UKFSLAM

UKFSLAM, or Unscented Kalman Filter SLAM, is a probabilistic robotics method for simultaneous localization and mapping that uses the Unscented Kalman Filter to estimate a robot’s trajectory together with the positions of map landmarks. It maintains a single Gaussian belief over an augmented state that includes the robot pose and all observed landmarks.

In UKFSLAM, the state is propagated through a nonlinear motion model using the unscented transform. A set of sigma points is generated from the current mean and covariance, passed through the motion function, and reintegrated to produce the predicted mean and covariance. When measurements are received, a nonlinear observation model links the robot pose and landmark locations to observed quantities such as range and bearing. The UKF then propagates the sigma points through this measurement function and updates the state and covariance to reflect the new information. Landmarks are added when unseen observations are made of new features, expanding the state vector and covariance accordingly. Data association, or matching measurements to known landmarks, is a critical component of the update step.

UKFSLAM offers advantages over EKF-SLAM by better handling nonlinearities in motion and observation models, since it uses the unscented transform rather than first-order linearization. However, it is computationally intensive because the sigma-point propagation scales with the state dimension, which grows as more landmarks are mapped. This can limit scalability for large maps. Like other Gaussian-based SLAM approaches, UKFSLAM assumes a Gaussian posterior and can suffer from inconsistency if data association is poor or the model is imperfect. It remains a useful framework in robotics research and scenarios requiring improved nonlinearity handling over EKF-SLAM.