Reproduce “FrameSkip: Learning from Fewer but More Informative Frames in VLA Training” 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
Bin Yu, Shijie Lian, Xiaopeng Lin, Zhaolong Shen, Yuliang Wei, Changti Wu, Hang Yuan, Haishan Liu and 3 more
FrameSkip introduces a data-layer framework for Vision-Language-Action (VLA) training that selects more informative frames from dense robot demonstrations, addressing temporal supervision imbalance. This approach increased the macro-average success rate from 66.50% to 76.15% across three benchmarks while utilizing only 20% of the original frames.
Key insights
- Not all frames in dense robot demonstrations contribute equally to VLA policy learning; critical interaction moments are significantly more informative.
- Reallocating temporal supervision by prioritizing informative frames can substantially improve VLA policy performance and robustness without modifying model architecture or objectives.
- A synergistic combination of action dynamics, visual-action changes, and task progress priors effectively identifies crucial frames for robotic manipulation tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.