Recently, ROV and AUV are often used in the maintenance of the offshore structures and the development of the ocean resources. These operations are carried out with the underwater manipulator. In this paper, we studied about the control system for the underwater manipulator considering the operations mentioned above. We try to apply the Learning Feed-Forward Controller (LFFC) to the Hybrid Position and Force control system of the under water manipulator. This control system consists of a feed-forward controller based on an inverce-dynamics model of the controlled system and a feedback controller like a Proportional, Integral and Derivative controller (PID). And this control system is adaptable to the nonlinear dynamics of multi-link manipulator and the change of dynamics both of the manipulator and of the controlled object. For the learning of the inverse-dynamics model we proposed a new method in order to hasten the learning speed. The experiments were carried out, it is shown that the new learning method succeeds in quick learning and that the control system is adaptable to the change of dynamics according to payload.