Reproduce “Future Representation Distillation for Imagination-free 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
Jaehyeong Jo, Sangwon Jang, Sihyun Yu, Seonghyeon Ye, Yekyung Nah, Hojung Jung, Kimin Lee, Sung Ju Hwang
DreamFree avoids generating future videos at deployment by training a visual encoder to predict future-aware features from a video model, then using those features to predict action chunks. It combines feature distillation with video-generation training. In DROID real-world tests, it achieves 64% average task success, versus 56% for DreamZero, across ten familiar task types with new objects or arrangements.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.