Walking Pattern Of Biped Robot Based On Neural Network Learning

The walking pattern design of biped robot has always been the focus of robot research.In order to realize the stable walking pattern of robot,this paper mainly uses the neural network with strong learning ability to plan.The subproblems are designed in detail,including trajectory generation in operating space,inverse kinematics of trajectory in joint space and dynamics of robot control.Finally,the effectiveness of the relevant network was verified by the simulation experiment of NAO robot.The main work and innovation points are summarized as follows:A dynamic control neural network system(DCNN)is designed to solve the trajectory generation problem of robot operating space.The whole space trajectory generation task is divided into two parts: time and space.In the first part,the intermediate trajectory with time characteristics is generated by the recurrent neural network(RNN).In the second part,the intermediate trajectory is mapped to the desired spatial shape by the feed forward neural network(FFNN).At the same time,the external adjustable signal and feedback control loop are added.Quadratic programming of trajectory space mapping is carried out by external adjustable signal,and the feedback control loop is self-adaptive for the whole system.A neural network solver for inverse kinematics(LMP-ZNN)was designed to solve the trajectory generation problem of robot in joint space.The inverse kinematics of Jacobian matrix is solved by Zhang neural network(ZNN).A special recurrent neural network(RNN)with implicit dynamics and a matrix error function are used to build the model.At the same time,facing the Moore-Penrose inverse problem of time-varying matrices in trajectory planning of biped robots,the ordinary matrix inverse is extended to the generalized matrix inverse(Moore-Penrose inverse).Then,ZNN is applied to the solution of the Moore-Penrose inverse to solve the singularity and real-time control problems in trajectory planning.Then Levenberg-Marquardt algorithm is introduced to realize the robustness of inverse kinematics solution by increasing damping factor and changing it in real time.Finally,the neural network solver of inverse kinematics is planned.Aiming at the dynamics problem of robot motion control,it is difficult to analyze the dynamics of complex or uncertain models,and there are some dynamics errors in driving the robot motion.A neural network adaptive controller with robust control is designed.The controller uses the general approximation property of the feed forward neural network(FFNN)to estimate the nonlinear dynamics of the robot by the way of adaptive learning,and its weight is updated adaptively.Adaptive control is used to solve the model and tracking error problems.