87,388 stars, 11,612 forks, MIT license, primary language Rust, still actively iterated -- that's the state of RuView (repo ruvnet/RuView) at the end of July 2026. What it does boils down to one sentence: turn ordinary WiFi signals into a spatial sensor for real-time presence detection, vital signs monitoring, and activity recognition, with zero cameras, zero wearables, and the ability to see through walls. A few years ago this was still a concept in academic papers. Now someone has packaged it as an open-source hardware kit that runs on a $9 ESP32 node. 87.4k stars says it hit a real need: cameras invade privacy, wearables are annoying, and WiFi sensing sits right in the gap between them.
1. What WiFi Sensing Is: Seeing a Room Without Cameras
Let's unpack the term. WiFi signals bouncing around your space get reflected and scattered by walls, furniture, and human bodies. Anyone in the room -- walking, breathing, even sitting still with their chest rising and falling -- creates faint but measurable disturbances in the radio waves. It's similar to radar, except it uses the WiFi band you already have, no dedicated RF hardware required.
RuView (also branded as π RuView) turns those disturbances into usable data. It advertises three core advantages: no camera (zero image pixels captured, privacy risk cut off at the source), no wearable (the person being sensed doesn't need to wear anything, lowering the psychological barrier), and through-wall (radio waves penetrate non-metallic walls, so a node in the living room can sense the bedroom). The sensing capabilities it covers include presence/occupancy, vital signs (breathing rate, heart rate, contactless), activity recognition (walking, sitting, gestures, falls), environmental mapping, and sleep quality. A single $9 ESP32 node exposes 21 entities to a smart home platform, effectively replacing a roomful of PIR sensors and cameras without touching the privacy question.
2. How It Works: CSI Signals Plus a Spiking Neural Network
The underlying signal is called Channel State Information (CSI) -- the WiFi physical layer's description of how a signal gets distorted through multipath propagation. Ordinary routers don't expose CSI to you. RuView reads it out using ESP32 sensors. Every human action in the space -- walking, sitting, breathing -- leaves a unique signature in the CSI time series, like an electrocardiogram except it's charting radio wave phase and amplitude changes.
What translates those signatures into "who's there, what are they doing, are they OK" is a pretrained model combo: RuView's own RuVector plus a Spiking Neural Network, which claims under-30-second adaptation to a new environment. A spiking neural network is not the same animal as a CNN or Transformer. Its neurons communicate via discrete spikes, closer to how biological neurons operate, with the advantage of high energy efficiency and a natural fit for running real-time signal streams on edge hardware. That matters for WiFi sensing -- the signal flows in continuously, and the event-driven nature of spiking networks lets you do real-time inference without burning a GPU.
The model weights are open on HuggingFace (ruvnet/wifi-densepose-pretrained), compressed to 8KB after 4-bit quantization, and can do microsecond-level inference on a Raspberry Pi. 8KB is smaller than a typical website favicon. The model is small enough to fit entirely in the ESP32's flash with no external accelerator needed. The entire inference chain stays local -- no cloud, no internet, works offline. A Cognitum Seed module handles persistent memory and encryption. The whole system is a pure edge architecture: data is born and consumed on the same device, never leaving your network. For privacy-sensitive scenarios, this isn't just a bonus feature -- it's an architectural guarantee that your breathing rate, sleep patterns, and activity traces never pass through any third-party server.
3. Real Numbers: 82.3% Accuracy, 8KB Model, $9 Node
Open-source projects are most easily led astray by "100% accuracy" claims. RuView handles this more honestly: the v2 encoder reports 82.3% accuracy on a time-triplet task. The old "100% presence detection" marketing has been retracted, replaced by this more conservative number. For an 87k-star project, that's actually a good signal -- being willing to pull numbers down into the real range means the maintainers care about credibility over marketing spin. Anyone who's used early-stage AI open source knows that the "99% accuracy" in many READMEs was cherry-picked on a clean test set and collapses the moment it hits a real environment. RuView voluntarily retracting inflated claims is a rare show of maturity.
Other key numbers: the model is 8KB after 4-bit quantization, running microsecond-level inference on a Raspberry Pi, which means it can run on $10-class hardware instead of a GPU. ESP32 mesh nodes go as low as $9 each; covering a three-bedroom apartment takes roughly 4-6 nodes, keeping total hardware cost under $60. Cognitum Seed handles persistent memory and encryption, the whole system runs pure-edge with no cloud and no network. These aren't lab-demo spec sheet numbers -- they're the cost structure of buying boards and soldering them yourself. For comparison: a commercial fall-detection camera runs $200+ and carries privacy baggage. A $9 ESP32 node captures no images and costs an order of magnitude less. The ESP32 mesh design is also worth noting -- nodes self-organize into a mesh rather than hanging off a central router in a star topology, so a single node going down doesn't break the others' sensing. That fault tolerance matters more than you'd think in a real deployment.
4. What It Senses: From Presence to Sleep Quality
Breaking the sensing capabilities apart, RuView covers a full spatial sensing stack:
- Presence/occupancy: is there a person in the room, more sensitive than PIR, not fooled by someone sitting still (PIR only reacts to moving heat sources and fails when a person stops moving -- a classic PIR pain point)
- Vital signs: breathing rate, heart rate, contactless, nothing strapped to the chest -- a dimension of capability cameras can't touch
- Activity recognition: walking, sitting, gestures, falls; fall detection is a hard requirement in eldercare
- Environmental mapping: drawing out the spatial structure, like a low-resolution radar map
- Sleep quality: how many times someone turned over at night, whether breathing is regular, without a wristband or a mattress sensor
Stacked together, these can replace a pile of sensors that each do one job. One node exposing 21 entities to a smart home platform is far lighter than a Z-Wave protocol stack. For someone used to writing automations in Home Assistant, a single ESP32 node can trigger "person enters, turn on lights," "no activity for 30 minutes, send notification," "fall detected, sound alarm" -- all without a camera in the bedroom.
5. Integration: Home Assistant, Apple Home, Matter
The breadth of smart home integration is another highlight. The official docs list four paths: Home Assistant over MQTT (most flexible, for tinkerers writing their own automations), Apple Home/HomePod via HAP-1.1 bridge (Apple's ecosystem is usually closed, so getting in means someone wrote a protocol adapter), Google Home natively supported, and Amazon Alexa over Matter (the universal standard, cross-vendor compatible). Each node exposes 21 entities, so once you've laid out a mesh network, every major smart home ecosystem can consume the data.
For someone already deep in Home Assistant, the MQTT path is the friendliest -- you define your own topics, write your own automations. Apple Home via HAP-1.1 bridge is the standout -- Apple's ecosystem typically doesn't welcome third-party hardware, so being able to bridge in means Siri can answer "is anyone in the bedroom?" Matter support means Alexa and Google Home share a single universal path, no per-vendor adapter needed. At 21 entities per node, a three-bedroom apartment with 4-6 nodes adds over a hundred usable sensor entities to your smart home platform, which opens up a lot of automation headroom.
6. Who It's For
Making the scenarios concrete:
Smart home tinkerers. If you're already running Home Assistant and have a drawer full of ESP32 boards, this is a new toy. $9 a node, 21 entities, cheaper than buying a pile of dedicated sensors, and it does through-wall sensing cameras can't. The Rust code is readable, and you can swap in your own models if you want.
Eldercare monitoring. This is RuView's most socially valuable application: put a node in an elderly parent's bedroom, and falls, breathing anomalies, or prolonged inactivity at night can all trigger alerts. No camera means no privacy awkwardness, no wearable means the elderly person doesn't have to learn or remember anything, and children can receive notifications remotely. Contactless vital signs are the key differentiator here -- the sensing happens passively, the person doesn't need to operate a device.
Privacy-sensitive scenarios. Cameras in bedrooms, bathrooms, and children's rooms are sensitive. WiFi sensing sidesteps image data entirely -- the raw signal is not a face or a video, and the psychological acceptance bar is much lower. For Airbnb hosts and hotel operators, this also sidesteps the legal risk of installing cameras.
7. Honest Boundaries: Hardware, CSI Access, Accuracy
Can't only praise. RuView has real barriers:
First, you need ESP32 hardware. This isn't a software project you can npm install and run. You have to buy boards, solder antennas, flash firmware. For someone with a pure software background, that's a wall. The ESP32 ecosystem is mature and well-documented, but actually deploying it still requires hands-on work.
Second, CSI access depends on firmware support. Not all ESP32 modules expose CSI data by default; it requires specific firmware compile configuration, and you need the right version to read clean CSI streams. This is the most common place to get stuck during initial deployment. The recommendation: buy the modules the official docs recommend, don't guess at part numbers.
Third, accuracy has honest limits. 82.3% is the v2 encoder's reported number on the time-triplet task, not "82.3% in all scenarios." Real-world multipath environments, wall thickness, and co-channel interference (your neighbor's WiFi on the same band) will drag the number down. The old "100%" claim has been retracted, which tells you the early data was optimistic. Users need to re-validate in their own environment.
Fourth, safety-critical features like fall detection cannot replace professional medical equipment before medical-grade certification is complete. An open-source solution is a good base for hacker experimentation and prototyping, not a medical device. If you want to use it on a real elderly person, test it yourself for a few months first.
87.4k stars doesn't mean a mature product. It means the direction has been validated and a community has formed around building it together. RuView's value is in pulling "WiFi sensing" out of academic papers and down to a $9-per-node reproducible engineering level, while the maintainers are willing to retract inflated data and report real accuracy numbers. For hackers willing to solder their own boards, it's one of the most worthwhile open-source sensing projects to follow in 2026. For those who just want to npm install, read the README first before ordering any ESP32s.
References
- RuView GitHub repo (87,388 stars, 11,612 forks, MIT, Rust): https://github.com/ruvnet/RuView
- Pretrained model wifi-densepose-pretrained (HuggingFace, 4-bit quantized, 8KB): https://huggingface.co/ruvnet/wifi-densepose-pretrained