FORWARD KINEMATICS SOLUTION OF STEWART PLATFORM USING NEURAL NETWORKS

The Stewart platform's unique structure presents an interesting proble m in its forward kinematics (FK) solution. It involves the solving of a series of simultaneous non-linear equations and, usually, non-unique , multiple sets of solutions are obtained from one set of data. In add ition, most effort usually result in having to find the solution of a 16th-order polynomial by means of numerical methods. A simple feed-for ward network was trained to recognise the relationship between the inp ut values and the output values of the FK problem and was able to prov ide the solution around an average error of 1.0 degrees and 1.0 mm. By performing a few iterations with an innovative offset adjustment, the performance of the trained network was improved tremendously. Two ext ra iterations with the offset adjustment reduced the average error of the same trained neural network to 0.017 degrees and 0.017 mm.