In recent years, robotics is starting to expand beyond industrial applications, and robots are starting to take part in tasks that require interacting with human beings. For this interactions to be natural for the users, it is necessary that the robots are capable of performing expressions autonomously. In situations where the robot is speaking, the non-verbal gestures performed by the robot must also support the communicative message expressed by the verbal component, and both components should be properly synchronized. In this work, we present a gesture prediction system for social robots based in one of the most significant advances in the area of deep learning: the transformer model. This solution will be compared with a previous system based on a combination of recurrent neural networks and conditional random fields. The results of the comparison conducted show that, as it is the case for other tasks in the field of natural language processing, transformers present a clear improvement for the task of predicting non-verbal expressions for social robots.