Commodity CSI: Capture Limits and Calibration

Research papers often use USRPs, Intel 5300 NICs (3×3 MIMO, 9 antenna pairs), or dense WiFi grids. Person-in-WiFi used 2× Intel 5300 sets with 30 subcarriers per link — 270 scalar measurements per frame. WiVi pushed 100 pkt/s with custom OpenWrt firmware on TP-Link routers, producing 100×114 CSI frames. Wavey uses a $5 ESP32. This post is about what that hardware actually delivers, where it falls short, and when it is enough.

What ESP-CSI exposes

Espressif’s ESP-CSI library reports per-packet CSI from the WiFi PHY:

• Subcarriers: 52–56 usable data subcarriers per 20 MHz channel (varies by chip and config).
• Antennas: typically 1 RX chain on most ESP32 modules (ESP32-S3 with external antenna can do better).
• Format: complex IQ per subcarrier — amplitude and phase as reported by the baseband.
• Rate: one CSI vector per received WiFi frame. Effective sampling rate = packet arrival rate, typically 20–200 Hz depending on traffic and configuration.

You do not control the transmitter — you sniff CSI from ambient WiFi traffic or a dedicated AP sending beacons/data to the node.

The phase noise floor

On a USRP with external clock sync, phase is stable enough to track sub-millimeter displacement. On COTS WiFi without sync, CARM showed CFO can reach 100 kHz — phase rotates by tens of radians per packet. On an ESP32:

• CFO (carrier frequency offset) between TX and RX rotates phase by degrees per packet.
• SFO (sampling frequency offset) adds a linear phase slope across subcarriers.
• Packet boundary jitter introduces random phase steps between consecutive samples.
• Quantization in the reported IQ values adds discretization noise.

For macro-motion HAR, amplitude is the primary signal on commodity hardware. The biological signal from breathing (0.1–0.5 Hz, sub-wavelength displacement) sits below the hardware noise floor in raw phase. Phase sanitization can recover micro-motion in controlled setups — remove CFO/SFO trends, work with differences, filter outliers. See the sensing pipeline for the preprocessing stack.

Amplitude variance and spectral features on amplitude spectrograms work well on ESP32 without heroic calibration. This is why occupancy and motion — which rely on amplitude dynamics — are the practical starting points.

Sampling rate constraints

At 50 packets/second, Nyquist says you can resolve signals up to 25 Hz. Walking (1–5 Hz) is well within range. Breathing (0.1–0.5 Hz) is fine. Heartbeat (0.8–2 Hz) is borderline — you need consistent packet rate and clean phase. CARM recommended ~800 CSI samples per second for fine activity segmentation; ESP32 deployments typically run at 20–200 pkt/s depending on traffic. That gap limits fine-grained HAR and impulse detection (falls, sit-downs) compared to research rigs.

Single antenna vs research rigs

Research systems with 3×3 MIMO (9 antenna pairs) extract spatial diversity — beamforming, angle of arrival, multi-person separation. Person-in-WiFi needed 9 TX-RX pairs to make pose estimation tractable. GoPose reconstructs 2D angle-of-arrival spectra from multi-receiver CSI. A single-antenna ESP32 sees one superposition of all scatterers. You get:

• Binary occupancy — yes, reliably.
• Motion intensity — yes.
• Coarse activity — yes, with labeled data.
• Person count — rough at best.
• Pose / localization — no, not without multiple nodes at different geometry.

Multiple ESP32 nodes at different positions partially compensate — each node sees a different superposition. Three nodes in a room give three perspectives, not one array, but enough spatial diversity for zone-level detection.

When cheap hardware is enough

The ESP32 is not a limitation for the tasks Wavey targets. It is a limitation for research-frontier tasks that need dense spatial sampling — and that is an honest scope boundary, not a product failure.

Further reading

• Getting started — deployment architecture
• The WiFi channel as a sensor — what the CSI matrix contains
• Detection ladder — task feasibility on commodity hardware