Looks like a classic case of the famous Jurissac Park quote: "Your scientists were so preoccupied with whether or not they could, they didn’t stop to think if they should."
Triangulation works accurately in open space, however this use case is inside a home with signals being blocked, bouncing, or attenuated by walls, mirrors, appliances, etc. On top of that, the goal is to use as few BLE rx'ers as possible (meaning not one per room). Given those constraints, a trained ML model actually does make sense.
"ML" is just a very overstated word for: "I have a handful of example data and a library that does basic math things"
The main difference between stats and ml is (i) community, (ii) stress on the details of compute and the focus on prediction accuracy rather than accurate recovery of parameters. A scheme that ensures good prediction even if the parameter is not recovered, for ML that's still a success.