Niantic, the creator of AR-IGR Pokémon Go and Ingress, announced about the development of new AI-model for navigation in the physical world. The basis for the technology was the data collected by millions of players. The company’s blog describes the concept of a large geospatial model (Large Geospatial Model, LGM), inspired by the approach of large language models (Large Language Model, LLM).
Modern language models have learned to analyze and generate the text, and visual ones – to understand and create two -dimensional images. LGM goes further: it allows you to simulate three -dimensional objects tied to specific geographical points, preserving their large -scale characteristics. Unlike ordinary 3D models that create abstract objects, the geospatial model forms a new generation, linking scenes and objects on a global scale. For example, she is able to “guess” how the back of the building looks, based on thousands of images of similar buildings around the world.
Demonstration video ( niantic spatial platform )
Technology is based on the Lightship Visual Positioning System (VPS) system, which provides accurate consolidation of virtual objects in the real world. One of the recent Niantic experiments is the Pokémon PlayGrounds function in Pokémon Go, where users can place Pokemon in certain places for interaction with other players. The function uses data collected from pedestrian prospects.
At the moment, Niantic has collected 10 million scanned locations around the world, of which more than a million are available through the VPS service. A weekly company receives about a million new scans, each of which contains hundreds of images. Data is received from the games and the Scaniverse application for 3D scanning.
Scanned locations on the map ( niantic )
The system is trained in more than 50 million neural networks with a total number of parameters over 150 trillion. Each local model is created for a certain area, however, the integration of all local data allows you to create global models that understand not only specific places, but also the relationship between them.