Kalman or Particle? Picking the Right State Estimator for Your Autonomous Robot

The short answer: if your robot's motion and sensor models are nearly linear and your noise is well-behaved, an Extended Kalman Filter (EKF) or Unscented Kalman Filter (UKF) will serve you well. If your system is strongly non-linear, your noise is non-Gaussian, or your robot must recover from losing track of where it is, a Particle Filter (PF) is the more robust choice—at a higher computational cost. This article lays out a decision framework for choosing among the three, focused specifically on mobile robot localization rather than filtering theory in general.

Introduction to State Estimation in Robotics

Every autonomous mobile robot must answer a deceptively hard question: where am I? Sensors are noisy, wheels slip, and motion models are only approximations of reality. A state estimator fuses these imperfect sources—odometry, range sensors, camera data—into a running belief about the robot's pose.

The choice of estimator is not academic. An IEEE comparison of Kalman and particle filters for underwater robot localization found that while both perform fairly well when abrupt sensor errors are rare, they diverge sharply under "jerky" sensor noise: the Kalman filter's location estimates hop around, while the particle filter continues to produce robust localization. That asymmetry—Gaussian assumptions versus arbitrary noise, single estimates versus multiple hypotheses—is what your decision ultimately hinges on.

Kalman Filters: EKF and UKF for Linearized Systems

Kalman filters represent the robot's belief as a single Gaussian distribution: one best estimate with one covariance describing its uncertainty. That elegance comes with two structural assumptions. The system model must be linear or linearizable, and the noise—both process noise and measurement noise—must be Gaussian. As Emma Benjaminson's technical explanation of particle filters puts it, Kalman filters "limit the user to thinking of the system as a Gaussian process, which may not be true in practice."

Real robots rarely fit that description exactly, so practitioners lean on two workarounds. The Extended Kalman Filter linearizes the non-linear motion and measurement models around the current state estimate. The Unscented Kalman Filter instead propagates a deterministic set of sigma points through the full non-linear model, capturing the non-linearity more faithfully without explicit linearization—generally more accurate than the EKF for non-linear systems, though still bound by the Gaussian assumption. Both degrade when the true uncertainty is multi-modal: two plausible corridors, two ambiguous doorways, one Gaussian can only average them into a confident, wrong answer.

Particle Filters: Handling Non-Linearity and Ambiguity

The particle filter drops the Gaussian representation entirely. Instead of one mean and covariance, it carries a cloud of weighted samples—particles—each representing a hypothesis about the robot's state. When the robot moves, each particle is shifted according to the control input, and noise is added to reflect uncertainty in how the state changed; as Benjaminson describes, the particles "don't all move the same amount," because the actions themselves carry uncertainty.

This sample-based representation is what gives particle filters their generality. Because the belief can take any shape, the filter handles non-linear motion models directly and non-Gaussian noise without approximation. Sebastian Thrun's survey of particle filters in robotics (presented at UAI) notes the approach's popularity in mobile robotics stems exactly from this flexibility—Monte Carlo Localization being its standard instantiation for the localization problem.

The trade-off is cost. Representing a complex distribution accurately requires enough particles, and the number needed grows with the dimensionality of the state space. Thrun's survey flags plain particle filters' "poor performance in higher dimensional spaces" as an often-criticized limitation, and notes practical fixes such as clustering particles with different resampling rates per cluster, which empirically increased robustness to map errors.

The 'Kidnapped Robot Problem' and Multi-Hypothesis Tracking

The sharpest illustration of where the approaches part ways is the kidnapped robot problem: the robot is lifted from its position and dropped somewhere else in the map, or equivalently, its accumulated belief is suddenly invalidated. A Gaussian filter committed to a single hypothesis has no principled way to rediscover itself—its covariance may be small and confidently wrong.

A particle filter can, because its particles can be spread across the map to re-explore candidate poses. But Thrun's survey adds an important qualification: generating particles from the motion model alone "has been recognized as insufficient under a range of conditions," including the kidnapped robot problem and cases where sensor accuracy is high relative to control accuracy. The practical remedy is hybrid sampling, in which a subset of samples is drawn directly from the measurement model. In data from a deployed mobile tour-guide robot, the extended hybrid-sampling version of Monte Carlo Localization outperformed plain MCL in exactly this kind of recovery scenario.

The lesson for your project: multi-hypothesis robustness is real, but it depends on particle diversity and your resampling strategy, not just on choosing PF over EKF.

Decision Framework: Choosing Your Robot's State Estimator

The matrix below classifies each filter against the four criteria that matter most for a localization stack: how it treats model non-linearity, what noise it assumes, how it handles ambiguity, and what it costs you computationally. Use it by matching your robot's characteristics—motion model, sensor behavior, failure modes—to each row's qualifications, not just its headline strengths.

State estimator decision matrix for mobile robot localization
Filter	System Model Linearity	Noise Distribution Assumption	Robustness to Ambiguity	Typical Computational Cost
Extended Kalman Filter (EKF)	Linearizes non-linear models around the current estimate; approximation can introduce errors in highly non-linear systems	Gaussian process and measurement noise; performance degrades significantly with non-Gaussian noise	Low—tracks a single hypothesis and cannot recover gracefully from invalidated beliefs	Low; scales with state dimension, no sampling population needed
Unscented Kalman Filter (UKF)	Captures non-linearity via deterministic sigma points, avoiding explicit linearization	Gaussian, propagated through the non-linear model; still limited by the Gaussian assumption	Low—still a single-hypothesis Gaussian belief	Moderate; more model evaluations than EKF, fewer than a large-particle PF
Particle Filter (PF)	Handles highly non-linear models directly; performance depends on particle count and resampling effectiveness	Arbitrary distributions, provided enough particles represent them accurately	High—maintains multiple hypotheses; recovery from failures like kidnapping depends on particle diversity and hybrid sampling	High; grows with particle count and degrades in high-dimensional state spaces

Read the matrix as a set of gates rather than a ranking. If your noise is genuinely Gaussian and your models are mildly non-linear, the EKF's or UKF's low cost wins. If a single row—non-Gaussian noise, multi-modal ambiguity, kidnapping-style failures—describes your environment, the particle filter's extra cost buys robustness the Kalman family structurally cannot provide.

Making the Right Choice for Your Robotic System

The selection of a state estimation filter for autonomous mobile robots should be driven by the inherent linearity of the system model, the nature of sensor noise (Gaussian vs. non-Gaussian), and the required robustness to localization ambiguities or failures. The IEEE comparison and Thrun's robotics survey converge on the same boundary: Kalman variants when the Gaussian, linearizable assumptions hold; particle filters when they do not.

Concretely, start your decision by characterizing your motion model's non-linearity and your sensors' error distribution under the conditions you actually deploy in—including the jerky, abrupt noise regimes where Kalman estimates hop and particle filters stay robust. The measurable indicator to watch is improved robot localization accuracy and reliability in your target environment, particularly under conditions of high uncertainty or non-ideal sensor data. If your first choice diverges under those conditions, the evidence above tells you which direction to move: toward more particles and hybrid sampling if robustness is the gap, or toward the cheaper Kalman variant if it is not.

Kalman or Particle? Picking the Right State Estimator for Your Autonomous Robot

Introduction to State Estimation in Robotics

Kalman Filters: EKF and UKF for Linearized Systems

Particle Filters: Handling Non-Linearity and Ambiguity

The 'Kidnapped Robot Problem' and Multi-Hypothesis Tracking

Decision Framework: Choosing Your Robot's State Estimator

Making the Right Choice for Your Robotic System

Sources

Tags

Related articles

How Industrial Robots Learn New Skills from Factory Workers: Architectures, Learning Loops, and Limits

Choosing Between LiDAR, Visual, and Multi-Modal SLAM for Your Industrial Mobile Robot

Choosing an RTOS for Robotics: How to Classify Your Tasks Before You Pick a Kernel

How Digital Twins Simulate, Optimize, and Predict Robotic Manufacturing Performance