The Daily Aus The AI model built on billions of Pokémon GO images
Mar 22, 2026
A discussion about using billions of player photos to build AI mapping and visual positioning tech. They explore how neural networks learn from in-game scans and what that enables for delivery robots on city streets. The conversation raises questions about consent, terms of service and the ethics of repurposing gameplay data.
AI Snips
Chapters
Transcript
Episode notes
Game Mechanics Turned Into A Mapping Goldmine
- Niantic used Pokemon Go's AR mechanics to collect vast real-world images that became valuable mapping data.
- Players worldwide took photos of landmarks and streets, producing roughly 30 billion images used to model locations in multiple conditions.
2016 Pokemon Go Craze Filled Streets And Caused Injuries
- Sam Koslowski recalled the 2016 Pokemon Go craze filling Sydney streets and causing many injuries from distracted walking and driving.
- He cited a 2016 US National Safety Council note reporting about 11,000 injuries in six months tied to the game.
Player Photos Trained A Street-Level AI Model
- Niantic trained about 50 million neural networks on player-contributed scans to build a large geospatial model powering a visual positioning system.
- The model fills image gaps to infer unseen views like the back of statues, enabling precise street-level localization where GPS is weak.