Neural Voxel Dynamics: Learning Implicit 3D Physics via Volumetric Feature Advection
Zican Wang, Niloy Mitra · 2026
Read Neural Voxel Dynamics: Learning Implicit 3D Physics via Volumetric Feature Advection by Zican Wang, Niloy Mitra (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-26410.txt.
Description
Zican Wang and Niloy Mitra want video models that respect real physics. Today's video generators often lose track of objects. They also break basic physical rules. Wang and Mitra move the prediction task out of flat images. They place it in a three-dimensional grid of small cubes called voxels.
Their system lifts features from a video model into that voxel grid. Features are compact summaries of what a frame shows. Depth estimated from a single camera view guides the lifting. A learned step then moves those features through space over time. It treats physics as the flow of material states. Training used only video and action signals, with no physics engine labels. The model stayed stable over long runs. It also looked physically plausible on three test collections, including solid objects moving through fluid.
Catalog record
| CREATOR | Zican Wang, Niloy Mitra |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 75 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.26410