Turn WiFi signals into spatial intelligence. Detect people, measure breathing and heart rate, track movement, and monitor rooms — through walls, in the dark, with no cameras. Just radio physics.
What This Does
WiFi signals bounce off people. When someone breathes, their chest moves the air, which subtly changes the WiFi signal. When they walk, the changes are bigger. This model learned to read those changes from a $9 ESP32 chip.
The v1 contrastive encoder shipped with a flat training loss (every epoch logged the
same 0.13517 — the optimizer was not actually learning), and its headline "100% presence
accuracy" was measured on a single-class recording (an overnight capture of one sleeping
person: 6,062 of 6,063 frames are labelled "present", 1 is "absent"). A constant
"yes" predictor scores 99.98% on that split — so the number is real but says nothing about
generalization. We are correcting that publicly rather than leaving it to stand.
v2 retrains the same 8 -> 64 -> 128 encoder with a working InfoNCE objective and reports
an honest, label-free, time-disjoint metric: held-out temporal-triplet accuracy =
P( d(anchor, temporal-positive) < d(anchor, temporal-negative) ), evaluated on the last 20%
of the recording by time (no leakage into training).
Encoder
Held-out temporal-triplet accuracy
Notes
Raw 8-dim features (no encoder)
66.4%
baseline
Random-init encoder
69.6%
untrained
v2 trained encoder
82.3%
+15.9 pts over raw, properly converged
Plain language: the embedding now reliably places two CSI snapshots taken moments apart
closer together than two taken far apart — i.e. it has learned the temporal structure of the
radio environment, which is exactly what a useful self-supervised sensing embedding should do.
v1, with its flat loss, was barely better than random on this same test.
Technical: 2-layer FC (BatchNorm + GELU) -> L2-normalized 128-dim embedding, 9,280
params, trained with InfoNCE (temperature 0.1, in-batch + temporal-far negatives), AdamW, 60 epochs.
Temporal positives within 2 s; negatives >30 s apart. Time-disjoint 80/20 split.
Trained on the same local capture (data/recordings/overnight-1775217646.csi.jsonl, 6,063 feature frames).
What v2 does not claim. This is one room, one capture, two nodes. The triplet metric
measures embedding quality, not downstream presence/vitals accuracy (which needs multi-class,
multi-room labelled data we don't yet have for this 2.4 GHz feature). For pose SOTA on a
public benchmark, see the separate 5 GHz model
ruvnet/wifi-densepose-mmfi-pose
(82.69% torso-PCK@20 on MM-Fi).
Benchmarks
Validated on real hardware (Apple M4 Pro + 2x ESP32-S3):
