Reproduce “Fast Generative Grasping via Lie Group-Constrained MeanFlow” 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

S. Talha Bukhari, Yi Wei, Ruiqi Ni, Zachary Kingston, Aniket Bera

GraspMF employs a Lie group-constrained MeanFlow formulation for generative robotic grasping, achieving millisecond-scale inference latency. The approach generates grasps with an 87.40% in-domain success rate and up to 39x faster than iterative state-of-the-art methods, transferring directly to real-world robotic systems.

Key insights

  • Riemannian MeanFlow, when constrained to Lie groups like SO(3) × R^3, can effectively learn to map noisy samples to the contact-feasible grasp manifold in a few steps.
  • Algebraic semigroup consistency on the Lie group flow map is essential for maintaining high grasp quality and preserving implicit contact constraints during few-step generation.
  • Parameterizing the network to directly predict the clean target grasp (endpoint parameterization) combined with SVD projection for rotations significantly enhances model stability and performance.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.