Robotic Target Classification And Grasping Based On Deep Learning

Robotic target classification and grasping problems can be divided into two parts:grasping detection and grasping execution.The graspping detection part recognizes the type of the graspping object from the image information of the camera and detects the graspping posture configuration of the end of the robotic arm;the graspping execution part of the robotic arm plans the trajectory of the robotic arm and controls it move to the target posture,so the target can be grabbed and placed to a certain position.First,a hardware system for target classifying and grabbing using a robotic arm is set up,and the coordinate conversion relationship between the coordinate system of the depth camera and the base of the robotic arm is obtained by hand-eye calibration.This thesis proposes and implements a classification-based grasp detection method based on deep learning and an obstacle avoidance planning algorithm for the UR3 manipulator on the designed hardware system.In the classification and detection part,this thesis combines a deep learning-based object detection algorithm,an ellipse fitting algorithm,and a point cloud plane fitting algorithm to propose a classification and grasping detection method based on deep learning.This method can accurately detect the target object's category and grasping configuration,and obtain the posture of the end of the robotic arm that is perpendicular to the surface of the target object,which is the basis for the robotic arm to complete the actual grasping task.This classification and grasp detection method has a certain success rate in the grasp detection and actual target grasping experiments designed in this theis.Using the point cloud plane fitting,this method can capture the target from the direction perpendicular to the surface of the grasping object.Compared to the grasping method with the direction perpendicular to the plane of the table,for the grasping target placed at different angles,the success rate of our proposed method is higher.In the grasping execution part of the robotic arm,in order to avoid the failure of the task due to the collision between the robotic arm and objects in the robotic workspace during the movement,this thesis proposes an obstacle avoidance planning method for the UR3 robotic arm.The types and positions of obstacles in the task are randomly changed.Therefore,the point cloud information of the depth camera is used to sense the unstructured environment of the working space in real time to generate a spatial probability map and map the environment map to the obstacle environment model in the working space of the robot arm.Based on this,the UR3 robotic arm obstacle avoidance planning method is proposed and implemented.For static grasping environment,it can accurately avoid obstacles in the grasping scene and reach the target position along the planned trajectory.For a dynamically changing scene,when the environment changes,the robotic arm can re-plan to obtain a new obstacle avoidance trajectory.