Neural networks for learning inverse-kinematics of redundant manipulators

The problem of learning, the inverse-kinematics of redundant manipulators is considered. It is argued that this is an ill-posed, one-to-many-type problem with infinite solutions. The problem is formulated as a constrained optimization problem to result in a unique solution. Using language multipliers, an energy function is defined to convert the latter into an unconstrained optimization problem. A learning algorithm for the adjustment of the connection weights in the network to minimize the energy function is presented. The trained network, upon receiving end-effector´s desired velocity, will produce unique minimum-norm velocity required for the joints