Repository navigation
Per-node presence classifier reports present on 99.6% of an empty room (ESP32-C6) #2522
Description
Summary
smooth_and_classify_node in wifi-densepose-sensing-server reported presence: true on 6,093 of 6,104 observations (99.8%) recorded while the room was empty, and on 99.6% of the held-out half. It is effectively always on for this link. Details and method are in docs/adr/ADR-426-ruv-fann-proof-programme.md, results 6 to 9 (PR #2519).
Evidence (MEASURED, 2026-11-10)
- One ESP32-C6 node (firmware 0.8.17, channel 10), one room, server started with
--source esp32. - Empty: the operator left for 32 minutes; 31.0 minutes kept after trimming 30 s at each end. No keyboard or mouse input occurred in that span.
- Occupied reference: keyboard or mouse input within the last 3 s (
scripts/input_activity_logger.py).
| Condition | Observations | Classifier said present |
|---|---|---|
| Empty room | 6,104 | 6,093 (99.8%) |
| Input-confirmed occupied | 429 | 429 (100%) |
During the empty stretch the label was present_moving for 6,048 observations.
Likely cause
The level is raw_classify(smoothed_motion) with adjusted = raw_motion - 0.7 * baseline. In a steady room the baseline converges to the raw score, so adjusted settles at about 30% of the raw motion score. On this link the raw score is high enough at rest that 30% of it stays above the 0.12 present_moving threshold indefinitely. The thresholds assume a motion-score scale this hardware does not have, and the per-feature direction is inverted here: variance, motion-band and breathing-band power were higher with the room empty than with a person seated.
What did work on the same recordings
Variability over a 30 s window, not level: the standard deviation of RSSI separated occupied from empty with AUC 1.00 in three sessions, and a head fitted on those windowed features (ruview.node-features/2) gave 0 false positives in 3,069 held-out empty observations at 91% sensitivity.
Limits
One node, one room, one afternoon. "Empty" is the operator's declaration. Occupied means seated and typing. ESP32-S3 nodes were not tested and may behave differently.
Suggested direction
Either calibrate the thresholds per link from a known-empty period, or base the per-node decision on windowed variability. Both need the motionless-person case tested before they replace the current rule.
Activity
ruvnet commentedon Nov 11, 2026
Correction: these recordings mixed transmitters
The node was capturing CSI from every transmitter, not from one. Looking at per-observation RSSI in the same recordings:
With capture restricted to the access point (
provision.py --filter-mac <bssid>), the access point's frames arrive at -30 to -39 dBm. So the source that appeared only when the operator was present is the access point sending data, which happens when someone is using the network.Consequences:
Both are being re-recorded with the MAC filter on. First observation under the filter, operator present and typing: the classifier reports
absenton 80 of 116 observations, which is a different failure from the one reported above. I will post the filtered empty-room comparison here.