The Many Maps of the Same Place

If you walk through Harbor Yard with a surveyor’s GNSS, a LiDAR scanner, a camera, and an OpenStreetMap editor, you will produce four different descriptions of the same forty acres. Each will be internally consistent. Each will be correct, by its own rules. And none of them will be interchangeable with the others.

That is not a problem to fix. It is the fundamental fact the entire field of HD mapping is built on top of.

The map you know

A navigation app shows you roads, names, and turn arrows on your phone. That map is designed for one task: helping a human driver decide where to turn. It does not need to know the exact width of the lane, the curvature of the curb, or the legal meaning of a faded yellow line. It only needs to be accurate enough that you recognize the intersection when you reach it.

That map is a geographic map. It is optimized for human recognition, rendered at screen scale, and tolerant of errors measured in meters. It is the most familiar representation. It is also the least useful for a robot.

The map a sensor builds

Now walk the same ground with a LiDAR scanner. The scanner does not see roads, names, or turn arrows. It sees points: millions of individual distance measurements, each one a coordinate in 3D space. Together those points form a point cloud, a dense geometric record of every surface the laser could reach.

A point cloud tells you, with centimeter precision, where the physical surfaces are. The warehouse wall. The asphalt. The loading dock edge. But it does not tell you which surfaces are roads and which are sidewalks. It does not tell you which side of the yellow line is legal to drive on. It records geometry without meaning.

The map a robot needs

Between the geographic map and the point cloud sits the map a robot actually needs. It needs precision (centimeters, not meters) and it needs meaning (semantics and rules, not just geometry). The formats that carry both, Lanelet2 and OpenDRIVE among them, are the subject of the middle third of this book.

For now, hold a simpler thought: a robot planning a route through Harbor Yard needs to know where the one-way loop begins, where it ends, and which dock faces the inbound lane. That information is not in the point cloud and not in the navigation app. It lives in a third representation, the semantic road network, which encodes lanes, connections, and rules as structured data.

Notice what that bundle implies: in formats like Lanelet2 and OpenStreetMap, the traced line and the logical tag live in the same file. The trace is the geometry; the tag is the meaning. The next chapter separates those into adjacent pipeline stages, and keeping them apart in your head will matter later.

The 3D model

Some use cases go further. A simulation environment needs not just the road network but a textured 3D model of every building, every container, every sign. A digital twin for site operations might layer live sensor feeds on top of that model. OpenUSD is the format built for such layered scene descriptions, and Part IV puts it to work as the structure behind the custom site assets you layer onto the map.

Each of these representations adds something the previous one did not carry. None of them replaces the others.

Why this matters now, not later

The instinct when encountering a new domain is to ask “which format should I learn first?” The question is premature. The right question is “which of my robot’s questions does this format answer?”

At Harbor Yard, a point cloud answers “what is physically here.” The Lanelet2 map answers “where can the tug go, and what are the rules along the way.” The OpenDRIVE export answers “how do I run this in a simulator first.” The OpenStreetMap extract answers “what did the world already know about this place before we arrived with a scanner.”

Four formats. Four different jobs. Same forty acres.

What comes next

The next chapter draws the full pipeline: Reality, to Capture, to Representation, to Semantics, to Rules, to Planning, to Motion. That pipeline is the master diagram for the rest of this book. Each Part that follows unpacks one stage. By the end of Part I, you will have the whole shape in your head before you touch a single format.