Neural Network Controller for a Stewart Platform

摘要

The report summarizes efforts to use cascaded neural nets to control a six leggedparallel link manipulator in the high speed, high accuracy domain. The Cerebellar Model Computer (CMAC) network structure has been extended from a standard single CMAC net to a cascaded network with coarse and veneer nets. The coarse and veneer networks have different input sensor quantization numbers and different output ranges. Together they can provide faster learned capability and the ability to capture both general trend and the fine details of unknown nonlinear mappings. The cascaded network structure was used to solve the forward kinematic problem for a Stewart platform. Both CMAC and back propagation networks can be trained to provide approximate forward mapping throughout the work space and the former needs fewer training iterations than the latter. Researchers investigated the use of the cascaded CMAC network for the dynamic control problem of Stewart Platform type robotic manipulators. A number of simulations were done to evaluate the performance of CMAC based controllers for parallel link mechanisms. Different generalization, quantization, learning rate, and memory size parameters were investigated. The use of CMAC networks for dynamic error compensation for multi-degree-of-freedom mechanisms was investigated and simulations show that the CMAC network can reduce the kinematic error to about one-third of the original error generated by a nominal approximate kinematic model.