Their model works on both. RGB only performance is comparable to other depth based methods which is why they emphasized it over the RGB-D version.

Direct paper link for ref: https://arxiv.org/pdf/2312.06741

Seems it can use both, if available.