Niantic plans a “Large Geospatial Model” trained on Pokémon Go player data
Niantic’s plan to build a “Large Geospatial Model” from Pokémon Go and Ingress player scans is prompting both excitement over new AR, robotics, and navigation capabilities and concern over how that data was gathered and will be monetized. Commenters probe the technical feasibility and limits of visual positioning at global scale, compare it to existing mapping efforts like Google Street View, and debate whether crowdsourced geospatial data and derived models should be treated as private corporate assets or part of a shared commons. Many also criticize opaque consent and compensation models, especially given that much of the training data was contributed by children and casual players via gamified tasks.
Scope and status of Niantic’s “Large Geospatial Model” (LGM)
- Several commenters stress this is mostly a plan/vision document, not proof a large unified model has been trained.
- Confusion is attributed to an editorialized HN title using past tense (“trained”) versus Niantic’s more aspirational language.
- Some see the post as positioning Niantic as an “AI company” to investors, leveraging its dataset more than demonstrated model capabilities.
How the tech works and what’s new
- Described as a large-scale evolution of existing Visual Positioning Systems (VPS): photo → camera pose, using dense 3D point clouds built from many scans.
- Key challenges mentioned: scale beyond room-sized point clouds, keeping localization robust to lighting/weather, and avoiding “hallucinated” wrong locations under uncertainty.
- LGM is framed as replacing explicit point-cloud databases plus feature matching with a single learned model; some are skeptical it will scale better.
Data sources, scale, and quality
- Niantic reportedly has ~10M scanned locations, ~1M “activated” for VPS, and ~1M new scans/week; people debate whether that implies multiple scans per location.
- Much data comes from deliberate “scan this POI” tasks in Pokémon Go/Ingress, not from casual AR battle use.
- Commenters note data quality issues: outdated or removed POIs, obstructed views, poor GPS in dense cities, night-time scans rejected, and users often scanning sidewalks or hands instead of targets.
Privacy, consent, and intelligence ties
- Strong privacy concerns: location + imagery can reveal habits, events, seasons, and trajectories; risk of deanonymization at sparsely visited places.
- Debate over whether models will encode only pose or richer “cultural” signals, given Niantic’s mention of broader applications.
- Multiple references to Niantic’s historic ties to In-Q-Tel/CIA and possible interest from intelligence agencies; others question how much extra value this adds beyond existing data holdings.
Ownership, fairness, and “free labor” debates
- Many players feel “tricked” into unpaid data collection; others counter that they received value via a free game, items, and exercise.
- Large argument over whether contributors to crowdsourced datasets deserve access to resulting models or datasets.
- Some propose geospatial data should be treated as a public commons; others argue Niantic added significant value by organizing and processing it.
- Questions raised about informed consent (especially for children) and whether ToS-based consent is ethically or legally adequate (GDPR mentioned).
Potential applications and risks
- Positive uses cited: AR navigation and HUDs, robotics and autonomous vehicles, indoor/outdoor relocalization, search-and-rescue, VR/AR world-building, procedural city/planet generation, GeoGuessr-like tasks.
- Darker possibilities: mass OSINT/geolocation of images, surveillance, military/intelligence targeting, AI-guided weapons.
- Some see it as a natural extension of prior work (Photosynth, NeRF-like models); others are impressed by the decade-long data-collection vision.