Reproduce “Think Like a World Model, Act Like a VLA: Distilling World-Model Representations into Compact Robot Policies” 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
Trung Dao, Sankalp Yamsani, Jaden Park, Joohyung Kim, Yong Jae Lee
THAW-VLA trains a compact vision-language-action policy using cached visual features from a frozen world model. During training, a projector aligns student features with teacher features alongside action learning from demonstrations; the teacher and projector are removed before deployment. The 0.8B student improves success on LIBERO and RoboCasa-GR1, including 97.9% versus 95.3% on LIBERO, but this does not establish causal understanding.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.