Use case | Scenes

Scenes — coming soon.

Start from a phone walkthrough of the environment your robot works in. The same pipeline returns a splat-accurate reconstruction of the exact space, an accurate collider mesh as its physics proxy, and sim-ready objects placed inside for the robot to pick up, push, open, and knock over — all composed into the layout you captured.

The result is a photorealistic, physically interactive digital twin of the environment your robot will actually operate in, built for training, validation, and generating the variations needed to cover edge cases you can’t practically stage.

Request a scene

Request one of our off-the-shelf scenes, scaled up into as many variations as you need.

User input

Describe what you want to createText input

A bedroom with a bed, two nightstands, and a wardrobe in the corner of the room...|

Upload an image of your sceneimage input

Upload an asset

Sim-to-real

The drift that breaks sim-to-real

A photoreal reconstruction that can be looked at but not interacted with

Visual fidelity ≠ physical fidelity

A policy is only as good as the world it observes. For sim2real transfer, a scene has to be faithful on two fronts at once: it has to look like the real place — geometry, materials, lighting — and it has to behave like it under physics. Most pipelines give you one or the other. A photoreal reconstruction you can look at but not touch. Or a scene assembled from nearest-match library props that behave nothing like the objects actually sitting in your workcell.

How it works

01Input

Walk the real site once with your phone or an off-the-shelf 360° camera. A few minutes of video is enough; no LiDAR, no tripod stations.

Step 01

A walkthrough capture of the environment, used as the input

02Build the environment

From a video, we generate the space to match what's in the frame — walls, floor, fixed surfaces, lighting. From a walkthrough, we estimate camera poses and train a 3D Gaussian splat: photoreal RGB from arbitrary viewpoints, with the real lighting, materials, and clutter preserved. Either way you also get the physics layer — an accurate collider mesh derived from the same reconstruction, so what the policy sees and what it collides with agree by construction.

Step 02

The captured space reconstructed, with each surface resolved

03Populate with sim-ready objects

The objects the robot needs to interact with are assets, not backdrop. Each one runs through the same pipeline as every asset we ship: geometry, PBR textures, collision mesh, physics tags (mass, friction, articulation), QA-gated and drop-tested individually. Objects are placed to match the reference or the real layout, then the scene is settled under physics before delivery.

Step 03

The scene broken out into individually manipulable sim-ready assets

04Export and train

You get a USD stage that's gravity-aligned, metric-scale, colliders applied and already settled — for a twin, the splat comes in as a NuRec asset with its collider mesh bound underneath as the physics proxy. It opens directly in NVIDIA Isaac Sim and Isaac Lab, with no alignment, rescaling, or collider authoring on your side.

Step 04

Final render of the environment in simulation

Prefer to start from nothing?

Describe a scene in text and we compose one around your prompt, generating as many distinct scenes from it as you need.

Off-the-shelf scenes

Start from a room we’ve already built

Each scene is live on Hugging Face with its sim-ready objects. Request it scaled up into as many variations as you need — layouts, objects and lighting.