Abstract: To solve the problem that the Kinect field of view cannot simulate the human body walking for a long time, a method based on the long short-term memory (LSTM) neural network prediction model to generating the cyclic gait of the humanoid robot was proposed. The one-dimensional time series of each joint angle was collected by Kinect when the human body walked, and the joint angle driving sequence of the humanoid robot was obtained through the humanoid gait balance model. The C-C method was used to determine the time delay and embedding dimension of the time series, and the phase space reconstruction of the joint angle sequence was used to obtain more eigenvalues of the time series. The joint angle prediction model based on the LSTM neural network was trained and passed by generating a sequence of joint angles for multiple gait cycles. Finally, the generated sequence was used to drive the NAO on WEBOTS platform to complete the walking motion of multiple gait cycles. This method effectively solves the problem of the visual field limitation of the somatosensory camera, enables the humanoid robot to complete the walking simulation of multiple gait cycles.