Reproduce “Contact-GraspNet: Efficient 6-DoF Grasp Generation in Cluttered Scenes” 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
Martin Sundermeyer, Arsalan Mousavian, Rudolph Triebel, Dieter Fox
Contact-GraspNet, developed by researchers from NVIDIA, DLR, TU Munich, and UW, efficiently generates 6-DoF parallel-jaw grasps directly from depth images of cluttered scenes. The method achieves over 90% grasp success rate on unseen objects in real-world tests and processes scenes in approximately 0.28 seconds.
Key insights
- Treating observed point cloud points as anchor points for grasps reduces the dimensionality of the 6-DoF prediction problem, simplifying learning and naturally handling partial observations.
- An end-to-end PointNet++-based architecture directly processing raw point clouds, without segmentation, improves robustness to occlusion and sensor noise in cluttered environments.
- Ablation studies confirm that using average distance loss for pose estimation, binned grasp width prediction, and training on local regions are critical for achieving high grasp quality and performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.