Abstract: Visual guided robotic grasping of soft gripper depends on correct grasp position, grasp angle and grasp depth, and therefore a complete grasp configuration model and a multi-task loss function for soft gripper are proposed. A two-stage deep learning network based on anchor and rotating blocks is designed to realize direct map from image to multi-gripper grasping. The performance of the network is analyzed by public cornell grasping dataset and self-built dataset. The results show that the two-stage network based on multi-task loss and anchor with rotated blocks improves the accuracy of multi-output grasp detection and increases the success rate of robotic grasping. Finally, the soft robotic grasping system is constructed and the robotic grasping experiment results show that the proposed method provides a certain robustness to vision error, achieves 96% grasp success rate at different fruits, and exhibits a good generalization ability to grasp fruit peel.

Key words: soft gripper; robotic grasping; deep learning; multi-task