초록 열기/닫기 버튼
In this paper, the optimal griping point learning is conducted through the neural network with the input of point cloud data. The SG-DNN (Stable Grasping-Deep Neural Network) algorithm, which was applied only to the shapes classified as the “original type”, is extended to the real object and additionally composed of cost functions. Through this, a study was conducted to extract the optimal point for stable griping. Griping objects were simplified and limited to 10 types of shapes in the SG-DNN algorithm, and 4 cost functions were designed accordingly in order to extract the optimal griping point. However, when applying the method for an actual object, there was a limitation in deriving the optimal griping point. The content of the text uses a sample extracted from coordinate points through the point cloud data of a real object as the input of the neural network model. In comparison to the conventional method, a learning data set including dense surface area coordinates is constructed, leading to the extraction of a precise gripping point. The cost function was supplemented with cost functions for height, width, and rotation, and was designed to fit the actual object while including the cost function for the mass center according to the material and the gripper depth. The performance of the algorithm was verified by griping experiments involving real objects.
