What's this about?
This paper (40 upvotes) recovers both the 3D shape of an object and the way it moves — where its hinges and joints are — from observing it in a single resting position, without ever watching it move. Existing methods require capturing the object in multiple articulation states.
What exactly is an "articulated object"?
Any object with parts that move relative to each other: a laptop with a hinge, a drawer that slides, a pair of scissors, a door. Reconstructing one means capturing not just its surface geometry but its kinematic structure — which parts rotate around which axes, and how far. Getting the shape is a solved problem; getting the joints has usually required video of the object being opened and closed from several angles.
Why is this worth noticing?
- Requiring motion observation is the practical bottleneck for building interactive digital twins at scale — you can photograph a warehouse of objects, but you can't manipulate every one of them on camera.
- This matters for robotics training, where simulators need thousands of objects a robot can plausibly open, pull, and turn, and hand-authoring each one's joints doesn't scale.