STNet: A Robot Grasping Prediction Model Based on Self-Normalized Attention

With the advancement of deep learning, robot grasping prediction has seen significant improvement in recent years. However, the current prediction models still face the challenge of poor generalization ability and a limited applicability in factory. In this study, the self-normalized attention network (STNet) is proposed to predict highly robust grasping of robot. STNet extracts multi-scale feature information by leveraging upsampling and downsampling techniques. Shifted-windows attention mechanism is used to extract and fuse global features and local features. Additionally, self-normalization is assigned to enhance the generalization capability of feature learning and improve the robustness of the model to cope with various grasping tasks. Experimental results on the Cornell grasping dataset demonstrate that STNet achieves an advanced level of accuracy, with 98.3% and 97.9% accuracy in image-wise (IW) and object-wise (OW) splitting respectively. In real robot grasping experiments, grasping postures of target object are predicted accurately, the grasp of irregular objects is achieved by this gasping method.This method provides a theoretical basis and application reference for the robot's grasping in real scenarios.