Robotic Grasping of Novel Objects from RGB-D Images by Using Multi-Level Convolutional Neural Networks

Inspired by human grasping, we propose an efficient robotic grasping detection method to get the optimal grasping rectangle from RGB-D images by developing multi-level Convolutional Neural Networks (CNNs). The multi-level CNNs consist of three levels with different structures and functions. The first level is designed to locate the grasped object roughly, and then determine the searching range of the grasping rectangles. The second level is designed to obtain the preselected grasping rectangles, rapidly find out usable ones and eliminate unusable ones by a small-scale CNN. The third level is designed to reevaluate the preselected grasping rectangles and catch much more features with a large-scale CNN, so as to evaluate every single preselected grasping rectangle accurately. A selection algorithm is presented further to obtain the optimal grasping rectangle. Tests on a large-scale grasping dataset validate that, the proposed robotic grasping detection method can significantly improve the accuracy of the grasping rectangle, compared with the existing methods based on CNNs. Grasping experiments are performed on a 5-DOF Youbot arm, and the experimental results indicate that the proposed method can find out the optimal grasping rectangle and successfully achieve novel object grasping.