How Pokémon Go is giving delivery robots an inch-perfect view of the world

2026-03-10 · Source: MIT Technology Review · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Data Science & Analytics · Depth: Intermediate, medium

Summary

Niantic Spatial, an AI spinout from Niantic, is leveraging a massive dataset of 30 billion urban images crowdsourced from "Pokémon Go" and "Ingress" players to develop a highly accurate visual positioning system. This system, trained on images tagged with centimeter-level location metadata from millions of game hotspots, allows devices to pinpoint their location and orientation based on visual input, even in GPS-denied environments. The company has partnered with Coco Robotics, a last-mile delivery robot startup operating in cities like Los Angeles and Helsinki, to integrate this technology into their fleet of approximately 1,000 robots. This collaboration aims to enhance robot navigation precision, ensuring reliable deliveries by overcoming GPS limitations in urban canyons and dense areas, ultimately improving robot autonomy and interaction within human-shared spaces.

Key takeaway

For robotics engineers developing autonomous navigation systems, especially for urban delivery robots, you should investigate integrating visual positioning systems like Niantic Spatial's. This technology offers centimeter-level accuracy in environments where GPS is unreliable, such as dense cityscapes with high-rises and underpasses, significantly improving robot reliability and operational efficiency for last-mile delivery and other human-shared space applications.

Key insights

Crowdsourced AR game data can train highly precise visual positioning systems for robot navigation.

Principles

• Visual positioning excels where GPS fails.
• More camera data improves mapping accuracy.

Method

Niantic Spatial trains its visual positioning model using 30 billion urban images from "Pokémon Go" and "Ingress" players, each tagged with precise location and orientation metadata, enabling centimeter-level localization from visual input.

In practice

• Integrate visual positioning for urban robot navigation.
• Utilize existing image datasets for spatial AI training.

Topics

• Visual Positioning Systems
• World Models
• Robotics Navigation
• Crowdsourced Geospatial Data
• Last-Mile Delivery

Best for: Machine Learning Engineer, Computer Vision Engineer, Robotics Engineer, AI Engineer, AI Product Manager

Related on AIssential

• Towards Next-Generation SLAM: A Survey on 3DGS-SLAM Focusing on Performance, Robustness, and Future Directions — 3D Gaussian Splatting offers a new paradigm for high-fidelity, efficient, and robust SLAM systems.
• Simulate real-world places with Project Genie and Street View — Project Genie now integrates Street View to create imaginative, interactive virtual worlds grounded in real-world locations.
• Turns Out Niantic Needed Your Pokemon Go Photos To Help Delivery Robots Navigate The World — Crowdsourced geospatial data from AR games can significantly enhance autonomous robot navigation accuracy.
• G2IA: Geometry-Guided Instance-Aware Retrieval and Refinement for Cross-Modal Place Recognition — G2IA improves cross-modal place recognition by combining geometry-guided retrieval with instance-aware refinement, overcoming modality gaps and perceptual aliasing.
• Measurement-Calibrated Multi-Camera Fusion for Vision-Based Indoor Localization — Explicitly characterizing single-camera errors can optimize multi-camera fusion for vision-based indoor localization.
• SurroundNEXO: Ego-Centric Metric Bridging for Spatially Consistent Geometry in Autonomous Driving — SurroundNEXO bridges low-overlap multi-camera views for metric depth prediction using ego-centric geometry and sparse LiDAR anchoring.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by MIT Technology Review.