Motion Control Of Antagonistic Bionic Joint Driven By Pneumatic Muscles Actuators Based On Reinforcement Learning

The hand is a highly intelligent organ and an important link in the interaction between human and environment.With the increasing demand for robot technology in society,the research and development of humanoid robots has developed rapidly.Driving technology is the key technology of bionic hand research and development.Most manipulators are mostly driven by DC motors,but due to the inherent characteristics of DC motors,these bionic hands are of high quality and have low safety and flexibility.In order to have supple joint properties,many researchers have focused on artificial muscles.At present,as a new type of standard industrial actuator,pneumatic muscles actuators is widely used as a driving element for objects such as bionic robots and manipulators because of its high power /weight ratio,good flexibility,low cost,and energy saving and other characteristics.In this paper,from the perspective of sports biomechanics,inspired by the skeletal muscle structure of human hands,a bionic joint is designed.The joint uses pneumatic muscles actuators to simulate human muscle characteristics to provide driving force for the joints,and the transmission mode is tendon transmission.By modeling and analyzing the joint,the reinforcement learning algorithm is applied to the joint to realize the motion control of the joint.Firstly,a single pneumatic muscles actuator characteristic test system is established.The static characteristics of the pneumatic muscles actuator is tested under the condition of constant pressure.A mathematical model of the single pneumatic muscles actuator is established and the characteristics of the pneumatic muscles actuator is analyzed.Based on this,a static model and a dynamic model of this bionic joint is established.According to the variable stiffness characteristics of the joint,a joint stiffness model and a joint position/stiffness' s calculating model are established.According to the characteristics of the tendon transmission,the conditions for ensuring the tension of the tendon rope during the movement are analyzed.This ensures the consistency between the mathematical model and the actual device state and lays the foundation for the realization of the joint position/stiffness control.Secondly,to solve the problem that the joint is difficult to establish an accurate mathematical model,a deep reinforcement learning algorithm is proposed for the joint motion control.The feasibility of applying reinforcement learning to the motion control of the joint with and without inertia is verified by studying the motion control methods of the joint.For the problem that reinforcement learning algorithm needs to collect a large number of samples in the actual device and interact with the environment for a long time,the idea of combining traditional nonlinear algorithm with reinforcement learning algorithm is proposed.The traditional non-linear algorithm is used to make sample data,shorten the running time of the algorithm,and speed up the training speed of the reinforcement learning algorithm.Finally,the joint control system is designed.A single joint experimental system platform is set up,and the motion control experiments of the joint are carried out.Through the comparison and analysis of experimental data,the accuracy of the mathematical model of the established pneumatic muscles actuator and bionic joints and the effectiveness of applying reinforcement learning algorithms to motion control of the joints is verified.