To solve the hysteresis caused by using torque sensors to control muscle force training equipment, a joint torque prediction model based on a group of antagonistic surface electromyography (sEMG) is designed in this article. Firstly, the rehabilitation training equipment is built to provide conditions for signal acquisition and experimental verification. sEMG is preprocessed and the variance characteristic of sEMG signal is selected as the neural network input. In addition, a dynamic recurrent neural network with the nonlinear auto-regressive model with exogenous inputs (NARX) is used in this study. A multi-step ahead prediction model (MSA) based on the actual values of joint moments and another model based on model prediction output ( MPO) are developed respectively. The torque prediction performance of MSA and MPO models is compared by isotonic and isometric test experiments. Experimental results show that there is a strong correlation between the predicted output value and the actual output value of the two models ( Pearson correlation coefficient is greater than 0. 95). As the number of advance prediction steps increases, the prediction accuracy of MSA model decreases. However, the advance prediction time increases. When n is less than 29 and 35, the prediction accuracy of MSA is significantly higher than that of MPO (p <0. 05). But the MPO model has advantages in cost and size. In summary, two models proposed in this article can accurately predict joint torques. In actual rehabilitation training equipment control, different torque prediction models can be selected according to application requirements.