Research On The Motion Planning Of Industrial Manipulator Based On Reinforcement Learning

Industrial manipulator takes an important part in industrial automation system.There are nonlinear coupling problems with the control of manipulator.Most of the traditional manipulator algorithms have the problems of complex modeling and poor environmental applicability,which cannot meet the complex requirements.In this thesis,a motion planning method for the industrial manipulator based on reinforcement learning is proposed for the real industrial environment.The algorithm can be applied in the real environment without complex kinematic modeling.The research contents of this paper are as follows:The basis of reinforcement learning algorithm is analyzed.And the motion control problems of manipulators are studied in the real environment.A modeling method of Markov Decision Process for the motion planning of manipulator is proposed based on the classical kinematics algorithm.To narrow the gap between simulation and reality,the simulation model is transformed from the realistic environment.A motion planning algorithm for manipulators based on trust region policy optimization method is proposed.The Actor-Critic framework is chosen to optimize the policy function and the value function.To ensure the safety and rationality of the manipulator control policy,A constrained method of reward function is proposed.Besides,the Hessian Free method is used to optimize the policy function.In this thesis,a reinforcement learning method of manipulator is proposed,which is based on the real environment to build simulation model and train in simulation environment.Finally,the practical of the motion planning and control algorithm is proved by experiments.The experiment platform is designed based on ROS.Analyze the experimental results and evaluate the performance of the algorithm.Results of the experiment show that the motion planning algorithm of manipulator based on reinforcement learning has the advantages of environmental applicability and strong robustness.And the control policy obtained by reinforcement learning can be optimized incrementally according to the environment.