Computer vision in robotics makes heavy usage of RGB-D data. However, collecting large manually annotated datasets is extremely time-consuming and therefore costly. A potential solution is to automatically generate synthetic datasets and to use them in order to make predictions on the real data. Nevertheless, the domain shift between the synthetic dataset (source domain) and the real data (target domain) partially invalidates the effectiveness of this solution, yielding an accuracy significantly lower than the one that would be obtained using labelled real data. In order to overcome this issue, multiple domain adaptation methods have been developed. These methods can also be employed in a multimodal scenario like RGB-D, but none of them exploits the existing relationship between modalities.