On solving the inverse kinematics problem using neural networks

Writing and solving the inverse kinematics equations of a robot is a cumbersome task. Furthermore, an analytic solution exists only for an ideal model and only if the structure of the robot meets certain criteria. If enhanced positioning precision is required, the robot needs to be calibrated. This eliminates most of the errors due to the differences in the ideal model and the real robot. Calibration methods require additional computations (usually of iterative nature) in the real-time cycle of the control system and are only capable of dealing with small differences between the ideal model and the real robot. In this paper a supervised learning based approach is proposed to solve the inverse kinematics problem and the calibration. Instead of creating and ideal model for a series of robots and calibrating each of them individually afterwards, the inverse kinematics function is learned using a neural network and so, it is tailored to one given robot, already including errors due to manufacturing and/or assembly tolerances. Moreover, it can also work for structures for which no analytic solution is possible. The preliminary results of this research are presented in this paper, covering two simple robot structures, a planar 2 DOF and a spatial 3 DOF structure, both with and without artificially introduced assembly errors (joint misalignments) which make analytical modeling unfeasible.