Vision-based Adjustment with Toleration of Error before Grasping

Automatic grasp has always been a fundamental topic in the field of robotics research. Only with the stable and high success rate grasp, can the robots behave well in some industrial scenes and man-machine interaction scenes which usually contain grasping tasks. However, the error of various devices caused by the diversity of the complicated environments mentioned above is always ignored by the traditional grasp frame assuming that the sensor and the actuator are precise. In this paper, we take error tolerance as the core idea, proposing a novel frame of grasp which includes global search and local adjustment. In addition to an under-actuated gripper, we also use a depth camera and an ordinary micro camera in the whole grasp system. Based on the data obtained by the cameras and an output from ResNet, the overall state of the gripper and the target object is estimated to make appropriate adjustments. This improvement enables the grasp system to adapt to more various unstructured environments. Finally, we conducted experiments on three common application scenarios proving the validity of the improvement.