How to fix ICP that does not converge¶

The symptom¶

registration_icp returns a fitness of 0.9 and the clouds are visibly misaligned; or the fitness is 0.05 and nothing moved; or it converges to a mirrored pose. On the Stanford bunny against a copy rotated 35° with 70 % overlap and 0.5 % noise, the tutorial call ends 32.0° off with a fitness of 0.98 (the experiment).

The parameters behind it¶

Parameter (Open3D)	What goes wrong	What to set
`init` (initial transform)	ICP is a local optimiser; its basin is a few degrees and a fraction of the object size. From the identity, a 35° offset converges to the nearest wrong minimum.	A global stage first: FPFH features + RANSAC. That is what `align()` does.
`max_correspondence_distance`	Too wide (0.02 m on a 0.15 m bunny): every point finds a partner, fitness looks excellent, the pose is wrong. Too tight: no correspondences, nothing moves.	Tie it to the data: 1.5× a voxel of 1.3 % of the bounding-box diagonal for the coarse pass, 0.4× for the fine pass.
`estimation_method`	Point-to-point zigzags and stalls on smooth surfaces.	Point-to-plane with target normals; point-to-point only when normals are unreliable (sparse LiDAR).
units	Millimetres read as metres make every distance 1000× too small.	`obj.frame.units`; `with_frame("sensor", "mm")`.
overlap	The 30 % of points without a counterpart pull the solution.	A tight fine-pass threshold; judge by inlier RMSE on the overlap, not by fitness.
symmetry	Cylinders, spheres, vertebral bodies have several equally good poses.	`ransac_n=4`, a prior pose, more context in the scan.

The one-liner that already handles it¶

import pardesline as pl

target, moving = pl.load("scan_a.ply"), pl.load("scan_b.ply")
aligned = target.align(moving)
print(aligned.transform["ransac_fitness"], aligned.transform["inlier_rmse"])

align = registration.ransac_icp_pipeline: voxel from the bounding box, FPFH at 5× the voxel, RANSAC with the mutual filter at 1.5× the voxel, then point-to-plane ICP at 0.4× the voxel. Every value and its reason: pl.describe("registration.ransac_icp_pipeline").

Still off? Walk the diagnostics¶

ransac_fitness below 0.3: the features are meaningless. The voxel is wrong for the data: pass voxel_size = 2–5× the point spacing. Check the units first.
Good RANSAC, bad ICP: the fine threshold is too wide for the noise or the overlap: icp_threshold_factor=0.2–0.3.
Flat scene: point-to-plane slides along planes: icp_estimation_method="point_to_point".
Converges to a mirror: symmetric object; verify with a target registration error on landmarks the algorithm did not see.
Different scale: ransac_with_scaling=True once (photogrammetry vs scanner), then fix the scale and run rigid again.

Measuring instead of guessing¶

The registration_eval.py script applies a known transform with noise and partial overlap to any cloud and prints fitness, inlier RMSE, rotation and translation error, Chamfer and HD95: the way to tune on your data before trusting a number.