When the Building Knows You Are Here: Wi-Fi as an Architectural Sense

You sit at your desk, absorbed in your work, minute after minute, until the lights suddenly go out above your head. Nobody has made a mistake. The building has simply decided that you do not exist. You wave your arms in the air like a castaway signaling a distant ship, and the lights return, reluctantly. Anyone who has sat in a closed meeting room, a small office, or a public restroom has lived this moment. It exposes a deep flaw in how our buildings “sense” us: they see motion, not people. Yet in that same ceiling, a few meters from the darkened lamp, a wireless device has been broadcasting its waves around you since morning. Recent research suggests that these waves can tell you are there, even when you sit still, and even when a wall stands between you and them. So what happens when an entire building becomes one large sense organ?

The Wave That Carries the Room’s Fingerprint

We think of a Wi-Fi network as a pipe that delivers data to our phones. But the waves that carry that data do not travel in a clean straight line. They strike walls, furniture, and human bodies. They reflect, scatter, and lose part of their energy to absorption, then reach the receiver along multiple overlapping paths. Some of these paths are fixed, created by walls, ceilings, and cabinets. Others vary, created by a body that walks, breathes, or sits. This difference is the foundation of wireless sensing.

In the early days, researchers relied on the received signal strength indicator, a crude measurement, like judging an entire orchestra by its overall volume alone. The turning point came with what is known as channel state information, a fine-grained measurement made available by modern Wi-Fi chips. This technology records the amplitude and phase of the wave for each sub-frequency separately, and for each pair of antennas. As the review by Wang and colleagues explains, the well-known Intel 5300 card provides thirty sub-frequencies per antenna pair, which amounts to as many as one hundred and eighty complex values with every data packet received. In this way the room becomes a digital fingerprint that changes with whoever is inside it.

Software then takes over. Raw readings are full of noise from hardware imperfections, environmental interference, and lost packets. They are therefore cleaned with filters that remove outlier values and retain the slow frequencies associated with human movement. Time-domain and frequency-domain features are then extracted and handed to machine-learning algorithms, which classify the state: an empty room, a person sitting still, or a person walking.

More important for architecture, these waves pass through common building materials, from gypsum board and wood to concrete, provided the signal is strong enough. This means that the walls we draw to divide space are not a barrier to this sense. They are part of the medium in which it operates. So how did we get here?

From Complex Experiments to a Five-Dollar Chip

The idea traces back to the middle of the first decade of this century, when researchers experimented with locating people who carried no device at all, relying on changes in signal strength between the nodes of a wireless network. But the decisive moment came in 2013 with the Wi-Vi system, presented by Adib and Katabi at the Massachusetts Institute of Technology. The two researchers used a technique that cancels static reflections across multiple antennas, removing the “noise” of walls and furniture and leaving only the trace of the moving body. With several programmable radios, they tracked a person walking behind a concrete wall. The review by Wang and colleagues regards this experiment as the cornerstone on which the field was built.

The idea then moved from a complex experiment to widespread practice. Once tools for extracting channel information from commercial Wi-Fi cards became available, expensive laboratory equipment was no longer necessary, and systems multiplied that detect falls and classify activities. In 2018, Di Domenico and colleagues showed that analyzing the Doppler spectrum alone, with only two features and a simple classifier, was enough to distinguish an empty room from one containing a stationary or moving person behind walls, with accuracy between 91 and 99 percent.

Another leap came with the “channel state information ratio” technique introduced by Zeng and colleagues. By dividing the reading of one antenna by the reading of another on the same receiver, the noise that affects both antennas cancels out and a clean signal remains. In this way the range for detecting breathing extended beyond eight meters through walls, and the “blind spots” that had disabled sensing at certain locations disappeared.

The frontier today looks more like an economic surprise. In 2025, Natarajan and colleagues showed that a small ESP32 microcontroller, priced at around five dollars and using a single link, can distinguish five states through walls: empty, walking, standing, sitting, and lying down, with accuracy between 97 and 99 percent. If the sensor costs this little, where in a building might it be planted?

The Room That Knows You Are Sitting

Let us return to the office lamp. The problem is that most conventional occupancy systems measure motion, while a person who reads, writes, or thinks barely moves. Wireless sensing, however, has been shown in several studies to distinguish a person sitting still from an empty room. The 2024 study by Zhang and colleagues achieved a presence-detection accuracy of 99 percent, and it held across different environments, bodies, and speeds of movement. This result changes the meaning of the word “occupancy” itself, because the building knows who is inside and not merely who is moving inside.

This has a direct effect on the economics of operation. Heating, ventilation, and air-conditioning systems often run on fixed assumptions about the number of occupants and their schedules. The reviews note that counting people through wireless sensing supports better air-conditioning and energy management based on actual occupancy. One to three people moving behind a thirteen-inch brick wall have been successfully counted. That means cooling and lighting an empty hall could become an avoidable choice rather than an accepted price of comfort.

The capabilities go beyond counting to understanding behavior. Activities such as walking, standing, sitting, and lying down have been classified, and in some studies even smoking, with an accuracy of 92.8 percent. An architect imagining this data could see in it what post-occupancy evaluation aims to capture: how a space is actually used, not how it was drawn on the plan. Which corner of the lobby has become a shortcut? Which hall does nobody sit in? And all of this without a single camera.

In health care settings, the significance multiplies. Breathing has been detected through a ten-centimeter wall at ranges of up to eight meters, with a mean error of no more than 0.47 breaths per minute, and such systems are also used to detect falls among older adults. A room that hears its occupant breathe without touching them sounds like a promise of care. But that promise collides with questions about the built environment itself.

When the Wall Becomes a Partner in Measurement

The laboratory is a gentle environment, and real buildings are not. When performance is measured through a single wall, accuracy falls from its best results in direct line of sight, which approach 99 percent, to between 85 and 95 percent, according to the survey review. With two walls, figures in some experiments remain in the range of 91 to 95 percent. Transmission weakens further with concrete, because it attenuates the signal and distorts its shape. The choice of wall material, which we usually treat as a structural, acoustic, and thermal decision, therefore also shapes the building’s “transparency” to its wireless sense.

A deeper problem is what the literature calls “domain shift.” A model trained in one room performs poorly in another, because every space has a unique fingerprint. A model loses about two percent of its accuracy even in the same room if the furniture is rearranged. Performance also degrades over time as the contents or the seasons of a place change without retraining. And chips behave differently from one manufacturer to another, making it hard to transfer a single model between devices.

Range has limits as well. Fine-grained breathing detection works in practice between three and eight meters, and most systems need suitable lines of sight between transmitter and receiver. Separating the signals of several people moving at once also remains difficult, although promising attempts have emerged using antenna diversity and multiple channel ratios.

What this means for the practitioner is that a building equipped with this technology is measured the way an acoustic space is measured: by materials, dimensions, and distribution. The position of the access point, the arrangement of furniture, and the thickness of partitions all become parameters in an equation of sensitivity, parameters that appear today on no drawing. But a sense organ that is affected by the details of the room can also be deceived.

A Building That Can Be Deceived, and One That Never Asks Permission

As wireless sensing moves closer to sensitive functions such as health and security, the question of its susceptibility to manipulation becomes pressing. In 2023, Song and colleagues showed that sensing systems built on deep learning are fragile in the face of false data injection. By corrupting the channel readings in time, by shuffling the order of packets, in space, by swapping the positions of antennas, or in value, by altering amplitudes, they drove activity-recognition accuracy from about 92 percent down to 2.5 percent on some datasets. These attacks can be carried out at any layer of the system, with no need for physical proximity.

Ambalkar and colleagues conducted a more unsettling experiment on a system for detecting breathing interruptions during sleep. They introduced imperceptible disturbances into the readings, and accuracy fell from 85 to 25 percent, even though the deceptive patterns resembled normal breathing. When examples of the attack were included in the model’s training, accuracy under the strongest attack rose to only 51 percent. The defense improved, but the problem was not settled. Imagine a hospital ward whose building believes that everyone is breathing regularly.

The institutional side, meanwhile, is moving quickly. The IEEE 802.11bf standard, whose working group began in September 2020, is the first unified framework for Wi-Fi sensing, and it defines roles and procedures that allow two-party and multi-party sensing among devices from different manufacturers. As Ropitault and colleagues explain, this standard marks the technology’s transition from the laboratory to commercial products. In other words, within a few years this sense could become a default feature of every router.

Here the greatest contradiction emerges. Wireless sensing is presented as a “privacy-respecting” alternative to the camera, yet it sees through walls, and it may detect people who never agreed to be detected. Developing methods that extract activity without revealing people’s identities remains an open research question, not a ready solution.

The next time the lights go out above your head while you sit, your small fidget may be the last remnant of the room’s privacy. If the day comes when the building knows you are there, still or moving, breathing calmly or heavily, alone or with others, will we call that care or surveillance? We have grown used to deciding for ourselves what a building sees of us: we define the wall as what blocks the view, and the door as what prevents entry. But waves that cross concrete and carry the fingerprint of our bodies turn that definition upside down. Perhaps the architect must be the first to answer a question no one trained them for: how do we design a space that knows we are here, while respecting our right not to be known more than we choose?

✦ ArchUp Editorial Insight

A building that learns to detect a seated body is not the product of a sensor. It is the product of an operating-cost problem that procurement never priced. Ventilation and lighting are specified against assumed occupancy, and the gap between assumption and actual presence is paid monthly by whoever owns the meter. This is the CAPEX-OPEX misalignment traced in The Hidden Cost of Breathing, where health and energy were externalized from the party who specified the system. Wireless sensing offers to close that gap with hardware already in the ceiling, and the cited research makes the offer credible: presence detected at 99 percent, and a five-dollar microcontroller classifying posture through walls. The same evidence records where it fails. Accuracy falls with wall material, degrades when furniture moves, and drops from roughly 92 percent to 2.5 percent under injected data, while the 802.11bf standard moves the capability toward default status in every router. The pattern is familiar. The standards body, the chip manufacturer, and the installer decide at a layer where consent is not a parameter, and they exit at handover. The occupant absorbs the consequence, whether a misread breathing pattern in a ward or a room that registers presence without agreement, and holds no seat where those terms were written. Partition thickness, access-point position, and finish material become sensitivity variables that no drawing records and no brief assigns.

References

• Di Domenico, S., De Sanctis, M., Cianca, E., Ruggieri, M. “WiFi-Based Through-the-Wall Presence Detection of Stationary and Moving Humans Analyzing the Doppler Spectrum.” IEEE Aerospace and Electronic Systems Magazine, 2018.
• Zeng, Y., Wu, D., Xiong, J., Zhang, D. “Boosting WiFi Sensing Performance via CSI Ratio.” IEEE Pervasive Computing, 2021.
• Zhang, Y., Wang, X., Wen, J., Zhu, X. “WiFi-Based Non-Contact Human Presence Detection Technology.” Scientific Reports, 2024.
• Natarajan, A., Krishnasamy, V., Singh, M. “Device-Free Human Activity Recognition in Through-the-Wall Scenarios Using Single-Link Wi-Fi Channel Measurements.” IEEE Sensors Journal, 2025.
• Wang, Z., Jiang, K., Hou, Y., et al. “A Survey on CSI-Based Human Behavior Recognition in Through-the-Wall Scenario.” IEEE Access, 2019.
• Ropitault, T., da Silva, C. R. C. M., Blandino, S., et al. “IEEE 802.11bf WLAN Sensing Procedure: Enabling the Widespread Adoption of WiFi Sensing.” IEEE Communications Standards Magazine, 2024.
• Song, J., Qian, C., Guo, Y., Hua, K., Yu, W. “Attack Evaluations of Deep Learning Empowered WiFi Sensing in IoT Systems.” IEEE INFOCOM Workshops, 2023.
• Ambalkar, H., Zhao, T., Wang, X., Mao, S. “Adversarial Attack and Defense for WiFi-Based Apnea Detection System.” IEEE INFOCOM Workshops, 2023.