Research On Gait Control Methods Of Quadruped Robot Based On Reinforcement Learning And CPG

Compared with wheeled robots and tracked robots,legged legged robots can flexibly plan landing points in complex terranes and effectively avoid and cross obstacles due to their high degree of freedom.Legged legged robots can help humans perform various tasks such as mountain search and rescue,field transportation and dangerous environment inspection.Quadruped robot has been studied by many teams because of its excellent stability and good dynamic performance.Compared with hydraulically driven quadruped robot,the compact and flexible motor driven quadruped robot stands out from the quadruped robot with its higher energy efficiency,environmental protection and more convenient maintenance.In order to realize the high adaptive motion ability of quadruped robot in complex environment,it is necessary to develop a reliable real-time control system and adopt a stable control method.In this paper,based on the motor driven quadruped robot platform designed by Shandong University,the overall control system of quadruped robot is designed,and the CPG control method and reinforcement learning control method are analyzed and verified by simulation experiments.The main research contents of this paper are as follows:(1)Platform construction and kinematics modeling of quadruped robot.Firstly,the mechanical structure design of quadruped robot is introduced.Secondly,based on the platform of quadruped robot,the hardware control system is built and the software communication architecture is developed and designed.Finally,the forward and inverse kinematics of quadruped robot is deduced by D-H method.(2)Control method based on central pattern generator.Firstly,the phase sequence of legs and feet of common locomotion gaits of quadruped robots was obtained by analyzing the rhythmic motion modes of various quadruped animals.Secondly,the dynamic characteristics of hopf oscillator are explored,and the CPG control network of quadruped robot is built using the oscillator.Finally,the parameters of the CPG network were adjusted by using the movement rules of quadruped animals to realize the output of multiple gait control signals.(3)Control method based on deep reinforcement learning.Firstly,the elements of interaction between quadruped robot and environment in reinforcement learning are introduced.Secondly,the design points of reinforcement learning controller,the design of state space,action space,termination conditions and reward function are proposed.Finally,the realization process of reinforcement learning controller and the network structure and parameters of reinforcement learning algorithm are designed.The above research content is verified by the physical experiment and simulation experiment.A position control experiment was carried out on a motor driven quadruped robot platform to verify the stability and real-time performance of the designed control system.The simulation experiment of CPG controller in pubullet simulation environment verifies the stability of the controller to generate multiple gait control signals.A reinforcement learning controller is used to train a quadruped robot,which proves its effectiveness.