acceptodds
Under review as a conference paper at ICLR 2027

When Cross-Modal Evidence Helps: Modality-Local vs. Cross-Modal Representations for Reliability Inference in State Estimation

Abstract

Learning-based state estimation increasingly relies on multimodal sensing under changing sensing conditions, yet it remains unclear when reliability should be inferred from a sensor's own history and when cross-modal evidence is necessary. We introduce a Local–Cross framework that separates modality-local temporal representations from cross-modal relational representations while keeping the prediction target and model capacity matched. Using UAV optical flow, IMU, and GPS state estimation, we study two estimator-facing reliability quantities: covariance for stochastic uncertainty and consistency for compatibility with the nominal sensing model. We find that covariance prediction consistently favors modality-local representations, whereas consistency prediction consistently benefits from cross-modal representations, both when each degradation occurs alone and when the two occur simultaneously. When integrated into a classical ESKF, modality-local covariance prediction combined with cross-modal consistency prediction improves state estimation under degraded sensing while preserving nominal behavior. These results suggest that cross-modal evidence should be used selectively according to the reliability quantity being inferred rather than by default.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Multimodal Representation Learning, Cross-Modal Representation, Reliability Inference, Sensor Degradation, Robust State Estimation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_PPwEOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper asks when reliability quantities in multimodal state estimation should be inferred from a sensor's own history and when they benefit from cross-modal evidence. It proposes a ``Local--Cross'' comparison for UAV optical-flow/IMU/GPS fusion. The Local variant feeds a task head only a modality's own causal TCN latent, while the Cross-Modal variant additionally uses the other two modalities' latents and learned directed residuals etm←n=htm−h^tm←ne_t^{m\leftarrow n}=h_t^m-\hat{h}_t^{m\leftarrow n}. Parameter counts are matched by adjusting hidden widths. Two reliability targets are studied: per-role covariance under stochastic covariance variation (SCV), and binary measurement consistency under sensing consistency disruption (SCD), including freezes, biases, drifts, and outliers; both are also evaluated under compound degradation. Experiments use PX4 SITL/Gazebo simulation with synthetic degradations injected in a 3×33\times3 SCV$\times$SCD design, and statistical comparisons use a family-stratified cluster bootstrap over 12 held-out trajectories. Across two matched-capacity tiers, Local achieves lower log-covariance MAE, whereas Cross-Modal achieves higher consistency AUROC/AUPRC. When integrated into a non-learned ESKF, combining Local covariance prediction with Cross-Modal consistency prediction (F3) reduces position and velocity error under SCD and compound degradation and improves several NEES-based calibration measures relative to Local consistency (F2), while leaving nominal performance essentially unchanged.

Critical strengths
  • The paper isolates a clear and meaningful question of whether reliability prediction benefits from modality-local or cross-modal information. The Local and Cross-Modal comparisons are tightly controlled, with the same prediction targets, data, optimization procedure, and closely matched parameter budgets. Evaluation is also statistically precise since it uses held-out trajectories, treats the clean trajectory as the unit of dependence, averages antithetic noise pairs, and reports family-stratified cluster-bootstrap confidence intervals. The inclusion of downstream NIS/NEES analysis further strengthens the evaluation beyond standalone prediction metrics.

  • The downstream evaluation is particularly useful. Continuous consistency modulation substantially improves state estimation under sensing consistency disruption relative to covariance-only adaptation and hard χ2\chi^2 gating, while leaving nominal performance essentially unchanged. More specifically, replacing Local consistency prediction with Cross-Modal consistency prediction (F3 vs. F2) gives an additional improvement under both SCD and compound degradation, showing that the representation-level finding has practical consequences for the estimator.

Critical weaknesses
  • The central finding is closely tied to how the benchmark is constructed. SCV changes a sensor's own noise covariance, so the target is naturally reflected in that sensor's local statistics. SCD, by contrast, is defined through violations of the nominal relationship among sensors. The result that covariance favors Local while consistency favors Cross-Modal is therefore largely expected from the definitions, rather than establishing a broader principle about when cross-modal evidence is useful.

  • The matched-capacity comparison does not fully isolate representation choice. Parameter matching is achieved by reducing the Cross-Modal hidden width (e.g., covariance Small: Local h=64h=64 vs. Cross h=49h=49), so the comparison changes both access to cross-modal information and local representational capacity. A same-width control or shared/frozen local encoder would make the conclusion that cross-modal information hurts covariance prediction much stronger.

  • The statistical conclusions are based on a single trained model per task/configuration. The trajectory bootstrap captures evaluation-set variation, but not training or initialization variance, which matters given the relatively small consistency gains (Local AUROC is already around 0.980.98).

  • The broader conclusion is supported only on one simulated UAV setup with synthetic degradations and one specific Cross-Modal construction. The downstream results are useful, but Cross-Modal covariance is never tested in the ESKF, and the filter remains far from statistically consistent under degradation despite the relative improvements.

Questions for authors
  • How much do the Local--Cross differences vary across independent training seeds? Can the authors retrain each configuration with several seeds and report uncertainty that incorporates both training and test-trajectory variation?

  • Parameter matching reduces the Cross-Modal hidden width relative to Local. Does Cross still underperform Local for covariance when the local encoder width is held fixed (e.g., Local h=64h=64 vs. Cross h=64h=64), even if total parameter counts are no longer matched?

  • Can the authors report consistency AUROC/AUPRC separately for each SCD type? In particular, is the Cross-Modal advantage concentrated in bias, drift, and coherent-incorrect measurements, while freezes/outliers are already detectable locally? If so, how should the claimed principle be interpreted beyond the observation that failures whose detection naturally depends on cross-modal disagreement benefit from relational evidence?

3 comments
AuthorsOpenReview ↗

Response to Reviewer PPwE: Training Variability, Degradation-Specific Findings, and Scope

We thank the reviewer for assessing our controls and downstream evaluation. Your questions prompted stronger tests of training variability, architecture, and degradation heterogeneity, and helped us refine the scope of the conclusions.

1. Does the finding follow from the benchmark definitions?

The SCV/SCD construction motivates the hypothesis, but does not determine the empirical ordering. Structured errors may leave detectable local temporal cues, and a cross-modal model can preserve a Local prediction pathway without using reference features.

The revised five-run results illustrate this distinction: Cross improves aggregate consistency AUPRC under CD, whereas SCD-only AUROC and AUPRC difference intervals include zero. Mechanism-specific results also include gains, uncertain differences, and reversals. Preserving the Local pathway reduces the covariance gap.

We frame the findings as an empirical reference for information selection, conditional on target, architecture, and degradation context, rather than a universal assignment of covariance to Local and consistency to Cross.

2. Training variability and capacity controls

Parameter-matched, width-matched, and fusion comparisons now use five runs. Their 95% CIs use 5,000 paired resamples crossing training runs with family-stratified trajectory clusters, accounting for both sources of variation (Appendix D).

Matching counts parameters connected to the task loss. Consistency Local/Cross counts are 120,307/121,103 (Small) and 198,343/197,123 (Large), differing by less than 1%; Local widths are 73/94 and Cross widths 50/64.

All four parameter-matched covariance contrasts favor Local. For Large CD, log-MAE is 0.1487 versus 0.1900: difference +0.0413, 95% CI [0.0300, 0.0512] (Table 3).

Consistency AUPRC differences are:

BudgetSCD-onlyCD
Small+0.0042; 95% CI [−0.0037, 0.0109]+0.0129; 95% CI [0.0060, 0.0198]
Large+0.0033; 95% CI [−0.0040, 0.0108]+0.0172; 95% CI [0.0095, 0.0253]

SCD-only AUROC intervals also include zero. We revised the abstract and conclusions to reflect this boundary (Table 9).

At fixed encoder/head hidden width h=64, covariance differences remain +0.0396 under SCV (95% CI [0.0304, 0.0485]) and +0.0410 under CD (95% CI [0.0288, 0.0524]). CD AUPRC improves by 0.0184 (95% CI [0.0116, 0.0262]); SCD-only AUPRC intervals include zero at both widths (Tables 4 and 10).

We compare Concat, Relational, and Preserved Local fusion. Preserved Local freezes the trained Local predictor and adds a zero-initialized residual branch; the initial checkpoint remains eligible for selection. Its CD covariance difference is +0.0017 (95% CI [−0.0017, 0.0055]). Thus, width alone does not explain the gap, but architecture materially affects it (Table 6).

3. Degradation-specific analyses

We extended your requested SCD breakdown to both targets: consistency AUROC/AUPRC for eight corruption mechanisms, and covariance error for five prediction roles under SCV/CD (Figure 7; Tables 23–26). The covariance analysis groups errors by predicted quantity, not by the role receiving the perturbation.

The consistency diagnostic set contains one SCD mechanism per case across 12 held-out trajectories. For Small CD, AUPRC differences are:

  • GPS freeze: +0.0206; 95% CI [0.0056, 0.0353].
  • OF incorrect velocity: +0.0490; 95% CI [0.0291, 0.0692].
  • GPS position bias: +0.000346; 95% CI [−0.0010, 0.0016].

OF incorrect velocity benefits at both capacities under SCD/CD. GPS freeze benefits at both capacities under CD, but its SCD-only intervals include zero. For Large SCD-only GPS position drift, Local is better: difference −0.0109, 95% CI [−0.0139, −0.0079].

These results clarify where aggregate findings hold. Mechanism-wise intervals are exploratory, unadjusted for multiple comparisons, and conditional on one training run. They do not establish a causal explanation for each difference.

4. Downstream effects, calibration, and scope

We extended the covariance replacement to a full factorial evaluation (Tables 17–22). Under CD, changing Local to Cross consistency at fixed Local covariance changes position/velocity MAE by −0.0298 m (95% CI [−0.0438, −0.0157]) and −0.0271 m/s (95% CI [−0.0383, −0.0160]).

Changing Local to Cross covariance at fixed Cross consistency changes position/velocity MAE by +0.0015 m (95% CI [−0.0026, 0.0057]) and +0.0046 m/s (95% CI [0.0016, 0.0076]). Source swaps hold width fixed; these intervals capture trajectory variation conditional on one training realization.

We agree that relative calibration gains do not establish consistency. CD position NEES decreases from 1681.1 for F2 to 886.8 for F3 but remains far above one. We report absolute diagnostics and acknowledge persistent overconfidence.

The evidence remains simulation-specific despite added architectures and noise shifts. We appreciate your time and thoughtful scrutiny, which made the study more rigorous and its conclusions more precise.

I thank the authors for their thorough rebuttal and substantial additional experiments, which have addressed many of my original concerns. While I do think some limitations remain regarding the generalizability of the findings beyond the simulated setting (as in for real-world generalization), I believe the additional evidence and also the more carefully scoped conclusions have strengthened the paper. I will therefore raise my score.

AuthorsOpenReview ↗

Thank You for the Reassessment

We thank the reviewer for the thoughtful reassessment and positive recognition of our revisions. We appreciate your constructive feedback and will retain clear statements of the remaining limitations regarding real-world generalization.

Reviewer_wd8VOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work introduces a Local--Cross framework that separates modality-local temporal representations from cross-modal relational representations while keeping the prediction target and model capacity matched. Experimental results demonstrate that local representations improve covariance prediction.

Critical strengths

This study investigates the advantages and disadvantages of local representations versus cross-modal representations in various reliability prediction tasks. Rather than presupposing one method to be superior at the outset, the authors carefully demonstrate through experiments the respective application domains of each approach. Under the established benchmark in this paper, the experimental conclusions are sounding.

Critical weaknesses
  1. The author classifies the degradation situations into three common types: SCV, SCD, and CD. Noise processing of clean sensor data is carried out manually, where SCV is generated through scaling, and SCD is constructed through freezing, drift, and other processes. These two types of degradation designs are beneficial for the clarity of experimental analysis. However, relatively simple degradation designs cannot fully reflect the real sensor conditions in reality. Some complex situations such as time-related deviation and non-Gaussian errors will couple with the common degradation mentioned in this article, making the actual state estimation difficult.

2.The covariance estimation of sensors and the processing of abnormal measurements have been studied by many researchers in this field. Under the proposed perception degradation design in this paper, the effects of the adaptive strategies and more robust state estimation strategies proposed by others have not been compared with this framework. Although this paper have achieved certain performance improvements in the existing experimental design, more detailed comparative experiments are needed to demonstrate the value of this method.

Questions for authors
  1. Can the author further explain whether the degradation generation mechanisms of SCV and SCD are consistent with the actual situations of GPS, IMU and other sensors? Is the noise generated manually comparable in terms of accuracy and precision to the real-world noise? Are there any existing studies that show this abstract modeling form can accurately represent the noise of real sensors?
  2. The current noise modeling is relatively simple. Can the author prove that this method still has considerable gain in different actual noise distributions?
  3. Can the author introduce other classic adaptive strategies or more robust state estimation strategies under this degradation design? Through comparison, make the gain of this Framework more comparable.
3 comments
AuthorsOpenReview ↗

Response to Reviewer wd8V: Sensor Modeling, Noise Shifts, and Adaptive and Robust Baselines

We thank the reviewer for recognizing the value of comparing Local and Cross-Modal representations without assuming either is uniformly preferable. Your comments helped us connect the controlled benchmark more clearly to practical sensor errors and established estimation strategies.

1. Physical interpretation of SCV and SCD

We added Appendix A.1 to explain the physical motivation and scope of the degradation models. SCV varies prescribed process or measurement covariance. GPS biases and drifts represent sustained or gradually increasing errors; freezing represents repeated use of stale measurements. OF offsets and outliers represent errors in velocity observations supplied to the estimator.

The appendix discusses studies of inertial stochastic errors, GNSS uncertainty and multipath/NLOS effects, and OF measurement errors. These motivate the error categories, but do not establish that our exact amplitudes, multipliers, and durations reproduce a particular sensor’s accuracy or precision. That correspondence would require measured error statistics under specified operating conditions.

We also clarify that the simulated OF channel uses finite differences of Gazebo world-frame position to obtain planar velocity, plus a binary quality indicator. Perturbations are applied to this velocity observation. It is an abstraction of OF-derived velocity, rather than an image-based optical-flow pipeline.

Your question further prompted clarification of the targets: covariance supervision specifies background covariance, whereas consistency labels identify observations affected by the SCD injection protocol. Structured errors can themselves increase observed dispersion. These are distinct benchmark prediction tasks, not an exhaustive decomposition of physical sensor errors.

2. Correlated and non-Gaussian noise

We added evaluations under AR(1) noise with correlation coefficient 0.8, Student-t(3) noise, and Gaussian-mixture noise with infrequent high-variance samples. Gaussian-trained models are tested without retraining or fine-tuning.

The distributions preserve prescribed marginal variance at the population level while changing temporal dependence or distribution shape. Trajectories, SCV schedules, SCD events, and antithetic pairing remain fixed; SCD outlier generators are unchanged. Every distribution is evaluated under all four regimes (Appendix D).

We extended this test to the adaptive and robust baselines, keeping their validation-selected settings fixed. This examines comparative performance under the same unseen conditions.

F3 has lower mean position and velocity errors than B1–B3 in every tested SCD and CD setting across all four distributions (Figure 8). Gains are not universal: under Gaussian-mixture noise and SCV, its position MAE is 25.6% higher than B3’s. Appendix E.4 reports absolute errors and paired differences with 95% CIs.

Relative gains also coexist with worse absolute accuracy: F3’s CD velocity MAE increases from 0.1787 m/s under Gaussian noise to 0.2818 m/s under AR(1). We therefore claim comparative evidence under specified synthetic shifts, not stable accuracy under arbitrary real-world noise.

3. Adaptive and robust baselines

We added:

  • B1: innovation-adaptive measurement covariance with nominal process covariance.
  • B2: Huber-weighted updates with nominal process and measurement covariance.
  • B3: the same Huber update with learned Local covariance.

B3 extends the requested classical comparisons by testing whether learned covariance combined with conventional robust weighting accounts for the gains attributed to consistency modulation.

Hyperparameters are selected exclusively on validation data and fixed during testing. Appendix D reports equations, search ranges, and selection criteria. All configurations share the ESKF implementation, including the gravity-axis attitude-gain scaling described in Appendix C.

Under Gaussian CD:

MethodPosition MAE (m)Velocity MAE (m/s)
F30.58450.1787
B11.36960.3694
B20.89450.3700
B30.96410.3715

Paired differences with 95% CIs appear in Appendix E.4. We also report nominal/SCV cases where advantages are smaller or absent.

We appreciate your emphasis on practical relevance and the time invested in identifying these gaps. Your suggestions led to broader comparisons and a clearer account of both the evidence and its limits.

Thank you for the detailed responses and additional experiments. I appreciate the expanded noise evaluations and comparisons with adaptive and robust baselines. However, the new experiments still rely on synthetic noise models and do not sufficiently establish their correspondence to real-world sensor errors. The results also show that the performance gains are not consistent across all noise conditions. Therefore, my original concerns remain, and I will maintain my initial rating.

AuthorsOpenReview ↗

Clarification of Scope and Empirical Guidance

We thank the reviewer for examining the additional experiments and acknowledging the expanded evaluations. We agree that synthetic noise tests do not establish correspondence to real-world sensor errors, and we retain this limitation explicitly.

We would like to clarify that our goal is to provide empirical guidance for information selection in reliability prediction. Simulation enables controlled comparisons under shared trajectories and perturbations. The nonuniform gains help identify the conditions under which particular design choices are beneficial and the limits of those benefits. We believe these comparisons provide useful evidence within the studied setting, while real-world generalization remains to be established.

Thank you again for your time and thoughtful assessment.

Reviewer_1J2xOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper compares modality-local and cross-modal representations for covariance and measurement-consistency prediction in UAV state estimation. It constructs synthetic sensing degradations using optical flow, IMU, and GPS, and integrates the predictions into a common ESKF. Local models achieve lower covariance error, while cross-modal models improve consistency discrimination; combining these predictions improves downstream estimation under degraded sensing.

Critical strengths
  1. The opposing Local–Cross trends for covariance and consistency provide an interesting empirical finding: different representation choices perform best for the two reliability targets. This pattern holds across both tested capacity and degradation regimes
  2. The combination of local covariance and cross-modal consistency translates this finding into a concrete estimator design. Compared with Local consistency, it reduces both position and velocity error under SCD and compound degradation, connecting reliability prediction to downstream estimation performance.
Critical weaknesses
  1. Matching total parameters gives Cross-Modal narrower encoders: 49 versus 64 and 64 versus 84 (Appendix B). Small Local still outperforms Large Cross-Modal at h=64 (SCV MAE: 0.0720 versus 0.1127, Table 3), so width alone is not a sufficient explanation. However, the relational branches and prediction-head inputs also change. The experiments therefore establish which tested configuration performs better, without separating the effect of cross-modal evidence from capacity allocation and fusion architecture. This distinction matters because the paper’s central question concerns the information needed for each reliability target.

  2. All learned ESKF configurations use Local covariance (Table 1). The F3–F2 comparison supports Cross-Modal consistency prediction, but cannot establish whether Local covariance also yields better state estimation than Cross-Modal covariance. Thus, the downstream gains validate only part of the proposed representation-selection rule.

Questions for authors
  1. With local-encoder and prediction-head hidden widths fixed, how do Local, direct feature concatenation, and relational Cross-Modal models compare? Does the covariance gap persist when the Cross-Modal model retains an unchanged local prediction pathway and can ignore reference features?
  2. With Cross-Modal consistency fixed, how does replacing Local covariance with Cross-Modal covariance affect ESKF position and velocity errors? This would clarify whether the covariance-prediction ordering also holds downstream.
1 comment
AuthorsOpenReview ↗

Response to Reviewer 1J2x: Architecture Controls, Downstream Attribution, and Presentation

We thank the reviewer for recognizing the empirical findings and downstream value. Your questions helped us distinguish information-source effects from architectural choices, and prediction accuracy from estimation accuracy.

1. Parameter, width, and fusion controls

We agree that parameter matching alone cannot isolate cross-modal information from capacity allocation and fusion design. We now combine task-specific parameter matching, equal-width comparisons, and fusion controls, each using five training runs. The 95% CIs jointly account for training-run and trajectory variation (Section 5.3; Appendix D).

Matching counts parameters connected to the task loss. For consistency, Local/Cross counts are 120,307/121,103 (Small) and 198,343/197,123 (Large), differing by less than 1%. Local widths are 73/94 and Cross widths are 50/64. Cross retains higher aggregate AUPRC under CD at both budgets; SCD-only AUROC and AUPRC difference intervals include zero (Table 9).

With local-encoder and prediction-head hidden widths fixed at h=64, the covariance gap persists. Cross-minus-Local log-MAE is +0.0396 under SCV (95% CI [0.0304, 0.0485]) and +0.0410 under CD (95% CI [0.0288, 0.0524]). Thus, narrower encoders do not fully explain the original result (Table 4).

Following your question about an unchanged Local pathway, we compare Local, direct concatenation, relational fusion, and Preserved Local. The latter freezes the trained Local encoder/head and adds a cross-modal residual branch with zero-initialized outputs. It initially reproduces Local exactly; this initial checkpoint is eligible for validation-based selection.

At h=64, covariance log-MAE differences from Local are:

ArchitectureSCVCD
Concat+0.0425; 95% CI [0.0307, 0.0537]+0.0482; 95% CI [0.0347, 0.0605]
Relational+0.0396; 95% CI [0.0304, 0.0485]+0.0410; 95% CI [0.0288, 0.0524]
Preserved Local+0.0035; 95% CI [0.000581, 0.0073]+0.0017; 95% CI [−0.0017, 0.0055]

Preserving Local substantially narrows the gap, with the CD interval including zero. We therefore qualify the finding by architecture rather than attributing it to cross-modal information alone (Table 6).

We also extended these controls to consistency: all three variants improve AUPRC under CD. Under SCD-only, Preserved Local improves AUROC and AUPRC, while Concat and Relational intervals include zero (Table 12). These controls match width, not parameters or computation; Tables 8 and 14 report architecture sizes and training budgets, including Local pretraining.

2. Covariance-source effects downstream

Your proposed replacement led us to evaluate all four covariance–consistency source combinations, extending the analysis to conditional effects, main effects, and interactions (Section 5.2; Tables 17–22).

Under CD, replacing Local covariance with Cross covariance at fixed Cross consistency changes:

  • Position MAE: +0.0015 m; 95% CI [−0.0026, 0.0057].
  • Velocity MAE: +0.0046 m/s; 95% CI [0.0016, 0.0076].

Local covariance therefore yields lower velocity error, while the position effect remains uncertain.

Replacing Local consistency with Cross consistency at fixed Local covariance changes:

  • Position MAE: −0.0298 m; 95% CI [−0.0438, −0.0157].
  • Velocity MAE: −0.0271 m/s; 95% CI [−0.0383, −0.0160].

This provides fuller attribution than F3–F2 alone. Source swaps hold width fixed within a shared ESKF; their 95% CIs quantify trajectory variation conditional on one training realization. We now distinguish covariance-prediction advantages from their metric-dependent downstream effects.

3. Presentation

We took seriously your assessment that parts were difficult to follow. We replaced dense main-text tables with visual comparisons of performance, effect sizes, and uncertainty, retaining numerical tables in the appendix. We also added an experimental roadmap and clearer subsection headings.

We appreciate your time and care. Your comments led to stronger controls, more precise conclusions, and a clearer presentation of the evidence.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response: Summary of Revisions and Appreciation

We sincerely thank all three reviewers for their careful reading and constructive feedback. We appreciate your recognition of the information-selection question and the value of connecting reliability prediction to downstream state estimation. Your comments helped us identify gaps in the experimental controls, clarify the interpretation of the results, and improve the presentation. The revised manuscript is substantially stronger because of this engagement.

The main revisions are:

  1. Expanded representation controls and training variability. Parameter-matched, width-matched, and fusion-architecture comparisons now use five training runs, with 95% CIs accounting jointly for training-run and test-trajectory variation. Parameter matching counts parameters connected to the task loss. We also compare direct concatenation, relational fusion, and a preserved Local prediction pathway.

  2. Complete downstream source attribution. We extended the requested covariance-source replacement to all four Local/Cross-Modal covariance–consistency combinations in the same ESKF, allowing conditional effects, main effects, and interactions to be examined.

  3. Degradation-specific analyses for both prediction targets. Beyond the requested breakdown of consistency performance by SCD mechanism, we analyze covariance errors by prediction role under SCV and compound degradation. These analyses identify where aggregate findings hold and where exceptions arise.

  4. Adaptive and robust baselines under distribution shift. We added innovation-adaptive covariance estimation, Huber updates, and a hybrid learned-covariance/Huber baseline. All are evaluated alongside our method under Gaussian, temporally correlated AR(1), Student-t(3)t(3), and Gaussian-mixture noise, with model weights and validation-selected baseline settings held fixed.

  5. Clearer diagnostics and presentation. We report absolute NIS/NEES statistics alongside paired comparisons and clarify the physical motivation and limitations of the synthetic sensor models. We replaced dense main-text tables with visual comparisons, retained detailed numerical results in the appendix, and added an experimental roadmap and clearer subsection headings.

The additional evidence also refined our conclusions. Local covariance prediction retains lower error in the parameter- and width-matched comparisons, but preserving the complete Local pathway substantially reduces the fusion-related gap. Cross-Modal consistency improves aggregate AUPRC under compound degradation, while the SCD-only difference intervals include zero in the parameter- and width-matched comparisons. Downstream, Cross-Modal consistency improves position and velocity accuracy under compound degradation; covariance-source effects differ between the two metrics. We have revised the claims to reflect these distinctions and explicitly acknowledge residual position overconfidence and the scope of the simulation-based evidence.

Together, the revisions provide a more complete empirical reference for selecting information sources in learned reliability prediction. Our individual responses explain how each concern was addressed and where the corresponding evidence appears in the revised manuscript.

Regardless of the final decision, we are grateful for the time, effort, and care that the reviewers have invested in this work. We value the reciprocal nature of peer review and sincerely appreciate your contributions to making the paper more rigorous, clearer, and more useful to other researchers.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.