Reproduce “POKE: Probe Offline, Keep Efficiently 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
Jaehyuk Lee, Yeajin Lee, Suhyun Bae, Donghun Lee
POKE estimates visual-token importance from predicted effects on a robot’s actions, then trains a small student to score observations. It derives labels offline using first-order action sensitivity rather than attention. In simulation, retaining 32 of 512 tokens yields 95.15% average LIBERO success on OpenVLA-OFT, versus 89.25% for VLA-Pruner, at similar reported speedups; real-robot performance remains untested.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.