ACM Artifacts AvailableACM Artifacts Evaluated FunctionalACM Artifacts Evaluated ReusableACM Results Reproduced
ACMSIGMOBILE
ACM MobiCom 2026

RF-LEGO: Modularized Signal Processing–Deep Learning Co-Design for RF Sensing via Deep Unrolling

Luca Jiang-Tao Yu, Chenshu Wu (The University of Hong Kong)

Motivation & Design Principles

Classical signal processing

Explicit operators.
Meaningful interfaces.

  • Inspectable frequency and angle spectra.
  • Blocks cascade into many sensing pipelines.
  • Handcrafted parameters require manual tuning.

Task-specific deep learning

Learned adaptation.
Entangled representations.

  • Adapts to complex measurements.
  • Often optimized end to end for one task.
  • Internal features can be hard to reuse.
RF-LEGO combines signal processing and deep learning via deep unrolling, preserving modularity, cascadability and interpretability.

Modularity

Replace one classical operator.

Cascadability

Connect learned and classical blocks.

Interpretability

Retain meaningful operators and outputs.

Primer on Deep Unrolling

Deep unrolling: classical signal-processing blocks become trainable operators, or solver iterations become trainable layers.

Operator unrolling learns a block’s parameters. Iterative unrolling maps solver iterations to layers with learnable update rules.

Three Trainable RF Modules

RF-LEGO Frequency Transformation

RF-LEGO Frequency Transformation: Bluestein's algorithm with fixed outer chirp multiplications and a learnable complex convolution filter.

Complex samples → frequency spectrum. Preserve Bluestein’s chirp multiplications; learn the complex convolution filter.

RF-LEGO Beamformer

RF-LEGO Beamformer: unrolled LASSO/ADMM iterations with trainable update coefficients, shrinkage thresholds and gates.

Array samples → angular spectrum. Retain the steering model; learn ADMM update coefficients, shrinkage levels and gates.

RF-LEGO Detector

RF-LEGO Detector: CFAR selection and testing with a trainable state-space model for noise and clutter estimation.

Spectrum → target detections. Preserve CFAR selection and testing; learn noise and clutter dynamics with a state-space model.

Case Study

RF-LEGO frequency transforms, beamformer and detector substitute classical front-end blocks in trajectory tracking, vital sign monitoring and human activity recognition.

Reuse the RF modules across three sensing pipelines.
Keep downstream algorithms and classifier architecture.

Trajectory Tracking

≈0.3 m median ATE

65 trajectories · 13 categories · 4 m × 4 m room

Figure 15(a–b): 4×4 m tracking room with mmWave radar, OptiTrack, and a head-mounted marker.
mmWave radar + OptiTrack ground truth
Measured RF-LEGO letter trajectories with SP and OptiTrack ground truth.

Vital Sign Monitoring

52.7% / 56.8% lower MAE

Infant simulator (20–100 BPM) and adults (6–20 BPM), versus SP

Figure 15(c–e): infant simulator, single-adult, and two-adult breathing experiments.
Breathing-rate MAE, reduced by 52.7% for the infant simulator and 56.8% for human adults, versus signal processing.
Infant simulator / adult breathing-rate MAE
CDF of breathing-rate MAE in single- and two-person adult trials, comparing RF-LEGO with signal processing.
Adult single- / two-person MAE CDF

Human Activity
Recognition

93.92% / 84.15%

Mean 7-way / 13-way
accuracy; SP:
91.88% / 79.31%
MCD-Gesture

7-way and 13-way recognition accuracy across in-domain, cross-user, cross-environment and cross-location settings.