Many 3D objects in virtual worlds stay motionless, even when they should respond to touch. A pillow should compress when pressed, and a door should swing when pulled. To make this possible, an object needs a way to describe all the positions or shapes it can take, called its configuration space, and the physical rules that govern how it moves between them.
The authors developed PRIO, a system that learns configuration spaces from many examples of objects moving and deforming. Given a static 3D shape, PRIO predicts its configuration space in a single step, without additional training. The predicted space is linear, meaning that each possible shape is a weighted sum of a few base shapes, which allows the equations of motion to be solved quickly without running a neural network at every moment.
PRIO works with common 3D formats like meshes and point clouds. In tests, it generated motion faster than existing 4D generation methods and followed user actions more faithfully. This brings interactivity to reconstructed and generated 3D scenes without per-object preparation.
