The joint torque prediction plays an important role in rehabilitation medicine, clinical medicine, sports training and other fields. The continuous and real-time torque prediction can make the human-computer interaction equipment better feedback and reproduce the intention of human motion. To provide a safe, active and comfortable rehabilitation training environment for patients and enhance the compliance of the human-computer interaction equipment, a novel method of joint torque prediction is proposed, which is based on an improved recursive cerebellar model neural network. In this method, muscle synergy analysis is used to reduce the dimensionality of surface electromyographic (sEMG) signals. Then, the reduced-dimension sEMG feature vector, joint angular velocity and joint angle are used as the input data of the prediction model. In addition, recursive unit and fuzzy logic rules are introduced into the cerebellar model neural network, while the wavelet function is used as membership function. Hence, the generalization ability of the network is optimized. The RWFCMNN model realizes the time series prediction of the dynamic torque of ankle dorsiflexion and plantarflexion in three states, non-fatigue, transitional fatigue and fatigue. The average Pearson correlation coefficient and the average normalized root mean square error between the predicted torque and the actual torque are 0. 933 5 and 0. 159 8, respectively. These numerical values verify the accuracy and effectiveness of this method for continuous prediction of lower limb joint torque.