Skip to content
All projects
Robotics & Control · AI & MLNov 2026

Robust LiDAR Odometry Bench — Planned

A planned benchmark of LiDAR odometry where the geometry stops constraining the motion: my own point-to-point ICP odometry against KISS-ICP on MulRan driving data, with a degeneracy indicator logged for every scan match. No experiments have been run yet.

roadmap steps done: the question and scope
1 of 10
roadmap steps done: the question and scope
results so far, by design: the README updates as each stage is tested
0
results so far, by design: the README updates as each stage is tested

Why

LiDAR odometry works well where the scene constrains all six degrees of freedom. In long corridors, tunnels, bridges and open roads, the geometry leaves some directions unconstrained, and scan matching can drift or slip along them. I am learning SLAM fundamentals (scan matching, pose estimation, observability) by building this from scratch.

Questions

  • Where does a simple ICP odometry fail first, and how does that compare with a strong open baseline, KISS-ICP?
  • Can a degeneracy indicator computed from the ICP problem itself, such as the eigenvalues or condition number of its 6×6 normal matrix, predict where the error grows?
  • Does a simple mitigation, such as limiting the update along weak directions, reduce drift on those segments?

Planned method

  • Data: MulRan, urban driving sequences from an Ouster LiDAR with ground-truth trajectories, favouring long straight or tunnel-like segments
  • Own pipeline: point-to-point ICP in Python and NumPy, with voxel downsampling, a constant-velocity initial guess and a local map, written for clarity rather than speed
  • Baseline: KISS-ICP, run with default parameters first
  • Metrics: absolute trajectory error and relative pose error with the evo toolbox, per segment
  • Indicator: the smallest eigenvalue and condition number of the normal matrix, logged per frame and plotted against the error

Status

Planned, in progress. The question and scope are defined; next are choosing MulRan sequences, a data loader, and reproducing a KISS-ICP trajectory on one sequence. There are no results yet, and the repository will report a mitigation honestly if it does not help.

Limits to expect

  • One person, one dataset and a handful of sequences, so findings will not generalise on their own
  • My ICP will be slower and less accurate than KISS-ICP; the aim is to understand failure modes, not to beat it
  • MulRan's ground truth is itself estimated, so small errors sit within its noise