Back to Project
V

Video-Point Model

Project Oct 21, 2026

A Cosmos3-based research system that jointly forecasts multi-view videos, persistent 3D point trajectories and optionally robot action chunks.

Visit primary link 3 sources attached
VPM method diagram connects multi-view history, instructions and point queries to future videos, actions and 3D point tracks.
The authors' architecture diagram explains joint video, action and point forecasting. It is an explanatory figure, not independent evidence of robot performance. Source: Byungwoo Jeon and Video-Point Model authors
multimodalverification

What happened lately

Video-Point Model timeline

Video-Point Model (VPM) learns future visual appearance together with the motion of persistent points in 3D. Its authors build on Cosmos3-Nano, adding multi-view camera conditioning and a trajectory head. An action-enabled extension also predicts robot action chunks.

The public package includes inference and training code, a VPM-DROID checkpoint, 10,000 simulated VPM-Data clips and the 1,200-clip VP-WorldBench. Code and VPM weights use CC BY-NC 4.0 for noncommercial use; Cosmos-derived portions and individual datasets carry their respective terms. Benchmark improvements are author-reported preprint results. The larger million-sample dataset and Cosmos3-Super version are planned releases, not part of the currently available package.