RuView — WiFi Sensing Models

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.

🆕 v2 update — honest re-benchmark + properly-converged encoder (2026-05-31)

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).

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.

EWC (catastrophic-forgetting protection): measured forgetting rate on the 4 protected tasks is 32.97% -- real and non-trivial, disclosed here rather than only in the raw training-metrics.json. Fisher information is computed and applied (protection strength 2000), but this is not "solved forgetting" -- a user fine-tuning per-room LoRA adapters should expect some drift on earlier tasks, not none.

Fixed v2.0.1 -- fixed the published training-metrics.json itself (2026-09-09)

Separately from the v1-to-v2 accuracy retraction above, an unrelated bug was found and fixed in the export code (scripts/train-ruvllm.js) that produces config.json and training-metrics.json for this file set (model.safetensors, model-q2/q4/q8.bin, presence-head.json, node-1/2.json) -- separate from the csi-embed-v2* files, which come from a different script.

The exported lossHistory and contrastive.durationMs were being read from a disconnected baseline pass (ContrastiveTrainer.train(), which computes loss but never updates the encoder) instead of from the loop that actually performs gradient updates on the encoder weights. Concretely: lossHistory published a flat, non-decreasing curve around 0.1352 for all 20 epochs, and durationMs reported ~12.4s -- both measuring the wrong computation.

This did not affect finalLoss/improvement/the 82.3% held-out temporal-triplet accuracy above -- those were already computed from the real training loop and are unchanged by this fix. It only affected the exported diagnostic curve and duration, which anyone auditing convergence from the published JSON would have been misled by. Re-run against the same real capture (data/recordings/overnight-1775217646.csi.jsonl) with the fix applied: lossHistory now shows a genuine decreasing curve (0.1037 -> 0.0654 across 20 epochs) and durationMs reports the real ~40s gradient-loop time. config.json's training.steps field (12212300 -- epochs x triplets, arithmetically consistent but read like optimizer-step count) is now reported as separate epochs/tripletUpdatesPerEpoch fields instead.

Repro: node scripts/train-ruvllm.js --data data/recordings/overnight-1775217646.csi.jsonl --output <dir> --epochs 20 in github.com/ruvnet/RuView (the aether-arena/staging/train_csi_embed.py path cited above is for the separate csi-embed-v2* files and was not touched by this fix). Full fix: ruvnet/RuView#1879.

v2 files & proof

• Encoder weights SHA-256: 3b37bca66e6050c50ccbc0f6e0501824f258bfdd8675dc0f4541b1e2e96feecd
• Repro: python aether-arena/staging/train_csi_embed.py in github.com/ruvnet/RuView
• 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):

Models in This Repo

Quick Start

# Download models
pip install huggingface_hub
huggingface-cli download ruv/ruview --local-dir models/

# Use with RuView sensing pipeline
git clone https://github.com/ruvnet/RuView.git
cd RuView

# Flash an ESP32-S3 ($9 on Amazon/AliExpress)
python -m esptool --chip esp32s3 --port COM9 --baud 460800 \
  write_flash 0x0 bootloader.bin 0x8000 partition-table.bin \
  0xf000 ota_data_initial.bin 0x20000 esp32-csi-node.bin

# Provision WiFi
python firmware/esp32-csi-node/provision.py --port COM9 \
  --ssid "YourWiFi" --password "secret" --target-ip YOUR_IP

# See what WiFi reveals about your room
node scripts/deep-scan.js --bind YOUR_IP --duration 10

Architecture

WiFi signals → ESP32-S3 ($9) → 8-dim features @ 1 Hz → Encoder → 128-dim embedding
                                                                    ↓
                                         ┌──────────────────────────┼──────────────────┐
                                         ↓                          ↓                  ↓
                                    Presence head            Activity head         Vitals head
                                    (v1 "100%" retracted)          (still/walk/talk)     (BR, HR)

The encoder converts 8 WiFi Channel State Information (CSI) features into a 128-dimensional embedding:

Training Details

No camera was used. Trained using self-supervised contrastive learning:

• Data: 60,630 samples from 2 ESP32-S3 nodes over 8 hours
• Method: Triplet loss + InfoNCE (nearby frames = similar, distant = different)
• Augmentation: 10x via temporal interpolation, noise, cross-node blending
• Supervision: PIR sensor, BME280, RSSI triangulation, subcarrier asymmetry
• Quantization: TurboQuant 2/4/8-bit with <0.5% quality loss
• Adaptation: LoRA rank-4 per room, EWC to prevent forgetting

17 Sensing Applications

Built on these embeddings (RuView):

Core: Presence, person counting, RF scanning, SNN learning, CNN fingerprinting

Health: Sleep monitoring, apnea screening, stress detection, gait analysis

Environment: Room fingerprinting, material detection, device fingerprinting

Multi-frequency: RF tomography, passive radar, material classification, through-wall motion

Hardware

Limitations

• Room-specific (use LoRA adapters for new rooms)
• Camera-free pose: 2.5% PCK@20 (camera labels improve significantly)
• Health features are for screening only, not medical diagnosis
• Breathing/HR less accurate during active movement

Citation

@software{ruview2026,
  title={RuView: WiFi Sensing with Self-Supervised Contrastive Learning},
  author={rUv},
  year={2026},
  url={https://github.com/ruvnet/RuView},
  note={Models: https://huggingface.co/ruv/ruview}
}

Links

• GitHub: https://github.com/ruvnet/RuView
• Cognitum Seed: https://cognitum.one
• RuVector: https://github.com/ruvnet/ruvector
• License: MIT