Research On The Inverse Kinematic Problem Of Robot Manipulators Based On Neural Networks

Solution to the inverse kinematics of robot manipulators is one of importantprocess to control it. But it is difficult to find common methods because ofsophistication of the problem. In this thesis, based on the ability of neural networksapproximating a nonlinear function, two kinds of typical neural networks are applied tosolve the inverse kinematic problem of manipulators.Firstly, research on the inverse kinematic problem of manipulators based onfeed-forward neural network. It was approved that three layers feed-forward neuralnetwork with a hidden layer of sigmoid function could approximate any continuousmultivariate real-valued function to any degree of accuracy. BP network is the typicalfeed-forward neural network. But conventional BP algorithm has defects on slowconvergent speed and easy convergence to a local minimum point of error function.After revealing the relations between active functions and convergent speed andaccuracy, we propose a kind of improved BP method. Then, on the base of improvedalgorithm, we establish the inverse kinematic model of manipulators through choosingreasonable type of active functions and different learning rates in each layer.Simulation experiments demonstrate that proposed method is effective to solve theproblem, and improves precision of solution as compared to the conventional BPalgorithm. Also its solution is the same accuracy as genetic algorithm.Secondly, research on the inverse kinematic problem of manipulators based onrecurrent neural network. We adopt second order recurrent neural network to study theinverse kinematic problem of three degree-of-freedom planar redundant manipulators,and put forward the method of neural networks based on the forward equations andJacobian matrix. According to its features, we design network model and givecorresponding algorithm. Simulation experiments show that the network can solve theinverse kinematic problem of manipulators, and it reaches to good precision ofsolution. The computation speed can meet the requirements for the manipulators' realtime control system.