Niantic Unveils AI Model Using Player-Generated Scans for Advanced Navigation

Niantic announced plans to create an AI model for navigating physical spaces. This model will use scans collected from players of its mobile games like Pokémon Go and users of the Scaniverse app. This approach differs from traditional data sources such as websites and videos, as it leverages data from a mobile gaming app.

Niantic’s Visual Positioning System (VPS) uses a single image from a phone to find its position and orientation. The system relies on a 3D map built from scans of various locations made by users in its games and the Scaniverse app. The model is called a “large geospatial model” (LGM), similar to large language models (LLMs) but focused on processing physical spaces through geolocated images.

Niantic has amassed over 10 million scanned locations globally, with approximately 1 million new scans added weekly. These scans come from a pedestrian perspective, capturing areas that cars and street-view cameras cannot access.

What are the benefits of using mobile game data for geospatial intelligence and navigation?

Interview with Dr. Emily Chang, Geospatial Intelligence Specialist

News Directory 3: Thank you for joining us, Dr. Chang. Niantic recently announced its plans to develop an AI model designed for navigating physical spaces through data collected from users of its mobile games and apps. Can you explain the significance of leveraging mobile gaming data for this purpose?

Dr. Emily Chang: Thank you for having me. Leveraging mobile gaming data represents a groundbreaking shift in how we understand and navigate our physical environment. Traditional sources of data, like street-view imagery or geospatial databases, often lack detail in pedestrian-centric areas. By utilizing scans from mobile gameplay—where users interact closely with their surroundings—Niantic is tapping into a rich source of geolocation data that provides insights that traditional methods cannot match.

News Directory 3: That’s fascinating. Could you elaborate on how Niantic’s Visual Positioning System (VPS) operates and its importance in the context of augmented reality?

Dr. Emily Chang: Absolutely. The VPS relies on analyzing a single image captured by a smartphone to determine its precise position and orientation within a given space. It creates a 3D digital map using user-generated scans, which enables interactive and immersive experiences in augmented reality. This method is crucial for applications that require accurate spatial awareness, especially in urban environments where conventional GPS might falter due to obstructions like tall buildings.

News Directory 3: Niantic has talked about its “large geospatial model” (LGM). How is this similar to large language models (LLMs), and what unique challenges does it face?

Dr. Emily Chang: The comparison to LLMs is intriguing. While LLMs process textual information by understanding context and semantics, the LGM focuses on physical spaces using geolocated images. Unique challenges include ensuring accuracy and comprehensiveness, as the model needs to recognize various angles and perspectives within a constantly changing environment. Achieving this requires robust training data and sophisticated algorithms to handle the intricacies of real-world locations.

News Directory 3: With over 10 million scanned locations and a million new scans added weekly, what are the implications of this data volume on artificial intelligence and geospatial analysis?

Dr. Emily Chang: The sheer volume of data presents both opportunities and challenges. On one hand, having extensive data allows for more nuanced models that can learn from diverse environments and urban layouts. On the other, it demands advanced computational techniques and infrastructure to process and interpret this data effectively. The integration of user-generated scans also leads to more democratized access to spatial data, which can fuel innovation across different sectors.

News Directory 3: You mentioned training over 50 million neural networks for this purpose. What does this look like in practice, and why is it necessary?

Dr. Emily Chang: Each neural network is trained to recognize specific locations or angles, which collectively allows the model to develop a comprehensive understanding of various environments. In practice, this means analyzing thousands of images across multiple perspectives to create detailed representations of physical spaces. The extraordinary number of parameters—over 150 trillion—enables the model to capture complex features and nuances in the environment, essential for precise navigation and interaction.

News Directory 3: what do you envision as the future applications of Niantic’s AI model in everyday life?

Dr. Emily Chang: The potential applications are profound. We might see enhanced navigation systems for pedestrians, smarter urban planning tools, and even applications in tourism, where users receive interactive, context-aware information about their surroundings. Additionally, this technology can significantly enhance immersive experiences in gaming and education, offering users new ways to engage with the world around them. As this technology evolves, it could transform our perception and interaction with physical spaces in unprecedented ways.

News Directory 3: Thank you for your insights, Dr. Chang. It’s exciting to think about the future of navigation and interaction shaped by this innovative approach.

The company trained over 50 million neural networks, each representing a specific location or angle. These networks convert thousands of mapping images into digital models of physical spaces. They contain over 150 trillion parameters, which help the networks recognize and interpret these locations. Niantic plans to integrate this information into a comprehensive model capable of understanding various locations from different angles.