We train a CycleGAN model that will generate "realistic" augmented images based on images coming from the Duckietown simulator. This is in an attempt to reducing the reality gap when transitioning from robot training in simulation to real life.