PaperlayerView exact revision on arXiv

Source-linked AI summary

When LiDAR Fades, Radar Takes Over: LiDAR Reliability-Aware Adaptive Localization Under Airborne Obscurants

Moumita Mukherjee, Avijit Banerjee, Anton Koval, George Nikolakopoulo

arXiv:2610.13788v1cs.RO

TL;DR

Airborne obscurants can degrade LiDAR measurements used for localization, while existing anomaly methods may not model degradation’s temporal evolution. The paper uses nominal-only temporal learning and reliability-aware LIO-RIO selection; dense-smoke ablations show reduced feature representations can remain effective and sometimes improve switching and trajectory agreement.

  • Problem

    Airborne particulates can degrade LiDAR returns, and common anomaly methods do not explicitly model the temporal evolution of sensor degradation.

    Source: §I · p. 1

  • Method

    A one-class LSTM autoencoder models sliding windows of multi-cue LiDAR reliability features using nominal training data, with reconstruction error supporting degradation detection.

    Source: B. One-Class Temporal Degradation… · p. 2

  • Results

    Across configurations, LMRIO-RIO RMSE ranged from 1.0553 to 2.1200 m, with the lowest value for C4; reduced representations could improve switching behavior and trajectory agreement in evaluated configurations.

    Source: D. Feature Ablation Study · p. 7, §V · p. 8

  • Takeaways & Limitations

    In the evaluated dense-smoke scenarios, reduced feature representations can remain effective and sometimes improve switching behavior and trajectory agreement.

    Source: §V · p. 8

Abstract

from arXiv · show

LiDAR-inertial odometry (LIO) is widely used for real-time localization and mapping, but its reliability can degrade substantially in subterranean environments affected by smoke, dust, and other airborne obscurants, whereas radar remains comparatively robust. This article presents a one-class temporal learning framework for LiDAR degradation detection and reliability-aware localization under such conditions. A broad set of degradation-sensitive cues, including spatial, cross-modal, geometric/motion, intensity, and measurement-recency features, is evaluated to characterize changes in LiDAR sensing quality. Their temporal evolution is modelled over sliding windows using an LSTM autoencoder trained exclusively on nominal operation, enabling deviations from healthy sensor behaviour to be detected without requiring labelled degradation data. The resulting LiDAR reliability estimate governs adaptive selection between LIO and radar inertial odometry (RIO), retaining LIO during reliable sensing and transferring localization authority to RIO when degradation is detected. The framework is evaluated using real-world ground robot data collected in GNSS-denied subterranean tunnels under nominal and dense-smoke conditions. Experimental results demonstrate degradation detection, adaptive modality switching, and continued localization output under severe LiDAR degradation. A feature family ablation study further shows that a reduced feature representation can improve degradation-aware localization performance in the evaluated dense-smoke scenarios, highlighting the importance of feature parsimony in one-class sensor degradation detection. The code will be released as open source to support reproducibility and further research.

I. INTRODUCTION

Airborne obscurants can reduce LiDAR returns and alter their distribution, challenging reliable localization in GNSS-denied subterranean settings. The article combines complementary LiDAR cues with temporal one-class learning and adaptive LIO-RIO selection.

I. INTRODUCTION

  • Smoke and other airborne particulates can reduce valid LiDAR returns and change their spatial distribution and density, while radar is comparatively robust.LiDAR nevertheless provides high-resolution geometric information.Source: §I · p. 1

A. Feature Extraction for Anomaly Detection

  • Voxel-based spatial statistics preserve point concentration and occupancy cues, complemented by temporal, geometric, cross-modal, inertial, and intensity features.The feature representation is designed to capture changes in sensing quality without relying on a single cue.Source: A. Feature Extraction for Anomaly… · p. 2

B. One-Class Temporal Degradation Detection with LSTM Autoencoders

  • An LSTM autoencoder models nominal feature sequences over sliding windows and detects emerging degradation through reconstruction-error deviations, without labelled failure data.This sequence-based approach addresses the limitation of methods that assess individual observations without modelling temporal evolution.Source: §I · p. 1, B. One-Class Temporal Degradation… · p. 2
  • The framework uses LiDAR health to retain LIO during reliable sensing and select RIO when LiDAR degrades or becomes unavailable.The feature representation and its ablation are evaluated under dense-smoke conditions.Source: B. One-Class Temporal Degradation… · p. 2

C. Notation

The paper formulates LiDAR reliability estimation and adaptive localization for a ground robot operating in GNSS-denied subterranean conditions. It defines a common IMU timeline and uses LiDAR health to select between LIO and RIO.

C. Notation

  • The notation indexes time on the common IMU timeline with k and indexes unique LiDAR scans with η.It also defines the reliability feature vector, temporal window, reconstruction error, anomaly threshold, and degradation state.Source: C. Notation · p. 2

II. PROBLEM FORMULATION

  • Under obscurants, LiDAR returns may degrade or become stale, so the robot estimates reliability from complementary cues and selects LIO or RIO accordingly.The target setting is a non-holonomic ground robot equipped with radar, a 3D LiDAR, and an IMU.Source: §II · p. 2

III. METHODOLOGY

  • The methodology synchronizes asynchronous sensor and odometry streams, encodes each LiDAR scan as twelve degradation-sensitive features, and processes consecutive features with a nominally trained LSTM autoencoder.Reconstruction error determines degradation and triggers a switch to RIO when LiDAR becomes unreliable.Source: §III · pp. 2–3

A. Sensor Synchronization

Asynchronous sensing and odometry streams are aligned to the IMU timeline using the most recent available measurement. LiDAR scans are represented with voxel-based statistics that summarize return amount and spatial distribution.

A. Sensor Synchronization

  • The IMU timestamps provide the common reference timeline, where the most recent available measurement is selected for each lower-rate stream.A zero-order hold aligns LiDAR, LIO, and RIO measurements; repeated LiDAR observations are excluded from the LSTM sequence.Source: A. Sensor Synchronization · p. 3

B. Voxel-Grid based LiDAR Pointcloud Representation

  • Each LiDAR scan’s point cloud is partitioned into cubic voxels of side length v to characterize its spatial distribution.The representation uses the minimum point-cloud coordinate to define the voxel partition.Source: B. Voxel-Grid based LiDAR Pointcloud… · p. 3
  • Occupied-voxel counts and point densities are summarized to characterize changes in both the amount and spatial distribution of LiDAR information.The density statistics are computed for points occupying each voxel.Source: B. Voxel-Grid based LiDAR Pointcloud… · p. 3

C. Degradation Sensitive Feature Extraction

The reliability representation augments voxel statistics with temporal, cross-modal, geometric, inertial, and intensity cues. These features measure LiDAR recency, estimator consistency, point-cloud motion, and return-intensity distributions.

C. Degradation Sensitive Feature Extraction

  • Voxel statistics are augmented with temporal, cross-modal, geometric, inertial, and intensity cues because spatial statistics alone may not distinguish degradation from scene or motion effects.LiDAR age captures staleness, while LIO-RIO pose discrepancy supplies a cross-modal consistency cue.Source: C. Degradation Sensitive Feature… · p. 3
  • Point-cloud centroid and its displacement between scans provide geometric and motion context for the reliability features.The representation also includes measured IMU angular-velocity magnitude.Source: C. Degradation Sensitive Feature… · p. 3
  • Median and interquartile range of valid LiDAR return intensities contribute to the twelve-dimensional reliability feature vector.The vector is associated with each unique LiDAR scan.Source: C. Degradation Sensitive Feature… · p. 3

D. Feature Standardization and Temporal Sequence Construction

Reliability features are standardized using statistics from nominal training data, then arranged into temporal windows of unique LiDAR observations.

D. Feature Standardization and Temporal Sequence Construction

  • Each reliability feature is standardized using feature-wise statistics computed exclusively from nominal training data.The standardization uses the healthy training dataset’s feature-wise mean and standard deviation.Source: D. Feature Standardization and… · p. 3
  • The framework synchronizes asynchronous LiDAR, IMU, LIO, and RIO measurements and encodes them as twelve-dimensional degradation-sensitive features.Consecutive normalized feature vectors form the temporal sequence used by the LSTM autoencoder.Source: Figure 2 · p. 4
  • Each temporal window contains W consecutive unique LiDAR observations.Repeated observations introduced by zero-order hold are removed before sequence construction.Source: D. Feature Standardization and… · p. 4

E. One-Class Temporal Degradation Detection

An LSTM autoencoder trained only on nominal operation detects temporal deviations through reconstruction error, with thresholding and smoothing used to form an anomaly decision.

E. One-Class Temporal Degradation Detection

  • The LSTM autoencoder is trained exclusively on nominal LiDAR operation as a one-class temporal model.Its parameters are learned by minimizing mean-squared reconstruction error over healthy training sequences.Source: E. One-Class Temporal Degradation… · p. 4
  • Reconstruction error for a complete temporal window is compared with an anomaly threshold derived from the upper tail of healthy errors.The threshold is based on nominal-training reconstruction errors.Source: E. One-Class Temporal Degradation… · p. 4
  • The online score is smoothed over the most recent S estimates to suppress isolated reconstruction-error fluctuations.Source: E. One-Class Temporal Degradation… · p. 4
  • The learned anomaly indicator is placed on the IMU timeline.The passage describes the indicator as being learned using LiDAR-indexed reconstruction-error estimates.Source: E. One-Class Temporal Degradation… · p. 4

F. Reliability-Aware LIO-RIO Selection

The reliability decision controls whether localization uses LIO or RIO, while a separate recency check addresses measurement stalls that reconstruction error alone cannot detect.

F. Reliability-Aware LIO-RIO Selection

IV. EXPERIMENTAL RESULTS

The evaluation used four subterranean robot traversals spanning nominal and dense-smoke conditions; without independent trajectory ground truth, its reported metrics measure agreement with comparison references rather than absolute accuracy.

IV. EXPERIMENTAL RESULTS

  • The framework was evaluated on a ground robot in a GNSS-denied real-world tunnel under nominal and dense-smoke conditions.Artificial smoke induced LiDAR-return attenuation and backscattering to test the reliability mechanism under progressive degradation and measurement loss.Source: §IV · p. 5
  • Four traversals covered P-shaped and round-tunnel trajectories under both nominal and dense-smoke conditions.Source: §IV · p. 5
  • Without independent trajectory ground truth, nominal-condition metrics use LIO as the reference and dense-smoke metrics use RIO, so they report estimate agreement rather than absolute localization accuracy.Source: §IV · p. 5

A. Experimental Protocol and Model Configuration

The evaluation tests LMRIO on nominal and dense-smoke tunnel traversals, using temporal LiDAR degradation detection and an independent recency criterion to select between LIO and RIO. LIO is retained during healthy sensing, while RIO maintains localization through degraded or stale LiDAR measurements.

A. Experimental Protocol and Model Configuration

  • The one-class LSTM autoencoder is trained only on nominal P-shaped data using twelve features arranged into 100-scan windows.The healthy reconstruction-error 99th percentile sets the anomaly threshold; online error is smoothed over the most recent 50 estimates.Source: A. Experimental Protocol and Model… · p. 5

B. Localization Performance Under Nominal and Dense-Smoke Conditions

  • In dense smoke, LMRIO selects RIO during LiDAR degradation or staleness, maintaining localization; under nominal conditions, it stays aligned with LIO without unnecessary switching.The P-shaped traversal tests progressive measurement degradation, while the round-tunnel traversal tests loss of LiDAR updates.Source: B. Localization Performance Under… · p. 5, Figure 3 · p. 6, Figure 4 · p. 6

C. Temporal Degradation Detection and Adaptive Modality Selection

  • Smoothed reconstruction error detects persistent departures from nominal feature dynamics, suppressing isolated fluctuations while retaining sustained degradation signals.In the P-shaped smoke traversal, persistent error growth triggers a switch to RIO before complete measurement loss.Source: C. Temporal Degradation Detection and… · pp. 5–6, Figure 5 · p. 6
  • LiDAR age is evaluated independently on the IMU timeline, and either the learned anomaly criterion or the 1.5 s staleness criterion activates the final degradation decision.This recency criterion handles stale or missing updates when reconstruction-error updates cease.Source: C. Temporal Degradation Detection and… · p. 7, Figure 6 · p. 7

D. Feature Ablation Study

The ablation shows that feature composition affects when localization switches to RIO and how the resulting trajectory agrees with the radar-inertial estimate. Reduced feature sets can detect degradation before LiDAR becomes stale, while timing, estimator usage, and trajectory agreement trade off.

D. Feature Ablation Study

  • The ablation compared the complete twelve-feature C1 against variants removing spatial, cross-modal, geometry/motion, or intensity features, plus C6 disabling the independent recency criterion.Feature-level variants were retrained on the same nominal P-shaped data; C6 reused C1 and changed only the final decision rule.Source: D. Feature Ablation Study · p. 7
  • The first LIO-to-RIO transition ranged from 74.44 to 125.00 s, while RIO usage ranged from 54.70% to 72.93%.The results indicate that feature composition affects switching behaviour.Source: D. Feature Ablation Study · p. 7
  • Without spatial statistics, C2 switched at 125.00 s, about 5 s before LiDAR staleness at 130.00 s; the round-tunnel run switched at about 35 s before staleness at about 90 s.The remaining cues therefore detected degradation before complete measurement loss in both reported traversals.Source: D. Feature Ablation Study · p. 7
  • Across configurations, LMRIO-RIO RMSE ranged from 1.0553 to 2.1200 m, with the lowest value for C4, alongside trade-offs in switching time, RIO usage, and trajectory agreement.The large LMRIO-LIO RMSE under dense smoke is attributed to severe LIO divergence rather than ordinary localization uncertainty.Source: D. Feature Ablation Study · p. 7

E. Quantitative Analysis

Pairwise trajectory comparisons characterize how LMRIO relates to LIO and RIO under nominal and dense-smoke conditions, but do not establish absolute trajectory accuracy. The conclusion reports continued localization under degraded or stale LiDAR and notes that reduced feature representations can remain effective, while identifying broader generalization and ground-truth evaluation as future work.

E. Quantitative Analysis

  • Under dense smoke, LMRIO-RIO RMSE was 1.1148 m versus 1.816 × 104 m for LIO-LMRIO; nominal LMRIO-LIO RMSE was 0.0121 m (RIO-LIO: 0.0650 m).Table V reports pairwise planar position RMSE; without independent external ground truth, these values measure agreement between estimates rather than absolute trajectory accuracy.Source: E. Quantitative Analysis · p. 7, Table V · p. 8
  • Table III highlights the reconstruction-error switching time closest to staleness, while Table IV lists the feature-family configurations used in the ablation.Source: Table III · p. 8, Table IV · p. 8

V. CONCLUSIONS

  • The LMRIO framework retains LIO during reliable sensing and transfers localization authority to RIO when LiDAR is degraded or unavailable, with continued output under corrupted and stale measurements.Real-world tunnel experiments covered nominal and dense-smoke conditions.Source: §V · p. 8
  • Reduced feature representations remained effective and improved switching behaviour and trajectory agreement in some evaluated configurations.The conclusion identifies generalization across environments and obscurant types, alternative one-class detectors, and independent trajectory ground truth as future work.Source: §V · p. 8
Loading 2610.13788v1…