In this paper, we study a cold-start heterogeneous-device localization problem. This problem is challenging, because it results in an extreme inductive transfer learning setting, where there is only source domain data but no target domain data. This problem is also underexplored. As there is no target domain data for calibration, we aim to learn a robust feature representation only from the source domain. There is little previous work on such a robust feature learning task; besides, the existing robust feature representation proposals are both heuristic and inexpressive. As our contribution, we for the first time provide a principled and expressive robust feature representation to solve the challenging cold-start heterogeneous-device localization problem. We evaluate our model on two public real-world data sets, and show that it significantly outperforms the best baseline by 23.1%-91.3% across four pairs of heterogeneous devices.