This letter proposes a dynamic system approach to learn point-to-point motions while keeping the stability of the dynamic system. The proposed approach is grounded on a Learning from Demonstration (LfD) method based on a neural network, which gets a better reproduction performance while guaranteeing the generalization ability. The proposed approach has been experimentally validated on the LASA dataset and by the “pick-and-place” task of Franka Emika robot, and experimental results demonstrate that: (1) compared with the state-of-the-art results, the trajectory generated by the proposed approach achieves higher accuracy (approximately 24.79%) in terms of the similarity with respect to the demonstration; (2) the proposed approach can handle high dimensional data and learn from one or more demonstrations; (3) the proposed approach can guarantee the performance regardless of the variation of starting points even in the case of high dimensional complex motions.