Skip to content

RF-MatID: Dataset and Benchmark for Radio Frequency Material Identification

Conference: ICLR 2026 arXiv: 2601.20377 Code: Available (project page) Area: AI Safety / Embodied AI / RF Sensing Keywords: RF sensing, material identification, UWB-mmWave, dataset benchmark, embodied AI

TL;DR

This paper introduces RF-MatID, the first open-source large-scale RF material identification dataset with wide frequency coverage (4–43.5 GHz) and diverse geometric perturbations, comprising 16 fine-grained material categories (5 superclasses) and 142K samples. A comprehensive benchmark is established across 9 deep learning models, 5 frequency protocols, and 7 data split settings.

Background & Motivation

Background: Material identification is a fundamental capability for embodied AI, currently dominated by optical sensors (cameras, hyperspectral imaging). RF-based methods exploit electromagnetic wave–material interactions to reveal intrinsic material properties (permittivity, conductivity, etc.), operating independently of illumination conditions and visual appearance.

Limitations of Prior Work: (1) All existing RF material datasets are proprietary, hindering fair algorithmic comparison; (2) COTS sensor frequency bands are narrow and fragmented (e.g., 77–81 GHz only), precluding systematic cross-band evaluation; (3) systematic evaluation of geometric perturbations (angle and distance variation) is absent, leaving real-world deployment robustness uncertain.

Key Challenge: Despite theoretical advantages of RF methods (strong penetration, illumination independence), the lack of research infrastructure (datasets and benchmarks) severely impedes the development and evaluation of learning-based approaches.

Goal: Construct the first open-source, wide-band, geometrically diverse RF material identification dataset and establish a complete benchmarking framework.

Key Insight: A custom UWB-mmWave sensing platform with continuous frequency coverage from 4 to 43.5 GHz is developed to systematically collect RF responses of 16 materials across varying distances (200–2000 mm) and angles (0–10°).

Core Idea: Enable standardized learning-based research in RF material identification through the first open-source wide-band RF dataset and systematic benchmark.

Method

Overall Architecture

RF-MatID is constructed at three levels: (1) Data collection — a custom UWB-mmWave sensing platform systematically acquires frequency-domain responses from 16 material classes over a distance–angle grid; (2) Data processing — dual-domain representation (frequency and time domains) with complex whitening normalization; (3) Benchmark evaluation — 5 frequency protocols × 7 data splits × 9 models.

Key Designs

  1. Sensing Platform and Data Acquisition:

    • Function: Wide-band RF signal acquisition covering 4–43.5 GHz.
    • Mechanism: A DRH40 double-ridge horn antenna paired with an MS46131A vector network analyzer collects complex responses \(H(f_i) = I(f_i) + jQ(f_i)\) at 2048 frequency bins per sample. Each material is measured on a distance grid of 200–2000 mm (50 mm steps) × angle grid of 0–10° (1° steps), yielding 142K samples in total.
    • Design Motivation: The 39.5 GHz bandwidth far exceeds the previous maximum of 4 GHz, simultaneously capturing centimeter-wave (3–30 GHz) penetration information and millimeter-wave Q-band (30–50 GHz) surface sensitivity.

  2. Dual-Domain Data Representation:

    • Function: Generate paired time-domain representations for each frequency-domain sample.
    • Mechanism: Frequency-domain data is represented as two-channel real tensors (real and imaginary parts); time-domain data is obtained via IFFT to produce signals of length 10240. Complex whitening is applied in the frequency domain to preserve phase relationships; standard normalization is used in the time domain.
    • Design Motivation: Experiments confirm that two-channel real representation outperforms complex-valued networks; frequency-domain captures frequency-selective attenuation while time-domain captures propagation delay.

  3. Frequency Protocols:

    • Function: Define 5 frequency allocation schemes to evaluate recognition capability across different bands.
    • Mechanism: P1 full band 4–43.5 GHz; P2 millimeter-wave 30–43.5 GHz; P3 centimeter-wave 4–30 GHz; P4 legally permitted commercial bands in the United States; P5 legally permitted bands in China.
    • Design Motivation: This is the first benchmark to incorporate regulatory constraints on frequency usage, enabling results to directly inform compliant system design.

  4. Seven Data Split Settings:

    • Function: Systematically evaluate model generalization across diverse deployment scenarios.
    • Mechanism: S1 standard random split (IID); S2 cross-distance OOD (mod1–3 covering different distance subsets); S3 cross-angle OOD.
    • Design Motivation: S1 assesses basic capability; S2/S3 assess robustness to distribution shift caused by sensor position variation in real-world deployment.

Material Taxonomy

5 superclasses → 16 fine-grained categories: Brick (over-fired clay brick, lightweight porous brick, volcanic brick); Glass (transparent acrylic, tempered glass, white opaque acrylic); Synthetic (melamine-faced board, mineral fiber board, PVC board); Wood (cedar sleeper, lauan plywood, red oak plywood); Stone (permeable paving stone, engineered stone, granite, concrete).

Key Experimental Results

Main Results (Protocol 1, Full Band 4–43.5 GHz)

Model S1 (IID) S2-mod1 (Cross-distance) S3-mod1 (Cross-angle) Notes
Baseline (proposed) 99.57 86.62 98.89 Simple CNN
LSTM-ResNet 99.84 97.12 99.69 Best IID
ConvNeXt 99.51 79.10 98.85 CV model
AirTac 96.81 91.36 98.12 RF-specific
Material-ID 99.28 95.67 97.63 RF-specific

Cross-Domain Robustness (S2 OOD, Protocol 1)

Model S2-mod1 S2-mod2 S2-mod3 Notes
LSTM-ResNet 97.12 49.95 71.00 Large drop at mod2
AirTac 91.36 86.95 65.41 Most robust across distances
ConvNeXt 79.10 64.19 63.52 Poor cross-domain generalization

Key Findings

  • IID performance is near-saturated: Most models exceed 99% under S1, offering little discriminability; with sufficient data, RF material identification is not inherently difficult.
  • Cross-distance domain shift is the primary challenge: Accuracy under S2-mod2 drops sharply to 50–87% across models, indicating that distance-induced signal attenuation poses the greatest deployment challenge.
  • AirTac is most robust under OOD settings: Although not the top IID performer, it exhibits the smallest OOD degradation, suggesting that RF-specific architectural design benefits robustness.
  • Frequency domain vs. time domain: Frequency-domain two-channel representation outperforms time-domain representation and complex-valued network processing.
  • Compliant frequency bands remain viable: Performance under P4/P5 (legally permitted bands) is lower than the full band but remains practically usable, validating real-world deployment feasibility.

Highlights & Insights

  • Landmark significance as the first open-source RF material dataset: Analogous to ImageNet's impact on visual recognition research, RF-MatID has the potential to standardize RF sensing research. The open-source policy constitutes the most significant contribution.
  • Frequency protocol design incorporates regulatory constraints: This is the first benchmark to integrate legal compliance into its design, directly facilitating the transition from research to deployment.
  • Systematic geometric perturbation: The grid-based distance and angle acquisition methodology can be adopted by other sensing modalities (e.g., LiDAR, ultrasound).

Limitations & Future Work

  • Limited material variety: 16 categories remain insufficient; real-world environments encompass far more materials, including liquids, textiles, and metals.
  • Single sensing platform: All data originate from one hardware setup; cross-device generalization is unknown.
  • Controlled indoor environment: Multipath interference, occlusion, and other real-world factors are not considered.
  • Narrow angle range: Only 0–10° is covered, whereas robotic manipulation may require 0–90°.
  • Absence of multimodal benchmark: As an embodied AI dataset, no paired visual or tactile data are provided for multimodal fusion research.
  • vs. VNA-based datasets (he2022accurate, shanbhag2023contactless): Broader frequency coverage (39.5 GHz vs. 4 GHz) and openly available.
  • vs. Wi-Fi/RFID datasets: Higher signal quality (coherent transceiver) but requires dedicated hardware.
  • Provides foundational data resources for material perception research in embodied AI, and can serve as an RF-branch baseline.

Rating

  • Novelty: ⭐⭐⭐⭐ First open-source wide-band RF material identification dataset, filling a critical gap.
  • Experimental Thoroughness: ⭐⭐⭐⭐ 9 models × 5 protocols × 7 splits, comprehensive coverage.
  • Writing Quality: ⭐⭐⭐⭐ Well-structured with thorough background exposition.
  • Value: ⭐⭐⭐⭐ Dataset contribution provides lasting impact to the field.