Single-sided through-wall sensing: next steps after a working ESP32-S3 2.4 GHz baseline

[goncalofrankefranco]

Hi Ruvnet,

Me and my team have now confirmed a working baseline with a single ESP32-S3 on 2.4 GHz using a normal omnidirectional router, and it can detect human presence through a wall from one side. The result is still coarse, more like a blob/disturbance than precise localization, but the proof of concept is real.

Our main use case is single-sided sensing from one side of the wall only. Given that baseline, we’d like guidance on how to improve the system in the most effective low-cost way.

A few questions:

Do you think adding more ESP32-S3 nodes would still help in a unilateral setup, or does the benefit only really appear in multi-sided / multistatic deployments?

Since we are currently working at 2.4 GHz, do you plan to implement support for Nexmon SDR / Raspberry Pi 3B+ or Pi 4 as a transmitter source in RuView, especially for experiments where the transmitter is controlled more precisely?

How do you see the tradeoff between 2.4 GHz, which seems to work better for wall penetration, and higher bands like 5.8 GHz or 24 GHz, which should improve spatial precision but may lose penetration through walls?

For the antenna geometry, how much benefit do you expect from using aluminum reflectors or a partial aluminum backplane to make the signal more directional on the single-sided setup? Would aluminum reflectors or backplanes meaningfully improve directivity, and what physical arrangement works best?

What would you do next after a working single-sided 2.4 GHz baseline to improve the project most effectively?

Hope to hear from you soon!

[ruvnet]

Hi Gonçalo — congrats on the working baseline. Through-wall presence detection from a single side with off-the-shelf ESP32-S3 + omni router is genuinely hard, so the fact that you're getting anything coherent is the load-bearing result. Some practical answers:

More nodes in a unilateral setup

Yes, but the gain shape is different from multistatic.

• Multistatic (nodes on opposite sides of the wall) buys you geometric diversity — different wall transmissions, low cross-correlation between paths. Roughly exponential in path independence.
• Unilateral (all nodes on your side of the wall) only adds angular diversity for the same TX-through-wall-back-out-through-wall path. You can do AoA via MUSIC/ESPRIT across the RX baseline → 2D localisation in the half-space behind the wall, but each new RX adds roughly linear improvement, not geometric.

What I'd target next: 3 RX nodes spread along a 2-3 m baseline, plus the existing TX. The signal/src/ruvsense/multistatic.rs fusion path already supports this — it's what powers the multistatic eigenvalue counter (field_model.rs). The geometry plumbing is in crates/wifi-densepose-ruvector/src/viewpoint/geometry.rs (Cramer-Rao bounds, geometric diversity index) so you can compute, for any node placement, what theoretical localisation accuracy you should expect.

Nexmon SDR / Pi 3B+/4 as transmitter

No current plan in this repo to integrate Nexmon — supported hardware is ESP32-S3 family + ESP32-C6 (mmWave) per ADR-028. But the question is what control over TX you actually need. For deterministic through-wall sensing the wins are:

1. Known waveform — already true if you control the AP-mode ESP32 (deterministic beacon interval, no hidden retransmits)
2. Frame-schedule control — true with two ESP32-S3s, one in AP mode beaconing on a fixed schedule
3. TX-side CSI extraction — this is where Nexmon-on-Pi adds something the ESP32 can't: you get CSI from the TX's perspective in addition to the RX's. Useful for joint TX+RX calibration. Not the highest-leverage next step IMO.

Pragmatic next-step on TX: a second ESP32-S3 in AP mode pinned to channel 6/11, fixed beacon interval, no other clients allowed. Cheap and deterministic. If you want CSI from both sides, then Pi+Nexmon is worth it.

2.4 vs 5.8 vs 24 GHz

Band	λ	Wall loss (drywall)	Theoretical resolution	Practical for through-wall?
2.4 GHz	12.5 cm	3-15 dB	~6 cm	Best — primary band
5.8 GHz	5 cm	6-25 dB	~2.5 cm	Marginal — fails on reinforced concrete or foil-backed insulation
24 GHz	1.25 cm	>40 dB	~6 mm	No — short-range LoS only (this is what the LD2410/MR60BHA2 mmWave sensors in the repo are for)

Don't pick one — fuse them. RuvSense's multiband.rs (multi-band CSI frame fusion + cross-channel coherence) is designed for exactly this. Stay 2.4 GHz for the through-wall geometry; if you have one antenna pair with line-of-sight to the wall on the same side, add a 5 GHz capture and let MultiBandCsiFrame combine them. The 5 GHz path won't go through the wall but it gives you a clean reference for the path that does.

Aluminum reflectors / backplane

Real and worth doing. The math:

• A flat metal reflector ¼-wavelength behind the antenna gives 5-6 dB forward gain (constructive reinforcement). For 2.4 GHz: place ~3.1 cm behind the antenna.
• A corner reflector (two flat sheets at 90° behind the antenna) gives ~9-12 dB. Same ¼-λ spacing.
• A flat backplane with bent side wings at ~60° approximates a corner reflector and is much easier to make.

5-7 dB SNR at the wall is realistic for ~30 minutes of DIY and a sheet of aluminium. The cost is angular coverage: you narrow the antenna pattern, so sources outside that cone get attenuated. For through-wall presence with a known target room this is a win (you actively want to reject in-room reflections). For multi-room sensing it's a problem.

For 2.4 GHz a real horn is impractical (12-15 cm aperture for meaningful directivity), but the corner-reflector approach is in the same ballpark with much less work.

What I'd do next, in order

1. Calibrate the empty target room — 10+ minutes of CSI with no one behind the wall, persist a field model (signal/src/ruvsense/field_model.rs, ADR-030). The eigenmodes of the empty room are the dominant variance; the residual after subtracting them is your signal. If you skip this step, every analysis below is fighting the room.
2. Add 2 more RX nodes with ~1 m baselines, run the existing multistatic fusion (MultistaticFuser in the same crate). You should immediately see better localisation precision per the Cramer-Rao bounds in viewpoint/geometry.rs.
3. Build a corner reflector for each RX. 30 min, ~5-9 dB.
4. Start recording continuously — through-wall pose is a hard inverse problem and the only thing that beats it is data + training. The wifi-densepose-train crate (ADR-016) consumes recorded JSONL and trains a model. The "blob/disturbance" you're seeing is an inverse-problem prior failing; the way out is dataset.
5. Try the AETHER re-ID embeddings (ADR-024) once you have enough data — they help the system stay locked on the same target through wall transmission, which is the failure mode that turns a real person into "a blob that drifts."

What not to bother with yet

• Going to 5 GHz only / abandoning 2.4 GHz — you'll lose penetration faster than you gain resolution.
• Big antenna arrays (8+ elements) — too far ahead of where the bottleneck is right now (data + room calibration).
• Dropping the edge ESP32 firmware in favour of pure Pi+Nexmon — Pi+Nexmon is a complementary capability, not a replacement.

Happy to talk through specifics if any of this lands wrong for your environment. The setup that makes this hard is mostly reinforced concrete, foil-backed insulation, and active interferers in your wall's pass-band — if you're hitting any of those, the priorities shift.

---

Generated by Claude Code