Reproduce “NEEDLEWORK: Offline Rewriting of Robot Data with Verified Local Stitches” 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
Juntao Ren, Yifan Hou, Shuran Song
NEEDLE improves policies by adding verified action paths between moments in demonstrations while retaining original actions. It proposes paths from RGB images and proprioception, then uses a learned verifier to select prefixes likely to reach targets. Across two real-robot tasks, gains average 21 percentage points; on near-optimal DexMimicGen tasks, gains over DP are at most four points, with Threading tied.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.