Reproduce “W2A: Distilling World Priors into Action Models for Real-Time Robotic Control” 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
Xiaopeng Zhang, Yongjin Mu, Yueyang Weng, Chengzhi Zhao, Yanjie Li
W2A transfers selected features from a pretrained video model into a robot action policy during training, using LoRA, PCA, and feature alignment. At deployment, it removes the video model and predicts actions from observations and instructions. On LIBERO, it reports 95.9% average success versus 91.6% for scratch-trained Mimic-video, with 55 ms per prediction and about 21× lower latency; these results are limited to tested conditions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.