Abstract: The complex spatial-temporal structure within human motion capture data often makes it extremely challenging to adapt the motion generation fields, ranging from data-driven character animation, sequence splicing and motion style fusion. Inspired by the popularity of deep learning theories in computer graphics, four kinds of deep generative models（i.e., restricted Boltzmann machine, recurrent neural network, convolutional neural network, deep reinforcement learning） and their hybrid models, are comprehensively surveyed for human skeletal motion generation, including studying their topology structures and theoretical optimizations. Meanwhile, we investigate the superiorities of these representative deep learning models in extracting the spatial-temporal motion features, and quantitatively compare their performances in different motion generation tasks. Finally, we carefully survey their potential challenges in complex motion generation, and discuss the future trends of recent deep learning models on robust motion generations.