Can we learn how objects deform without knowing the physics?

Simulating how objects bend, stretch, and move over time typically requires knowing the physics equations—too rigid for diverse real-world scenarios. NeuROK learns a latent kinematic space from a large 4D dataset that captures all plausible deformations of an object, then uses a transformer decoder to generate realistic temporal sequences. By working in this learned latent space rather than raw geometry, the method sidesteps the need for predefined physical models, generalizing across object types and scales where traditional system-identification approaches fail.