Position and differential kinematic neural control of robot manipulators: a comparison between two schemes

Two schemes for kinematics robot control based on neural networks are presented and compared. The first scheme is straightforward-it tries to find directly the inverse coordinate vector given a desired cartesian position. The second scheme is a differential approach-it tries to find the inverse Jacobian using a context network and a cartesian feedback. This scheme takes advantage of the function decomposition of the differential kinematics, since we are using information about the structure of the kinematic equations. The neural networks used are three layer feedforward networks trained with a modified backpropagation error algorithm, whose hidden layer is nonlinear (based on sigmoid or gaussian neurons), and its output layer is linear. We simulate both schemes and compare their results, concluding that the first scheme provides a good approximation for the kinematic problem without cartesian position feedback; while the differential scheme combined with a good cartesian coordinates measurements, gives a high precision robot motion