Reinforcement Learning-Based Control Method For Precision Assembly Of Robot Shaft Holes

In order to make robots more intelligent in the assembly process in industrial manufacturing,which can adapt to complex environments,new tasks,and unknown models,an intelligent assembly framework system is proposed,in which robots can autonomously generate intelligent assembly strategies when performing assembly tasks.This paper includes the study of the flexibility-based control strategy,the study of the deep reinforcement learning assembly strategy and the simulation training and related comparison experiments,aiming at improving the assembly accuracy and intelligence of industrial robots in the working process.When a robotic arm is used as a tool,a kinematic model of the arm is needed to enable it to execute control commands accurately.To ensure that the robotic arm can efficiently and accurately complete commands,it will be tested and verified using a forward and inverse kinematic solution to provide basic support for the design of a guide controller.The contact problem in the assembly task will be analyzed,and a robot arm guide controller with a softening effect will be proposed,and the impedance model will be rewritten to solve the workpiece conflict problem by controlling the robot arm position,thus serving to soften the system,improve the assembly efficiency,robustness,and avoid the loss of workpiece accuracy due to excessive contact force.The deep reinforcement learning algorithm is used in the face of the assembly task,and the TD3 algorithm is chosen for its excellent results in dealing with the continuous action space problem.The algorithm is analyzing and processing high-dimensional data by combining reinforcement learning with deep neural networks.And the algorithm is modified for the applicable assembly task.Finally,the TD3 algorithm is optimized and improved for the problems on the assembly task,and by optimizing the network parameters and improving the data update processing,it is able to improve the efficiency and stability in dealing with the shaft-hole assembly problem.The specific optimization includes adaptive guidance,neural network pre-training,etc.By changing the algorithm strategy decision method,thus improving the efficiency of the training of the assembly task and accelerating the convergence of the training.In the experimental part,due to the different data update efficiency between simulation and physical experiments,corresponding data migration learning algorithms are designed.This paper is mainly done by combining Coppelia Sim simulation platform and pycharm and building JAKA Zu 7 robot experiment platform.Long-term payoff design is done based on training,assembly tasks are considered in layers,and reward function rules are developed.And the effect of the robotic arm guide controller on the soft handling when high contact forces are generated by the workpiece assembly is verified.Finally,a multi-control group experiment is conducted to experiment the multiple aspects of the deep reinforcement learning flexible assembly strategy.From the experimental conclusion showing the multi-control group experiments for shaft parts in different positions and different initial assembly positions,it is finally concluded that the assembly success rate reaches 100% when the shaft parts to be assembled are within 1mm of the position and 13 o of the initial offset angle,and the success rate is as high as 92% within 3mm of the initial assembly range.The feasibility,stability,and efficiency of the application of the proposed intelligent assembly method in the precision assembly of shaft bores were verified.