Pokémon Go data fuels geospatial AI for drones

Millions of Pokémon Go players unknowingly fed a geospatial AI that could steer delivery drones.

The arc of the story begins in May 2025 when Niantic spun out Niantic Spatial to build what the company calls real world foundation models. The move followed Niantic’s decision to sell its flagship game, Pokémon Go, to Scopely. The team reports that Niantic Spatial plans to train a large geospatial model, a 3D representation of the physical world powered by geolocated images gathered from real world scans. The scale is nontrivial: the model is trained on billions of real world images captured by millions of players, plus data from users of the Scaniverse app. The claim from Niantic Spatial is stark: ground scans were one component to help train the models, but the models themselves are the product of training, not copies of the underlying scans. The sources emphasize that the scanned data largely covered public points of interest such as statues and fountains rather than private spaces or sensitive locations.

In practical terms, the geospatial model aims to help machines navigate complex environments. The team points to navigation technology for delivery robots as a primary use case, with a nod to potentially broader, dual use applications such as military drones. That caveat (dual use potential) appears explicitly in the conversation around the project, underscoring a broader tension in AI development: consumer facing data can seed capabilities that experts worry could be repurposed for surveillance or combat scenarios. The reports describe a deliberate stance: the output is meant to be a model, not a direct dump of the scans, which Niantic Spatial frames as a safeguard for IP and privacy while enabling downstream robotics tasks.

From a practitioner’s perspective, the engineering constraints are evident. The sweep of data from billions of user captured images to scans from the Scaniverse app must be converted into a coherent 3D world that can support reliable localization, mapping, and obstacle avoidance. That transition is nontrivial: real world scenes vary with weather, time of day, and crowd activity, and the model must generalize across cities, neighborhoods, and landmarks. The claim that the models are trained on, rather than copies of, scans matters for intellectual property and rights management: it offers a path to leveraging user generated visuals without circulating raw data, but it also raises questions about what ends up in the trained representation and what remains behind the scenes for privacy and consent.

Another practical angle is governance. The dual use risk is not merely theoretical: if a geospatial backbone can guide drones with low level precision, it becomes a strategic asset for multiple players in logistics and security. That reality nudges stakeholders toward tighter disclosure, consent regimes, and perhaps more granular control over what kinds of environments are used for training. The Niantic Spatial team’s framing (public POIs and geolocated imagery used to train a world model) points to a path where consumer AR platforms can seed advanced robotics capabilities while attempting to keep the underlying data footprint bounded.

What to watch next, from an engineering perspective: first, how Niantic Spatial validates navigational reliability in edge cases (dynamic crowds, construction zones, and multistory environments). Second, how the company negotiates data rights as the model scales, especially if additional data sources are introduced. Third, how policymakers and industry partners handle dual use concerns without stifling innovation in AR and autonomous systems. And fourth, the benchmarks by which these geospatial models prove their worth in real world tasks like drone flight planning and last mile delivery beyond the promise of a large geospatial model.