acceptodds
Under review as a conference paper at ICLR 2027

When Does Learned Inertial Odometry Fail?

Abstract

Learned inertial odometry estimates velocity directly from IMU sequences and reduces the drift caused by inertial integration, but its cross-domain generalization remains limited. Existing studies mainly focus on improving target-domain accuracy, while aggregate metrics provide little insight into why failures occur. We analyze target-domain failures by relating them to source-domain motion coverage, source prediction error, and model internal states. Our results show that source motion coverage alone does not determine prediction failure. Some failures occur under motion conditions that are rare in the source domain, while others arise under source-covered conditions near source low-error motion regions. For the latter, failed predictions tend to remain biased toward the source-domain velocity associated with similar motion. This indicates that motion–velocity relations learned from source data may become invalid after domain shift. We further show that different failure types favor different target-domain adaptation data, and that motion and internal-state information provide complementary cues for failure detection. These results link the data regularities that enable direct velocity prediction to the cross-domain failures of learned inertial odometry.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Deep inertial odometry, Distribution shift, Motion-conditioned analysis, Learned representations, Failure diagnosis

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

This paper looks at the generalisation of learnt inertial odometry models and explores the sources of error when training a model using a source domain and then applying it to a target domain. It provides interesting insights into the sources of error, shows that the target errors are observed not only in high error source areas, but also in low error source areas with similar motion but different velocities. It also provides ideas on how different types of target domain adaptation data could be used to reduce different types of error.

Critical strengths
  • Critical review into cross domain failures in the context of data-driven inertial odometry estimation

  • Could inspire similar reviews in odometry systems using other modalities

  • A mixture of obvious (expected) and non-obvious (interesting) results e.g. ---- for example an expected result is that when you have a high error region in the source domain you expect high errors in the same region of the target domain ----- less obvious result that in low error motion regions in the source domain, sometimes one still sees high error motion regions in the target domain, and the explanation is the difference in velocities in that region

  • evaluation across multiple datasets spanning cars, drones, humans and animals

  • demonstration of recovery rates by calibration data

Critical weaknesses

In Table 20- the recovery gains from calibration data depend on the evaluated category, with some seeing benefits and others deterioration. The approach to recovery using calibration data is interesting but does not seem to guarantee performance improvements in a reliable way.

Questions for authors

In Table 20- the recovery gains from calibration data depend on the evaluated category, with some seeing benefits and others deterioration. If this is the case is there a good approach to add additional sessions from the target domain without risking deterioration in regions / categories that have low errors?

Is the analysis done for inertial odometry easy to expand and use in other types of odometry?

Can one use domain knowledge on the shift between source and target domains to improve failure detection?

5 comments
AuthorsOpenReview ↗

A1: Non-failure samples mitigate calibration deterioration at a cost to failure recovery

We agree that readout calibration degrades predictions for some originally non-failed target windows. But including non-failure windows alleviated this deterioration.

Mean absolute velocity error of originally non-failed target windows increased from 0.1172m/s0.1172 m/s before calibration to 0.5730–0.7708m/s0.5730–0.7708 m/s under failure-only calibration, as shown by the 0% NF-share rows below. Then we kept the total calibration budget fixed and replaced 50% of the calibration windows with original non-failure windows sampled to balance motion regions. We compared C1 (source-sparse failures), C2 (source-covered failures in source high-error regions), C3 (source-covered failures in source low-error regions), and a mixture of all three.

Effect of including non-failure calibration windows on OxIOD

\hline \text{Failure pool} & \text{NF share (\\%)} & \text{Overall EPE (m/s)} & \text{NF EPE (m/s)} & \text{New failures (\\%)} \\\\ \hline \mathrm{C1} & 0 & 0.5751 & 0.7708 & 88.11 \\\\ \mathrm{C1} & 50 & 0.5594 & 0.4142 & 69.92 \\\\ \hline \mathrm{C2} & 0 & 0.4543 & 0.6802 & 84.45 \\\\ \mathrm{C2} & 50 & 0.5938 & 0.3567 & 38.88 \\\\ \hline \mathrm{C3} & 0 & 0.5438 & 0.7185 & 85.02 \\\\ \mathrm{C3} & 50 & 0.5436 & 0.3054 & 41.67 \\\\ \hline \text{Mixed} & 0 & 0.4357 & 0.5730 & 65.09 \\\\ \text{Mixed} & 50 & 0.5379 & 0.3286 & 43.77 \\\\ \hline \end{array}$$ Footnote:NF denotes windows that were non-failures before calibration. NF share is their proportion in the calibration set; Mixed uses C1, C2, and C3. NF calibration windows are sampled with approximately equal motion-region quotas, capped by availability and without replacement. Each condition uses 1,600 calibration windows per fold and the same held-out sessions. Results are means over 10 sampling repetitions. Overall EPE is evaluated on all \$32,866 \$ test windows. NF EPE and new-failure percentages are evaluated on the \$1,754 \$ original NF test windows. Before calibration, overall EPE was 0.8136 m/s and NF EPE was \$0.1172 m/s \$. The result indicates that including non-failure windows reduced both the mean error and new-failure rate among original non-failures across all four groups. However, recovery of original failures declined, and overall error increased for C2 and mixed calibration. **These results show a trade-off between preserving low-error predictions and correcting failures. Including non-failure windows mitigates deterioration but does not restore their pre-calibration error levels.**
AuthorsOpenReview ↗

A2: The framework could extend to other type of odometry by adapting the interpretable features and failure criteria to their inputs and outputs.

An important aspect of our approach is to use physically meaningful and intuitively understandable motion features to relate model failures to observable motion conditions, informing failure-risk assessment in practical use.

For extensions to visual and LiDAR odometry, similarly observable and interpretable features, such as scene appearance and geometric structure, could be used to examine when models are prone to error and whether large errors can still occur under conditions common in the source domain. Failure criteria would also need to be adapted to outputs such as pose and displacement.

This could support a similar failure-analysis framework for identifying high-risk conditions and assessing a model’s range of applicability.

AuthorsOpenReview ↗

A3: Domain knowledge provides complementary cues for failure detection

Domain knowledge can be expressed through measurable motion descriptors to identify conditions requiring closer monitoring. We design five features based on IMU kinematics, describing rotational intensity, angular-velocity changes, and acceleration fluctuations (Section 2.1, Eqs. (1)–(2)). Sections 3.1 and 3.2 use these features to analyze source prediction difficulties and their cross-domain behavior. For Car and Human, some high-error regions involve rapid changes in angular velocity and acceleration; for Dog, the studied high-error region exhibits strong rotation. Drone shows that prediction error does not always increase with motion intensity. These findings connect prediction risk to specific motion conditions that warrant attention even within the source domain.

For target-domain failure detection, Section 4.2 compares motion features, internal model features, and their combination. Table 6 shows that combining them achieves higher AUROC than either feature set alone on KITTI, OxIOD, and QuadSLAM. For example, AUROC on OxIOD increases from 0.803 with internal features alone to 0.871 with the combined detector. This demonstrates that descriptors based on motion knowledge can complement internal model responses and improve failure detection.

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

This paper studies cross-domain failures in learned inertial odometry. I think the question of when these leaned inertial predictions fed to the EKF/Facto Graphs for inertial odometry fail is an important, valuable and interesting question. It represents IMU windows using five motion descriptors, uses a Gaussian KDE to determine source coverage, and defines failures using a source-validation error threshold. It then analyzes failures under source-sparse and source-covered motion, studies whether velocity remains available in the learned representation, and evaluates failure detection using motion and internal model features.

Critical strengths
  1. The paper asks an important question: aggregate target-domain performance does not explain when or why learned inertial odometry fails.

  2. The paper attempts to analyze failures using source coverage, source prediction behavior, and internal model features rather than reporting only aggregate errors. I think it is a good initial attempt at this hard problem.

Critical weaknesses

Presentation:

The problem is clear after careful reading, but important terms such as SourceTrain, LiLY, densMAP, Gaussian KDE, and Scott’s bandwidth should be introduced more clearly. Table 4 is difficult to understand without reconstructing the categories from the text. Figure 2 has labeling and color inconsistencies, and line 313 contains an unresolved figure reference. More detailed captions would improve readability, by enabling the reader to understand the main results by reading through the tables and understanding the figures.

The approach taken:

  • The definitions of failure and motion coverage are not sufficiently well calibrated.
  • The “relative velocity error” behaves like absolute error below 1 m/s and relative error above it. Equation 4 defines failure as a relative tail event: a small source-error distribution can still produce “failures,” while a poor source model can have large errors below its threshold.
  • The use of model- and carrier-specific thresholds also makes comparisons difficult.
  • It is also unclear whether the five descriptors measure physical motion shift or changes in sensor noise, bias, frame, sampling rate, filtering, and resampling.
  • The temporal-difference features may amplify noise and interpolation effects.
  • The KDE coverage score therefore establishes coverage only in the chosen feature space, not necessarily physical motion coverage.
  • Speed alone is not sufficient to characterize source coverage.
  • The region partition, above-median high-error definition, and matching thresholds appear arbitrary, and “motion matching” relies on five lossy summary statistics.
  • Absolute velocity is also fundamentally ambiguous from a finite IMU window.
  • Section 4 additionally requires target ground-truth velocities to construct failure labels. It therefore detects failures in held-out sessions from an already labeled target dataset, rather than in a genuinely unseen and unlabeled domain.
  • Temporal correlation, sequence length, and other possible confounders require more careful sensitivity analysis.
  • We lack understanding of the confounders in the analysis. Even if the paper presented the case for why this problem is so hard to tackle rigorous it would have been a good contribution.
Questions for authors
  1. How sensitive are the conclusions to the five motion descriptors, KDE bandwidth, feature scaling, coverage threshold, regional partition, sampling rate, sensor noise, and sequence length?

  2. Why is the source-validation 90th percentile an appropriate failure threshold? Do the conclusions remain valid under fixed absolute-error thresholds or continuous-error analysis, and how would this definition extend to methods that predict displacement or IMU biases instead of velocity?

  3. Can the failure detector operate without labeled target velocities? If not, should it be described as supervised failure detection within a known target domain rather than cross-domain failure detection?

    1. For curiosity, what do the authors think about algorithms that have symmetry based architectures like RIO, EqNIO and Lie-event preprocessing being analysed under this current approach? In particular, could some effects identified here as physical motion shift instead arise from the fixed-rate representation or resampling procedure?
  4. Address all the weaknesses mentioned above.

18 comments
AuthorsOpenReview ↗

Clarifying the Analysis Framework and Strengthening the Presentation

We thank all reviewers for their comments on the soundness of our work. Your feedback highlighted that the original manuscript did not clearly connect the analysis framework, experimental evidence, and main conclusions. We have therefore reorganized the paper and revised its definitions, terminology, and presentation to make the reasoning easier to follow.

The revised paper proceeds from defining the analysis framework to studying source-domain performance, analyzing cross-domain failures, and evaluating practical applications. Section 2 defines the motion features used to describe IMU inputs, the velocity-error criterion for prediction failure, and source coverage as a measure of support from similar source motions. Section 3.1 then partitions the motion-feature space and compares regional prediction errors across source training and held-out validation trajectories. This establishes which prediction difficulties already exist within the source domain. Section 3.2 extends the analysis to target data, focusing on failures despite source support and non-failures under sparse source support. For covered failures, we assess whether they occur in source high-error regions and whether similar motion features correspond to different ground-truth speeds across domains. For sparse non-failures, we analyze ground-truth speeds and absolute prediction errors. Section 4 evaluates whether these findings can guide target-domain adaptation and help detect unreliable predictions.

We have also revised the research questions and contributions in the Introduction to align with this structure and the experimental findings. Transitions between Sections 2 and 3 now clarify the question addressed by each analysis. Detailed data-processing, matching, and evaluation procedures are provided in the appendix to support verification and reproducibility. We have checked regional definitions, coverage groups, and associated statistics across the main text and appendix, and revised the discussion to distinguish empirical observations from possible explanations. These changes make the evidence supporting each conclusion and its scope more explicit.

We have attached the revised manuscript and appendix for your review.

AuthorsOpenReview ↗

A1.1: Sensitivity to the Motion Representation

We kept the model predictions and failure labels fixed, varied the KDE bandwidth, feature scaling, and coverage threshold, and compared the failure rates of source-covered and source-sparse windows across settings.

1. Sensitivity to the Choice of Motion Representation

To assess how the three conclusions in the introduction depend on the choice of motion descriptors, we replaced the original five descriptors with LiLY’s frozen 20-dimensional representation. We revisited the three parts of the analysis in turn: source coverage, prediction difficulty in the source domain, and failure recovery.

(1) Source coverage. We compared the coverage results obtained with the five motion descriptors and the 20-dimensional LiLY representation, using the same source reference windows and KDE procedure for both.

Table1. Prediction outcomes by source coverage under two representations

\hline & \text{Covered failures} & \text{Covered failures} & \text{Sparse non-failures} & \text{Sparse non-failures} \\\\ \text{Target dataset} & \text{5D} & \text{LiLY 20D} & \text{5D} & \text{LiLY 20D} \\\\ \hline \text{KITTI} & \text{3,394/4,006 (84.72\\%)} & \text{2,545/3,149 (80.82\\%)} & \text{44/226 (19.47\\%)} & \text{52/1,083 (4.80\\%)} \\\\ \text{KAIST} & \text{1/385 (0.26\\%)} & \text{8/678 (1.18\\%)} & \text{286/293 (97.61\\%)} & \text{—} \\\\ \text{OxIOD} & \text{22,816/24,352 (93.69\\%)} & \text{14,983/15,940 (94.00\\%)} & \text{218/8,514 (2.56\\%)} & \text{797/16,926 (4.71\\%)} \\\\ \text{QuadSLAM} & \text{257/471 (54.56\\%)} & \text{540/837 (64.52\\%)} & \text{177/697 (25.39\\%)} & \text{94/331 (28.40\\%)} \\\\ \hline \end{array}$$ Under both representations, KITTI, OxIOD, and QuadSLAM contain failed source-covered windows and non-failed source-sparse windows. The first conclusion therefore remains supported. **(2) Prediction difficulty.** We replace them with LiLY's frozen 20-dimensional representation. High- and low-error groups are redefined using the prediction errors of neighboring source training samples. This tests whether prediction difficulties identified in the training data remain observable on validation data after changing the representation. The results are reported below. **Table 2. High- and low-error groups in the LiLY 20D representation** $$\begin{array}{lllrrrrr} \hline \text{Carrier} & \text{Model} & \text{Type} & e_{\mathrm{ref}} & \text{Train }N & \text{Train median }e & \text{Val }N & \text{Val median }e \\\\ \hline \text{Car} & \text{LiLY} & \text{High} & 0.1011 & 11{,}011 & 0.1705 & 3{,}062 & 0.2020 \\\\ & & \text{Low} & 0.1011 & 12{,}522 & 0.0675 & 4{,}286 & 0.0755 \\\\ \hline \text{Drone} & \text{LiLY} & \text{High} & 0.1485 & 7{,}257 & 0.2309 & 1{,}436 & 0.4106 \\\\ & & \text{Low} & 0.1485 & 6{,}956 & 0.1050 & 891 & 0.3003 \\\\ \hline \text{Human} & \text{LiLY} & \text{High} & 0.0762 & 9{,}341 & 0.2123 & 2{,}719 & 0.1989 \\\\ & & \text{Low} & 0.0762 & 19{,}929 & 0.0343 & 5{,}019 & 0.0264 \\\\ \hline \text{Dog} & \text{LiLY} & \text{High} & 0.0591 & 7{,}333 & 0.1664 & 3{,}743 & 0.1309 \\\\ & & \text{Low} & 0.0591 & 7{,}582 & 0.0068 & 2{,}558 & 0.0221 \\\\ \hline \end{array}$$ **Using LiLY’s frozen 20D representation, we also found higher median errors in the source-defined high-error group on all four held-out source validation sets**, as do the five-dimensional motion features. That means the second conclusion remains valid. **(3) Failure recovery.** We redefine source coverage and high- and low-error groups using LiLY's frozen 20-dimensional representation and repeat the recovery experiment on OxIOD. **Table3. OxIOD failure recovery using groups defined in the LiLY 20D representation** $$\begin{array}{lrr} \hline \text{Calibration sample selection} & \text{C1 recovery (\\%)} & \text{C2 recovery (\\%)} \\\\ \hline \text{C1 failures} & \mathbf{41.56} & 51.53 \\\\ \text{C2 failures} & 1.74 & \mathbf{94.97} \\\\ \text{Random failure windows} & 12.94 & 90.96 \\\\ \text{Random target windows} & 12.91 & 90.11 \\\\ \hline \end{array}$$ After regrouping with the 20-dimensional representation, the choice of calibration samples still affects recovery. Both evaluated failure groups achieve their highest recovery rates when calibrated using failures from the same group. **These results support the conclusion that calibration sample selection affects which failures are recovered**. These comparisons show that covered failures, sparse non-failures, and source-domain differences in prediction difficulty also occur with a learned representation.
AuthorsOpenReview ↗

A1.2: Sensitivity to KDE bandwidth (1)

(1)The KDE bandwidth affects the estimation of source-domain coverage and may therefore affect Conclusion 1. To evaluate this sensitivity, we vary the bandwidth (h) from 0.1 to 0.8 and compare source-domain coverage with absolute velocity error following the procedure used in the lower panel of Figure 1. The results are reported in the following table.

Table4. Error statistics across KDE bandwidths and source-density bins

\hline h & \text{B1} & \text{B2} & \text{B3} & \text{B4} & \text{B5} & \text{B6} & \text{B7} & \text{B8} \\\\ \hline \text{Car / KITTI} & & & & & & & & \\\\ 0.1 & 4.4541 & 5.5597 & 7.0719 & 7.9384 & 7.2700 & 8.3840 & 9.1547 & 8.6612 \\\\ 0.2 & 4.1492 & 5.0825 & 7.3673 & 8.2705 & 9.0240 & 6.9888 & 8.4904 & 8.7328 \\\\ 0.4 & 4.5574 & 7.5572 & 7.8166 & 7.8948 & 9.1714 & 9.1517 & 8.6649 & 9.0097 \\\\ 0.8 & 5.7095 & 7.2285 & 9.2791 & \text{--} & \text{--} & \text{--} & \text{--} & \text{--} \\\\ \hline \text{Drone / KAIST} & & & & & & & & \\\\ 0.1 & 0.2663 & 0.4046 & 0.3184 & 0.3784 & 0.0002 & 0.0002 & 0.0282 & 0.0002 \\\\ 0.2 & 0.2111 & 0.4076 & 0.4188 & 0.2781 & 0.3911 & 0.2574 & 0.1582 & 0.0002 \\\\ 0.4 & 0.1632 & 0.1947 & 0.3324 & 0.3700 & 0.2763 & 0.0001 & 0.0002 & 0.0013 \\\\ 0.8 & 0.0912 & 0.1820 & 0.2736 & 0.1990 & 0.0002 & \text{--} & \text{--} & \text{--} \\\\ \hline \text{Human / OxIOD} & & & & & & & & \\\\ 0.1 & 0.8863 & 0.8356 & 0.8011 & 0.8075 & 0.8173 & 0.8398 & 0.8309 & 0.7497 \\\\ 0.2 & 0.8927 & 0.8567 & 0.8177 & 0.8253 & 0.8354 & 0.8384 & 0.7979 & 0.7355 \\\\ 0.4 & 0.8984 & 0.8571 & 0.8339 & 0.8357 & 0.8326 & 0.8100 & 0.7568 & 0.7332 \\\\ 0.8 & 0.8957 & 0.8453 & 0.8358 & 0.8302 & 0.8024 & 0.7392 & \text{--} & \text{--} \\\\ \hline \text{Dog / QuadSLAM} & & & & & & & & \\\\ 0.1 & 0.7142 & 0.6117 & 0.5466 & 0.5919 & 0.4855 & 0.5392 & 0.4261 & 0.0963 \\\\ 0.2 & 0.7103 & 0.7142 & 0.6969 & 0.5467 & 0.6110 & 0.5503 & 0.4664 & 0.0986 \\\\ 0.4 & 0.6160 & 0.7206 & 0.7095 & 0.7060 & 0.5606 & 0.6367 & 0.4907 & 0.0523 \\\\ 0.8 & 0.2966 & 0.6544 & 0.7296 & 0.7147 & 0.7060 & \text{--} & \text{--} & \text{--} \\\\ \hline \end{array}$$ Footnote: $�0�–�1�$ represent increasing levels of coverage density. **The overall trend is consistent with the lower panel of Figure 1. Conclusion 1 therefore remains valid.** (2)**The KDE bandwidth also affects how samples are assigned to the source-covered and source-sparse groups.** We fix the coverage threshold at $Q_{95}$ and compare the group sizes across bandwidths, along with their membership overlap with the assignments obtained at $h=0.4$. **Table5. Stability of source-coverage assignments under different KDE bandwidths** $$\begin{array}{lrrrrr} \hline \text{Target} & h & \text{Covered }N & \text{Sparse }N & \text{Covered Jaccard (\\%)} & \text{Sparse Jaccard (\\%)} \\\\ \hline \text{KITTI} & 0.1 & 3{,}459 & 773 & 94.25 & 72.70 \\\\ & 0.2 & 3{,}787 & 445 & 96.34 & 75.74 \\\\ & 0.4 & 3{,}670 & 562 & 100.00 & 100.00 \\\\ & 0.8 & 4{,}161 & 71 & 88.20 & 12.63 \\\\ \hline \text{KAIST} & 0.1 & 278 & 400 & 58.65 & 51.00 \\\\ & 0.2 & 341 & 337 & 71.94 & 60.53 \\\\ & 0.4 & 474 & 204 & 100.00 & 100.00 \\\\ & 0.8 & 642 & 36 & 73.83 & 17.65 \\\\ \hline \text{OxIOD} & 0.1 & 18{,}312 & 14{,}554 & 71.75 & 50.51 \\\\ & 0.2 & 23{,}254 & 9{,}612 & 90.96 & 76.06 \\\\ & 0.4 & 25{,}504 & 7{,}362 & 100.00 & 100.00 \\\\ & 0.8 & 27{,}431 & 5{,}435 & 92.97 & 73.80 \\\\ \hline \text{QuadSLAM} & 0.1 & 168 & 1{,}000 & 26.26 & 53.15 \\\\ & 0.2 & 297 & 871 & 46.77 & 61.19 \\\\ & 0.4 & 635 & 533 & 100.00 & 100.00 \\\\ & 0.8 & 995 & 173 & 63.65 & 32.21 \\\\ \hline \end{array}$$ The effect of bandwidth on group membership varies across datasets. Most settings remain close to the baseline, while some produce clear changes.
Reviewer_jhJWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates cross-domain failures in learned inertial odometry using source motion coverage, source regional prediction errors, and model internal representations. It introduces five motion descriptors to distinguish source-sparse target failures from failures occurring under source-covered conditions. For a subset of covered failures, predictions are substantially closer to matched source velocities than to target ground truth. The paper further studies category-dependent recovery through frozen-feature readout adaptation and supervised failure detection using motion and internal features. Experiments cover four carrier–target settings and two models, with readout adaptation evaluated on OxIOD.

Critical strengths
  1. A useful diagnostic perspective beyond aggregate odometry errors. The distinction between source-sparse failures, source-covered high-error regions, and residual covered failures helps identify heterogeneous failure behavior. In particular, Table 4 shows markedly different compositions across targets, motivating more targeted analysis than a single domain-level performance number.

  2. Session-separated evaluation and informative adaptation trade-offs. The readout experiments hold out complete sessions, and the failure detectors use session-level evaluation with bootstrap intervals. Appendix E.3 also explicitly reports newly introduced failures and deterioration outside the favored category. These results provide valuable evidence about the limitations of targeted adaptation, rather than reporting recovery alone.

Critical weaknesses
  1. The motion-matching analysis does not sufficiently distinguish a changed predictive relationship from information discarded by the descriptors. The five descriptors summarize intensity, temporal variation, and rotational composition, but omit substantial directional and temporal information available to the odometry models. Consequently, similar motion descriptors do not establish similar model inputs or recurrent contexts. A source–target velocity difference conditional on these descriptors could arise even if the relationship between richer IMU inputs and velocity remains unchanged.

Moreover, Tables 5 and 17 analyze samples already selected for prediction failure and source–target speed mismatch. This selection does not mathematically guarantee source proximity, but it makes appropriate controls essential. Comparisons against a global source velocity reference, shuffled neighbors, and matched non-failures would help establish whether the observed proximity specifically reflects a learned local motion–velocity relationship. Without these controls, the proposed mechanism remains plausible but insufficiently isolated.

  1. The evaluation protocol and adaptation evidence leave important alternative explanations unresolved. Appendix B.1 states that SourceTrain regional errors use one-second windows with model state reset for each window, whereas SourceVal uses ten-second contexts. For a recurrent model, this changes the effective information available for prediction. Since source regional errors determine the failure taxonomy, it is important to establish that the categories remain stable under identical inference contexts and to specify the target protocol precisely.

The readout results also require a narrower interpretation. In Table 20, low-error-region-rich calibration increases LiLY’s median errors in C1 and C2 by 311.1 and 393.7 percentage points and converts 84.9% of original non-failures into failures. These effects do not invalidate category-specific recovery, but they substantially limit its practical implication. Furthermore, selecting entire sessions enriched in a category can also select different speed distributions or recording conditions. Repeated, controlled calibration selections are needed to separate failure-category effects from session composition.

Questions for authors
  1. Does the source-proximity result persist with richer directional and temporal descriptors, or matching that accounts for the model’s effective input context? How does it compare with global-source and shuffled-neighbor reference baselines?

  2. What are the exact inference-context and hidden-state-reset protocols for each target dataset? When SourceTrain, SourceVal, and Target use identical protocols, how much do the regional labels and Table 4 decomposition change?

  3. Under repeated calibration-session selections, do the category-specific recovery advantages remain after controlling for speed and recording conditions? Please report variability and overall held-out error alongside recovery and newly introduced failures.

  4. How sensitive are the central findings to the failure threshold? Table 10 gives substantially different source-relative thresholds across carriers—for example, 87.23% for LiLY–Drone versus 36.24% for LiLY–Human. Do the qualitative conclusions persist with several source quantiles and a common absolute-EPE criterion?

6 comments
AuthorsOpenReview ↗

A1: Source-Proximity Analysis

First, we replace the five hand-crafted motion features with LiLY’s learned 2020-dimensional internal representation and repeat the source-coverage analysis. Failures remain present in source-covered regions, while some samples in source-sparse regions remain correctly predicted. The results are reported in Tables 11–33. We also repeat the analysis using a 1717-dimensional descriptor set that explicitly includes directional and temporal features, obtaining the same main conclusions.

Table 1. Prediction outcomes by source coverage under two representations

\hline \text{Target} & \text{Covered failure (5D)} & \text{Covered failure (20D)} & \text{Sparse non-failure (5D)} & \text{Sparse non-failure (20D)} \\\\ \hline \text{KITTI} & 3{,}394/4{,}006\\;(84.72\\%) & 2{,}545/3{,}149\\;(80.82\\%) & 44/226\\;(19.47\\%) & 52/1{,}083\\;(4.80\\%) \\\\ \text{KAIST} & 1/385\\;(0.26\\%) & 8/678\\;(1.18\\%) & 286/293\\;(97.61\\%) & \text{--} \\\\ \text{OxIOD} & 22{,}816/24{,}352\\;(93.69\\%) & 14{,}983/15{,}940\\;(94.00\\%) & 218/8{,}514\\;(2.56\\%) & 797/16{,}926\\;(4.71\\%) \\\\ \text{QuadSLAM} & 257/471\\;(54.56\\%) & 540/837\\;(64.52\\%) & 177/697\\;(25.39\\%) & 94/331\\;(28.40\\%) \\\\ \hline \end{array}$$ **Table 2. Target-domain errors by source-defined difficulty in the LiLY 20D representation** $$\begin{array}{lrrrr} \hline \text{Target dataset} & \text{High-error group }(N) & \text{Median error (\\%)} & \text{Low-error group }(N) & \text{Median error (\\%)} \\\\ \hline \text{KITTI} & 1{,}434 & 99.65 & 2{,}798 & 100.00 \\\\ \text{KAIST} & 264 & 28.46 & 414 & 0.11 \\\\ \text{OxIOD} & 16{,}514 & 83.10 & 16{,}352 & 83.69 \\\\ \text{QuadSLAM} & 988 & 58.26 & 180 & 37.27 \\\\ \hline \end{array}$$ **Table 3. OxIOD failure recovery using groups defined in the LiLY 20D representation** $$\begin{array}{lrr} \hline \text{Calibration sample selection} & \text{C1 recovery (\\%)} & \text{C2 recovery (\\%)} \\\\ \hline \text{C1 failures} & \mathbf{41.56} & 51.53 \\\\ \text{C2 failures} & 1.74 & \mathbf{94.97} \\\\ \text{Random failure windows} & 12.94 & 90.96 \\\\ \text{Random target windows} & 12.91 & 90.11 \\\\ \hline \end{array}$$ The higher-dimensional representation changes sample distances, coverage assignments, and numerical results, but the main conclusions remain unchanged. **This supports the five-dimensional motion features as a simplified analysis tool that provides intuitive and practical evidence at lower computational cost.** The choice of calibration data still affects recovery, and refitting only the velocity prediction layer while freezing the feature extractor still corrects some failures. **These recovery results indicate that the frozen representation retains information useful for target-domain velocity prediction.** Regarding the model’s effective input context, we checked both the architecture and numerical outputs of the evaluated LiLY checkpoint. Its deterministic velocity prediction depends only on the current IMU window. Grouping the same windows individually or into ten-window sequences produces identical velocity predictions. The matching window therefore covers the input context used for this prediction. Second, we test whether these failures arise because the model follows the velocity relationship of similar source samples. We compare the local motion-matched source reference with the coordinate-wise median velocity of the full source pool, the corresponding median restricted to source low-error regions, and shuffled local reference vectors. The shuffling permutes the local reference vectors across the same target samples over $100$ repetitions; it does not reconstruct or shuffle the neighbor sets. **Table 4. Predictions closer to the source reference than to the target ground truth** $$\begin{array}{lrr} \hline \text{Source reference} & \text{KITTI (\\%)} & \text{OxIOD (\\%)} \\\\ \hline \text{Local motion-matched} & 100.00 & 95.15 \\\\ \text{Global source} & 100.00 & 95.42 \\\\ \text{Global low-error source} & 100.00 & 95.32 \\\\ \text{Shuffled local references} & 100.00 & 95.22 \\\\ \hline \end{array}$$ Replacing the local source reference with a global reference, or shuffling the local references, leaves the proportion of predictions closer to the source reference nearly unchanged. **The current results therefore do not show that covered failures are caused by predictions favoring the velocities of specific similar source samples.** We have narrowed the causal interpretation in the revised paper accordingly.
AuthorsOpenReview ↗

A2: Window Packing Leaves Predictions and Failure Decomposition Unchanged

We have clarified the “LiLY inference protocol” in Appendix B.1B.1 (Data and evaluation windows). The frozen LiLY model predicts velocity independently for each 11-second IMU window. SourceTrain, SourceVal, and all four target datasets use the same effective input length, with no recurrent hidden state passed between windows. The “1010-second contexts” in the original text refer to packing 1010 one-second windows together and do not provide additional historical information.

We verified that single-window inference, ten-window packing, and window-order permutation produce identical velocity outputs, with a maximum absolute difference of 00 in the checked inputs. Changing the packing protocol therefore leaves the predictions, regional labels, and Table 44 failure decomposition unchanged.

AuthorsOpenReview ↗

A3: Category-Specific Recovery Gains Persist After Controlling for Recording and Speed, with Overall Performance Trade-offs

We add a category-specific recovery experiment and update Figure 33 in the main paper with its results. The table below reports the recovery rates shown in Figure 33, together with their sampling standard deviations. Table 55 in Section 44 reports the overall test error and newly introduced failures, allowing the within-class recovery gains to be considered alongside their costs. Table 5. Fine-tuning outcomes

\hline \text{Fine-tuning data} & \text{New failures} & \text{Recovered failures} & \text{Mean absolute velocity error change (m/s)} \\\\ \hline \text{C1} & 1{,}542 & 20{,}971 & -0.2408 \\\\ \text{C2} & 1{,}474 & 17{,}340 & -0.3531 \\\\ \text{C3} & 1{,}492 & \mathbf{21{,}079} & -0.2729 \\\\ \hline \text{Random failures} & 1{,}142 & 16{,}220 & \mathbf{-0.3846} \\\\ \text{Random target windows} & \mathbf{1{,}076} & 15{,}581 & -0.3762 \\\\ \hline \end{array}$$ **Table 6. Recovery rates across fine-tuning and test failure categories** $$\begin{array}{lrrr} \hline \text{Fine-tuning data} & \text{C1 recovery (\\%)} & \text{C2 recovery (\\%)} & \text{C3 recovery (\\%)} \\\\ \hline \text{C1} & \mathbf{88.6} & 36.0 & 92.6 \\\\ \text{C2} & 30.3 & 48.4 & 84.9 \\\\ \text{C3} & 32.4 & \mathbf{62.9} & \mathbf{98.0} \\\\ \hline \end{array}$$ After matching recording sessions and speed, class-specific sampling still achieves higher within-class recovery than random failure sampling. However, it also introduces more new failures and produces a higher overall error. **The within-class recovery advantage therefore remains after controlling for these factors, supporting the use of failure categories to guide fine-tuning data selection.**

Comments on the paper

7 comments
AuthorsOpenReview ↗

A1: Source Coverage Is Unchanged

(1) Source coverage is independent of the prediction-error definition.

Source coverage depends only on the five motion descriptors and their source-domain density, as defined in Section 2.32.3, Eqs. (5)(5)–(6)(6). With the descriptors and density estimator fixed, changing the failure threshold, velocity-normalization floor, or error definition leaves the source-covered and source-sparse assignments unchanged.

(2) Source high- and low-error regions depend on continuous relative velocity errors rather than the failure threshold.

These regions are defined using regional median relative velocity errors in Section 3.13.1, Eq. (7)(7), with the error defined in Section 2.22.2, Eq. (3)(3). Changing only the failure threshold therefore leaves the region labels unchanged, whereas changing the normalization floor may alter them. We recompute the labels under different normalization floors and measure whether source windows retain their original high- or low-error assignments.

The table reports label agreement (%)(\%) for LiLY’s source-training windows in confirmed regions, relative to the original floor of 1 m/s1\,\mathrm{m/s}. All windows receive equal weight, and unconfirmed regions are excluded.Table1. Failure-label agreement under different velocity floors

\hline \text{Carrier} & \text{Windows} & \text{Floor 0.1 m/s (\\%)} & \text{Floor 0.5 m/s (\\%)} & \text{Floor 2 m/s (\\%)} \\\\ \hline \text{Car} & 22{,}600 & 53.00 & 78.44 & 100.00 \\\\ \text{Drone} & 13{,}520 & 92.37 & 95.98 & 79.32 \\\\ \text{Human} & 28{,}731 & 86.59 & 99.48 & 100.00 \\\\ \text{Dog} & 14{,}548 & 88.66 & 89.09 & 100.00 \\\\ \hline \end{array}$$ High/Low labels remain unchanged for all carriers when only the failure threshold varies. Under alternative normalization floors, most settings preserve the original labels for most windows. However, smaller floors can produce substantial changes, as illustrated by the 53.00% agreement for Car at 0.1 m/s. (3)Absolute velocity error directly measures the difference between predicted and ground-truth velocity and does not involve the velocity normalization floor. The failure threshold determines which samples are labeled as failures. We therefore evaluate its effect on the main conclusions. Changing the threshold from $Q_{90}$ to $Q_{80}$ alters the sample assignments and numerical results, but the main qualitative conclusions remain unchanged.
AuthorsOpenReview ↗

A2: The Principal Conclusions Remain Valid.

We replace the five hand-crafted motion features with LiLY’s learned 2020-dimensional internal representation and repeat the source-coverage analysis. Failures remain present in source-covered regions, while some samples in source-sparse regions remain correctly predicted. The results are reported in Tables 11–33. We also repeat the analysis using a 1717-dimensional descriptor set that explicitly includes directional and temporal features, obtaining the same main conclusions.

Table 1. Prediction outcomes by source coverage under two representations

\hline \text{Target} & \text{Covered failure (5D)} & \text{Covered failure (20D)} & \text{Sparse non-failure (5D)} & \text{Sparse non-failure (20D)} \\\\ \hline \text{KITTI} & 3{,}394/4{,}006\\;(84.72\\%) & 2{,}545/3{,}149\\;(80.82\\%) & 44/226\\;(19.47\\%) & 52/1{,}083\\;(4.80\\%) \\\\ \text{KAIST} & 1/385\\;(0.26\\%) & 8/678\\;(1.18\\%) & 286/293\\;(97.61\\%) & \text{--} \\\\ \text{OxIOD} & 22{,}816/24{,}352\\;(93.69\\%) & 14{,}983/15{,}940\\;(94.00\\%) & 218/8{,}514\\;(2.56\\%) & 797/16{,}926\\;(4.71\\%) \\\\ \text{QuadSLAM} & 257/471\\;(54.56\\%) & 540/837\\;(64.52\\%) & 177/697\\;(25.39\\%) & 94/331\\;(28.40\\%) \\\\ \hline \end{array}$$ **Table 2. Target-domain errors by source-defined difficulty in the LiLY 20D representation** $$\begin{array}{lrrrr} \hline \text{Target dataset} & \text{High-error group }(N) & \text{Median error (\\%)} & \text{Low-error group }(N) & \text{Median error (\\%)} \\\\ \hline \text{KITTI} & 1{,}434 & 99.65 & 2{,}798 & 100.00 \\\\ \text{KAIST} & 264 & 28.46 & 414 & 0.11 \\\\ \text{OxIOD} & 16{,}514 & 83.10 & 16{,}352 & 83.69 \\\\ \text{QuadSLAM} & 988 & 58.26 & 180 & 37.27 \\\\ \hline \end{array}$$ **Table 3. OxIOD failure recovery using groups defined in the LiLY 20D representation** $$\begin{array}{lrr} \hline \text{Calibration sample selection} & \text{C1 recovery (\\%)} & \text{C2 recovery (\\%)} \\\\ \hline \text{C1 failures} & \mathbf{41.56} & 51.53 \\\\ \text{C2 failures} & 1.74 & \mathbf{94.97} \\\\ \text{Random failure windows} & 12.94 & 90.96 \\\\ \text{Random target windows} & 12.91 & 90.11 \\\\ \hline \end{array}$$ The higher-dimensional representation changes sample distances, coverage assignments, and numerical results, but the main conclusions remain unchanged. **This supports the five-dimensional motion features as a simplified analysis tool that provides intuitive and practical evidence at lower computational cost.** The choice of calibration data still affects recovery, and refitting only the velocity prediction layer while freezing the feature extractor still corrects some failures. **These recovery results indicate that the frozen representation retains information useful for target-domain velocity prediction.**
AuthorsOpenReview ↗

A3: Reference Controls Limit the Source-Proximity Interpretation; Readout Adaptation Reveals Useful Velocity Information

We added controls that replace or shuffle the source velocity references and found that the proportions of predictions closer to the source reference remained nearly unchanged. The original Table 55 therefore cannot rule out selection effects. We revised Section 3.23.2 to remove the interpretation that these failures are caused by prediction bias toward source-domain velocities.

The readout adaptation results in Section 4.14.1 show that some failures can be corrected by refitting the velocity readout while keeping the internal representation frozen. This suggests that the representation retains useful velocity information and that target-domain adaptation can correct some predictions. We also added an explanation at the end of Section 4.14.1 to connect these findings with the analysis in Section 3.23.2.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.