Reproduce “OpenWAM: An Open Framework for Composable 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
Heng Yu, David D. Yuan, Juze Zhang, Changan Chen, Yao Feng, Michelle Baldonado, Steve Cousins, Li Fei-Fei and 2 more
OpenWAM tests links between video predictions and robot actions, including reuse of action models. It compares generation routes in a shared framework and trains separate dynamics models on demonstrations and counterfactual transitions. After target-task video adaptation, a frozen local-context inverse dynamics model achieves 84.0% mean closed-loop success across four held-out LIBERO-90 simulation targets, versus 47.0% for full-context conditioning.
Key insights
- Several video–action interaction designs achieve strong benchmark performance; no single design dominates every simulated suite.
- For inverse-dynamics transfer, training on counterfactual transitions is far more effective than demonstration-only training, while simply providing broader context does not ensure transfer.
- Counterfactual training helps forward dynamics distinguish among alternative outcomes from the same starting state, not just predict average future appearance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.