Reproduce “Dune: A Zeroth-Order Optimizer with Time-Conditioned Historical Error Regression for VLA Test-Time 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
Chengyang Dong, Ze Chen, Kai Ye, Chutian Wang, Heng Xu, Wei Mo, Maomao Li, Jiayan Xu and 6 more
Dune adapts pretrained vision-language-action robot policies during deployment using historical regression without global backpropagation. It predicts layer-specific error signals from output residuals and flow-matching stages, then blends them with current zeroth-order perturbation estimates. On LIBERO, its mean success rate is 61.4% versus SignSGD’s 52.0% after 24,000 steps; AdamW reaches 78.3%, and Dune’s predictor adds no forward passes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.