RuView is an open-source edge-AI platform that turns a $9 ESP32 into a through-wall sensing radar using WiFi Channel State Information (CSI). Source: https://github.com/ruvnet/RuView
Key points
- Why CSI matters: WiFi signals already fill rooms, reflecting off walls, furniture, and people. CSI—the internal channel-modulation data from WiFi chips—changes as bodies move, chests rise with breathing, and blood pulses with heartbeats, turning ordinary WiFi into a spatial sensor.
- Tiny models: The quantized model is just 8 KB at 4-bit precision, enabling microsecond-level inference on a Raspberry Pi. A pretrained model is published at https://huggingface.co/ruvnet/wifi-densepose-pretrained, with 100% presence-detection accuracy on the validation set.
- ESP32 mesh: multi-node networking, frequency hopping across 6 WiFi bands, even using neighbors' routers as free radar illuminators.
- Spiking neural networks (SNN): adapt to new environments within 30 seconds, more power-efficient than conventional DNNs.
- Cognitum Seed (optional): persistent memory layer with vector storage, kNN, and an Ed25519 witness chain that cryptographically authenticates every measurement.
- RuVector: underlying CSI processing engine (https://github.com/ruvnet/ruvector).
- Total BOM cost: roughly $140.
- Home Assistant: MQTT publishing, 21 entities (11 raw signals + 10 semantic states), 3 starter Blueprints.
- Apple Home / HomePod: HAP-1.1 bridge with Siri voice queries.
- Google Home / Alexa: via HA bridge or Matter endpoint.
- SmartThings: Matter Bridge.
- Through-wall range of ~5 meters depends on signal quality.
- Multipath-rich environments (metal, mirrors) can cause interference.
- Best pose-estimation accuracy requires camera-supervised fine-tuning.
- WiFi band availability varies by region, limiting some features.
- RuView: https://github.com/ruvnet/RuView
- Pretrained model: https://huggingface.co/ruvnet/wifi-densepose-pretrained
- RuVector engine: https://github.com/ruvnet/ruvector
Sensing capabilities
| Capability | Method | Accuracy / speed | |---|---|---| | Breathing rate | 0.1-0.5 Hz bandpass + zero-crossing count | 6-30 BPM, real time | | Heart rate | 0.8-2.0 Hz bandpass + zero-crossing count | 40-120 BPM, real time | | Presence detection | Phase-variance fallback + pretrained head | <1 ms | | 17-keypoint pose estimation | Candle engine, cog-pose-estimation | 8.4 ms cold start (Pi 5) | | Fall detection | Phase-acceleration threshold + 3-frame debounce | <200 ms | | Multi-person counting | Adaptive P95 normalization + dedup | Real time, self-calibrating | | Through-wall sensing | Fresnel-zone geometry + multipath modeling | ~5 meters |
Technology stack
105 edge modules
Pluggable modules span healthcare (sleep monitoring, apnea screening, elderly activity anomalies), security (presence, intrusion, fall detection), commercial (foot traffic, queue length, meeting-room occupancy), and industrial use (cleanroom tracking, equipment monitoring). Each runs on the ESP32 itself, no cloud required.
Smart home integration
Semantic states include someone-sleeping, possible-distress, room-active, elderly-inactivity-anomaly, and fall-risk-elevated—inferred scenarios rather than raw data.
Privacy by architecture
No camera, no wearable, no app, fully local processing, and Ed25519-signed measurements. Privacy is structural: visual spying is physically impossible because there is no visual sensor.
Limitations
Conclusion
RuView is not the first CSI-sensing project, but it packages the concept into a complete product stack: $9 hardware, edge AI, smart home integration, and privacy-first design. Example scenario: an elderly parent living alone can be monitored for breathing anomalies via Home Assistant—no camera, no wristband, just an ESP32 plugged into a wall outlet.
References