Grasping pose detection based on self-attention and convolutional neural networks

Abstract To address the challenges commonly encountered in existing robotic arm grasping networks, this study introduces a lightweight grasping pose detection network, SA-Grasp. This network employs a deep convolutional neural network as its backbone structure, utilizes the Coordinate Attention to suppress the interference of irrelevant information, and incorporates a self-attention module to aggregate global image information. Experimental results demonstrate that SA-Grasp achieves impressive grasping detection accuracies of 98.37% and 97.19% on the image-wise and object-wise subsets of the Cornell dataset, respectively, and a grasping detection accuracy of 96.33% on the Jacquard dataset. Furthermore, the network requires a mere 21 ms to detect a single image. When applied to a real-world planar grasping experimental platform for 180 grasping experiments across nine types of objects, SA-Grasp achieves a grasping success rate of 94.44%, thereby verifying its reliability.