Reproduce “SAVLA: Symmetry-Aware Vision-Language-Action Models for Robotic 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
Junle Li, Weixian Waylon Li, Fuxiang Wu, Fusheng Hao, Fengxiang He
SAVLA augments a frozen vision-language-action backbone with learned image canonicalization and an SO(3)-equivariant flow-matching action head, making robot commands transform consistently with scene rotations. On LIBERO, it raises average success from 86.5% to 91.6%, improves rotated-scene performance substantially, and retains an advantage with fewer demonstrations and on a small real-robot evaluation.
Key insights
- Canonicalized visual conditioning and an equivariant action head are complementary: removing canonicalization reduced LIBERO-Goal performance from 90.4% to 46.0%, while adding canonicalization alone reached 63.5%.
- The largest remaining limitation is visual warping rather than rotation-angle prediction; a true canonical view achieved 97.3% versus 90.4% with the learned estimate.
- Encoding rotational structure is especially useful in data-limited settings, where the model can avoid learning separate action patterns for every scene orientation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.