GRiN-Drive: Control in EM and Vision

• GRiN-Drive is a design principle leveraging spatially varying refractive indices to enhance system-level performance in domains like lens antennas, quantum RF sensing, and reconfigurable optics.
• It enables precise phase control, impedance matching, and beamforming through engineered GRIN profiles, validated by empirical scaling laws and additive manufacturing methods.
• In computer vision, GRiN-Drive extends to zero-shot monocular depth estimation by incorporating geometry-aware diffusion for state-of-the-art metric reconstruction.

“GRiN-Drive” is best understood as an Editor’s term for a family of research directions in which a graded-index (GRIN) medium is used to drive a target function: wideband collimation and matching in lens antennas, local field enhancement in quantum RF sensing, mechanically tunable focusing and beamforming, or nonlinear spatiotemporal control in GRIN multimode fibers. A separate usage appears in computer vision, where GRIN denotes “Geometric RIN” and is connected to an autonomous-driving setting through zero-shot metric monocular depth estimation rather than through gradient-index electromagnetics. The term therefore does not denote a single standardized device class; it denotes a recurring design logic in which a spatially varying refractive or effective index becomes the actuator of system-level performance (Wang et al., 2023, Garcia et al., 2020, Tishchenko et al., 3 Dec 2025, Kaboutari et al., 26 Jun 2025, Arabí et al., 2017, Guizilini et al., 2024).

1. Scope and principal realizations

Across the literature, GRiN-Drive refers to GRIN-mediated control at several physical scales. In microwave and millimeter-wave systems it appears in Luneburg or flat GRIN lens antennas, where the index profile supplies phase equalization, impedance transformation, and aperture efficiency. In quantum sensing it denotes a passive GRIN lens that concentrates RF energy at a vapor cell and thereby enhances electromagnetically induced transparency (EIT) signatures. In layered optics it denotes mechanical reconfiguration through lateral displacement of GRIN layers. In multimode fiber physics it denotes engineered GRIN structures that drive Kerr beam self-cleaning or geometric parametric instability. In computer vision, the same label is plausibly extended to a driving-oriented depth system because GRIN there names a model architecture, not a refractive medium (Garcia et al., 2020, Tishchenko et al., 3 Dec 2025, Kaboutari et al., 26 Jun 2025, Krupa et al., 2018, Arabí et al., 2017, Guizilini et al., 2024).

Domain	Mechanism	Representative manifestation
Lens antennas	GRIN phase control and matching	14–40 GHz flat lens; 31%–72% aperture efficiency (Garcia et al., 2020)
3D-printed Luneburg lenses	Gyroid artificial dielectric	Maximum frequencies of 20, 25, 33, and $>40$ GHz (Wang et al., 2023)
Quantum RF sensing	Luneburg-type field concentration	EIT splitting approximately doubled at 2.2 and 3.6 GHz (Tishchenko et al., 3 Dec 2025)
Reconfigurable optics	Lateral layer shifts	Focal-point control for $[0,0,0]$, $[0,-5,+5]$, $[0,-10,+10]$ mm (Kaboutari et al., 26 Jun 2025)
GRIN multimode fibers	Nonlinear modal dynamics	Self-cleaning threshold $\sim 2.5\ \text{kW}$ at 1064 nm (Krupa et al., 2018)
Axially modulated GRIN fibers	Moiré-like self-imaging control	Modified GPI gain through periodic core modulation (Arabí et al., 2017)
Monocular depth for driving	Pixel-level diffusion with geometry	Zero-shot metric depth across eight datasets (Guizilini et al., 2024)

This breadth suggests that the unifying content of GRiN-Drive is not a specific geometry but a control principle: spatial index variation is treated as the main design variable through which focusing, sensitivity, modal redistribution, or geometric inference is driven.

2. Electromagnetic foundations of GRIN-driven devices

The canonical electromagnetic realization is the Luneburg lens. Its radial permittivity profile is given by

$\epsilon(r) = 2 - \left(\frac{r}{R}\right)^2,$

with refractive index

$n(r) = \sqrt{2 - \left(\frac{r}{R}\right)^2}.$

This profile focuses a plane wave from any incidence direction to a point on the opposite surface. In additive-manufactured or metamaterial implementations, the continuous profile is approximated by artificial dielectrics whose local geometry determines an effective permittivity $\epsilon_{\text{eff}}(r)$ and hence an effective index $n_{\text{eff}}(r)=\sqrt{\epsilon_{\text{eff}}(r)}$ (Wang et al., 2023, Tishchenko et al., 3 Dec 2025).

In the 3D-printed Luneburg implementation, the lens is tessellated by cubic gyroid unit-cells of side length $l_{uc}$, and the permittivity is varied by changing gyroid wall thickness. The host polymer is Rogers Radix™ with $[0,0,0]$0, while the minimum effective permittivity achievable with structural integrity is $[0,0,0]$1. The ideal Luneburg profile is therefore truncated wherever the nominal $[0,0,0]$2 falls below 1.2; those regions are printed at $[0,0,0]$3 (Wang et al., 2023).

A second electromagnetic formulation appears in wideband flat GRIN lenses based on intrinsically matched unit-cells. There, each unit-cell contains a phase-delaying core bracketed by top and bottom matching sections. The lens is divided into concentric rings, and each ring is assigned a unit-cell from a library characterized over frequency and incident angle. The design objective is uniform phase at the lens output aperture while maintaining high transmission magnitude $[0,0,0]$4. Aperture efficiency is written as

$[0,0,0]$5

and ring selection is performed by choosing the unit-cell whose phase delay most closely matches the required local phase correction (Garcia et al., 2020).

A related problem is impedance transformation in layered GRIN media. Because non-magnetic dielectric realizations couple phase velocity and impedance through $[0,0,0]$6,

$[0,0,0]$7

physically uniform matching layers become non-commensurate transmission-line sections. The systematic taper-design method introduces an effective permittivity $[0,0,0]$8 to equalize the total electrical length of an ideal commensurate taper and a physically uniform taper at the cutoff frequency, thereby restoring predictable Klopfenstein-type behavior (Wang et al., 2021).

3. Lens-antenna synthesis, bandwidth, and operating limits

In 3D-printed GRIN lens antennas, the central empirical constraint is the relation between periodic feature size and guided wavelength. Four identical 10 cm Luneburg lenses were printed using gyroid unit-cells of 12.5, 10, 7.5, and 5 mm and measured over the K- and Ka-band. The maximum operating frequencies were found to be 20, 25, 33, and $[0,0,0]$9 GHz, respectively. With

$[0,-5,+5]$0

the 7.5, 10, and 12.5 mm unit-cells were approximately $[0,-5,+5]$1 on a side at their maximum usable frequencies. Because the cubic gyroid cell can be decomposed into eight octants, the corresponding sub-unit is $[0,-5,+5]$2 on a side, yielding the stated design guideline for additive-manufactured GRIN media (Wang et al., 2023).

That result is significant because it converts a print-resolution question into an electromagnetic scaling law. The relevant breakdown mechanism is not stated as an abstract homogenization threshold alone; it is identified experimentally through the frequency at which lens gain begins to reduce as frequency is increased. A plausible implication is that GRiN-Drive, in this regime, is bounded not only by the intended index profile but by the largest periodic feature that can still behave as an effective dielectric (Wang et al., 2023).

A more mature synthesis pipeline appears in the flat wideband GRIN lens literature. A unit-cell library built from matched phase-delay cells is reused to design lenses of different diameter, focal distance, and feed. The demonstration lens has an 8-inch diameter, nominal thickness 1.2 inches, focal distance 5 inches, and operates from 14 GHz to 40 GHz. Its measured aperture efficiency ranges from 31% to 72% over a 2.9:1 bandwidth, and the antenna is reported as suitable for proposed 5G MMW bands and Ku- and Ka-band fixed satellite services (Garcia et al., 2020).

The matching-taper problem extends this bandwidth engineering. A 9-layer physically uniform Klopfenstein taper was designed for GRIN lens matching, and the fabricated structure achieved return loss better than 15 dB from 8 to 78 GHz. The same work proposed an approximate efficiency formula for taper-matched lenses,

$[0,-5,+5]$3

which predicts aperture efficiency without full-wave simulation and agrees with full-wave results within about 5% across several lens designs (Wang et al., 2021).

Taken together, these studies define a practical GRiN-Drive workflow for antennas: derive the required phase law, select or synthesize unit-cells whose phase and transmission satisfy the local requirement, and ensure that fabrication-induced periodicity remains below the empirically validated frequency limit (Garcia et al., 2020, Wang et al., 2023, Wang et al., 2021).

4. Field enhancement, sensing, and mechanical tunability

In quantum RF sensing, GRiN-Drive denotes the use of a passive GRIN metamaterial lens to increase the local electric field at a sensor rather than the use of a resonant metallic structure. The demonstrated system integrates a GRIN Luneburg-type metamaterial lens with a Cesium vapor-cell Rydberg receiver. The vapor cell is a Thorlabs GC19075-CS quartz reference cell of diameter 19 mm and length 75 mm, placed at the focal region of the lens. Probe and coupling lasers at 852 nm and 509 nm create the EIT configuration, while far-field excitation is provided at 2.2 GHz and 3.6 GHz with 11 dBm transmit power (Tishchenko et al., 3 Dec 2025).

The Luneburg-type lens has diameter $[0,-5,+5]$4 mm, radius $[0,-5,+5]$5 mm, and a cubical lattice with $[0,-5,+5]$6 at 3.5 GHz. It is 3D printed in eight segments using PLA, whose refractive index is given as $[0,-5,+5]$7 at 3.5 GHz. The focusing gain is defined by

$[0,-5,+5]$8

and the maximum measured focusing gain is approximately 8.42 dB at the focal point at 3.6 GHz. In the integrated receiver, the EIT splitting is approximately doubled at both 2.2 GHz and 3.6 GHz, consistent with a reduction of the minimum detectable field according to

$[0,-5,+5]$9

The literature presents this as an ultrawide-bandwidth, non-resonant sensitivity enhancement for quantum RF receivers (Tishchenko et al., 3 Dec 2025).

Mechanical reconfiguration provides a different interpretation of GRiN-Drive. A prospective concept forms a GRIN lens from corrugated magneto-dielectric layers whose refractive-index profiles are expressed in a basis of first-kind Chebyshev polynomials,

$[0,-10,+10]$0

The composite phase profile is altered by laterally shifting layers with respect to one another. The analyzed shift configurations are $[0,-10,+10]$1 mm, $[0,-10,+10]$2 mm, and $[0,-10,+10]$3 mm. Geometrical-optics predictions place the focal points at approximately 134 mm, 125 mm, and 119 mm, whereas full-wave simulations give focal positions at approximately 110 mm, 87 mm, and 79 mm. The ordering is preserved in both models: increasing opposite shifts decreases focal distance (Kaboutari et al., 26 Jun 2025).

This comparison is important because it shows both the utility and the limitations of GRIN-mediated mechanical control. The concept is validated numerically, but discrepancies arise from modeling approximations, structural granularity, and truncation of shifted basis functions at the aperture edges. A plausible implication is that mechanically driven GRIN systems are viable as tunable phase plates, provided that calibration is performed against full-wave or experimental data rather than against geometrical optics alone (Kaboutari et al., 26 Jun 2025).

5. GRIN multimode fibers and nonlinear dynamics

In optical fibers, GRiN-Drive refers to the use of graded-index structure to drive nonlinear self-organization. In a standard commercial 52/125 GRIN multimode fiber with nearly parabolic index profile,

$[0,-10,+10]$4

Kerr nonlinearity induces beam self-cleaning: energy is transferred toward low-order modes, producing a robust central bell-shaped beam. For an 11 m fiber pumped at 1064 nm with 500 ps pulses, the self-cleaning threshold is approximately $[0,-10,+10]$5 input peak power (Krupa et al., 2018).

The same experiments show that spatial beam cleaning is accompanied by nonlinear polarization dynamics. At low power, after 11 m propagation, the degree of linear polarization (DOLP) is approximately 0.1. As the input peak power reaches 4 kW, DOLP rises to approximately 0.26, a 2.5-fold increase, while the polarization azimuth changes by more than $[0,-10,+10]$6 as power increases. When a 13 $[0,-10,+10]$7m aperture is scanned across the beam at 4 kW, the DOLP at the beam center exceeds 0.6. These observations are interpreted as nonlinear re-polarization associated with self-cleaning, followed at higher powers by nonlinear depolarization due to polarization rotation and time averaging (Krupa et al., 2018).

A second fiber-based realization introduces an engineered longitudinal modulation of the GRIN core radius,

$[0,-10,+10]$8

so that the intrinsic self-imaging period

$[0,-10,+10]$9

interacts with the external modulation period $\sim 2.5\ \text{kW}$0. When $\sim 2.5\ \text{kW}$1 is close to $\sim 2.5\ \text{kW}$2, the system develops a Moiré-like pattern that modifies the geometric parametric instability (GPI) gain observed in homogeneous GRIN fibers. The consequence is a richer parametric spectrum in which principal GPI bands split into additional sub-bands whose positions depend on the ratio $\sim 2.5\ \text{kW}$3 (Arabí et al., 2017).

These two bodies of work establish a fiber-specific meaning of GRiN-Drive: the GRIN profile is not merely a passive conduit but the mechanism that organizes modal interference, effective nonlinearity, and instability growth. In one case the outcome is a cleaner, more polarized beam; in the other it is tunable frequency conversion through engineered self-imaging (Krupa et al., 2018, Arabí et al., 2017).

6. Autonomous-driving usage in monocular depth estimation

A separate usage appears in computer vision. Here GRIN denotes “Geometric RIN,” an efficient diffusion model for zero-shot metric monocular depth estimation that is explicitly connected to an autonomous-driving setting. The model operates on sparse unstructured training data and conditions the diffusion process with image features and 3D geometric positional encodings derived from camera intrinsics,

$\sim 2.5\ \text{kW}$4

Depth is represented in log space over the range $\sim 2.5\ \text{kW}$5 using

$\sim 2.5\ \text{kW}$6

with inverse mapping back to metric depth (Guizilini et al., 2024).

The architecture uses Recurrent Interface Networks rather than a UNet, with local and global conditioning streams. At training time, pixels with missing depth are discarded, and the model trains on random subsets of valid pixels, which allows direct use of sparse LiDAR supervision. Training data include Waymo Open Dataset, Lyft Level 5, ArgoVerse 2, Large-Scale Driving, Parallel Domain, TartanAir, OmniData, and ScanNet. Zero-shot evaluation spans eight indoor and outdoor datasets, including KITTI, DDAD, nuScenes, VKITTI2, NYUv2, SunRGBD, and DIODE (Guizilini et al., 2024).

The reported results establish state of the art in zero-shot metric monocular depth even when trained from scratch. On KITTI, GRIN reports AbsRel 0.046, RMSE 2.251, and $\sim 2.5\ \text{kW}$7 0.983; on DDAD, AbsRel 0.093, RMSE 5.307, and $\sim 2.5\ \text{kW}$8 0.922; on nuScenes, AbsRel 0.138, RMSE 7.217, and $\sim 2.5\ \text{kW}$9 0.857; and on VKITTI2, AbsRel 0.074, RMSE 3.501, and $\epsilon(r) = 2 - \left(\frac{r}{R}\right)^2,$0 0.937. Inference uses DDIM with 10 timesteps and is reported at 0.8 seconds per $\epsilon(r) = 2 - \left(\frac{r}{R}\right)^2,$1 image on a single A100 GPU (Guizilini et al., 2024).

This usage is conceptually distinct from gradient-index electromagnetics, yet it retains the same “drive” idea in a looser sense: geometry-aware conditioning drives dense metric structure from a single RGB image. The literature therefore contains both a photonic and a computer-vision sense of GRiN-Drive, and they should not be conflated.

7. Design rules, limitations, and prospective directions

Several cross-cutting design rules recur across the GRiN-Drive literature. In additive-manufactured lens media, the most explicit is the empirical homogenization rule that the relevant gyroid sub-unit should remain at approximately $\epsilon(r) = 2 - \left(\frac{r}{R}\right)^2,$2 on a side. In wideband antennas, matching is improved by embedding impedance transformers into each unit-cell and, when fabrication imposes physically uniform layers, by using an effective-permittivity construction to design non-commensurate Klopfenstein tapers with predictable ripple and cutoff (Wang et al., 2023, Garcia et al., 2020, Wang et al., 2021).

Limitations are equally recurrent. In 3D-printed GRIN lenses, smaller unit-cells increase print time, complexity, cost, and can reduce yield, while minimum wall thickness imposes a lower bound on achievable $\epsilon(r) = 2 - \left(\frac{r}{R}\right)^2,$3. In mechanically reconfigurable lenses, structural granularity and edge truncation shift focal positions away from geometrical-optics predictions. In quantum RF sensing, the dielectric lens is large relative to the vapor cell and requires careful alignment of focal position and polarization. In modulated GRIN fibers, the theoretical treatment neglects linear intermodal coupling, higher-order dispersion, Raman scattering, and other real-fiber perturbations. In diffusion-based depth estimation, inference cost remains high for real-time driving deployments (Wang et al., 2023, Kaboutari et al., 26 Jun 2025, Tishchenko et al., 3 Dec 2025, Arabí et al., 2017, Guizilini et al., 2024).

The forward directions identified in the literature are likewise domain-specific. Higher-resolution additive manufacturing, alternative unit-cell geometries, and multi-material printing are suggested as routes to operation into W-band and beyond. Mechanical GRIN devices are discussed as scalable to infrared wavelengths through MEMS actuation and appropriate transmissive materials. Fiber studies point toward optimized axial modulation patterns and fuller multimode simulations. The depth-estimation work points toward progressive distillation, one-step diffusion, and possible joint inference of camera parameters and 3D points (Wang et al., 2023, Kaboutari et al., 26 Jun 2025, Arabí et al., 2017, Guizilini et al., 2024).

A plausible synthesis is that GRiN-Drive denotes a research style rather than a closed technology stack. Its characteristic move is to elevate the GRIN profile—implemented through artificial dielectrics, layered corrugations, fiber core geometry, or geometry-aware conditioning—from a passive background parameter to the central mechanism that determines usable bandwidth, focus location, local field strength, instability spectrum, or metric scene reconstruction.