverified provenance, not as clean. Do not let the easy denial close the file.
  1. Design for a GPS-denied baseline if your operating area has any jamming history at all. The trigger is concrete: if your deployment zone has any recorded history of jamming, spoofing, or signal loss (and a dense urban canyon counts), make visual or inertial positioning the primary mode and treat GPS as the backup, not the reverse. A camera-against-world-model fallback stops being a nice-to-have the moment your environment can be degraded by anyone who wants to.
  1. If you sit on first-person scan data, decide now what it is worth and who can have it. The companies with the largest ground-level scan corpora, not the ones with the best sensors, are the ones this deal makes valuable. If you are collecting that kind of data, you are holding an asset class that just got a price. Either lock down who can license it, or be honest with your users that you are building something you intend to sell.

The drones are the spectacle. The contract language is the lesson. One of those you can do something about before early 2026, and it is not the drones.

Sources

  • https://dronexl.co/2026/06/09/pokemon-go-scans-niantic-vantor-military-drone-navigation/
  • https://spacenews.com/vantor-partners-with-niantic-spatial-on-gps-free-navigation-for-defense-market/
  • https://en.wikipedia.org/wiki/Niantic_Spatial