Research On Robot Autonomous Grasp Detection Algorithm Based On Convolutional Neural Network

Robotic grasp is one of the most widely used robotic tasks and also one of the most challenging technologies in robotic operation.The current mature robotic grasp systems are aimed at the structured operating environment,and rely on the pre-acquired object model to plan grasp process.The type,shape,pose,size,and color of the grasp object and the application scenario is relatively fixed.The system is lack of flexibility and robustness.In order to adapt to the application requirements of the unknown target in the unstructured environment,the key is how to perform the self-detection of the captured pose for the unknown target.In this paper,grasp detection algorithm of unknown target objects is studied.Robot grasp detection algorithm based on convolutional neural network is proposed and verified by experiments.The main contents include:The representing method of robot grasp pose based on RGB-D image is studied,by which the robot 3D space grasp detection problem is transformed into 2D image detection problem.A directed rectangular box representation method on the image plane is presented to represent the 3D space of the robot.Taking the direction of the normal direction of the surface of the object,combined with the RGB-D image data,the unique parameters of the directed rectangular frame are mapped to the 3D space of the robot to capture the pose parameters.A grasp detection algorithm based on sliding window detection is proposed,whose accuracy,compared with the same type of algorithm,is improved.Based on the convolutional neural network model,a discriminative confidence model for grasp detection is designed,by which the candidate grasps generated by the sliding window algorithm are ranked and the optimal grasp is obtained.The algorithm achieved a 91.3% grasp detection accuracy on the Cornell grasp dataset.A grasp detection algorithm based on deep residual network is proposed.The accuracy of grasp detection is maintained while the detection speed is further improved.The convolutional neural network model based on deep residual network is used to learn the RGB-D image data and the optimal grasping pose relationship end-to-end.The data augmentation and transfer learning technology is used to accelerate the learning process and avoids the overfitting problem.The algorithm's accuracy reached 90.1%,and the detection speed was increased to 0.37 sec/frame compared to the sliding window based algorithm.A verification experiment platform based on Kinect depth camera and UR5 robot is designed for the familiarity and unknown target objects.The experimental results show that the grasp detection accuracy and robotic grasp success rate of the algorithm have reached a superior level,and the robot autonomous grasp process is flexible and robust.