Using your neighbor's WiFi to see through walls

What it does RuView is an edge-computing system that reads Channel State Information from ESP32 radios—priced at roughly $9 per node—to detect human presence through walls, estimate breathing and heart rate, recognize activities like falls, and infer 17 body keypoints. It exposes all of this as smart-home entities via Home Assistant, Apple HomeKit, Google Home, and Amazon Alexa without requiring cameras, wearables, or cloud connectivity. Each node ships 21 entities, including semantic states such as possible-distress, bed-exit, and elderly-inactivity-anomaly.

The interesting bit The mesh doesn’t rely solely on your own access point; it hops six WiFi channels and treats neighboring routers as free radar illuminators. A 4-bit quantized embedding model fits in 8 KB and runs on a Raspberry Pi, while spiking neural networks locally adapt to a new room in under 30 seconds. Every measurement is cryptographically attested via an Ed25519 witness chain—an unusual level of paranoia for a home sensor.

Key highlights

• Claims 82.3% held-out temporal-triplet accuracy on presence detection using a self-supervised contrastive encoder; the authors explicitly retracted an earlier “100%” single-class figure.
• Achieves 82.69% torso-PCK@20 on the MM-Fi pose-estimation benchmark, beating MultiFormer and CSI2Pose, with an 8.4 ms cold-start on a Pi 5.
• 105 edge-module “Cogs” run directly on ESP32 hardware; the full stack with an optional Cognitum Seed is listed at a $140 total bill of materials.
• Through-wall sensing works up to roughly 5 meters, though the README notes this is signal-dependent.
• Smart-home integrations include a Matter bridge, HA-DISCO MQTT publisher, and native HAP-1.1 support for Siri voice queries.

Caveats

• The README bundles an enormous scope—pose estimation, vital signs, world-model prediction, 3D point-cloud fusion, and 105 edge modules—making it hard to tell which features are fully integrated versus experimental.
• The project publicly retracted an earlier “100% presence” accuracy claim, replacing it with the 82.3% figure; the sheer volume of benchmarks and architecture-decision records suggests a fast-moving, possibly unstable surface.
• Some headline features, like the OccWorld future-occupancy predictor, require an RTX 5080 and 3.4 GB VRAM, so “edge-only” depends heavily on which module you actually run.

Verdict Privacy-obsessed smart-home builders and RF tinkerers should look here—it’s a rare fully offline, camera-free presence system with honest benchmarking. If you just want a simple motion sensor, this is massive overkill, and the ESP32 mesh plus optional Seed node is not plug-and-play.