Research on Pose Estimation Algorithm for Visual Robot Grasping Target Objects Based on Unet-DGN-L2

In recent years, advancements in robotics have significantly heightened interest in robotic arm grasping, a critical capability for intelligent robotic systems. Nevertheless, achieving autonomous and precise grasping remains a formidable challenge. Despite extensive research exploring diverse neural network architectures integrated with deep learning for robotic grasping, the achieved accuracy levels frequently fail to meet practical requirements. This study proposes a novel Unet-DGN-L2 neural network architecture, which employs the Unet framework as the primary mechanism for feature extraction while integrating a decoupled grasping network to generate pixel-wise grasping representations. To address the issue of overfitting, L2 regularization is incorporated, resulting in a robust model for predicting grasping poses. Evaluated on the Cornell dataset, the proposed Unet-DGN-L2 architecture achieves a grasping accuracy of 86%, demonstrating substantial improvements in autonomous and precise object grasping. This advancement enhances the applicability of intelligent robots in real-world scenarios, particularly in fields such as industrial automation, thereby contributing to the progression of robotic manipulation technologies.