Your Wi-Fi Is Watching—Here’s What It Can See

Imagine Wi-Fi-Fi router not just connecting devices, but discerning your movements through walls. A recent live test demonstrated this: three roughly $10 ESP32-S3 boards, running open-source firmware and a trained AI model, classified a creator’s movements—sitting, standing, walking, squatting—with 79.5% accuracy.

This capability stems from Channel State Information (CSI). When your router transmits a Wi-Fi packet, it divides the data across dozens of distinct subcarriers. The receiving device measures how each subcarrier’s signal strength and phase shift as they travel, bouncing off walls, furniture, and even your body.

Movement subtly alters these signal paths. Your body, largely water, absorbs and reflects Wi-Fi signals, creating unique interference patterns that the ESP32-S3 boards capture. This raw, rippling signal data is then fed to the AI model.

Crucially, this system doesn't "see" you like a camera. Instead, it infers broad movement patterns from these signal changes. The AI, initially trained by correlating Wi-Fi data with ground-truth pose information from a camera, learns to recognize specific actions without ever forming a visual image.

RuView’s learning process mirrors a teacher-student dynamic. A standard webcam, running a pose detection model, acts as the "teacher," labeling human activities like sitting or walking. Concurrently, three ESP32-S3 boards record Channel State Information (CSI)—the subtle signal changes caused by movement—from their intersecting Wi-Fi paths.

These CSI recordings, paired with the webcam’s labels, train a second AI model. This "student" model learns to predict specific activities solely from the Wi-Fi data, effectively translating invisible signal disruptions into recognizable movements. Once training concludes, the webcam is removed, leaving the Wi-Fi model to operate independently.

Creator's DIY setup involved three roughly $10 ESP32-S3 boards positioned around a living room, generating multiple intersecting signal paths. Approximately nine minutes of labeled movement data were collected, teaching the model to interpret various actions.

Initial results demonstrated strong performance for static or predictable movements:

• Sitting: 93% accuracy
• Walking: 70% accuracy
• Room empty: 96% accuracy

However, the model struggled with smaller, more nuanced gestures, detecting waving with only 13% accuracy. This suggests that while large-scale body movements significantly alter Wi-Fi signals, subtle limb actions may not generate sufficient CSI perturbation for reliable detection.

A critical look at RuView’s reported accuracy reveals important nuances. While the project claims a benchmark accuracy of a little over 82% on the MM-Fi dataset—a robust collection from 40 people across four rooms in Singapore—the home test results warrant closer inspection.

Reviewers found the model’s testing methodology somewhat optimistic; it often evaluated performance using examples from the same recordings used during training. When tested on an unfamiliar room, RuView’s reported accuracy plummeted to a mere 11%, highlighting a significant generalization problem.

Furthermore, the widely circulated viral demo, featuring a real-time skeleton tracking a person, reportedly generated that skeleton from webcam video, with the Wi-Fi portion simulated. This key caveat underscores that one room and one participant do not establish reliable, general-purpose tracking. For those interested in the project's architecture or replicating experiments, the RuView GitHub Repository offers comprehensive details.

The disparity between benchmark claims and real-world performance emphasizes the challenge of robust Wi-Fi-based human sensing. While promising, the technology still grapples with adapting to diverse environments and individuals.

Wi-Fi sensing presents a fascinating duality. Legitimate applications like camera-free occupancy sensing, home alarms, and fall detection offer significant benefits, particularly in environments where cameras are undesirable. These applications, however, demand rigorous testing and robust safeguards to prevent misuse.

The privacy implications are substantial. Wi-Fi signals can penetrate some walls, allowing for presence and movement detection without recording identifiable video. While this isn't precise spying at arbitrary distances, it raises concerns about invisible surveillance. The technology could reveal sensitive patterns of life without explicit consent or awareness.

This sensing capability is not a covert feature of everyday routers. The IEEE 802.11bf standard specifically addresses Wi-Fi sensing, but its existence signifies an industry effort to standardize dedicated sensing functionalities, not to enable human tracking via current, off-the-shelf equipment. Practical limitations remain: significant changes to a room’s layout or the introduction of additional people typically necessitate retraining the model to maintain accuracy.

Can Wi-Fi really detect people through walls?

Wi-Fi signals can pass through some walls, and changes in those signals can reveal presence or movement. Accuracy depends on the room, hardware, and setup.

What is CSI in Wi-Fi sensing?

Channel State Information (CSI) records how Wi-Fi subcarriers change as signals travel between devices, including shifts caused by objects or people.

How accurate was the RuView home test?

The creator reported 79.5% accuracy across eight movements in a separate live test, but the experiment involved one person, one room, and a model trained there.

Does Wi-Fi sensing require a camera?

A camera can provide labels to train a model, but once trained, the sensing model can make predictions from Wi-Fi data alone.