Reproduce “FineGrasp: Towards Robust Grasping for Delicate Objects” 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

Yun Du, Mengao Zhao, Tianwei Lin, Yiwei Jin, Chaodong Huang, Zhizhong Su

FineGrasp, developed by Horizon Robotics, presents a grasp detection method that enhances robotic handling of small and delicate objects in cluttered environments. The system achieves state-of-the-art performance, including a 91% real-world grasp success rate for delicate objects, by introducing architectural improvements and leveraging a novel synthetic dataset.

Key insights

  • The inability of robots to grasp delicate objects stems from a scale imbalance during training, where larger objects dominate the learning process.
  • Applying graspability label normalization at the instance level rather than globally is crucial for ensuring effective supervision and learning for small objects.
  • Leveraging geometric priors like surface normals and dynamically aggregating features across multiple scales significantly improves the accuracy of grasp pose and orientation estimation.

Summary by alphaXiv.

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