Reproduce “Structured Latent Dynamics for Long-Horizon Embodied 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
Kevin Chen, Yu-Wei Zhan, Haoran Liang, Sen Cui, Changshui Zhang
HWM replaces a world-model latent transition with learned dynamics: it maps hidden states to coordinates and momenta, evolves them with structured updates, then maps them back for visual prediction. In controlled simulation, 30-step jerk RMSE falls from 14.135 to 11.401 after 10-step training. RoboCasa success averages 68.5% versus 66.7% across 24 tasks and three seeds, with task-level variation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.