AI Can See Without Cameras: How WiFi and Radio Signals Are Redefining Machine Vision
For decades, machine vision meant cameras. Computer vision systems relied on pixels of light, struggling in darkness, through smoke, or behind walls. That constraint is dissolving. Artificial intelligence can now reconstruct human poses, track movement, measure vital signs, and even map hidden objects by reading the subtle disturbances that people and objects create in everyday radio signals. WiFi was the first practical demonstration. It is no longer the ceiling.
The principle is straightforward yet powerful. When a WiFi router transmits, its radio waves scatter off everything in the environment. The way those waves arrive at a receiver—captured as Channel State Information, or CSI—contains detailed information about the space. Amplitude and phase variations across subcarriers act like a fingerprint of the physical scene. A person walking, breathing, or even shifting posture creates measurable multipath changes. Deep learning models trained on paired radio and visual data learn to translate those changes into usable spatial understanding: presence, activity, body keypoints, or three-dimensional shape.
Early academic work established the foundation. Researchers at Carnegie Mellon University showed that commodity WiFi signals could support dense human pose estimation. Their system mapped phase and amplitude from multiple antennas into body surface coordinates, achieving an average precision of 87.2 percent at the 50 percent IoU threshold for detection in controlled indoor settings. It could handle multiple people and worked without lighting or direct line of sight. Performance was lower than pure camera methods and dropped when environments changed, highlighting the generalization challenge that still occupies researchers. Yet the core idea was proven: radio waves carry enough structure for AI to recover pose.
Open-source efforts have accelerated practical deployment. Projects building on the CMU research now run on inexpensive hardware such as ESP32 boards costing only a few dollars. These systems extract CSI in real time, estimate full-body skeletons with 17 or more keypoints, and recover breathing rates between 6 and 30 breaths per minute along with heart rates in the 40–120 range. Recent models trained on public benchmarks like MM-Fi have reported torso-normalized PCK@20 scores above 82 percent in matched conditions, outperforming earlier published baselines. Edge-optimized versions fit into tens of kilobytes and run in fractions of a millisecond, making always-on sensing feasible on low-power devices. Honest evaluations also reveal limits: zero-shot transfer across rooms or subjects often collapses, requiring short calibration. Still, the trajectory from research paper to deployable code has been rapid.
Generative AI has pushed the technology further from detection toward true reconstruction. At MIT, researchers developed Wave-Former, a system that uses millimeter-wave signals—closely related to the frequencies in modern WiFi—to recover the three-dimensional shape of completely occluded everyday objects. Traditional radar-style approaches typically capture only the top surface because of specular reflections. Wave-Former proposes candidate surfaces from the radio measurements, then applies a transformer-based generative model trained with physics-aware synthetic data to complete the full geometry. In tests, it raised recall from 54 percent to 72 percent while holding high precision, reconstructing cans, boxes, utensils, and other common items hidden behind cardboard, drywall, wood, or fabric. A related approach reconstructs entire indoor scenes by analyzing reflections from people moving through the space, using a single stationary sensor rather than a mobile robot. These advances turn sparse radio observations into coherent spatial models usable by robots or smart environments.
Standardization is catching up. IEEE 802.11bf, published in September 2025, formalizes WLAN sensing procedures. It defines how devices exchange sensing capabilities, set up measurement transmissions, and report results across both sub-7 GHz and 60 GHz bands. The amendment does not prescribe specific algorithms—those remain vendor or application specific—but it creates an interoperable foundation so that future access points and clients can treat sensing as a native function rather than a research hack.
The larger horizon is Integrated Sensing and Communication, or ISAC, a core pillar of 6G. Instead of separate radar and communication systems, future cellular networks will use the same signals and infrastructure for both. Base stations become distributed sensors capable of detecting people, vehicles, drones, and environmental features even when those targets carry no device. Early work is already appearing in 5G software upgrades for specialized uses such as drone detection near stadiums, while full native ISAC remains a 6G feature. Industry partnerships, including collaborations between major equipment makers and operators, are testing human presence sensing, network optimization, and low-altitude airspace monitoring. The same radio infrastructure that delivers data will increasingly deliver spatial awareness at city scale.
Applications span multiple domains. In homes and assisted living, camera-free fall detection and activity monitoring protect privacy while providing safety. Hospitals can track patient movement and vital signs without wearable sensors or visual recording. Warehouses and factories gain robots that locate items inside boxes or behind obstacles. Security systems can detect intruders through walls without capturing identifiable images. Disaster response teams may one day locate survivors under rubble using ambient signals. Smart buildings can adjust lighting, heating, and access based on precise occupancy without invasive surveillance.
These capabilities bring corresponding risks. Radio sensing does not record faces or clothing, which reduces certain privacy harms, yet it can still reveal routines, presence patterns, gait, and physiological data. A low-cost node placed near an existing access point can map movement through walls. Open tools lower the barrier for both beneficial and malicious use. Cross-environment generalization remains imperfect, and performance can degrade with interference or complex multipath. Regulatory frameworks for always-on radio sensing are still emerging.
Cameras will continue to dominate tasks requiring fine visual detail and color. For presence, motion, pose, through-obstacle awareness, and privacy-sensitive monitoring, however, radio-based AI sensing is already competitive and in some dimensions superior. The signals that once only carried data now carry information about the physical world itself. WiFi proved the concept was possible. Generative models, standardized protocols, and the coming generation of cellular networks are turning it into infrastructure. Machine vision is expanding beyond the camera. The radio environment is becoming the new sensor.
About The Author
