The physiological electrical signal of human body is the direct response from human behavior intention. By analyzing and interpreting the physiological electrical signal of human body, human subject consciousness can be effectively recognized. In order to enable people with physical disabilities to effectively control peripherals according to their own intentions, this research is based on the classification and recognition of sEMG hand movements. The position of sEMG electrode on the muscle is determined. sEMG signals of 8 kinds of common hand movements in daily life are successfully collected and data collection of sEMG signals is completed. The VGG16 is used to construct the convolutional neural network structure suitable for sEMG gesture recognition. At the same time, a highly biomimetic robotic hand is built according to the bone structure of human hand to complete 8 gestures as an actuator, and STM32 microprocessor is used to control it. Through the comparison and analysis with two VGG16 networks with different convolution kernel sizes, the results show that the VGG16 network structure with convolution kernel size of 3×3 adopted by this method has a high accuracy for the recognition of 8 kinds of hand movements.