arxiv: 2604.02706 · v1 · submitted 2026-04-03 · 💻 cs.RO

Recognition: no theorem link

ALIVE-LIO: Degeneracy-Aware Learning of Inertial Velocity for Enhancing ESKF-Based LiDAR-Inertial Odometry

Seongjun Kim , Daehan Lee , Junwoo Hong , Sanghyun Park , Hyunyoung Jo , Soohee Han

Authors on Pith no claims yet

Pith reviewed 2026-05-13 20:35 UTC · model grok-4.3

classification 💻 cs.RO

keywords LiDAR inertial odometrydegeneracy detectionneural velocity predictionerror state Kalman filtersensor fusionpose estimationautonomous navigation

0 comments

The pith

ALIVE-LIO compensates for LiDAR degeneracy by selectively fusing neural network velocity predictions into an error-state Kalman filter.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a hybrid LiDAR-inertial odometry system that detects when LiDAR measurements lose reliability in environments like corridors and uses a neural network to supply body-frame velocity estimates at those moments. These estimates are incorporated into the classical error-state Kalman filter only when needed, preserving the filter's uncertainty modeling while addressing gaps in sensor observability. This matters for autonomous robots that must navigate real indoor or urban spaces where standard LIO drifts because LiDAR cannot observe motion in all directions. The approach is tested on multiple datasets and shows lower drift than competing methods in most sequences.

Core claim

ALIVE-LIO employs a neural network to predict body-frame velocity and fuses this prediction into the ESKF only upon detecting degeneracy in LiDAR observations, thereby providing state updates along unobservable directions and reducing pose drift in challenging environments.

What carries the argument

Degeneracy-aware selective fusion of neural velocity predictions into the ESKF

Load-bearing premise

The neural network must deliver accurate velocity predictions specifically in the degenerate directions, and the degeneracy detector must activate fusion without adding bias to the estimate.

What would settle it

Running the system on a sequence in a degenerate environment where the neural velocity prediction error is larger than the natural LiDAR drift would show increased error compared to standard ESKF-LIO.

Figures

Figures reproduced from arXiv: 2604.02706 by Daehan Lee, Hyunyoung Jo, Junwoo Hong, Sanghyun Park, Seongjun Kim, Soohee Han.

**Figure 2.** Figure 2: Pipeline of ALIVE-LIO. TABLE I NOTATIONS USED IN THIS PAPER Notation Description W (·) A vector (·) in the world frame B(·) A vector (·) in the body frame (·)m Measured raw sensor data (·) NN Output of the neural network xb κ κ-th update of x in IESKF δx Error state in IESKF x State of x resulting from update degeneracy and suppresses its effects within the IESKF, producing an odometry estimate along the n… view at source ↗

**Figure 4.** Figure 4: Distribution of gravity W g, acceleration bias ba and gyro bias bg estimated by the LIO system. The z-component of gravity is shown after removing the well-known constant magnitude. 9.81 m/s2 [15], [18]. However, because many LIO systems do not know the initial roll and pitch, the rotation of the starting frame is set to the identity [19], [32], [33]. W g is initialized by remaining stationary for a few se… view at source ↗

**Figure 5.** Figure 5: Vehicle motion: Estimated trajectory visualization under degenerate conditions (X–Y plane in meters). The symbols [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗

**Figure 6.** Figure 6: Handheld motion: Estimated trajectory and velocity visualization on [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗

**Figure 7.** Figure 7: Plots of the estimated trajectories, velocities, and acceleration biases for degenerate cases. The left column shows the trajectory (X–Y plane in meters), [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗

**Figure 8.** Figure 8: Performance comparison under abundant and sparse visual features [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗

read the original abstract

Odometry estimation using light detection and ranging (LiDAR) and an inertial measurement unit (IMU), known as LiDAR-inertial odometry (LIO), often suffers from performance degradation in degenerate environments, such as long corridors or single-wall scenarios with narrow field-of-view LiDAR. To address this limitation, we propose ALIVE-LIO, a degeneracy-aware LiDAR-inertial odometry framework that explicitly enhances state estimation in degenerate directions. The key contribution of ALIVE-LIO is the strategic integration of a deep neural network into a classical error-state Kalman filter (ESKF) to compensate for the loss of LiDAR observability. Specifically, ALIVE-LIO employs a neural network to predict the body-frame velocity and selectively fuses this prediction into the ESKF only when degeneracy is detected, providing effective state updates along degenerate directions. This design enables ALIVE-LIO to utilize the probabilistic structure and consistency of the ESKF while benefiting from learning-based motion estimation. The proposed method was evaluated on publicly available datasets exhibiting degeneracy, as well as on our own collected data. Experimental results demonstrate that ALIVE-LIO substantially reduces pose drift in degenerate environments, yielding the most competitive results in 22 out of 32 sequences. The implementation of ALIVE-LIO will be publicly available.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

ALIVE-LIO adds selective fusion of a learned body-frame velocity into ESKF only on detected degeneracy, and the reported gains on 22/32 sequences are plausible but rest on unshown details about the network and detector.

read the letter

The main thing to know is that this paper trains a neural net to predict velocity and feeds it into a standard ESKF-based LIO only when a degeneracy detector flags that LiDAR observability has dropped in certain directions. That targeted hand-off is the actual addition on top of existing work, and the experiments claim it cuts pose drift enough to come out ahead on 22 of 32 sequences from public and custom data. They also say the code will be released, which helps anyone who wants to check the numbers themselves. What works is the decision to keep the probabilistic filter rather than swap it for a full learned odometry system; that keeps consistency properties intact while patching the practical failure mode in corridors or narrow-FOV setups. The hybrid angle is straightforward and addresses a recurring pain point in real deployments. The soft spots are the missing pieces that make the central claim hard to evaluate from the abstract alone. There is no description of the network architecture, training data or loss, the exact degeneracy metric, or how the NN covariance is aligned with the ESKF innovation. Without those, it is difficult to confirm that the velocity predictions are accurate precisely on the axes the detector marks as unobservable, or that false triggers do not inject bias. The stress-test concern about consistency under mismatch is fair and not obviously resolved in the summary. The results are presented as competitive, but without error bars, per-sequence breakdowns, or more explicit baseline comparisons the improvement could be smaller than it appears once the full implementation is examined. This paper is for robotics researchers who already run ESKF LIO and need a lightweight way to handle degeneracy without rebuilding the whole pipeline. A reader who cares about hybrid classical-learning filters would find the selective-fusion pattern worth looking at. I would send it to peer review; the idea is concrete enough and the problem is practical enough that referees can usefully press on the implementation details and consistency arguments.

Referee Report

3 major / 2 minor

Summary. The paper proposes ALIVE-LIO, a degeneracy-aware LiDAR-inertial odometry framework that integrates a deep neural network to predict body-frame velocity and selectively fuses these predictions into an error-state Kalman filter (ESKF) only when degeneracy is detected, aiming to reduce pose drift in environments like corridors or single-wall scenarios where LiDAR observability is lost. It evaluates the method on public and self-collected datasets, claiming the most competitive results in 22 out of 32 sequences.

Significance. If the central experimental claims hold after addressing the gaps in validation, the work would offer a practical hybrid approach that preserves the probabilistic consistency of classical ESKF while leveraging learned velocity estimates for degenerate directions, potentially improving robustness for robotics applications in challenging real-world settings.

major comments (3)

[Abstract and §4] Abstract and §4: The headline claim of superior performance on 22 of 32 sequences is presented without architecture details for the neural network, training procedure, explicit degeneracy metric, error bars on pose errors, or full baseline comparisons, leaving the load-bearing experimental support for the fusion strategy unverified.
[§3.2–3.3] §3.2–3.3: The degeneracy detector and selective fusion rule into the ESKF are described at a high level, but no analysis or ablation is given on false-positive rates (which could inject biased updates) or false-negative rates (which would leave drift uncorrected), directly undermining the consistency guarantee of the filter.
[Table 1 and §4.2] Table 1 and §4.2: Reported pose errors lack per-axis breakdowns or covariance consistency checks, so it is impossible to confirm that the NN predictions are accurate precisely along the LiDAR-unobservable directions flagged by the detector.

minor comments (2)

[§2] §2: The state vector and covariance notation in the ESKF formulation would benefit from an explicit variable table to improve readability.
[Figure 3] Figure 3: The degeneracy visualization could include quantitative thresholds or example innovation values to clarify when fusion is triggered.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback. We will revise the manuscript to supply the requested details on the neural network, detector performance analysis, and error reporting. Our point-by-point responses follow.

read point-by-point responses

Referee: [Abstract and §4] The headline claim of superior performance on 22 of 32 sequences is presented without architecture details for the neural network, training procedure, explicit degeneracy metric, error bars on pose errors, or full baseline comparisons, leaving the load-bearing experimental support for the fusion strategy unverified.

Authors: We agree that these details are necessary for full verification. In the revised manuscript we will add the neural network architecture and training procedure to §3.1, state the explicit degeneracy metric in §3.2, include error bars on all pose-error tables, and expand the baseline comparisons in §4 to directly support the selective-fusion claims. revision: yes
Referee: [§3.2–§3.3] The degeneracy detector and selective fusion rule into the ESKF are described at a high level, but no analysis or ablation is given on false-positive rates (which could inject biased updates) or false-negative rates (which would leave drift uncorrected), directly undermining the consistency guarantee of the filter.

Authors: We accept this criticism. We will insert a new ablation subsection in §4 that reports false-positive and false-negative rates of the degeneracy detector on all evaluated sequences and quantifies their effect on filter consistency, thereby confirming that selective fusion does not compromise the ESKF's probabilistic guarantees. revision: yes
Referee: [Table 1 and §4.2] Reported pose errors lack per-axis breakdowns or covariance consistency checks, so it is impossible to confirm that the NN predictions are accurate precisely along the LiDAR-unobservable directions flagged by the detector.

Authors: We will revise Table 1 to provide per-axis error breakdowns and add covariance-consistency analysis (NEES statistics) in §4.2. These changes will explicitly show that the network predictions are accurate in the directions identified as degenerate by the detector. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected in derivation chain

full rationale

The paper's core approach integrates a separately trained neural network for body-frame velocity prediction into a standard ESKF, with selective fusion triggered by a degeneracy detector. No load-bearing step reduces by the paper's own equations to a fitted parameter renamed as prediction, a self-definition, or a self-citation chain. The ESKF structure, degeneracy detection logic, and NN training remain independent of the target pose-drift results, satisfying the criteria for a self-contained derivation.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The method assumes standard Kalman filter consistency properties and relies on a neural network whose weights are fitted to training data; no new entities are postulated.

free parameters (1)

degeneracy detection threshold
Parameter used to decide when LiDAR observability is lost and the network prediction should be fused.

axioms (1)

domain assumption ESKF maintains consistent probabilistic estimates when fused measurements are available and unbiased
Invoked implicitly when the selective fusion is claimed to preserve filter properties.

pith-pipeline@v0.9.0 · 5554 in / 1171 out tokens · 36712 ms · 2026-05-13T20:35:41.674414+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

7 extracted references · 7 canonical work pages

[1]

FusionPortableV2: A unified multi-sensor dataset for generalized slam across diverse platforms and scalable environments,

H. Wei, J. Jiao, X. Hu, J. Yu, X. Xie, J. Wu, Y . Zhu, Y . Liu, L. Wang, and M. Liu, “FusionPortableV2: A unified multi-sensor dataset for generalized slam across diverse platforms and scalable environments,”The Int. J. Robot. Res., vol. 44, no. 7, pp. 1093–1116, 2025

work page 2025
[2]

Hong Kong UrbanNav: An open-source multisensory dataset for benchmarking urban navigation algorithms,

L.-T. Hsu, F. Huang, H.-F. Ng, G. Zhang, Y . Zhong, X. Bai, and W. Wen, “Hong Kong UrbanNav: An open-source multisensory dataset for benchmarking urban navigation algorithms,”NA VIGATION: J. Inst. Navig., vol. 70, no. 4, 2023

work page 2023
[3]

UrbanNav: An open-sourced multisensory dataset for benchmarking positioning algorithms designed for urban areas,

L.-T. Hsu, N. Kubo, W. Wen, W. Chen, Z. Liu, T. Suzuki, and J. Meguro, “UrbanNav: An open-sourced multisensory dataset for benchmarking positioning algorithms designed for urban areas,” inProc. Int. Tech. Meet. Satell. Division of the Inst. Navig., St. Louis, Missouri, Sep. 2021, pp. 226–256

work page 2021
[4]

Heterogeneous LiDAR dataset for benchmarking robust localization in diverse degenerate scenarios,

Z. Chen, Y . Qi, D. Feng, X. Zhuang, H. Chen, X. Hu, J. Wu, K. Peng, and P. Lu, “Heterogeneous LiDAR dataset for benchmarking robust localization in diverse degenerate scenarios,”The Int. J. Robot. Res., 2025

work page 2025
[5]

R3LIVE: A robust, real-time, RGB-colored, LiDAR-inertial-visual tightly-coupled state estimation and mapping pack- age,

J. Lin and F. Zhang, “R3LIVE: A robust, real-time, RGB-colored, LiDAR-inertial-visual tightly-coupled state estimation and mapping pack- age,” inProc. IEEE Int. Conf. Robot. Automat., 2022, pp. 10 672–10 678

work page 2022
[6]

Fast-LIVO: Fast and tightly-coupled sparse-direct LiDAR-inertial-visual odometry,

C. Zheng, Q. Zhu, W. Xu, X. Liu, Q. Guo, and F. Zhang, “Fast-LIVO: Fast and tightly-coupled sparse-direct LiDAR-inertial-visual odometry,” inProc. IEEE/RSJ Int. Conf. Intell. Robot. Syst., 2022, pp. 4003–4009

work page 2022
[7]

Fast-LIVO2: Fast, direct LiDAR–inertial–visual odometry,

C. Zheng, W. Xu, Z. Zou, T. Hua, C. Yuan, D. He, B. Zhou, Z. Liu, J. Lin, F. Zhu, Y . Ren, R. Wang, F. Meng, and F. Zhang, “Fast-LIVO2: Fast, direct LiDAR–inertial–visual odometry,”IEEE Trans. Robot., vol. 41, pp. 326–346, 2025

work page 2025