Research On Control Of Robotic Manipulator Based On Reinforcement Learning

This thesis concentrated on the control problem for robotic manipulators. Reinforcement learning is learning what to do - how to map situations to actions - so as to maximize a numerical reward signal. Reinforcement learning does not require a complete model of the environment. In order to complete the task, Agents interact with the environment, and access to knowledge. Reinforcement learning has been used more and more in artificial intelligence and robotics control.The manipulator control was researched and the existing control methods were analyzed in this thesis. To some extent of simplification, a multi-agent control system was translated to a 4-DOF planar manipulator, and the state of the environment variables was divided into the state of deviation angles and minimum distances.Considering the disadvantages of reinforcement learning such as slow convergence and easiness appearance of"dimension disaster", continuous state space was analyzed and processed into discretization, and finally a discrete state space suitable for reinforcement learning was built up as a substitution. Simultaneously, the system credit assignments, especially the time and structure credit assignments were analyzed in this thesis.Through the comparison of a number of main reinforcement learning methods, the implement process of the algorithm was found and given. And a simulation platform was developed for the control system. And two simulation experiments were carried out using general reinforcement learning methods and the given method in this thesis. By comparing the results of two experiments, the method was proved to be effective and rationality.