Research On Motion Planning Of Manipulator Base On Reinforcement Learning

In 2015,China released the made in China 2025 plan,in which the manipulator in industrial automation as the main technical field is focused.In the modern industry,the application of the manipulator is still in the state of artificial teaching.In the future,the industrial manipulator needs to be intelligent,that is to say,the manipulator needs to have a certain ability of autonomous motion planning.Therefore,this thesis studies the robot arm in the model free and multi-object environment,through the reinforcement learning algorithm to combine the push and grab action,and plans the appropriate strategy to complete the grab task.Based on the investigation of the current research situation at home and abroad,through reading a large number of books,papers and references at home and abroad.In this thesis,starting from the modeling basis of the manipulator,the mathematical basis and kinematic problems involved in the modeling,as well as the reinforcement learning algorithm are briefly described.This thesis points out the problems in the research of manipulator motion planning,and then puts forward the idea of applying reinforcement learning algorithm to solve the difficulties and problems in the environment of model free and multi-obj ect.There are two main innovative methods in this thesis:first,through human's habit of grabbing objects in multi-object environment,we propose a mechanical arm behavior mode which combines the push action and grabbing action.By pushing the action to rearrange the disordered objects,the space for grabbing is made for the manipulator and the actuator.At the same time,the grabbing action can move the objects and make the pushing action more accurate.Secondly,in the DQN algorithm without model reinforcement learning,two full convolution networks are used to train the push action and the grab action respectively.In this way,the cooperative action of pushing and grabbing becomes possible,and the manipulator learns the pushing action strategy that makes grabbing successful.In this thesis,the training and testing are carried out in the simulation and real environment.The environment is constructed by the random combination of many kinds of building blocks and the fixed special combination,so that the robot arm can be fully trained and tested through these environments to ensure that the final strategy can achieve the desired results.Finally,experiments show that the robot arm can quickly learn the cooperative strategy of pushing and grabbing action by using reinforcement learning algorithm,and can complete the grabbing task efficiently in the multi-object environment.