Reproduce “PRISM: Primitive Routing via In-context Skill Mixing for Retentive VLA Adaptation” 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
Jiachen Tao, Gengyu Zhang, Haoxuan Wang, Bing Liu, Yan Yan
PRISM addresses retention and skill switching in vision-language-action robot policies. It gives each behavior a low-rank adapter and uses a router to mix them based on visual context and instruction. Composite training freezes skill adapters but updates residuals and shared modules. RoboCasa365 success is 56.43% atomic and 20.00% composite, versus 17.86% and 17.50% for a two-stage baseline; one real-robot rollout succeeds, not establishing a success rate.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.