In 2016, people began venturing outside to catch and train Pokémon using their smartphones. Pokémon Go was the game that popularized the concept of Augmented Reality (AR) on a large scale. Although the initial excitement has faded, the game still enjoys some popularity today.
What most players of Pokémon Go didn’t realize was that they were helping the developer, Niantic, build a Large Geospatial Model with their Pokémon snapshots. Niantic recently introduced this concept in a blog post, stating that “spatial intelligence is the next stage in the development of AI models.”
Large Geospatial Models (LGM) are to 3D models what Large Language Models (LLM) like ChatGPT are to text and language. Just as LLMs need to be fed enormous amounts of text to produce coherent language, LGMs require millions of images to create three-dimensional views. These models go beyond simple 3D models, capturing “3D entities” that are anchored to specific geographical locations and possess metric quality.
A geospatial AI model understands how a captured image relates to “millions of other scenes around the world.” It has a form of spatial intelligence, learning from past observations and applying that knowledge to new locations, even if they are only partially observed. This capability aims to solve the problem computers face when trying to visualize unseen perspectives, such as the back of a building when only the front is visible. LGMs are expected to be useful in future applications for video games and maps.
So where does Niantic get the training data for its LGM? The data comes from its AR games like Pokémon Go and Ingress, as well as from the app Scaniverse, which allows users to scan and share objects in 3D. For the past five years, Niantic has been working on a Visual Positioning System (VPS). This system can recognize images based on a 3D map of their environment and is already implemented in some Niantic games. Scanning an image within these games can earn players in-game rewards.
In Pokémon Go, VPS is integrated into a feature called Pokémon Playgrounds, which is currently being tested. Players can leave a Pokémon at real-world locations, where other users can find and photograph them. Niantic openly states its ambition to map the entire world, referring to “experiences on a planetary scale.”
These ambitions have sparked concerns and criticism. One user expressed fears that “Pokémon Go is being used to build an AI system that will almost inevitably be used by automated weapon systems to kill people.” Despite these concerns, Niantic continues to pursue its vision of creating a comprehensive geospatial model.
This development in AI and geospatial technology represents a significant step forward in how digital environments are mapped and interacted with. As technology evolves, it will be fascinating to see how these models are applied in various industries and what new possibilities they unlock.