An improved YOLOv5 algorithm for enhancing pre-classification of soft gripper grasp

As a popular direction in the field of robotics, soft robots have many applications, for example, rescue, medical treatment, and human-computer interaction because of their flexibility and high adaptability. As a representative of soft robots, the soft hand is most used in object grasping. Therefore, it is necessary to explore and study soft hand-grasping technology. Object detection technology based on deep learning brings a new opportunity to the development of soft hand. It can pre-classify objects to be grasped in a non-contact way, so that a soft hand can better determine the grasping posture. Based on the YOLOv5 algorithm, this paper proposes an improved YOLOv5 algorithm to enhance the pre-classification ability of soft hands. A SamAM attention module has been introduced into the YOLOv5 structure to weigh different hierarchical features, making it more suitable for grasping situations. The experimental results show that on VOC2007, COCO, and common food data sets, the improved algorithm has different degrees of improvement in the evaluation performance index.