Quadruped robot control method based on model predictive control optimization reinforcement learning

Abstract

The invention relates to the field of robot intelligent control, in particular to a quadruped robot control method based on model predictive control optimization reinforcement learning. The method comprises the following steps: establishing a dynamic model according to physical parameters of a physical model machine, and converting the dynamic model into a state-space equation; optimizing model predictive control according to the state-space equation, and deploying the optimized model predictive control on the physical model machine; and establishing a reinforcement learning model, wherein the reinforcement learning model interacts with environment and model predictive control to train the physical model machine at the same time. According to the invention, meaningless data generated in a training process is reduced based on model predictive control optimization reinforcement learning, the requirement for computing power is reduced through model guide training, and the method can be directly deployed in physical model machine training to reduce the training process, so that the problems that an algorithm based on value and strategy has high requirements for data and computing power, expensive computer equipment is needed for pre-training to preliminarily achieve control effect of a robot, and long-time training is needed after the algorithm is deployed to a physical model machine are solved.

Description

technical field [0001] The invention relates to the field of robot intelligent control, in particular to a quadruped robot control method based on model predictive control optimization reinforcement learning. Background technique [0002] Common quadruped robots have twelve degrees of freedom and complex structures. Quadruped robots perform better than wheeled robots on unstructured terrain, so the working environment is often in an unpredictable situation, and traditional control algorithms are difficult to adjust to adapt. Therefore, reinforcement learning is used in the control strategy of quadruped robots. Its self-learning ability can reduce the difficulty and cost of development while improving its adaptability. The reinforcement learning of quadruped robots is mostly a model-free algorithm based on value and strategy. It has extremely high requirements on data and computing power, and often requires expensive computer equipment for pre-training to initially achieve th...

AI Technical Summary

Problems solved by technology

[0005] In order to solve the problem that the value-based and strategy-based algorithms in the prior art require extremely high data and computing power, expensive computer equipment is required for pre-training to achieve the initial control effect of the robot, and a long period of time is required after deployment to the physical prototype training problem

Images