Reproduce “Dynamics-Informed Action Learning for Goal-Conditioned Latent Planning” 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
Yinuo Jiang, Guangzhi Wang, Yiran Wang, Junfu Pu, Cheng Cheng, Honghao Chen, Cheng Cheng, Ying Shan
DIAL trains a visual world model on its own predicted states and learns inverse dynamics to initialize cross-entropy method search with a goal-directed action sequence. This initialization changes the search mean, while sampling and refinement remain. At a 50-step goal offset, DIAL reaches 78.0% macro success versus LeWM’s 63.0% under matched evaluation conditions; each plan spans 25 primitive steps, so reaching the goal requires replanning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.