The dataset bridges perception and wireless-channel modeling by synchronizing 2 kHz geometry/CSI streams with modality-native sensor records (cameras, LiDAR, radar, IMU, GNSS). That alignment makes it practical to study how visual and geometric cues correlate with short-timescale channel dynamics and to train models that predict or condition channel estimates on perception inputs.
What Sets It Apart
- Tight time alignment: CARLA geometry and Sionna-derived CIR/CSI use a 2 kHz clock and per-frame mapping so sensor records can be mapped to high-rate channel frames for cross-modal learning.
- Controlled propagation variety: four propagation profiles (core, multipath, scattering_mild, scattering_medium) let researchers test sensitivity to scattering, ray-budget, and interaction depth while preserving consistent scenario structure.
- Multi-sensor urban driving stack: multi-view RGB/depth, high-channel-count LiDAR, roadside radar, IMU and GNSS are packaged per-sample as deterministic tar archives for selective extraction and integrity checking.
- Reproducible simulated provenance: generated with CARLA 0.9.16 and Sionna RT, and distributed with checksums and generation metadata to support deterministic experiments.
Who It's For and Tradeoffs
Great fit if you need a small, fully documented simulated corpus to prototype perception-to-channel feature extraction, CIR/CSI prediction, synchronization/fusion methods, or robustness studies across speed and propagation profiles. Look elsewhere if you require large-scale real-world channel measurements, broader geographic diversity, long-duration traces, or production-ready wireless datasets—the release is intentionally compact (100 validated 1‑s samples across four CARLA towns) and remains a simulated benchmark with layered third-party licensing.
Practical notes
The dataset emphasizes stratified comparisons (Town × motion-state × profile) rather than treating individual samples as identically configured RF reruns. Users should account for simulation limits, avoid data-split leakage across related runs, and validate model conclusions on real measurements before deployment.
