Reproduce “FunGrasp: Functional Grasping for Diverse Dexterous Hands” 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

Linyi Huang, Hui Zhang, Zijian Wu, Sammy Christen, Jie Song

FunGrasp transfers human grasp guidance to robot hands by estimating a pose from an RGB-D image, retargeting its posture and contact points, and training a feedback policy in simulation for hardware. On the Allegro hand, it succeeds on 74% of grasps across 20 objects unseen during policy training, using three references per object; it requires known object meshes.

Key insights

  • A multi-component loss function for human-to-robot grasp retargeting effectively translates human-like functional postures and precise contact points to diverse robot hands.
  • Privileged learning with an LSTM-based encoder can robustly reconstruct unobservable contact information (forces, states) from proprioception, enabling sim-to-real transfer without real-world contact sensors.
  • System identification and gravity compensation are critical for bridging the sim-to-real gap, significantly improving real-world grasping success rates for high-DoF dexterous hands.

Summary by alphaXiv.

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