Reproduce “Zero-Shot VLA Adaptation with Retrieved Guidance” 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
Subin Kim, Anh-Quan Pham, Dinesh Jayaraman
RAGu uses retrieved demonstrations as spatial guidance to steer a frozen vision-language-action (VLA) policy, rather than replaying trajectories or adding examples to its context. It warps motions to target scenes and selects references compatible with predicted actions. On five real-world tasks, RAGu raises π0.5-DROID task success from 20% to 80% across 50 rollouts, but depends on approximately matching task demonstrations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.