Beyond BLE: UWB Ranging and Real Localization
BLE tells you a tag is nearby. UWB tells you it is 3.42 metres away — and then multipath tells you it is lying.
🎯 By the end you'll understand
Why converting a received-signal-strength reading into a distance produces errors of a factor of two, and what RSSI is good for. How ultra-wideband measures time of flight, and why 500 MHz of bandwidth is the whole trick. The arithmetic of single-sided two-way ranging, why a twenty-parts-per-million crystal turns it into a three-metre error, and how double-sided ranging cancels that to first order. How anchor geometry multiplies your ranging error into a position error, and why a metal cabinet biases a range long rather than making it noisy. What a real UWB module costs a battery-powered tag. And the honest conclusion for Beacon.
KEY INSIGHTDistance is a time measurement or it is a guess
The tag was reporting a confident 3.42 metres. Moving it ten centimetres to the left made it 3.41. Moving it ten centimetres to the right, behind the corner of a steel tool cabinet, made it 4.31 — and it stayed 4.31, calmly, for as long as anyone watched. It was not noisy. It was consistently wrong.
There are exactly two ways for a radio to measure distance. You can measure how much the signal weakened, and invert a propagation model to get a range — cheap, universally available, and hopeless indoors, for reasons Lesson 4.5 has already given you. Or you can measure how long the signal took, and multiply by the speed of light — expensive, requiring wide bandwidth and precise clocks, and accurate to centimetres when it works. Ultra-wideband is the second approach made practical. Everything interesting about it follows from one number: light travels 30 centimetres in a nanosecond, so ranging to 10 cm means resolving arrival times to about a third of a nanosecond.
UWB's headline accuracy is a line-of-sight accuracy, and its two enemies are not noise. The first is clock error, which a naive ranging protocol converts into metres of systematic bias — and which the right protocol removes almost entirely, for free, by adding one message. The second is non-line-of-sight bias, which is always positive, because a reflected path is longer than a direct one. Both produce answers that look stable and precise and are simply wrong. A centimetre-accurate system with a metre of bias is more dangerous than a metre-accurate system that admits it.
CORE CONCEPTWhy RSSI is not a distance, in one line of algebra
The log-distance model from Lesson 4.5 says received power falls as PL(d) = PL(d₀) + 10·n·log₁₀(d/d₀). Invert it and you get the formula that appears in every "BLE indoor positioning" tutorial:
d = d₀ × 10(P₀ − RSSI) / (10·n)
It is correct algebra applied to a model that does not hold. Differentiate it and the problem is immediate: a measurement error of ΔdB produces a multiplicative distance error of 10Δ/(10n). At an indoor exponent of n = 3, a 10 dB error gives 1010/30 = 2.15×. Your three-metre estimate is somewhere between 1.4 m and 6.5 m.
And 10 dB is not a pessimistic error, it is a typical one. Indoor multipath swings received power by 15–20 dB over displacements of a few centimetres. A human body in the path costs 3–6 dB. Antenna orientation costs more. The reference power P₀ varies between units by several decibels because antennas and PCB assemblies vary. And n is a property of the building that you almost certainly did not measure.
Coarse, monotone, heavily-averaged proximity. "This tag is being heard by the gateway in the east workshop and not by the one in the west store" is a reliable, useful statement — it is a zone assignment, derived from which receivers hear it at all, not from a number of metres. Averaging dozens of readings over minutes and comparing them against a fingerprint survey of the same building can do somewhat better. What RSSI cannot do, at any amount of filtering, is tell you which of two shelves a metre apart something is on. If your product needs metres, you need a time measurement; if it needs rooms, you already have what you need and UWB is an expensive way to buy it twice.
HOW IT WORKSWhat ultra-wideband actually is
Ultra-wideband, in the sense used by IEEE 802.15.4z and the modules you can buy, means transmitting extremely short pulses that occupy at least 500 MHz of bandwidth, in bands around 6.5 GHz (channel 5 is centred at 6489.6 MHz) and 8 GHz (channel 9 at 7987.2 MHz). The power spectral density is very low — this is a signal that lives under the noise floor of narrowband receivers, which is how it is permitted to share spectrum at all.
Bandwidth is the entire point, and the reason is a fundamental one: a signal's time resolution is inversely proportional to its bandwidth. A 1 MHz signal cannot distinguish arrival times closer than about a microsecond, which is 300 metres. A 500 MHz signal has a pulse envelope of a couple of nanoseconds, and a receiver correlating against a known preamble can estimate the leading edge of that pulse far more finely than its width — which is how commercial parts reach the ten-centimetre class in line of sight.
The second consequence of wide bandwidth is the useful one: reflections separate. A narrowband receiver sums the direct path and every reflection into one complex amplitude and calls it a signal — that sum is exactly the multipath fading of Lesson 4.5. A UWB receiver sees the channel impulse response as a series of distinct arrivals separated in time, and can pick the first one, which is the direct path if there is one. That is why UWB is robust where narrowband ranging is not — and it is also why the failure mode is so specific: when there is no direct path, the first arrival is a reflection, and the range is biased long by exactly the extra distance that reflection travelled.
| Technique | Typical accuracy | Infrastructure | Tag energy | Answers |
|---|---|---|---|---|
| BLE RSSI, single reading | Factor of ~2 in distance | None beyond the gateway | Free — it comes with the packet | Near / far, badly |
| BLE, which gateways hear it | Zone-level | One gateway per zone | Free | Which room |
| BLE direction finding (AoA) | Degrees of bearing | Multi-antenna locators | Low — the tag still just advertises | Which direction, from a fixed point |
| UWB two-way ranging | ≈ 10 cm line of sight | Surveyed anchors, wired or powered | Milliamp class while ranging | Which shelf |
Note the fourth column, because it is the one that decides things in this course. A UWB ranging exchange is several messages with receive windows, and a UWB transceiver draws roughly a hundred milliamps while transmitting or receiving. That is not a BLE advertisement's twenty microjoules; it is a different order of device.
HOW IT WORKSTwo-way ranging: the naive version and the one that works
Measuring time of flight one way requires the two devices to share a clock to sub-nanosecond precision, which in practice means wiring them together or giving each a GPS-disciplined oscillator. Two-way ranging avoids that by measuring a round trip on one device's clock and subtracting the other device's processing delay.
Single-sided TWR is the obvious protocol. The tag sends a poll at T1 on its own clock. The anchor receives it at T2, spends T_reply processing, and replies at T3, all on its clock. The tag receives the reply at T4. Then
ToF = (T_round − T_reply) / 2 = ((T4 − T1) − (T3 − T2)) / 2
and the flaw is visible if you look at whose clock measures what. T_round is measured on the tag's crystal; T_reply on the anchor's. If the two crystals differ in frequency by a fractional offset e, the anchor's measurement of its own reply delay is wrong by e × T_reply, and half of that error lands directly in the time of flight. Put numbers on it:
| Crystal offset | Reply time | Time-of-flight error | Range error |
|---|---|---|---|
| 20 ppm | 1 ms | 10 ns | 3.0 m |
| 20 ppm | 100 µs | 1 ns | 0.30 m |
| 2 ppm (TCXO) | 1 ms | 1 ns | 0.30 m |
| 20 ppm, double-sided | 1 ms | < 0.1 ps | < 0.1 mm |
Read the first row again. A ±20 ppm crystal is an ordinary, entirely respectable part, and a one-millisecond reply time is what you get if the anchor's firmware does anything at all between receiving and replying. Together they produce three metres of error on a radio whose whole selling point is ten centimetres. The two classical fixes are to make T_reply very short and tightly controlled — which means doing the turnaround in hardware, not in an interrupt handler — or to measure the carrier frequency offset and correct for it, which good transceivers support.
Double-sided TWR fixes it structurally. The tag sends a FINAL message after receiving the response, so a second round trip is measured with the roles reversed. The estimator combines both round trips and both reply times in a ratio whose leading clock-offset terms cancel:
ToF = (Tround1 × Tround2 − Treply1 × Treply2) / (Tround1 + Tround2 + Treply1 + Treply2)
The residual error is second order in the offset, so 20 ppm becomes 4 × 10⁻¹⁰ rather than 2 × 10⁻⁵ — which is to say, gone. The cost is one extra message per ranging exchange: about 50 % more air time, 50 % more energy, and a protocol that fails differently if the FINAL is lost. Essentially every serious deployment uses it, and any product claiming centimetre accuracy from single-sided ranging with commodity crystals is claiming something it cannot deliver.
SEE ITTWR ranging simulator
A 12 × 8 m room. Place the tag, choose where the anchors were screwed to the walls, pick a protocol, and set the crystal offset and reply time. The widget computes each range with its clock error, its ranging noise floor, and any obstruction bias, then solves for position and reports how far off it is. Three separate lessons are hiding in it, and they arrive in this order: protocol beats hardware, geometry beats protocol, and obstruction beats everything.
Do these in order. One: select single-sided TWR and leave the crystal at 20 ppm with a 1 ms reply. Three metres of error on a centimetre-accurate radio, and because every anchor has its own crystal, each link is wrong by a different amount — so the position lands over a metre from the truth even with textbook geometry. Switch to double-sided: the error collapses to the noise floor, and you did it with one extra message, not a better crystal. Two: stay on double-sided and switch the anchors to "three clustered in a corner". The ranges are still accurate to centimetres and the position error jumps to half a metre or more, because the dilution of precision multiplies it. Nothing about the radio changed. Three: put the steel cabinet in front of anchor A. One range is now biased long by most of a metre, the solver pulls the whole estimate away from that anchor, and the error is not noise — it is a stable, confident, wrong answer, which is the failure mode from the opening quote. Four: read the last readout. A tag ranging once a second is drawing milliamps, which is where this lesson meets Beacon.
CORE CONCEPTGeometry, dilution, and the one-sided lie
Once you have several ranges, position comes from trilateration: find the point whose distances to the known anchors best match the measurements. In practice it is a small least-squares problem, and the interesting question is not how to solve it but how much it amplifies your errors.
That amplification is the dilution of precision. Build a matrix H whose rows are the unit vectors from the estimated position to each anchor; then the covariance of the position estimate is proportional to (HᵀH)⁻¹, and the geometric dilution of precision is the square root of its trace. The rule of thumb that follows is worth memorising:
position error ≈ range error × GDOP
Anchors surrounding the tag from different directions give unit vectors that span the plane well, and a GDOP near 1. Anchors clustered together, or all on one side, give nearly parallel unit vectors, an ill-conditioned matrix, and a GDOP of 4, 8, or worse. The same 10 cm ranging system delivers 10 cm or 80 cm of position error depending on nothing but where somebody put the brackets. This is why anchor placement is engineering, not installation, and why it belongs in a site survey rather than in a fitter's judgement.
Two geometry traps beyond dilution. Collinear anchors — three in a straight line — leave the solution mirror-ambiguous: a tag at (x, y) and its reflection across that line produce identical ranges, and only a prior (a wall, a previous position, a fourth anchor off the line) breaks the tie. And anchor survey error propagates directly: your position estimate cannot be more accurate than your knowledge of where the anchors are. Measuring anchor coordinates with a tape to ±5 cm and then advertising ±10 cm positioning is arithmetic that does not work.
When the direct path is blocked, the receiver's first detected arrival is a reflection. A reflection travelled further than the direct path would have, so the measured range is always too long — never too short. This has three consequences that catch people. It is systematic: averaging a thousand measurements converges beautifully on the wrong answer. It is stable: the estimate does not jitter, so every quality metric you might compute from variance says the reading is excellent. And it is asymmetric: because errors are one-sided, a least-squares solve that assumes zero-mean noise is biased by the blocked link and pulls the whole position away from it.
The mitigations are not statistical. Good transceivers expose diagnostics from the channel impulse response — the amplitude of the first detected path against the total received power — and a low ratio is a genuine indicator that the first path is weak and probably not direct. Use it to discard the anchor rather than to weight it. Then, beyond that, the answer is physical: more anchors, placed so that no likely tag position is blocked from all but one of them.
IN PRACTICEBring-up notes: the REYAX RYUW122
The module used for this lesson's bench work is a REYAX RYUW122 on a separate breakout board — deliberately not on the Beacon PCB, and the canon says so explicitly. It is a UWB transceiver with an AT-command interface over a UART, which makes it a good teaching part: you can drive the whole ranging protocol from a serial terminal before writing a line of firmware, and the module hides the sub-nanosecond timestamping that would otherwise be the entire project.
Three things from an actual bring-up, offered as field notes rather than as datasheet claims. REYAX revises the AT command set between firmware versions, so verify each against the reference for the revision you have.
The UART wiring was straight-through, not crossed
Serial bring-up instinct says cross the lines: module TXD to host RXD. On the breakout used here the silkscreen labels are from the host's point of view, so the working connection was straight through — the pin marked TXD on the breakout went to the pin marked TXD on the microcontroller. This is not a claim about all UWB modules or even all revisions of this one; it is a reminder that a silkscreen label tells you what the board's designer was thinking, not what the signal is doing. When a UART is silent, swapping the two wires is a ten-second experiment and should be the first one, before you suspect baud rates or firmware.
The mode numbering was the opposite of the intuitive reading
AT+MODE selects whether the unit behaves as a tag or as an anchor. On the firmware in hand, 0 selected Tag and 1 selected Anchor — which is the reverse of what most people assume, since anchors are usually described first in the documentation and "0" reads as "the base station". Two units both configured as anchors, or both as tags, produce exactly the same symptom as a wiring fault: commands are accepted, nothing ranges, and no error is reported. Read the mode back from the module rather than trusting what you sent, and label each physical unit with a sticker.
The reported bearing is not a bearing you can use alone
The module reports an angle alongside the distance. Treating it as a compass heading does not survive contact with a room: the derived angle has a front-to-back ambiguity — the same measurement is consistent with a target in front of the array and one behind it — and the accuracy degrades sharply away from boresight and in the presence of reflections. It is usable as one constraint among several: combined with a known wall, a previous position, or a second anchor's range it disambiguates fine. It is not usable as "the tag is 34 degrees to the left". If your product design has an arrow on a screen, that arrow needs more than one module's angle output behind it.
# Unit A — configure as the anchor. Read every setting back; do not trust the echo. > AT+MODE=1 < +OK > AT+MODE? < +MODE=1 # 1 = Anchor on this firmware. Confirm before wiring anything else. > AT+ADDRESS=ANCHOR_A < +OK > AT+NETWORKID=BEACONLAB < +OK # Unit B — the tag. Same network id, different address, opposite mode. > AT+MODE=0 # 0 = Tag. The numbering is the reverse of the obvious guess. < +OK > AT+ADDRESS=TAG_01 < +OK > AT+NETWORKID=BEACONLAB < +OK # A ranging exchange, seen from the anchor's UART: < +ANCHOR_RCV=TAG_01,342,-34 # distance in cm, then an angle — treat the angle with suspicion < +ANCHOR_RCV=TAG_01,341,-33 < +ANCHOR_RCV=TAG_01,431,-31 # tag moved 10 cm behind a steel cabinet: +90 cm, and it STAYS there # Symptoms and what they actually meant, on this bench: # silence, no echo at all → UART wiring; straight-through worked, crossed did not # +OK to everything, no ranges → both units in the same mode # stable distance, wrong value → non-line-of-sight bias, not a calibration problem
That last comment line is the one worth carrying away. A stable wrong reading looks like a calibration offset, and the natural response is to subtract it. Do that and you have hard-coded the geometry of one afternoon into your product.
BEACONBeacon: why the answer is still BLE
UWB is genuinely better at the thing it does. It is also, for Beacon, the wrong tool, and the argument is worth making with the project's own requirements rather than as a general opinion.
| Requirement | What UWB would do to it |
|---|---|
| R1 — six months on 1000 mAh | A double-sided exchange with four anchors is twelve messages of transmit and receive. Even at one fix a minute rather than one a second, the ranging energy is multiples of Beacon's entire 1.8 mC report, and continuous ranging is milliamps. This requirement fails first and fails hardest. |
| R5 — BOM ≤ IDR 150,000 (~USD 9.5) at 25 units | A UWB module alone is a substantial fraction of that budget before the antenna, and it does not replace the BLE radio, which is already inside the ESP32-C3-MINI-1 at no extra cost. |
| R8 — 35 × 25 mm, 4-layer | A second radio needs its own antenna, its own keep-out, and its own RF layout on a board that already has an antenna keep-out constraint at one edge (Lesson 5.7). |
| Infrastructure | Anchors must be installed, powered, and surveyed to better accuracy than the positioning you are promising. Beacon's entire deployment story is "stick it on the cart"; anchors turn that into a site-survey project per customer. |
And then the question that decides it: what does the product actually need to know? Beacon's job, from the canon, is to report temperature, humidity, battery, and whether the asset has moved, once a minute, to a nearby gateway. "Which room is this cart in" is answered by which gateway hears it. "Has it moved" is answered by the LIS2DH12 for two microamps. Neither question needs a distance in metres, and both are already answered by hardware Beacon has.
So Beacon stays BLE, and this lesson is here for two reasons that are not "add UWB". The first is that you now know why the RSSI number in your gateway logs must not become a distance in your dashboard — that temptation is real, the code is four lines, and the result is a product that confidently reports metres it cannot measure. The second is that when a customer eventually asks for shelf-level location, you will be able to say what it costs: a second radio, powered anchors, a survey, a protocol choice that is not negotiable, and a tag that is charged weekly rather than twice a year. That is a design conversation, and having it with numbers is the difference between a scoping meeting and a year of disappointment.
WEAKNESSWhere this bites
Every BLE stack exposes an RSSI, the log-distance inversion is one line, and the resulting number looks like a distance. It is not one: a 10 dB measurement error — routine indoors, and smaller than the swing you get from turning the tag over — becomes a factor of 2.15 in the reported range at n = 3. The honest uses are zone assignment from which receivers hear the tag at all, and heavily averaged proximity with an explicit near/far/very-far vocabulary. If a specification says "distance in metres" and the hardware is a narrowband radio, the specification is wrong, and it is much cheaper to say so at requirements time than after a dashboard has been built on it.
Single-sided TWR is simpler, uses one message fewer, and is what most first implementations do. With a ±20 ppm crystal and a reply time of a millisecond it carries three metres of systematic error — twenty to thirty times the radio's actual capability. The symptom is a system that ranges beautifully when both units happen to be at the same temperature and drifts as the room warms up, which sends people looking for calibration procedures and antenna delay tuning. Antenna delay calibration is real and necessary, but it is a fixed offset of tens of centimetres; it cannot explain metres that move. Check the protocol before you check the hardware, and if you must use single-sided ranging, make the turnaround short and hardware-timed rather than handled in firmware.
"Ten centimetre accuracy" is measured in clear air. In a warehouse of steel racking, or with a tool cabinet between the tag and one anchor, the first detected arrival is a reflection and the range is biased long — stably, confidently, and by an amount that depends on the room's geometry rather than on anything you control. A system validated in an empty lab and deployed into a workshop will not degrade gracefully; it will report precise wrong positions. Validate in the actual environment, use the transceiver's first-path diagnostics to discard suspect links rather than average them, and place enough anchors that no plausible tag position is blocked from all but one.
KEY INSIGHTHow this connects
- Wireless Fundamentals: dBm, Link Budget, and Antennas (4.5) — the multipath that made RSSI vary by 20 dB over centimetres is the same multipath that UWB resolves in time rather than summing in amplitude.
- BLE: Advertising, Connections, and Data Design (4.6) — the fourteen bytes that already answer Beacon's real questions, for a fraction of a millicoulomb.
- OTA Updates, Provisioning, and Security Basics (4.9) — next, and the last of the stage: how a fleet of 25 units gets fixed after it ships, and how not to brick it.
- Capstone: Beacon Pilot Deployment and Post-Mortem (6.9) — where the real path-loss exponents, gateway placements, and metal cabinets of an actual building stop being assumptions.
You now know what a distance measurement costs and what it is worth. For Beacon, the answer is that the cheaper measurement already answers the question.
GLOSSARYKey terms introduced
- Ultra-wideband (UWB)
- Radio using at least 500 MHz of bandwidth, in bands around 6.5 and 8 GHz under IEEE 802.15.4z. The bandwidth is what gives nanosecond-scale time resolution and separates reflections from the direct path.
- Time of flight (ToF)
- The propagation delay between two devices. At the speed of light, 1 ns is 30 cm, so ten-centimetre ranging requires resolving arrival times to roughly a third of a nanosecond.
- Single-sided TWR
- Ranging from one round trip: ToF = (T_round − T_reply)/2. Its error is e × T_reply / 2, so a 20 ppm offset with a 1 ms reply produces about 3 m of systematic range error.
- Double-sided TWR
- Ranging from two round trips with a third message, combining both reply times so that clock offset and drift cancel to first order. Residual error is second order in the offset — effectively zero.
- Anchor and tag
- The fixed, surveyed units and the mobile one. Anchors are usually powered; the tag carries the battery, which is why the tag's message count is the energy-relevant number.
- Dilution of precision (GDOP)
- The factor by which anchor geometry multiplies range error into position error, from the trace of (HᵀH)⁻¹ over the unit vectors to each anchor. About 1 for surrounding anchors, 4 or more for clustered ones.
- NLOS bias
- The one-sided positive range error when the direct path is blocked and the first detected arrival is a reflection. Systematic and stable, so averaging does not remove it and variance-based quality metrics do not detect it.
- First-path diagnostics
- Transceiver-reported metrics derived from the channel impulse response, such as first-path amplitude against total received power. A low ratio flags a probable non-line-of-sight link, which should be discarded rather than down-weighted.
ACTIVE RECALLCheck your understanding
Answer in your own words first, then reveal.
1Your single-sided TWR system reads 3 m too far. Someone proposes a better crystal. Someone else proposes a different protocol. Quantify both.›
The error in single-sided ranging is e × T_reply / 2 in time, which is c × e × T_reply / 2 in distance. With e = 20 ppm and T_reply = 1 ms that is 3 × 10⁸ × 2 × 10⁻⁵ × 10⁻³ / 2 = 3 m, which matches the symptom.
Better crystal: a 2 ppm temperature-compensated oscillator is a tenfold improvement, so the error falls to 30 cm. That is a real gain, it costs money on every unit, and 30 cm is still three times the radio's line-of-sight capability. Different protocol: double-sided TWR makes the residual second order in the offset, so at the same 20 ppm the clock contribution becomes about 4 × 10⁻¹⁰ of the reply time — sub-millimetre, which is to say gone. It costs one extra message per exchange: roughly 50 % more air time and energy, and no hardware change at all.
The protocol wins by a wide margin, and the general lesson generalises: when an error is structural, fix the structure before you buy better components. The crystal upgrade would be the right answer only if you were forced into single-sided ranging by an energy budget that could not afford the third message.
2Two installations use identical hardware and identical firmware. One reports positions to 12 cm, the other to 70 cm. What is the most likely difference, and how would you confirm it without visiting?›
Almost certainly anchor geometry. Position error is approximately range error multiplied by the geometric dilution of precision, and GDOP depends only on where the anchors are relative to the tag. Anchors surrounding the tag give a GDOP near 1; anchors clustered on one wall or in one corner give 4 or more. The same 12 cm ranging accuracy therefore yields 12 cm or 50–100 cm of position error with nothing else changed.
To confirm remotely: you already know the anchor coordinates because the system needs them to solve, so compute the GDOP at several representative tag positions in each installation and compare. If the good site is near 1 and the bad site is near 5, that is the answer and the fix is a bracket and a ladder. Also check the reported per-anchor residuals: geometric dilution shows as a large position error with small range residuals, whereas a non-line-of-sight problem or an anchor survey error shows as one link with a persistently large residual. The two look identical in the position output and completely different one level down.
3A colleague says the ranging is noisy behind a metal cabinet and proposes averaging over thirty seconds. What is wrong with the diagnosis and with the fix?›
The diagnosis is wrong because the error is not noise. When the direct path is blocked, the first detected arrival is a reflection, and a reflection travelled further than the direct path — so the range is biased long, always, by an amount set by the room's geometry rather than by chance. The reading is typically very stable, which is why it does not look like a fault.
The fix is wrong for the same reason: averaging reduces zero-mean noise, and this error has a mean. Thirty seconds of averaging converges precisely on the wrong answer and, worse, makes every variance-based quality metric report that the measurement is excellent. It converts an obvious problem into a confident one.
What to do instead: read the transceiver's first-path diagnostics — the amplitude of the first detected path relative to total received power — and use a poor ratio to discard that anchor from the solve rather than to weight it, since a biased measurement contributes bias no matter how lightly you weight it. Beyond that the answer is physical: place enough anchors that no plausible tag position is blocked from all but one, and validate in the real environment rather than an empty lab.
4A customer asks for Beacon to report "how far the cart is from the door, in metres". Give the answer you would actually give.›
Start with what is possible today at no cost: Beacon can tell you reliably which gateway's zone the cart is in, and whether it has moved since the last report. If the door has its own gateway, "the cart is in the doorway area" is available now and is accurate, because it comes from which receivers hear the tag rather than from a distance calculation.
Then be explicit about what cannot be done: the received signal strength is not a distance. Indoor multipath moves it by 15–20 dB over centimetres, which becomes a factor of two or more in any inverted range estimate. A dashboard reporting metres from RSSI would be reporting a number that is wrong by more than the width of the room, stably enough to look trustworthy.
Then price the real answer: metre-level positioning means ultra-wideband. That is a second radio on the tag, which does not fit the current board, the bill of materials at 25 units, or the six-month battery requirement — a tag doing double-sided ranging with four anchors even once a minute costs multiples of its entire current report, and continuous ranging puts it in the milliamp class. It also means powered anchors installed and surveyed to better than the accuracy being promised, which turns a stick-on tag into a site-survey project per customer. That is a different product, and I would want to know whether "which zone, and has it moved" genuinely fails to answer the underlying question before committing to it.