To improve the ability of overcoming the spatial disturbance

and reduce the joint torque fluctuations and energy consumption during operation

a learning strategy from demonstration based on dynamics constraints for space manipulators is proposed. This strategy is divided into two phases. Phase 1 is Gaussian process-based learning from demonstration

in which the motion model of the task is obtained by utilizing Gaussian process based on the kinesthetic demonstrations. Then

the desired trajectory distribution of the current task is reproduced using the model according to the environment. Phase 2 is the design of dynamics-constraint-based controller. The input of this controller is the trajectory distribution from phase 1

and the outputs are the joint desired torques. This controller is used to generate smoother joint control torques

while ensuring that the trajectory of manipulator can meet the task requirements. Finally

the strategy is verified by the on-orbit locating bolts task with Tiangong-2 space manipulator. Compared with the strategy of traditional learning from demonstration combined with computed torque controller

the joint torques peak-peak value of the large load joint is reduced by 45%

the number of peaks is reduced by 40%

and the energy consumption is reduced by 31%. Besides

the joint torques

accelerations and velocities are much smoother.