Motion Control Of Quadruped Inspection Robot In Oil And Gas Pipeline Station Based On Deep Reinforcement Learning

With the continuous growth of domestic oil and gas pipelines and stations along the way,the use of robots instead of manual inspection has become the focus of research.Oil and gas stations with complex pipelines,stairs and steps put forward high requirements for the movement ability of inspection robots.Quadruped robot has high mobility and good obstacle crossing ability,and has a good application prospect in oil and gas station inspection.However,there are many difficulties in the research of quadruped robot in balance control,gait speed control and motion decision-making in complex environment.At present,the relevant control theory is not mature.Therefore,studying the base motion algorithm of quadruped robot and exploring efficient and flexible obstacle crossing strategy will effectively improve the intelligent management level of oil and gas station.This paper mainly studies the motion control method of quadruped robot.The robot control system is divided into three control levels: high,medium and low.The motion control method of quadruped robot is studied for each control level.This paper mainly studies the motion control of quadruped robot,divides the robot control system into three control levels: high,medium and low,and carries out the research on the motion control method of quadruped robot according to the theories of robot kinematics,central pattern generator(CPG)and deep reinforcement learning(DRL).Firstly,the single leg kinematics model for the base control of the robot is established,the forward kinematics formula and velocity Jacobian are established by using the improved D-H coordinate method,the inverse kinematics equation of the three degree of freedom link is derived by using the geometric method,and the existence of the inverse kinematics solution and the singularity of the velocity Jacobian are analyzed.Secondly,integrate the single leg kinematics equation,establish the overall coordinate system based on the body,and adjust the rotation matrix and homogeneous transformation matrix to realize the three degree of freedom attitude control of the robot.Based on foot trajectory planning and CPG method,a new combined control method is designed.The test shows that this method can realize the basic walking function of the robot,and the free conversion of walk and trot gait can be carried out in the process of walking.In addition,a stair climbing control strategy is designed,and the modeling and execution of single step action in the strategy are realized.The climbing of stairs with determined specifications is realized in the simulation.Thirdly,the DRL algorithm is applied to the robot patrol simulation control,the depth network model of the improved DQN algorithm,and the action reward function is designed and adjusted according to the robot learning effect in the simulation environment.The target tracking task under one-dimensional and two-dimensional state parameter input is realized successively in the simulation environment.Finally,the physical prototype is used to verify the single leg and overall control of the robot.The test shows that the physical prototype can turn,traverse,height adjustment and gait conversion.At the same time,the stair climbing experiment shows that the control strategy proposed in this paper can make the robot pass through the hollow stairs with a step height of 0.2m and a slope of 35 °.