Reproduce “FlowVLA: Enabling World Dynamics Understanding through Pluggable 3D Flow Learning” 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
Junhyeong Go, Mingyu Jung, Woobin Im, Jongbin Ryu
FlowVLA trains an image-based robot policy with 3D scene-motion supervision reconstructed from RGB demonstrations. During training, it predicts selected point displacements and links them to the robot’s end-effector motion; the auxiliary heads and 3D inputs are removed for deployment. On Seer’s LIBERO-Long suite, success rises from 80.4% to 87.8%; Hang Cup success rises from 46.7% to 86.7% in the reported real-world setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.