Reproduce “Grasp-Oriented Non-Prehensile Manipulation via Learning a Graspability Field” 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
Licheng Zhong, Gim Hee Lee
Researchers at the National University of Singapore developed GOMP, a robotic policy that uses a learned graspability field to perform non-prehensile manipulation, preparing objects for stable grasping without requiring predefined target poses. This system achieved a 75.5% manipulation success rate in simulation and a 62% total success rate on a real robot, with 97% grasp success once manipulation was complete.
Key insights
- Guiding non-prehensile manipulation toward an abstract 'graspable set' via a continuously defined field is more effective and robust than targeting a single, explicit object pose.
- A smooth, differentiable graspability field, enhanced with temporal anchoring, provides a stable and consistent learning signal, preventing oscillations and improving manipulation trajectories.
- The learned graspability distance can serve as an effective, scale-invariant criterion for autonomously transitioning from manipulation to grasping, unifying the pipeline.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.