nvidia/aerial-isac-pusch-hest

Aerial ISAC PUSCH Channel Estimates

Uplink PUSCH DMRS channel estimates from a 5G CBRS cell running indoors on the NVIDIA Aerial testbed, paired with camera-derived floor positions of a person walking through the cell: 1.19 million estimates over 18 runs, nine with a person in the area and nine recorded empty as a background reference.

Plan view in the label coordinate frame, with the recorded track of a clear run and of an obstacle run.

Dataset Description:

This dataset provides uplink PUSCH DMRS channel estimates captured by a private indoor 5G CBRS cell, paired with camera-derived floor positions of a single person moving through the monitored area. It comprises 18 recordings of roughly two minutes each, all made in a working open-plan office on an NVIDIA campus: nine have a person walking the area in varied patterns, and nine cover the same geometry recorded empty as a background reference. In runs 15 to 18, the last two of each condition, a black cloth screen stands inside the area as a static obstacle.

The capture is bistatic and UE-collaborative: a commercial handset runs an ordinary uplink data session, and the O-RU estimates the channel from the PUSCH DMRS of every scheduled slot, exactly as it must to decode the data. The person carries no radio and is sensed only through the changes they cause in that channel. What is published is the physical layer's own channel estimate, one step past the antenna port, which is what a dApp or an ML model receives from L1 in production.

Repository structure

data/aerial-isac-pusch-hest-NNN/     one folder per run, NNN being the run number
├── pusch_dmrs_hest.h5   10-18 GiB   channel estimates and per-record radio metadata
├── labels.jsonl         7-12 MiB    floor position per labelled record, target runs only
└── metadata.json        5 KiB       machine-readable run summary
data/runs.json                       index over all 18 runs
assets/                              figures used by this page

This collection is suitable both for training and for evaluation. It is not intended for individual identification: the ground-truth videos are not published, no image-domain coordinates are released, and the labels carry no identity beyond a run number. This dataset is ready for commercial or non-commercial uses.

Dataset Owner(s):

NVIDIA Corporation

Dataset Creation Date:

2025-10-28

Version:

1.0.0

Previous Version(s): N/A

License/Terms of Use:

Creative Commons BY 4.0. See the LICENSE file for the full terms.

Intended Usage:

The dataset supports the development and benchmarking of CSI-based indoor sensing models. A model can regress a person's floor position from standard channel estimates, detect presence by contrast with the background runs, or be tested for robustness to a static obstruction. Classical processing applies just as well, since delay profiles and Doppler spectra follow directly from the frequency-domain estimates and their timestamps. This is also the measurement campaign behind the localization results of reference [1], where the cuSense dApp was trained on five of the target runs, took its static channel template from a background run, and reached a mean error of 77 cm on unseen runs. The dataset is not intended for identifying individuals.

Dataset Characterization

Data Collection Method
Hybrid: Automatic/Sensors/Manually Collected

Labeling Method
Hybrid: Automated, Manually-Labelled

Radio and video were recorded simultaneously in an indoor lab on an NVIDIA campus. The transmitter is an unmodified Samsung Galaxy S22 handset running an iperf uplink session; the receiver is a Foxconn RPQN CBRS 4800 4T4R O-RU feeding NVIDIA Aerial CUDA-Accelerated RAN L1. The published estimates are the cuPHY output for each PUSCH occasion, so these are the same estimates the physical layer used to decode the data. Ground truth comes from an iPhone 14 filming the area from the O-RU stand.

Acquisition
Waveform	5G NR uplink PUSCH DMRS, band n48 CBRS at 3.65 GHz, 100 MHz (273 PRB, 30 kHz SCS)
Duplexing	TDD, DDDDDDSUUU pattern, three uplink slots per half-frame
Estimator	cuPHY two-stage LS + MMSE, frequency domain, pre-equalization
Capture	1,187,660 records over 35.1 min in 18 runs, about 555 records per second
Record	complex64, [4 rx antenna, 1 layer, 3276 subcarrier, 3 DMRS symbol]
Hardware	Samsung Galaxy S22 UE, Foxconn RPQN CBRS 4800 4T4R O-RU, single cell
Platform	NVIDIA ARC-OTA release 1.6 (now Aerial Testbed): Aerial CUDA-Accelerated RAN 25-2 L1 on a GH200 server, OpenAirInterface L2/L3 and core, PTP grandmaster clock
Geometry	Bistatic, UE-collaborative, 17.0 m between the radios, 6.78 x 10.05 m target area
Traffic	iperf UDP uplink, 100 Mbps target
Ground truth	iPhone 14 at 59.94 fps with a burned-in UTC timestamp; the videos are not published
Site	Open-plan office on an NVIDIA campus

Labels are camera-derived, not radio-derived, so they are independent of any sensing model benchmarked against them. For every video frame an Ultralytics YOLOv8 person detector was run and the highest-confidence box kept, a single subject walking in each run, and the box centroid was projected onto the floor through a planar homography fitted to the hand-annotated corners of the taped rectangle in that run's video. The detector was an internal tool only, and neither its code nor its weights (AGPL-3.0) is redistributed here. What is published is its numeric output: floor coordinates in metres, a detector confidence and the frame index.

Dataset Format

Modality: Radio channel estimates, Structured labels (JSON/JSONL)
Format: HDF5 (complex64), JSONL, JSON, PNG

Per-file sizes and roles are listed under Repository structure above, and the HDF5 layout is described in The estimates below.

Dataset Quantification

Record Count: 1,187,660 PUSCH channel-estimate records across 18 runs, 260 GiB of complex64; 629,964 labelled records derived from 63,745 camera frames; nine runs are empty-room references carrying no labels

Feature Count: Each record holds a channel estimate over 4 receive antennas, 1 layer, 3276 subcarriers and 3 DMRS symbols, with per-record radio metadata: timestamp, system frame and slot, PRB allocation, symbol allocation, layer and DMRS port, modulation and coding scheme, transport block size, and radio identity. Each label row carries the record index, frame number, TAI timestamp, floor position in metres, detector confidence, and distance outside the marked area

Total Data Storage: ~280 GB

Getting started

The whole release is 260 GiB, so a single run is usually the right unit to start with:

pip install -U "huggingface_hub[cli]"
hf download nvidia/aerial-isac-pusch-hest --repo-type dataset --local-dir aerial-isac-pusch-hest \
  --include "data/runs.json" "data/aerial-isac-pusch-hest-002/*"

Each record is chunked on its own, so reading a random record costs about 1.6 ms and touches only its 307 KiB. Nothing beyond h5py and numpy is needed:

import h5py, json

RUN = "data/aerial-isac-pusch-hest-002"

with h5py.File(f"{RUN}/pusch_dmrs_hest.h5") as f:
    est, md = f["channel_estimates"], f["metadata"]

    i = 2726                                  # record index
    h = est[i]                                # [4, 1, 3276, 3] complex64
    n_sc = int(md["num_prbs"][i]) * 12        # subcarriers actually allocated
    sc0 = int(md["start_prb"][i]) * 12        # carrier index of the first valid one
    h = h[:, :, :n_sc, :]                     # drop the zero padding
    print(h.shape, md["timestamp"][i], dict(f.attrs)["channel_estimator"])

rows = map(json.loads, open(f"{RUN}/labels.jsonl"))
labels = {r["record_index"]: r for r in rows}       # join on record_index

The estimates

channel_estimates is the cuPHY output for one PUSCH occasion: a two-stage least squares plus MMSE estimate over the DMRS positions, interpolated across the allocated subcarriers, with no interpolation in time. The last axis holds the three DMRS symbol positions of that slot, all populated, so each record carries three snapshots within the same 0.5 ms.

Records arrive once per scheduled uplink slot, roughly 555 per second: 30 kHz numerology gives 0.5 ms slots, and the TDD pattern plus the scheduler leave about one PUSCH every 1.8 ms. A 60 fps camera frame therefore spans about nine records.

Zero padding. The subcarrier axis is fixed at 3276, the full 273 PRB carrier, but the scheduler did not always allocate all of it: 81% of records use 273 PRB and 18% use 261. Valid samples are packed from index 0, so a record holds num_prbs * 12 meaningful subcarriers followed by zeros, and start_prb * 12 gives the carrier offset of the first of them. Reference [1] reaches the same active bins without the metadata, by masking wherever |H| falls below 1e-10.

metadata/ holds one value per record, uncompressed:

Dataset	Meaning
timestamp	Record time in seconds, TAI scale, see Clocks
record_id	Sequential index as written by the capture
sfn, slot	System frame number and slot inside the frame
num_prbs, start_prb	PUSCH allocation, see the padding note above
num_symbols, start_sym	PUSCH duration in symbols, 12 or 13, starting at 0
layers, dmrs_ports, dmrs_scrm_id	1 layer on 1 DMRS port throughout, scrambling identity 51
mcs_index, tb_size	Modulation and coding scheme and transport block size, both moving with link adaptation
rnti	Ephemeral radio identity of the UE for that connection

File attributes carry the radio configuration, the estimator identity and the provenance of the file; dict(f.attrs) prints all of them.

The files are stored one record per chunk with the HDF5 byte-shuffle filter and gzip level 4, both core filters, so any HDF5 reader opens them with no plugin and the values are exactly the complex64 the estimator produced.

Labels

The nine even-numbered runs have a person in the area and a camera recording. For each of those runs, labels.jsonl carries one row per labelled radio record:

{"record_index": 2726, "frame_number": 2699, "timestamp_tai": 1761679834.2685,
 "x_m": 0.3988, "y_m": 3.4289, "confidence": 0.8866, "dist_outside_m": 0.0}

Field	Meaning
record_index	Index into channel_estimates, the join key
frame_number	Camera frame this record falls in; about nine records share one frame
timestamp_tai	metadata/timestamp of that record, repeated for convenience
x_m, y_m	Position of the person in the floor frame of Scene and geometry
confidence	Detector confidence for that frame, 0.25 to 0.94
dist_outside_m	Distance from the taped rectangle, 0 when inside

Across the nine target runs labels cover 92% of records, per run 64% to 97%. The uncovered records precede the first or follow the last labelled frame, or fall in frames where the detector lost the person; they are valid radio data without a position. The odd-numbered runs carry no labels by construction and are the background reference for the empty room.

Label quality

Ground truth is a 60 fps camera and a homography annotated by hand, so treat a position as good to a few centimetres at best and as an average over the 16.7 ms the frame covers. 97.3% of rows sit inside the taped rectangle and 99.4% within 0.2 m of it, the remainder being the person stepping on or just over the tape plus homography error near the far corners, where the camera view is most oblique.

Run 4 is the exception: one stretch of about 5.6 s has the person up to 1.4 m outside the rectangle, and 4.5% of its rows lie more than 0.4 m out, against zero elsewhere in the release. In run 16 the detector locked onto a second person in the background for the first 13 frames, so those 129 rows were dropped and the records are unlabelled. dist_outside_m gives the distance for every row, so you can set whatever threshold your task needs.

Labels are continuous coordinates rather than grids. The grid block of metadata.json records the 0.2 m resolution they were produced at, 35 x 52 cells from the taped origin, for models that want them rendered onto a grid. Reference [1] trains on a padded version of that grid, 42 x 63 cells of the same size with its origin at (-0.69, -1.11) m, each one-hot target cell smoothed by a Gaussian of sigma 8 cells.

Scene and geometry

The cell is an occupied open-plan office with cubicles, partitions and a support column. A 6.78 x 10.05 m rectangle taped on the floor defines the label frame: origin at the corner nearest the camera, x along the 6.78 m edge, y along the 10.05 m edge, z up, right-handed.

Quantity	Value
Taped area	6.78 x 10.05 m, origin at its near corner
O-RU (receiver)	(-2.74, -3.02, 2.0) m, on a stand behind the near corner
UE (transmitter)	(7.01, 10.86, 0.67) m, on a table just past the far corner
Camera	co-located with the O-RU on the same stand
Whiteboard, all runs	1.07 x 1.68 m, standing diagonally in front of the UE
Cloth screen, runs 15 to 18	2.13 x 1.73 m, spanning x 3.38 to 5.51 m at y 5.81 m

The testbed as deployed: the lab from both ends, the UE, and the O-RU on its stand.

Two fixtures stand in the room. A mobile whiteboard sits diagonally in front of the UE for every run, so it is part of the static background rather than a variable. The obstacle of runs 15 to 18 is a folding screen of four black cloth panels on a thin metal frame, placed across the middle of the area to break the camera's line of sight; cloth scatters weakly, so expect it to perturb the channel rather than block it the way a metal panel would.

Distances were tape-measured to the nearest inch and converted, so treat the positions as good to a few centimetres. The image-to-floor mapping is fitted separately for each run, from the corners of the taped rectangle as they appear in that run's video, so a small camera shift between runs does not propagate into the positions. Ground truth was filmed by an iPhone on the O-RU stand, running an app that burns a UTC timestamp into every frame; the fixed camera visible under the O-RU in the photo was not used for these labels.

The marked area and the space immediately around it were kept clear during every run, background ones included, but movement by other occupants further back in the room cannot be excluded, so the empty-room channel is close to but not perfectly stationary.

Clocks

Radio timestamps are TAI seconds from the PTP grandmaster that disciplines the cell, while the camera stamped UTC. TAI ran 37 s ahead of UTC in October 2025, and the record-to-frame assignment in labels.jsonl already accounts for it, so nothing needs applying. The residual after that 37 s is 1 to 2 ms of phone clock skew, and the assignment spreads over 16 ms peak to peak, one camera frame period, so alignment is good to about ±8 ms.

Runs

Runs alternate: an empty background recording, then the same geometry with a person walking varied patterns, zigzag, lawnmower, circular and random. Runs 15 to 18 repeat that with the cloth screen standing inside the area as a static obstacle.

Run	Person	Screen	Records	Duration	Labels	Coverage
1	no	no	58,400	104 s	-	-
2	yes	no	72,940	131 s	70,165	96%
3	no	no	55,580	99 s	-	-
4	yes	no	83,800	145 s	74,467	89%
5	no	no	60,200	108 s	-	-
6	yes	no	75,320	134 s	72,548	96%
7	no	no	54,400	97 s	-	-
8	yes	no	73,920	130 s	71,214	96%
9	no	no	55,400	106 s	-	-
10	yes	no	73,780	128 s	71,541	97%
11	no	no	54,600	96 s	-	-
12	yes	no	74,760	133 s	71,257	95%
13	no	no	67,000	118 s	-	-
14	yes	no	74,000	134 s	71,560	97%
15	no	yes	48,600	88 s	-	-
16	yes	yes	73,400	127 s	46,980	64%
17	no	yes	48,160	84 s	-	-
18	yes	yes	83,400	146 s	80,232	96%

data/runs.json is the machine-readable version of this table.

Contributing

This project is currently not accepting contributions.

Citation

If you use this dataset in your research, please cite the paper that describes the campaign, the testbed and the dApp trained on it:

@article{villa2026programmable,
    title = {{Programmable and GPU-Accelerated Edge Inference for Real-Time ISAC on NVIDIA Aerial Testbed}},
    author = {Villa, Davide and Belgiovine, Mauro and Hedberg, Nicholas and Polese, Michele and Dick, Chris and Melodia, Tommaso},
    journal = {arXiv:2512.06493 [cs.NI]},
    year = {2026}
}

Reference(s):

[1] D. Villa, M. Belgiovine, N. Hedberg, M. Polese, C. Dick, and T. Melodia, "Programmable and GPU-Accelerated Edge Inference for Real-Time ISAC on NVIDIA Aerial Testbed," arXiv:2512.06493 [cs.NI], 2026. arXiv PDF

[2] NVIDIA Corporation, "Aerial ISAC SRS I/Q." Raw uplink SRS I/Q with a synchronized video, recorded on the same testbed. Hugging Face

[3] NVIDIA Corporation, "Aerial Sample Apps." GitHub repo

[4] NVIDIA Corporation, "Aerial CUDA-Accelerated RAN." GitHub repo

[5] NVIDIA Corporation, "Aerial Testbed." NVIDIA Docs

Ethical Considerations:

The campaign was recorded on private NVIDIA premises with the participant's knowledge. The videos used to derive ground truth are not published, and neither are per-frame pixel coordinates: only floor positions in metres, a detector confidence and a frame index. Labels carry no identity, and the only identifiers in the release are a run number and the ephemeral 5G radio values the capture recorded, rnti and DMRS scrambling identity 51, both scoped to the connection. The marked area was kept clear during the background runs, but movement by other occupants further back in the room cannot be excluded and is disclosed above rather than edited out.

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. Developers should work with their internal developer teams to ensure this dataset meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

Please report quality, risk, security vulnerabilities or NVIDIA AI Concerns here.