I have looked at some of the literature on WiFi sensing ((ab)using consumer WiFi equipment as janky radar). It has some weird quirks. Almost everyone seems to use hopelessly outdated hardware for no obvious reason - ESP32s (there was a paper on this) and Intel AX200s can dump CSI information. Also, while plenty of work uses deep learning and some of it is even using modern architectures, nobody seems to have caught onto the Bitter Lesson. I saw one paper doing useful transfer learning - and even then their "pretraining" was 20 minutes on a consumer GPU. Everyone uses tiny (<100k samples) datasets. Probably one organization willing to deploy a lot of devices for data collection and more than 20 minutes of computing power could beat all previous work and make something people actually want.