Detecting Graspable Rectangles of Objects in Robotic Grasping

Most convolutional neural network based grasp detection methods evaluate the predicted grasp by computing its overlap with the selected ground truth grasp. But for typical grasp datasets, not all graspable examples are labelled as ground truths. Hence, directly back propagating the generated loss during training could not fully reveal the graspable ability of the predicted grasp. In this paper, we integrate the grasp mapping mechanism with the convolutional neural network, and propose a multi-scale, multi-grasp detection model. First, we connect each labeled grasp and refine them by discarding inconsistent and redundant connections to form the grasp path. Then, the predicted grasp is mapped to the grasp path and the error between them is used for back-propagation as well as grasp evaluation. Last, they are combined into the multi-grasp detection framework to detect grasps with efficiency. Experimental results both on Cornell Grasping Dataset and real-world robotic grasping system verify the effectiveness of our proposed method. In addition, its detection accuracy keeps relatively stable even in the circumstance of high Jaccard threshold.