Quadruped robot motion control method based on reinforcement learning and position increment

Abstract

The invention provides a quadruped robot motion control method based on reinforcement learning and position increment, which relates to the field of quadruped robot control and comprises the following steps: acquiring motion environment information, quadruped robot attitude information and sole position information; on the basis of the acquired information, generating a sole position in each preset time step when the quadruped robot moves, and calculating the variation of the sole position in each time step; with the maximum moving distance in a single time step as a constraint, accumulating the time steps to obtain a sole position track; controlling the quadruped robot to execute corresponding actions on the basis of the plantar position track in combination with a preset reward function, so as to enable the quadruped robot to keep motion balance; in order to solve the problem that motor damage is caused by large sudden change of the position of a motor generated in an existing quadruped robot motion control method, sudden change of a control command is avoided by restraining the foot bottom position variable quantity of the quadruped robot in each time step, and the capacity of the quadruped robot for passing through the complex terrain is enhanced.

Description

technical field [0001] The invention relates to the field of quadruped robot control, in particular to a quadruped robot motion control method based on reinforcement learning and position increment. Background technique [0002] Quadruped robots are widely used in surveillance patrols, environmental reconnaissance, and transportation supply scenarios. On the other hand, the flexibility and adaptability of quadruped robots make their dynamic characteristics more complex, which makes it a great challenge to realize animal-like motion of quadruped robots. The basis and premise of the realization of the specific functions of the robot. [0003] The motion control of quadruped robots is mainly divided into two categories: model-based methods and reinforcement learning-based methods. [0004] (1) Traditional modeling control method [0005] The traditional modeling control method performs feature extraction according to the robot state information and obtains valuable informati...

AI Technical Summary

Problems solved by technology

However, due to the nonlinearity of the neural network, the position of the motor directly generated by the neural network will undergo a large mutation, and the motor needs to output a huge torque to track the target position, which will easily cause physical damage to the motor

Although this problem can be alleviated by introducing a constrained reward function for motor output torque or speed, this also greatly increases the difficulty of reward function design and parameter adjustment, making it impossible for methods based on reinforcement learning to obtain good performance in motion control. Strategy

Images