Research On Robot Multiple Peg-in-hole Assembly Method In Continuous Action Space

With the gradual improvement of industrial production technology,more and more robot equipment plays an important role in production.Improving assembly efficiency,product quality,and reducing cost consumption are industry needs facing the industrial field.At present,most industrial robot assembly operations are used in structured environments.When faced with high-precision requirements,fragile materials,complex geometric constraints,and weak stiffness parts of the workpieces,it is difficult to ensure product effects,lacking robustness and autonomous learning capabilities.Peg-in-hole assembly is a typical case in many assembly processes.Among them,the interaction of multiple peg-in-hole assembly components is more complicated,the contact state is changeable,and there are problems such as uncertainties such as contact force jump.It is necessary to improve the assembly strategy of the robot to improve self-learning ability to achieve compliant control.In this paper,with the multiple peg-in-hole assembly process as the research background,combined with robotics,deep reinforcement learning and other theories,the assembly strategy of the robot in the continuous motion space was studied,and the algorithm was verified in the multiple peg-in-hole flexible assembly platform,which has achieved good results.The major research contributions of this paper are as follows:(1)The assembly process of multiple peg-in-hole was described.The multiple peg-in-hole assembly process was introduced,and the process was divided into the hole searching stage and the insertion stage for analysis.(2)The physical platform and simulation platform of multiple peg-in-hole assembly were built.In this paper,the system composition and experimental design of each platform were introduced,and the design of reward function in different environments and the application of template matching method were studied,so as to provide platform support for algorithm verification.(3)Carried out theoretical research on deep reinforcement learning,and proposed a flexible assembly system method based on NAF and DDPG algorithms.First,combined with the robot's multiple peg-in-hole assembly process,the deep reinforcement learning algorithm was analyzed and explained,and the significance of the continuous motion space in the application of robot research was analyzed.Then the principle and training process of NAF and DDPG algorithms were studied,the network structure and training design were designed,and the multiple peg-in-hole assembly learning model was constructed.Finally,the training process image and assembly success rate were analyzed.The experimental results showed that the proposed assembly method could effectively improve the system learning ability,and guide the robot to complete the assembly task more accurately.(4)In view of the problems of empirical playback methods occupying more resources and limited convergence performance,A3C algorithm research was carried out.The concept of multithreading was introduced to improve the learning ability,and some shortcomings of previous algorithms using experience playback mechanism in continuous action space were analyzed.For example,the system learning process is too long.In this chapter,the significance of algorithms were summarized in reducing system storage space,shortening calculation time,and improving training efficiency.At the same time,it put forward ideas for improvement.Finally,it sorted out and summarized the main research content and work,summed up the current shortcomings of this paper and the direction of improvement that can be carried out in the future.