Whether you're looking for an angel investor, a growth advisor, or just want to connect — I'm always open to great ideas.
Get in TouchNiantic's Pokémon GO generated 30 billion user photos that are now training delivery robots for real-world navigation. The biggest gamification-to-AI training pipeline ever raises questions about consent, data rights, and the future of crowd-sourced robotics.
AI, startups & growth insights. No spam.
TL;DR: That Pokémon you caught near a coffee shop in 2016 contributed to one of the most consequential AI training datasets ever assembled outside a tech giant's internal walls. Niantic's spinoff, Niantic Spatial, has trained a visual positioning system on 30 billion ground-level images collected by Pokémon GO players over nearly a decade — and is now deploying it to help delivery robots navigate city sidewalks to within a few centimeters of precision. The partnership with Coco Robotics, announced in March 2026, is the first major commercial deployment of a model built entirely on gamified crowd-sourced data. The privacy questions it raises are ones the broader tech industry has largely avoided answering.
Pokémon GO launched in July 2016 and became a cultural phenomenon faster than almost any technology product in history. Within weeks, hundreds of millions of people were walking through cities, parks, and neighborhoods with their phone cameras pointed at the physical world to catch digital creatures overlaid on real locations. That core mechanic — augmented reality anchored to real-world geography — required the app to understand where players were standing, what they were looking at, and how physical landmarks related to each other in three-dimensional space.
From the beginning, Niantic was building something more than a game. The company's core technology was a geospatial platform: a system for mapping the physical world at ground level using consumer-grade phone cameras, in places and at angles that satellite imagery could never capture. Every scan, every photo, every augmented reality interaction added another data point to a map that was being built crowd-sourced, at scale, across nearly every major city on the planet.
The dataset accelerated dramatically in 2020, when Pokémon GO introduced what it called "Field Research" — a feature that prompted players to scan real-world statues, storefronts, fountains, and local landmarks with their cameras in exchange for in-game rewards. Points. Rare Pokémon. Power-ups. The mechanic was straightforward: walk up to a landmark, scan it, get something valuable in the game. What players were less aware of was that those scans were being stitched together into 3D models of physical spaces that Niantic's systems were using to build a persistent, centimeter-accurate map of urban environments around the world.
By early 2026, that dataset had reached 30 billion images — clustered around the "hot spots" that Pokémon GO had established as important in-game locations: battle arenas, PokéStops, gyms. Those locations correspond precisely to the kinds of places where delivery robots need to operate: restaurant entrances, apartment building lobbies, store fronts, urban intersections. The overlap is not coincidental. It is structurally baked into how Niantic designed the game.
Niantic's story in 2025 is one of the more striking strategic pivots in recent tech history. In May 2025, the company sold its games division — including Pokémon GO, Pikmin Bloom, and Monster Hunter Now — to Scopely, a mobile gaming company. The Pokémon GO brand and its hundreds of millions of players went with that sale. What Niantic kept was the underlying technology: the geospatial AI platform, the 30 billion image dataset, and the team that built it.
The spun-off entity is called Niantic Spatial Inc., funded with $250 million — $200 million from Niantic's own balance sheet and $50 million from Scopely as part of the transaction. John Hanke, Niantic's founder and former CEO, leads Niantic Spatial as chief executive. Brian McClendon, former VP of Engineering at Google where he helped build Google Maps and Google Earth, serves as chief technology officer. Thomas Gewecke, previously a senior executive at Warner Bros., is chief operating officer.
McClendon has been direct about the strategic logic behind the pivot. "Everybody thought that AR was the future, that AR glasses were coming," he said in an interview with MIT Technology Review. "And then robots became the audience." The insight is blunt but accurate: the Visual Positioning System technology Niantic had been building for augmented reality headsets is, it turns out, precisely what autonomous delivery robots need to navigate city sidewalks with the kind of precision that GPS cannot provide.
The sale of the games division effectively freed Niantic Spatial to pursue robotics partnerships and enterprise licensing without the constraints of consumer game operations. The 30 billion images stayed with Niantic Spatial. The gameplay mechanics, the brand, and the players went elsewhere. It is one of the more efficient asset separations in recent tech history.
GPS is accurate to within a few meters under good conditions — closer to 10-15 meters in urban canyons surrounded by tall buildings. That level of imprecision is acceptable when you are navigating by car and looking for a parking garage. It is completely unacceptable when you are a sidewalk delivery robot trying to stop in front of the correct building entrance, not three doors down, without blocking pedestrian traffic or occupying the wrong pickup zone outside a restaurant.
Niantic Spatial's Visual Positioning System (VPS) bypasses GPS entirely for the final positioning problem. Instead of relying on satellite signals, it compares the live camera feed from a robot's onboard camera against the 30 billion image database in real time. The system identifies visual landmarks in the frame — a specific storefront facade, a particular statue, a distinctive intersection corner — and triangulates position based on what it recognizes. The result is positioning accuracy measured in centimeters, not meters.
The technical challenge is significant. The system must recognize locations across different lighting conditions, different seasons, different times of day, with different pedestrians and vehicles in the frame. A storefront that looks one way at noon looks completely different at midnight. A park bench surrounded by leaves in summer looks different in winter. The 30 billion image dataset, collected over years across all seasons and conditions, provides the variation needed to make the system robust to these changes.
The VPS does not replace GPS for routing and long-range navigation — it supplements it. A delivery robot still uses GPS and mapping services to plan its route from a restaurant to a customer's address. The VPS kicks in for the last few dozen meters: finding the exact pickup spot, stopping precisely at the correct door, identifying the right building entrance from the street. It is the difference between "somewhere on this block" and "exactly here."
Coco Robotics is a Los Angeles-based company that builds and operates sidewalk delivery robots for restaurant and retail last-mile delivery. Founded in 2020, Coco currently deploys approximately 1,000 robots across cities including Los Angeles, Chicago, Jersey City, Miami, and Helsinki, integrating with platforms like Uber Eats, DoorDash, and Wolt. The company has completed more than 500,000 zero-emission deliveries and serves over 3,000 merchants.
The partnership with Niantic Spatial, announced in March 2026, is the first large-scale commercial deployment of the VPS technology outside Niantic's own applications. Coco's robots integrate the VPS directly: when a robot approaches a restaurant for pickup, the VPS identifies the correct entry point and positions the robot at the designated spot — not approximately there, but within the specific area the restaurant has defined for robot pickup. When the robot arrives at the customer's location, the same system positions it precisely outside the correct door.
The practical improvement is more meaningful than it sounds. Centimeter-level precision determines whether a robot blocks foot traffic, whether it parks in the correct pickup zone that a restaurant has operationally integrated into its workflow, and whether a customer can easily access their delivery without the robot having stopped awkwardly far from the entrance. These small positioning errors compound at scale across thousands of deliveries per day.
Coco's stated goal is to scale its fleet to 10,000 robots globally by the end of 2026. At that scale, the Niantic VPS partnership is not a feature — it becomes infrastructure. Every robot in every new city needs to know precisely where it is on a sidewalk, and the visual database Niantic has built provides that foundation without requiring Coco to instrument every street corner with its own sensors.
The headline fact is straightforward: hundreds of millions of Pokémon GO players contributed to a 30 billion image dataset that is now training commercial AI for delivery robots. The more contested question is whether they meaningfully consented to that outcome.
Niantic's position is that participation in location scanning was voluntary. Players "could choose to submit anonymized scans of public places to help improve VPS," and scans are not connected to player accounts in the final training dataset. The company notes that its privacy policies have always indicated the possibility of using data for purposes beyond gaming, and that Field Research scanning was an opt-in activity rewarded with in-game items.
Critics offer a different framing. The terms of service that governed data use in 2016, 2018, and 2020 did not describe — because Niantic had not yet decided — that the data would be spun off into a separate commercial AI company and licensed to robotics firms. Consent is meaningful when the consenting party understands what they are agreeing to. "You agreed to the terms of service" is a weaker form of consent when the commercial application was not foreseeable to the person generating the data.
The gamification mechanic adds another layer. Field Research scanning was structured as an exchange: scan a landmark, get a reward. That framing positions the player as an active, willing participant. But the asymmetry of information between what Niantic knew it was building and what the average player understood they were contributing to is significant. Most players scanning a statue for a Rare Candy were not thinking about robot navigation. Most were thinking about their next raid battle.
This dynamic is not unique to Niantic. Waze users contribute traffic data that Google (which acquired Waze in 2013) uses for navigation products across its entire ecosystem. Google Street View drivers contributed imagery that now underpins multiple AI training pipelines. Facebook photo uploads trained facial recognition systems whose commercial applications users did not anticipate. What makes the Niantic case notable is the directness of the chain: a consumer game, a specific in-game mechanic designed explicitly to collect spatial data, and a commercial robotics product that would not exist without that specific data.
Regulators are beginning to pay attention. The EU's GDPR framework requires that data collected for one purpose not be repurposed for materially different uses without fresh consent. Whether training a commercial robotics VPS from gaming data constitutes a "materially different purpose" from improving augmented reality experiences is an open legal question in European courts. In the United States, no equivalent federal framework exists, leaving the question governed primarily by terms of service and FTC enforcement discretion.
Niantic Spatial's approach is distinctive but not unique. Several companies are building navigation and positioning systems from large-scale real-world data collection, each with different sources, methods, and commercial structures.
Tesla has built its entire autonomous driving stack on fleet learning: every Tesla on the road contributes camera data back to Tesla's training pipeline, creating a dataset of billions of miles across real-world driving conditions. Tesla's approach uses its consumer product as the data collection vehicle, with customers implicitly consenting through their purchase agreements and software terms. The scale is enormous — millions of vehicles contributing simultaneously — but the data is car-centric and highway-biased, less useful for sidewalk navigation.
Waymo has taken the opposite approach: a smaller, controlled fleet with extraordinarily detailed sensor data per vehicle, including lidar point clouds, high-definition radar, and multi-angle camera arrays. Waymo's dataset is richer per mile but narrower in geographic coverage, reflecting a deliberate decision to prioritize quality and safety validation over breadth. Waymo's partnership with DoorDash for autonomous food delivery uses car-based delivery (food in the trunk of a Jaguar I-Pace), not sidewalk robots — a different last-mile solution for different urban contexts.
DoorDash is pursuing a multi-modal autonomous delivery strategy: sidewalk robots through its partnership with Serve Robotics in Los Angeles, drone delivery in select markets, and its own internally developed delivery bot called Dot, being tested in Phoenix. DoorDash's approach is to deploy multiple modalities and determine which works best in which context, rather than betting on a single platform.
None of these competitors has access to anything analogous to the Niantic dataset: 30 billion ground-level urban images collected across nearly a decade, clustered precisely around the commercial and civic locations where delivery robots operate. Tesla's fleet data is from roads, not sidewalks. Waymo's data is sensor-rich but geographically narrow. DoorDash's robots are still accumulating operational data. Niantic Spatial's dataset is, for the specific problem of sidewalk-level urban positioning, arguably the most relevant training resource that exists.
The Niantic case raises an obvious question that the tech industry has so far avoided answering directly: which other consumer applications are sitting on datasets that could become the next commercial AI training pipeline?
Google Street View is the most obvious analog — billions of images of streets and buildings captured by camera cars and backpack-wearing operators — but it is already Google's internal asset, already powering multiple AI products, and its collection method is well understood by regulators.
Strava has GPS traces from hundreds of millions of athletes worldwide, including precise path data through parks, trails, and urban neighborhoods that could train pedestrian and cycling navigation systems. Strava already sells aggregated anonymized data to city planners through its Metro product.
Waze (Google): real-time traffic data from 140+ million users, including road conditions, hazards, and speeds, already integrated into Google Maps. The secondary commercial value is baked into the acquisition logic.
Ingress (also Niantic): the predecessor game to Pokémon GO that seeded the original PokéStop location database. Ingress players effectively designed the geographic infrastructure that made Pokémon GO's data collection possible.
Snap and Instagram: billions of geo-tagged images and videos captured in specific locations, already used for augmented reality features like Snap's Local Lenses. The training potential for location-specific visual models is substantial.
The common thread is gamification or social reward: mechanisms that give users a reason to contribute data that would otherwise require paid collection. In each case, the gap between what users understood they were contributing to and what the data ultimately enabled is the central ethical and regulatory tension.
Labeled training data is among the most expensive inputs in modern AI development. Companies like Scale AI have built billion-dollar businesses on the back of human annotation: paying contractors to label images, transcribe audio, and identify objects in video frames so that machine learning models can learn from structured examples.
Niantic's approach inverted this model entirely. Instead of paying annotators, it paid players — in virtual currency worth fractions of a cent per scan — to generate labeled, geo-anchored, contextually rich data at a scale that would have cost hundreds of millions of dollars to produce through conventional annotation pipelines. The "payment" was a Rare Candy or a Stardust bonus. The data was ground-level photographic coverage of every significant urban location on Earth.
The economics are extraordinary. A conservative estimate: 30 billion images at even $0.01 per image annotation cost would represent $300 million in data production value. Niantic's actual cost to produce that dataset was the ongoing maintenance of a free-to-play game — a business that generated its own revenue through in-app purchases while simultaneously building the dataset. The game funded the data collection while the data collection made the game more accurate.
This is the template that makes the Niantic case strategically significant beyond the specific Coco Robotics partnership. It demonstrates that consumer applications can serve as training data collection systems at a scale, quality, and cost structure that dedicated data collection cannot match. The question for the industry — and for regulators — is whether this template is legitimate, and if so, what disclosure obligations come with it.
Niantic Spatial's stated mission is to build spatial intelligence that helps machines understand, navigate, and engage with the physical world. The Coco Robotics partnership is the first commercial deployment, but it is unlikely to be the last.
The VPS technology is directly applicable to any autonomous system that needs centimeter-accurate positioning in urban environments: delivery drones identifying rooftop landing zones or balcony drop points, humanoid robots navigating indoor commercial spaces, autonomous wheelchairs locating building entrances, AR glasses overlaying persistent digital content on physical locations. Each of these use cases benefits from the same underlying capability: knowing exactly where you are based on what you can see.
Niantic Spatial has $250 million in funding and a dataset that took a decade to build and cannot easily be replicated. The company's competitive moat is not the VPS algorithm itself — that is reproducible. It is the 30 billion images: the geographic breadth, the temporal depth (multiple seasons, multiple years), and the ground-level perspective clustered around exactly the locations where autonomous systems need to operate.
Whether that moat is defensible depends on questions that are still being answered. Can competitors aggregate comparable datasets through different means — partnerships with mapping companies, purpose-built scanning programs, or deals with other app developers? Will regulatory action on data rights constrain how the existing dataset can be commercialized? Will Niantic Spatial's head start in robotics partnerships compound into durable market position, or will the technology become commoditized as the underlying AI capabilities mature?
Those questions will be answered over the next two to three years. What is already clear is that the Pokémon GO player who scanned a fountain in a city park for a Stardust bonus in 2020 contributed, without knowing it, to a commercial AI system guiding delivery robots through city sidewalks in 2026. That chain of events — from consumer game mechanic to robotics infrastructure — is the story of how AI training data actually gets built, at scale, outside the walls of the hyperscalers.
Sources:
Meta's Chief AI Scientist Yann LeCun just closed the largest seed round in European history for AMI Labs, backed by Bezos, NVIDIA, and Samsung. His bet: JEPA-based world models that learn physical laws will outperform LLMs on the path to AGI.
Oracle commits $50B in FY2026 capex for AI infrastructure, more than doubling last year's spend, as its cloud backlog surges to $553B. First to deploy NVIDIA Blackwell at scale and part of the $500B SoftBank-OpenAI mega-deal, Oracle is reinventing itself as an AI infrastructure powerhouse.
Microsoft's Copilot Cowork routes tasks to Anthropic Claude or OpenAI models based on capability, executing actions across Outlook, Teams, and Excel. With 500M+ addressable users, this is the largest enterprise agent deployment ever.