Scene2Wave

Scene2Wave is a physics-consistent scene-to-wave data infrastructure for wireless world modeling. This repository is its first public reference instantiation: a synchronized CARLA traffic release compiled through RF Scene Compiler and simulated with Sionna RT. Version 1.1.1 contains 300 formally validated one-second episodes spanning 12 CARLA Towns, five motion states, four radio profiles, multi-view cameras, Birdview, LiDAR, Radar, IMU, GNSS, poses, path-level CIR, and CSI. The public release is a qualified dataset artifact; it does not imply that the complete generation toolchain or editable simulator assets are included.

v1.1.1 metadata compatibility patch

Version convention: v1.1.0 denotes the immutable 300-episode data payload; v1.1.1 denotes the current public release, comprising those byte-identical TAR archives plus additive kinematics and paper-audit metadata. Version 1.1.1 adds an explicit kinematics sidecar at metadata/kinematics/pose_velocity_v1_1_1.npz. The original velocity is exposed as tx_route_command_velocity_mps and rx_route_command_velocity_mps; the corresponding tx_pose_derived_velocity_mps and rx_pose_derived_velocity_mps arrays are centered finite differences over a total 10-ms span for interior states, with documented one-sided episode boundaries. They are offline geometric diagnostics, not temporally causal forecasting inputs; causal tasks should derive backward differences from positions available at or before the anchor. Per-frame closure errors and documented 0.5-m/s diagnostic consistency flags are included. A false flag warns against treating the legacy route-command velocity as a pose derivative; it does not invalidate pose, sensing, or RF data.

The mismatch was caused by the older waypoint-motion writer using a route orientation for velocity while position followed the interpolation-segment chord on some curves. Current generation code derives both from the same chord and gates pose--velocity closure before publication. The fix applies to future generated datasets; v1.1.1 does not silently rewrite historical sample data. See metadata/kinematics/README.md and patch_manifest.json for the exact algorithm, field meanings, thresholds, and audit statistics.

The Scene2Wave Project Content rightsholder is Pengcheng Laboratory. Based on the current release configuration and the authors' current understanding, the dataset uses official CARLA assets and the official Sionna software stack; no separately sourced scene assets are known to be included.

Authors

Mengfan Zheng, Liwen Jing, Tingting Yang, Li Sun, Yuxuan Shi, Ping Zhang, Leiyang Xu, and Jianjun Chen.

Dataset summary

Property	Value
Formal episodes	300
Towns	Town01, Town02, Town03, Town04, Town05, Town06, Town07, Town10HD, Town11, Town12, Town13, Town15; 25 episodes each
Motion states	static, 10, 20, 40, 60 km/h; 60 episodes each
Profiles	core 120, multipath 60, scattering mild 60, scattering medium 60
Carrier frequency	3.7 GHz
Bandwidth	15.36 MHz
Subcarriers	512 at 30 kHz spacing
CSI sampling rate	2 kHz
Episode duration	1 second
CARLA	0.9.16
Sionna	0.19.2

Propagation profiles within each motion state

Each motion state (static, 10kmh, 20kmh, 40kmh, or 60kmh) contains 60 episodes: five distinct CARLA runs from each of the 12 Towns. Within every Town × motion-state bucket, those five runs are allocated as two core and one each of multipath, scattering_mild, and scattering_medium. Thus each motion state contains 24/12/12/12 episodes respectively, and the complete release contains 120/60/60/60.

All four profiles enable LOS, specular reflection, diffraction, and edge diffraction. num_samples below is the Sionna ray-launching budget per source, not the number of dataset examples. max_depth is the maximum interaction depth used by the path solver.

Profile	Per Town × state	Per state	num_samples	max_depth	Scattering	Material profile / scale	scat_keep_prob
core	2	24	1,000,000	2	disabled	none / inactive	inactive
multipath	1	12	2,000,000	3	disabled	none / inactive	inactive
scattering_mild	1	12	2,000,000	3	enabled	urban_37ghz / 0.30	0.0002
scattering_medium	1	12	2,000,000	3	enabled	urban_37ghz / 0.60	0.0005

The profiles have the following intended meanings:

• core is the baseline propagation setting. Its shallower interaction depth and smaller ray budget emphasize dominant LOS, specular, and diffracted components while keeping the baseline representative and computationally bounded.
• multipath disables diffuse scattering like core, but doubles the ray budget and raises the interaction depth from 2 to 3. It is intended to retain more higher-order reflected/diffracted paths and represent a denser non-scattering multipath condition.
• scattering_mild uses the multipath solver budget and enables the urban_37ghz material-scattering model at scale 0.30. It introduces a moderate diffuse component for robustness and material-scattering sensitivity studies.
• scattering_medium raises the same material-scattering scale to 0.60 and increases scattering-path retention. It represents the more difficult diffuse-scattering condition included in the main training distribution.

material_scattering_scale is a relative multiplier applied to the configured urban material-scattering profile; it is not a percentage of total received power. Likewise, scat_keep_prob controls stochastic retention of candidate scattering paths for tractable simulation and is not the physical probability that a surface scatters.

These are independently generated main_training episodes, not four RF reruns of identical CARLA geometry. Profile comparisons should therefore use stratified or aggregate analysis across Town and motion state; a difference between two individual episodes must not be attributed solely to the profile.

Sensor suite and acquisition configuration

The CAV is a CARLA vehicle.tesla.model3. The roadside unit (RSU) is mounted 5 m above the sampled road surface. The following table records the formal release configuration; counts and rates are per episode unless stated otherwise.

Platform / modality	Configuration	Nominal output
CAV RGB cameras	4 views: front, rear, right, left; 1280 × 720; 110° horizontal FOV; camera height 2.4 m	20 Hz, 20 frames/view
CAV depth cameras	4 views co-located with the RGB cameras; 1280 × 720; 110° horizontal FOV	20 Hz, 20 frames/view
CAV LiDAR	128 channels; 120 m range; 360° horizontal FOV; −10° to +90° vertical FOV; mounted at 2.5 m	20 Hz, target 30,000 points/frame
CAV IMU	acceleration noise std. dev. 0.1 m/s²/axis; gyroscope noise std. dev. 0.002 rad/s/axis; gyro bias 0.001 rad/s/axis; seed 42	100 Hz, 100 records
CAV GNSS	latitude, longitude, and altitude noise std. dev. set to zero	10 Hz, 10 records
RSU RGB cameras	4 cardinal views: east, north, west, south; 1280 × 720; 110° horizontal FOV	20 Hz, 20 frames/view
RSU depth cameras	4 views co-located with the RSU RGB cameras; 1280 × 720; 110° horizontal FOV	20 Hz, 20 frames/view
RSU LiDAR	128 channels; 120 m range; 360° horizontal FOV; −50° to +50° vertical FOV	20 Hz, target 30,000 points/frame
RSU Radar	100 m range; 120° horizontal × 30° vertical FOV; yaw chosen to face the road	20 Hz, 10,000-point ray budget/frame; stored detections vary by scene
Birdview RGB	top-down camera at 130 m; 1280 × 1280; 90° FOV; yaw 0°	20 Hz, 20 frames

The 1-second recorded interval begins after a 0.3-second simulator warm-up. CARLA geometry/pose and CIR/CSI use a 2 kHz clock (2,000 samples), while the physical sensors retain the native rates above. Streams are therefore time-aligned but not rate-equal. Use each archive's carla/alignment_index.json: camera, LiDAR, Radar, and other sensor records are mapped to the nearest 2 kHz geometry/CSI frame. Do not assume that ordinal file indices from two modalities identify the same timestamp.

The LiDAR point counts are generation targets rather than a promise that every serialized cloud has exactly that size. Similarly, the Radar setting is a ray budget; CARLA writes only returned detections, so the number of stored Radar points depends on scene content. Per-frame poses and timestamps are stored with the sensor data. metadata/generation_config.yaml indexes the two generation batches, and their complete machine-readable policies are under metadata/generation_configs/.

Every formal episode passed CARLA trajectory and collision qualification before wireless ray tracing. The release contains no rejected candidates, visual QA panels, videos, logs, cache keys, or editable simulator assets.

Packaged layout

To keep the Hub repository usable, each episode is distributed as one deterministic, uncompressed tar archive instead of hundreds of thousands of small files:

Scene2Wave/
├── README.md
├── LICENSE
├── ATTRIBUTION.md
├── THIRD_PARTY_NOTICES.md
├── CITATION.cff
├── DATA_PROVENANCE.md
├── dataset_info.json
├── schema.md
├── checksums.sha256
├── metadata/
│   ├── samples.jsonl
│   ├── source_samples.jsonl
│   ├── kinematics/
│   │   ├── pose_velocity_v1_1_1.npz
│   │   ├── patch_manifest.json
│   │   ├── schema.json
│   │   └── README.md
│   ├── paper_reproducibility/
│   ├── json_schema.json
│   ├── generation_config.yaml
│   ├── generation_batches.json
│   ├── generation_configs/
│   └── raw_dataset_schema.md
└── data/main_training/<Town>/<state>/<profile>/<sample_id>.tar

Each tar has sample_metadata.json, carla/, and sionna/ at its member root. metadata/samples.jsonl adds archive_path, archive_bytes, and archive_sha256 to the public episode-index record. source_samples.jsonl preserves the original unpacked-tree index.

The Hugging Face Dataset Viewer is intentionally configured for the lightweight episode_index/episodes split backed by metadata/samples.jsonl; it browses episode metadata without attempting to decode the large deterministic TAR archives. Use the Scene2Wave Space for precomputed multimodal visualization and the archive-reading example below for payload access.

Load one episode without extracting the full dataset

import io
import json
import tarfile
from pathlib import Path

import numpy as np

root = Path("Scene2Wave")
records = [
    json.loads(line)
    for line in (root / "metadata/samples.jsonl").read_text().splitlines()
    if line.strip()
]
record = records[0]

with tarfile.open(root / record["archive_path"], mode="r:") as archive:
    member = archive.extractfile("./sample_metadata.json")
    metadata = json.load(member)

    npz_name = next(
        name for name in archive.getnames()
        if name.startswith("./sionna/") and name.endswith("_paths.npz")
    )
    payload = archive.extractfile(npz_name).read()
    with np.load(io.BytesIO(payload), allow_pickle=False) as arrays:
        print(record["sample_id"], arrays.files)

print(metadata["radio"]["carrier_frequency_hz"])

For repeated access, extract only the required episode:

mkdir sample
tar -xf data/main_training/Town01/10kmh/core/Req_Town01_dynamic_10kmh_t001_rfmain.tar -C sample

See schema.md for archive handling and metadata/raw_dataset_schema.md for the complete sensor, alignment, radar, and wireless field reference.

Integrity

checksums.sha256 covers every episode archive and all release metadata:

sha256sum --check checksums.sha256

Intended uses

• multimodal wireless-channel representation learning;
• CIR/CSI estimation and prediction;
• perception-to-channel feature extraction;
• synchronization, fusion, and robustness research;
• controlled comparisons across Town, speed, and propagation profile.

Limitations

• This is simulated data and does not reproduce every real sensor, material, traffic, weather, hardware, calibration, or regulatory condition.
• Only the 12 documented CARLA Towns and RF configurations are represented.
• Dataset splits must avoid leakage between closely related scenarios when evaluating generalization.
• Empty or outage links require explicit handling even though the formal release records valid CIR for all indexed episodes.
• Users remain responsible for validating conclusions against real-world data.

License and third-party materials

This repository uses a layered license. To the extent Pengcheng Laboratory owns the applicable rights, Scene2Wave Project Content is licensed under CC BY 4.0. CARLA-derived elements and other third-party materials retain their respective terms. Read LICENSE, ATTRIBUTION.md, and THIRD_PARTY_NOTICES.md; the Hugging Face metadata therefore uses license: other rather than applying one license to every layer.

Citation

Until a dataset paper or DOI is assigned, cite this release using the metadata in CITATION.cff and include the repository URL and version. Preserve the author order listed above.