Semi-autonomous teleoperation based on Learning from Demonstration is an effective method of remote operation of space manipulators, especially in the scenarios with limited communication and repeated operation problems. However, the joint trajectories generated by those methods may not be suitable for space manipulator control. In this paper, we present a novel semi-autonomous teleoperation method, which has been used for the TianGong-2 manipulator system. The proposed method can not only reproduce trajectories for the tasks according to the current environment, but also generate the smoother and smaller joint control torques. To implement the method, we first collected kinesthetic demonstrations by the space manipulator teaching platform as prior knowledge. Then, based on these kinesthetic demonstrations, we designed the joint control commands with Dynamics Constraint Learning from Demonstration algorithm. We finally evaluated our method with the simulation and on-orbit experiment by locating the dexterous hand to the pre-screwing bolt. Our results show a significant reduction of joint control torque fluctuations and peak-to-peak values, and also can reduce energy consumption.