Reproduce “The Right Future for Action: Learning Action-Relevant Predictive States in 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
Qiwen Gu, Jifan Li, Bingjie Gao, Rui Chen, Jing Tang, Xiangxiang Chu, Junqiao Zhao
ARPS makes the action-facing state explicit by compressing intermediate visual features into 32 learned tokens and training them to predict absolute future representations and their changes. At inference, only the current observation reaches the action expert. On LIBERO-Plus, ARPS scores 87.3% without adaptation versus 48.1% for Fast-WAM, but it does not lead every shift category.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.