Traditional rehabilitation training equipment of rigid body can not adapt the specific training needs for different patients, and can not integrate the adjuvant therapy of functional myoelectricity stimulation. To solve the above mentioned problems, a automatic rehabilitative manipulator system with variable rigidity is developed to realize active and passive training. A new adaptive elastic coupling driver model was built. Then the built model was integrated with myoelectric signal torque estimator designed by the nonlinear regression mathematical model and support vector regression analysis. Using digital signal processor NIsbRIO9642 as the control core, the system realized rehabilitation training exercise supported by exoskeleton such as impedance movement, equal length movement and isotonic movement. The experiments show that the developed manipulator can realize automation of multiple training models for patient training with different needs. In addition, the developed manipulator has the features of comfort and safety.