We propose a new statistical modeling approach that we call Sequential Adversarial Auto-encoder (SAAE) for learning a synthesis model for motion sequences. This model exploits the adversarial idea that has been popularized in the machine learning field for learning accurate generative models. We further propose a conditional variant of this model that takes as input an additional information such as the activity which is performed in a sequence, or the emotion with which it is performed, and which allows to perform synthesis in context.