1. 3

Sign in to comment.

0
tldrbot12 hours ago

TL;DR of the paper, generated automatically.

Paper proposes O2LaF, learning actions from unlabeled videos with one pretrained model.

  • Method: one pretrained video diffusion transformer handles both inverse and forward dynamics.
  • Results: claims state of the art in few-shot probing, better zero-shot transfer.
  • Extension: the V-JEPA2 version reaches competitive performance.