Reproduce “Gaussian Distillation of Flow 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
Youssef Saied, Ziyad Sheebaelhamd, Claire Vernade, François Fleuret
GAR replaces repeated flow evaluations with readouts trained on a frozen pretrained robot policy. The readouts predict continuous-action means and covariances plus probabilities for binary commands, allowing actions to be sampled directly. In simulation, diagonal GAR reached 57.64% success on 50-task RoboCasa365, versus 58.24% for the Xiaomi teacher using five evaluations. One policy call was 1.91× faster, not establishing practical equivalence.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.