Robotic Grasp And Assembly Method Based On Intelligent Perception And Learning

With the improvement of industrial manufacturing techniques,robots are widely used in human daily production and life.However,the production level of robots is limited by the change in the production environment.It is difficult to perform flexible production.By sensing the external environment,the robot can adjust its posture according to the information.Therefore,a robot grasping and assembling method based on intelligent perception and learning is proposed.In further,a robot grasping and assembling system is designed on this basis and verified by experiments.According to the above content,this paper carried out a robot grasping and assembling method based on intelligent perception and learning,which mainly included the following aspects:Firstly,at the level of intelligent perception,a five-dimensional grasp pose representation method was constructed based on RGB-D images,and an intelligent perception method based on contact state was designed.On the level of execution planning,the framework of robot motion planning is constructed based on the Rapidly Exploring Random Trees(RRT)method.Move IT! software package is used to control robot motion planning.Kinect V2 camera is used as a sensor to obtain the image,and hand-eye calibration is used to complete the coordinate transformation of the robot.Finally,a robot grasping and assembling system is constructed.Secondly,for unknown irregular objects under an unstructured environment,this paper proposes a grasp position detection method based on a cascaded convolutional neural network,implements an end-to-end robotic autonomous grasping method.Mask-RCNN is used for extracting the grasping feature and the grasp position candidate bounding-boxes.In order to guarantee the generalization ability and improve the detection rate,we estimate the grasp angle based on Y-Net.To solve the problem of insufficient accuracy of grasping position,Q-Net is proposed to acquire the grasp quality distribution.Finally,the robot finishes the robot grasping task according to the best grasp posture.Thirdly,taking the gear assembly station of the retarder assembly line as a research background,an online modeling learning and parameter optimization method for the complex robot assembly process are proposed to solve the problems of low assembly success rate and low efficiency using current offline methods.For the complex and changeable robotic assembly process of gear,the dynamic model of contact states and robot motion is established based on Gaussian process regression(GPR).An improved particle swarm optimization algorithm based on a generative adversarial algorithm is used in online learning to generate an optimization strategy of assembly key parameters.Support vector data description is utilized to find out the new assembly data.The method finally realizes the online modeling and parameter optimization of the assembly process.Finally,Experiments are validated in the dataset and real environment respectively.The experimental results show that the method could quickly calculate the robot posture.Compared to the previous methods,it has considerable improvement in grasp accuracy and speed.It can be applied to clutter object scenarios and other scenarios.It has strong stability and robustness.Experiment results of the gear and spline shaft show that the GAPSO-GPR method is superior to the manual and offline methods in assembly success rate and efficiency.It can be used for the online assembly of gears with different batches and specifications,which can meet the actual production demand of the retarder assembly line.