acceptodds
Under review as a conference paper at ICLR 2027

DIAL: Drift-aware Continual Alignment for Implicit LiDAR Localization

Abstract

Implicit LiDAR localizers compress a scene into a metric field stored in network parameters, but this field becomes stale as environments evolve. Repeated pose annotation and full retraining are costly across recurring revisits. Continual updates face two coupled problems: supervision is missing where the historical model fails, and unconstrained adaptation can overwrite a coordinate readout that remains valid elsewhere. We present the first study of continual learning for implicit LiDAR localization. We find that the historical localizer supports both needs: its predictions reveal where the implicit map has changed, while its metric readout preserves valid coordinate associations when old scans are unavailable. DIAL’s Temporal Consistency-guided Pseudo-label Pipeline (TCPP) cross-validates retrieval and scene-coordinate evidence over time to recover pseudo poses for changed or unseen frames. Head-Decoupled Encoder Regularization (HDER) freezes the readout and adapts selected encoder blocks when old scans are inaccessible. With historical memory, Memory-Conditioned Rehearsal (MCR) reinstates old-scene constraints; Redundancy-Aware Exemplar Replay (RAER) selects the replay needed as route coverage expands. On NCLT, pseudo labels for about 20% of revisit frames suffice for maintenance. A buffer-based DIAL route updates the localizer for the first NCLT environmental change in about 12 minutes of continual optimization; repeated maintenance then sustains localization across roughly one year of NCLT revisits, while segmented QEOxford approaches full-retraining accuracy. Code will be released.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Continual Localization,Implicit Localization, Scene Evolution, Metric Memory

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_zpPNOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DIAL, a framework for continual implicit LiDAR localization under scene evolution. It combines TCPP, HDER, MCR, and RAER to balance adaptation to changed environments with the preservation of historical localization accuracy. Experiments on NCLT and QEOxford demonstrate reduced maintenance costs, improved long-term localization, and performance approaching joint training under segmented route expansion.

Critical strengths

The paper addresses a practically important problem. It studies how implicit LiDAR localization models can be maintained as environments evolve, rather than assuming a static scene or repeatedly retraining from scratch. Its consideration of different historical-data availability and route-expansion scenarios makes the problem formulation relevant to long-term robotic deployment.

Critical weaknesses

TCPP's diagnostic reliability remains insufficiently established. Since only SCR-rejected observations enter the update set, smooth systematic drift may escape detection. Although Table 6 and Figure A2 provide indirect evidence of pseudo-label quality and end-to-end improvements, the paper lacks quantitative analysis of diagnostic false positives and false negatives.

Questions for authors

It would be helpful if the authors could provide quantitative or qualitative evidence of TCPP’s region-selection reliability, or explain why such an evaluation is infeasible. This would clarify whether temporal residuals reliably indicate localization degradation rather than ordinary trajectory irregularities.

2 comments
AuthorsOpenReview ↗

Response to Reviewer zpPN

Thank you for your review and recognition of DIAL's practical value. We have submitted a revised manuscript, with changes highlighted in orange-yellow.

Main result. On 05-26, TCPP selects 20.82% of observations and detects 97.29% of SCR failures above 5m. The recovered labels reduce mean error on that selected set from 88.19m/41.33deg to 2.61m/4.35deg. We report diagnostic errors separately from label accuracy below.

R3.1 [W1/Q1]: Diagnostic false positives and negatives

Concern: The paper lacks quantitative analysis of diagnostic false positives and negatives.

We use GT only to audit the saved routing decisions. A true positive is a selected frame whose historical SCR 3D translation error exceeds the tolerance; a false positive is selected but at or below that tolerance; a false negative exceeds it but is unselected. Precision (P), recall (R), false-positive rate (FPR), and false-negative rate (FNR) are percentages.

Table R7. Fixed-gate diagnosis and full-route mean TE (m), SCR/TCPP.

RevisitR, 5mR, 10mP, 5mFPR, 5mFNR, 5mSCR/TCPP
05-2697.2998.5053.4710.942.7119.19/0.82
10-2884.3896.5052.6011.3115.627.83/1.82
11-1661.7986.3750.8110.5438.215.06/2.32

On 10-28/11-16, 88.20%/87.34% of missed >5m failures lie within 5-10m. Missed >10m failures account for 0.24%/0.72% of all frames, respectively; the >5m failure-conditional FNRs remain 15.62%/38.21%. Both tolerances evaluate the same decisions without GT-based tuning. P measures selection precision, not localization accuracy. Full-route >5m rates are 11.44%/12.95%/14.98%; selection concentrates such failures by 4.67/4.06/3.39 times, respectively.

Diagnosis precedes the current update: 10-28 uses post-05-26 SCR; 11-16 uses post-10-28 SCR, already updated twice. On 11-16, full-route mean TE falls from 5.06m to 2.32m after fusion (Table R7).

R3.2 [Q1]: Normal motion versus localization errors

Question: Do temporal residuals indicate localization degradation or ordinary trajectory irregularity?

To test normal-motion confounding directly, we restrict to accurate SCR predictions (TE <=2m): selection then indicates a false alarm. Turns use GT heading changes >=30deg over ten steps on each side; straight motion is <=10deg, with both displacements >1m.

Table R7a. Selection rates among accurate SCR predictions (TE <=2m).

RevisitTurns selected (%)Straight selected (%)
05-268.815.35
10-285.883.63
11-164.673.40

Thus >91% of accurate turn frames remain unselected, although turns raise false selection. This tests the complete gate over a 20-step span, not every instantaneous sharp turn.

For selected degraded frames (SCR TE >5m), mean TE falls from 163.02 to 3.77m, 59.36 to 10.14m, and 35.39 to 10.71m (Table R8a). This is recovery quality, separate from diagnostic recall/false alarms in R7. Smooth global offsets can still escape detection.

R3.3 [W1/Q1]: Selected-region supervision

Concern: End-to-end gains alone do not establish region-selection reliability.

Diagnostic precision alone does not determine label quality. We therefore measured the actual selected H5 poses on the same timestamps as historical SCR.

Table R8. Mean TE/RE (m/deg) on TCPP-selected training frames, not the full route.

RevisitHistorical SCRTCPP H5
05-2688.19/41.332.61/4.35
10-2832.47/36.306.28/11.28
11-1619.31/25.367.08/13.85

Table R8a. TCPP-selected TP/FP at 5m. Mean TE is in m; TCPP closer is the percentage of frames whose TCPP TE is smaller than SCR TE, not a mean-error reduction.

RevisitGroupSCR meanTCPP meanTCPP closer (%)
05-26TP163.023.7798.98
05-26FP2.211.2886.41
10-28TP59.3610.1485.46
10-28FP2.641.9979.53
11-16TP35.3910.7175.31
11-16FP2.693.3373.53

The selected high-error groups improve substantially. False positives can also receive better labels, but not uniformly: on 11-16 their mean TE rises from 2.69m to 3.33m despite most individual errors decreasing. This separates selection precision from label quality without treating every selected frame as beneficial.

Full-route TCPP mean TE is 0.82m, 1.82m, and 2.32m, respectively; this is separate from the harder training subsets. The submitted Table 6 additionally tests training utility: with its matched label budget, TCPP gives aggregate TE/RE of 1.73/4.92 versus 2.35/5.23 for retrieval+ICP. The new measurements directly test routing and label quality, complementing that downstream evidence.

Thank you for the experiments and clarifications. My main concerns have been addressed, and I will keep my score unchanged.

Reviewer_zBi6OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

DIAL studies continual maintenance of implicit LiDAR localizers. It selects unreliable observations using temporal consistency, recovers pseudo poses through retrieval and ICP, and combines selective parameter updates with historical replay.

Critical strengths

I appreciate the practical problem of adapting a deployed localizer while preserving historical accuracy. The distinction between changed revisits and route expansion provides a useful framework for choosing what to update.

Critical weaknesses
  1. My main concern is that local trajectory smoothness is insufficiently validated as a diagnostic of localization failure. Eqs. 1 and 9 can accept globally shifted or constant predictions. Figure A2 reports full-trajectory pseudo-label accuracy, but does not quantify missed failures or label quality on the selected training subset. This leaves the reliability of supervision routing unclear.
  2. I also find the evidence for HDER’s specific stability mechanism incomplete. Table 4 compares full-revisit naive adaptation with HDER trained on TCPP-selected frames, and Table A3’s zero-prior row reuses that baseline. Without matched simple fine-tuning and freeze-only controls, the effects of selection, freezing, and regularization cannot be separated.
Questions for authors
  1. Could the authors report TCPP’s precision/recall for high-error frames and translation/rotation errors on the selected training subset?
  2. Under identical TCPP supervision and update budgets, how do simple fine-tuning, freezing without the encoder prior, and full HDER compare?
1 comment
AuthorsOpenReview ↗

Response to Reviewer zBi6

Thank you for your review and recognition of DIAL's practical value. We have submitted a revised manuscript, with changes highlighted in orange-yellow.

Main results. TCPP reduces selected-pool mean TE/RE on 05-26 from 88.19m/41.33deg to 2.61m/4.35deg. Under identical supervision and steps, aggregate TE is 9.221m with fine-tuning, 5.273m with freeze-only, and 2.707m with HDER.

R2.1 [W1/Q1]: Diagnostic reliability

Concern: Smooth global errors may pass the gate. What are its precision and recall?

Smooth global offsets or constant predictions can pass the local temporal gate. We therefore quantify detection coverage and missed failures directly.

We audited saved masks against GT without retuning: a reference failure has pre-update SCR 3D TE above the tolerance. R is recall; FPR/FNR are false-positive/negative rates. All rates are percentages; 5m/10m evaluate the same decisions. P is selection precision, not localization accuracy: selected frames within 5m count as false positives. Full-route >5m rates are 11.44%/12.95%/14.98%; selected pools are 4.67/4.06/3.39 times enriched in failures.

Table R4. Diagnosis and full-route mean TE (m), SCR/TCPP.

RevisitR, 5mR, 10mP, 5mFPR, 5mFNR, 5mSCR/TCPP
05-2697.2998.5053.4710.942.7119.19/0.82
10-2884.3896.5052.6011.3115.627.83/1.82
11-1661.7986.3750.8110.5438.215.06/2.32

11-16 uses the twice-updated SCR (see below). Its >10m failure recall is 86.37%; full-route TE falls from 5.06m to 2.32m after TCPP recovery.

R2.2 [W1/Q1]: Selected-training-subset accuracy

Concern: Figure A2 does not quantify TE/RE on the selected training subset.

We evaluated the actual training H5 on identical SCR/TCPP timestamps. TE/RE denotes mean translation/rotation error in m/deg.

Table R5. Selected supervision versus full-route pseudo poses.

RevisitSelected SCR TE/RESelected TCPP TE/REFull-route TCPP TE/RE
05-2688.19/41.332.61/4.350.82/2.59
10-2832.47/36.306.28/11.281.82/4.35
11-1619.31/25.367.08/13.852.32/6.48

The selected sets contain 5,504, 5,308, and 4,424 labels, respectively. Thus the 6-7m errors concern the hard training subsets, not entire trajectories. Both selected-pool TE and RE improve on all three revisits, although noisy labels remain. SCR is the inherited model before each update: initial for 05-26, post-05-26 for 10-28, and post-10-28 for 11-16.

R2.3 [W2/Q2]: Matched HDER controls

Question: Under identical supervision and steps, how do fine-tuning, freeze-only and HDER compare?

We trained all three from the same historical checkpoint with the same 05-26 TCPP H5, batch size 50, seed 7, data loading, augmentation, AdamW/LR schedule, and 5,500 steps (50 epochs), without replay. Fine-tuning updates encoder and head. Freeze-only fixes the head and Conv1-3, with no encoder prior. HDER has the same frozen groups and adds the encoder prior: flat layer weights, no adaptive scaling, and a 0.03 -> 0.0125 -> 0.03 schedule. The new runs use four workers, 0.25m voxels, LR 0.001, weight decay 0.01, and one RTX 4090. 05-11 is evaluation-only.

Table R6. New matched controls; mean TE/RE (m/deg).

VariantHistoricalEvolved/unseenAggregate
TCPP-matched fine-tuning8.479/11.39111.447/11.7509.221/11.481
Freeze-only5.146/9.1325.652/9.7245.273/9.280
HDER, matched rerun2.325/6.0553.855/7.2952.707/6.365

Table R6a. Per-sequence matched controls; mean TE/RE (m/deg).

SequenceFine-tuningFreeze-onlyHDER
02-128.884/11.7575.441/9.5952.325/6.016
02-197.928/10.5035.033/8.7922.318/5.971
03-318.624/11.9144.965/9.0092.332/6.179
05-1111.447/11.7505.652/9.7243.855/7.295

Historical averages 02-12/02-19/03-31; evolved/unseen is 05-11; aggregate averages all four. HDER lowers TE and RE on every evaluated route versus both controls. Aggregate TE falls by 70.6% relative to fine-tuning and 48.7% relative to freeze-only. This separates freezing from the added encoder prior under fixed selection and budget. These single-seed reruns are separate from the paper's DIAL-H model (2.59/6.03 aggregate).

R2.4 [W2]: Submitted Tables 4 and A3

Concern: The full-revisit naive baseline cannot isolate the encoder prior.

Table A3's original "0 (naive update)" row is not HDER with only its prior removed: both supervision and trainable groups differ. The new freeze-only row supplies that control. The revised Tables 4/A3 distinguish full-revisit naive adaptation from the HDER variants. Table A3 now lists it as an external reference, with the caption and accompanying text clarifying the distinction.

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

The paper proposes DIAL that continually updates the implicit LiDAR localizer under environmental changes and route expansion. DIAL uses a temporal consistency guided pipeline (TCPP) to identify unreliable scene coordinates and recover pseudo poses for unseen or changed frames. It also introduces HDER, MCR, and RAER to preserve historical geometry. The experimental results retain accuracy on long-term localization benchmarks under a small amount of optimization.

Critical strengths

The paper addresses a useful problem of long-term drift in implicit LiDAR localization, and proposes to continually adapt without rebuilding the localizer. Tab. 2 indicates that the method reduces evolved/unseen translation error from 15.92m to 2.91m on the same PELoc backbone, while preserves accuracy of the localizer in historical regions. The efficacy of the proposed modules is also justified through detailed ablation studies.

Critical weaknesses

Eq.1 only measures local path inconsistency, while Eq. 9 measures deviation from a local trajectory centroid. Therefore, a smooth trajectory with incorrect poses can pass the temporal gate and become an anchor point under Eq.2; a sharp but precise turning would be ignored.

The authors are suggested to provide statistical correlations between the high-localization-error frames and the error distribution of accepted anchors to see if the smoothness is sufficient for identifying the erroneous frames. The experiments do not quantify how often the gate misses high-error regions or accepts erroneous anchors.

Questions for authors

The hardware, optimizer, learning rate, batch size, and the number of update steps should be clarified in the manuscript for reproducing the paper.

2 comments
AuthorsOpenReview ↗

Response to Reviewer jcMR (1/2)

Thank you for your review and recognition of DIAL's practical value. We have submitted a revised manuscript, with changes highlighted in orange-yellow.

Main result. On 05-26, the gate detects 97.29% of historical SCR errors above 5m. On the selected training frames, TCPP reduces mean translation/rotation error from 88.19m/41.33deg to 2.61m/4.35deg.

R1.1 [W1/W2]: Smooth errors and missed failures

Concern: Smooth wrong poses can pass the gate; missed failures need quantification.

Smooth global offsets or constant predictions can pass the local temporal gate. We therefore quantify detection coverage and missed failures directly, separately from retrieval/ICP recovery. The gate detects 97.29% of >5m SCR failures on 05-26 and 98.50%/96.50%/86.37% of >10m failures across the three revisits (Table R1).

We audited saved decisions without retuning. A reference failure has historical SCR 3D TE above the tolerance; GT is used only for evaluation. R is recall; FPR/FNR are false-positive/negative rates; all rates are percentages. P is selection precision, not localization accuracy: selected frames within 5m count as false positives. Compared with full-route >5m rates of 11.44%/12.95%/14.98%, the selected pools are 4.67/4.06/3.39 times enriched in failures.

Table R1. Diagnosis and full-route mean TE (m), SCR/TCPP; fixed gate at both tolerances.

RevisitR, 5mR, 10mP, 5mFPR, 5mFNR, 5mSCR/TCPP
05-2697.2998.5053.4710.942.7119.19/0.82
10-2884.3896.5052.6011.3115.627.83/1.82
11-1661.7986.3750.8110.5438.215.06/2.32

Among missed >5m failures on 10-28/11-16, 88.20%/87.34% lie within 5-10m. Tables R2a/R2b then examine the consequences for accepted poses and propagation seeds, comparing their original and final errors.

Pre-update SCR: initial for 05-26, post-05-26 for 10-28, post-10-28 for 11-16. Thus 11-16 starts from a twice-updated model; TCPP still reduces full-route error.

R1.2 [W2]: Accepted-anchor error distribution

Concern: How often are erroneous anchors accepted, particularly in high-error regions?

We audited both pose-source pools before fusion, with retrieval-first attribution. Ret denotes retrieval-gate-accepted poses; SCR-only denotes SCR-gate-accepted poses without accepted retrieval at that timestamp. These are route-wide pools, not counts of propagation-boundary anchors or selected training labels. B@5 is the percentage with TE >5m.

Table R2. Pre-fusion source pools; TE in m, B@5 in %.

RevisitSourcenMean TEP90 TEB@5
05-26Ret215621.010.300.64
05-26SCR-only18332.272.671.80
10-28Ret199491.530.311.01
10-28SCR-only25083.225.9817.38
11-16Ret159482.780.561.82
11-16SCR-only54603.586.9323.28

Table R2a. Retrieval-accepted pool restricted to SCR TE >5m: before/after full postprocessing on identical frames. TE in m; B@5 in %.

RevisitPose sourceMean TEP90 TEB@5
05-26Retrieval7.500.405.90
05-26Final TCPP0.820.404.80
10-28Retrieval9.203.489.35
10-28Final TCPP1.991.127.77
11-16Retrieval24.98102.6218.64
11-16Final TCPP4.2820.1216.38

R2a measures full postprocessing, separately from gate acceptance. To check propagation itself, we reconstructed the unique seeds retained for ICP-filled intervals from saved anchor inputs and the archived propagation order.

Table R2b. Propagation seed poses before/after postprocessing, at identical timestamps. TE in m; B@5 in %.

RevisitStagenMean TEP90 TEB@5
05-26Original38024.016.1110.53
05-26Final TCPP3802.493.386.32
10-28Original49519.2314.6330.91
10-28Final TCPP4955.237.3328.48
11-16Original66518.0113.9342.56
11-16Final TCPP6655.7910.9439.70

These seeds have heavier errors than the route-wide pool. Full postprocessing reduces mean and P90 errors on all three revisits, while residual >5m errors remain.

AuthorsOpenReview ↗

Response to Reviewer jcMR (2/2)

R1.3 [W1]: Accurate sharp turns

Concern: Accurate sharp turns may be rejected by the temporal gate.

Using GT directions over ten acquisition steps on each side, we define turns as heading changes of at least 30deg and straight motion as at most 10deg, requiring both displacements to exceed 1m. Among accurate SCR frames (TE <=2m), the selected fractions are:

Table R3. Selection of accurate predictions near turns versus straight motion.

RevisitTurns (%)Straight (%)
05-268.815.35
10-285.883.63
11-164.673.40

More than 91% of accurate turn frames remain unselected. This tests the complete gate; normal turns modestly increase false selection.

R1.4 [Q1]: Reproducibility

Question: Please specify hardware, optimizer, LR, batch size and update steps.

The original NCLT DIAL-H and DIAL-Mhead runs and new controls use AdamW (initial LR 0.001, effective weight decay 0.01), batch 50, seed 7, 0.25m voxels, FP16, and five-epoch warmup followed by cosine LR. All train on the 05-26 TCPP pool (5,504 labels), with 110 batches/epoch and 5,500 optimizer updates over 50 epochs.

Table R3a. Original maintenance runs versus fresh controls.

RunUpdated parametersReplay
Paper DIAL-HEncoder except Conv1-3None
Paper DIAL-MheadHead onlyRAM raw buffer
New fine-tuningEncoder and headNone
New freeze-only/HDEREncoder except Conv1-3None

DIAL-Mhead freezes the encoder and uses 02-18 replay at stride 4, weight 1, once every four current steps (including step 0 each epoch): 28 replay batches/epoch, 1,400 replay batches, within the same 5,500 optimizer updates. Its archived training-process time is 712.173s (~12min), including buffer loading but excluding pseudo-label generation/evaluation. DIAL-H fixes the head and uses an encoder prior; the new HDER schedule is 0.03 -> 0.0125 -> 0.03.

The head-only run and new controls use one RTX 4090 (24GB). New controls use PyTorch 1.11.0+cu113 and MinkowskiEngine 0.5.4. These optimization details are summarized in Appendix G.2.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Overall Author Response — DIAL

We sincerely thank the Area Chair and all reviewers for their time, careful evaluation, and constructive suggestions. Their feedback prompted additional analyses and clearer reporting of DIAL's diagnostic reliability, pseudo-label quality, and continual-update controls. We have incorporated the relevant results and clarifications into the revised manuscript and detailed responses.

  • To jcMR: We quantified TCPP's detection recall, precision, false positives, and missed failures, including the limitations of temporal checks under smooth pose errors. We examined accepted pose sources and ICP propagation seeds, tested whether accurate sharp turns trigger false selection, and clarified optimization settings and update costs.

  • To zBi6: We reported translation and rotation errors for the TCPP-selected training labels and added matched comparisons of fine-tuning, freeze-only adaptation, and HDER. Under the same supervision and training budget, HDER performed better than both controls.

  • To zpPN: We reported diagnostic false-positive and false-negative rates, examined normal-turn confounding, and evaluated recovered labels separately for correctly and incorrectly selected regions. These analyses distinguish the reliability of selecting frames from the accuracy of the supervision recovered for them.

We are grateful for the reviewers' engagement and welcome any further questions, concerns, or requests for clarification during the discussion.

Area_Chair_516uOpenReview ↗

Please share your post-rebuttal review comments

Dear Reviewers,

Thanks for reviewing this paper. Please take a closer look at the feedback from the authors and see if it has addressed your concerns and if you have follow-up questions for the authors. The discussion session will end soon!

Best, AC for paper 5769

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.