Reproduce “OnlyAct: Extracting Action Representations from Foundation Models for World Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Runjia Li, Minghao Chen, Junyu Xie, Philip Torr, Andrea Vedaldi, Tomas Jakab
OnlyAct learns action representations from unlabeled video to condition video world models and transfer demonstrated motion across scenes. It adapts a pretrained V-JEPA 2 encoder with contrastive frame-pair examples, while real-action control still needs labeled examples. On unseen RLBench video-generation transfer, it scores 0.686 in transfer direction versus 0.572 for DreamDojo; this is offline, not closed-loop control.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.