Temporal human pose prediction from a 3D human skeleton sequence is vital for robot applications such as autonomous control and human-robot interaction. Recent pose prediction methods generally make predictions using GCN. However, because all frames of human poses are processed at once using a GCN, it is necessary to wait for pre-processing until all input frames are available, and intermediate predictions cannot be obtained until post-processing is complete. In addition, when predicting not in the time domain but in the frequency domain, if the input/output time is less than a few seconds, the number of sampling points is extremely small and the frequency resolution is low. In this study, we propose Recurrent Graph Convolutional Network (RGCN) and its application to a pose prediction. The advantages of RNN and GCN for sequential predictions in the time domain without frequency transformation are combined into RGCN to address the problem of existing methods. Through evaluation on public datasets, we confirmed that the accuracy of the proposed model using RGCN, which predicts sequentially in the time domain with simple processing and fewer parameters, is comparable to that of latest prediction method.