Does localization always need a prebuilt map?

No. It can use external references or be part of a joint mapping process, although many indoor methods use an existing map.

What is the difference between pose and position?

Pose includes orientation as well as position.

Can a robot be confident and wrong?

Yes. Incorrect models, ambiguous observations, or implementation errors can produce overconfident estimates.

AI Glossary / Reinforcement learning and robotics / Intermediate

Localization

Definition

Localization is the process of estimating a robot's position and orientation relative to a reference frame or environment. It combines available motion and sensor information to determine where the robot is, often while representing uncertainty. In this glossary category, the term concerns robot pose estimation rather than adapting software or content for a language or region.

What it means

A robot needs more than a destination: it needs an estimate of its current pose to interpret maps, place sensor observations, and execute movements. Pose usually includes position and orientation, such as x, y, and heading for a robot moving on a flat floor.
Localization is an inference problem because measurements are incomplete and noisy. A camera can see similar-looking corridors, wheel motion can slip, and a landmark can be hidden. A useful estimate therefore includes an assessment of confidence or alternative possibilities, not merely a single displayed dot.

How it works

Many localization systems predict pose changes from motion information and then correct that prediction using observations. Measurements may come from range sensors, cameras, satellite positioning, inertial sensors, wheel encoders, or other references. The appropriate combination depends on the environment and task.
When a map is available, observations can be compared with what the robot would expect to see at candidate poses. An estimator balances those observations against the motion prediction and their uncertainty. Filters, particle methods, and optimization-based approaches offer different representations and computational trade-offs.
Position tracking starts near an estimated pose and follows its changes. Global localization must recover pose with much less initial knowledge. Recovery after an unexpected displacement is harder than routine tracking and should not be assumed from tracking performance alone.
Mapping estimates the environment; localization estimates the robot's pose. Simultaneous localization and mapping, or SLAM, addresses their joint estimation.

Three practical examples

Indoor service: a mobile robot aligns range measurements with a building map to estimate which corridor it occupies and which direction it faces.
Inspection: a robot combines motion estimates with recognizable environmental features to reduce accumulated drift.
Docking preparation: a navigation system uses an estimated pose to approach a docking area, with a more precise local procedure where the application requires it.

Think of it this way

Think of finding yourself on a hiking map. Counting steps gives an approximate change in location, while recognizing a bridge or junction can correct the estimate. If several junctions look alike, there may be multiple plausible locations. A robot faces the same distinction between knowing how it moved and knowing where that movement places it in the wider environment.

Worked example

Consider an original one-dimensional localization example. Before a new observation, a robot's position x has a Gaussian estimate with mean 4 meters and variance 1 square meter. An independent position sensor reports z = 6 meters with Gaussian measurement-error variance 3 square meters. Assume the sensor measures x directly and has no bias.
The scalar Kalman gain is K = 1/(1 + 3) = 1/4. The updated mean is 4 + (1/4)(6 - 4) = 4.5 meters. The updated variance is (1 - 1/4) times 1 = 0.75 square meters, giving a standard deviation of approximately 0.866 meters.
The estimate moves only half a meter toward the sensor report because the prior is more precise than that measurement. Equal averaging would give 5 meters, which is not the correct update under these stated uncertainties.
If the measurement-error variance were instead 0.25 square meters, K would be 1/(1 + 0.25) = 0.8. The updated mean would become 5.6 meters and the variance 0.2 square meters. A more precise observation receives more weight.
These calculations illustrate uncertainty-weighted correction, not a full robot localization stack. Real pose estimates include orientation and often correlated variables. Reusing the same observation as if it were an independent new reading would double-count evidence and create unjustified confidence. An incorrect noise model or unmodeled sensor bias can also make the reported uncertainty misleading.

Illustrative case study

Fictional case: Aspen Inspection deploys a test robot in a building with several similar corridors. The team observes that a single pose marker sometimes looks stable even when the robot is aligned with the wrong corridor.
Engineers evaluate localization against independently measured reference positions and inspect multiple-hypothesis behavior during initialization. They add tests for changed furniture, obscured landmarks, and unexpected repositioning, then define how navigation should respond when localization is uncertain.
This fictional case illustrates why a convincing display is insufficient evidence of correct localization. It does not report a successful deployment, a measured accuracy improvement, or a guarantee that the robot can recover from every disturbance.

Common misconceptions and mistakes

Localization is not synonymous with GPS: satellite positioning is only one possible source of information. Odometry estimates motion and can drift; it does not automatically establish an accurate global pose. A map can be correct while the robot is incorrectly localized within it.
Small reported uncertainty does not prove that the estimator is correct, especially under wrong assumptions or repeated structures. More sensors do not help automatically if their errors are correlated, their timing is wrong, or their coordinate transforms are misconfigured. Position accuracy also does not substitute for orientation accuracy.

Business use

Localization supports dependable navigation, repeatable inspection, and meaningful association of observations with places. Evaluation should match operational needs: a robot crossing a broad room and one aligning with a narrow station may need different accuracy. Teams should measure errors against independent references, test difficult initialization and recovery cases, and examine how uncertainty affects downstream behavior. Typical-case accuracy alone can hide operationally important failures.

Check your understanding

Why does the first worked update move to 4.5 meters rather than 5 meters?

Answer: The sensor's error variance is three times the prior variance, so it receives a gain of one quarter. The update weights information by the stated uncertainties rather than averaging the two means equally.

Related terms

Learn next

Learn the distinction between position, orientation, and coordinate frames. Then study prediction and measurement correction before comparing Gaussian pose estimates with particle representations that can maintain several possible locations.

Sources

Nav2, State Estimation, localization, SLAM, and odometry: https://docs.nav2.org/rolling/getting_started/navigation_concepts/state_estimation/ Simo Sarkka, Bayesian Filtering and Smoothing (2013), Bayesian estimation and Kalman filtering: https://users.aalto.fi/~ssarkka/pub/cup_book_online_20131111.pdf

Independent review checked Nav2 positioning scope and Kalman correction.

Previous: Locality-sensitive hash Next: Locally adaptive regression kernel (LARK)