Visual failure
Dense smoke reduces transmittance and contrast, invalidating sparse photometric residuals.
IROS 2026 · Pittsburgh
RADAR · LiDAR · VISION · IMU · WHEEL
FIRE-LIVWO
Robust LiDAR-Inertial-Visual-Wheel Odometry via Failure-Immune mmWave Radar Enhancement
1 China University of Mining and Technology · 2 Information Institute, Ministry of Emergency Management · † Corresponding author
Real-world underground deployment
Inspect full result ↗
01 · Motivation
When visibility fades,
geometry may fail next.
Underground coal mines combine abrupt visual failure from smoke and dust with weak geometric observability in long, self-similar corridors. FIRE-LIVWO treats them as different failure modes and responds with the sensor that remains informative.
Geometric failure
Long, smooth, repetitive tunnels leave LiDAR registration underconstrained along critical directions.
Coupled failure
Fixed fusion weights cannot safely handle observations whose reliability changes across modalities and time.
Abstract
FIRE-LIVWO is a tightly coupled multimodal odometry framework based on an iterated error-state Kalman filter. It unifies 4D mmWave radar, LiDAR, and visual features in a shared VoxelMap; adds radar Doppler velocity for smoke-filled scenes; and tightly couples wheel odometry with nonholonomic constraints and online lever-arm compensation for geometrically degenerate corridors. Online visual and geometric observability analysis dynamically adjusts measurement weights and fusion modes.
02 · Method
One estimator.
Five sensing streams.
All measurements meet inside one IESKF pipeline. A shared VoxelMap supports heterogeneous geometry and sparse direct vision, while the detection layer activates radar and wheel constraints only when their complementary information is needed.
Radar–LiDAR geometric residuals, photometric residuals, Doppler velocity, and wheel constraints update a unified state estimate.
Heterogeneous VoxelMap
Radar and LiDAR features share point-to-plane geometry with sparse visual map points in one map representation.
Radar Doppler velocity
Pointwise Doppler constraints retain motion observability when dense smoke compromises camera and LiDAR measurements.
Wheel odometry + NHC
Nonholonomic motion and online lever-arm compensation suppress drift along geometrically weak tunnel axes.
Dual observability tests
Visual transmittance and geometric Hessian spectra drive continuous, interpretable modality switching.
Nominal LIV switches to radar-assisted LIVR in smoke, wheel-assisted LIVW in weak geometry, or LIVRW when both failures coexist.
LiDAR + IMU + Vision
Add radar Doppler
Add wheel + NHC
Radar + wheel
03 · Results
One route.
Two failure boundaries.
A Husky A200 traverses three real underground tunnels containing dense smoke, feature-sparse sections, and strongly repetitive geometry. Twenty points measured by a total station provide trajectory reference.
FIRE-Full reaches the tunnel endpoint with the lowest average localization error; the evaluated ablations and baselines drift or fail under visual or geometric degradation.
Average localization error ↓
Full multimodal fusion wins the complete sequence.
Methods marked as failure cannot output the full trajectory.
FIRE-LIVWFailureFAST-LIVO2Failure
04 · Video
From smoke entry
to tunnel endpoint.
The one-minute IROS supplementary video shows the field platform, multimodal framework, adaptive switching, and side-by-side mapping behavior under real visual and geometric degradation.
05 · Citation
Build on
FIRE-LIVWO.
If this work helps your research, please cite our IROS 2026 paper.
@inproceedings{hu2026firelivwo,
author = {Kun Hu and Menggang Li and Kaidi Wu and
Zhiwen Jin and Yingjie Zhao and Chaoquan Tang and
Eryi Hu and Gongbo Zhou},
title = {{FIRE-LIVWO}: Robust {LiDAR}-Inertial-Visual-Wheel
Odometry via Failure-Immune mmWave Radar Enhancement},
booktitle = {2026 IEEE/RSJ International Conference on
Intelligent Robots and Systems (IROS)},
year = {2026}
}