In this article, contrary to popular studies on human motion learning, we focus on addressing the problem of humanoid robot motions directly. Performances of different kernel functions with principal components analysis (PCA) in Gaussian process dynamical models (GPDM) are investigated to build efficient humanoid robot motion models. A novel kernel-PCA-GPDM method is proposed for building different types of humanoid robot motion models. Compared with the standard-PCA-GPDM and auto-encoder-GPDM methods, our proposed method is more efficient in humanoid robot motion modeling. In this work, three types of NAO robot motion models are studied: walk-model, lateral-walk model, and wave-hand model, where motion data are collected from an Aldebaran NAO robot using magnetic rotary encoder sensors. Using kernel-PCA-GPDM method, the motion data are first projected from the high 23-dimension observation space to a 3-dimension low latent space. Then, three types of humanoid robot motion models are learned in the 3D latent space. Compared with other kernel-PCA-GPDM or auto-encoder-GPDM methods, our proposed novel kernel-PCA-GPDM method performs efficiently in motion learning. Finally, we realize humanoid robot motion representation to verify the motion models that we build. The experimental results show that our proposed kernel-PCA-GPDM method builds efficient and smooth motion models.