In order to improve the accuracy of gesture recognition, a two-flow recurrent neural network (RNN) architecture based on dynamic Adam optimization was proposed for gesture classification of surface electromyography (sEMG). The two-flow RNN was used to capture the spatiotemporal characteristics of sEMG signals, and the dynamic Adam optimizer was introduced to adjust the learning rate adaptively, which significantly improvesd the classification accuracy and model convergence speed. In the comparison experiment, the accuracy of the proposed model reaches 92.86%, which is more than 10% higher than that of traditional algorithms and single-stream RNN model, and has good classification performance and good generalization ability. The experimental results fully validate the effectiveness and superiority of the proposed model in sEMG-based gesture recognition, providing strong technical support for high-accuracy and robust electromyography-based interactive systems.