Niantic Spatial, Inc., the AI spinout from mobile game developer Niantic, Inc., is training a world model on 30 billion geolocated images crowdsourced from Pokémon Go and Ingress players over the past decade, according to an exclusive report from MIT Technology Review published April 17, 2026. The dataset — assembled without any explicit AI training consent from the players who photographed landmarks while catching virtual Pokémon — is the largest street-level visual corpus ever assembled. No competitor can replicate it, and the timeline to build something comparable starts at a decade.
What a World Model Actually Is
A world model is an AI system trained to understand, represent, and predict the physical environment in three-dimensional space. This is categorically different from large language models, which process token sequences, or diffusion models, which learn statistical distributions of pixels. A world model learns geometry: how streets connect, where buildings relate to each other, how scenes change across lighting conditions and time of day.
The training data requirements are specific. Useful spatial training data must be street-level (not satellite), multi-angle (the same location photographed from multiple viewpoints), precisely geolocated, and time-varying — captured across different seasons, hours of day, and years. Pokémon Go players, across a decade of gameplay, produced all four categories at a scale that no deliberate data collection effort has matched.
How 100 Million Players Built It Without Knowing
The mechanism was Niantic’s Wayspot submission system. Since Pokémon Go launched in July 2016, players have photographed and geolocated real-world structures — parks, transit stations, churches, murals, storefronts, and public art — to nominate them as in-game PokéStops and Gyms. Each submission included GPS coordinates, device orientation, and on newer iPhone models, LiDAR depth data. Players were incentivized by in-game rewards. Niantic collected the data.
The resulting corpus covers 190 countries and is weighted toward pedestrian-scale urban environments. It captures the same landmarks thousands of times across different angles, seasons, and lighting conditions. Google Street View, by comparison, is captured from vehicle-mounted rigs at fixed intervals. Pokémon Go photos were taken by people, at human height, from every walkable angle. The coverage density in major cities — where the game’s player base concentrated — is unlike anything in any competing dataset.
MegaOne AI has previously covered how autonomous exploration systems like Nomad are pushing the limits of AI-driven spatial navigation; the limiting factor for all of these systems is training data quality, and Niantic Spatial just announced it holds the world’s best supply.
Why Street-Level Beats Satellite — and Beats Street View
Satellite imagery is irrelevant to the primary applications of world models. An autonomous robot navigating a warehouse corridor cannot use overhead imagery. An AR headset anchoring a digital overlay to a building facade needs to know what that facade looks like from 1.7 meters off the ground, not from 400 kilometers up.
Street View is closer to the right data type but has structural limitations. It is captured from vehicles, which means pedestrian plazas, building interiors, hiking trails, and transit platforms are largely absent. It provides single-pass coverage per update cycle — not the multi-angle, multi-temporal data required for accurate 3D reconstruction. And its per-location density is insufficient for the temporal modeling that world models require.
Niantic’s corpus has a different character. A popular PokéStop in a major city has been photographed thousands of times from dozens of angles across every season and time of day over ten years. That temporal and angular density allows a world model to generalize from a location — to predict what it looks like from an unseen angle or at an unseen time. That predictive spatial capability is exactly what autonomous systems need and have never had at this scale.
Niantic Spatial: The Corporate Spinout
Niantic carved its AI and spatial computing work into a separate entity, Niantic Spatial, Inc., as part of a deliberate restructuring. The separation allows Niantic Spatial to raise capital from deep-tech and enterprise investors without diluting Niantic’s gaming business, and it repositions the dataset as a licensable infrastructure asset rather than an internal competitive moat kept behind closed doors.
The spinout structure mirrors the playbook used across several major AI pivots. The core defensible asset is not the model architecture — it is the data. Niantic Spatial’s structural position is the 30 billion image corpus, not any particular model trained on it. Licensing access to that corpus, or to spatial API capabilities derived from it, is the clear commercial path. The AI infrastructure race increasingly favors whoever controls foundational data layers, not just compute or model architecture, and Niantic Spatial is entering that race holding a card no one else has.
The Applications: Robotics, Autonomous Vehicles, and AR
The three most immediate commercial applications for a world model trained at this scale are autonomous vehicles, general-purpose robotics, and augmented reality navigation.
Autonomous vehicles rely on HD maps — manually annotated representations of road environments that cost approximately $1,000 per kilometer to produce and require regular updates as environments change. A world model capable of generating spatial representations from raw imagery cuts that cost materially and enables operation in unmapped environments. Every autonomous vehicle company faces this constraint at scale.
Humanoid and mobile robotics faces the sim-to-real gap: systems trained in simulated environments fail in physical spaces because simulations cannot capture real-world complexity and variation. A world model trained on actual pedestrian environments closes this gap directly. The current cohort of humanoid robotics companies — collectively valued in the tens of billions — all need this kind of foundation data, and none of them have it.
Augmented reality is Niantic’s home domain. The company’s existing Lightship AR platform already uses visual positioning to anchor digital content to physical locations. A world model improves positioning accuracy and enables persistent AR: digital objects that stay anchored to real locations across sessions and devices, at the fidelity that Apple Vision Pro, Meta Quest, and next-generation AR glasses require. This is the application Niantic has been building toward for a decade, and the dataset was being assembled the entire time.
The Privacy Problem No One Consented To
Niantic’s terms of service, like most consumer app agreements, grants broad rights to user-submitted content for improving its services. Whether training a commercial AI world model falls within that scope is a live legal question that privacy advocates and plaintiffs’ attorneys are likely already analyzing.
The players who submitted 30 billion photos to Pokémon Go over the past decade were playing a mobile game. They had no indication their submissions would become training data for commercial AI products licensed to robotics and autonomous vehicle companies. Many of those images incidentally capture people, license plates, building security infrastructure, and private property visible from public areas. The consent question is structurally identical to ongoing litigation around image generation models trained on artist portfolios without compensation.
The Humans First movement has been specifically targeting these consent gaps in AI training pipelines, and Niantic Spatial is entering a regulatory environment significantly more hostile than the one that existed when Pokémon Go launched. The EU AI Act’s data governance provisions impose explicit obligations on foundation model providers regarding training data provenance and documentation. How Niantic Spatial demonstrates valid consent for a decade-old consumer dataset will face scrutiny — and the answer is not obvious.
The Moat No Competitor Can Buy
The most consequential fact about this dataset is not its size. It is its irreproducibility.
Building a comparable corpus would require deploying millions of motivated, GPS-equipped human photographers to every pedestrian environment on earth, over multiple years, capturing the same locations from multiple angles under varying conditions. That cannot be manufactured on any timeline that matters for the current AI infrastructure buildout. It required a compelling consumer application, ten years of sustained gameplay, and a global user base that Niantic spent years and hundreds of millions of dollars cultivating.
Even companies deploying $10 billion in AI infrastructure cannot purchase a substitute for a dataset that requires a decade of human behavior to generate. Tesla’s autonomous driving data was accumulated from 5 million vehicles over years of fleet operation. Waymo‘s spatial data is locked inside a closed commercial ecosystem. Niantic Spatial’s corpus is the only large-scale street-level spatial dataset assembled from pedestrian-scale, multi-angle perspectives with global coverage — and the only one structured as a licensable asset available to third parties.
MegaOne AI tracks 139+ AI tools across 17 categories. The spatial AI category — localization, world modeling, AR positioning, robotic navigation — is one of the fastest-consolidating segments in the current landscape. Niantic Spatial holds the most significant structural data advantage of any company operating in that space. The players built it over ten years while hunting Pikachus. The company owns it entirely. The next decade of physical-world AI will be trained on it.
