Research On Robot Grasping Based On Improved Grasp Quality Convolutional Neural Network

Robotic grasping is a key step in tasks such as handling,sorting,and assemblage.At present,the accuracy of robotic grasping is affected by many factors such as unknown object information,grasping planning design and dynamic environment,which resulting in the ability to grasp the unknown objects without meeting the application requirements.It is of great significance to study the efficient,accurate and reusable gripping method for improving the working efficiency of the robot arm and reducing the production cost.In order to solve the problem that robots are difficult to achieve effective grasping of unknown objects,the following research is carried out:(1)Based on the Yumi robot's self-acquisition hardware system,the transformation of the camera coordinate system and the robot coordinate system is analyzed,the hand-eye calibration model and the robot kinematics model of the depth camera are established.(2)The grasping planning process based on the Grasping point Quality judgment Convolutional Neural Network(GQ-CNN)is analyzed,and the four-dimensional attitude parameters of the robot fixture are designed.Aiming at the problem that GQ-CNN cannot effectively evaluate the candidate points of complex objects,an improved multi-scale Grasping point Quality judgment Convolutional Neural Networks(GQ-CNN1)is proposed,which replaces large-scale by using multiple small-scale convolution kernels that improves the accuracy of the network to evaluate the quality of the grasping points.The experiment proves that the improved GQ-CNN1 performance is better than GQ-CNN,and solves the object grasping problem in complex environments to some extent.(3)Aiming at the problem that the GQ-CNN1 network cannot effectively evaluate the small object candidate grasping points,an improved Grasping point Quality judgment Convolutional Neural Networks(GQ-CNN2)is proposed.By combining the low-level feature information into the high-level feature information,reduce the features information lost due to pooling,thus improving the network's accurate judgment of small object grasping points.The experiment proves that the improved GQ-CNN2 network performance is better than GQ-CNN1 and the original network,which can better solve the object grasping problem in complex environments.In this paper,the problem that the robot is difficult to achieve effective grasping of unknown objects which solved by our proposed algorithm.The experimental results prove the effectiveness of it,and our results have a great significance for improving the accuracy of robot grasping.