Research On The Motion Stability Of Quadruped Robots On Slopes Based On Reinforcement Learnin

The motion mode of the foot robot has the characteristics of discontinuous contact between the foot and the ground,which makes the foot robot able to adapt to the complex movement environment,but also leads to the poor motion stability of the foot robot.One of the problems that need to be solved is to maintain the motion stability of legged robots in the uneven terrain.The slope terrain is a typical non-flat terrain,and four-legged robots often need to move in the slope environment.Aiming at the above problems,the motion stability of quadruped robot in slope environment is studied in this paper.Aiming at the problem that the projection of the body’s centroid on the supporting surface is far away from the intersection of diagonal supporting lines,which affects the walking stability of quadruped robot,a method to adjust the robot’s posture and equivalent leg length parameters is proposed.Webots simulation software was used to verify the simulation results.The simulation results showed that the roll Angle and yaw Angle of the quadruped robot on slope walking could be reduced by the method of centroid adjustment.The roll Angle variance of the robot was reduced by 15% and yaw Angle variance was reduced by 55% compared with that before adjustment.The Central pattern generator(CPG)control network constructed by Hopf oscillator is used as the motion control method of quadruped robot.The vestibular biological reflection mathematical model is introduced into the CPG control network,so that the slope movement of quadruped robot can adjust the robot posture according to the slope Angle information.To stabilize the body’s center of gravity and improve the motion stability of the robot’s slope motion.The combination of traditional control method and reinforcement learning method is the mainstream method for robot motion control at present.In this paper,CPG and reinforcement learning are combined to design two reinforcement learning control methods.Through comparison of simulation results,both methods can improve the motion stability of robots in slope environment.Reinforcement learning control method based on foot trajectory planning converts the output x and y of Hopf oscillator into the foot trajectory of quadruped robot through the conversion formula.This foot trajectory is taken as the reference trajectory of reinforcement learning,and Deep deterministic policy gradient(DDPG)is used for learning to improve the motion stability of the robot’s slope environment.Through Webots simulation verification,slope Angle is set to 20°.The simulation results show that: after training,the pitch Angle variance decreases by 69.2%,roll Angle variance decreases by 59.0%,and yaw Angle variance decreases by 58.0% when the robot goes uphill.When the robot goes downhill,the pitch Angle variance decreases by 8.9%,roll Angle variance decreases by 14.9% and yaw Angle variance decreases by 33.3%The reinforcement learning control method based on the optimization of CPG parameters adopts the CPG control network with vestibular feedback as the bottom controller and DDPG as the high controller to improve the stability of the robot’s slope environment motion by optimizing the CPG network parameters.Webots simulation software was used for simulation verification,and slope Angle was set at 20 °.The simulation results showed that: after training,the pitch Angle variance,roll Angle variance and yaw Angle variance were reduced by 48.7%,67.0% and 50.0% when the robot went uphill.When the robot goes downhill,the pitch Angle variance,roll Angle variance and yaw Angle variance decrease by 45.2%,32.0% and 48.8%.