A Novel Grasping Approach with Dynamic Annotation Mechanism

The Grasping of unknown objects is a challenging but critical problem in the field of robotic research. However, existing studies only focus on the shape of objects and ignore the impact of the differences in robot systems which has a vital influence on the completion of grasping tasks. In this work, we present a novel grasping approach with a dynamic annotation mechanism to address the problem, which includes a grasping dataset and a grasping detection network. The dataset provides two annotations named basic and decent annotation respectively, and the former can be transformed to the latter according to mechanical parameters of antipodal grippers and absolute positioning accuracies of robots. So that we take the characters of the robot system into account. Meanwhile, a new evaluation metric is presented to provide reliable assessments for the predicted grasps. The proposed grasping detection network is a fully convolutional network that can generate robust grasps for robots. In addition, evaluations based on datasets and experiments on a real robot show the effectiveness of our approach.