An Automatic Grasp System with Sensor Feedback Based on Soft Gripper

In this paper, we present a automatic robotic grasp system that is capable of grasping objects with a wide range of sizes and shapes firmly and delicately. The two main components of grasping object are grasp selection and grasp execution. The key novel feature of our proposal is that we define three grasp primitives which are able to handle diverse objects. In grasp selection process, the algorithm needs to determine a better grasp primitive for each grasp position. This phase does not rely on precise object model but just on the local length and width information of object. For different grasp primitives, there are different sensors triggered. We utilize decision-making tree method to build a rules set which guides grasp execution. The pressure value of suction cups, tactile information and the drive motor angle are needed in training process. We capture over 3000 data groups and labels in total to train the rules set with cross validation method. For decision-making tree, the true positive rate and true negative rate are 92.74% and 97.84% respectively. Our results also show that the gripper can grasp paper cups of diverse diameters delicately without crushing them. Finally, we display the ability of our system by grasping diverse objects automatically.