Unscented Kalman Filter (UKF)

The unscented transform is the deterministic sampling technique at the heart of the UKF. Instead of linearizing nonlinear functions, it selects a minimal set of sigma points that capture the true mean and covariance of the state distribution. These points are propagated through the nonlinear process and measurement models, and the transformed points are used to compute a new mean and covariance. This provides a more accurate approximation of how the true probability distribution evolves, especially for highly nonlinear systems where the EKF's Jacobian-based linearization fails.

A major practical advantage over the Extended Kalman Filter (EKF) is that the UKF does not require the analytical derivation or numerical computation of Jacobian matrices. The EKF's linearization depends on these Jacobians, which can be complex to derive for intricate models and a source of implementation error. The UKF only requires the ability to evaluate the nonlinear functions f(x) and h(x) as black boxes. This simplifies implementation, reduces code complexity, and eliminates errors introduced by incorrect Jacobians, making the UKF more robust for rapid prototyping and deployment of complex models.

While the UKF assumes the final estimated state is Gaussian (represented by a mean and covariance), the unscented transform provides a more accurate capture of the posterior distribution's moments (mean and covariance) after nonlinear transformation. This makes it more effective than the EKF for systems where the state distribution becomes skewed or heavily transformed. For problems with significant non-Gaussian characteristics, the UKF often serves as a higher-performance alternative before one must resort to the computational expense of a Particle Filter. It achieves a favorable trade-off between accuracy and computational cost for many practical robotic estimation problems.

The UKF scales predictably with state dimension n. The standard algorithm uses 2n+1 sigma points. This linear scaling is often more computationally efficient than the Particle Filter, which may require an exponentially increasing number of particles to maintain accuracy in high dimensions. For state estimation in robotics—such as Visual-Inertial Odometry (VIO) or LiDAR-Inertial Odometry (LIO) where the state may include position, orientation, velocity, and sensor biases—the UKF provides a tractable solution. Its deterministic nature also avoids the sampling degeneracy issues that can plague particle filters.

The UKF is often more consistent than the EKF, meaning its estimated covariance more accurately reflects the true error of the state estimate. The EKF's linearization can lead to overly optimistic (too small) covariance estimates, causing the filter to become overconfident and ignore new measurements. By better approximating the true nonlinear transformation, the UKF maintains a more realistic uncertainty bound. This derivative-free property also makes it applicable to systems with discontinuities or hard constraints where derivatives are undefined, broadening its applicability in robotic control and planning.

The primary alternative, the Extended Kalman Filter, linearizes nonlinear models using a first-order Taylor expansion. This introduces error that grows with system nonlinearity. The UKF, in contrast, is accurate to the third order for Gaussian inputs (and at least second order for non-Gaussian) for capturing the posterior mean and covariance. For systems with strong nonlinearities in either the process model (e.g., aggressive drone dynamics) or measurement model (e.g., bearing-only sensors), the UKF typically provides superior estimation accuracy and filter stability. It is the estimator of choice when linearization errors are a primary concern and the state dimension is moderate.

The Kalman filter is the foundational optimal recursive algorithm for state estimation in linear Gaussian systems. It operates in a two-step predict-update cycle:

• Prediction: Projects the current state and its uncertainty forward using a linear process model.
• Update: Corrects the prediction with a new sensor measurement using a linear measurement model. It is computationally efficient and provides the minimum mean square error estimate, but its core limitation is the strict requirement for linearity, which the UKF was designed to address.

The Extended Kalman Filter (EKF) is the traditional solution for nonlinear estimation. It linearizes the nonlinear process and measurement models around the current state estimate using a first-order Taylor series expansion. The standard Kalman filter equations are then applied to this local linear approximation.

Key drawbacks vs. UKF:

• Can introduce significant error for highly nonlinear systems due to crude linearization.
• Requires derivation and implementation of Jacobian matrices, which is error-prone.
• Can diverge if the local linearity assumption is violated, whereas the UKF's unscented transform more accurately captures the true mean and covariance.

A particle filter is a sequential Monte Carlo method that represents the state's probability distribution with a set of random samples called particles. Each particle has a weight representing its likelihood.

Comparison with UKF:

• Strength: Handles arbitrary nonlinearities and non-Gaussian noise far better than UKF.
• Weakness: Computationally expensive (curse of dimensionality); requires many particles for high-dimensional states.
• Use Case: The UKF is often preferred for moderate-dimensional, near-Gaussian problems for its fixed computational cost, while particle filters are the tool of choice for multimodal distributions (e.g., global localization).

The Error State Kalman Filter (ESKF) is a sophisticated variant that estimates the error in a nominal state, not the full state itself. The nominal state is propagated using a full nonlinear model, while the small error state is estimated using a Kalman filter (often an EKF) applied to linearized error dynamics.

Key advantages:

• Excellent numerical stability, especially for attitude estimation on manifolds like SO(3).
• The error state typically operates near zero, where linearization is more valid.
• Often used in conjunction with or as an alternative to the UKF for inertial navigation and visual-inertial odometry (VIO).

Sensor fusion is the overarching engineering discipline of combining data from multiple, often heterogeneous, sensors (e.g., IMU, LiDAR, camera, GPS) to produce a state estimate that is more accurate, complete, and reliable than any single sensor could provide. The UKF is a primary algorithmic tool for achieving fusion.

Architectural Patterns:

• Tightly-Coupled Fusion: Raw or low-level sensor data (e.g., pixel features, raw IMU readings) are fused directly in the UKF's measurement update. Higher accuracy, more complex.
• Loosely-Coupled Fusion: Sensors first produce independent state estimates (e.g., pose from visual odometry), which are then fused by the UKF. More modular, potentially less accurate.

The covariance matrix is a fundamental mathematical object in probabilistic estimation. It is a square, symmetric matrix that represents:

• Uncertainty (Variance): The diagonal elements represent the estimated error variance for each state variable (e.g., position uncertainty).
• Correlation (Covariance): The off-diagonal elements indicate how errors in one state variable are related to errors in another (e.g., correlation between x-position and y-velocity).

The UKF's unscented transform is specifically designed to propagate this covariance matrix through nonlinear functions more accurately than the EKF's linearization. The matrix is central to the filter's gain calculation and outlier rejection via metrics like the Mahalanobis distance.