Reproduce “HINT-WAM: Hindsight-Informed Training for 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
Phuong Tuan Dat, Qi Li, Xinchao Wang
HINT-WAM adapts robot policies by using demonstration futures to train a teacher that guides the student along action-denoising states generated by its own sampler. It uses demonstration observations, without rewards or new environment rollouts during training. On LIBERO-90, it reaches 85.18% success, versus 78.88% for SFT, a 6.30-point gain across 90 tasks; evaluation uses held-out initial configurations, not unseen tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.