Is Monte Carlo localization the same as SLAM?

No. Standard MCL localizes within a given map; SLAM jointly estimates pose and map.

Must particles lie on grid centers?

No. Their pose coordinates can be continuous.

Does resampling always improve accuracy immediately?

No. It reallocates the representation and introduces randomness; excessive resampling can reduce diversity.

AI Glossary / Reinforcement learning and robotics / Advanced

Monte Carlo localization

Definition

Monte Carlo localization is a robot localization method that represents uncertainty about pose using a set of weighted samples called particles. It updates those samples using motion and sensor models, typically with resampling. As a particle-filter implementation of probabilistic localization, it can represent several possible locations without storing a dense grid over the entire pose space.

What it means

Each particle is a hypothesis about where the robot is and how it is oriented. A cloud of particles can describe a concentrated estimate or several competing possibilities, such as similar-looking corridors.
The particles are not physical objects or separate robots. Their positions and weights approximate a probability distribution. A finite cloud can miss important possibilities, so increasing or concentrating particles is a computational choice with practical limitations rather than a guarantee of correct localization.

How it works

A common bootstrap-filter cycle starts with particles representing the current belief. The motion model moves each particle according to the measured or commanded movement, including modeled uncertainty. This produces a prediction of possible new poses.
Each predicted particle receives a weight based on how well the latest sensor observation agrees with the map at that pose. The weights are normalized. Resampling then selects particles in proportion to their weights, producing repeated copies of more supported hypotheses and fewer copies of weakly supported ones.
After resampling, weights are commonly reset to equal values; subsequent noisy motion spreads the hypotheses again. Resampling schedules and particle-count adaptation vary by implementation. Particle depletion, misleading sensor likelihoods, and missing support can prevent recovery.
A pose estimate may be extracted from a dominant cluster. Simply averaging widely separated hypotheses can place the reported robot in an implausible location between them.

Three practical examples

Global localization: particles begin broadly distributed across a known map while observations narrow the plausible starting poses.
Indoor tracking: a particle cloud follows a mobile robot using motion information and range observations.
Ambiguous environments: several particle clusters retain alternative corridor or orientation hypotheses until additional evidence distinguishes them.

Think of it this way

Think of placing many small markers on a map, each representing a possible location. Move the markers according to your uncertain travel, then favor those whose surroundings match what you observe. Some markers are copied and others disappear. If every marker near the true location disappears, simply copying the remaining wrong markers will not bring the truth back.

Worked example

Consider an original one-dimensional illustration with four equally weighted particles initially at positions -1, 0, 3, and 4 meters. Assume a deliberately simplified, deterministic motion of +1 meter. The predicted particles become 0, 1, 4, and 5 meters. Real localization normally models motion uncertainty as well.
Suppose a map-based sensor model assigns likelihoods 0.1, 0.2, 0.6, and 0.1 to the latest observation at these four predicted positions. Multiplying by the prior weight 1/4 gives 0.025, 0.05, 0.15, and 0.025. Their sum is 0.25, so the normalized weights are 0.1, 0.2, 0.6, and 0.1.
The weighted mean position is 0 times 0.1 + 1 times 0.2 + 4 times 0.6 + 5 times 0.1 = 3.1 meters. This is a summary of the particle distribution, not proof that the robot occupies that exact position.
For an illustrative systematic resampling step, choose initial offset 0.08 from the interval [0, 0.25). The four evenly spaced selection points are 0.08, 0.33, 0.58, and 0.83. Against cumulative weights 0.1, 0.3, 0.9, and 1.0, they select particles at 0, 4, 4, and 4 meters.
Assigning equal weights 1/4 now gives mean 3 meters for this resampled realization. It need not equal the weighted mean 3.1 exactly: resampling adds sampling variability. The original weights' effective sample size diagnostic is 1/(0.01 + 0.04 + 0.36 + 0.01) = 50/21, approximately 2.38. Four particles are used only to expose the arithmetic, not as a suggested deployment count.

Illustrative case study

Fictional case: Willow Navigation evaluates particle localization for a service robot in a building with repeated door layouts. During initialization, the particle cloud forms two distinct clusters instead of immediately identifying one corridor.
The team checks the clusters against independent reference positions and tests whether additional observations distinguish the alternatives. Engineers also evaluate unexpected repositioning and inspect whether the implementation introduces new hypotheses when the existing cloud is inconsistent with observations.
They select particle settings using accuracy, recovery behavior, and computing load together. This fictional case illustrates validation priorities; it does not claim that a particular particle count or algorithm reliably resolves every ambiguity.

Common misconceptions and mistakes

More particles do not automatically fix a wrong map, sensor calibration, or likelihood model. Resampling does not create new evidence, and repeated copies are not independent confirmations of a location. A low effective sample size measures weight concentration, not localization error directly.
A tight cloud can be confidently wrong. Recovery from a large unexpected displacement may need additional mechanisms beyond the basic filter. Finally, orientation requires appropriate treatment: ordinary arithmetic averaging can fail near the angular wraparound or across separate pose clusters.

Business use

Monte Carlo localization can support navigation where uncertainty is multimodal and a single Gaussian estimate is inadequate. Teams should assess startup time, position and orientation error, failure detection, recovery, and processing cost. Tests should include the actual map quality, sensor occlusions, and repeated structures expected in operation. A visually compact particle cloud should be treated as an estimator output to validate, not as evidence of success by itself.

Check your understanding

Why are three copies of the four-meter particle not three independent sensor observations?

Answer: All three descend from the same weighted hypothesis during resampling. The copies allocate representation to that hypothesis; they do not supply additional measurement evidence.

Related terms

Markov localization
Importance sampling
Resampling
Effective sample size
Localization
Particle filter
Robot pose
Adaptive Monte Carlo localization

Learn next

Work through prediction, weighting, and resampling with a few particles. Then compare resampling methods and examine how multimodal pose estimates differ from a single averaged location before exploring adaptive particle counts.

Resampling

Sources

Dieter Fox, Wolfram Burgard, Frank Dellaert, and Sebastian Thrun, Monte Carlo Localization: Efficient Position Estimation for Mobile Robots (1999): https://www.cs.cmu.edu/~thrun/papers/fox.aaai99.pdf Frank Dellaert, Dieter Fox, Wolfram Burgard, and Sebastian Thrun, Monte Carlo Localization for Mobile Robots (1999): https://sites.cc.gatech.edu/ai/robot-lab/online-publications/dellaert_frank_1999_2.pdf

Independent review checked particle weighting and recovery limitations.

Previous: Monte Carlo integration Next: Monte Carlo planning