Reproduce “Relational Spatial Alignment for Vision-Language-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
Chaoran Zhu, Hengyi Wang, Chenglin Cui, Yazhe Wan, Taein Kwon, Changjae Oh
RIGID transfers spatial relationships from a 3D teacher to a vision-language-action model by aligning attention between image patches rather than matching feature vectors. The reported analysis finds that direct feature alignment reduces measured representation rank, while relational alignment retains it. On LIBERO-Plus, OpenVLA-based RIGID reports 75.0% average success across four task suites, versus 70.3% for Spatial Forcing, but scores lower under language perturbation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.