UWB Real-Time Location Systems in Factories: IEEE 802.15.4z vs BLE Channel Sounding

A forklift crosses a bay at walking speed, and your digital twin shows it six metres from where it actually is. That gap is not a software bug. It is the physics of radio signal strength in a steel building, and it is why the UWB real-time location system (RTLS) became the default answer for factories that need decimetre-level tracking of tools, carts, and people. Ultra-wideband (UWB) measures time of flight instead of signal strength, and a 499.2 MHz wide pulse lets a receiver separate the direct path from the reflections that wreck narrowband methods.

The picture changed in 2024. Bluetooth Core 6.0 added Channel Sounding, a link-layer ranging procedure that brings phase and round-trip-time measurement to ordinary 2.4 GHz Bluetooth LE radios. Vendors now pitch it as “good enough” for jobs once reserved for UWB. Whether that holds depends on the geometry, the tag count, and what you intend to feed into the twin.

What this covers: the time-of-flight math behind two-way ranging (TWR) and time-difference-of-arrival (TDoA), what IEEE 802.15.4z actually standardised, how anchor placement drives error, where Channel Sounding competes and where it does not, and how to stream positions into a digital twin over MQTT.

Context and Background

Indoor positioning has been tried with almost every radio. Wi-Fi and Bluetooth received signal strength indication (RSSI) fingerprinting delivers metre-scale error at best, and degrades whenever a pallet moves, because RSSI is a function of path loss, shadowing, and antenna orientation rather than of distance. Bluetooth angle-of-arrival (AoA) improves on that with antenna arrays but still depends on a narrowband signal that cannot separate overlapping paths. Optical and ultrasonic systems are accurate but demand line of sight and dense infrastructure. UWB occupies the niche between: radio, so it passes through light obstructions, but wideband, so timing is precise.

The key quantity is time. Radio waves travel at 299,792,458 m/s, roughly 0.3 m per nanosecond. A receiver that timestamps arrival with 1 ns uncertainty therefore carries about 30 cm of range uncertainty, and 100 ps maps to about 3 cm. A signal with 499.2 MHz of bandwidth has a main pulse lobe of around 2 ns, which is why UWB can resolve a direct path from a reflection that arrives a few nanoseconds later. A 2 MHz Bluetooth LE signal cannot, and neither can a 20 MHz Wi-Fi channel.

The IEEE 802.15.4 family defines the high-rate pulse (HRP) UWB physical layer. According to Keysight’s overview of the IEEE 802.15.4 HRP UWB standard, the standard defines 16 channel numbers across a sub-GHz band, a low band (3494.4 to 4492.8 MHz), and a high band (6489.6 to 9484.8 MHz). The 802.15.4z-2020 amendment added the scrambled timestamp sequence (STS), an AES-128 based ciphered sequence that strengthens ranging integrity. In practice, industrial hardware centres on channels 5 (6.5 GHz) and 9 (8 GHz). Qorvo’s DW3000 family lists exactly those two channels, TWR, TDoA, and phase-difference-of-arrival (PDoA) support, and states a ranging precision “within 10 cm” with plus or minus 5 degrees angular measurement.

Two ecosystems sit above the radio. The FiRa Consortium defines interoperability profiles, and its Core 3.0 release in January 2025 added hybrid UWB scheduling, dedicated data transfer, and a one-to-many double-sided TWR protocol used for the Car Connectivity Consortium Digital Key. Consumer silicon also matters because it sets component volume: 9to5Mac’s report on the iPhone 15 states the second-generation Apple U2 chip offers up to three times the range of its predecessor. Factory RTLS rarely uses phone silicon, but the volume of consumer UWB lowers chip cost and hardens the standard.

This article stays at the physical and system level. It leaves out vendor price lists and any unpublished accuracy claim, because those change by quarter and are rarely reproducible. Where a figure is illustrative rather than sourced, the text says so. If you need the Bluetooth fundamentals first, our complete technical guide to Bluetooth Low Energy covers the stack up to the link layer that Channel Sounding extends.

Reference Architecture of a Factory UWB RTLS

A UWB real-time location system consists of tags, fixed anchors, a time-synchronisation fabric, a location engine, and an integration layer. Tags emit or answer UWB frames. Anchors timestamp them. The engine converts timestamps to positions using TWR or TDoA, filters the result, and publishes it. Accuracy is set by timestamp quality, anchor geometry, and multipath, in that order of controllability.

Figure 1: Reference architecture of a UWB real-time location system, from tags and anchors to the digital twin and lakehouse.

Figure 1 shows the data path. Tags on assets and people transmit frames. Ceiling-mounted anchors receive them and send timestamps over wired Ethernet with Power over Ethernet (PoE), which also carries the synchronisation signal in most industrial deployments. A location engine solves for position. A filter smooths the output and applies map constraints. The result is published to an MQTT broker, from which both the twin and long-term storage subscribe.

Tags, anchors, and the role of the infrastructure

Anchors are fixed at surveyed coordinates, and every position estimate is relative to those coordinates. An anchor error of 5 cm in the survey becomes a 5 cm bias in every position it influences, so the survey is part of the system accuracy budget, not a commissioning footnote. Tags are cheap and battery powered. Their cost, size, and update rate are the main design variables for the asset population, while anchors are the capital cost of coverage.

Two operating models dominate. In the TWR model, the tag actively ranges with each anchor by exchanging several frames. In the TDoA model, the tag transmits a single frame, and the anchors, which share a common clock, record when it arrives. The choice determines battery life, tag density, and infrastructure cost, and it is the most consequential decision in the design.

Why wide bandwidth buys accuracy

Timing resolution is inversely related to bandwidth. A receiver estimates arrival time by correlating the received signal against a known template, and the sharpness of the correlation peak scales with bandwidth. With 499.2 MHz, the peak is narrow enough that the leading edge of the channel impulse response (CIR) can be located to within a fraction of a nanosecond under good signal-to-noise ratio.

This is not the same as saying UWB “sees through walls”. The direct path is attenuated by obstructions, sometimes below the level of a strong reflection. The advantage of wide bandwidth is that the receiver can distinguish the two arrivals in time, then choose the earliest credible one. Narrowband systems see the sum of the two and cannot undo it.

The 802.15.4z contribution: secure timestamps

Before 802.15.4z, ranging relied on a preamble whose structure was public. Researchers showed this allowed distance-reduction attacks, where an attacker transmits a signal that makes a receiver believe a frame arrived earlier. The Keysight overview and the paper Secure Ranging with IEEE 802.15.4z HRP UWB both describe the STS as a field generated using AES-128 in counter mode, producing pseudo-random pulse polarities that an outsider cannot predict. The same paper shows the synchronisation preamble remains known and therefore still attackable, and proposes a receiver design that relies on the STS-derived channel estimate to bound false acceptance.

For a factory this matters less than for car keys, since most RTLS threats are operational rather than adversarial. It still matters in two cases: access-controlled zones where position gates a machine interlock, and any system where a spoofed position could cause a safety decision. For general asset tracking, the practical benefit of 802.15.4z is the standardised frame structure and the multi-vendor interoperability it enables.

Anchor synchronisation as the hidden cost of TDoA

TDoA needs every anchor to share a time base to within a fraction of a nanosecond, because a 1 ns synchronisation error becomes a 30 cm range-difference error. There are three common methods. Wired distribution of a reference clock over Ethernet or dedicated cable gives the best stability. Wireless sync, where a master anchor broadcasts a frame and others timestamp it, avoids cabling but must correct for clock drift between frames. A hybrid applies wireless sync with wired backhaul of timestamps.

Clock drift is the quiet enemy. Crystals commonly drift in the range of parts per million (ppm), and at 10 ppm a clock gains 10 microseconds per second. Between sync frames sent every 100 ms, the unmodelled drift is 1 microsecond, about 300 metres of apparent range. Practical systems estimate each anchor’s clock offset and skew continuously with a Kalman-type filter, and the residual after correction is what limits TDoA accuracy. This is why TDoA systems publish sync intervals as a headline specification.

Ranging Mathematics: ToF, TWR, and TDoA

Time of flight (ToF) ranging converts a measured propagation delay into distance with d = c x ToF. Two-way ranging cancels the unknown clock offset between devices by exchanging frames in both directions. TDoA instead measures differences in arrival time at synchronised anchors and solves for the tag on hyperbolas. TWR scales poorly with tag count but needs no anchor sync; TDoA scales well but depends on it.

Single-sided and double-sided TWR

In single-sided TWR (SS-TWR), the tag sends a poll at its local time, the anchor receives it, waits a known reply delay, and responds. The tag measures the round-trip time Tround and subtracts the anchor’s reply time Treply. The ToF is (Tround – Treply) / 2. The problem is that Tround is measured by the tag clock and Treply by the anchor clock, and the two oscillators disagree.

The error is easy to quantify. If the anchor reply delay is Treply and the fractional clock mismatch is e (for example 10 ppm is e = 10 x 10^-6), the ToF error is approximately Treply x e / 2. With an illustrative reply delay of 300 microseconds and a 10 ppm mismatch, the error is 1.5 ns, about 45 cm. Shortening the reply delay shrinks it, but hardware latency sets a floor. Some chips estimate carrier frequency offset from the received frame and correct for it, which is how SS-TWR reaches useful accuracy in many consumer products.

Double-sided TWR (DS-TWR) adds a third message, so both sides measure a round trip and a reply delay. Figure 2 shows the exchange. Combining the two round trips and two reply delays in the asymmetric formula cancels the first-order clock error regardless of the reply delays, which is why DS-TWR is the default for accuracy-critical ranging and why FiRa’s one-to-many Digital Key protocol uses the double-sided form.

Figure 2: Double-sided TWR exchange. Two round trips with two reply delays cancel first-order clock offset between tag and anchor.

The cost is airtime and energy. Each range to each anchor needs three frames, plus a result report if the tag does not compute position itself. To obtain a 2D fix, the tag needs ranges to at least three anchors; a 3D fix needs four, and a fifth improves redundancy. Frame airtime at the common 6.8 Mb/s rate is short, on the order of a hundred microseconds or less for a minimal frame (illustrative, depending on preamble length and payload), but the tag still wakes its radio many times per fix. Battery life falls in proportion.

Capacity: why TWR does not scale to thousands of tags

Every TWR fix consumes channel time. If a fix requires N anchors times 3 frames at about 150 microseconds each (illustrative), and scheduling guard times double that, then a single fix occupies roughly 3 ms for four anchors. A channel can then support on the order of 300 fixes per second before saturation, which is 300 tags at 1 Hz or 30 tags at 10 Hz. Real systems schedule ranging slots and can reuse channels spatially, but the order of magnitude explains why high-density deployments choose TDoA. The calculation is a model for intuition, not a product specification.

TWR retains two advantages. Anchors need no precise synchronisation, so installation is faster and wireless-only anchors are practical. And the tag knows its own position, which suits applications where the mobile device decides, such as a handheld tool that must verify it is inside a work cell before enabling torque.

TDoA: one frame, many receivers

In uplink TDoA, the tag transmits a blink frame, and every anchor within range records the arrival time on its own clock. After clock correction, the difference in arrival time between anchors i and j is (d_i – d_j) / c, where d is the tag-to-anchor distance. Each time difference constrains the tag to a hyperbola with the two anchors as foci. In 2D, two independent differences (three anchors) intersect at a point; in 3D, three differences (four anchors) are required. Additional anchors over-determine the system, and the solver uses nonlinear least squares or an extended Kalman filter.

Figure 3: Uplink TDoA. A master anchor synchronises the others, a tag blink is timestamped by all, and a solver intersects hyperbolas.

The tag in this mode is a transmitter only, so it uses a tiny fraction of the energy of TWR. A tag sending one short blink per second can reach multi-year battery life on a coin cell in the vendor literature, though the exact figure depends on blink length, battery chemistry, and sleep current, and I have not verified a specific number for this article. The scaling advantage is structural: adding tags adds airtime only for their own blinks, and every anchor hears every blink at no extra cost.

The limit is uncoordinated access. Tags transmit blinks without listening, so collisions occur. For a pure ALOHA channel, throughput peaks at 18.4 percent utilisation, and a blink survives only if no other blink overlaps within twice the frame duration. With an illustrative blink of 200 microseconds and 1,000 tags blinking once per second, offered load G is 0.2 and the survival probability is exp(-2G), about 67 percent. Randomised blink intervals and a few retransmissions recover the loss, but the model shows why update rate per tag trades off against tag count.

Geometry: why the same ranging error gives different position error

Ranging error does not translate one to one into position error. The geometric dilution of precision (GDOP) is the multiplier that depends on where the anchors sit relative to the tag. For a well-surrounded tag, GDOP is near 1 to 1.5 in 2D. For a tag outside the anchor polygon, or where anchors are nearly collinear, GDOP rises quickly and the solution stretches along the weak direction.

A simple rule of thumb is position error roughly equals GDOP times the ranging standard deviation. With an illustrative 10 cm ranging error, a GDOP of 1.5 gives about 15 cm, while a GDOP of 5 at a corridor end gives 50 cm. Those numbers are not measurements, they show why anchor layout deserves more design effort than anchor model selection.

Vertical geometry is the usual weakness. Anchors mounted at similar heights on a ceiling provide almost no vertical baseline, so z error can be several times worse than x and y. Teams that need floor discrimination add anchors at different heights, use known tag mounting height as a constraint, or fuse a barometric sensor. State this limit at design time, because stakeholders often assume that a 3D system is isotropic.

Multipath and non-line-of-sight

Multipath produces two different failure modes. In line of sight (LOS) with reflections, UWB’s time resolution usually lets the receiver pick the first path correctly, and error stays within a few centimetres. In non-line-of-sight (NLOS), the direct path is blocked or heavily attenuated, the first detectable path may be a reflection or a delayed, dispersed version of the direct path, and the measured range is biased long. The bias is positive because radio can only arrive late, never early, which is a useful asymmetry.

Detection methods look at the CIR. A receiver can compare first-path power with total received power, or examine the delay spread, and flag the range as suspect. Many chips expose diagnostics that the location engine uses to down-weight or reject NLOS ranges. With TDoA, one NLOS anchor biases one time difference, and robust solvers such as RANSAC-style subset selection or iteratively reweighted least squares can discard it when redundancy exists. This is another argument for five or more anchors per zone.

Metal-heavy environments add a subtler issue. Large reflective surfaces create strong specular paths, and moving machinery makes the multipath pattern time-varying. A static survey will not reveal this. Plan a walk test with a reference tag on a surveyed trolley, record the error distribution, and look at percentiles rather than the mean. A mean error of 15 cm with a 95th percentile of 80 cm is a different system from one with a 95th percentile of 30 cm.

BLE Channel Sounding: What It Measures and What Is Sourced

Bluetooth Channel Sounding (CS) is a Bluetooth Core 6.0 feature that measures distance between two LE devices using phase-based ranging (PBR) across many 2.4 GHz channels, plus round-trip time (RTT) packets that add a security check and resolve ambiguity. It targets proximity and ranging between a pair of devices. It is not, by itself, a multi-anchor RTLS.

CNX Software’s summary of the Core 6.0 announcement confirms the two methods and dates the Core 6.0 specification announcement to September 2024. The Bluetooth SIG’s own Channel Sounding talk states that PBR operates across multiple frequency tones in the 2.4 GHz band, that RTT measurements mitigate man-in-the-middle distance manipulation, and that RSSI-based methods are limited to errors above 1 metre. It does not give a headline accuracy figure for CS.

How phase-based ranging works

In PBR, one device sends an unmodulated continuous-wave tone at frequency f, and the other measures its phase and returns a tone. Since the signal travels out and back, the measured phase is proportional to 2d: phi = 2 pi f (2d / c) plus unknown constant offsets. The offsets are removed by exchanging tones in both directions and combining them. Doing this at many frequencies gives phase as a function of frequency, and the slope of that line is proportional to distance: d = -(c / 4 pi) x dphi/df.

The unambiguous range follows from the frequency step. A step of delta-f changes round-trip phase by 4 pi delta-f d / c, which wraps at 2 pi when d = c / (2 delta-f). For a 1 MHz step, that is about 150 m, comfortably more than a factory bay. The phase slope averages over all measured tones, so noise falls as more tones are included, and a multipath reflection appears as a deviation from the straight line.

The fundamental limit is bandwidth. The tones span at most the 2.4 GHz band, roughly 80 MHz. Resolution in the time domain is about c / (2B), which for 80 MHz is near 1.9 m for separating two paths. UWB with 499.2 MHz gives about 0.3 m for the same measure. When two paths are closer than the resolution, phase measurements mix them, and the slope estimate has a bias that depends on the relative phase and amplitude of the reflection. This is the physics behind the claim that CS is more sensitive to multipath than UWB.

Accuracy claims: what can and cannot be cited

Third-party explainers circulate figures. One, from Hubble Network’s guide, describes mode 1 as RTT at about 1 m, mode 2 as PBR with sub-centimetre theoretical precision subject to ambiguity, mode 3 as a combination giving 10 to 50 cm, and notes that multipath and antenna orientation reduce real-world performance. This is a vendor-side explainer, not a measurement study, and I treat the 10 to 50 cm range as a claim rather than a verified result. CNX Software’s source says only that the specification “promises sub-meter accuracy”.

I could not verify, in this run, an independent peer-reviewed measurement of CS accuracy in an industrial environment. Treat any single number in marketing material as a best case in a benign environment until you reproduce it on your own floor. The sensible procedure is the one used for UWB: a surveyed reference trolley, many positions, and percentile reporting.

Architecture differences that matter more than accuracy

CS is a connection-oriented procedure between an initiator and a reflector. Each device measures against one peer at a time. That is natural for a phone locating a tag or a door lock verifying a key, but a floor with 2,000 tags needs an architecture that schedules ranging events over many connections. UWB TDoA requires no connection at all, and a single blink serves all anchors.

Hardware cost is the strongest argument for CS. As Hubble’s guide notes, it is a controller-level feature on Bluetooth 6.0 chipsets and needs no second radio, so a Bluetooth tag can add ranging without a UWB chip. For tags already carrying BLE for telemetry, the incremental bill of materials is small. That changes the economics of coarse zone tracking, ergonomics tools, and phone-based workflows.

Comparison and Decision Matrix

UWB TDoA fits dense, many-tag, decimetre-class tracking. UWB TWR fits secure, per-device ranging. BLE Channel Sounding fits phone and low-cost tag ranging where sub-metre is enough. BLE RSSI or AoA fits room or zone presence. The right choice follows tag count, accuracy percentile, and whether phones are in the loop.

Figure 4: Decision tree for indoor positioning technology in a factory, keyed on tag count, accuracy need, and phone support.

Dimension	UWB TDoA (802.15.4z HRP)	UWB TWR (802.15.4z, FiRa)	BLE Channel Sounding	BLE RSSI or AoA
Measured quantity	Arrival time differences	Round-trip time	Phase slope plus RTT	Signal strength or angle
Bandwidth	499.2 MHz class channels	499.2 MHz class channels	About 80 MHz span of tones	1 to 2 MHz
Tag energy per fix	Lowest, one blink	Higher, multiple frames	Moderate, connected procedure	Low
Tag scaling	Excellent	Limited by airtime	Limited by connections	Good
Infrastructure	Synchronised anchors	Anchors, no precise sync	Any BLE 6.0 peer	BLE gateways
Multipath resistance	Strong, first-path detection	Strong	Weaker, phase mixes paths	Weak
Phone support	Limited to UWB phones	UWB phones via FiRa	Any BLE 6.0 phone	Nearly all phones
Typical use	Fleet and WIP tracking	Secure access, tool zones	Proximity, lock, find-my	Presence, zones
Accuracy claim status	Vendor spec, for example within 10 cm ranging for DW3000 family	Same	Marketing claims, 10 to 50 cm cited, not independently verified here	Metre-class

Read the table column by column, not row by row. The matrix says nothing about price, because component and installation prices vary by region and volume, and I have not verified current figures. It also does not claim that CS cannot reach decimetre accuracy; it says the evidence available here does not establish it for a cluttered factory.

Where each technology wins

UWB TDoA wins when the question is “where is every cart, every tool, every person, continuously?” Update rates of several hertz per tag across thousands of tags, with decimetre-class accuracy, are well matched to uplink TDoA. Examples include work-in-progress tracking on an assembly line, tool tracking for torque traceability, and proximity alarms between forklifts and pedestrians.

UWB TWR wins when a device must decide on its own, with a secure range. Access to a machine cell, vehicle-to-key proximity, and tool-to-station authentication fit this. FiRa’s profile and the STS give a standard path to interoperability. Apple’s U1 and U2 chips and Android UWB phones add consumer endpoints, but support varies by model, so check the specific device list before building a phone-based workflow.

Channel Sounding wins on cost and ubiquity. A BLE 6.0 tag that can range to a phone or a fixed BLE node adds proximity features cheaply, and for use cases that need “within a metre and secure” it may replace UWB. It is a mistake to treat the two as direct substitutes. Treat them as complementary layers, with BLE for wake-up and coarse zone, and UWB for the fine fix.

Hybrid designs

A common pattern combines BLE advertising for presence with UWB for precision. The tag sleeps with the UWB radio off, wakes on a BLE trigger when it enters a zone, and starts blinking at a higher rate. This cuts UWB airtime and battery use, and it keeps the TDoA collision model in a good operating range. CS may take the BLE-trigger role with better distance gating than RSSI, which is a realistic near-term role.

Integrating Positions into a Digital Twin over MQTT

A twin needs positions as a stream of timestamped, quality-tagged events, not as raw dots. Publish filtered positions to a hierarchical MQTT topic structure, include a covariance or accuracy estimate, and keep raw timestamps available for replay. Let the twin decide how to use uncertain data.

The integration layer sits behind the filter in Figure 1. Our MQTT and Sparkplug B reference architecture describes the broker topology, and the same unified namespace pattern fits location data: one topic per asset, with a stable schema.

Topic and payload design

Use a topic such as site/plant1/area/bay3/asset/forklift-17/location. Keep the payload small and typed. Include the coordinate frame identifier, because the twin and the RTLS frequently disagree about the origin, and a sign error on the y axis is a common first-week bug.

{
  "ts": "2026-10-06T08:45:00.120Z",
  "frame": "plant1-floor0-v3",
  "x": 41.27, "y": 12.84, "z": 0.12,
  "cov": [[0.02, 0.0], [0.0, 0.03]],
  "hdop": 1.4,
  "anchors_used": 6,
  "los_flags": 0,
  "source": "uwb-tdoa"
}

Publish with QoS 0 for high-rate position updates, because a lost sample is replaced by the next, and QoS 1 for zone entry and exit events, which are state transitions that must not vanish. Use retained messages for the last known location, so a twin that restarts can rebuild current state immediately.

import json, time
import paho.mqtt.client as mqtt

client = mqtt.Client(mqtt.CallbackAPIVersion.VERSION2)
client.connect("broker.plant1.local", 1883)
client.loop_start()

def publish_fix(asset, fix):
    topic = f"site/plant1/area/{fix['area']}/asset/{asset}/location"
    payload = json.dumps({
        "ts": fix["ts"], "frame": "plant1-floor0-v3",
        "x": fix["x"], "y": fix["y"], "z": fix["z"],
        "hdop": fix["hdop"], "anchors_used": fix["n"],
        "source": "uwb-tdoa",
    })
    client.publish(topic, payload, qos=0, retain=True)

Filtering and fusion before the twin

Raw TDoA fixes are noisy and occasionally jump. A motion-model filter smooths them. A Kalman filter with a constant-velocity model suits carts and people. A particle filter handles map constraints, such as walls and one-way aisles, better, at higher compute cost. Our walk-through of Kalman filter vs particle filter vs EnKF state estimation for digital twins covers the trade-off in detail, and the same reasoning applies to position streams.

Fusion with inertial data helps in the gaps. A tag with an accelerometer and gyroscope can dead-reckon through a short NLOS shadow, and the UWB fix corrects the drift when it returns. Fusion with machine state helps too: a forklift whose mast is raised, or a conveyor that reports a part present, supplies constraints the radio cannot.

Twin binding and history

On the twin side, a platform such as Eclipse Ditto can hold the latest position as a feature of the asset’s Thing. Our Eclipse Ditto digital twin tutorial shows how to bind MQTT messages to features. For history, write positions in batches to columnar tables. Trajectory data grows quickly: 2,000 tags at 2 Hz is 4,000 rows per second, or about 345 million rows per day, which is a good fit for an open table format. See our comparison of Apache Iceberg versus Paimon lakehouse table formats for how to choose one for streaming ingest.

Time alignment deserves care. The RTLS timestamps are in the location engine’s clock, the MES events are in another, and the twin has its own. Synchronise all three with NTP or PTP and record the sync source in the payload, otherwise you cannot correlate a position with a process event to better than the clock offset, which can be larger than the position accuracy you paid for.

Trade-offs, Gotchas, and What Goes Wrong

The most common failure is not the radio. It is a mismatch between the accuracy promised in a datasheet and the accuracy delivered at the 95th percentile on a real floor. Datasheet figures describe line-of-sight ranging in a controlled setting. A plant adds metal racks, moving vehicles, human bodies, and tags worn at awkward orientations, and each one shifts the error distribution.

Body shadowing. A tag on a worker’s chest ranges through the torso whenever the person faces away from an anchor. That is NLOS, and ranges are biased long. Mount tags on shoulders or helmets, and expect worse performance for tags in pockets.

Antenna orientation. UWB antennas are not isotropic, and chip delay depends on the angle of arrival. Calibrated systems store antenna delay tables, and uncalibrated ones show angle-dependent bias of several centimetres. A tag lying flat on a pallet differs from one standing upright.

Sync fragility in TDoA. A failed switch or a loose cable can degrade the sync of a subset of anchors without any alarm, and positions drift or jump in one area. Monitor sync residuals per anchor as a first-class metric, and alert on them.

Interference and coexistence. UWB is low power and coexists well with narrowband radios, but dense deployments can interfere with each other if channels, preamble codes, and schedules are not planned. Adjacent plants or neighbouring vendors’ systems on the same channel are a real hazard, so record channel and preamble-code plans in your asset register.

Regulation. UWB emission limits and permitted bands differ by country, and rules have changed over time. I did not verify current regulatory limits in this run, so confirm channel availability and power limits with the local regulator and the vendor before specifying hardware for a multi-country rollout.

Channel Sounding specific risks. CS is new, and chip and stack support is still maturing as of this writing. Its phase method is sensitive to multipath within its resolution, antenna design affects the result, and connection-based ranging limits scaling. Treat early silicon results as provisional, and test on candidate hardware rather than reference designs.

Security is not free. STS makes distance reduction attacks hard on a UWB link, but provisioning keys, rotating them, and protecting the anchor network are your responsibility. A position stream also reveals where people are, which is personal data in many jurisdictions. Decide retention, anonymisation, and works-council consultation before go-live, not after.

Over-trusting the twin. A twin that renders a forklift at a precise coordinate gives a false sense of certainty. Surface the covariance, and let downstream logic, such as collision alerts, use uncertainty-aware thresholds.

Practical Recommendations

Start from the use case, not the technology. If the requirement is “all assets, continuously, decimetre-class”, plan UWB TDoA with wired sync. If it is “secure, per-device, phone-aware proximity”, plan UWB TWR under FiRa profiles. If it is “within a metre, cheap, and phone-friendly”, evaluate Channel Sounding with a pilot, and plan for UWB as a fallback.

Budget anchors for geometry, not coverage. Surround each zone, avoid collinear layouts, add height diversity where z matters, and keep five or more anchors in view of every working position so a solver can reject an NLOS range. Survey anchor coordinates to better than the accuracy you want and record the survey method.

Pilot with a reference. Put a tag on a surveyed trolley or a robot with an independent ground truth, walk or drive a grid of positions, and report mean, median, 95th and 99th percentile error per zone. Repeat after a production shift, because the environment changes when the plant is running.

Design the data path once. Publish filtered positions with covariance to MQTT, keep raw timestamps for replay, and write history to a table format built for streaming.

Checklist:

• Define the accuracy requirement as a percentile per zone, not a mean.
• Choose TDoA for density and battery life, TWR for secure per-device ranging, CS for low-cost proximity.
• Plan anchor geometry for GDOP, height diversity, and at least five anchors in view.
• Monitor anchor synchronisation residuals and set alerts.
• Calibrate antenna delays and test tag orientations and body shadowing.
• Confirm local UWB regulation and a channel and preamble-code plan.
• Publish position, covariance, frame identifier, and clock source over MQTT.
• Decide privacy, retention, and works-council consultation for people tracking.
• Pilot Channel Sounding on your target silicon before committing to it.

Frequently Asked Questions

How accurate is a UWB real-time location system in a factory?

Vendor datasheets cite ranging precision of around 10 cm for UWB chips, for example Qorvo’s DW3000 family lists “within 10 cm” for ranging. Installed position accuracy is typically worse, because it also depends on anchor geometry, NLOS, and synchronisation. Expect to measure it yourself and report percentiles. Line-of-sight zones with good anchor coverage can approach the chip figure, while shadowed zones will not.

What is the difference between TWR and TDoA in UWB positioning?

TWR measures the round-trip time between a tag and each anchor by exchanging several frames, so no anchor synchronisation is needed but each tag consumes airtime and energy. TDoA has the tag send one blink, and synchronised anchors measure arrival-time differences, which scales to many tags and saves tag battery. TDoA needs precise anchor clock synchronisation, and a 1 ns error equals about 30 cm.

What did IEEE 802.15.4z add to UWB?

The 802.15.4z-2020 amendment added the scrambled timestamp sequence, an AES-128 based ciphered sequence that protects ranging against distance-reduction attacks, along with ranging-oriented frame and mode enhancements. It underpins FiRa interoperability and secure ranging in consumer and automotive products. For factory tracking, its main value is standardised, secure, multi-vendor ranging rather than a change in raw accuracy.

Can BLE Channel Sounding replace UWB for indoor positioning?

For proximity and ranging between two devices, possibly, where sub-metre accuracy is acceptable and phones are involved. For dense multi-tag tracking with decimetre accuracy, the evidence is weaker. Channel Sounding uses about 80 MHz of 2.4 GHz spectrum and is more sensitive to multipath, and it is connection-oriented. Independent industrial measurements were not available to verify in this run, so pilot it first.

How many anchors does a UWB RTLS need?

Mathematically, three anchors give a 2D fix and four a 3D fix. Practically, plan for at least five anchors in view of every working position, so the solver can reject an NLOS range and the geometry stays strong. Coverage per anchor depends on environment and link budget, so the count is determined by the survey and pilot rather than a fixed area rule.

How do I send UWB position data to a digital twin?

Filter the raw fixes with a Kalman or particle filter, then publish them to an MQTT topic per asset with timestamp, coordinate frame, position, covariance, and quality flags. Use QoS 0 for streaming positions, QoS 1 for zone events, and retained messages for last known state. The twin subscribes and binds the data to the asset model, while a lakehouse stores the history.

Further Reading

Internal:

• Bluetooth Low Energy complete technical guide for the BLE stack beneath Channel Sounding.
• MQTT and Sparkplug B reference architecture for the broker and namespace pattern used to carry positions.
• Kalman filter vs particle filter vs EnKF for digital twin state estimation for filtering position streams.
• Apache Iceberg vs Paimon lakehouse table formats for storing high-volume trajectory history.
• Adjacent architecture reads on distributed ledger settlement, outside the positioning topic: bank stablecoin consortium architecture and Circle Arc stablecoin-native Layer 1 architecture.

External:

• Keysight: An overview of the IEEE 802.15.4 HRP UWB standard
• Secure Ranging with IEEE 802.15.4z HRP UWB (arXiv 2312.03964)
• Bluetooth SIG: High-accuracy, low-power, secure ranging using Channel Sounding
• FiRa Consortium Core 3.0 announcement

By Riju — about