RF-LEGO: Modular RF Design Paradigm

• RF-LEGO is a multifaceted concept uniting modular RF connectors for UAV arrays, batteryless RFID-based reconfigurable surfaces, and plug-and-play deep-unrolling for signal processing.
• It emphasizes modularity through self-aligning mechanical docking, selective passive element aggregation, and structured, trainable SP-DL modules.
• Empirical results show enhanced phase coherence, improved bandwidth and detection rates, and scalable performance across UAV, RFID, and sensing applications.

“RF-LEGO” denotes three distinct research constructs in recent wireless-systems literature: a LEGO-inspired RF connector for UAV swarm-based phased arrays, a wireless, batteryless, RF-powered reconfigurable surface built from commodity RFID tags, and a modularized signal processing-deep learning co-design framework for RF sensing via deep unrolling (Debnath et al., 30 Jul 2025, Vardakis et al., 2021, Yu et al., 11 Apr 2026). Taken together, this suggests that RF-LEGO is presently a paper-specific designation rather than a single universally standardized term. The shared motif is modular composition: mechanical self-alignment and rigid docking in airborne arrays, selective aggregation of passive backscatter elements in reconfigurable surfaces, and plug-and-play “LEGO bricks” in structure-aligned deep-unrolled sensing pipelines.

1. Terminological Scope and Research Context

In the cited literature, RF-LEGO refers to three different objects with different problem settings, hardware assumptions, and performance criteria. One line of work places the RF-LEGO connector at the heart of a UAV swarm-based phased-array concept. A second realizes a passive reconfigurable surface by tiling ultra-low-cost Gen2 RFID tags into a two-dimensional array. A third proposes RF-LEGO as a modular co-design framework that transforms interpretable SP algorithms into trainable, physics-grounded DL modules through deep unrolling (Debnath et al., 30 Jul 2025, Vardakis et al., 2021, Yu et al., 11 Apr 2026).

Usage	Definition	Source
RF-LEGO connector	LEGO-inspired RF connector for UAV swarm-based phased arrays	(Debnath et al., 30 Jul 2025)
RF-LEGO surface	Wireless, batteryless, RF-powered reconfigurable surface built from commodity RFID tags	(Vardakis et al., 2021)
RF-LEGO framework	Modularized SP-DL co-design for RF sensing via deep unrolling	(Yu et al., 11 Apr 2026)

The terminological overlap can invite a misconception that these works describe one evolving platform. They do not. The first is an RF interconnect and airborne array architecture, the second is a batteryless reconfigurable surface, and the third is a deep-unrolling methodology for wireless sensing. Their only commonality, as stated in the papers, is a strong emphasis on modularity, reusability, and structured composition.

2. RF-LEGO as a UAV Swarm RF Connector and Phased-Array Enabler

In the UAV-swarm setting, the LEGO-inspired RF-LEGO connector is a compact, non-threaded, hands-free docking interface designed to guarantee precise alignment, sub-millimeter inter-element spacing accuracy, and a continuous, low-loss RF path from DC well into the UHF band—even during mid-flight docking maneuvers. Each connector half consists of a rectangular slotted patch of silver ($$\sigma\approx1.27\times10^{-4}\,\Omega\cdot\mathrm{m}$$) printed on a $1.6\,\mathrm{mm}$-thick FR4 substrate ($\epsilon_r=4.4$, $\tan\delta=0.02$), with a continuous copper ground plane underneath; one end of the patch is shorted to ground via a plated-through via, the other is fed by an SMA bulkhead connector. A $2\,\mathrm{mm}$-thick neodymium magnet is bonded to the rear face of each FR4 board, while 3D-printed LEGO-inspired alignment brackets in a convex/concave stud-and-socket arrangement mate with sub-millimeter tolerance and eliminate vertical and lateral misalignment. Carbon-fiber docking rods ($\sigma\approx1\times10^4\,\mathrm{S/m}$) attach to the magnets and fix the inter-element spacing $d_{\mathrm{ele}}$ to within $\pm0.5\,\mathrm{mm}$ (Debnath et al., 30 Jul 2025).

The docking mechanism is explicitly non-threaded and hands-free. Permanent magnets in each half cause the connector halves to “snap” into place when the UAVs approach, and the convex/concave bracket system passively corrects residual offset. Measurement of $S_{12}$ versus misalignment shows that even a $6\,\mathrm{mm}$ vertical offset degrades insertion loss to below $1.6\,\mathrm{mm}$0; with the LEGO brackets engaged, $1.6\,\mathrm{mm}$1 and $1.6\,\mathrm{mm}$2 remains within $1.6\,\mathrm{mm}$3 of the aligned value. The mechanical baseline imposed by the carbon-fiber rods locks the UAVs into a rigid, phase-coherent formation with $1.6\,\mathrm{mm}$4 fixed to better than $1.6\,\mathrm{mm}$5. In the $1.6\,\mathrm{mm}$6 demonstrator, $1.6\,\mathrm{mm}$7; larger arrays use multiple rods to maintain $1.6\,\mathrm{mm}$8.

The connector was refined through a multi-stage full-wave HFSS optimization that balanced bandwidth, insertion loss, and form factor. Stage 0 used an open-patch baseline with $1.6\,\mathrm{mm}$9 and air gap $\epsilon_r=4.4$0, yielding $\epsilon_r=4.4$1 and $\epsilon_r=4.4$2 bandwidth $\epsilon_r=4.4$3. Stage 1 closed the gap, giving $\epsilon_r=4.4$4. Stage 2 used a narrow patch with $\epsilon_r=4.4$5, $\epsilon_r=4.4$6, giving $\epsilon_r=4.4$7, $\epsilon_r=4.4$8. Stage 3 added a short-circuit via and reduced the effective length to $\epsilon_r=4.4$9, giving $\tan\delta=0.02$0, $\tan\delta=0.02$1. Stage 4 introduced a narrow capacitive slot, producing $\tan\delta=0.02$2, $\tan\delta=0.02$3, and bandwidth from DC to $\tan\delta=0.02$4 and beyond. The optimization was organized around

$\tan\delta=0.02$5

where $\tan\delta=0.02$6 is the dielectric path length within each patch and $\tan\delta=0.02$7 the air gap, together with

$\tan\delta=0.02$8

and the impedance model

$\tan\delta=0.02$9

Measured RF performance at $2\,\mathrm{mm}$0 included $2\,\mathrm{mm}$1 versus simulated $2\,\mathrm{mm}$2, and $2\,\mathrm{mm}$3, implying $2\,\mathrm{mm}$4 and

$2\,\mathrm{mm}$5

The $2\,\mathrm{mm}$6 bandwidth spans DC up to $2\,\mathrm{mm}$7 for the $2\,\mathrm{mm}$8 patch, and can be extended to $2\,\mathrm{mm}$9 with a $\sigma\approx1\times10^4\,\mathrm{S/m}$0 patch or $\sigma\approx1\times10^4\,\mathrm{S/m}$1 with a $\sigma\approx1\times10^4\,\mathrm{S/m}$2 patch by scaling slot length.

Once mechanically latched, the RF-LEGO connectors tie together the feed network so that all element excitations share a single oscillator reference. This eliminates independent LO drift and phase-lock overhead: mechanically linking $\sigma\approx1\times10^4\,\mathrm{S/m}$3 UAVs effectively creates one monolithic $\sigma\approx1\times10^4\,\mathrm{S/m}$4-element array. The array factor is written as

$\sigma\approx1\times10^4\,\mathrm{S/m}$5

where $\sigma\approx1\times10^4\,\mathrm{S/m}$6 are the element positions fixed by docking rods, $\sigma\approx1\times10^4\,\mathrm{S/m}$7 the programmed phase shifts, and $\sigma\approx1\times10^4\,\mathrm{S/m}$8. To avoid grating lobes, the architecture enforces

$\sigma\approx1\times10^4\,\mathrm{S/m}$9

for all steering angles.

Experimental validation was reported in both stationary and in-flight settings. In an anechoic chamber, two-element arrays mounted on low-dielectric foam produced beam patterns at $d_{\mathrm{ele}}$0 that closely match simulation within $d_{\mathrm{ele}}$1, and each Yagi element had measured $d_{\mathrm{ele}}$2 around $d_{\mathrm{ele}}$3. In flight, two identical UAVs approached and docked in mid-air at $d_{\mathrm{ele}}$4, then undocked at $d_{\mathrm{ele}}$5; continuous RF path was verified by commanding phase shifts and observing stable received power at three ground receivers at $d_{\mathrm{ele}}$6, with peak power fluctuations below $d_{\mathrm{ele}}$7 over $d_{\mathrm{ele}}$8. Scalability was quantified by array gain increasing from $d_{\mathrm{ele}}$9 for $\pm0.5\,\mathrm{mm}$0 to $\pm0.5\,\mathrm{mm}$1 for $\pm0.5\,\mathrm{mm}$2 at steer $\pm0.5\,\mathrm{mm}$3, and by frequency scaling through reduced patch length.

3. RF-LEGO as a Wireless, Batteryless, RF-Powered Reconfigurable Surface

In the RFID-based work, RF-LEGO realizes a passive reconfigurable surface by tiling ultra-low-cost Gen2 RFID tags into a two-dimensional array. The proof-of-concept uses $\pm0.5\,\mathrm{mm}$4 tags on a $\pm0.5\,\mathrm{mm}$5 rectangular grid with inter-tag spacing $\pm0.5\,\mathrm{mm}$6, $\pm0.5\,\mathrm{mm}$7, approximately $\pm0.5\,\mathrm{mm}$8 at $\pm0.5\,\mathrm{mm}$9, to keep mutual coupling negligible. The tags are standard batteryless EPC-Gen2 inlays (Zebra Z-Perform 1500T) whose antenna is a small dipole-loop printed on PET and matched at $S_{12}$0 by a two-element L-network. In steady state, the antenna input impedance is $S_{12}$1; through tuning components, two discrete load impedances $S_{12}$2 realize two distinct reflection coefficients $S_{12}$3. The antenna’s structural scattering parameter is measured as $S_{12}$4. Control and power are provided by a modified SDR-based RFID reader using a USRP N200+RFX900 daughterboard: one SDR transmits a continuous-wave carrier at $S_{12}$5 to power and query the tags, and another SDR at $S_{12}$6 receives the assisted source-destination link (Vardakis et al., 2021).

The control mechanism inverts framed-slotted Aloha by forcing controlled collisions to synthesize arbitrary subgroups of tags into backscatter reflectors. The sequence is explicit: the reader sends CW so all tags harvest energy and go Ready; a series of Select commands with mask/Action parameters asserts each tag’s SL flag individually, building an active set $S_{12}$7; a single-slot Query then causes all tags with $S_{12}$8 to transmit their 6-bit Preamble and 16-bit RN16 simultaneously, so that the destination observes a coherent sum with envelope proportional to

$S_{12}$9

where $6\,\mathrm{mm}$0. To move to a new configuration, the reader issues a deassert-all Select and repeats the procedure. With $6\,\mathrm{mm}$1 active tags, each configuration takes

$6\,\mathrm{mm}$2

which in the reported setup with $6\,\mathrm{mm}$3 yields $6\,\mathrm{mm}$4–$6\,\mathrm{mm}$5.

The theoretical model writes the baseband received signal as

$6\,\mathrm{mm}$6

and the design objective is to maximize instantaneous power

$6\,\mathrm{mm}$7

For the $6\,\mathrm{mm}$8-load-per-tag optimization problem, the paper gives the closed-form structure

$6\,\mathrm{mm}$9

In the $1.6\,\mathrm{mm}$00 case, transitions occur only when $1.6\,\mathrm{mm}$01 passes

$1.6\,\mathrm{mm}$02

and sorting the $1.6\,\mathrm{mm}$03 boundary angles takes $1.6\,\mathrm{mm}$04, yielding global optimum complexity $1.6\,\mathrm{mm}$05 rather than $1.6\,\mathrm{mm}$06. For general $1.6\,\mathrm{mm}$07, the paper states an $1.6\,\mathrm{mm}$08 bound.

Experimentally, when configured for constructive beamforming, the received power at $1.6\,\mathrm{mm}$09 rose by up to $1.6\,\mathrm{mm}$10 above the direct-link alone. For a different geometry and exhaustive exploration of $1.6\,\mathrm{mm}$11 active tags, the measured peak gain reached $1.6\,\mathrm{mm}$12. Average-gain simulations over $1.6\,\mathrm{mm}$13 Monte Carlo channel realizations at $1.6\,\mathrm{mm}$14 show $1.6\,\mathrm{mm}$15–$1.6\,\mathrm{mm}$16 improvement for $1.6\,\mathrm{mm}$17 tags and $1.6\,\mathrm{mm}$18 loads; moving to $1.6\,\mathrm{mm}$19 via varactor buys an extra $1.6\,\mathrm{mm}$20. The paper also states that gain saturates as $1.6\,\mathrm{mm}$21 grows, even at perfect CSI, because the end-to-end backscatter SNR remains low with $1.6\,\mathrm{mm}$22 for each element. Imperfect channel estimates, modeled with MMSE and pilot fraction $1.6\,\mathrm{mm}$23 of the coherent block, degrade $1.6\,\mathrm{mm}$24 by several dB when $1.6\,\mathrm{mm}$25, making estimation overhead critical for large arrays.

A common misconception is to read this work as a conventional high-gain RIS result. The paper explicitly argues otherwise: even with perfect channel estimation, the weak nature of backscattered links limits the performance gains, even for large number of surface elements. This makes RF-LEGO here a low-cost and batteryless reconfigurable surface, but not an unrestricted passive beamforming mechanism with linear gain growth.

4. RF-LEGO as Modularized SP-DL Co-Design via Deep Unrolling

In the sensing framework, RF-LEGO is defined as a modular co-design framework that transforms interpretable SP algorithms into trainable, physics-grounded DL modules through deep unrolling. It is positioned between pure SP and end-to-end DL: classical methods such as FFT, beamformers, and CFAR detectors are interpretable and modular but brittle in low-SNR or multipath environments, whereas purely end-to-end DL models are task-specific and monolithic and lack stage-wise interpretability. RF-LEGO introduces three deep-unrolled modules for critical RF sensing tasks—frequency transform, spatial angle estimation, and signal detection—while preserving core processing structures and mathematical operators. The paper identifies three claimed properties: modularity, cascadability, and structure-aligned interpretability (Yu et al., 11 Apr 2026).

The RF-LEGO FT begins from the discrete Fourier transform

$1.6\,\mathrm{mm}$26

and uses Bluestein’s algorithm to rewrite the DFT as a single convolution,

$1.6\,\mathrm{mm}$27

RF-LEGO FT leaves the outer chirp multipliers intact but replaces the fixed convolution with a small complex-valued convolutional layer whose kernel $1.6\,\mathrm{mm}$28 is initialized to $1.6\,\mathrm{mm}$29 and then trained. Optionally, a nulling head implemented as $1.6\,\mathrm{mm}$30 suppresses residual leakage. The resulting operator is

$1.6\,\mathrm{mm}$31

with learnable kernel $1.6\,\mathrm{mm}$32, optional shrinkage threshold $1.6\,\mathrm{mm}$33, and mild complex-valued nonlinearities.

The RF-LEGO Beamformer starts from a ULA sparse AoA model

$1.6\,\mathrm{mm}$34

and the LASSO formulation

$1.6\,\mathrm{mm}$35

Classical ADMM updates

$1.6\,\mathrm{mm}$36

$1.6\,\mathrm{mm}$37

are preserved structurally but modified by learnable, iteration-dependent quantities: a diagonal preconditioner $1.6\,\mathrm{mm}$38, step size $1.6\,\mathrm{mm}$39, shrinkage level $1.6\,\mathrm{mm}$40, and gating vector $1.6\,\mathrm{mm}$41. The unrolled layer becomes

$1.6\,\mathrm{mm}$42

$1.6\,\mathrm{mm}$43

$1.6\,\mathrm{mm}$44

$1.6\,\mathrm{mm}$45

By default, $1.6\,\mathrm{mm}$46 iterations are used.

The RF-LEGO Detector recasts CFAR’s selection-testing pipeline as a compact discrete-time state-space model with latent state $1.6\,\mathrm{mm}$47: $1.6\,\mathrm{mm}$48 The matrices $1.6\,\mathrm{mm}$49 are small trainable matrices initialized to approximate a sliding-window average plus linear test. The paper interprets this as a trapezoidal discretization of the continuous SSM $1.6\,\mathrm{mm}$50, allowing the model to learn clutter/noise adaptation implicitly while preserving the high-level CFAR workflow.

The architecture is explicitly modular. RF-LEGO FT uses a single complex-valued 1D convolution layer of kernel size $1.6\,\mathrm{mm}$51, one mild nonlinearity, and an optional nulling head. RF-LEGO Beamformer uses $1.6\,\mathrm{mm}$52 unrolled ADMM layers with sigmoid gating and softplus step sizes. RF-LEGO Detector uses a single SSM unrolled for $1.6\,\mathrm{mm}$53 time steps, with $1.6\,\mathrm{mm}$54 as the reported window length. The cascaded pipeline may stack FT $1.6\,\mathrm{mm}$55 Beamformer $1.6\,\mathrm{mm}$56 Detector, FT $1.6\,\mathrm{mm}$57 Detector, or Beamformer $1.6\,\mathrm{mm}$58 Detector, and the paper stresses that intermediate outputs remain in the same signal domain, so no bespoke adapters are needed.

Training uses $1.6\,\mathrm{mm}$59 synthetic frames per module. For FT, the data are random spectra with spectral leakage patterns and AWGN at $1.6\,\mathrm{mm}$60–$1.6\,\mathrm{mm}$61 SNR, inverse-DFT to time domain of length $1.6\,\mathrm{mm}$62. For Beamformer, the data are random ULA snapshots with $1.6\,\mathrm{mm}$63 sources, $1.6\,\mathrm{mm}$64 antennas, grid size $1.6\,\mathrm{mm}$65 over $1.6\,\mathrm{mm}$66, and AWGN. For Detector, the data are $1.6\,\mathrm{mm}$67–$1.6\,\mathrm{mm}$68 Hann- or Hamming-shaped peaks in AWGN with window length $1.6\,\mathrm{mm}$69. Implementation is in PyTorch on NVIDIA RTX 4090, with AdamW, learning rate $1.6\,\mathrm{mm}$70, weight decay $1.6\,\mathrm{mm}$71, batch $1.6\,\mathrm{mm}$72, and dropout $1.6\,\mathrm{mm}$73. Reported computational cost is $1.6\,\mathrm{mm}$74, $1.6\,\mathrm{mm}$75 parameters for FT; $1.6\,\mathrm{mm}$76, $1.6\,\mathrm{mm}$77 parameters for Beamformer; and $1.6\,\mathrm{mm}$78, $1.6\,\mathrm{mm}$79 parameters for Detector.

5. Empirical Performance, Cascadability, and Downstream Applications

The sensing paper evaluates RF-LEGO using real-world data for Wi-Fi, millimeter-wave, UWB, and 6G sensing, including mmWave from TI IWR1843 FMCW $1.6\,\mathrm{mm}$80–$1.6\,\mathrm{mm}$81, UWB from XeThru X4A02 at $1.6\,\mathrm{mm}$82, and Wi-Fi CSI from Intel AX200; public datasets include UWCR, OPERAnet, UWB-Context, and DeepSense 6G. Metrics are Peak-to-Side-lobe Ratio (PSLR), Peak-to-Average Power Ratio (PAPR), Mean Absolute Error (MAE), and Detection Rate (DR) at fixed False Alarm Rate $1.6\,\mathrm{mm}$83 (Yu et al., 11 Apr 2026).

At the module level, RF-LEGO FT on mmWave range spectra reports $1.6\,\mathrm{mm}$84 versus $1.6\,\mathrm{mm}$85 for FFT, $1.6\,\mathrm{mm}$86 versus $1.6\,\mathrm{mm}$87, and MAE reduction of $1.6\,\mathrm{mm}$88. For Doppler FT on mmWave, UWB, and Wi-Fi, the reported values are $1.6\,\mathrm{mm}$89 versus $1.6\,\mathrm{mm}$90, $1.6\,\mathrm{mm}$91 versus $1.6\,\mathrm{mm}$92, and MAE reduction of $1.6\,\mathrm{mm}$93. RF-LEGO Beamformer on mmWave AoA reduces MAE from $1.6\,\mathrm{mm}$94 for classical LASSO to $1.6\,\mathrm{mm}$95, a $1.6\,\mathrm{mm}$96 reduction; DA-MUSIC achieves $1.6\,\mathrm{mm}$97 but the paper states that it relies on EVD, has unstable gradients, and lacks explicit angle spectrum, whereas RF-LEGO retains the interpretable spectrum and stable training. RF-LEGO Detector on UWB ToF reports $1.6\,\mathrm{mm}$98 at $1.6\,\mathrm{mm}$99, versus $\epsilon_r=4.4$00; DL-CFAR reaches $\epsilon_r=4.4$01, and loose coupling gives $\epsilon_r=4.4$02.

The framework is evaluated in cascaded pipelines as well. For FT $\epsilon_r=4.4$03 Detector and Beamformer $\epsilon_r=4.4$04 Detector, Detection Rates at $\epsilon_r=4.4$05 improve over pure SP by $\epsilon_r=4.4$06 for range on mmWave, $\epsilon_r=4.4$07 for Doppler, and $\epsilon_r=4.4$08 for angle. Relative to cascaded DL and loose-coupling baselines, the average uplift is $\epsilon_r=4.4$09 and $\epsilon_r=4.4$10, respectively. Ablation results indicate that removing activation in RF-LEGO FT reduces PSLR by $\epsilon_r=4.4$11 for range and $\epsilon_r=4.4$12 for Doppler, while removing the nulling head changes little; for the Beamformer, removing iteration connection increases MAE from $\epsilon_r=4.4$13 to $\epsilon_r=4.4$14, removing gating gives $\epsilon_r=4.4$15, and reducing antennas from $\epsilon_r=4.4$16 to $\epsilon_r=4.4$17 raises MAE to $\epsilon_r=4.4$18. For the Detector, removing activation reduces DR by about $\epsilon_r=4.4$19 and increases FAR by $\epsilon_r=4.4$20, while fixing $\epsilon_r=4.4$21 changes DR from $\epsilon_r=4.4$22 to $\epsilon_r=4.4$23 and increases FAR by $\epsilon_r=4.4$24.

Micro-benchmarks on public datasets report that RF-LEGO matches or outperforms SP and DL baselines in all metrics. Under multiple targets ($\epsilon_r=4.4$25–$\epsilon_r=4.4$26) in mmWave, it retains high PSLR/PAPR and MAE advantage without fine-tuning. Fine-tuning with $\epsilon_r=4.4$27–$\epsilon_r=4.4$28 real data recovers most gains, with range MAE reduction of $\epsilon_r=4.4$29. Interface-sensitivity tests, in which white noise is injected at the module boundary, show that RF-LEGO pipelines degrade more gracefully than pure SP. Reported edge inference latency on Jetson Orin Nano / Raspberry Pi 4 / ESP32-P4 is $\epsilon_r=4.4$30 for FT versus $\epsilon_r=4.4$31 for FFT, $\epsilon_r=4.4$32 for Beamformer versus $\epsilon_r=4.4$33, and $\epsilon_r=4.4$34 for Detector versus $\epsilon_r=4.4$35.

The downstream case studies preserve the same pattern: only the upstream RF-LEGO modules are changed, while downstream filters, estimators, or classifiers remain identical. In trajectory tracking with mmWave plus EKF, on $\epsilon_r=4.4$36 human trajectories in a $\epsilon_r=4.4$37 room with OptiTrack ground truth, the median Absolute Trajectory Error is $\epsilon_r=4.4$38 versus $\epsilon_r=4.4$39 for SP, and Relative Trajectory Error is likewise halved. In vital-sign monitoring, using an infant simulator at $\epsilon_r=4.4$40–$\epsilon_r=4.4$41 and humans at $\epsilon_r=4.4$42–$\epsilon_r=4.4$43, the $\epsilon_r=4.4$44th-percentile MAE is below $\epsilon_r=4.4$45 versus $\epsilon_r=4.4$46 for SP, with overall MAE reduction of $\epsilon_r=4.4$47 for the simulator and $\epsilon_r=4.4$48 for humans. In human activity recognition on the MCD-Gesture mmWave benchmark, the reported accuracy is $\epsilon_r=4.4$49 versus $\epsilon_r=4.4$50 for the 7-way task and $\epsilon_r=4.4$51 versus $\epsilon_r=4.4$52 for the 13-way task.

6. Comparative Interpretation, Misconceptions, and Open Directions

Across the three works, the “LEGO” metaphor is attached to modularity, but the substrate of modularity differs. In the UAV connector, modularity is literal mechanical composition through magnetic self-alignment and a guided stud-and-socket interface. In the RFID surface, modularity is the selective aggregation of passive backscatter elements into active sets. In the deep-unrolling framework, modularity means trainable blocks that preserve classical input/output contracts and can be concatenated as plug-and-play “LEGO bricks” (Debnath et al., 30 Jul 2025, Vardakis et al., 2021, Yu et al., 11 Apr 2026).

A second recurrent theme is that each work constrains flexibility in order to preserve a desired systems property. The UAV connector constrains geometry so that inter-element spacing remains fixed and phase coherence is mechanically enforced. The RFID surface constrains element states to discrete reflection coefficients and relies on Gen2 signaling, which keeps the system batteryless and commodity-based but limits configuration speed and gain. The sensing framework constrains network form to mirror classical SP structure, which preserves structure-aligned interpretability but does not guarantee that internal learned parameters correspond to physically meaningful quantities. The paper explicitly names this limitation “information alienation.”

Several misconceptions are therefore addressed directly by the source material. RF-LEGO in the sensing framework does not claim full transparency of learned weights; its interpretability is structure-aligned. RF-LEGO in the RFID work does not remove the low-SNR bottleneck of end-to-end backscatter links; the paper states that gains saturate and channel-estimation overhead becomes critical for large arrays. RF-LEGO in the UAV work does not rely on complex over-the-air synchronization loops; rather, mechanically linking $\epsilon_r=4.4$53 UAVs effectively creates one monolithic $\epsilon_r=4.4$54-element array with a shared oscillator reference.

The future directions reported in the papers are correspondingly domain-specific. For the RFID surface, the paper lists multi-band operation through redesign of the tag antenna and matching network, higher-order reconfiguration via multi-bit varactor or switched-capacitor networks, faster control through custom RFID commands, active amplification powered by harvested energy, and integration in which groups of tags share a single switch control. For the sensing framework, the paper proposes extending the RF-LEGO library with unrolled modules for clutter suppression via robust PCA, MIMO joint comms + sensing, Doppler-angle coupling, and ray-tracing-informed unrolling; it also proposes hybrid unrolling with physical priors such as nonnegativity, sparsity, and Toeplitz structure, as well as combining RF-LEGO blocks with large-language-model “sensor LMs” for cross-modal reasoning over RF streams. For the UAV architecture, scalability is already framed in terms of gain and frequency by adjusting array element density per UAV, UAV dimensions, and patch length.

Taken together, these works indicate that RF-LEGO is less a single technology than a recurring design language for RF modularization. A plausible implication is that the name is likely to remain context-dependent unless later literature consolidates it into a broader taxonomy. At present, the term spans airborne phased-array interconnects, batteryless reconfigurable surfaces, and deep-unrolled sensing modules, and its meaning must therefore be resolved from the surrounding research domain rather than from the label alone.