Reproduce “GAP-RL: Grasps As Points for RL Towards Dynamic Object Grasping” on OpenResearch

About this paper

Pengwei Xie, Siang Chen, Qianrun Chen, Wei Tang, Dingchang Hu, Yixiang Dai, Rui Chen, Guijin Wang

GAP-RL introduces a reinforcement learning framework for dynamic object grasping that transforms 6D grasp poses into a gripper-agnostic Gaussian point representation, processed by a hierarchical GraspGroupNet. This system achieves an 18-34% higher success rate on dynamic grasping benchmarks with novel objects and motions, while also demonstrating improved sim-to-real transferability compared to baseline methods.

Key insights

• Representing 6D grasp poses as abstract, randomly generated Gaussian points, rather than raw object features or fixed keypoints, significantly enhances generalization capabilities to novel objects and diverse dynamic motions.
• The hierarchical GraspGroupNet effectively captures both intra-grasp details (e.g., orientation) and inter-grasp relationships among multiple grasp candidates, improving the robustness of the learned policy.
• Operating on abstracted grasp features rather than raw visual data intrinsically reduces the sim-to-real domain gap, making policies more robust to real-world sensor noise and facilitating seamless deployment.

Summary by alphaXiv.