Bluetooth Direction Finding Explained: AoA vs AoD RTLS Architecture, IQ Sampling, Antenna Arrays and Real-World Accuracy
Vendor pages promise “centimeter-level” accuracy, yet a 2025 Imperial College study of a five-anchor classroom deployment measured walking-path errors averaging about one metre, with outliers of eight. Both statements can be true. Bluetooth direction finding is a physics-limited measurement, and its error budget is set far more by geometry, multipath and calibration than by the radio chip. That gap matters in 2026 because buyers now choose between three Bluetooth-adjacent location options: Angle of Arrival (AoA), Angle of Departure (AoD), and the newer Channel Sounding ranging in Core 6.x. Ultra-wideband (UWB) sits alongside them. This article takes the mechanism-first view. It walks the Constant Tone Extension bit by bit, shows how IQ samples become angles, lays out a production real-time location system (RTLS) reference architecture, and reads the published accuracy data honestly.
What this covers: CTE packet anatomy, IQ sampling maths, antenna arrays, AoA vs AoD, angle estimators, connectionless vs connected modes, MQTT backhaul, capacity and battery maths, failure modes, a deployment checklist and how direction finding compares with Channel Sounding and UWB.
Context and Background
Bluetooth Low Energy (BLE) location began with received signal strength indication (RSSI). A receiver reports how loud a beacon is, and software converts loudness into distance with a path-loss model. RSSI is cheap and universal, but it is noisy. Body shadowing, antenna orientation and reflections easily move a reading by 10 dB, which translates into metres of range error. Proximity and room-level presence work well with RSSI. Pallet-level or aisle-level tracking does not.
Direction finding was added in Bluetooth Core Specification 5.1, published in early 2019. The Bluetooth SIG’s own technical overview by Martin Woolley, first dated 20 March 2019, describes the feature in detail. Instead of measuring amplitude, the receiver measures phase. A known, unmodulated tone called the Constant Tone Extension (CTE) is appended to a packet. An antenna array samples that tone on different elements, and the phase differences between elements reveal the direction the wave came from. Two or more bearings from known locator positions intersect at the tag.
Before 5.1, the Finnish company Quuppa had already built a proprietary BLE angle-of-arrival system, and it still markets “from 10 cm precision” and tag battery life “up to 10 years” (vendor claims, conditions not stated). Core 5.1 standardised the over-the-air part so that commodity BLE silicon from Silicon Labs, Nordic Semiconductor, Texas Instruments and others could emit and sample CTEs. Positioning engines, antenna arrays and calibration remain vendor territory, which is why accuracy varies so widely between products.
The landscape changed again on 3 September 2024, when the SIG released Core 6.0 with Channel Sounding. Channel Sounding measures distance rather than angle, using phase-based ranging (PBR) and round-trip time (RTT). It has since been refined in Core 6.2 and Core 6.3, the latter dated 5 May 2026. We compared Channel Sounding with UWB in detail in our factory RTLS analysis of IEEE 802.15.4z UWB versus BLE Channel Sounding. This post fills the gap underneath that one: how the angle-based half of Bluetooth location actually works. For the primary source, read the Bluetooth SIG direction finding technical overview.
The thesis here is specific. Bluetooth direction finding is best understood as an infrastructure-heavy, tag-light system. Almost all cost, complexity and accuracy live in the locators, their placement and the software that fuses their bearings. Teams that budget for tags and treat locators as commodity “gateways” are the teams that end up with metre-level results.
Bluetooth Direction Finding Reference Architecture
Bluetooth direction finding locates a BLE device by measuring the phase of a Constant Tone Extension across an antenna array. In AoA, a single-antenna tag transmits and multi-antenna locators compute bearings. In AoD, array beacons transmit and the mobile device computes its own bearing. A positioning engine intersects bearings into coordinates.

Figure 1: A production Bluetooth AoA RTLS reference architecture, from tag to digital twin.
Figure 1 reads left to right. A tag with one antenna transmits packets carrying a CTE. Each locator in range samples that CTE across its array. Locators either convert IQ samples into an angle on board or forward raw IQ data for central processing. Angle reports flow over a message broker, usually MQTT, to a positioning engine. The engine triangulates, filters and publishes positions to applications and to a site model or digital twin.
The tag is deliberately dumb
In an AoA system the tag needs nothing special beyond a Core 5.1 controller that can append a CTE. It needs no array, no RF switch and no knowledge of where it is. That is the core economic argument for AoA RTLS. A site may have thousands of tags and only dozens of locators. Pushing complexity into the fixed infrastructure keeps the per-tag bill of materials close to a normal BLE beacon.
The cost of that choice is that the tag cannot help with accuracy. It cannot correct its own antenna orientation, and it cannot report its own multipath environment. Every error correction must happen downstream. This is the opposite of the AoD model, where the mobile device carries the burden and the infrastructure is simple beacons with arrays.
The locator is where accuracy is won or lost
A locator combines three subsystems. The first is an antenna array, typically a uniform rectangular array (URA) of patch antennas. The second is an RF switch network driven by the radio’s antenna-switch pins. The third is a radio that captures IQ samples and a processor that turns them into angles.
Concrete reference designs show the range of array sizes. Silicon Labs’ BG22 Dual Polarized Antenna Array Radio Board (BRD4191A) is a 4×4 array, 160 x 160 mm, with 16 dual-polarized patches. Each patch has a vertical and a horizontal element, giving 32 switchable antennas through a chain of RF switches. u-blox’s ANT-B10 board used eight patch antennas driven by a NINA-B411 module, although u-blox now lists its XPLR-AOA explorer kits as no longer available. Dual polarization matters because a tag’s antenna orientation is unknown. A single-polarized array can lose most of the signal when a tag lies flat on a pallet.
Element spacing is bounded by physics. To avoid ambiguous phase readings, adjacent elements should be no more than half a wavelength apart. At 2.4 GHz the wavelength is about 12.5 cm, so half-wavelength spacing is roughly 6.2 cm. MathWorks’ Bluetooth toolbox documentation states this constraint as a maximum spacing of λ/2. Wider spacing gives sharper angular resolution but causes grating lobes, which are false directions that look identical to the true one.
The positioning engine turns bearings into coordinates
A single bearing only says “the tag is somewhere along this ray”. With two locators the rays intersect in 2D. With three or more, and with both azimuth and elevation, the engine can solve 3D position by least squares and reject inconsistent bearings. Silicon Labs describes exactly this in its Real-Time Location (RTL) library documentation. The library takes locator positions and AoA estimates as inputs and outputs a triangulated tag position, using spatial and temporal filtering and averaging across frequencies to suppress reflections.
The engine is also where tracking lives. Raw positions jitter, so production systems run a Kalman filter or particle filter with a motion model. That filter is the same machinery used to keep models aligned with sensors, which we covered in our guide to data assimilation for digital twins with Kalman filters. A forklift tag should be modelled with different process noise than a tag on a slow-moving tote.
Why MQTT is the default backhaul
Silicon Labs’ reference implementation uses MQTT to gather angle data from locators. That choice lets the host applications run either on one PC through localhost or across a distributed network. MQTT suits the pattern well. Locators are publishers, the positioning engine is a subscriber, and topics can be organised by locator and tag.
At scale, the broker design matters. One positioning engine instance can become a bottleneck when hundreds of locators each report many tags per second. Shared subscriptions let several engine workers split the load, provided that all reports for one tag land on the same worker. We discuss that pattern in our article on MQTT 5 shared subscriptions and broker clustering. The QoS choice also matters: angle reports are perishable, so QoS 0 is usually fine, while position outputs feeding alarms may justify QoS 1.
Deeper Analysis: From Radio Wave to Position
This section walks the Bluetooth direction finding signal chain in order: the packet, the sampling, the angle maths, the connection modes, and finally capacity and power.
Anatomy of the Constant Tone Extension

Figure 2: How a Constant Tone Extension is appended after the CRC, with guard, reference and switch or sample slots.
Figure 2 shows a normal BLE packet followed by the CTE. The packet header carries a CTEInfo field that tells the receiver a CTE follows, how long it is and what type it is. The CTE itself starts after the CRC. According to the SIG’s technical overview, it consists of a 4 µs guard period, an 8 µs reference period, and then alternating switch slots and sample slots. The CTE length runs from 16 µs to 160 µs, set in 8 µs units, so the allowed values are 2 to 20. Silicon Labs’ stack API documents exactly this “Range: 2 to 20”.
The CTE is a run of binary ones that is not whitened. Excluded from the CRC, MIC and whitening, it produces a steady tone. Nordic’s documentation notes that on the LE 1M PHY this tone sits 250 kHz above the carrier, and on LE 2M it sits 500 kHz above. Direction finding works on LE 1M or LE 2M but not on the LE Coded PHY, which rules out long-range coded links for angle measurement.
Slots are either 1 µs or 2 µs long. The SIG overview states that 2 µs slots are mandatory to support and 1 µs slots are optional. Silicon Labs encodes the CTE type as 0 for AoA, 1 for AoD with 1 µs slots and 2 for AoD with 2 µs slots.
How many IQ samples you actually get
During the reference period, the receiver stays on one antenna and takes 8 IQ samples at 1 µs intervals. After that, only the sample slots produce data; switch slots are discarded while the RF switch settles. Silicon Labs gives a clean formula for the sample count:
N = 8 + (L − 12) / 2 / S
Here L is the CTE length in microseconds and S is the slot length. For the maximum 160 µs CTE with 2 µs slots, N = 8 + 148 / 4 = 45 samples. With 1 µs slots, N = 8 + 148 / 2 = 82 samples. That is the entire raw material for one angle estimate. A 16-element array sampled with 2 µs slots therefore sees each element only two or three times per packet. This is why engines average across many packets.
Each IQ sample is a point in the complex plane. I is the in-phase component and Q the quadrature component, so the phase is atan2(Q, I) and the amplitude is the vector length. The standard HCI report carries 8-bit signed I and Q values. Zephyr’s locator sample prints them from a BT_DF_IQ_SAMPLE_8_BITS_INT report and also supports a vendor-specific 16-bit format on Nordic controllers.
Frequency offset compensation is the first trap
The tone is not at exactly 250 kHz. Transmitter and receiver crystals disagree, adding an offset that Silicon Labs puts on the order of 10 kHz. Between consecutive 1 µs samples, the phase therefore advances by roughly 90 degrees, but the real value can vary between 80 and 100 degrees.
If you ignore that drift, it masquerades as an antenna phase difference. The reference period exists to solve this. Because all 8 reference samples come from one antenna, their phase progression measures the true rotation per microsecond. Silicon Labs “strongly recommends” estimating this per packet, by averaging or median-filtering the 8 phase differences. The estimate is then scaled to the gaps between later samples: multiplied by 2 for 1 µs slots and by 4 for 2 µs slots, because switch slots sit between sample slots. Only after subtracting that predicted rotation do the residual phase differences reflect geometry.
From phase difference to angle
For two elements spaced d apart, a plane wave arriving at angle θ from the array axis produces a phase difference ψ = 2πd·cosθ / λ. Solving gives θ = arccos(ψλ / 2πd). For AoD, MathWorks writes the same relation with sine, β = arcsin(ψλ / 2πd), because the angle is measured from the array normal.
The sensitivity of this formula is uneven, and that unevenness drives real deployments. With d = λ/2, ψ = π·cosθ, so the derivative is dψ/dθ = −π·sinθ. When the wave arrives broadside (θ near 90 degrees), a small angle change produces a large phase change, and estimates are precise. Near endfire (θ near 0 degrees), the derivative approaches zero, and noise in ψ turns into large angle errors. The 2021 experimental study by researchers using Texas Instruments CC26X2R LaunchPads observed exactly this: angular standard deviation rose near 0 degrees, especially indoors. Practical rule: a locator is most accurate for tags roughly in front of it, not off to the side.
Two-element phase comparison is the simplest estimator. Production systems use the whole array. Classic beamforming (Bartlett) steers a virtual beam across candidate angles and picks the peak power. Capon or minimum-variance beamforming sharpens peaks by suppressing interference. MUSIC, introduced by R. O. Schmidt in IEEE Transactions on Antennas and Propagation in 1986, splits the sample covariance into signal and noise subspaces and searches for steering vectors orthogonal to the noise subspace. MUSIC can resolve a direct path from a strong reflection when the array has enough elements and enough snapshots, which is one reason vendors favour 4×4 arrays. Its catch is that coherent reflections, which are copies of the same signal, defeat the subspace split unless the engine applies spatial smoothing.
Connectionless versus connection-oriented modes

Figure 3: The connectionless AoA flow, from periodic advertising sync to a published position.
Core 5.1 defines two ways to deliver a CTE for Bluetooth direction finding. In connection-oriented mode, one device sends an LL_CTE_REQ and the peer replies with an LL_CTE_RSP that carries the CTE. In connectionless mode, the CTE rides on periodic advertising packets, specifically the AUX_SYNC_IND PDUs. All four combinations of AoA or AoD with either mode are valid, and controller support for each is optional. The SIG overview expects AoA to be used mainly with connections and AoD mainly connectionless, but most commercial AoA asset-tracking systems actually use connectionless periodic advertising, because a locator cannot hold a connection with thousands of tags.
Figure 3 shows the connectionless sequence. The tag runs extended advertising that includes SyncInfo, which points scanners at its periodic train. A locator creates a periodic sync, then follows the tag’s AUX_SYNC_IND packets on the data channels. For each packet it switches antennas through the CTE, captures IQ, compensates frequency offset and estimates azimuth and elevation. It publishes those angles and the RSSI to the broker, and the positioning engine intersects bearings from several locators.
The periodic-sync requirement is a real scaling constraint. Each locator must maintain a sync per tag, and controllers cap the number of concurrent syncs. Silicon Labs offers a proprietary alternative, “Silicon Labs CTE”, which attaches the CTE to extended advertising so locators need no sync. Its API notes that only one CTE per event is currently supported in that mode. The trade-off is standards portability: a non-Silicon Labs locator will not sample those packets.
Zephyr’s open-source samples make the configuration concrete. The beacon sample sets CTE_LEN to 0x14, which is 160 µs, and sends 5 CTEs per periodic advertising event. Its default periodic interval comes from Zephyr’s slow-interval constants, 0x0320 to 0x03C0 in 1.25 ms units, which is 1.0 to 1.2 seconds. The locator side enables CTE sampling on a periodic sync like this, taken from the upstream Zephyr sample:
const static uint8_t ant_patterns[] = { 0x1, 0x2, 0x3, 0x4, 0x5,
0x6, 0x7, 0x8, 0x9, 0xA };
const struct bt_df_per_adv_sync_cte_rx_param cte_rx_params = {
.max_cte_count = 5,
.cte_types = BT_DF_CTE_TYPE_ALL,
.slot_durations = 0x2, /* 2 us switch and sample slots */
.num_ant_ids = ARRAY_SIZE(ant_patterns),
.ant_ids = ant_patterns,
};
err = bt_df_per_adv_sync_cte_rx_enable(adv_sync, &cte_rx_params);
/* IQ reports arrive in the periodic-sync callback */
static void cte_recv_cb(struct bt_le_per_adv_sync *sync,
struct bt_df_per_adv_sync_iq_samples_report const *report)
{
for (uint8_t i = 0; i < report->sample_count; i++) {
int8_t I = report->sample[i].i;
int8_t Q = report->sample[i].q;
/* phase = atan2(Q, I); hand to the angle estimator */
}
}
Note what this Bluetooth direction finding sample does not do: it prints IQ data but computes no angle. Angle estimation, calibration and positioning are left to the application, which is the same split Silicon Labs makes between its raw CTE API and its RTL library.
Capacity maths (illustrative)
The following numbers are illustrative, derived from first principles rather than measured. On LE 1M, each octet takes 8 µs. A periodic packet with a short payload, say 31 octets including preamble, access address, header, payload and CRC, takes about 248 µs. Adding a 160 µs CTE brings it to roughly 410 µs. Five CTEs per event means five chained transmissions, so about 2 ms of airtime per tag per event, plus inter-frame spacing.
At one event per second, 1,000 tags consume around 2 seconds of airtime per second, spread across 37 data channels. Channel load per channel is therefore modest in theory. In practice the binding limits are the locator’s concurrent sync count, its radio time (it can only listen to one channel at once), and processing time for IQ sets. If a locator can track only a few dozen synced tags at a time, a dense zone needs either more locators, shorter CTEs, fewer CTEs per event or a slower update rate. Size this with your vendor’s published per-locator tag limit, and measure it in a pilot.
Battery maths (illustrative)
Battery life is also illustrative here and depends heavily on silicon. Assume a tag radio draws about 5 mA while transmitting at 0 dBm. If periodic plus extended advertising adds up to roughly 3 ms of active radio time per second, average transmit current is around 15 µA. Add a few microamps of sleep current and overhead, and the total might approach 20 µA. A 220 mAh coin cell would then last around 11,000 hours, or roughly 15 months, before derating for temperature and cell ageing.
The levers are clear from the arithmetic. Halving the update rate nearly doubles life. Using one CTE per event instead of five cuts airtime substantially but reduces samples per estimate. Motion-triggered rate changes, using an accelerometer to slow down when stationary, are usually the largest single win. Vendor claims of multi-year tag life generally assume such duty cycling.
What the accuracy data actually says
For Bluetooth direction finding accuracy, vendor marketing and peer-reviewed measurements tell different stories, so it helps to separate them.
| Source | Setup | Reported result | Type |
|---|---|---|---|
| Bluetooth SIG feature page | General claim | “centimeter-level location accuracy” | Marketing claim |
| Bluetooth SIG RDF technical overview | Conclusion | “sub-meter accuracy levels” | Overview claim |
| Silicon Labs location services page | General claim | “sub-meter location accuracy” | Vendor claim |
| Quuppa product page (Siemens site) | General claim | “from 10 cm precision” | Vendor claim |
| arXiv 2103.04121 (2021) | TI CC26X2R + BOOSTXL-AOA, two anchors, 10 m x 10 m | Angle error 1.83° indoor, 0.48° outdoor; position error 36.5 cm indoor, 22 cm outdoor | Measured |
| arXiv 2501.08805 (2025) | Five u-blox XPLR-AOA-2 anchors, ~55 m² room | Static horizontal 0.33–1.25 m; walking mean 1.06 m, max 8 m | Measured |
Methodology caveat: the two studies use different hardware, geometry and ground truth, so they are not a head-to-head benchmark. The 2021 paper’s abstract quotes an outdoor mean of 0.28°, which conflicts with 0.48° in its results section; 0.28° appears to be the standard deviation. The 2025 paper’s abstract claims sub-meter kinematic accuracy, while its kinematic averages exceed one metre.
The pattern is consistent. Angle errors of one to two degrees are achievable in controlled conditions. Static tags close to well-placed anchors reach a few tens of centimetres. Moving tags in real rooms see around a metre, with occasional large outliers. The 2025 study also found that systematic bias, between 0.44 m and 0.95 m per axis, dominated random error of about 0.2–0.3 m. That finding points directly at calibration, not at the radio, as the main lever.
A useful geometric rule for any Bluetooth direction finding deployment converts angle error to position error. Lateral error is roughly range times tan(angle error). At 10 m, a 2-degree error is about 35 cm; at 20 m it is about 70 cm. This is why locator spacing, not chip choice, sets the accuracy ceiling.
AoA vs AoD: Choosing the Right Mode
The AoA versus AoD decision is mostly about who needs to know the position. In AoA, the infrastructure knows where every tag is, which suits asset tracking, worker safety and inventory. In AoD, the mobile device knows where it is, which suits wayfinding and privacy-preserving navigation, because the device never transmits.
AoD flips the hardware burden. Beacons carry arrays and switch antennas while transmitting. The receiver needs only one antenna but must know the beacon’s switching pattern and array geometry, and it must run the angle estimation itself. That last requirement is why AoD has not become “indoor GPS” on phones. A phone app would need raw IQ access from the Bluetooth controller. To our knowledge, mainstream mobile operating systems do not expose CTE IQ samples to apps, which leaves AoD mainly for purpose-built receivers such as AGVs (automated guided vehicles) and dedicated handhelds. The SIG’s own direction finding page lists AGV guidance and venue navigation as AoD use cases.
| Criterion | AoA (Angle of Arrival) | AoD (Angle of Departure) |
|---|---|---|
| Array location | Locator (fixed) | Beacon (fixed) |
| Mobile side | Single-antenna tag | Single-antenna receiver that computes angles |
| Who knows position | Infrastructure | Mobile device |
| Privacy | Tags are tracked | Receiver stays silent |
| Scaling limit | Locator sync and compute capacity | Effectively unlimited receivers |
| Mobile compute | None | IQ processing on device |
| Typical use | Asset and people tracking | Navigation, AGV guidance |
| Phone support | Phone can act as a tag only if it can append a CTE | Needs IQ API access |
In short: choose AoA when you need a central picture of many cheap tags, and AoD when many receivers must self-locate privately. Hybrid sites sometimes run both on the same array hardware, because the Silicon Labs BG22 array kit supports both modes.
Bluetooth Direction Finding vs Channel Sounding and UWB

Figure 4: A decision tree for choosing between AoA, AoD, Channel Sounding, UWB and plain RSSI.
Figure 4 starts with the same question as the previous section: who computes the position. If the mobile device must self-locate, AoD is the Bluetooth answer. If the infrastructure computes, the next question is required accuracy. Zone-level needs are served by RSSI or proximity. Sub-meter needs with many low-cost tags point to AoA. Applications that need tens of centimetres plus resistance to relay attacks point to Channel Sounding or UWB.
Channel Sounding is a different measurement. Instead of measuring angle with an array, two devices exchange tones across many channels and compute distance. The SIG’s 3 September 2024 announcement described phase-based ranging for precision and RTT-based distance bounding to counter man-in-the-middle relay attacks. Core 6.2, the last release of 2025, added DFT-based detection of amplitude-modification attacks on ranging. Core 6.3 added inline transfer of phase-aligned tones and PHY-specific RTT accuracy reporting. Bluetooth direction finding has no comparable security story. A CTE is just a tone, and nothing in it authenticates the transmitter.
The practical difference is that Channel Sounding gives distance, not direction, and needs a connection between the two devices. Ranging to three or more anchors yields position by trilateration, but each anchor needs a ranging session. Direction finding gives direction from one locator, and connectionless AoA scales to many tags without connections. Some newer designs combine them: Channel Sounding for distance plus AoA for bearing from a single anchor.
UWB, standardised in IEEE 802.15.4z, measures time of flight with very short pulses, which makes it naturally resistant to multipath and capable of decimetre accuracy. It costs more per tag and draws more power. The digital-key ecosystem shows how these technologies get layered: our explainer on the Aliro digital key standard describes BLE for discovery and UWB for secure ranging, side by side.
Trade-offs, Gotchas, and What Goes Wrong
Multipath is the dominant error in Bluetooth direction finding. Indoors, a tag’s signal arrives directly and via reflections from racks, walls, vehicles and people. The array sees a superposition, and the phase it measures is a weighted blend of directions. The 2021 TI-based study measured indoor angle error almost four times larger than outdoor, attributing the gap to multipath. Metal-rich warehouses are worse than offices. Mitigations include more array elements, subspace estimators with spatial smoothing, averaging across many channels, and simply mounting locators so that the direct path dominates.
Calibration debt accumulates silently. Every array has per-element phase and gain offsets from cables, switches and manufacturing tolerances. Every installed locator also has a mounting orientation, and a few degrees of rotation become a fixed bias in every bearing. The 2025 Imperial study’s finding that bias outweighed random error is the warning. Locators bumped by forklifts or remounted during maintenance must be re-surveyed.
Polarization and body blocking. A tag on a lanyard rotates and gets shadowed by its wearer. Dual-polarized arrays help, but a body between tag and locator can make a reflection stronger than the direct path. People-tracking deployments need more locator overlap than asset-tracking ones.
Endfire and height. Because angle sensitivity collapses near endfire, tags far off to the side of a ceiling locator are poorly measured. Very high ceilings worsen elevation geometry and stretch range, so the same angle error costs more distance.
Clock drift and short CTEs. A short CTE reduces airtime but leaves fewer samples, and the offset estimate from 8 reference samples becomes the limiting noise. Do not shorten CTEs to save battery without re-measuring accuracy.
Anti-patterns. Treating locators as cheap gateways and spacing them like Wi-Fi access points. Shipping without a site survey. Running angle estimation centrally over lossy backhaul without timestamps. Forgetting that a tag is a broadcaster, so its periodic train can be passively tracked by anyone with a locator, which is a privacy issue for people tracking.
Practical Recommendations
Start from the use case, not the chip. Decide whether you need zone, sub-metre or decimetre accuracy, whether tags or receivers compute position, and whether ranging security matters. If you need sub-metre tracking of many low-cost assets and can invest in fixed infrastructure, AoA is the right Bluetooth direction finding mode. If you need secure ranging, look at Channel Sounding or UWB first.
Then design the locator grid as the product, because Bluetooth direction finding accuracy is set by geometry. Plan overlap so that every point in the coverage area is seen by at least three locators at favourable angles, preferably within their broadside cone. Mount arrays with known orientation, survey their positions with a total station or laser tool, and record that survey as configuration under version control. Pilot one representative zone, including the worst metal-rich area, and measure static and walking errors against surveyed ground truth before rolling out.
Finally, engineer the data path. Put timestamps on IQ or angle reports at the locator, use MQTT topic design that keeps one tag’s reports together, and run the tracking filter with motion models per asset class. Feed positions into your site model or digital twin only after filtering. If you run a Thread-based building network alongside, plan coexistence in the 2.4 GHz band; our OpenThread border router architecture guide covers that side of the stack.
Deployment checklist:
- Define accuracy targets as percentiles, for example P90 error, not as averages.
- Survey every locator’s position and rotation; re-survey after any physical change.
- Keep tags within each locator’s broadside cone where possible; avoid endfire geometry.
- Use dual-polarized arrays for tags with uncontrolled orientation.
- Start with 160 µs CTEs and tune shorter only after measuring the accuracy cost.
- Confirm per-locator sync and tag capacity with the vendor and in a pilot.
- Budget tag battery life with motion-triggered update rates.
- Log raw IQ for a sample of tags so you can re-run estimators offline.
Frequently Asked Questions
What is Bluetooth direction finding?
Bluetooth direction finding is a feature introduced in Bluetooth Core 5.1 that lets a receiver determine the direction of a BLE signal. A transmitter appends a Constant Tone Extension, an unmodulated tone of 16 to 160 µs, to a packet. A device with an antenna array samples the phase of that tone on different elements. The phase differences reveal the angle of arrival or departure. Combining bearings from several known locators gives the device’s position, typically for indoor asset tracking or navigation.
What is the difference between BLE AoA and AoD?
In Angle of Arrival (AoA), the mobile tag has one antenna and the fixed locator has an antenna array, so the infrastructure computes where each tag is. In Angle of Departure (AoD), the fixed beacon has the array and the mobile device has one antenna, so the device computes its own position. AoA suits asset and people tracking with many cheap tags. AoD suits navigation and AGV guidance where privacy matters, but it needs IQ access on the receiver.
How accurate is Bluetooth AoA positioning in practice?
Vendors claim sub-meter to centimetre-level accuracy, but peer-reviewed results vary with geometry and multipath. A 2021 study measured about 36.5 cm average error indoors with two anchors in a 10 m by 10 m area. A 2025 Imperial College study found static errors of 0.33 to 1.25 m and walking errors averaging about 1.06 m, with outliers to 8 m. Expect tens of centimetres in favourable layouts and around one metre in cluttered, dynamic spaces without careful calibration.
Do I need Bluetooth 5.1 hardware for direction finding?
Yes. The CTE and IQ sampling procedures are defined in Core 5.1 and later, and controller support is optional even then, so you need silicon that specifically supports CTE transmission or reception. Examples include Silicon Labs EFR32BG22 and BG24 parts and Nordic nRF52833 and nRF5340 devices used in the Zephyr samples. Tags for AoA need only CTE transmit support. Locators and AoD receivers need IQ sampling plus antenna-switch control, and the arrays are separate hardware.
Is Bluetooth Channel Sounding replacing direction finding?
Not directly. Channel Sounding, introduced in Core 6.0 in September 2024, measures distance using phase-based ranging and round-trip time, and adds relay-attack resistance that direction finding lacks. Direction finding measures angle and scales to many connectionless tags. They answer different questions and can be combined: Channel Sounding for range and AoA for bearing. For secure access and phone-centric use cases, Channel Sounding is likely to grow faster. For dense tag-based asset tracking, AoA remains practical.
How many locators does a Bluetooth AoA RTLS need?
There is no universal ratio, because accuracy depends on range, ceiling height and clutter. A useful rule is that every point should be seen by at least three locators at favourable angles. Position error grows with range times angle error: a 2-degree error costs about 35 cm at 10 m and 70 cm at 20 m. Start from your accuracy target, work back to a maximum locator-to-tag distance, and validate the spacing in a pilot zone before full rollout.
Further Reading
- UWB real-time location systems in the factory: IEEE 802.15.4z versus BLE Channel Sounding — the ranging-based counterpart to this angle-based guide.
- Aliro digital key standard explained: UWB, NFC and BLE credentials — how BLE and UWB are layered for secure access.
- OpenThread border router and Thread mesh architecture — the 2.4 GHz building network that often shares space with BLE RTLS.
- Bluetooth mesh architecture: managed flooding, provisioning and low-power nodes — the other major Bluetooth infrastructure pattern in buildings.
- The IoT, digital twin and PLM complete overview — where location data fits in the wider digital thread.
- External: Bluetooth SIG, Bluetooth Direction Finding: A Technical Overview and Silicon Labs, Interpreting IQ Samples.
By Riju — about
