Active engagement of human nervous system in the rehabilitation training is of great importance for the neurorehabilitation and
motor function recovery of nerve injury patients. To this goal, the human motion intention should be detected and recognized in real time, which can be implemented by modeling the relationships between sEMG signals and the associated joint torques. However, present sEMG-torque modeling methods, including neuromusculoskeletal and black-box modeling methods, have their own deficiencies. Therefore, a hybrid modeling method based on the neuromuscular activations and Gaussian process regression (GPR) algorithm is proposed. Firstly, the preprocessed sEMG signals are converted into neural and muscular activations by the neuromusculoskeletal modeling method. The obtained muscle activations together with the associated joint angles are then transformed into the adjacent joint torques by a GPR algorithm to avoid the complicated modeling process of the muscle contraction dynamics, musculoskeletal geometry, and musculoskeletal dynamics. Moreover, the undetermined parameters of neuromuscular activation and GPR models are calibrated simultaneously based on an optimization algorithm designed in this study. Finally, the performance of the proposed method is demonstrated by validation and comparison experiments. It can be seen from the experiment results that, a high accuracy of torque prediction can be obtained using the proposed hybrid modeling method. Meanwhile, when the difference between the test and calibration trajectories is not very big, the joint torques for the test trajectory can be predicted with a high accuracy as well.