Localization, Mapping & SLAM

robot localization

Robot localization is a robot figuring out where it is on a map it already has. Imagine waking up in a hotel and being handed the floor plan, but with no dot saying you are here. You glance around — a window on your left, a long hallway ahead, an exit sign to the right — and match what you see against the plan until only one spot fits. That match, the robot's best guess of its position and which way it is facing, is what localization delivers. In robotics this guess of position-plus-heading is called the robot's pose.

It works by comparing two things: what the robot's sensors report right now, and what the map says it should see from each possible spot. A laser scanner might report a wall two meters ahead and a doorway to the left; the robot asks, of all the places on this map, which ones would produce that exact view? Usually several spots roughly fit, so the robot does not keep a single answer but a spread of likely poses, and every time it moves and senses again, the spread tightens around the true location. Wheel counters and motion sensors say how far it rolled, while the laser or camera corrects the drift that those counters slowly accumulate.

The reason it matters is that almost nothing else a robot does works without it. A delivery robot cannot route to room 304, a warehouse robot cannot stop at the right shelf, and a vacuum cannot avoid cleaning the same strip twice unless each first knows where it stands. Crucially, localization assumes the map is already given and trusted; the harder cousin where the robot must build the map and locate itself at the same time is SLAM.

A museum guide robot starts each morning at its charging dock, then drives the halls. By matching its laser readings to the building map it always knows it is, say, four meters into Gallery 2 facing the Monet — so it can stop and start its tour at exactly the right painting.

With a trusted map in hand, the robot's only job is to find the one spot that explains what it sees.

A special hard case is the kidnapped-robot problem: the robot is switched off, carried somewhere random, and switched back on with no idea where it is — it must localize from scratch with no clue about its starting point.

Also called
pose estimation in a known map位置估计定位