This is a demo release with 20 clips. The full dataset (~5,000 clips) will be available upon paper acceptance at NeurIPS 2026 Evaluations & Datasets Track.

Overview

Rad-R is the first mmWave radar dataset combining:

Raw ADC captures from a TI MMWCAS-RF-EVM 77 GHz cascaded radar (12 TX × 16 Rx = 192 virtual channels)
Controlled hardware fault annotations across 6 fault types at 3 severity levels
Synchronized companion sensors: IMU (BNO055), temperature (DS18B20), GPS (NEO-M9N), camera
4 out-of-distribution evaluation splits: device, day, scene, severity
RadRBench: standardized benchmarking system with 5 evaluation tasks

Fault Taxonomy

Fault Type	S0 (Healthy)	S1 (Mild)	S2 (Moderate)	S3 (Severe)
healthy	✓ baseline	—	—	—
vibration	—	<2 m/s² RMS	2-5 m/s²	>5 m/s²
misalignment	—	2-4° yaw	5-8°	>9°
blockage	—	<30% coverage	30-70%	>70%
rx_degradation	—	<3 dB drop	3-8 dB	>8 dB
thermal_stress	—	ΔT 5-10°C	10-20°C	>20°C
rf_interference	—	Far/off-axis	Partial align	Near boresight

Data Format

Metadata (viewable above)

Each row in the dataset viewer shows per-clip metadata including fault type, severity, sensor readings (IMU, temperature, GPS), device ID, and scene ID.

HDF5 Files (downloadable)

Each clip in the HDF5 files contains:

Field	Shape	Type	Description
`iq`	(1, 64, 256, 16)	complex64	Raw IQ tensor (frames, chirps, samples, rx)
`rd`	(224, 224)	float32	Range-Doppler map (log-normalized)
`ra`	(224, 224)	float32	Range-Azimuth map (beamformed)
`microdoppler`	(224, 224)	float32	Micro-Doppler spectrogram

Plus per-clip attributes: fault_type, severity, device_id, scene_id, session_id, run_id, sha256, and all sensor context values.

File Structure

radr.h5        — Full dataset (per-clip HDF5 groups, gzip compressed)
train.h5       — Pre-split flat arrays for fast training (14 clips)
val.h5         — Validation split (3 clips)
test.h5        — Test split (3 clips)
index.json     — Lightweight metadata sidecar
splits.json    — Split definitions
data/          — CSV metadata for HuggingFace viewer

Quick Start

Option 1: radr Python package

pip install radr

from radr import RadRDataset, RadRBench, build_model

# Load dataset
ds = RadRDataset("path/to/data", representation="rd")
sample = ds[0]  # {"input": (1, 224, 224), "fault_label": int, "severity_label": int}

# Build and benchmark a model
model = build_model("resnet18")
results = RadRBench.evaluate(model, "path/to/data", representation="rd")
RadRBench.print_summary(results)

Option 2: HuggingFace datasets

from datasets import load_dataset
ds = load_dataset("gtaxcenter/radr", split="train")

Available Models (via `build_model()`)

Model	Input	Type
`resnet18`	RD/RA/μD maps	Traditional CNN
`vit_small`	RD/RA/μD maps	Vision Transformer
`iqcnn`	Raw IQ	1D CNN
`radrnet`	Raw IQ	Hierarchical Mamba (ours)
`clip`	RD maps (RGB)	CLIP ViT-B/32 fine-tune
`openclip`	RD maps (RGB)	OpenCLIP ViT-L/14 LoRA
`llava_lite`	RD maps	LLaVA vision encoder
`qwen2vl`	RD maps	Qwen2-VL vision encoder
`paligemma`	RD maps	PaliGemma SigLIP
`resnet18_fused`	RD+RA+μD	Multi-rep gated fusion

Radar Configuration

Parameter	Value
Hardware	TI MMWCAS-RF-EVM (AWR1243 4-chip cascade)
Frequency	77 GHz
Bandwidth	2.53 GHz
ADC samples	256 per chirp
Loops	64 (Doppler dimension)
TDM-MIMO	12 TX × 16 Rx = 192 virtual channels
Range	0–15.2 m, resolution 0.059 m
Velocity	±21.3 m/s

Sensor Suite

Sensor	Model	Sample Rate	Measurements
IMU	Bosch BNO055	100 Hz	3-axis accel, gyro, magnetometer, euler angles
Temperature	2× DS18B20	1 Hz	Board temp (PCB-DSP gap), ambient
GPS	u-blox NEO-M9N	10 Hz	Lat, lon, altitude, speed
Camera	RPi Camera v3	10 fps	1280×720 RGB

Citation

@inproceedings{mallick2026radr,
  title={Rad-R: A Real-World Raw-ADC Dataset and Benchmark for mmWave Radar Robustness},
  author={Mallick, Mainak and Yim, Junghwan and Choi, Seung-Kyum},
  booktitle={NeurIPS Evaluations \& Datasets Track},
  year={2026}
}

License

CC BY-NC-SA 4.0

Rad-R: A Real-World Raw-ADC Dataset and Benchmark for mmWave Radar Robustness