Does odometry require wheels?

No. Visual and other motion-estimation methods can supply odometry.

Can it work without a prebuilt map?

Yes, although the result is relative and may drift.

Does combining sensors guarantee accuracy?

No. Incorrect calibration, correlated errors, and unsuitable models can still produce misleading estimates.

AI Glossary / Reinforcement learning and robotics / Intermediate

Odometry

Definition

Odometry is the estimation of a robot's change in position and orientation from measurements of its motion over time. It builds a relative movement estimate by combining successive increments. Wheel encoders are a common source, but cameras, inertial sensors, and other sensing arrangements can also contribute. The resulting estimate generally accumulates error unless additional information constrains it.

What it means

A robot often needs to know how far it has moved since the last control update, even when it has no reliable absolute position fix. Odometry supplies that local movement history. Starting from an initial pose, the system repeatedly estimates a small displacement and composes it with the previous pose.
This is different from simply counting total distance. Orientation determines the direction in which movement changes position. It is also different from knowing the robot's location accurately on a building map: a smooth relative estimate can gradually drift away from the true mapped position.

How it works

Wheel odometry converts measured wheel rotations into traveled distances using wheel geometry. A vehicle motion model then converts those distances into changes in the robot's pose. For an ideal differential-drive robot, the average of the left and right wheel distances gives the center's traveled arc length, while their difference divided by wheel separation gives the heading change.
The computation assumes the wheel measurements and model describe actual motion. Slipping, wheel deformation, incorrect dimensions, and timing errors violate that assumption. Small systematic errors can accumulate substantially over a long route.
Visual odometry estimates motion from successive camera observations. Inertial measurements can help estimate changes in motion, but biases and integration errors also require attention. Sensor fusion combines complementary information under an explicit estimation model.
Navigation systems often maintain a continuous local motion estimate while a separate localization process relates that estimate to a global map. These roles let local control remain smooth even when global position estimates are corrected.

Three practical examples

Delivery robots: wheel-based movement estimates help the controller track short motions between map-based position updates.
Inspection devices: visual odometry estimates camera movement through an area where satellite positioning is unavailable.
Mapping: relative motion estimates provide initial relationships between successive sensor observations before other mapping constraints refine the trajectory.

Think of it this way

Think of estimating your route by remembering each step and turn from a starting point. The method can work well over a short distance, but a slightly wrong stride length or missed turn gradually shifts the estimated location. Recognizing a known doorway supplies a different kind of information that can help correct the accumulated mistake.

Worked example

Consider an ideal two-wheel differential-drive robot with wheel separation b = 0.50 meters. It starts at x = 0, y = 0, heading along the positive x-axis. During one interval, its left wheel travels 0.40 meters and its right wheel travels 0.60 meters, with constant wheel speeds and no slipping.
The center's arc length is d = (0.40 + 0.60)/2 = 0.50 meters. The heading change is Δθ = (0.60 − 0.40)/0.50 = 0.40 radians, about 22.92 degrees counterclockwise. Because the wheel speeds are constant, the center follows a circular arc with radius R = d/Δθ = 1.25 meters.
The exact pose increment for these assumptions is Δx = R sin(Δθ) = 1.25 sin(0.40), approximately 0.4868 meters, and Δy = R[1 − cos(Δθ)], approximately 0.0987 meters. The final estimated pose is therefore approximately (0.4868, 0.0987, 0.40).
Adding the entire 0.50 meters to x and then turning would instead produce (0.50, 0, 0.40), which describes a different sequence of movement. The center's arc length is also not the same as its straight-line displacement, approximately 0.4967 meters.
If both wheels had traveled 0.50 meters equally, the heading change would be zero and the robot would move straight to (0.50, 0, 0). Actual wheel slip could invalidate either estimate even if encoder readings were recorded perfectly.

Illustrative case study

Fictional case: WillowDelivery finds that a robot's estimated position slowly shifts to one side after repeated corridor runs. The team checks wheel dimensions, encoder scaling, timestamps, and the surface conditions under which the drift occurs. It compares the relative trajectory with independent position measurements instead of assuming every discrepancy comes from the map.
After changing a calibration parameter, the team repeats straight, turning, and closed-loop routes under representative loads. This fictional example illustrates a diagnostic process, not a measured improvement or a claim that calibration eliminates all odometry error.

Common misconceptions and mistakes

An encoder measures wheel rotation, not guaranteed ground displacement. A higher update rate does not remove a wrong wheel radius. Smooth output does not prove accurate global location, and resetting the starting coordinates does not fix the underlying error source. Odometry is also not synonymous with SLAM: SLAM jointly estimates a trajectory and a map, potentially using odometry as one input.

Business use

Reliable local movement estimates support navigation quality, docking consistency, and investigation of motion problems. Evaluate both translation and heading errors over relevant routes, surfaces, speeds, and payloads. Record sensor timing and calibration settings so failures can be reproduced. Where operations depend on a global location, assess the localization and correction process as well as the quality of relative odometry.

Check your understanding

Why does the example robot turn left?

Answer: The right wheel travels farther than the left.

What determines the direction of a position update?

Answer: The robot's orientation during that movement.

Why is the final x-coordinate less than 0.50 meters?

Answer: The center moves along an arc rather than entirely along the x-axis.

Related terms

Pose
Localization
wheel encoder
differential drive
visual odometry
inertial measurement unit
sensor fusion
dead reckoning
SLAM
coordinate frame

Learn next

Next, examine how heading uncertainty affects position over longer distances. Then compare wheel calibration errors with slipping and learn how external observations can constrain accumulated drift.

Sources

PAL Robotics, Odometry glossary entry: https://docs.pal-robotics.com/edge/glossary Caltech ME/CS 133, Kinematics and Odometry for a Differential Drive Vehicle: https://robotics.caltech.edu/wiki/images/d/d3/DifferentialDriveKinematics.pdf Nav2 documentation, State Estimation: https://docs.nav2.org/rolling/getting_started/navigation_concepts/state_estimation/

Independent review checked no-slip kinematics and relative-pose scope.

Previous: Odds ratio Next: Odor source localization