Reinforcement Learning-based Search Strategy For High-precision Peg-in-Hole Tasks

Peg-in-Hole assembly is a common operation in the field of mobile phone production and testing,usually done manually.Automated assembly by robots not only improves efficiency,but also reduces labor costs.But robots execute assembly tasks are still a major challenge in recent years.For peg-in-hole assembly,the traditional method is to solve the contact force model of the peg and hole based on static analysis,which requires a lot of time and effort.And for high-precision peg-inhole assembly,the accuracy of the model is so low that the assembly cannot be completed due to the noise of the sensor,the positioning error of the arm and others.The model-free reinforcement learning control method need not to build a model,reduces the influence of errors such as sensors,and improves assembly accuracy.In this paper,for the high-precision axial hole assembly task,taking the optical fiber assembly in the 3C field as the research goal,analyzing the task requirements and selecting the appropriate actuator and torque sensor,the experimental platform under the real scene is built,and a basis is proposed.A high-precision axial hole assembly strategy search algorithm for reinforcement learning.According to the different characteristics of the assembly process,it is decomposed into two steps of finding and inserting holes.The hole is moved by moving the plug to the center of the jack by changing the position of the plug in the plane of the jack,and the target of the insertion is by changing the plug.In the forward direction,move the plug to the target depth.Different Markov decision processes are designed for the above two steps,and the low variance Actor-Critic method is introduced.The simulation environment was built under the ROS system and experiments were carried out.The results show that the high-precision shaft hole assembly can be completed.Compared with other algorithms,the proposed algorithm enables the robot to learn how to select the best action,less execution steps,high success rate and complete the shaft hole assembly under a small contact force in a short training time.Finally,the experiments in the real scene also prove the effectiveness of the algorithm,and test the performance of the algorithm under different initial conditions.When the time for setting the arm to move one step is 0.2s,the assembly time of the shaft hole is at least 4.4s,the number of execution steps is 11-14 steps,and the assembly success rate is 98%.Since the algorithm proposed in this paper uses the relative position between the shaft holes as a state input,the robot can be used for porous assembly without repeated training.The experiment proves that the algorithm has very important practical significance for modern factory production assembly.