Reproduce “BiGraspFormer: End-to-End Bimanual Grasp Transformer” 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

Kangmin Kim, Seunghyeok Back, Geonhyup Lee, Sangbeom Lee, Sangjun Noh, Kyoobin Lee

BiGraspFormer, developed by researchers at Gwangju Institute of Science and Technology, introduces an end-to-end transformer framework for bimanual grasping that directly generates coordinated 12-DoF grasp poses from object point clouds. The framework achieves a 89.67% top 1% success rate in simulation under normal force and operates with an inference time under 0.05 seconds.

Key insights

  • The Single-Guided Bimanual (SGB) strategy effectively decomposes the complex 12-DoF bimanual grasp problem by leveraging features from initial single-arm grasp proposals, thereby simplifying the search space and enhancing coordination.
  • An end-to-end transformer design, with its attention mechanisms, allows for inherent coordination between the two robot arms, addressing limitations of modular pipelines that often produce uncoordinated grasps.
  • The framework can efficiently learn to generate diverse and stable bimanual grasps without explicit object models, improving generalization to novel objects and significantly reducing inference time compared to prior multi-stage approaches.

Summary by alphaXiv.

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