Reproduce “Failure Prediction and Classification in Vision-Language-Action Models Via Calibrated Latent Probing” 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
Michal P. Podolinsky, Yunhao Yang, Daniel Little, Hong Lu, Rohan Siva, Olin Goudey, Neel P. Bhatt, Matthias Scheutz and 1 more
VLA-TUG monitors a vision-language-action model’s internal representations with a lightweight probe trained on pre-failure signals to predict failure scores and modes. Learn-then-Test calibrates alarms against false-alarm and early-detection targets when calibration and test episodes are exchangeable. On unseen LIBERO-X held-out-skill tasks it scores 39.0% versus 30.1% for SAFE-LSTM, but results vary across settings and shift weakens guarantees.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.