KAIST's SE(3)-LIO Cuts Drone Navigation Error by 2x in Rough Terrain
A new LiDAR-inertial odometry method jointly propagates rotation and translation on SE(3), cutting trajectory error by more than half on rough terrain.
ProTakeaways
- SE(3)-LIO propagates 6-DoF poses jointly on the SE(3) manifold instead of separating rotation and translation.
- Uncertainty-aware motion compensation accounts for correlation between predicted poses when undistorting LiDAR points.
- Cuts trajectory error more than 2x versus FAST-LIO2 on a 2.4 km rough-terrain sequence.
- Runs at about 20 ms per scan, competitive with the fastest baselines while beating accuracy across three datasets.
- Accepted at ICRA 2026; code is open-source and built on the FAST-LIO2 and VoxelMap frameworks.
- Currently limited to LiDAR plus IMU; visual-inertial and radar extensions are listed as future work.
Autonomous drones, ground robots, and handheld scanners all rely on the same trick to know where they are: fuse a fast inertial measurement unit (IMU) with a slower LiDAR. The IMU predicts motion between LiDAR scans and corrects distortion in the point cloud caused by the sensor moving during a sweep. A new method called SE(3)-LIO from researchers at KAIST and Hanwha Aerospace rethinks how that IMU integration is done, and the payoff shows up in the numbers.

Where standard IMU integration breaks down
Most modern LiDAR-inertial odometry (LIO) stacks, including FAST-LIO2 and LIO-SAM, follow the preintegration recipe: represent a 6-DoF pose as rotation and translation on separate manifolds, then integrate each independently. Because rotational variation never feeds into how position advances, a robot that rotates and moves at the same time (which is basically always) sees its predicted position drift.
The second issue is more subtle. To undo the smear that appears in a LiDAR sweep, each point is transformed back to a common reference pose using a relative transformation between two IMU-predicted poses. Those predicted poses carry errors, so the relative transformation has uncertainty too. MA-LIO was the first method to incorporate that uncertainty into measurement noise during motion compensation, but it overestimates the uncertainty by neglecting the correlation between predicted poses.