An implementation of PropSplat on the grid: measure the received signal between every Wi-Fi radio on the grid's nodes, train the model on a grid server, and see what it learned about the room, without and with calibration of the radios, and with one chip family or all of them.
The method is from W. Bjorndahl, M. P. Singh, F. Nouri and J. Camp, "PropSplat: Map-Free RF Field Reconstruction via 3D Gaussian Propagation Splatting", arXiv:2605.08035, May 2026. This page and its package are an independent implementation from the paper's Section III, not the authors' code.
grid: 129 nodes, 205 radios of five chip families, 2 rounds. Training with 15 seeds on srv5, all radios and ath9k only. Both ran through propsplat_run.sh, the sequence cosmos-tutorial run propsplat calls; the notebook ran on the shipped sample.PropSplat (Bjorndahl et al., arXiv:2605.08035) predicts the path loss of a link from the positions of its two ends. It starts from a distance law with one learned exponent, then adds a cloud of 3D Gaussians. Each Gaussian holds a loss offset in dB, and a link collects the offset of every Gaussian near its straight line. Training moves and reshapes the Gaussians until the predictions match the measurements.
Every run trains the model twice: on the raw links, and with calibration, which first removes a fixed offset per receiver and per transmit chip learned from the training links. The tutorial shows both results, and compares a survey of every radio with one of the ath9k radios alone.
| Where | grid |
| Nodes | any number of grid nodes, every Wi-Fi radio on each; the shipped survey has 129 nodes and 205 radios |
| Server | one grid server for training, CPU only |
| Images | propsplat.ndz on the nodes, baseline.ndz on the server |
| Radio | 2.4 GHz, 1 Mbit/s injected frames, channel 1, 6 or 11 |
| Time | imaging about 20 min, survey about 4 min per round for 129 nodes, training see below |
| Results | a notebook, also readable with no reservation on the shipped sample |
grid - see /wiki/public/getting-started.| Domains | grid |
| Nodes | any number, named in OMF_NODES |
| On the console | python3, ansible-playbook, ssh, tar, curl, xargs, shuf |
cosmos-tutorial check propsplat verifies, on each node:
Not in the package - you provide:
Power the nodes on, load the images, then power them on again: omf load cannot wake a grid node that is off, and it leaves every node off when it finishes.
cd /usr/share/cosmos-experiments/propsplat
NODES=$(python3 propsplat_lib.py nodes)
omf tell on -t "$NODES"
omf load -t "$NODES" -i propsplat.ndz -r 0 -o 1200 -b 600
omf tell on -t "$NODES"
omf tell on -t srv5
omf load -t srv5 -i baseline.ndz -r 0 -o 1200 -b 900
omf tell on -t srv5
export OMF_NODES=$NODES
Nodes that fail to image or boot should be removed from OMF_NODES. Retry them with omf load -t system:topo:checkinfailed, after omf tell on on the same set.
The tutorial is installed on the console. There is nothing to download.
cosmos-tutorial check propsplat
cosmos-tutorial run propsplat # ansible
cosmos-tutorial jupyter propsplat # jupyter
cosmos-tutorial run propsplat --mode teardown # teardown
Training is one job per split, calibration and seed: 2 × 2 × propsplat_seeds jobs, each on its own CPU. Time depends on propsplat_iterations and the server; see What you should see. Keep the whole run inside your reservation: a run stopped early fails the contract.
There is no manual mode: the survey is a loop over every node and radio, and a hand copy of it would drop the checks that make the data trustworthy.
Change any of these with cosmos-tutorial run propsplat --params my.yml. An unknown key or a wrong type is refused. cosmos-tutorial params propsplat prints the file.
| Key | Default | Meaning |
|---|---|---|
propsplat_server |
srv5 |
The grid server that trains the model. |
propsplat_links |
`` | Where the links come from: "" measures them; a directory holding nodes.csv and links.csv skips the survey. "sample" is the survey shipped with this package. |
propsplat_channel |
0 |
Survey. |
propsplat_rounds |
3 |
|
propsplat_frames |
200 |
|
propsplat_gap_s |
0.005 |
|
propsplat_rx_exclude_ids |
168c:0034, 168c:0033 |
Chip ids whose RSSI is not physical: AR9462 (0034) reads around +10 dBm and AR958x (0033) -45 to -1 dBm, so those radios only transmit. |
propsplat_rx_max_dbm |
0 |
Any other radio that reads a sender above this, in dBm, only transmits too. |
propsplat_min_frames |
50 |
|
propsplat_drivers |
0 items | Train on the radios with these drivers only, e.g. [ath9k]; [] keeps every radio. |
propsplat_splits |
2 keys | Control variables of the model. |
propsplat_seeds |
3 |
Repeats of every (split, calibration) job. |
propsplat_n_gaussians |
2000 |
|
propsplat_iterations |
12000 |
|
propsplat_tx_power_dbm |
20.0 |
nominal: a monitor-mode radio cannot set or report it |
propsplat_device |
cpu |
cpu, cuda, or auto (a GPU if torch sees one); a P100 was slower than 20 CPUs |
propsplat_workers |
0 |
0: one job per CPU, or every job at once on a GPU |
The survey needs no association. Each node sends 200 broadcast frames from each of its radios, with its row, column and radio number in the source address, and every radio on the other nodes captures them in monitor mode. One capture per radio therefore holds every sender. RSSI comes from the radiotap header.
Grid nodes carry one to three radios. propsplat-radio loads every driver it supports and numbers each node's radios in bus order. On 2026-11-04 the survey found:
| Driver | Chips | Radios | Sends | Receives |
|---|---|---|---|---|
ath9k |
AR928X, AR922X, AR9462, AR958x | 75 | 74 | 48 trusted (see below) |
iwlwifi |
Intel 6250, Intel 7260 | 71 | 71 | 71 |
ath10k_pci |
QCA988x | 53 | none | 53 |
rt2800usb |
Ralink USB | 4 | 4 | 4 |
carl9170 |
AR9170 USB | 2 | 2 | 2 |
ath10k radios receive in monitor mode, but their stock firmware sends no injected frames. A node with two radios is two transmitters in one place, which separates a radio's offset from its location. Links between radios on the same node are left out. Splits hold out whole nodes, so no radio at a test position is trained on.
Three things about the radios shape every result:
propsplat_rx_exclude_ids). Any other radio that reads a sender above propsplat_rx_max_dbm (0 dBm) only sends too.The model is the paper's: 2,000 Gaussians, trained on links from a random 70% (dense) or 12.6% (sparse) of the transmitting nodes and scored on the rest. Localisation is the paper's test too: each held-out transmitter is found by matching what the receivers heard against the model's map.
Calibration answers the third point. It fits the distance law to the training links, then takes the mean leftover of each receiver and of each transmit chip as that radio's offset and removes it. A held-out node has never sent, so on the transmit side only its chip type is known. Everything is learned from training links only.
This is the survey shipped with the package (propsplat_links: sample): grid, 2026-11-04, 129 nodes, 205 radios, 151 transmitters, 178 trusted receivers, 60,930 readings over 2 rounds on channel 1. Training used the defaults, 2,000 Gaussians and 12,000 iterations, with 15 seeds.
The 12.2 dB swap difference splits by chip pair into a fixed offset, which calibration can remove, and a spread around it, which it cannot:
| Chip pair | Links | Offset (dB) | Spread (dB) |
|---|---|---|---|
| ath9k → Intel | 3,317 | 13.6 | 8.0 |
| Intel → Intel | 2,430 | 0.2 | 8.3 |
| ath9k → ath9k | 1,081 | -0.8 | 6.0 |
| Intel → Ralink USB | 283 | -8.2 | 6.1 |
| ath9k → Ralink USB | 186 | 5.3 | 4.8 |
| AR9170 USB → Intel | 140 | -7.1 | 6.7 |
| ath9k → AR9170 USB | 94 | 20.6 | 4.3 |
Offsets of up to 21 dB between chip families are what calibration is for. The spread is not: two Intel radios disagree by 8.3 dB link by link, against 6.0 dB for two ath9k radios.
Mean and spread over the 15 seeds:
| Split | Calibration | MAE (dB) | RMSE (dB) | p95 (dB) | Localisation, median (m) |
|---|---|---|---|---|---|
| dense, 70% | without | 6.03 ± 0.44 | 8.17 ± 0.94 | 16.6 | 1.56 |
| dense, 70% | with | 3.82 ± 0.32 | 6.11 ± 1.04 | 14.1 | 1.18 |
| sparse, 12.6% | without | 6.98 ± 0.37 | 9.35 ± 0.47 | 18.4 | 2.48 |
| sparse, 12.6% | with | 5.25 ± 1.16 | 7.60 ± 1.83 | 14.9 | 1.93 |
Calibration removes about 2 dB of error on both splits and about a quarter of the localisation error.
The same survey, trained on the ath9k radios alone (propsplat_drivers: [ath9k]): 75 radios, 74 transmitters, 10,802 readings. With calibration:
| Split | Radios | RMSE (dB) | p95 (dB) | Localisation, median (m) |
|---|---|---|---|---|
| dense, 70% | ath9k only | 4.90 ± 0.42 | 10.3 | 1.35 |
| dense, 70% | every radio | 6.11 ± 1.04 | 14.1 | 1.18 |
| sparse, 12.6% | ath9k only | 6.27 ± 1.00 | 12.7 | 2.24 |
| sparse, 12.6% | every radio | 7.60 ± 1.83 | 14.9 | 1.93 |
The two numbers move in opposite directions:
The two link errors are scored on different test links, each run on its own held-out transmitters, so read them as what each survey supports rather than as one model scored twice. The ath9k-only result agrees with the 2026-11-02 survey of 73 ath9k nodes (dense, with calibration: 4.37 dB RMSE, 1.11 m).
Measured on the grid's servers:
| CPU platform | Device | Jobs | Time |
|---|---|---|---|
| 2× Intel Xeon E5-2698 v3, 2 × 16 cores, 64 logical CPUs | CPU, one job per CPU | 60 | 36 min |
| 2× Intel Xeon E5-2640 v4, 2 × 10 cores, 20 logical CPUs | CPU, one job per CPU | 12 | 15.5 min |
| same, with an NVIDIA Tesla P100 12 GB | GPU, all 12 jobs at once | 12 | 20 min |
| Intel Xeon Gold 6126, 4 logical CPUs | CPU, 4 jobs at a time | 12 | about 66 min |
The GPU is faster per job (168 s for one job alone), but the model trains on batches of 128 links, too small to keep a GPU busy, so a server's CPUs win. That is why propsplat_device defaults to cpu; set it to cuda to try a GPU.
For scale, the paper reports 4.74 dB dense and 6.51 dB sparse RMSE on its BLE data, a different building with different radios.
The notebook shows the survey, the noise numbers and the chip-pair table, both result tables and the comparison, a predicted coverage map, the learned Gaussians, and one link broken down Gaussian by Gaussian.
Change one parameter at a time with --params my.yml.
| Parameter | Question |
|---|---|
propsplat_drivers |
Which radios help? Train on one chip family, e.g. [ath9k] or [iwlwifi], and compare. With propsplat_links: sample this needs no reservation. |
propsplat_seeds |
How much do the results move between random choices of held-out nodes? More seeds give tighter statistics; jobs grow with them. |
propsplat_n_gaussians, propsplat_iterations |
Does a smaller or shorter-trained model generalise better? On the 2026-11-02 survey, with calibration, 2,000 Gaussians at 12,000 iterations reached 2.07 dB training RMSE but 4.27 dB on held-out transmitters; 50 Gaussians gave 2.91 and 2.93 dB (dense, one seed). |
propsplat_splits |
How little data does the model need? |
propsplat_channel |
Does a busier channel change the survey? |
propsplat_rounds, propsplat_frames |
How much survey time does a stable reading need? |
contract.json in the run directory passes only when every training job finished. Check repeatability_db and reciprocity_db, and read the table: calibration should lower the error on both splits.
cosmos-tutorial run propsplat --mode teardown
This stops the captures and the training unit. Nothing else needs tidying: the next experimenter's image wipes the nodes.
| Symptom | Cause |
|---|---|
omf load reports many nodes "failed to check in" while omf stat says POWERON |
The nodes were off. omf load resets nodes but does not power them on, and records them as on anyway. Run omf tell on on the set, then load again. |
Nodes answer ssh but refuse the root key after an omf load was interrupted |
They are still in the imaging environment. Power-cycle them: omf tell offh, wait 20 s, omf tell on. |
omf load fails at once with Frisbee: User '<you>' does not exist |
The imaging server cannot look up accounts. Report it to the testbed staff; it is not your account. |
A server never checks in during omf load |
Servers can take over 10 min. srv4 did not check in within 14 min on 2026-11-02; use another server and set propsplat_server. |
| ath10k radios receive but never appear as transmitters | Expected: their stock firmware sends no injected frames. |
| RSSI above 0 dBm | The dataset step marks that radio as send-only and prints it. Add its chip id to propsplat_rx_exclude_ids once confirmed. |
A node has fewer radios than lspci shows |
propsplat-radio printed the radio that would not enter monitor mode, in survey/radio-final/<node>.err. Its driver may be missing, or its firmware. |
contract failed at the end of a run |
Training was stopped or a job died. out/job-*.json holds what finished. |
Package cosmos-experiments-propsplat 0.1.0-3 - on every console it is released to (cosmos-tutorial list shows the installed version). The code is installed on the console in /usr/share/cosmos-experiments/propsplat/; the programs in its node/ subdirectory are copied to the nodes by every run, and the nodes load propsplat.ndz for the rest of what they need. To change anything, cosmos-tutorial copy propsplat puts an editable copy in the current directory, and cosmos-tutorial run propsplat then runs that copy instead of the installed one (cosmos-tutorial which propsplat shows which will run). Repository (GitLab account needed, not needed to run or change it): packages/cosmos-experiments-propsplat/usr/share/cosmos-experiments/propsplat.