TL;DR
SE(3)-LIVOM is a robust, real-time SLAM system designed for legged robots navigating extreme environments, from narrow collapsed buildings to wide-open snowfields. By combining data from two LiDARs, a camera, and an IMU with a novel adaptive downsampling strategy and an online plane-based backend, it achieved the top spot in the IROS 2026 COMFORT Challenge with a remarkable average trajectory error of just 1.40 cm.
What Problem Does This Paper Address?
Legged robots operating in the "wild"—such as search-and-rescue sites or construction zones—face a dual challenge for state estimation. First, the environments are highly variable: a narrow corridor provides very few geometric constraints for a LiDAR, while an open field generates a massive amount of data that can overwhelm a real-time processor. Second, aggressive motion from walking or stumbling can cause standard algorithms to lose track.
Existing methods often struggle to balance computational speed with global consistency. A fixed downsampling resolution might be too coarse for indoor navigation (losing detail) or too fine for outdoor use (causing lag). SE(3)-LIVOM aims to solve this by creating a system that automatically adjusts its processing load to the environment while maintaining high precision through multi-sensor fusion.
Methods
The architecture of SE(3)-LIVOM is split into a multi-threaded frontend and backend, as illustrated in the pipeline below:

The Frontend: Adaptive Multi-Sensor Fusion
- Multi-LiDAR Integration: The system merges scans from two different LiDARs (a Hesai XT32 and a downward-looking Livox Mid-360). This setup is crucial for "localizability"; where one LiDAR might see only a featureless wall in a corridor, the other catches the floor or ceiling, preventing the system from "slipping."
- Adaptive Downsampling: This is the "secret sauce." Instead of using a fixed grid size, the system adjusts the voxel resolution at every scan. If the number of inlier points ($n_k$) deviates from a target ($N=2000$), the resolution is scaled up or down. This ensures the LiDAR update always takes roughly 35 ms, regardless of the scene's complexity.
- Visual Refinement: After the LiDAR update, a camera image is used for a "photometric update," refining the robot's state using visual patches attached to the LiDAR map points.
The Backend: Global Consistency
- Plane Bundle Adjustment (PBA): To correct long-term drift, the backend groups scans into submaps and performs a global optimization that aligns planes across these submaps.
- Incremental Pose Graph: Using iSAM2, the system integrates odometry, gravity (from the IMU), and inter-submap factors to provide a smooth, globally consistent trajectory at 10 Hz.
Datasets
The authors evaluated their system using the GrandTour dataset, a benchmark specifically designed for legged robots in challenging terrains. The dataset includes six distinct test scenes:
- ARC-2 & ARC-7: Collapsed buildings/corridors.
- CON-4: Construction sites.
- EIG-1: Indoor environments.
- SNOW-2: Open snowfields.
- SPX-2: Industrial stairs and structures.
Key Insights and Ideas
- Localizability is Key: The authors found that using a single LiDAR in narrow corridors led to significant drift (up to 19 seconds of "blindness"). Adding a second, differently-angled LiDAR reduced this vulnerability to just 3 seconds.
- Efficiency vs. Accuracy: Figure 2 in the paper (shown below) demonstrates that adaptive downsampling provides the "best of both worlds." It matches the accuracy of high-resolution fixed grids (0.1m) while maintaining the processing speed of low-resolution grids (0.5m).
- SE(3) Manifold Propagation: By performing IMU propagation directly on the SE(3) manifold (handling rotation and translation together), the system remains stable even during the jerky, aggressive movements typical of legged locomotion.

Key Results
SE(3)-LIVOM was the overall winner of the COMFORT Challenge. Quantitative highlights include:
- Precision: Achieved an average Absolute Trajectory Error (ATE) of 1.40 cm across all test scenes, outperforming the baseline FAST-LIVO2 (2.16 cm).
- Real-time Performance: The frontend consistently processed scans in 28 to 42 ms, fitting comfortably within the 100 ms scan period of the sensors.
- Robustness: The system used a single configuration for all scenes, proving it doesn't require manual tuning to move from a building interior to a snowfield.
Limitations and Future Work
While the system is highly effective, the authors noted a bottleneck in the backend. In very large or complex scenes (like CON-4), a single round of Plane Bundle Adjustment can take up to 19 seconds. Although the system is designed so that late backend updates don't delay the immediate pose output, this lag could eventually lead to larger drift before a correction is applied. The authors suggest that developing a faster backend for large-scale scenes is the primary direction for future work.
Summary updated