Online Abstraction Calibration

• Online abstraction calibration is the process of recalibrating reduced predictive, geometric, or physical representations during system operation to enhance downstream inference and control.
• It employs both passive and active methods, such as bounded-memory updates and observability-aware planning, to adjust calibration on non-i.i.d. or adversarial data streams.
• Applications span robotics, sequential forecasting, sensor fusion, and scientific instrumentation, where recalibration of abstraction layers optimizes performance without modifying core models.

“Online abstraction calibration” (Editor’s term) denotes online procedures that recalibrate reduced predictive, geometric, or physical representations while a system is operating. The surveyed literature does not use a single standardized label; instead it speaks of online calibration, recalibration, self-calibration, detector calibration, or blind calibration. Across these lines of work, the calibrated object ranges from predictive quantile functions and score maps, to sensor extrinsics and temporal offsets, to detector drift-time constants, Gamma maps and relay-wall geometry in confocal NLOS, and latent channel parameters in compressed sensing (Deshpande et al., 2021, Gupta et al., 2023, Nobre et al., 2019, Rohr et al., 2017, Pan et al., 2021, Gabrié et al., 2019). This suggests a unifying view in which calibration often acts on an intermediate abstraction of the underlying system—such as a black-box score, a compact probabilistic forecast, a bounded set of informative motion segments, or a compressed prior over nuisance parameters—rather than on raw data alone.

1. Scope and conceptual structure

A central feature of this literature is that “online” does not mean a single algorithmic pattern. In sequential forecasting, it means that empirical coverage or probability calibration is enforced along the realized adaptive sequence. In robotic self-calibration, it often means bounded-memory estimation over selected trajectory segments, with updates triggered by new informative motion. In detector and instrumentation pipelines, it means that calibration constants are produced and fed back within the validity interval of the underlying physical process. In blind calibration, it means that the posterior over latent calibration parameters becomes the prior for the next sample, so past information is retained without storing all raw observations (Deshpande et al., 2021, Nobre et al., 2019, Rohr et al., 2017, Gabrié et al., 2019).

The calibrated object also varies substantially. Some systems recalibrate uncertainty representations rather than the base model itself. In Bayesian optimization, the calibrated quantity is the predictive quantile function $Q_t$ and its recalibrated map $Q_t \circ R_t$, not the latent GP posterior parameters themselves. In online post-hoc binary classification, the calibrated object is a scalar score transformation $m_t(f(x_t))$, where the base classifier is left unchanged. In high-dimensional forecasting, calibration is enforced on distributions over a finite outcome space by randomizing among a small family of sub-forecasters that predict empirical outcome frequencies over recent windows (Deshpande et al., 2021, Gupta et al., 2023, Peng, 12 Apr 2025).

A second recurring axis is whether calibration is passive or active. Passive methods recalibrate from whatever data arrive. Active methods plan or select data collection actions to improve observability. The clearest example is observability-aware active calibration for ground robots, where the trajectory itself is optimized by maximizing the minimum eigenvalue of the Fisher Information Matrix (FIM). By contrast, bounded-memory robotic self-calibration methods such as FastCal remain passive with respect to motion, but become observability-aware during inference through segment selection and truncated updates (Wang et al., 16 Jun 2025, Nobre et al., 2019).

A third axis is the level of abstraction used to define correspondence or consistency. Some methods use compact semantic or physical surrogates rather than raw measurements. Physics- and semantic-informed multi-sensor calibration uses road users, drivable area, sky constraints, and cycle consistency across camera, LiDAR, and radar. CalibRefine uses object-level correspondences and a planar homography proxy rather than direct $SE(3)$ extrinsic regression. Confocal NLOS imaging uses the line-of-sight component of a transient as a built-in calibration signal for wall geometry, galvanometer mapping, and temporal jitter (Hayoun et al., 2022, Cheng et al., 24 Feb 2025, Pan et al., 2021).

2. Sequential probabilistic prediction and decision-making

In sequential model-based decision-making, the online calibration problem is usually posed as calibration of predictive quantiles or probabilities under adaptive, non-i.i.d. data collection. In Bayesian optimization, the central claim is that the uncertainty that matters is not merely a model’s internal posterior variance, but whether predictive quantiles are calibrated over time—for example, whether an “80% predictive interval” really contains the outcome about 80% of the time. The online criterion is

$$\frac{1}{T}\sum_{t=1}^T \mathbb I\{y_t \le Q_t(p)\} \to p,$$

and the recalibrated target is

$$\frac{1}{T}\sum_{t=1}^T o_t(y_t,R_t(p)) \to p,\qquad o_t(y_t,p)=\mathbb I\{y_t\le Q_t(p)\}.$$

The recalibrator is updated by Follow-The-Regularized-Leader,

$$R_t(p)\in\arg\min_q\left[\psi(q)+\sum_{s=1}^{t-1}\ell_{sp}(y_s,q)\right],$$

with quantile pinball loss

$\ell_{tp}(y_t,q)=(q-Q_t^{-1}(y_t))(o_t(y_t,q)-p).$

For arbitrary sequences, the resulting recalibrated quantiles satisfy

$$\left| \frac{1}{T}\sum_{t=1}^T \mathbb I\{y_t \le Q_t(R_t(p))\} - p \right| \le \frac{1+\eta}{\eta T},$$

and Bayesian optimization is then run on $M_t \circ R_t$ rather than on the uncalibrated model $Q_t \circ R_t$0 (Deshpande et al., 2021).

This literature is explicit that posterior uncertainty and calibrated uncertainty are different notions. Model misspecification, non-Gaussianity, non-stationarity, and action-dependent feedback can all invalidate nominal posterior quantiles. A common misconception is therefore that a Bayesian posterior variance is automatically a reliable exploration signal. The Forrester example in calibrated Bayesian optimization is used precisely to show the opposite: an overconfident GP can fail to explore the true optimum, while recalibration widens intervals and changes the acquisition trajectory (Deshpande et al., 2021).

The same theme appears in online recalibration via Blackwell approachability. There, an arbitrary online predictor producing probabilities $Q_t \circ R_t$1 is transformed into calibrated probabilities $Q_t \circ R_t$2 while controlling regret under a strictly proper scoring rule. Calibration is bucketed, and regret is measured as

$Q_t \circ R_t$3

The key achievement is a joint calibration–accuracy tradeoff: $Q_t \circ R_t$4 obtained by formulating recalibration as an approachability problem over a vector payoff containing both calibration coordinates and a regret coordinate (Okoroafor et al., 2023).

A further generalization addresses multiclass outcomes in high dimension. There the forecaster predicts $Q_t \circ R_t$5, and calibration error is

$Q_t \circ R_t$6

The constructive result is a randomized strategy that becomes $Q_t \circ R_t$7-calibrated after

$Q_t \circ R_t$8

days, together with a lower bound

$Q_t \circ R_t$9

for general high-dimensional calibration. The algorithm uses a small family of forecasters operating at different temporal scales, each predicting empirical outcome frequency over recent windows, and the analysis controls calibration error by a telescoping entropy drop across scales (Peng, 12 Apr 2025).

3. Score-based recalibration and calibration of predictive abstractions

Online post-hoc calibration for binary classification makes the abstraction layer explicit. A pretrained probabilistic classifier $m_t(f(x_t))$0 is not retrained; instead the system learns an online mapping $m_t(f(x_t))$1 and predicts

$m_t(f(x_t))$2

The paper on Online Platt Scaling states that $m_t(f(x_t))$3 is a compact feature summarizing $m_t(f(x_t))$4, and explicitly argues that this is exactly an abstraction layer for calibration. Standard Platt scaling

$m_t(f(x_t))$5

is converted into an online calibrator by replacing batch logistic regression with online logistic regression on the two-dimensional pseudo-feature

$m_t(f(x_t))$6

With ONS, the method competes with the best fixed Platt scaler in hindsight, with deterministic finite-sample regret bounds such as

$m_t(f(x_t))$7

under score clipping assumptions (Gupta et al., 2023).

This work is careful to separate regret to a parametric calibrator from absolute calibration. Low regret to the best fixed Platt scaler does not imply that the forecasts are calibrated, because the Platt family itself may be misspecified. The paper addresses this by adding calibeating, which bins examples according to the expert forecast and corrects forecasts within bins either by empirical tracking or by adversarial hedging. The tracking variant TOPS predicts the historical average label within the current OPS bin; the hedging variant HOPS runs Foster’s randomized procedure within each bin. For HOPS, the expected calibration error satisfies

$m_t(f(x_t))$8

while preserving much of OPS sharpness (Gupta et al., 2023).

This clarifies a general structural point. Calibration of abstractions can be parametric, nonparametric, or hybrid. OPS is parametric and low-dimensional. TOPS and HOPS add a nonparametric correction layer on top of the parametric abstraction. The approachability-based recalibrator is more adversarially oriented and produces an explicit calibration–regret frontier. High-dimensional multiclass calibration replaces geometric discretization of the full simplex with a compressed temporal abstraction built from recent empirical frequencies. These are distinct constructions, but they all recalibrate a reduced representation rather than the full original covariate space (Gupta et al., 2023, Okoroafor et al., 2023, Peng, 12 Apr 2025).

A plausible implication is that “abstraction calibration” is most effective when the abstraction retains decision-relevant structure but is small enough for online guarantees or bounded-memory updates. The score $m_t(f(x_t))$9, the quantile function $SE(3)$0, and the recent-window empirical distribution are three different examples of this tradeoff (Gupta et al., 2023, Deshpande et al., 2021, Peng, 12 Apr 2025).

4. Robotic self-calibration, observability, and active excitation

In robotics, online abstraction calibration usually appears as online self-calibration of sensor extrinsics, with explicit handling of observability and drift. FastCal formulates calibration as a loosely coupled alignment problem over sensor-relative poses already produced by upstream odometry or SLAM. Instead of jointly optimizing poses, landmarks, and extrinsics, it solves a calibration-only regression

$SE(3)$1

which for one sensor pair yields a $SE(3)$2 least-squares system. Informativeness is scored by Gaussian differential entropy

$SE(3)$3

and only a bounded queue of informative motion segments is retained. Observability is handled by a truncated-SVD update on the whitened Jacobian, so only numerically observable singular directions are updated. Drift is handled by exponential time decay,

$SE(3)$4

with stale segments removed when their decay falls below threshold. The method is reported to run up to an order of magnitude faster than similar self-calibration algorithms for camera-to-camera extrinsics, and was integrated on a real platform for three weeks and roughly $SE(3)$5 km (Nobre et al., 2019).

An earlier monocular SLAM system already established several of the same principles: retain only the top $SE(3)$6 informative trajectory segments, perform self-calibration at any time, and detect calibration changes probabilistically. Its change detector compares recent and long-term calibration posteriors through a multivariate Behrens–Fisher test with null hypothesis

$SE(3)$7

uses significance level $SE(3)$8, and requires $SE(3)$9 consecutive rejections before declaring a recalibration event. In the reported zoom-event experiment, the system achieved $$\frac{1}{T}\sum_{t=1}^T \mathbb I\{y_t \le Q_t(p)\} \to p,$$0 m translation error over $$\frac{1}{T}\sum_{t=1}^T \mathbb I\{y_t \le Q_t(p)\} \to p,$$1 m traveled, i.e. $$\frac{1}{T}\sum_{t=1}^T \mathbb I\{y_t \le Q_t(p)\} \to p,$$2 distance-traveled error, while recalibrating intrinsics online (Keivan et al., 2014).

The active-calibration extension makes observability not just an inference issue but a motion-planning objective. For a 2D ground robot with wheel odometry, LiDAR, and microphone array, the full extrinsic vector is

$$\frac{1}{T}\sum_{t=1}^T \mathbb I\{y_t \le Q_t(p)\} \to p,$$3

and the trajectory is parameterized by clamped B-splines,

$$\frac{1}{T}\sum_{t=1}^T \mathbb I\{y_t \le Q_t(p)\} \to p,$$4

Observability is quantified by the FIM

$$\frac{1}{T}\sum_{t=1}^T \mathbb I\{y_t \le Q_t(p)\} \to p,$$5

and the planner solves

$$\frac{1}{T}\sum_{t=1}^T \mathbb I\{y_t \le Q_t(p)\} \to p,$$6

subject to trajectory feasibility constraints. During execution, an EKF updates extrinsics recursively, and the next spline segment is replanned using the updated estimate and covariance. In both simulation and real-world experiments, active planning outperformed fixed figure-8, circle, line, and random trajectories, especially on weakly observable directions and especially for the microphone array (Wang et al., 16 Jun 2025).

A recurring misconception in this area is that online self-calibration can overcome fundamental identifiability limits. The robotics papers are explicit that it cannot. FastCal freezes unobservable dimensions rather than hallucinating them, and the active method treats poor FIM eigenvalues as a planning problem rather than a defect that estimation alone can solve. In other words, robustness to degeneracy means stable behavior under insufficient excitation, not recovery of fundamentally unobservable parameters (Nobre et al., 2019, Wang et al., 16 Jun 2025).

5. Multimodal autonomous-system calibration beyond passive extrinsics

A different strand of work calibrates richer multimodal abstractions in autonomous systems. One approach performs integrated temporal–spatial online calibration for camera–LiDAR fusion by modeling the transform from LiDAR at time $$\frac{1}{T}\sum_{t=1}^T \mathbb I\{y_t \le Q_t(p)\} \to p,$$7 to camera at time $$\frac{1}{T}\sum_{t=1}^T \mathbb I\{y_t \le Q_t(p)\} \to p,$$8 as

$$\frac{1}{T}\sum_{t=1}^T \mathbb I\{y_t \le Q_t(p)\} \to p,$$9

with residual error contributions from both temporal motion compensation and extrinsic offset. The method uses inertial pre-integration, environmental line features, dynamic target elimination by adjacent point-cloud matching, and an iterative search-based alignment score over projected LiDAR line points. On KITTI, it reports a $$\frac{1}{T}\sum_{t=1}^T o_t(y_t,R_t(p)) \to p,\qquad o_t(y_t,p)=\mathbb I\{y_t\le Q_t(p)\}.$$0 improvement over ROS soft synchronization in temporal calibration, automatic disturbance correction within $$\frac{1}{T}\sum_{t=1}^T o_t(y_t,R_t(p)) \to p,\qquad o_t(y_t,p)=\mathbb I\{y_t\le Q_t(p)\}.$$1 s, and spatial accuracy of $$\frac{1}{T}\sum_{t=1}^T o_t(y_t,R_t(p)) \to p,\qquad o_t(y_t,p)=\mathbb I\{y_t\le Q_t(p)\}.$$2 (Wang et al., 2022).

Joint camera–LiDAR–radar calibration extends this further by using pairwise semantic and physical constraints together with global cycle consistency. The pairwise transforms $$\frac{1}{T}\sum_{t=1}^T o_t(y_t,R_t(p)) \to p,\qquad o_t(y_t,p)=\mathbb I\{y_t\le Q_t(p)\}.$$3 are optimized by

$$\frac{1}{T}\sum_{t=1}^T o_t(y_t,R_t(p)) \to p,\qquad o_t(y_t,p)=\mathbb I\{y_t\le Q_t(p)\}.$$4

with a cyclic constraint

$$\frac{1}{T}\sum_{t=1}^T o_t(y_t,R_t(p)) \to p,\qquad o_t(y_t,p)=\mathbb I\{y_t\le Q_t(p)\}.$$5

The corresponding self-supervised model predicts all pairwise extrinsics from a single frame using the same pairwise and global losses,

$$\frac{1}{T}\sum_{t=1}^T o_t(y_t,R_t(p)) \to p,\qquad o_t(y_t,p)=\mathbb I\{y_t\le Q_t(p)\}.$$6

Quantitative results show that both optimization and SSL outperform an object-tracking baseline, that joint methods outperform pairwise methods, and that joint SSL is best overall (Hayoun et al., 2022).

CalibRefine adopts a different abstraction: it does not estimate a conventional $$\frac{1}{T}\sum_{t=1}^T o_t(y_t,R_t(p)) \to p,\qquad o_t(y_t,p)=\mathbb I\{y_t\le Q_t(p)\}.$$7 extrinsic transform, but a planar homography from LiDAR ground-plane coordinates to image coordinates,

$$\frac{1}{T}\sum_{t=1}^T o_t(y_t,R_t(p)) \to p,\qquad o_t(y_t,p)=\mathbb I\{y_t\le Q_t(p)\}.$$8

Its pipeline combines a Common Feature Discriminator, coarse homography estimation from object-level correspondences, iterative online refinement every $$\frac{1}{T}\sum_{t=1}^T o_t(y_t,R_t(p)) \to p,\qquad o_t(y_t,p)=\mathbb I\{y_t\le Q_t(p)\}.$$9 frames, and an attention-based residual homography

$$R_t(p)\in\arg\min_q\left[\psi(q)+\sum_{s=1}^{t-1}\ell_{sp}(y_s,q)\right],$$0

Evaluation is reported in terms of AED and RMSE, with final results of AED $$R_t(p)\in\arg\min_q\left[\psi(q)+\sum_{s=1}^{t-1}\ell_{sp}(y_s,q)\right],$$1 and RMSE $$R_t(p)\in\arg\min_q\left[\psi(q)+\sum_{s=1}^{t-1}\ell_{sp}(y_s,q)\right],$$2 on Dataset 1, and AED $$R_t(p)\in\arg\min_q\left[\psi(q)+\sum_{s=1}^{t-1}\ell_{sp}(y_s,q)\right],$$3 and RMSE $$R_t(p)\in\arg\min_q\left[\psi(q)+\sum_{s=1}^{t-1}\ell_{sp}(y_s,q)\right],$$4 on Dataset 2, outperforming the listed targetless baselines and manual initialization in those comparisons (Cheng et al., 24 Feb 2025).

These methods show that online calibration in autonomous systems increasingly operates on mid-level abstractions: object centers, drivable area, semantic classes, line features, homographies, and cycle-consistent pairwise transforms. This suggests a shift away from purely geometric target-based procedures toward targetless, scene-driven calibration objectives that can be evaluated continuously during deployment (Wang et al., 2022, Hayoun et al., 2022, Cheng et al., 24 Feb 2025).

6. Streaming scientific instrumentation, NLOS systems, and latent-channel calibration

In large-scale scientific instrumentation, online calibration appears as a systems problem in which calibration constants must be derived from reconstructed data and fed back before the relevant physical conditions drift. In ALICE HLT, TPC drift-time calibration is built into a real-time reconstruction farm. The drift velocity model is

$$R_t(p)\in\arg\min_q\left[\psi(q)+\sum_{s=1}^{t-1}\ell_{sp}(y_s,q)\right],$$5

where $$R_t(p)\in\arg\min_q\left[\psi(q)+\sum_{s=1}^{t-1}\ell_{sp}(y_s,q)\right],$$6 is the theoretical drift velocity from temperature, pressure, and time, and $$R_t(p)\in\arg\min_q\left[\psi(q)+\sum_{s=1}^{t-1}\ell_{sp}(y_s,q)\right],$$7 is the correction factor inferred from TPC–ITS misalignment. The pipeline accumulates statistics over many events, merges calibration products centrally, preprocesses an updated TPC transformation map, and feeds it back via ZeroMQ to the front of the reconstruction chain. In first deployment, cluster $$R_t(p)\in\arg\min_q\left[\psi(q)+\sum_{s=1}^{t-1}\ell_{sp}(y_s,q)\right],$$8-position disagreement with offline improved from as much as $$R_t(p)\in\arg\min_q\left[\psi(q)+\sum_{s=1}^{t-1}\ell_{sp}(y_s,q)\right],$$9 to less than $\ell_{tp}(y_t,q)=(q-Q_t^{-1}(y_t))(o_t(y_t,q)-p).$0, corresponding to an improvement by about a factor $\ell_{tp}(y_t,q)=(q-Q_t^{-1}(y_t))(o_t(y_t,q)-p).$1 (Rohr et al., 2017).

A companion HLT paper clarifies that this required two enabling abstractions: fast TPC tracking and asynchronous processing support. The tracker is based on a Cellular Automaton and a Kalman filter, was ported to CUDA, OpenCL, and CPU with a single common code base, and on a GTX980 achieved a reported tracking time of $\ell_{tp}(y_t,q)=(q-Q_t^{-1}(y_t))(o_t(y_t,q)-p).$2 ms in the given benchmark table. Framework extensions added event-synchronous custom sources, out-of-band ZeroMQ feedback, and asynchronous execution of long-running calibration tasks so the event pipeline remained loop-free and nonblocking (Rohr et al., 2017).

In confocal NLOS imaging, online calibration is embedded directly into the acquisition model. The measured transient is decomposed into hidden and line-of-sight components,

$\ell_{tp}(y_t,q)=(q-Q_t^{-1}(y_t))(o_t(y_t,q)-p).$3

and the LOS term

$\ell_{tp}(y_t,q)=(q-Q_t^{-1}(y_t))(o_t(y_t,q)-p).$4

is reused as a built-in calibration signal. From LOS peak timing and amplitude, the system estimates relay-wall geometry, galvanometer voltage-to-angle mapping, Gamma maps, temporal jitter, and a usable hidden-scene bounding box. The relay wall is fit as

$\ell_{tp}(y_t,q)=(q-Q_t^{-1}(y_t))(o_t(y_t,q)-p).$5

using the online loss

$\ell_{tp}(y_t,q)=(q-Q_t^{-1}(y_t))(o_t(y_t,q)-p).$6

This is a particularly clear example of calibration from a reused intermediate abstraction: the LOS return, normally treated as nuisance, becomes the calibration channel (Pan et al., 2021).

Blind calibration for compressed sensing makes the same idea probabilistic. Measurements depend on unknown sensor-wise calibration parameters $\ell_{tp}(y_t,q)=(q-Q_t^{-1}(y_t))(o_t(y_t,q)-p).$7, and online cal-AMP processes one sample at a time while carrying forward a posterior over calibration variables. The core recursion is posterior-as-prior: $\ell_{tp}(y_t,q)=(q-Q_t^{-1}(y_t))(o_t(y_t,q)-p).$8 For gain calibration, the entire history compresses into two per-sensor sufficient-statistics-like quantities $\ell_{tp}(y_t,q)=(q-Q_t^{-1}(y_t))(o_t(y_t,q)-p).$9 and $$\left| \frac{1}{T}\sum_{t=1}^T \mathbb I\{y_t \le Q_t(R_t(p))\} - p \right| \le \frac{1+\eta}{\eta T},$$0, updated recursively as new samples arrive. The paper derives state evolution for both offline and online variants, and reports that online cal-AMP is somewhat less sample-efficient than offline joint processing but remains highly effective, with strong agreement between AMP trajectories and state-evolution predictions (Gabrié et al., 2019).

Across these physical-system examples, the calibrated object is rarely the raw instrument alone. It is a reduced model component—drift velocity correction, transformation map, LOS-derived Gamma map, or latent channel parameter—that mediates between raw observations and downstream inference. That is why these papers fit naturally into an abstraction-calibration perspective (Rohr et al., 2017, Pan et al., 2021, Gabrié et al., 2019).

7. Recurring principles, misconceptions, and open constraints

Several design principles recur across the surveyed literature. One is bounded-memory or compressed-state online updating. FastCal keeps only a bounded queue of informative segments. Online Platt Scaling calibrates only a scalar score. High-dimensional online calibration replaces simplex discretization by a logarithmic family of temporal sub-forecasters. Online cal-AMP carries forward only an updated prior or its compressed parameters. ALICE HLT distributes only updated conditions objects rather than replaying the full run (Nobre et al., 2019, Gupta et al., 2023, Peng, 12 Apr 2025, Gabrié et al., 2019, Rohr et al., 2017).

A second principle is explicit treatment of non-i.i.d. or adversarial data. Sequential Bayesian optimization calibration proves coverage guarantees for arbitrary sequences. OPS and HOPS are designed for drift and adversarial outcomes. The approachability-based recalibrator derives calibration and regret guarantees without stochastic assumptions. High-dimensional multiclass calibration is likewise adversarially sequential. This is a substantive departure from classical batch calibration, where exchangeability or static train/test assumptions are often implicit (Deshpande et al., 2021, Gupta et al., 2023, Okoroafor et al., 2023, Peng, 12 Apr 2025).

A third principle is observability or identifiability awareness. In robotics, this appears as TSVD freezing of unsupported directions, multivariate change detection, or FIM-maximizing active trajectory generation. In multimodal vehicle calibration, it appears more indirectly through semantic filtering, block-based correspondence diversity, and global cycle consistency. In scientific instrumentation, the analogue is validity-interval budgeting: the calibration loop must complete before detector conditions drift too far (Nobre et al., 2019, Keivan et al., 2014, Wang et al., 16 Jun 2025, Hayoun et al., 2022, Rohr et al., 2017).

Several misconceptions are repeatedly corrected by the literature. Calibration is not the same as posterior variance estimation, not the same as regret to the best parametric scaler, and not the same as global geometric consistency under an uninformative motion regime. Another common misunderstanding is that an online method necessarily updates the full underlying model. Many successful systems instead recalibrate a post-hoc map, a correction factor, a homography, or a compact latent prior while leaving the primary predictor or reconstruction chain intact (Deshpande et al., 2021, Gupta et al., 2023, Okoroafor et al., 2023, Cheng et al., 24 Feb 2025).

Important constraints also remain. Some methods are strongest only for piecewise-constant changes rather than gradual drift. Some rely on line-rich or object-rich environments. Some require a known sound source or a planar relay wall. Some online calibration objects are entangled with the assumptions used during calibration and therefore are not directly transferable to offline reconstruction under different conditions. Homography-based calibration proxies are narrower than full rigid-body calibration, and recursive filters can be less forgiving than smoothing-based estimators under strong nonlinearities (Keivan et al., 2014, Wang et al., 2022, Pan et al., 2021, Rohr et al., 2017, Cheng et al., 24 Feb 2025, Wang et al., 16 Jun 2025).

Taken together, the literature supports a broad but technically coherent interpretation of online abstraction calibration: online maintenance of reduced predictive, geometric, semantic, or physical representations whose correctness is essential for downstream inference and control. The calibrated representation may be a quantile map, a score transform, a trajectory-selected information subspace, a detector correction factor, a homography, or a latent channel prior; what unifies these cases is that calibration is performed during operation, under bounded resources, and with explicit attention to coverage, observability, or physical consistency (Deshpande et al., 2021, Nobre et al., 2019, Rohr et al., 2017, Gabrié et al., 2019).