Real-world foundation models for physical AI
Robots, agents, and autonomous systems need world models grounded in physics and geometry, not imagination
Establishing ground truth for physical AI
We're building for the 80% of economic activity beyond screens — so delivery robots navigate city streets safely, agents can inspect critical infrastructure, and field teams can operate where GPS is interrupted. Trained on billions of images, terrain, drone, and satellite data, our models power three core capabilities: reconstruct, localize, understand.
Reconstruct
We build geometrically accurate digital twins from smartphones, 360° cameras, drones, satellites, and LiDAR — output as meshes and Gaussian splats. We created SPZ, the open-source file format for Gaussian splats now used across the industry.
Localize
Our Visual Positioning System turns aerial and ground capture devices into a precise positioning and orientation sensor, delivering accurate localization almost anywhere, including where GPS fails — powered by 50 million neural networks and 150 trillion parameters.
Understand
We are building spatial intelligence: an AI interface to the physical world that you can talk to — one that can understand, measure, and reason about every point in 3D. This marks the shift from physical AI that navigates the world to physical AI that interprets it: understanding what it sees, why it matters, and how to respond/what to do next.
Our research is at the frontier of physical AI
We’re building the foundational models that enable AI to operate in the real world, powering precision, automation, and real-time understanding across physical operations.
Inventing the Intelligent Spaces of Tomorrow
We are a global, multidisciplinary team advancing AI, computer vision, and graphics publishing at top conferences and collaborating with leading research institutions worldwide.
Learn more about Inventing the Intelligent Spaces of Tomorrow
Improving Generalization of Scene Coordinate Regression Through Query Pre-Training
ICCV 2025
Localization that works in places it has never seen. ACE-G learns from tens of thousands of environments, so it recognizes where a camera is even under new angles or lighting it wasn't trained on. That means positioning that holds up when conditions change, without recapturing the scene.
Learn more about Improving Generalization of Scene Coordinate Regression Through Query Pre-Training
Feed-Forward View Synthesis with Georeferenced Images
CVPR 2026
3D scenes built from the sky down, not just the ground up. Cross-View Splatter combines street-level photos with satellite imagery to fill the gaps ground photos alone leave behind. That means faster, cheaper 3D coverage of large outdoor spaces without capturing every angle by hand.
Learn more about Feed-Forward View Synthesis with Georeferenced Images
Language-Guided Object Placement in Real 3D Scenes
ICCV 2025
Drop the right object in the right spot using plain language. PlaceIt3D takes a 3D scan and a text instruction, then figures out where an asset actually fits based on space and geometry. That means planning equipment or layouts inside a real scanned environment without manual 3D modeling.
Learn more about Language-Guided Object Placement in Real 3D Scenes
Inventing the Intelligent Spaces of Tomorrow
Improving Generalization of Scene Coordinate Regression Through Query Pre-Training
ICCV 2025
Feed-Forward View Synthesis with Georeferenced Images
CVPR 2026
Language-Guided Object Placement in Real 3D Scenes
ICCV 2025
Trusted by industries that demand real-world intelligence
Robotics
Robots lose position indoors and in GPS-degraded environments. We provide continuous, accurate localization without depending solely on GPS.
Defense
Mission outcomes hinge on ground truth. We deliver 3D spatial intelligence for decision advantage – fuse multi-sensor data, search every area of interest, and localize in DDIL environments.
Oil and Gas
Complex facilities run on outdated models and fragmented data. We deliver a current, precise digital twin of any facility so field operators can visualize, locate, and query every asset in 3D.
What's New
AI for the Real World
LLMs and World Models aren't enough, we're focused on the missing piece
Read more→
Scaniverse for Enterprises
Scaniverse is the front door to our geospatial AI services
Read more→
