Reproduce “ForceDelta-VLA: Distilling Force-Conditioned Action Corrections for Contact-Rich Manipulation” 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
Ju Dong, Yu Fu, Jian Chen, Yimeng Liu, Haocheng Zhao, Lei Zhang, Kaixin Bai, Liding Zhang and 4 more
ForceDelta-VLA separates reusable task motion from fast contact-driven pose corrections. It trains corrections from paired predictions of a frozen force-conditioned teacher, then applies them using recent force history, robot state, and a cached reference. Across nine real-robot tasks, it achieves mean success of 82.2%, versus 54.4% for ForceVLA and 70.6% for direct teacher execution. Its corrections do not replace task-level replanning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.