Rights-cleared data from real spaces
We capture diverse physical spaces in photoreal 3D and turn them into environments robots and world models can train in
Who it's for
Coverage
Homes inside and out, kitchens, living rooms, hotel rooms and suites, commercial floors and construction sites, captured across 25 cities in 8 countries and growing 20% a month
A real space in the delivery format, splat, lidar, collision, poses and USD
Five layers, one coordinate frame, so a pixel in any image resolves to a point in the room. Hover a layer to see the files
Fisheye and undistorted frames from three cameras, operator and rig masked out
camera_0..2/fisheye + perspective framesmasks/human and device masks, per frameWhere every frame was taken from, at 10 Hz, with IMU at 200 Hz
poses.csvcamera-to-world, TUM formathigh_frequency_poses.csvIMU, 200 Hzcolmap/cameras.bin · images.bin · points3D.binPhotoreal splat for perception, lidar and collision proxy for contact, navmesh for the floor
splat.plyGaussian splatcloud.laslidar point cloud, metriccollision.objdecimated proxynavmesh.objwalkable floorEvery object named and boxed, in 2D and as an oriented 3D box
instances.jsonclass · instance id · source framebbox2d · bbox3dcentre, size, quaternionLoads into Isaac Sim as delivered, splat for what the robot sees, proxy for what it touches
scene.usdzUSD scene, proxy split from splatcamera.jsonintrinsics + extrinsicsThe same scene serves a simulation team and a world-model team, they just read different layers
The scene is a static, photoreal environment: splat for what the robot sees, collision proxy for what it touches, navmesh for where it can go. Objects to manipulate are added as separate assets on top, so you control what is graspable
World-model and perception teams take the registered still frames with camera poses, depth rendered from the lidar cloud, and re-renders of the splat along any camera path you choose
Running in simulation today
Research areas
SLAM, exploration and path planning through rooms nobody staged
Approach, position and reach in real kitchens, suites and living rooms
Predict a room from a viewpoint the camera never visited
Train and benchmark against lidar-backed metric geometry
Train in a photoreal twin, test in the room it was captured from
Supervise monocular depth and camera pose from registered captures
Objects in cluttered, lived-in interiors rather than staged sets
Tie instructions to objects that have a real position and size
Multi-room tasks across a whole property, inside and out
Scroll through the five layers a training pipeline reads. Perception reads the splat and the point cloud behind it; physics reads the collision proxy; evaluation reads the camera track and the labelled instances
Schematic stand-in drawn from the real layer stack. Real captures load the same five layers, ask for the sample pack to run them yourself
Raw fisheye and undistorted perspective frames from three cameras, with human and device masks so the operator never ends up in the scene
COLMAP sparse reconstruction for both image sets. Device trajectory at 10 Hz, IMU at 200 Hz. Delivered camera-to-world in TUM format
Gaussian splat PLY for photoreal rendering, lidar point cloud for structure, a decimated collision mesh and a per-scene navigation mesh
Per-frame instance labels: class, instance id, source image, 2D box, and a 3D oriented box with centre, size and quaternion, fused from imagery, poses and 3D coordinates
A USD scene with the collision proxy separated from the splat, plus camera intrinsics and extrinsics as JSON. Splats carry perception; proxies carry contact
Start with the sample pack, the shortest path to knowing whether this data fits your stack
Twenty-five scenes in the delivery format. You pick which files you need, so we do not send you anything you will not open
Bulk access to the library by scene type, with the full delivery schema: imagery, poses, splat, collision, semantics
Environments you need and we don't have yet. We run the capture; you get the same delivery format
For labs publishing on reconstruction, world models or sim-to-real. Scenes plus poses, on citation terms
Roughly 1,500 scenes released publicly on citation terms, a successor to what ScanNet gave the field. Downloadable, no email
Query the library by room type, geography and modality, and pull scenes straight into your pipeline
We turn what comes back from a capture into assets a lab can load the same day
One to two days on site with a PortalCam
Images, trajectory and lidar solved into one frame
Gaussian splat for appearance, lidar surface for structure
Collision proxies at the level physics needs, plus a navmesh
Instances fused from images, poses and 3D coordinates
USD scene, splat for perception, proxies for contact
Every capture adds a rights-cleared scene to the library. 8,000 today, growing 20% month over month, with an API on the way
Start here