Neural Voxel Dynamics: Learning Volumetric Feature Advection for 3D Physics in V-JEPA Latent Space

Name: Neural Voxel Dynamics: Learning Volumetric Feature Advection for 3D Physics in V-JEPA Latent Space
Uploaded: 2026-10-06T07:41:29.765688+00:00
Duration: 713 s
Description: Neural Voxel Dynamics learns 3D physical dynamics by shifting from 2D image prediction to a 'lifted' 3D volumetric latent space derived from video features. It uses monocular depth priors to create a dense voxel grid and employs Volumetric Feature Advection, a flow-matching approach, to simulate complex interactions like rigid body motion implicitly.

Neural Voxel Dynamics learns 3D physical dynamics by shifting from 2D image prediction to a 'lifted' 3D volumetric latent space derived from video features. It uses monocular depth priors to create a dense voxel grid and employs Volumetric Feature Advection, a flow-matching approach, to simulate complex interactions like rigid body motion implicitly.

Open the full episode