Reproduce “FAVR: Flow-Aware Visual Residual Alignment for Vision-Language-Action Policies” 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
Qixuan Wu, Xiaole Yang, Yuguang Yang, Canyu Chen, Linlin Yang, Haixia Pan, Guodong Guo, Baochang Zhang
FAVR adds video-derived visual guidance to a frozen vision-language-action policy by querying a state with the current action feature and gating visual channels before the frozen action head. It refits the gate at each flow step rather than reusing one gate throughout an action chunk. On five RoboTwin 2.0 tasks, success is 82.0% versus 72.0%, supporting refitting in this setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.