This is cool work; however, whoever approved this as PUBLICATION should be shot.
No software in sight. This is especially bad since "machine learning" is cited as a cornerstone of the detection. Hardware papers behind paywall or not yet existing (and may never exist). Nothing even remotely useful for somebody to replicate or evaluate.
If this appeared on the web from a company other than Google, everybody would dismiss it as pure marketing fiction.
Maybe I'm being overly harsh, but one of the big problems with 60GHz type stuff is evaluation that what you have is reliably signal. Especially with "machine learning", who knows what the system is latching onto.
Given the accuracy, could it read your fingerprints while you're doing the gesture? From the paper, it looks like the average displacement accuracy is 0.4mm. I wonder if that's accurate enough to read your fingerprints. If it could, you could ensure that only your gestures are recognized and not someone elses!
Well, resolution for fingerprint scanners is roughly 500 dpi [0]; 1 in^2 = 645 mm^2; so.. maybe, I'm not sure!
But, that said, I believe there are many >1 mm^2 resolution features that can basically uniquely identify users, like, the shape of a hand (individual finger geometry, etc), the length of a limb, the particular subcutaneous layout of veins, face geometry, etc. It's interesting to ponder about, at least!
Currently looking around for cheap (and arduino-amenable) single chip radar systems whereby I might try to hack together a more low tech prototype of Soli
edit: this link seems to work https://dl.acm.org/citation.cfm?id=2925953