A Multi-Task Grasp Detection Network Based on Global Channel–Spatial Attention Mechanism for Industrial Workshop Environments

To improve the accuracy and robustness of target grasp detection and classification for robotic manipulators in unstructured environments such as workshops, a multi-task grasp detection algorithm based on a global channel–spatial attention mechanism is proposed. Through the global channel attention module, the network is able to extract richer semantic information while reducing redundant feature information, thereby improving overall performance. The effectiveness of the proposed algorithm is validated on the Cornell dataset and a self-built dataset. The grasp prediction accuracy reaches 99.5% on the Cornell dataset, while on the self-built dataset the grasp detection accuracy reaches 98.5% and the object classification accuracy reaches 99.5%. To further verify the effectiveness of the algorithm, grasping experiments are conducted on an AUBO-i3 robotic manipulator. The results show that the average grasping accuracy reaches 95% for unseen daily items and common workshop objects, which is 8.75% higher than the baseline GGCNN network. These results indicate that the proposed algorithm can effectively perform grasping tasks for robotic manipulators in unstructured environments.