6-DoF Grasping Pose Generation Based on Two-Stage Keypoint Regression

This study proposes a grasping pose gen-eration framework based on two-stage keypoint regression, which takes RGB-D images as input, predicts the 2D perspective projection of predefined 3D keypoints of the gripper, and uses PnP algorithm to restore the 3D grasping pose. Unlike traditional single-stage keypoint regression methods, we introduce gripper center information and adopt an adaptive keypoint offset mechanism to dynamically correct complex back-grounds and multi-target scenes, thereby improving the accuracy of keypoint localization. The experimental results show that this method achieves high accuracy in grasping pose generation on the Primitive Shape dataset.