Pokémon GO’s Hidden Role in Training the Next Generation of Autonomous Robots
Billions of images captured by Pokémon GO players are helping build detailed digital maps that researchers and AI companies can use to improve how autonomous robots navigate complex pedestrian environments.
In 2021, Niantic added a feature to Pokémon GO called AR Mapping. Players pointed their phones at PokéStops and walked a slow circle, capturing a few seconds of video for minor in-game rewards. It was optional, and largely unloved, a chore task buried in an app most people opened for Pokémon, not photogrammetry.
AR Mapping wasn't the main event, though. It was the smallest and most visible piece of a much larger design, one built around getting players to cover more ground, revisit the same landmarks, and do it at different times of day, for years. Niantic Spatial's own account of its training data reflects that pattern. The images aren't spread evenly across a city; they cluster densest around what the company calls hot spots, the same PokéStops and Gyms the game was already routing players toward for unrelated reasons. Ten years of that produced roughly 30 billion ground-level images, according to the company. It now calls the resulting system a Large Geospatial Model, the core of a Visual Positioning System built on that imagery. That system currently guides delivery robots down sidewalks in Helsinki.
Playing it myself
Before writing this, I spent time actually playing Pokémon GO rather than just reading Niantic's technical papers about it, because the "players unknowingly generated a dataset" framing only tells you the outcome, not the mechanism. Once you're inside the game, the mechanism is not subtle.
There's an in game feature called Field Research in the game. Basically daily mini-quests you get from spinning PokéStops and other in game thing like walking 2 km or somethiimes, and if you do them you get an encounter with something genuinely hard to find, a Mew, a regional-exclusive, whatever everyone's chasing that month. If you're the kind of player trying to fill out your Pokédex, that's not some side detail, that's the whole reason you're playing. Nobody's doing the task because they enjoy the walk over to the PokéStop. You do it because you want the Pokémon waiting at the end, and the walking is just what it costs you to get there. Special Research works on the same idea but bigger, these multi-stage quest chains tied to legendary and mythical Pokémon events that stretch across days or sometimes weeks, so you're paying that same toll over and over, at different spots around town, whenever the game decides to hand you the next step. Then there's Routes, which Niantic added back in July 2023, and it takes the whole thing a step further. You record a walking path between two PokéStops and publish it, and now other people can follow it too, and everyone who walks it, you included, gets Stardust and XP for the trouble. It basically turns players into unpaid mapmakers for each other.
Some of the mechanics are less about where you go and more about what your phone is doing while you're there. The game tracks elevation, not just flat GPS coordinates, which matters for a model trying to understand 3D space rather than a 2D map. Location permission has to be set to "precise," not "approximate." The game outright won't function on the coarser setting most privacy-conscious phone users default to. And feeding, petting, or opening a gift from your Buddy Pokémon, a core part of daily play, requires AR mode turned on, meaning the camera is active and reading the environment around you, not just your coordinates.
Then there's timing. Niantic runs recurring bonus events that specifically reward playing at hours the game doesn't naturally get much traffic, double Stardust windows and similar bonuses timed for evenings and nights, when a city block looks nothing like it does at 2 p.m. If the goal is a model that recognizes a location in any lighting condition, a static daytime-only dataset is much less useful than one with the same corner photographed at noon and at 9 p.m. An incentive that specifically pulls players out after dark solves that gap without ever mentioning that's what it's doing.
None of this required a conspiracy. Every one of these mechanics has an obvious, honest gameplay justification: walking is central to Pokémon as a franchise, elevation makes AR placement look more realistic, night events keep a five-year-old game from going stale. But the cumulative effect of all that ordinary design was a dataset with precise coordinates, elevation, camera orientation, and lighting variation, gathered at almost no marginal cost, from people who were there to catch a Squirtle.
Coco Robotics
Coco Robotics runs about 1,000 delivery robots, sized to carry up to eight large pizzas or four bags of groceries, across Los Angeles, Chicago, Jersey City, Miami, and Helsinki. CEO Zach Rash says the fleet has completed more than half a million deliveries covering a few million miles in all weather.
Coco became Niantic Spatial's first robotics partner in March 2026. The problem it's solving is mundane and specific: GPS is unreliable in dense cities, where radio signals bounce off glass and steel and a phone's blue dot can drift 50 meters off course, enough to put a robot on the wrong side of the street. Niantic Spatial's positioning system compares what a robot's cameras see against its trained model to pin location down to within centimeters, using visual landmarks instead of satellite signal alone.
For Coco, better localization means fewer robots stuck at the wrong door or drifting into traffic, the kind of failure that decides whether a delivery robot business survives contact with an actual city. Rival Starship Technologies uses a comparable approach, building its own 3D maps from onboard sensors. Niantic Spatial's argument is that a decade of crowdsourced, multi-angle imagery is a head start no competitor can easily replicate.
Correcting the ownership story
The popular version of this story says Google bought Niantic and is now using Pokémon GO to build robots. That's backward. Niantic began in 2010 as an internal incubator inside Google, started by Keyhole founder John Hanke, the same engineer behind the technology that became Google Earth. Google spun Niantic out as an independent company in October 2015, during the restructuring that created Alphabet, backing it with about $30 million in seed funding alongside Nintendo and The Pokémon Company. Pokémon GO launched the following year under that independent company, not under Google.
Google hasn't had a stake in Niantic's decisions since 2015, and has none in Niantic Spatial, which exists because Niantic sold its games business to Scopely, a studio owned by Saudi Arabia's Public Investment Fund, for about $3.5 billion in 2025. The robot-navigation business is owned by Niantic's original investors, not by Google and not by Scopely. If anything, the real ownership story is odder than the popular one: a Google-incubated mapping project spent a decade gathering data through a game, then split between a Saudi-backed gaming company and an independent AI spinout with no remaining Google stake at all.
The consent question
Someone who spun a PokéStop for AR Mapping credit in 2021 was playing a scavenger-hunt mechanic for a handful of PokéCoins. Nothing in that exchange suggested the footage would end up training commercial infrastructure sold to robotics companies five years later. It's not the first time Niantic has faced this question. Senator Al Franken pressed the company in 2016 over whether parents had enough information to meaningfully consent to their kids' location data being collected, and a separate 2016 California lawsuit accused Niantic of placing PokéStops on private property without owners' consent. The specifics change; the pattern doesn't. Broad terms of service, a feature nobody remembers agreeing to, and a use case invented years after the fact.
The Vantor question
Niantic Spatial partnered with Vantor in December 2025. Vantor is the defense and intelligence firm that rebranded from Maxar Intelligence, and the deal combines ground-level positioning with aerial navigation for drones operating where GPS is jammed or spoofed. Vantor's existing work includes the National Geospatial-Intelligence Agency's Global Enhanced GEOINT Delivery program, a contract worth an estimated $70 million. When the Dutch outlet Trouw reported on the partnership, followed by Kotaku and Slashdot, the obvious question followed: did Pokémon GO scans end up training navigation systems bound for military drones.
There's no public evidence they did. But there's also no clean answer, because the chain runs through the company rather than around it. Pokémon GO scans helped train Niantic Spatial's early models, and those models are the foundation the company is now adapting for GPS-denied environments generally, Vantor included. That's a different claim than "your Pokémon GO photos are on a drone," and it's the one that actually holds up. It was still enough to put "Pokémon GO" and "military drones" in the same headline across gaming press for a week.
How Niantic Spatial responded
Facing that coverage, Niantic Spatial narrowed its claim rather than disputing the history. A company spokesperson told Kotaku that Pokémon GO data isn't shared with Niantic Spatial now that the game belongs to Scopely, and that AR scans were submitted voluntarily under whatever terms of service applied at the time. Separately, the company confirmed to reporters that Pokémon GO data was one input used to train an early version of its model, before that data-sharing arrangement ended with the Scopely split. On Vantor specifically, the company said it had nothing further to add. It told IGN that data-sharing isn't part of the Vantor agreement, and that it wants its models used by partners "in a responsible manner" consistent with human rights principles.
What a player can actually check
Niantic Spatial's privacy policy allows users to opt out of contributing further scans through in-app settings, and states that shared imagery is anonymized, including blurring, before being used to train models. That control only works going forward. Nothing in the policy describes a way to remove imagery already absorbed into the trained system. Anyone who played between 2021 and the 2025 Scopely sale contributed to a dataset that has since changed hands and purpose more than once, with no lever to pull their share back out.
Where it stands
Niantic Spatial's case: a decade of voluntarily submitted, GPS-tagged imagery is legitimate infrastructure, and it's already working, robots navigating real sidewalks in five cities, Helsinki included. The case made by Kotaku, Slashdot, and the gaming press that picked up Trouw's reporting: a reward mechanic built around catching cartoon monsters was never a meaningful basis for consent to train navigation systems now being adapted for defense-adjacent uses a decade later, whatever the current contract language says.
Both can be true. The scans were submitted under the rules in place at the time, and almost nobody submitting them could have guessed where the resulting model would end up. The Vantor partnership is still early. Niantic Spatial maintains Pokémon GO imagery isn't part of it, and no public document contradicts that. What isn't in dispute is that the dataset itself, 30 billion images, a decade of collection, no way for any single player to withdraw their piece of it, isn't going anywhere.
