Interrogating a CSI Sensing Vendor: Where the Pipeline Breaks and What Survives

Ask the vendor what happens to presence detection when the access point that initiates sensing sessions reboots. Then listen for whether the answer describes an architecture or apologises for the lack of one. CSI sensing chains radio firmware, a measurement protocol, a transport path and an inference model, and each link fails in its own way. With a shortlist and little lab time, the useful questions are the ones that show which link takes the others down with it.

Who owns the measurement session

Under the Wi-Fi sensing amendment, IEEE 802.11bf, a sensing initiator sets up measurements with responders, schedules sounding and collects reports. Ask whether that role can move to another AP or a client station when the initiator disappears, and how long it takes to set up the measurements again. The risk behind the question is a hidden single point of failure: if every measurement instance is tied to one AP's association state, a firmware crash quietly ends sensing across the whole zone while the network still looks healthy, because data traffic roams and sensing does not. Ask too whether the system tells 'no motion' apart from 'no measurements', because a model that reads an empty stream as an empty room is the worst failure you can buy. A vendor that exposes heartbeat and staleness flags on its output API has thought about this, and one that only shows accuracy curves has not.

Calibration that outlives a firmware update

CSI is not a clean physical measurement. Raw subcarrier phase carries carrier frequency offset, sampling time offset and automatic gain control steps, and vendors remove these with sanitisation stages tuned to one chipset and driver build. The failure that matters is a routine AP firmware update that changes the AGC table or the antenna mapping. The CSI amplitude distribution shifts, and a classifier trained on last quarter's data starts reporting occupancy that isn't there or missing falls. Ask whether the vendor detects this shift at runtime, whether it can fall back to a coarser amplitude-only or variance-based detector while it recalibrates, and whether recalibration needs a site visit or runs from an empty-room baseline captured automatically. Ask also who controls the firmware schedule, because if the network team patches APs on its own timetable, sensing accuracy depends on a change process the sensing vendor never sees.

Degraded mode, stated in writing

Hold the vendor to a written degraded mode: if contended airtime stretches the sensing report interval beyond 10 ms, they should name which detection classes are still valid at the slower rate, because gesture recognition fails long before presence detection does.

Cellular fallback and the fusion layer

Several shortlisted designs pair Wi-Fi CSI with 5G-Advanced positioning or sensing on a private cell, and sell the pairing as redundancy. That redundancy only holds if the fusion layer treats each input as independently fallible: weighted by confidence, aligned to a common clock, and able to run on one input without retraining. Ask where fusion runs, because an edge box that takes in both streams is itself a single point of failure unless each radio can publish a lower-confidence result directly. Ask for the timestamp alignment budget too, since misaligned streams fused in good faith produce confident wrong answers, which are harder to catch than a dropout. With limited lab time, apply one test: in the single session you get, unplug the initiator AP and then the fusion host, and treat any vendor whose output goes stale without flagging it as a single point of failure, whatever its bench accuracy. As an estimate to bring into negotiation, expect one to two weeks of re-baselining after any major firmware change unless the vendor demonstrates automated recalibration, and write that cost into the support contract.

Back to CSISensing · Domains

← Back to CSISensing
Engineering & Architecture Brief

Interrogating a CSI Sensing Vendor: Where the Pipeline Breaks and What Survives

Ask the vendor what happens to presence detection when the access point that initiates sensing sessions reboots. Then listen for whether the answer describes an architecture or apologises for the lack of one. CSI sensing chains radio firmware, a measurement protocol, a transport path and an inference model, and each link fails in its own way. With a shortlist and little lab time, the useful questions are the ones that show which link takes the others down with it.

Who owns the measurement session

Under the Wi-Fi sensing amendment, IEEE 802.11bf, a sensing initiator sets up measurements with responders, schedules sounding and collects reports. Ask whether that role can move to another AP or a client station when the initiator disappears, and how long it takes to set up the measurements again. The risk behind the question is a hidden single point of failure: if every measurement instance is tied to one AP's association state, a firmware crash quietly ends sensing across the whole zone while the network still looks healthy, because data traffic roams and sensing does not. Ask too whether the system tells 'no motion' apart from 'no measurements', because a model that reads an empty stream as an empty room is the worst failure you can buy. A vendor that exposes heartbeat and staleness flags on its output API has thought about this, and one that only shows accuracy curves has not.

Calibration that outlives a firmware update

CSI is not a clean physical measurement. Raw subcarrier phase carries carrier frequency offset, sampling time offset and automatic gain control steps, and vendors remove these with sanitisation stages tuned to one chipset and driver build. The failure that matters is a routine AP firmware update that changes the AGC table or the antenna mapping. The CSI amplitude distribution shifts, and a classifier trained on last quarter's data starts reporting occupancy that isn't there or missing falls. Ask whether the vendor detects this shift at runtime, whether it can fall back to a coarser amplitude-only or variance-based detector while it recalibrates, and whether recalibration needs a site visit or runs from an empty-room baseline captured automatically. Ask also who controls the firmware schedule, because if the network team patches APs on its own timetable, sensing accuracy depends on a change process the sensing vendor never sees.

Degraded mode, stated in writing

Hold the vendor to a written degraded mode: if contended airtime stretches the sensing report interval beyond 10 ms, they should name which detection classes are still valid at the slower rate, because gesture recognition fails long before presence detection does.

Cellular fallback and the fusion layer

Several shortlisted designs pair Wi-Fi CSI with 5G-Advanced positioning or sensing on a private cell, and sell the pairing as redundancy. That redundancy only holds if the fusion layer treats each input as independently fallible: weighted by confidence, aligned to a common clock, and able to run on one input without retraining. Ask where fusion runs, because an edge box that takes in both streams is itself a single point of failure unless each radio can publish a lower-confidence result directly. Ask for the timestamp alignment budget too, since misaligned streams fused in good faith produce confident wrong answers, which are harder to catch than a dropout. With limited lab time, apply one test: in the single session you get, unplug the initiator AP and then the fusion host, and treat any vendor whose output goes stale without flagging it as a single point of failure, whatever its bench accuracy. As an estimate to bring into negotiation, expect one to two weeks of re-baselining after any major firmware change unless the vendor demonstrates automated recalibration, and write that cost into the support contract.

Related Architecture Specifications (ISACECOSYSTEM): isacgrid.com · isacnetworks.com · isacradio.com

More Technical Articles on csisensing.com

Phase Sanitization Algorithms for Multipath Channel State Information in Commodity Wi-Fi →Deep Convolutional Neural Networks for Device-Free Human Activity Classification from CSI →CSISensing: Reference Architecture, Interoperability & Deterministic Telemetry →