From my understanding, the authors meant zero-shot communication (in the title) or stitching (in the text), where two NN components trained in different setups can be stitched together without further finetuning. This is just one useful application of the shared relative representation proposed in the paper.