Abstract: An indirect iterative learning control scheme for a general nonlinear system to track a trajectory with a finite time interval is presented, which uses the least square algorithm to identify the linearized model of the nonlinear system in terms of previous experiences of repetitive tracking. A segmented learning scheme is proposed to keep the learning only in the linear approximation region. How to avoid control singularity is discussed. A high efficient modification is proposed to ensure non-zeso of the determinant of the estimated input matrix. The scheme is applied to robot movement imitation. Even without any knowledge about camera-robot relationship, the controller faces no singularity. This might be a new way for robot trajectory programming by means of a camera instead of hard coding.