LiDAR odometry

On this page 7

LiDAR odometry is the process of estimating a robot’s incremental motion and pose by aligning consecutive laser range scans and computing the rigid transformation between them. In robotics perception, it serves as a geometric motion sensor that works in darkness, ignores surface texture and does not depend on wheel contact with the ground.

Unlike a full mapping system, LiDAR odometry only answers how far and in which direction the sensor moved since the previous scan. Because each estimate is chained onto the last, small errors accumulate as drift, which is why it is usually paired with loop closure, global localization or sensor fusion in a complete navigation stack.

How LiDAR odometry estimates motion

LiDAR odometry works by registering each new point cloud against a reference, either the previous scan or a local map built from recent scans. The registration step searches for the rotation and translation that best overlays matching geometry, and that transform becomes the motion estimate for the interval between scans.

A typical pipeline runs in a fixed sequence at the scan rate of the sensor, commonly 10 to 20 Hz for spinning units. Each stage has a distinct purpose, and weaknesses in any one of them propagate directly into the pose estimate.

StageWhat it doesWhy it matters
PreprocessingRemoves noise, ego returns and out-of-range pointsReduces false correspondences
Motion compensationCorrects points captured during sensor movementPrevents warped scans at speed
DownsamplingApplies voxel grids or feature selectionKeeps runtime within the scan period
RegistrationFinds the transform aligning scan and referenceProduces the core motion estimate
Map updateInserts the aligned scan into a local mapProvides a stable reference for the next scan
Stages of a typical LiDAR odometry pipeline

Motion compensation, also called deskewing, is often underestimated. A spinning LiDAR needs roughly 50 to 100 ms to complete one sweep, so a vehicle moving at 10 m/s travels up to a meter during a single scan. Without correcting each point by its timestamp, walls appear bent and registration converges to a biased solution.

Registration methods and algorithm families

The choice of registration method defines the accuracy, speed and robustness of LiDAR odometry. Most approaches descend from the Iterative Closest Point (ICP) algorithm, which alternates between finding nearest-neighbor correspondences and solving for the transform that minimizes their distance.

Practical systems refine this idea with better error metrics and smarter point selection. Point-to-plane ICP minimizes distance along local surface normals and converges faster than point-to-point matching, while feature-based methods extract edges and planar patches to reduce computation.

ApproachCore ideaStrengthWeakness
Point-to-point ICPMinimizes distances between matched pointsSimple and generalSlow convergence, needs good initial guess
Point-to-plane ICPMinimizes distance along surface normalsFast and accurate on structured scenesRequires reliable normals
Generalized ICPModels local covariance on both cloudsRobust to sparse or noisy dataHigher computational cost
NDTMatches points against voxel distributionsSmooth cost function, wide basinSensitive to voxel size
Edge and plane featuresAligns extracted geometric featuresReal-time on modest hardwareDegrades in unstructured terrain
Common registration approaches compared

The Normal Distributions Transform (NDT) represents the reference as Gaussian cells rather than raw points, which gives a smoother optimization landscape. Feature-based pipelines such as the LOAM family popularized the split between a fast odometry thread and a slower mapping thread, a design that many modern systems still follow.

LiDAR odometry is one of several relative motion estimators, and each has a characteristic failure mode. Understanding these differences explains why production robots almost always combine more than one source rather than trusting a single estimator.

MethodMeasured quantityTypical driftMain failure mode
LiDAR odometryScene geometryAbout 0.5 to 2 percent of distanceGeometrically degenerate scenes
Visual odometryImage features or intensitiesAbout 1 to 3 percent of distanceLow light, blur, textureless surfaces
Wheel odometryWheel rotationOften several percentSlip on loose or wet ground
Inertial integrationAcceleration and angular rateGrows quickly without correctionBias integration over time
Relative motion estimators in mobile robots

Drift figures depend heavily on sensor resolution, environment and tuning, so the ranges above indicate order of magnitude rather than guarantees. LiDAR odometry also differs from SLAM: odometry estimates local motion, whereas SLAM maintains a globally consistent map and corrects accumulated error through loop closure.

Configuring LiDAR odometry for real robots

Good results depend on matching parameters to the sensor and the environment. A setting that works in a warehouse may fail in a forest, so engineers usually tune against recorded datasets before deploying on hardware.

ParameterTypical rangeEffect of increasing it
Voxel size0.1 to 1.0 mFaster but loses fine geometry
Correspondence distance0.5 to 3 mWider convergence, more outliers
Maximum iterations10 to 50Better convergence, higher latency
Local map windowTens to hundreds of scansMore stable reference, more memory
Minimum range filter0.5 to 2 mRemoves robot body returns, loses close detail
Key tuning parameters and their effects

A strong initial guess dramatically improves convergence. Many systems therefore use an inertial measurement unit (IMU) for prediction and deskewing, creating LiDAR-inertial odometry. Tightly coupled variants estimate IMU biases jointly with pose and remain stable during aggressive motion where pure scan matching would fail.

  • Use per-point timestamps whenever the driver provides them.
  • Calibrate the LiDAR-to-IMU extrinsic transform carefully, because a few degrees of error causes systematic drift.
  • Monitor registration fitness scores and reject estimates that fall below a threshold.

Limitations and common failure modes

The most serious weakness of LiDAR odometry is geometric degeneracy. In long corridors, tunnels, open fields or along flat facades, the scan constrains motion in some directions but not others, so the solver cannot observe translation along the corridor axis.

SymptomLikely causeMitigation
Pose slides along a corridorDegenerate geometryDegeneracy detection and IMU or wheel fusion
Bent walls in the mapMissing motion compensationDeskew with timestamps and IMU
Sudden pose jumpsMoving objects or wrong matchesDynamic point removal and outlier rejection
Slow vertical driftSparse ground returnsGround constraints or barometric aiding
Noisy estimates in rain or fogSpurious returns from particlesIntensity and range filtering
Symptoms, causes and mitigations

Dynamic environments pose another risk, since crowds or traffic can dominate the scan and pull the solution toward their motion. Finally, unbounded drift is inherent: without loop closure or absolute references such as GNSS, position error keeps growing with distance traveled.

In plain words

A LiDAR sensor spins and measures distances to everything around it many times per second, building a three-dimensional snapshot of the surroundings. LiDAR odometry compares each new snapshot with the previous one and asks a simple question: how must the robot have moved for these two pictures to line up?

It resembles walking through a dark room with your hand on the furniture. By feeling how the edges shift, you know how far you moved. The method is precise over short distances, but tiny errors add up over long trips, so robots combine it with other sensors to stay accurate.

A real-world example

An inspection robot surveys a 400-meter underground utility tunnel where GNSS is unavailable and wet concrete makes wheels slip. Engineers mount a 32-channel spinning LiDAR running at 10 Hz together with an IMU and run tightly coupled LiDAR-inertial odometry with a 0.3 m voxel grid.

Along most of the tunnel, cable trays and side niches provide enough structure for reliable registration. In one featureless 60-meter stretch, the system detects degeneracy along the tunnel axis and temporarily weights wheel odometry more heavily. At the exit, the accumulated position error is about 1.5 meters, well under one percent of the distance, which is sufficient to tag each detected crack to the correct tunnel segment.