Skip to content

Video Prediction Policy

A robot policy that predicts future video frames

Visit official site

What Video Prediction Policy is

The model is described as a robot policy built on video diffusion models that predicts future image sequences.

What you can do with it

  • Predict future frames of a manipulation task
  • Combine diverse human and robot datasets
  • Explore tasks with a unified video training objective

Application scenarios

  • Test on the CALVIN and MetaWorld benchmarks
  • Run Panda arm and XHand manipulation
  • Place, stack, press and hand over objects

Pros & cons

✓ What we like

  • The material lists CALVIN, MetaWorld, Panda arm and XHand tests
  • Multiple manipulation tasks such as placing, stacking and pressing are named

! What to watch out for

  • The approach depends on video diffusion models
  • Training uses combined human and robot datasets

FAQ

What does it predict?

The material states it predicts future image sequences using predictive visual representations.

Which results are reported?

The material reports a 28.1% relative improvement on Calvin ABC-D and a 28.8% higher success rate.

Last reviewed: 2026-11-15

More AI video generation tools

View all →

How we review