See motion through Wi-Fi.

Papers & resources.

Three preprints — MORIC, DoRF, and DoRF++ — and the UTHAMO dataset they are evaluated on. Scan or click through for the full text and the data.

Sensing people, without sensors on people.

Fall detection in elder care, gesture control, presence sensing, contactless breathing monitoring — human activity recognition powers them all. Cameras must see and record you; wearables must be worn and charged. Wi-Fi is already in the room: it works in the dark, through walls, and never films anyone.

Why Wi-Fi for sensing?

A camera goes blind in the dark and behind walls — and records everything it sees. A wearable only works while it's worn and charged. The Wi-Fi signal is already everywhere in the room, and it senses without watching.

Recognizing activity with Wi-Fi alone.

Two ordinary Wi-Fi devices, a person moving between them. The motion disturbs the signal on its way from transmitter to receiver — and from those disturbances alone, a model can tell a wave from a push. Through walls, in the dark, with nothing worn and nothing filmed.

It works — until the user or the environment changes.

Train and test on the same person in the same room, and accuracy looks solved. Hand the system to a new user or a new room, and the very same gesture arrives as a very different signal — accuracy can fall toward chance. Poor generalization is what keeps Wi-Fi sensing out of the real world.

The channel is a complex number.

Between every Wi-Fi transmitter and receiver sits a wireless channel. From known training symbols in each frame, the receiver estimates how that channel transformed the signal on every subcarrier — the Channel State Information. Each CSI value is complex: a magnitude, how much the signal was attenuated, and a phase, how much it was delayed.

Motion writes itself into the phase.

When the hand moves, every reflected path it touches gets longer or shorter. The path’s phase rotates, and the rate of that rotation is a Doppler shift in frequency — positive as the hand approaches, negative as it recedes. The channel becomes a recording of velocity.

One gesture, many echoes.

Indoor Wi-Fi never travels one way. Every wall and surface reflects it, and every reflected path is touched by a moving hand. The receiver hears the same motion many times — from many directions at once.

Every path is a 1-D camera.

Each path observes only the projection of the motion onto its own direction — a one-dimensional Doppler signal. The catch: nobody knows where these cameras point.

Use every echo, not just one.

Earlier Doppler methods collapse all multipath into a single velocity estimate. MORIC — our first proposed method — separates the CSI by propagation delay and extracts a Doppler projection from every component, then classifies them as an unordered, possibly repeated set. Shuffle the projections however you like: the answer stays the same.

More views, better transfer.

Treating the projections as a set pays off exactly where Wi-Fi sensing struggles: on people the model has never seen. MORIC substantially outperforms single-estimate Doppler and CSI-magnitude pipelines in cross-user accuracy.

MORIC PAPER

arxiv.org/abs/2506.12997

Puzzle pieces, never assembled.

MORIC treats the views as an unordered, possibly repeated set of motion descriptors and classifies them together — a big step for cross-user generalization. But each piece is interpreted on its own; the scene they describe together is never reconstructed.

Borrowed from computer vision.

Neural radiance fields (NeRF) reconstruct a full 3-D scene from ordinary 2-D photos taken from different viewpoints — even when the camera poses are unknown. DoRF brings the same idea to Wi-Fi sensing: its photos are 1-D Doppler projections, its cameras are multipath reflections, and the scene it reconstructs is the motion itself.

From pictures to motion.

The NeRF recipe carries over almost term for term. Camera viewpoints become multipath reflections; 2-D photos become 1-D Doppler projections; unknown camera poses become unknown viewing directions; and the reconstructed 3-D scene becomes the recovered 3-D motion.

One motion to explain them all.

DoRF treats the views as observations of a single latent 3-D motion — and recovers the motion and each view's direction jointly, the Wi-Fi analogue of unposed NeRF in computer vision.

A radiance field for Doppler.

The recovered motion is re-projected onto a uniform grid of directions on the unit sphere: every viewpoint, evenly covered, whatever the room's reflections happened to be. Each grid point senses its own 1-D Doppler signal — a virtual camera of its own.

Read on the sphere, recognized.

A spherical Transformer consumes the field directly — respecting its geometry and rotational symmetries — and names the activity. DoRF++ outperforms state-of-the-art Wi-Fi HAR methods in cross-user generalization, from a single access point.

The sphere pays off.

On users the model has never seen, DoRF++ pushes cross-user accuracy well beyond MORIC — the uniform, geometry-aware sphere turns scattered views into generalization that holds, from a single access point.

DORF++ PAPER

arxiv.org/pdf/2608.08381

The bandwidth ceiling.

Wi-Fi bandwidth is limited. At 80 MHz the delay resolution is about 12.5 ns — several meters of path length — so reflections arriving closer together than that cannot be told apart: they merge into one effective path. A merged path mixes several viewing directions at once, and its Doppler projection comes out blurred, like a photo from a shaking camera.

The whole story of DoRF, end to end.

One gesture, its echoes, their projections, the sphere, the model — and the lamp obeys.

UTHAMO: gestures from real rooms.

To test generalization honestly, we collected UTHAMO — a challenging hand-gesture dataset recorded with commodity Wi-Fi hardware. Participants perform four gestures — circle, left–right, up–down, and push–pull — while CSI is recorded from a single multi-antenna access point.

UTHAMO DATASET

ieee-dataport.org →

Latest updates.

Cite this work.

@article{hasanzadeh2026dorfpp,
  title   = {DoRF++: Spherical Representation Learning over Doppler
             Radiance Fields for Robust Wi-Fi Sensing},
  author  = {Hasanzadeh, Navid and Valaee, Shahrokh},
  journal = {arXiv preprint arXiv:2608.08381},
  year    = {2026}
}

@article{hasanzadeh2025dorf,
  title   = {DoRF: Doppler Radiance Fields for Robust Human Activity
             Recognition Using Wi-Fi},
  author  = {Hasanzadeh, Navid and Valaee, Shahrokh},
  journal = {arXiv preprint arXiv:2507.12132},
  year    = {2025}
}

@article{hasanzadeh2025moric,
  title   = {MORIC: CSI Delay-Doppler Decomposition for Robust
             Wi-Fi-based Human Activity Recognition},
  author  = {Hasanzadeh, Navid and Valaee, Shahrokh},
  journal = {arXiv preprint arXiv:2506.12997},
  year    = {2025}
}

@data{qjfg-s580-25,
  title     = {UTHAMO: A Multi-Modal Wi-Fi CSI-Based Hand Motion Dataset},
  author    = {Hasanzadeh, Navid and Djogo, Radomir and
               Salehinejad, Hojjat and Valaee, Shahrokh},
  publisher = {IEEE Dataport},
  doi       = {10.21227/qjfg-s580},
  year      = {2025}
}

Get in touch.

Questions about the method, the dataset, or reproducing the results are welcome — as are collaborations on Wi-Fi sensing that has to work outside the lab.