NeuROK: Generative 4D Neural Object Kinematics

Researchers introduced NeuROK, a transformer-based framework for learning latent object kinematics from 4D data.

The model uses a transformer encoder-decoder to generate plausible object deformations over time.

It replaces hand-crafted dynamics with data-driven latent kinematics, aiming to make 4D animation and physical simulation more scalable.