Research On Robotic Grasping System Based On Deep Learning

In recent years,due to the proposal of "German Industry 4.0" and"Made in China 2025" planning,intelligent robots' research has entered a rapid development period.As one of the particularly important topics,the robot grasping system has also been paid attention to by many researchers.Based on the six-degree-of-freedom serial robot,this paper studies the robot grasping system based on deep learning.The system can actively identify the image's target position and provide the grasping pose and motion trajectory for the robot.The main work content and research results of this paper include the following aspects.Kinematic modeling of robot:the forward and inverse kinematics models of the robot are established using the Modified D-H method,and the correctness of the models is verified.In MATLAB,we use the Robotics Toolbox to establish the robot's simulation model and analyze the working space at the end of the robot by the Monte Carlo method.Robot trajectory planning:We designed a time-optimal trajectory planning algorithm that simultaneously searches the optimal path based on cubic uniform B-spline interpolation and genetic algorithm.Simultaneously,the simulation and comparison experiments based on Matlab verify the effectiveness of the proposed algorithm and its superiority relative to other similar algorithms.Target recognition based on deep learning:We made use of the 3D industrial camera to create a robot object grasping and recognition dataset.At the same time,we used the Labelme tool to mark the object category and boundary of the images in the dataset.The Mask R-CNN model was trained and tested on our own dataset,and the average recognition accuracy of the model was 83.33%.Grasp detection based on deep learning:This paper creates a grasp detection dataset that matches the target recognition model and annotates the appropriate grasp position corresponding to the target.The MultiGrasp model is used as the basic model of grasp detection,and the dataset produced in this paper is used to train and test the model.The average test accuracy of the grasp detection model reaches 94.12%.System simulation experiment platform:The upper computer debugging software is written based on C#language,and the system simulation experiment platform is built on this basis.The platform is used to simulate the process of grasping the target object to verify each module algorithm's feasibility.In summary,this paper completely simulated the process of robot grasping detection and verified the model's detection performance in the multi-target complex environment.The model operation results show that our model can accurately identify and detect the target object and its grasping position in the single target scene or multi-target scene and guide the robot to complete the grasping operation.