Indoor radio-map prediction relies on pathloss models, but real-world measurements are too scarce to train predictors that generalize across many buildings. A new arXiv paper suggests that large-scale synthetic pretraining could address this bottleneck, offering a way to build useful models without exhaustive data collection.
The catch, according to the abstract, is that real building layouts are also in short supply. That scarcity may limit how well synthetic pretraining can transfer to actual environments, since the model still needs realistic spatial structures to learn from.
The paper is at an early stage, with the abstract only outlining the problem and proposed direction. Further details on the synthetic data generation method or experimental results are not included in the available text.