Reproduce “MaskVLA: Visual Masking Against Trajectory Overfitting of Vision-Language-Action Model” 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
Yuxuan Jiang, Jiaying Huang, Ge Wang, Shenhao Yan, Jiahao Yang, Chengsi Yao, Qi Liu, Qing Zhao and 4 more
MaskVLA adapts vision-language-action policies by randomly zeroing the entire overhead-camera image during 10% of training timesteps, while retaining wrist views, language, and robot state; complete inputs return at inference. It reaches 75.0% mean success across four hard RoboTwin tasks, versus 55.0% for OpenVLA-OFT. Real-world results also improve across four ALOHA tasks, but these results do not establish broad deployment generalization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.