Pokémon Go data fuels real world navigation AI

Pokémon Go scans power an AI that could steer delivery robots and drones.

A decade after the mobile game mania, a new AI company is using billions of real world images captured by millions of players to train navigation systems for delivery robots and potentially military drones. Niantic Spatial, a spinoff born in May 2025 after Niantic sold its licensed titles like Pokémon Go to the Saudi-backed Scopely, is building what it calls a large geospatial model. The concept is simple in theory and big in practice: turn crowdsourced scans of streets, statues, fountains, and other visible landmarks into a 3D model of the real world that machines can interpret and navigate.

The crucial hinge is data provenance. A Niantic Spatial spokesperson told Ars that ground scans were one component used to train the company’s real world foundation models, AI systems designed to learn to recognize and interpret physical spaces. The claim is that the models are the product of that training, not a copy of or direct access to the underlying scans, which were public points of interest such as statues and fountains. In other words, the raw scans aren’t mirrored back to users or stored as a personal dataset; the value comes from learning to generalize from broad, public visuals across neighborhoods.

What makes this arrangement notable is the scale and the potential reach. The team reports that the training mix draws on billions of real world images and on data captured by users of Scaniverse, Niantic’s companion app. The goal is a robust 3D representation of physical spaces that can power navigation for autonomous delivery devices, robots that must understand sidewalks, crossings, and landmarks well enough to operate in bustling real world environments. It’s a quiet intensification of a trend: crowdsourced perception data fueling models that aim to bridge perception and planning in the real world.

For practitioners, the case highlights two hard engineering constraints. First, data quality and coverage are uneven. Crowdsourced imagery is strong where players congregate or travel, but gaps remain in many urban and rural areas. The result is a model whose accuracy varies by geography, which has obvious implications for safety and reliability in autonomous systems. Second, there is dual use risk. A geospatial model trained on public spaces can be adapted for civilian delivery use, but the same foundation could influence military or surveillance applications if the model becomes a widely deployed navigation translator for machines in the field. That creates a governance problem: who gets to deploy, when, and under what constraints?

Beyond these constraints, the Niantic Spatial approach spotlights a broader industry shift. The promise of real world foundation models is to let robots and drones reason about unfamiliar streets using a shared, learned representation rather than reinventing perception from scratch. But the practical path to reliability will hinge on data provenance, model safety, and measurable improvements in robustness under diverse conditions, things that benchmarks and public disclosures have yet to fully spell out.

Looking ahead, engineers will want to watch how Niantic Spatial validates its model performance across different neighborhoods, weather conditions, and times of day. They’ll also watch for governance guardrails: how Niantic and Scopely handle user consent, data minimization, and any regulatory scrutiny tied to dual-use capabilities. If the model proves capable and scalable, it could redefine how autonomous systems perceive the world from crowdsourced visuals, channeling a popular game's data into a new class of industrial navigation tools, and raising questions about who benefits, who bears risk, and how the data trail is secured.