3× ESP32-S3 presence-detection build — a control method to separate real presence from channel drift, my honest results so far, and where I'd love guidance

[bcarpio]

First: thanks for open-sourcing this. I'm building toward exactly the use case this is perfect for — whole-home presence without any indoor cameras (I run a local Home Assistant setup and deliberately keep cameras outside only). WiFi CSI is the one modality that could give me per-room "someone's here" without a lens, so I really want to get this working, and I've got the hardware set up and staying up.

Posting this as its own thread to (a) share a control experiment I think is useful for anyone measuring presence, (b) give honest numbers from my room, and (c) ask the folks who've gotten further what I should try next.

One distinction I wanted to nail down for myself, since a lot of the discussion elsewhere turns on it: "the nodes stream real CSI" and "presence detection works" are two separate claims. I confirmed the first solidly; the second I haven't cracked yet, and I'd like help figuring out whether that's placement, phase, calibration, or something else.

Setup

• 3× ESP32-S3-DevKitC-1 (WROOM-1 N16R8), firmware v0.6.7, re-provisioned --edge-tier 0 (raw passthrough), unique --node-id, --tdm-slot 0/1/2 --tdm-total 3.
• Server ruvnet/wifi-densepose:v0.8.4-esp32, --net=host, --source esp32, /health → "source":"esp32", all three nodes in one bedroom on the same AP.

Confirmed it's real hardware, not sim mode: I tcpdump'd UDP frames from the three ESP32 IPs and hand-decoded them against the ADR-018 spec in #34 — magic 0xC5110001, correct header, int8 I/Q, 64-subcarrier HT20 (some 128), RSSI matching each board's placement, subcarrier profiles with the expected DC-null shape and real temporal variation. So the pipeline genuinely works end to end. 🎉 That part was really satisfying to see.

A control experiment I'd suggest for everyone

The thing that tripped me up (and might be tripping up some positive reports): a naive occupied-vs-empty test is confounded by time-of-day channel drift. So I ran three windows:

• 5.9 h occupied overnight (asleep in the room)
• a fresh empty baseline after leaving
• an occupied-again window 37 min after the empty one — same time of day, so drift is controlled

Amplitude window features, balanced classes, AUC (imbalance-robust), 20–30 draws:

Node	night-occupied vs empty (confounded)	occupied-again vs empty (drift-controlled)
node1	1.000	1.000
node2	0.657	0.607
node3	0.963	0.581
all 3	0.752	0.576

The confounded column looks great (node3 0.96!). But the controlled column is the honest one, and it tells a cautionary tale:

• node3's 0.96 was mostly overnight drift — controlled, it drops to ~chance.
• node1's perfect 1.0 was a link artifact — during the empty window its RSSI was −59 dBm vs −87 dBm in both occupied windows (a 28 dB swing that reverses when re-occupied = a re-association, not a person).
• Controlled, stable-link, combined: AUC ≈ 0.58, so with amplitude-only features I'm not yet separating presence from noise.

This lines up with the server's own adaptive classifier for me: after room calibration (recorded train_empty_room/train_still_presence/train_walking/train_active_movement, correctly split into absent/present_still/present_moving/active), training accuracy was 0.45 on 4 classes, and a known-empty room still read "present" 31/31 frames. So I don't think I'm doing something obviously wrong — I think I'm missing an ingredient.

(Two useful robustness notes: the per-node AUCs don't depend on cross-node TDM sync, so imperfect TDM isn't the cause; and the same-time-of-day control neutralizes the documented "fan/AC false-positive" caveat, since a stationary interferer is present in both windows.)

Where I think the signal might be — and where I'd love input

I'm fairly convinced the info is there and I'm just not extracting it. My leading suspects, and I'd love to know which to prioritize:

1. Phase. I used amplitude only and threw away phase — but breathing/presence is often in the phase and the 0.1–0.5 Hz band. Is phase the main lever here? Any recommended CSI phase-sanitization (linear detrend / conjugate-multiply between antennas) for these boards?
2. Placement. Mine were ~4 m apart, one bedroom, not rigorously to the placement guide (height/diagonal). How much does getting this right actually move presence AUC in your experience?
3. The pretrained encoder. I dug into ruvnet/wifi-densepose-pretrained — it's an 8→64→128 MLP on the 8 summary features (not raw CSI), and its 82.3% is heldout_triplet_acc (random encoder ~69.6% on the same metric), so I set expectations that it's a representation model rather than a presence classifier. Is there a presence-specific model/head I should be loading instead, and what's the current jsonl→rvf path (saw Training does not start – /ws/train/progress returns 404 and no model is generated #1233/Training is badly documented #1322 — is there a working converter yet)?
4. Feature window / cadence. Would longer windows, or the breathing-band spectral features specifically, be the unlock over my crude window stats?

Offer

I've got the 3-node rig up and staying up, a clean overnight raw-CSI capture (749k frames, ADR-018 parsed), and Python parse/analysis scripts I'm happy to share so results are comparable. I'll gladly re-run with whatever placement/firmware/feature changes people suggest and post updated controlled numbers. If a phase-aware pipeline is the key, I've even got idle GPUs to throw at training a room-specific model.

Genuinely want to get this to a reliable boolean "room occupied" — that alone would be hugely useful for camera-free home automation. What would you try next?