Reproduce “RynnWorld-Latent: A Cross-Embodiment World Model Conditioned on Universal Latent Actions” 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
Haoyu Zhao, Yinzhou Tang, Zixuan Wang, Minghao Zhu, Dongchi Huang, Xingyue Zhao, Qin Zhao, Kehan Li and 6 more
RynnWorld-Latent learns video prediction conditioned on latent descriptions of visual change, allowing training from human, robot, and simulation videos with missing or incompatible controls. A frozen RynnLAM encoder supplies descriptors to global attention and frame-aligned conditioning. Synthetic mid-training raised success on LIBERO-Plus from 53.6% to 79.4%, but does not establish physical validity or isolate added data from extra compute.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.