Introducing OVER Research

Over the past years we have built one of the world’s largest collections of real-world 3D data: hundreds of thousands of indoor and outdoor environments, captured at metric scale and still growing through the OVER network. Today we are launching OVER Research, a division dedicated to Spatial Intelligence and Physical AI, and to putting that data to work. It follows directly from our Physical AI Thesis, where we argued that real-world spatial data, more than compute, is becoming the last scarce asset for the next generation of AI systems. The thesis made the argument; OVER Research exists to test it, with experiments, models and published results.

Our work covers:

→ Large Geospatial Models → Machine perception and spatial reasoning → Visual localization and real-time navigation → 3D reconstruction and scene understanding → Simulation and training environments for robotics

Why 3D maps are training data

A mapped environment contains much more than a visual reconstruction. Each capture combines multi-view images, depth, camera poses, geometry and metric scale: several aligned representations of the same physical place. That alignment is what allows a machine to learn where objects and surfaces sit, estimate distances, recognize navigable space and keep track of its own position while moving through a scene.

Data like this cannot be scraped. Language models learned from decades of text that already existed on the internet, but embodied AI has to learn from places that someone first has to capture. Reality cannot be crawled, it has to be reconstructed. Simulation only covers part of the gap, because real spaces are full of exactly what simulators struggle to reproduce: changing light, reflections, occlusions, irregular geometry, people moving through the frame, and the visual diversity that separates one location from the next.

This is where our dataset stops being a product and becomes a research instrument. Working directly on the OVER 3D Maps Dataset, we can study how existing models behave when trained and evaluated on real spatial data, how much relatively small subsets improve their performance, and how those gains scale as more scenes are added. We can also turn mapped environments into simulation-ready spaces where robots and embodied agents learn to navigate and act before operating in the real world. The same questions sit underneath Visual Positioning Systems, digital twins, robotics and Large Geospatial Models, which is why we treat them as research problems and not only as engineering ones.

Research in the open

OVER Research will publish technical reports, papers, open datasets, models and code through the OVER Research platform and our releases on Hugging Face and GitHub. We are opening collaborations with universities and teams working on robotics, machine perception, spatial computing and embodied AI, and what we learn will feed back into how the OVER network reconstructs and processes environments for developers and enterprise partners.

The first projects arrive in the coming weeks: dedicated releases showing how our data improves existing models, supports real-time spatial understanding and trains intelligent systems inside real 3D environments.

Explore OVER Research ↓ https://www.overthereality.ai/research

Read our Physical AI Thesis ↓ https://www.overthereality.ai/the-last-scarce-asset