Reproduce “OneWorld: Learning Consistent Physics Across Actions in 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
Ke He, Yichen Ding, Bin Yang
OneWorld evaluates whether action-conditioned futures from one scene can share a physical explanation, and uses that compatibility signal in training and sampling. In Push Cube, Push Rope, and Pour Water simulations, it reduces the mean shared-world gap from 0.02382 to 0.00374, an 84.3% reduction versus ACWM-DiT, while improving M-MSE. The results test simulator-specific consistency, not real-world physics or planning success.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.