NeuROK: Generative 4D Neural Object Kinematics

Researchers introduced NeuROK, a data-driven framework for simulating dynamic object behavior in 4D.

It learns a latent kinematic state space and uses a transformer encoder-decoder trained on a large curated 4D dataset.

The model can generate plausible object deformations without predefined physical rules.

This could broaden neural scene modeling, robotics, and graphics by reducing reliance on hand-designed physics models.