Research On A Robot Arm Throwing Subject Based On Reinforcement Learning

At present,the research of robot arm throwing subject mainly focuses on the optimization of the trajectory,but the research of its release action is less.Especially in the use of reinforcement learning method,this problem is particularly obvious.In recent years,Reinforcement learning is widely used in various fields.However,when some actions have special restrictions,it is difficult to get effective results by using reinforcement learning algorithm directly.The throwing problem of our subject has the characteristics of this kind of problem.In our subject,reinforcement learning is used to deal with the problem of robot arm throwing task with release action.Because the release action has the special limitation that it will end the round directly after execution,the release action will affect the strategy learning.In this paper,a reinforcement learning method based on Bayesian optimization is proposed to solve the task,and its effectiveness is proved in practice.Firstly,this paper introduces the development of reinforcement learning and the research status of manipulator control,as well as the research significance of manipulator control based on reinforcement learning.It is found that in many researches of robot arm throwing task based on reinforcement learning,the release action is simplified.Therefore,this paper will face up to this problem,and study how to use reinforcement learning to deal with the throwing problem without simplifying the release action.Secondly,this paper introduces the knowledge of reinforcement learning in detail.Including the basic knowledge of reinforcement learning,classification methods,etc.,and the development of some mainstream algorithms and the advantages and characteristics of each algorithm are introduced.Then,this paper studies the problem of robot arm throwing task with release action.In this paper,a robot arm throwing task with release action is concreted firstly,and then based on v-rep simulation platform,the simulation environment is built,and the reinforcement learning algorithm is directly connected for experiments to verify that the problem invalidates the algorithm.By analyzing the causes of the problems,a reinforcement learning method combined with Bayesian optimization is proposed.This method optimizes the actions with different levels of restriction to avoid mutual interference in the learning process.Finally,the feasibility of this method is verified by experiments.The final training model can complete the task well.