RuView: Turning Wi-Fi into a $7 Through-Wall Radar with ESP32

> One router and a thumb-sized development board—no camera, no wearable—can detect which room you are in, whether you are walking or sitting, your breathing rate, heart rate, and even whether you experience apnea during sleep.

RuView, a project that recently trended on GitHub, delivers an counterintuitive physical fact to developers: the Wi-Fi signals emitted by your router already reflect off your body. Each breath causes chest movement; each heartbeat causes micro-vibrations. Both modulate the 2.4 GHz electromagnetic waves. The question was never whether sensing was possible, but how to extract the embedded information.

RuView's answer is deceptively simple: a $7 ESP32-S3 board runs open-source firmware that captures Wi-Fi CSI (Channel State Information), feeds the data to a locally executed small model, and outputs 17 body keypoint pose estimates, respiration rate, heart rate, fall detection, and sleep staging.

The Physics: Wi-Fi Signals Are Already Radar

The first step in understanding RuView is abandoning the intuition that only cameras can "see."

Radar works by emitting electromagnetic waves, receiving reflections, and extracting target position and velocity from time delay and frequency shift. A Wi-Fi router emits thousands of electromagnetic pulses per second; your body—mostly water—reflects part of the energy. Conventional Wi-Fi chips care only about decoding 0s and 1s and discard reflections as noise. CSI is exactly what gets discarded: it records the amplitude and phase of each subcarrier at every moment—a fine-grained snapshot of how the signal has been distorted in space.

When a person walks through a room, the CSI waveform traces a clear trajectory. When a person sits still and breathes, the waveform exhibits periodic micro-fluctuations of 12–20 cycles per minute. When a person sleeps, the slower breathing cycle is joined by a higher-frequency micro-fluctuation at 60–100 cycles per minute that corresponds to heart rate. After an FFT, these appear as two real peaks in the spectrum.

RuView connects the chain from physics to semantics: ESP32 captures CSI → preprocessing denoises the signal → a trained model infers semantic events ("someone is sleeping," "possible fall," "bathroom occupied").

Engineering: 21 Entities per Smart-Home Node

RuView is not a paper—it is a working system. Each ESP32 node outputs 21 entities: 11 raw signals (CSI amplitude, phase, SNR, etc.) plus 10 inferred semantic states.

The 10 semantic states are worth examining individually because they reveal what "spatial intelligence" actually means:

• someone-sleeping: someone is asleep
• possible-distress: possible abnormal state
• room-active: the room is active
• elderly-inactivity-anomaly: abnormal stillness for an elderly person
• meeting-in-progress: a meeting is in progress
• bathroom-occupied: bathroom is in use
• fall-risk-elevated: fall risk is elevated
• bed-exit: has left the bed
• no-movement: no movement
• multi-room-transition: moved between rooms

These are not low-level events such as "motion detected"; they are semantically enriched inferences. "Elderly inactivity anomaly" is not "no motion," but rather "activity is normally expected at this time, yet no micro-motion has been detected for two continuous hours." "Fall risk elevated" is not "a fall happened," but rather "the gait CSI pattern resembles the 30-second pre-fall pattern observed in history."

This is a three-layer architecture of sensing → inference → semantics, isomorphic to computer vision's pixel → feature → object-recognition pipeline. The difference is that RuView's "pixels" are CSI time series rather than RGB frames.

Privacy: Surveillance Without a Lens

The most important discussion around RuView is not technical; it is its redefinition of "surveillance."

The problem with cameras is not that they capture images; it is that those images can be replayed, stolen, or used to train facial-recognition models. RuView has no lens. Its output is "there is someone in the room, breathing 16 times per minute, lying down"—data useless to facial recognition yet invaluable for the use case of "is an elderly person living alone safe?"

This is a critical distinction: the privacy cost of surveillance depends not on what can be sensed but on whether the sensed signal can be reversed into personal biometric features. CSI signals are highly semantic abstractions—they tell you "a person is breathing" but cannot tell you who that person is. That is fundamentally different from a camera.

That said, CSI is not absolutely anonymous. Research has shown that gait CSI patterns can support person identification, because each person's walking rhythm, stride, and weight shift differ. RuView explicitly acknowledges this in its documentation and attaches privacy policy, uncertainty, source, and witness records to every sensing event. This reflects a "governance-first" engineering attitude rather than "build first, ask questions later."

MetaHarness: An Honesty Check for AI Claims

A subtle but important component of the RuView repository is MetaHarness, an AI operator that guides humans and AI agents through deployment, calibration, training, and validation, and performs honesty checks on every sensing claim.

The reason this matters: Wi-Fi sensing has an old problem—95% lab accuracy can drop to 60% in your living room. Router placement, wall materials, other reflectors, and models trained on lab data all contribute. Without field calibration and verification, claimed accuracy is an empty check.

MetaHarness requires each claim to carry a source citation, certainty verification, and field-data validation. This aligns with a recurring theme in the "the model already knows" series: a claim is not capability; claim plus evidence is capability. The RuView authors built this directly into the system design rather than leaving it to the user.

Cross-Domain Analogy: Electric Fish and Bats

RuView's underlying physics has direct biological parallels. Weakly electric fish navigate and hunt in the murky Amazon by emitting electric pulses and sensing the reflected electric field. Bats use ultrasonic echolocation. Neither relies on lens-based vision; both use actively emitted signals plus fine analysis of reflections to construct an environmental model.

RuView is the engineered version of this strategy: actively emit Wi-Fi signals, analyze CSI reflections, and construct a spatial model. Electric fish and bats spent millions of years optimizing this capability; RuView caught up in a few years using open-source hardware and deep learning.

RuView's "Unified RF World Model" attempts to unify Wi-Fi CSI, millimeter-wave radar, UWB, and cellular sensing into a single scene model. This mirrors multimodal sensory fusion in biology: owls combine vision and hearing, sharks combine smell and electroreception. Single modalities have blind spots; multimodal fusion is robust.

Why This Is Happening Now

Wi-Fi sensing is not new. A 2013 MIT WiTrack paper already demonstrated through-wall tracking using Wi-Fi. For more than a decade, the field was blocked by three bottlenecks:

1. Hardware barrier — extracting CSI required specialized SDRs or modified drivers. ESP32-S3 is the first low-cost consumer chip that outputs CSI directly. 2. Model barrier — traditional methods relied on hand-crafted features with poor generalization. Transformer architectures and LoRA make training on small datasets feasible. 3. Ecosystem barrier — there was no standardized semantic output layer. RuView's 21-entity model plus Home Assistant/Matter/HomeKit integration puts Wi-Fi sensing into standard smart-home protocols for the first time.

When all three bottlenecks loosen simultaneously, the breakthrough happens. $7 hardware + open-source firmware + pretrained models + smart-home integration takes "through-wall sensing" from the laboratory to a purchasable DIY kit.

A Deeper Observation

RuView illustrates a cross-domain principle: the best sensor is one you have already deployed but are not yet using.

Wi-Fi routers have always emitted signals; those signals have always been reflected. For decades we used only the layer that decodes 0s and 1s and discarded the finer-grained CSI as noise. RuView's contribution is not "inventing a new sensing modality"; it is "recovering information that was thrown away."

This is the same pattern as SOPHIA's residual stream direction, Möbius RoPE's topological intervention, and the mantis shrimp's phonon shield: instead of doing the same thing harder, solve the problem at a different layer. Cameras add more lenses; RuView uses signals that already exist.

The corollary: many "free sensors" around you are being wasted. Bluetooth advertisement packets, ambient light sensors, microphone background noise, accelerometer micro-vibrations—all contain hidden information that no one is decoding. RuView provides a reusable engineering template for the "decode already-existing signals" approach.

$7, one ESP32, one open-source repository. Seeing you breathe through a wall, with no lens. This is not science fiction; it is GitHub trending in 2026.

---

• Project repository: https://github.com/ruvnet/RuView
• Model repository: https://huggingface.co/ruvnet/wifi-densepose-pretrained
• Author blog: https://ruview.blog/wifi-csi-sensing-open-source