Accelerated convergence in the inverse kinematics via multilayer feedforward networks

A hybrid approach to the iterative solution of robotic manipulators is presented. In this method a multiplayer feedforward network (MFN) is trained to provide an approximate solution to the inverse kinematic problem of a robot. This approximate solution is then utilized as the initial guess to the iterative procedure which provides the solution within some specified tolerance. Examples involving the PUMA 560 robot are given. Since the time taken by the MFN to provide the initial estimate is only a fraction of the time taken by the iterative procedure taken to reach a solution, this combination can be usefully employed in order to decrease the time for obtaining the solution to the inverse kinematic problem in robotics.<>