Powered byRadioVerum
ACM MobiCom ’26

RadioSight

Predictive mmWave XR Network Optimization from Dynamic Neural Radio Fields

Lihao ZhangPaul KudybaZhenlin AnHaijian Sun

University of Georgia

RadioSight overview. Left: the physical world, with an access point, an XR user whose predicted path is drawn as a dashed arrow, and a person walking past. Right: the neural radio field, where the system finds a reflected path to the user and switches to it.
RadioSight keeps a radio twin of the room up to date while people move, and picks each user's beam before the link breaks.

Walk an XR headset through a room we measured at 28 GHz, add people who block the signal, and get what the headset would see: its camera view, the spatial spectrum and the power of every receive beam.

Get started Paper (ACM) arXiv

01 · Features

A measured room you can query.

Put headsets and people into a room we measured at 28 GHz, and get data in the same shape as our beam scans.

01

Any head pose

The camera view, the 28 GHz spatial spectrum and the power of all 55 receive beams, from anywhere in the room.

02

Recorded motion

Replay the walk we took while measuring the room, or five recorded XR activities: walking, gaming, interacting, navigating.

03

People who block

Animated people walk planned routes around the furniture, or go where you send them, and shadow the signal.

04

Several XR users

Each headset sees the others: who is in view, where, and how much of them is visible.

05

Your own agents

Drive every headset and person from Python, step by step, and react to what they measured.

06

Data in one folder

Beam scans, spectra, camera images, arrival directions, line of sight and the base-station beam, plus a video of the run.

02 · Previews

Recorded motion, people in the way.

Each clip is generated, not filmed: floor map · what the headset sees · the 28 GHz spectrum it receives. Brighter means stronger.

Recorded head motion

Measurement walkThe walk we took while measuring the room.
WalkingA recorded XR activity, replayed in this room.
GamingQuick head turns while someone walks by.

People who block

Someone walking pastA person walks between the headset, the base station and the tripod.
Crossing the line of sightThe headset stands still; a person keeps walking through its beam.
Two XR usersEach headset sees, and is blocked by, the other people.
03 · Get started

A few commands to your first run.

Linux with an NVIDIA GPU. Every run also writes a short page with its video and how to load the data.

pip install torch==2.4.1 --index-url https://download.pytorch.org/whl/cu124
pip install gsplat==1.5.3 --extra-index-url https://docs.gsplat.studio/whl/pt24cu124
git clone https://github.com/Pervasive-Intelligence-Lab/radiosight.git && cd radiosight && pip install -e .
radiosight-sim download all --dest twins
radiosight-sim run scenarios/nlos_tx1_recorded_walk_pedestrian.yaml --out runs/demo

Scenarios and Python API

04 · Cite

Using the twin? Please cite the paper.

Lihao Zhang, Paul Kudyba, Zhenlin An, and Haijian Sun. ACM MobiCom 2026, Austin, TX, USA.

BibTeX
@inproceedings{zhang2026radiosight,
  title     = {RadioSight: Predictive mmWave XR Network Optimization from Dynamic Neural Radio Fields},
  author    = {Zhang, Lihao and Kudyba, Paul and An, Zhenlin and Sun, Haijian},
  booktitle = {Proceedings of the 32nd Annual International Conference on Mobile Computing and Networking (MobiCom '26)},
  year      = {2026},
  address   = {Austin, TX, USA},
  publisher = {ACM},
  doi       = {10.1145/3795866.3844465}
}