Reproduce “Self-Distilling Future Video Representations 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

Huiwon Jang, Suhyeok Jang, Hyeonmo Kang, Seungku Kim, Jinwoo Shin, Sihyun Yu

FRED trains a video model to produce single-pass future features that guide robot actions. It aligns noise-generated student features with an exponential-moving-average teacher’s features of noisy observed futures, using recent observation history. On DROID-trained policies tested in a new environment, its 14B backbone reaches an 81.7% long-horizon task-progress rate versus 61.7% for DreamZero. With human-video co-training, FRED reports a 50.0% task score on transfer tasks, allowing partial credit.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.