Device-free wireless sensing has recently attracted significant interest due to its potential to support a wide range of immersive human-machine interactive applications. However, data heterogeneity in wireless signals and data privacy regulation of distributed sensing have been considered as the major challenges that hinder the wide applications of wireless sensing in large area networking systems. Motivated by the observation that signals recorded by wireless receivers are closely related to a set of physical-layer semantic features, in this paper we propose a novel zero-shot wireless sensing solution that allows models constructed in one or a limited number of locations to be directly transferred to other locations without any labeled data. We develop a novel physical-layer semantic-aware network (pSAN) framework to characterize the correlation between physical-layer semantic features and the sensing data distributions across different receivers. We then propose a pSAN-based zero-shot learning solution in which each receiver can obtain a location-specific gesture recognition model by directly aggregating the already constructed models of other receivers. We theoretically prove that models obtained by our proposed solution can approach the optimal model without requiring any local model training. Experimental results once again verify that the accuracy of models derived by our proposed solution matches that of the models trained by the real labeled data based on supervised learning approach.