Reproduce “Native Action-Prior Learning from Videos for World Action 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
Zhaochong An, Fei Zhang, Menglin Jia, Duncan Frost, Zijian Zhou, Yikai Wang, Xudong Wang, Aditya Patel and 5 more
NAVA-WAM trains a robot policy from videos without action labels by learning transitions inside its action-generating model, then uses labeled robot demonstrations to adapt it to executable control. In a RoboTwin ablation with 25 demonstrations per task, it achieves 72.0% success, versus 60.6% and 65.0% for representation-based and latent-action-based pretraining, showing improved action-label efficiency under that protocol, not label-free control.
Key insights
- Future-state prediction can provide action-relevant supervision to the action-generating module itself, without a separate visual-representation or latent-action learning stage.
- Corrupting the future visual state before it reaches the action stream is intended to prevent the stream from simply copying the prediction target.
- The pretrained action stream is used at inference without generating future video: visual features are cached, and only actions are iteratively denoised.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.