Research On Technology Of Robot Peg-in-Hole Assembly Based On Learning Method

With the development of science and technology,robots are increasingly used to replace various human operating tasks such as welding,handling,painting,assembly.Among them assembly robot is an important part of industrial application in aviation parts,engine,windshield large objects assembly,and assembly of electronic components and other small parts.Peg-in-hole assembly is a typical basic assembly unit in robot assembly operations,such as bearing,PCB board and engine room rivet assembly are peg-in-hole assembly tasks.When the robot performs the peg-in-hole assembly task,it faces the following difficulties:For the small assembly parts,the robot has some difficulties in assembly accuracy;The non-structural environment interferes with the robot and affects the assembly success rate of the robot,such as the obstacles in the robot workspace and the position movement of the hole.There is error in the positioning of the hole,which leads to large interaction force when each assembly part is in contact with each other,resulting in part wear.Therefore,it is of great significance to study the technology of intelligent robot pegin-hole assembly in unstructured environment.Based on robots to complete the demand of the high-precision assembly task,peg-in-hole assembly tasks can be divided into three phases:the mobile phase,searching phase and insertion phase.Carry out the pegin-hole assembly robot technology based on the method of learning research,including mobile phase trajectory optimization method study,hole strategy learning of manipulator considering the prior knowledge and strategy of people in the loop of insertion.Firstly,the trajectory optimization method based on teaching and learning was studied in view of obstacles in the working space and joint stability factors in the moving phase.According to the obstacle avoidance requirements of the robot arm in the moving phase,considering the proximity between the generated trajectory and the teaching trajectory,the smoothness factor of the generated trajectory motion,the obstacle avoidance performance index is designed.An obstacle avoidance method for manipulator end based on dynamic motion primitive algorithm is proposed.The coupling terms of obstacle avoidance are designed based on the functions of traditional static and dynamic situation fields,and the path planning of manipulator’s obstacle avoidance in the moving phase is realized.According to the stability index of joint motion,the coupling term is designed to optimize the joint trajectory.In the simulation environment,the end trajectory of the learned manipulator can be adjusted according to the position and size of obstacles.Compared with the traditional method,the average distance between the generated trajectory and the teaching trajectory is reduced by at least 3.09%,and the average acceleration of the trajectory is reduced by 12.60%.Secondly,aiming at the sensor noise and manipulator positioning error in the hole seeking phase,a hole seeking strategy learning method based on reinforcement learning is studied.Aiming at the problems of low learning efficiency and low success rate of traditional deep reinforcement learning algorithm,this paper improved the method by adding human teaching data and using priority experience playback instead of random playback to improve the training speed and efficiency of reinforcement learning model.According to the position of axle hole and contact force,the reward function is designed to realize the learning of hole finding strategy.It is verified in V-REP simulation environment,and the success rate of hole finding task is increased from 75.5%to 89.5%.Thirdly,gaussian mixture model was used to encode the data of peg velocity and peg-hole contact force in the process of assembly,and the conditional probability distribution of peg speed and peg-hole contact force was obtained,which can guide the insertion task.and the probability distribution was used to guide the insertion task.In order to improve the success rate of the insertion task,this paper proposed that when the insertion task failed,the robot control mode was switched to manual teaching mode to guide the robot to continue to complete the insertion task,and the parameters of Gaussian mixture model were updated to improve the success rate of the insertion strategy.The experimental results show that the success rate of insertion has been improved.Finally,the peg-in-hole assembly platform was built,and the end obstacle avoidance experiment,hole seeking strategy verification experiment and insertion strategy verification experiment of the manipulator were designed respectively.The experimental data were then analyzed to verify the validity and reliability of the proposed method.The experimental results show that the manipulator can realize the learning of peg-in-hole assembly skills,and the end of the manipulator can avoid obstacles in the moving phase.The success rate of hole searching is increased from 65%to 87%,and the success rate of insertion is increased from 70%to 90%.