Rad-R is the first publicly released radar dataset combining raw ADC captures from a 4-chip TI MMWCAS-RF-EVM cascade radar (12 Tx × 16 Rx = 192 virtual channels, 77 GHz) with physically induced hardware fault annotations at multiple severity levels, synchronized with companion sensors (BNO055 IMU, DHT22 board/ambient temperature, three GPS streams, Intel RealSense D435 camera with two co-recorded streams).

Captured fault classes

Class	Severity	Specification	File
Healthy	S0	Baseline (no induced fault)	`healthy.h5`
Misalignment	S1 (mild)	5° yaw via precision jig + inclinometer	`yaw0.h5`
Misalignment	S2 (severe)	10° yaw	`yaw2.h5`
Vibration	S1 (mild)	20 Hz drive (eccentric DC motor on radar mount)	`vib1.h5`
Vibration	S2 (severe)	40 Hz drive	`vib2.h5`
Blockage	S1 (mild)	30 % polycarbonate coverage on PP radome	`blockage1.h5`
Blockage	S2 (severe)	60 % polycarbonate coverage on PP radome	`blockage2.h5`
Rx degradation	S1 (mild)	6/16 Rx antennas copper-taped	`degrade0.h5`
Rx degradation	S2 (severe)	10/16 Rx antennas copper-taped	`degrade1.h5`

Severity is gated by measured physical quantities (BNO055 RMS, inclinometer angle, corner-reflector dB drop, single-tone calibration sweep) — not just the procedural setting.

Repository layout

synced_hdf5/                    # Sensor-sync metadata, one HDF5 per capture (~25 MB total)
  healthy.h5                    #   Per-radar-frame timestamps + IMU + DHT22 + GPS + camera paths
  yaw0.h5  yaw2.h5
  vib1.h5  vib2.h5
  blockage1.h5  blockage2.h5
  degrade0.h5  degrade1.h5
training_cache.h5               # Pre-processed RD maps + raw IQ subset (~2.4 GB)
                                #   1800 frames (200 sampled per capture)
                                #   rd_map: (1800, 224, 224) f32 dB
                                #   iq:     (1800, 64, 256, 16) complex64 (Tx=0 slice)
                                #   fault_label, severity_label, capture, frame_idx
LICENSE                         # CC BY 4.0
croissant.json                  # ML metadata (Croissant 1.0 standard)
radr.py                         # HuggingFace `datasets` loader

Note on the released IQ. training_cache.h5 ships the single-transmitter (Tx=0) IQ slice (1800, 64, 256, 16) to stay within the HF size quota. The full 192-virtual-channel TDM-MIMO cube used for the headline RadrNet results is reconstructed from the raw *_data.bin archive with the released build script (see below); the full raw archive is deposited on Zenodo at camera-ready.

How to load (training cache, recommended)

import h5py
import numpy as np

with h5py.File("training_cache.h5", "r") as f:
    rd      = np.asarray(f["rd_map"])          # (1800, 224, 224) f32
    iq      = np.asarray(f["iq"])              # (1800, 64, 256, 16) complex64 (Tx=0 slice)
    fault   = np.asarray(f["fault_label"])     # (1800,) int (0=healthy, 1=vib, 2=misalign, 3=block, 4=rxdeg)
    sev     = np.asarray(f["severity_label"])  # (1800,) int (0=healthy, 1=mild, 2=severe)
    capture = np.asarray(f["capture"]).astype(str)

How to load (synced metadata + raw .bin pipeline)

The 9 synced_hdf5/*.h5 files contain per-radar-frame wall-clock timestamps and synchronized companion-sensor streams, but NOT the raw IQ — that lives in TI mmWave Studio *_data.bin capture directories (~324 GB total, available via Zenodo deposit on request).

If you have the raw .bin directories alongside the synced HDF5s, the matlab-to-python pipeline (in the code repo) reads the cascaded ADC, applies range/Doppler FFT, and produces the same RD maps as in training_cache.h5 (and the full 192-channel MIMO cube used for RadrNet).

Hardware

Spec	Value
Platform	TI MMWCAS-RF-EVM (4× AWR1243)
Center frequency	77 GHz
Bandwidth	~2.5 GHz (256 ADC samples × 79 MHz/μs)
Tx × Rx	12 × 16 = 192 virtual channels
Range resolution	0.06 m
Max range	15.2 m
Frame rate	10 Hz (5 min × 9 captures = 45 min total recording)
Companion camera	Intel RealSense D435 (two streams: `camera2`, `camera4`)
Radome	1/16″ polypropylene panel (always installed)

Companion sensors (per-frame aligned)

Bosch BNO055 IMU — 3-axis accel + gyro + temperature, ~33 Hz
DHT22 — board + ambient temperature, ~1 Hz
u-blox NEO-M9N GPS — three decoded streams (legacy, bag-decoded, trips-DB-fused)
Intel RealSense D435 camera — two co-recorded streams (camera2 and camera4 stream IDs in the synced HDF5s); PNG paths + per-frame timestamps (the camera PNG frames themselves are not redistributed in this HF release)

Evaluation tasks

Within-clip fault classification (5-way: healthy + 4 fault classes)
Controlled cross-severity generalization (train one severity, test the other)
Severity classification (3-way: 0=healthy, 1=mild, 2=severe)
Chirp-wise anytime and few-shot label-budget protocols
(Future: cross-day / cross-device / cross-scene OOD splits — deferred to a follow-on capture session)

Baselines benchmarked (in the paper)

ResNet-18, ViT-Small, ConvNeXt-V2, DeiT-III, Swin-T, PVTv2, EfficientNetV2, and RadrNet (hierarchical Mamba on raw IQ, with capture-invariant and dual-stream variants).

License

CC BY 4.0. You may use, distribute, and adapt this dataset, including for commercial purposes, provided you give attribution to the authors.

Citation

@inproceedings{radr2027,
  title   = {Rad-R: A Real-World Raw-ADC Dataset and Benchmark for mmWave Automotive Radar Robustness},
  author  = {Anonymous},
  booktitle = {WACV Evaluations \& Datasets Track},
  year    = {2027}
}

Limitations

Single capture device, single day, overlapping scenes. Full OOD splits along device / day / scene axes deferred to a follow-on capture campaign.
Two severity levels per fault (mild / severe). The taxonomy supports up to 4 levels (S0–S3) but only S1–S2 are captured here.
Thermal stress and RF interference are part of the dataset specification but not yet captured. Planned for a future release.
Raw .bin not redistributed via this HF repo (300 GB exceeds quota). Available on request; will be deposited on Zenodo at camera-ready.

Downloads last month: 49

Total file size:

39.4 GB