The Art of Spatial Segmentation

A hyper-dense point cloud is not a better map. It is a tax on every computation the robot runs. More points mean slower scan matching, higher memory pressure, and an increased risk that the localizer will match against something transient: a parked truck, a stack of pallets, a bush that grew three inches after a rainstorm.

The skill is knowing what to keep and what to throw away.

Signal and noise

Every point in a scan belongs to one of two categories:

Signal. Rigid, static geometry that remains identical visit after visit. Concrete curbs. Steel H-beams. Building corners. Utility poles. These are spatial anchors. The robot can count on them.

Noise. Transient or geometrically unstable data. Parked delivery vans. Stacks of wooden pallets. Seasonal vegetation. Sensor artifacts from rain, dust, or reflective surfaces. These exist today and are gone or shifted by Tuesday.

What happens when you keep the noise

When the scan matcher tries to align the robot’s live LiDAR feed against the map, it searches for correspondences between points. Every noise point in the map is a candidate for a false match. The matcher wastes cycles on irrelevant geometry, slows down, and in the worst case locks onto a false positive and reports high confidence in a wrong position.

A scan matcher that falls below its fitness threshold will warn you:

Warning: Scan match fitness score below threshold [0.45]

If you see this consistently, you are probably asking the robot to localize against noise.

Harbor Yard: the bushes that broke the shuttle

The north perimeter of Harbor Yard has a heavy-duty chain-link fence on one side and a row of decorative bushes on the other.

The fence is high-signal. Its posts are sunk in concrete. Its geometric signature is identical in July and January. It is ground truth.

The bushes are geometric lies. During the dry season they match the map. After three days of rain, they grow, sag, and shift. A robot trying to localize against them will weave, because its map says “bush at this position” and its LiDAR says “bush three inches left.”

The team that deployed the Harbor Yard shuttle learned this the hard way. The robot worked perfectly on Monday. By Friday, after a heavy rain on Wednesday, it was drifting unpredictably. The code was fine. The sensor was fine. The map was trying to localize against a biological growth that no longer matched the scan. They were treating noise as signal.

What to keep, what to discard

Keep (signal)	Discard (noise)
Concrete pillars and H-beams	Parked delivery trucks and vans
Building corners and wall facets	Stacks of pallets and crates
Utility poles and bollards	Seasonal vegetation and shrubs
Retaining walls and permanent fencing	Puddles, snowbanks, and debris
Fixed industrial signage	Moving pedestrians and site staff

Segmentation in practice

In Veer Studio, segmentation means running a pipeline that isolates vertical features, the geometry with high entropy that scan matching algorithms like NDT and ICP rely on. Flat pavement looks the same everywhere. A vertical pillar is a distinct fingerprint.

A clean segmentation pass tells you what you gained:

[INFO] Point cloud segmentation complete.
[INFO] Static points: 14,202
[INFO] Noise points removed: 89,401

A high noise-removal count is not a problem. It means you gave the robot the signal and threw away the fever dream.

Try this

[5 min] Look around the room or building you are in. Identify three objects that would be signal (rigid, permanent, geometrically distinct) and three that would be noise (transient, deformable, likely to move). For each noise object, imagine the robot trying to localize against it a week from now. What changed?