Quadruped robot gait control method based on reinforcement learning and CPG controller

Abstract

The invention discloses a quadruped robot gait control method based on reinforcement learning and a CPG controller. The quadruped robot gait control method comprises the steps that S1, building a single-leg model and an overall model of a quadruped robot; S2, constructing an actor neural network and a reviewer neural network to train the single-leg model; S3, determining a relative sequential relationship between legs of the quadruped robot and constructing a controller model according to leg movement characteristics; and S4, constructing a foot end motion trail model of the quadruped robot from an output signal of the controller model through mapping transformation, and taking the output of the foot end motion trail model as a gait control signal of the quadruped robot to drive the quadruped robot. According to the method, it is guaranteed that legs of the quadruped robot move strictly according to the time sequence, the gait characteristic of mammals is achieved, meanwhile, it is guaranteed that the trunk of the robot is stable in the moving process, the pitch angle and the lateral angle change range is small, and overturning is effectively prevented; the leg movement has self-adaptability, and a complex control and planning algorithm does not need to be artificially designed.

Description

technical field [0001] The invention relates to the technical field of robots, in particular to a method for controlling the gait of a quadruped robot based on reinforcement learning and a CPG controller. Background technique [0002] In the field of gait control of quadruped robots, the central rhythm controller CPG is one of the simplest and most classic control methods. First, a suitable rhythm controller is constructed to assign a controller to each leg of the quadruped robot, so that the robot Exercise according to a certain sequence, and complete simple actions such as forward or backward. However, using this method has great limitations: 1. It is difficult to further modify and intervene in the leg motion planning after the controller is set. 2. Since only the timing problem is considered, the setting of the joint motion of the robot is not necessarily reasonable, resulting in a large change in the pitch angle and lateral angle of the robot torso during the motion pr...

AI Technical Summary

Problems solved by technology

However, using this method has great limitations: 1. It is difficult to further modify and intervene in the leg motion planning after setting up the controller

2. Since only the timing problem is considered, the setting of the joint motion of the robot is not necessarily reasonable, resulting in a large change in the pitch angle and lateral angle of the robot torso during the motion process, which does not meet the requirements for smooth motion. robot capsize

But there are still some deficiencies: 1. The training neural network is too large and the reproducibility is low; these robot motion control networks are often trained by a large number of engineers and researchers for several months to achieve the effect. Reward and punishment functions are also extremely complex and difficult to transfer and reproduce; 2. The footed robots trained by existing reinforcement learning algorithms have abandoned a characteristic of footed robots: timing

The movement of the legs according to a certain sequence is a characteristic of mammals, and it is also a necessary condition for mammals to show faster and stronger ground adaptability. Robots controlled by reinforcement learning algorithms that remove the sequence will have poor adaptability The problem

Images