The control of a robot with unactuated joints is much more difficult than a conventional robot. Because the unactuated joints can't be controlled directly, it is an incomplete controllable system. Usual smooth feedback control is not usable for such system. A method of learning control is proposed in this paper, that is dynamics feedforward control by learning. In this way, we can control an under actuated robot with high precision. The effectiveness of the method has been confirmed by experiment with an under-actuated robot. Method, implementation and experimental results of the control method are described in this paper.