Embodied manipulation
We built a robotic-arm digital twin from CAD alone — no robot, no camera.
Atelier can now reason about, simulate and manipulate physical objects through robotic actuators operating inside a shared spatial context. The first of those actuators — a 6-DoF arm — orbits live in your browser below, rendered as a 3D Gaussian splat. There was no physical robot, and nothing was ever photographed.
SO-101 robotic arm — interactive 3D Gaussian splat, running in your browser.
Why this is unusual
3D Gaussian Splatting normally needs a real object and dozens of photographs taken from every angle. Here we had neither — no arm in the room, no camera. We started from a CAD model of an open-source SO-101 arm and ended with a photoreal-pipeline spatial asset you can orbit and zoom in a browser tab.
How it was made
The trick is synthetic capture: render the CAD model from many virtual viewpoints with known camera poses, then train a Gaussian splat from those views. Because the poses are exact, the reconstruction is clean — and the whole thing runs on cloud GPU in hours, not weeks.
Why it matters for Physical AI
You can stand up an interactive spatial twin of a machine before it is ever on the floor — to plan a workcell, brief a team, or rehearse a deployment. That is core to how we think about operational spatial runtimes: bridging design, simulation and the real world so the loop closes faster.
It is one small piece of Real2Sim2Real — and a reminder that the gap between a CAD file and a usable spatial twin is now measured in hours.