How Pokemon Go's maker is building geospatial models to guide robots

Data generated by millions of Pokemon Go players is now powering Niantic's geospatial AI models that help autonomous robots navigate complex urban environments

When Niantic launched the augmented reality (AR) game Pokemon Go in 2016, players stepping into the streets to catch virtual creatures were doing more than simply playing a game.

 

In the process, they were generating vast amounts of visual data — more than 30 billion images — that would later become a building block for artificial intelligence systems.

 

Millions of users scanning neighbourhoods in search of Pokemon helped create the dataset now used to build a visual positioning system (VPS), a technology capable of guiding autonomous robots to precise locations in dense urban environments.

 

The relevance of that data has become clearer in recent developments.

 

Last year, Niantic underwent a major restructuring, selling key titles — including Pokemon Go, Pikmin Bloom and Monster Hunter Now — to mobile gaming company Scopely for $3.5 billion.

 

At the same time, it spun off its geospatial AI business into a new entity, Niantic Spatial, backed by an initial $250 million in funding.

 

Last week, this geospatial entity partnered with Coco Robotics to enable autonomous delivery of small consignments guided by its spatial AI and VPS infrastructure.

 

Why is a game developer producing geospatial models?

 

AI can process text, code and images with considerable accuracy, but interpreting the physical world through conventional digital maps remains a challenge. For machines to navigate real environments reliably, they require a different kind of mapping system — one that bridges the physical and digital worlds and overcomes the limitations of dominant navigation technologies such as the Global Positioning System (GPS).

 

Over the years, Niantic Spatial has used data in the form of images produced by millions of its users. It collected visual and location data from players of Pokemon Go and Ingress, another phone-based augmented reality game. Using this, the company built a VPS — a technology that determines a device’s location based on visual inputs from its surroundings — according to MIT Technology Review.

 

Why is a geospatial model necessary?

 

Geospatial refers to linking visual information from the physical environment to locations within cities. The company said it is developing a Large Geospatial Model (LGM) built on a proprietary database of more than 30 billion posed images. This will add spatial reasoning capabilities to Large Language Models (LLMs) through “a spatially grounded and semantically rich understanding of physical environments”.

 

Where can this be used?

 

VPS enables robots to reach designated locations accurately and autonomously. For instance, Niantic Spatial serves as the core infrastructure partner for Coco Robotics’ delivery fleet.

 

To guide robots with greater accuracy in dense urban environments, the company’s model has been trained on a large dataset of images. “The promise of last-mile robotics is immense, but the reality of navigating chaotic city streets is one of the hardest engineering challenges,” according to John Hanke, chief executive officer (CEO), Niantic Spatial.

 

Why use VPS when we have GPS?

 

The system is particularly useful in areas where GPS struggles with location precision, such as dense urban neighbourhoods where multiple buildings within a single block can weaken GPS signals. In some cases, there is a complete loss of signal. “The new release of the product delivers best-in-class reliability, flexible data inputs from a wide range of 2D and 3D data sources, and global coverage without pre-scanning,” Niantic Spatial said in a blog post.