One-Shot Calibration in Sensing & Robotics

• One-shot calibration is a family of methods that consolidate parameter identification into a single acquisition or global optimization, reducing the need for repeated measurements while leveraging strong model priors.
• In geometric sensing and robotics, techniques such as structured-light imaging and pseudo-LiDAR enable accurate calibration with minimal data, achieving low RMS errors and enhanced precision.
• The approach extends beyond imaging to optical, quantum, and algorithmic domains by using differentiable models and global parameter tuning to yield computational savings and robust performance.

Searching arXiv for recent and canonical papers on “one-shot calibration” across sensing, robotics, imaging, and hardware calibration. One-shot calibration denotes a family of calibration procedures that compress parameter identification into a single acquisition, a single pass through the environment, or a single global optimization run, rather than relying on multi-image sequences, repeated poses, or sequential per-component tuning. The term is therefore domain-specific. In geometric sensing, it can mean one structured-light image per calibration pose, one reconstructed fringe image, or one image–LiDAR pair; in instrumentation, it can mean a single static camera arrangement or a single measurement step; in hardware optimization, it can mean one global training run over a finite dataset; and in uncertainty quantification, it can mean calibrating prediction sets from extremely scarce labeled data. The unifying idea is not the elimination of modeling, but the concentration of informative measurements so that calibration becomes operationally minimal while remaining statistically or physically constrained (Huang et al., 2018, Cocheteux et al., 2023, Korkiakoski et al., 2013, Zheng et al., 2024, Waldron, 8 Jan 2026, Bousquet et al., 2017).

1. Meanings and scope of the term

The literature uses “one-shot calibration” in several non-equivalent ways. In camera–projector geometry, “single-shot-per-pose” means that, for each pose of the calibration target, only one structured light image needs to be projected and captured to establish dense camera–projector correspondences (Huang et al., 2018). In camera–LiDAR extrinsic estimation, it means a single image–LiDAR pair processed in a single forward pass through a learned pipeline, with no iterative ICP or multi-frame accumulation (Cocheteux et al., 2023). In 3D single-pixel imaging, it means full calibration from one single reconstructed fringe image of a calibration object (Ma et al., 2024). In UWB anchor calibration, it means calibrating stationary anchor positions using a single traversal of a mobile platform through the environment (Yuan et al., 2024).

The same phrase also appears in calibration problems that are not primarily geometric. In wavefront control, the method of differential optical transfer functions is “one-shot” in hardware because it uses a single, static focal-plane camera and no moving parts, even though two intensity images are required in principle and HDR compilations are used in practice (Korkiakoski et al., 2013). In continuous-variable quantum key distribution, “one-time” calibration denotes identifying the shot-noise unit in only one step rather than via a traditional two-time evaluation (Zhang et al., 2019). In post-training quantization, one-shot calibration means a single, short post-training optimization pass on a small data slice that supports runtime switching among several bit-widths without repeated optimization (Xu et al., 13 Feb 2026). In photonic integrated circuits, one-shot calibration means one global training run over measurement datasets to estimate all nonlinear phase–current relations, beam splitter reflectivities, losses, and coupling efficiencies simultaneously (Zheng et al., 2024).

Two further usages broaden the term beyond physical instrumentation. “Toward Optimal Run Racing: Application to Deep Learning Calibration” defines one-shot calibration as selecting, in a single massively parallel launch, the best deep-learning configuration for a dataset while provably minimizing wasted compute by early-stopping non-optimal runs; here calibration refers to algorithm and hyperparameter configuration selection, not probabilistic calibration of predictive confidence (Bousquet et al., 2017). By contrast, “CAOS: Conformal Aggregation of One-Shot Predictors” uses calibration in the conformal sense: computing thresholds or p-values so that prediction sets satisfy a target marginal coverage level in the one-shot regime (Waldron, 8 Jan 2026). A common misconception is therefore that one-shot calibration always refers either to sensor extrinsics or to confidence calibration; the published usage is substantially broader.

2. Geometric calibration from single acquisitions

A canonical geometric instance is the camera–projector system of “A Single-shot-per-pose Camera-Projector Calibration System For Imperfect Planar Targets.” The method uses two independent De Bruijn sequences, one for horizontal stripes and one for vertical stripes, so that every stripe intersection carries a unique codeword decodable from a single captured image. With $k=4$ colors and order $n=3$, each sequence length is $k^n=64$, and the grid dimension is $m \times m$ with $m=k^n+2=66$. The initial calibration uses Zhang’s planar method for the camera, structured-light correspondences for the projector, and pose-wise composition for camera–projector extrinsics. Robustness against imperfect board planarity is then obtained by bundle adjustment over camera parameters, projector parameters, and per-node board coordinates, with cost

$$J = \sum_{j=1}^{N} \sum_{i=1}^{M^j} \left(\| \mathbf{x}^j_\mathrm{c}(i) - \pi_\mathrm{c}(\mathbf{x}^j_\mathrm{m}(i)) \|^2 + \| \mathbf{x}^j_\mathrm{p}(i) - \pi_\mathrm{p}(\mathbf{x}^j_\mathrm{m}(i)) \|^2 + \lambda^j_i \|\hat{\mathbf{x}}^j_\mathrm{m}(i) - \dot{\mathbf{x}}^j_\mathrm{m}(i)\|^2 \right).$$

On real data, the proposed method reported camera, projector, and stereo RMS reprojection errors of $0.35$, $0.64$, and $0.51$ pixels, respectively, and lower reconstruction errors than homography-only and multi-shot baselines on a paper box, plaster bust, and folded paper board (Huang et al., 2018).

A different single-image construction appears in “Calibration of 3D Single-pixel Imaging Systems with a Calibration Field.” There, a standard white cube and a learned “Calibration Field” map one fringe image $I_c$ to a dense pointmap $n=3$0, where $n=3$1, $n=3$2, and $n=3$3. A second module, FriAM, recovers absolute phase $n=3$4 from the same image, giving the observable SPDG coordinate $n=3$5. From the resulting dense $n=3$6 and $n=3$7 correspondences, the projector matrix $n=3$8 and reduced SPDG parameter vector $n=3$9 are estimated. The paper explicitly compares image counts: Ma’s method uses 143 images, Niu’s method 296 images, and the proposed method 1 image. Reported real-data accuracy includes $k^n=64$0 mm RMSE on the calibration cube and $k^n=64$1 mm on a standard sphere (Ma et al., 2024).

Single-shot extrinsic calibration can also be achieved by moving the environment, rather than the target, into a calibrated infrastructure. “Single-Shot is Enough: Panoramic Infrastructure Based Calibration of Multiple Cameras and 3D LiDARs” reconstructs a panoramic room with piece-wise planar walls and floor, CCTag fiducials, and stereo SfM. Each camera is localized from one image by triangle-based AP3P+RANSAC and multi-camera Levenberg–Marquardt refinement; each LiDAR is localized from one scan by corner-direction alignment and point-to-plane ICP against a plane-densified map. Cross-sensor extrinsics are then obtained by composition, $k^n=64$2. The reported average localization times are approximately $k^n=64$3 s for a camera and $k^n=64$4 s for a LiDAR on an i7-8700K CPU, with calibration accuracy comparable to Kalibr, Zhang 2004, and Jeong 2018 on the tested platforms (Fang et al., 2021).

A further variant is catadioptric single-image calibration. “MirrorNeRF: One-shot Neural Portrait Radiance Field from Multi-mirror Catadioptric Imaging” estimates camera pose from a planar homography induced by red hexagon corner markers, restores per-pixel reflected rays using a rasterized sphere-mirror proxy, and then learns a neural warping radiance field whose per-mirror latent codes compensate residual mirror-array misalignment. This couples one-shot geometric calibration with a reconstruction model; the reported checkerboard mean reprojection error is $k^n=64$5, versus $k^n=64$6 for a NeRF baseline and $k^n=64$7 for a vertex-color baseline, and synthetic rendering quality improves to $k^n=64$8 PSNR, $k^n=64$9 SSIM, and $m \times m$0 LPIPS with 25 mirrors (Wang et al., 2021).

3. Initialization-free multisensor calibration in robotics

“PseudoCal: Towards Initialisation-Free Deep Learning-Based Camera-LiDAR Self-Calibration” formalizes camera–LiDAR extrinsics as $m \times m$1, with $m \times m$2 parameterized by roll–pitch–yaw and learned directly rather than via quaternions or exponential maps. The method first converts a monocular RGB image into pseudo-LiDAR by metric depth estimation with GLPN and back-projection,

$m \times m$3

then compares pseudo-LiDAR and true LiDAR in Pillars space via PseudoPillars, followed by two refinement modules, UniCal-M and UniCal-S. Training uses a composite loss

$m \times m$4

with $m \times m$5, $m \times m$6, $m \times m$7, and $m \times m$8. The coarse stage is trained over roll/pitch/yaw ranges $m \times m$9 and translation range $m=k^n+2=66$0 cm, enabling operation under extreme decalibration and even disjoint fields of view. On KITTI, the final reported MAE is $m=k^n+2=66$1 rotation and $m=k^n+2=66$2 cm translation, whereas PseudoPillars alone gives $m=k^n+2=66$3 and $m=k^n+2=66$4 cm (Cocheteux et al., 2023).

The one-shot property in this setting is not merely single-frame inference. It also depends on avoiding externally supplied initial extrinsics. Prior methods summarized in the same source rely on projecting LiDAR into the camera image using a good initial guess and fail under large misalignment or null overlap. PseudoCal instead works in full 3D through pseudo-LiDAR and Pillars fusion, and the paper explicitly states that inference is a single forward pass per module without iterative optimization or ICP (Cocheteux et al., 2023).

A related large-scale deployment problem appears in “Large-Scale UWB Anchor Calibration and One-Shot Localization Using Gaussian Process.” Here one-shot calibration means calibrating stationary anchor positions from one round of UWB range sampling paired with continuous-time LiDAR-Inertial Odometry over environments of roughly $m=k^n+2=66$5 square meters. The anchor-range model is

$m=k^n+2=66$6

where $m=k^n+2=66$7 is modeled with a Gaussian Process using a Matérn kernel. Calibration alternates between anchor-position optimization and GP fitting, with robust filtering for NLoS ranges. The same calibrated anchors are then used to gate the search range for LiDAR loop-closure descriptors by enforcing

$m=k^n+2=66$8

On the reported campus-scale dataset, the proposed CT-LIO + GP method achieves average anchor RMSE of about $m=k^n+2=66$9 m across A1–A9, while RTK-GPS-based calibration yields average anchor error of about $$J = \sum_{j=1}^{N} \sum_{i=1}^{M^j} \left(\| \mathbf{x}^j_\mathrm{c}(i) - \pi_\mathrm{c}(\mathbf{x}^j_\mathrm{m}(i)) \|^2 + \| \mathbf{x}^j_\mathrm{p}(i) - \pi_\mathrm{p}(\mathbf{x}^j_\mathrm{m}(i)) \|^2 + \lambda^j_i \|\hat{\mathbf{x}}^j_\mathrm{m}(i) - \dot{\mathbf{x}}^j_\mathrm{m}(i)\|^2 \right).$$0 m and some anchors never obtain GPS lock. Descriptor search time is reduced from about $$J = \sum_{j=1}^{N} \sum_{i=1}^{M^j} \left(\| \mathbf{x}^j_\mathrm{c}(i) - \pi_\mathrm{c}(\mathbf{x}^j_\mathrm{m}(i)) \|^2 + \| \mathbf{x}^j_\mathrm{p}(i) - \pi_\mathrm{p}(\mathbf{x}^j_\mathrm{m}(i)) \|^2 + \lambda^j_i \|\hat{\mathbf{x}}^j_\mathrm{m}(i) - \dot{\mathbf{x}}^j_\mathrm{m}(i)\|^2 \right).$$1 ms to about $$J = \sum_{j=1}^{N} \sum_{i=1}^{M^j} \left(\| \mathbf{x}^j_\mathrm{c}(i) - \pi_\mathrm{c}(\mathbf{x}^j_\mathrm{m}(i)) \|^2 + \| \mathbf{x}^j_\mathrm{p}(i) - \pi_\mathrm{p}(\mathbf{x}^j_\mathrm{m}(i)) \|^2 + \lambda^j_i \|\hat{\mathbf{x}}^j_\mathrm{m}(i) - \dot{\mathbf{x}}^j_\mathrm{m}(i)\|^2 \right).$$2 ms, and localization success rises from about $$J = \sum_{j=1}^{N} \sum_{i=1}^{M^j} \left(\| \mathbf{x}^j_\mathrm{c}(i) - \pi_\mathrm{c}(\mathbf{x}^j_\mathrm{m}(i)) \|^2 + \| \mathbf{x}^j_\mathrm{p}(i) - \pi_\mathrm{p}(\mathbf{x}^j_\mathrm{m}(i)) \|^2 + \lambda^j_i \|\hat{\mathbf{x}}^j_\mathrm{m}(i) - \dot{\mathbf{x}}^j_\mathrm{m}(i)\|^2 \right).$$3 for STD alone to about $$J = \sum_{j=1}^{N} \sum_{i=1}^{M^j} \left(\| \mathbf{x}^j_\mathrm{c}(i) - \pi_\mathrm{c}(\mathbf{x}^j_\mathrm{m}(i)) \|^2 + \| \mathbf{x}^j_\mathrm{p}(i) - \pi_\mathrm{p}(\mathbf{x}^j_\mathrm{m}(i)) \|^2 + \lambda^j_i \|\hat{\mathbf{x}}^j_\mathrm{m}(i) - \dot{\mathbf{x}}^j_\mathrm{m}(i)\|^2 \right).$$4 for STD + calibrated anchors (Yuan et al., 2024).

Taken together, these robotics papers show that one-shot calibration often replaces repeated acquisition with strong priors on geometry or learned latent structure: pseudo-LiDAR compensates for the lack of an extrinsic initializer, while a GP over range residuals compensates for NLoS bias under sparse large-scale coverage. This suggests that the practical meaning of “one-shot” in robotics is usually “single data-collection episode plus heavy model structure,” not “model-free instantaneous estimation.”

4. Optical and electromagnetic instrumentation

In wavefront control, “Calibrating a high-resolution wavefront corrector with a static focal-plane camera” uses localized diversity at the pupil edge and differential optical transfer functions. With pupil function

$$J = \sum_{j=1}^{N} \sum_{i=1}^{M^j} \left(\| \mathbf{x}^j_\mathrm{c}(i) - \pi_\mathrm{c}(\mathbf{x}^j_\mathrm{m}(i)) \|^2 + \| \mathbf{x}^j_\mathrm{p}(i) - \pi_\mathrm{p}(\mathbf{x}^j_\mathrm{m}(i)) \|^2 + \lambda^j_i \|\hat{\mathbf{x}}^j_\mathrm{m}(i) - \dot{\mathbf{x}}^j_\mathrm{m}(i)\|^2 \right).$$5

the dOTF is the difference between the OTF of a diverse image and the OTF of a reference image, and for sufficiently localized diversity the result contains two shifted copies of the complex pupil field. The method is “one-shot” in hardware because a single static camera and fixed optics are used and no moving parts are required, but in practice each of the two images is compiled from 5–6 exposure times and about 40 reference/diverse pairs are recorded to obtain HDR and averaging. The reported calibration supports a wavefront corrector with about $$J = \sum_{j=1}^{N} \sum_{i=1}^{M^j} \left(\| \mathbf{x}^j_\mathrm{c}(i) - \pi_\mathrm{c}(\mathbf{x}^j_\mathrm{m}(i)) \|^2 + \| \mathbf{x}^j_\mathrm{p}(i) - \pi_\mathrm{p}(\mathbf{x}^j_\mathrm{m}(i)) \|^2 + \lambda^j_i \|\hat{\mathbf{x}}^j_\mathrm{m}(i) - \dot{\mathbf{x}}^j_\mathrm{m}(i)\|^2 \right).$$6 degrees of freedom, determines element locations with spatial resolution $$J = \sum_{j=1}^{N} \sum_{i=1}^{M^j} \left(\| \mathbf{x}^j_\mathrm{c}(i) - \pi_\mathrm{c}(\mathbf{x}^j_\mathrm{m}(i)) \|^2 + \| \mathbf{x}^j_\mathrm{p}(i) - \pi_\mathrm{p}(\mathbf{x}^j_\mathrm{m}(i)) \|^2 + \lambda^j_i \|\hat{\mathbf{x}}^j_\mathrm{m}(i) - \dot{\mathbf{x}}^j_\mathrm{m}(i)\|^2 \right).$$7 of the pupil diameter, and improves Strehl from $$J = \sum_{j=1}^{N} \sum_{i=1}^{M^j} \left(\| \mathbf{x}^j_\mathrm{c}(i) - \pi_\mathrm{c}(\mathbf{x}^j_\mathrm{m}(i)) \|^2 + \| \mathbf{x}^j_\mathrm{p}(i) - \pi_\mathrm{p}(\mathbf{x}^j_\mathrm{m}(i)) \|^2 + \lambda^j_i \|\hat{\mathbf{x}}^j_\mathrm{m}(i) - \dot{\mathbf{x}}^j_\mathrm{m}(i)\|^2 \right).$$8 to $$J = \sum_{j=1}^{N} \sum_{i=1}^{M^j} \left(\| \mathbf{x}^j_\mathrm{c}(i) - \pi_\mathrm{c}(\mathbf{x}^j_\mathrm{m}(i)) \|^2 + \| \mathbf{x}^j_\mathrm{p}(i) - \pi_\mathrm{p}(\mathbf{x}^j_\mathrm{m}(i)) \|^2 + \lambda^j_i \|\hat{\mathbf{x}}^j_\mathrm{m}(i) - \dot{\mathbf{x}}^j_\mathrm{m}(i)\|^2 \right).$$9 when driving a Fast & Furious controller (Korkiakoski et al., 2013).

“WALOP-South: A Four-Camera One-Shot Imaging Polarimeter for PASIPHAE Survey. Paper II -- Polarimetric Modelling and Calibration” defines one-shot operation as measuring $0.35$0, $0.35$1, and $0.35$2 in a single exposure through four simultaneous analyzer channels at $0.35$3, $0.35$4, $0.35$5, and $0.35$6. The basic estimators are

$0.35$7

but field-dependent instrumental polarization and HWP retardance make direct inversion insufficient. The calibration therefore fits per-field polynomial mappings,

$0.35$8

anchored by a calibration linear polarizer and by unpolarized or polarized references. Zemax modeling and lab validation indicate $0.35$9 accuracy in $0.64$0 across more than $0.64$1 of the field of view and $0.64$2 in a narrow high-cross-talk strip (Maharana et al., 2022).

“Calibrating photon counts from a single image” gives a different instrumental meaning of one-shot calibration: detector gain is estimated from the actual image for which the photon count is desired by quantifying out-of-band Fourier information. If $0.64$3 is the average spectral power beyond a threshold $0.64$4, and $0.64$5 is the Fourier-area fraction used, then the image variance is estimated as

$0.64$6

leading, when offset and read noise are known, to

$0.64$7

When offset is unknown, the paper uses a tile-based mean–variance regression to estimate the intercept $0.64$8. The method is reported to be comparable in precision to traditional gain-calibration procedures requiring tens of calibration images, with relative gain error generally well below $0.64$9 and for cameras below $0.51$0 on the tested devices (Heintzmann et al., 2016).

Microwave imaging provides yet another operational pattern. “S-parameter calibration procedure for multiport microwave imaging systems” performs a single one-port ECal calibration on one reference analyzer port, transfers that reflection calibration to all other antenna ports under the assumption that all antennas interact with homogeneous phantoms in the same way, and then uses the unknown-thru technique to calibrate inter-port transmission. This is “near one-shot” because it avoids conventional per-port SOLT or TRL insertion and does not perturb RF components, but still uses three distinct phantoms and reciprocal pair measurements to solve the remaining error terms (Kasper et al., 2019).

5. Quantum, photonic, and hardware-global formulations

“One-Time Shot-Noise Unit Calibration Method for Continuous-Variable Quantum Key Distribution” reframes calibration as normalization under a trusted detector model. Traditional two-time evaluation measures electronic noise with LO off and total detector variance with LO on, then subtracts the former from the latter. The proposed one-time model instead measures

$0.51$1

in a single step with LO on and the signal path blocked, and normalizes the detector output by

$0.51$2

The resulting normalized output is then exactly mapped to the entanglement-based model with trusted losses. The paper reports that a $0.51$3 SNU error can drastically reduce the key rate beyond about $0.51$4 km, and gives proof-of-principle experimental secret key rates of $0.51$5 kbps in the asymptotic regime and $0.51$6 kbps in the finite-size regime over $0.51$7 km (Zhang et al., 2019).

“Global calibration of large-scale photonic integrated circuits” replaces one-by-one phase-shifter tuning by a differentiable end-to-end model of the entire PIC. Each thermo-optic phase shifter $0.51$8 is parameterized by

$0.51$9

each beam splitter by a reflectivity $I_c$0 and transfer efficiency $I_c$1, and the full circuit by an ordered product of discrete-time quantum walk operators. Rather than calibrating element after element, the method jointly optimizes all $I_c$2, beam splitter parameters, and output coupling efficiencies over random current configurations. For the demonstrated depth-12 quantum walk PIC, the chip contains 66 thermo-optic phase shifters and 78 beam splitters, and the reported average $I_c$3 distance over 225 test configurations is $I_c$4 (Zheng et al., 2024).

“QuEPT: Quantized Elastic Precision Transformers with One-Shot Calibration for Multi-Bit Switching” transfers the one-shot notion into post-training quantization. One-shot calibration here means a single, short optimization pass on a small data slice that learns low-rank compensation and clipping thresholds jointly across several target bit-widths. The framework combines Multi-Bit Cascaded Low-Rank Adapters, with

$I_c$5

and a multi-bit MAE reconstruction objective over a sampled low/mid/high-bit triple. After that single pass, runtime switching among uniform and mixed-precision settings uses only the appropriate rank slice and stored scales or zero-points. On ImageNet, QuEPT reports ViT-S top-1 values $I_c$6 across W4/5/6/7/8, and on LLaMA2-7B reports W4A4 perplexities $I_c$7 on WikiText2 and $I_c$8 on C4, with calibration times of about 17–38 minutes for the reported ViTs on a single RTX 3090 (Xu et al., 13 Feb 2026).

These cases share a structural feature: calibration is no longer localized at a single connector, detector, or phase shifter. Instead, one-shot refers to a global fit in which a trusted model or differentiable simulator absorbs the interactions among components. A plausible implication is that one-shot calibration becomes more attractive as the system graph becomes denser, because sequential calibration increasingly accumulates error.

6. Statistical and algorithmic interpretations

In “Toward Optimal Run Racing: Application to Deep Learning Calibration,” one-shot calibration is a sequential multiple-testing problem over massively parallel training runs. All $I_c$9 configurations are launched at once up to horizon $n=3$00, each run is fit online with an ensemble of 11 parametric learning-curve models, and posterior predictive distributions for final performance are used to stop unpromising runs. The key stopping rule is

$n=3$01

with run $n=3$02 stopped when $n=3$03. The paper’s “clever-halt” operator sets $n=3$04 to the $n=3$05-th best conservative predicted asymptotic performance across runs, with

$n=3$06

so that cumulative risk remains controlled without introducing a new free hyperparameter. Reported compute savings are often $n=3$07–$n=3$08, including $n=3$09 on CIFAR-Adagrad and $n=3$10 on CIFAR-Adam at $n=3$11 (Bousquet et al., 2017).

“CAOS: Conformal Aggregation of One-Shot Predictors” addresses a different statistical problem: prediction-set calibration when each labeled exemplar induces its own one-shot predictor. For dataset $n=3$12, CAOS aggregates the $n=3$13 smallest one-shot nonconformity scores,

$n=3$14

then calibrates by leave-one-out scores

$n=3$15

and threshold

$n=3$16

Under exchangeability and self-score optimality, the resulting set $n=3$17 satisfies finite-sample marginal coverage at least $n=3$18. On one-shot facial landmarking, CAOS reduces average set size relative to split conformal baselines while maintaining target coverage; for example at $n=3$19, CAOS reports coverage $n=3$20 and set size $n=3$21, versus $n=3$22 and $n=3$23 for SCOS Avg (Waldron, 8 Jan 2026).

These two papers show that one-shot calibration in machine learning can mean either selecting a model configuration from a single parallel race or calibrating uncertainty sets when only one demonstration per predictor is available. The term therefore retains the “single pass” intuition, but the calibrated quantity may be optimization policy, not sensor geometry or radiometric scale.

7. Recurring design patterns, assumptions, and limits

Across domains, one-shot calibration is rarely assumption-free. In camera–projector calibration it depends on spatial multiplexing by De Bruijn patterns and on bundle adjustment that explicitly absorbs target non-planarity (Huang et al., 2018). In PseudoCal it depends on accurate camera intrinsics, metric monocular depth, and training over extreme decalibration ranges (Cocheteux et al., 2023). In 3D SPI it depends on a digital twin, a known calibration cube, at least two visible cube faces, and a fringe marker that anchors the $n=3$24th order for phase unwrapping (Ma et al., 2024). In dOTF wavefront calibration it depends on Nyquist sampling, localized diversity at the pupil edge, high dynamic range, and system stability during a 2–4 minute acquisition (Korkiakoski et al., 2013).

The same pattern holds in other fields. WALOP’s single-exposure polarimetry still requires field-dependent polynomial calibration, CLP observations, and offset anchoring with standards or Moon-sky patches (Maharana et al., 2022). The CV-QKD one-time model relies on trusted detector assumptions and LO stability (Zhang et al., 2019). Global PIC calibration assumes a correct differentiable circuit model and acknowledges that thermal cross-talk can require explicit extension (Zheng et al., 2024). QuEPT supports only a predefined set $n=3$25 of bit-widths; adding a new bit-width outside $n=3$26 requires re-running the one-shot calibration (Xu et al., 13 Feb 2026). CAOS requires $n=3$27 and its coverage proof assumes self-score optimality (Waldron, 8 Jan 2026). Single-image photon-count calibration assumes band-limited image content, approximately white noise, adequate sampling, and no saturation (Heintzmann et al., 2016). Large-scale UWB anchor calibration can fail when the platform never comes within about 20–30 m of an anchor (Yuan et al., 2024).

Taken together, these works suggest that one-shot calibration is best understood as an engineering trade: measurement burden is reduced by shifting difficulty into code design, simultaneous sensing, differentiable modeling, or stronger prior structure. The concept is therefore not a single method but a recurring strategy for compressing calibration effort without discarding the physical, statistical, or geometric constraints that make the inverse problem identifiable.