// the find
cocel-postech/genz-icp
GenZ-ICP: SOTA robust LiDAR odometry (IEEE RA-L 2025)
GenZ-ICP is a LiDAR odometry algorithm that extends KISS-ICP with adaptive weighting to handle degenerate environments — corridors, tunnels, and other geometry-poor spaces where standard ICP drifts or fails. Published in IEEE RA-L 2025, it targets robotics engineers who need reliable pose estimation in environments that break simpler odometry systems. The core novelty is detecting degeneracy on the fly and down-weighting affected DoFs rather than bailing out.
The degeneracy handling is the real contribution — pre-tuned configs for corridors, tunnels, and indoor/outdoor transitions ship out of the box, which saves the usual painful per-environment tuning. The C++ core with Python pybind bindings is the right architecture: fast where it needs to be, scriptable where convenient. Both ROS1 and ROS2 are supported with launch files and rviz configs, so dropping this into an existing robot stack is straightforward. pip install works now (as of May 2026), which eliminates the CMake-from-scratch barrier for evaluation.
The pre-tuned configs imply you still need to pick the right one for your environment — there's no automatic environment classification, so someone hitting a partially-degenerate mixed outdoor-corridor scene is back to manual tuning. The LIO extension (GenZ-LIO) is a separate paper and repo, so if you need IMU fusion you're on your own for integration. Documentation on what the adaptive weighting actually does to output covariance is thin — if you're fusing this into a larger state estimator, trusting the uncertainty estimates is risky without digging into the paper. Windows support is absent; it's Linux-only, which is fine for most robotics work but worth knowing.