The centimeter number is real, but only under conditions that rarely survive a production floor. In an open hall with clear line of sight, four well-placed anchors, and two-way ranging, I have seen repeatability around 5 to 10 cm and dynamic RMSE in the same range. Move the same tag into a metal rack aisle and the number degrades fast — 20 to 30 cm RMSE is typical, and worse, the error becomes a stable bias rather than noise, so averaging does not fix it.
The reason is first-path detection. UWB ranging measures the leading edge of the pulse, and in a metal environment the direct path is often blocked or attenuated while a reflected path arrives a fraction of a nanosecond later. The receiver then locks onto the reflection and reports a position that is consistently off by tens of centimeters. That is the NLOS trap LunaWei mentioned: a stable but wrong position is harder to catch than a dropout, because the system looks healthy.
Three things have helped me most. First, anchor geometry: keep anchors high, spread out, and away from metal shelving so the PDOP stays low, and avoid putting all anchors on one wall. Second, use two-way ranging rather than TDOA where you can, because it is less sensitive to anchor clock sync drift, and log the channel impulse response so you can see whether the first path is actually present. Third, never mount a tag directly on metal — even a 5 to 10 mm spacer changes the antenna pattern and the reported range noticeably, and calibrating the antenna delay per tag type matters more than the module datasheet suggests.
For acceptance, I would report fixed-point repeatability, a dynamic trajectory RMSE along a known path, and the share of NLOS-flagged fixes, separately for the open area and the rack area. If the open-area number is 10 cm and the rack-area number is 30 cm, that is honest — and it tells you where to add an anchor rather than where to tune an algorithm.