| Grasping is not only the most common means of robot technology in real life,but also one of the most challenging tasks in robot operation.It is of great significance for liberating human hands and improving industrial productivity.Traditional robot grasping operations are mostly carried out in a structured environment,which requires the pre-acquired model of the object in advance and plan the grasping trajectory of the manipulator offline.The applicable scene is relatively fixed and the features of the object such as shape,color and size are single,lacking flexibility and robustness.In order to realize the precise grasping of arbitrary objects by the robot in an unstructured environment,the first problem to be solved is the grasping pose detection.Based on the RGB-D camera,this paper deeply studies the detection algorithm of grasping pose based on convolutional neural network.The efficient and robust grasping pose detection of the known object and the unknown object is realized and verified by the grasping experiments respectively.The main research contents include:This paper studies the representation method of robot grasping pose based on depth image,and uses two-dimensional image detection instead of three-dimensional space grasping pose detection.By combining the representation of multiple contact points with the 2D rotating rectangular frame,the two-point grasping pose representation method is proposed to represent the three-dimensional grasping pose of the robot,which is more in line with the pinching grasping operation of the two-finger parallel gripper.By combining with the depth data and limiting the fetching direction,the 2D fetching parameters are uniquely mapped back to the 3D grasping space.This paper proposes a two-stage grasping pose detection algorithm for known objects,which combines target detection with grasping pose detection.Using small sample data set to train the network model,which has a high adaptability to the change of object pose.Performing target detection based on SSD＿Mobile Net network model,and the threshold segmentation technology is used to separate the object from the background,and the principal component analysis method can be used to determine the grasping pose of multiple objects at the same time.Compared with the traditional method,it can effectively reduce the time of database collection and achieve the grasp success rate of 87%,with excellent performance.This paper proposes an end-to-end grasp detection algorithm of unknown objects.The fully convolutional network model is trained by the simulation data set and the grasp position,Angle and quality of each position in the depth map are directly output,which greatly reduces the collection time of data set and improves the accuracy of grasping pose detection.The grasp quality prediction function is used to evaluate the quality of grasping pose.The accuracy of the algorithm on the verification set reaches 0.899,which is higher than the accuracy of the original GQCNN.The average time for grasping pose detection of unknown objects is 0.65 s,which is 31.6% higher than the original network,and the grasp quality value are concentrated,all above 0.79,it is a great improvement over the previous one,which is unstable and the overall value is small.This paper built a robot grasp experiment platform to verify the performance of the grasping pose detection algorithm by grasping unknown objects of different shapes.The experimental results show that the accuracy of pose detection and the success rate of robot grasp are high,and the process is flexible and stable. |
