Real-Time 3-DoF Robotic Grasp Detection via Keypoint Regression

Abstract

Although vision-based grasping methods demonstrate strong perceptual capabilities in robotic grasping tasks, they often require direct regression of the grasp angle. Such regression is susceptible to the angle’s periodicity and discontinuity, leading to unstable model training. To address this issue, this paper proposes a 3-degree-of-freedom grasping method based on keypoint regression. The method directly predicts two grasp contact points as keypoints and computes the grasp center, grasp orientation, and gripper opening width from their geometric relationship, thereby avoiding angle regression. Experimental results show that the proposed method achieves an inference time of 0.02 s per image with only 2.6M parameters on the Cornell dataset, and attains a grasping accuracy of about 92% with 0.03 s latency on the Jacquard dataset. Furthermore, real-world planar grasping experiments demonstrate that the method can predict effective grasp keypoints from RGB images and guide a manipulator to grasp objects, validating its feasibility and practicality.