Collision phenomena are very fast and nonlinear, thus, it is difficult to control a manipulator by collision phenomena. Therefore, in the past, manipulators moved slowly in order to avoid collision. However, the need for high-speed operation has been increasing, making it is indispensable to control manipulators by collision phenomena. With such fast phenomena, it is effective to use learning control in a forward manner. In this paper, we have proposed a learning control method to optimize the weighted least-squares criterion of learning errors. This method is applied in order to obtain a unique control gain by the Riccati equation which has a state dimension equal to the sampling number. It is shown that the convergence of learning error can be readily assured because the present learning rule consists of a steadystate Kalman filter. Based on this learning control method, experimental results of force control with a collision phenomena are reported.