State estimation is the art of inferring the hidden internal variables of a system—such as the precise position of a robot or the temperature inside a chemical reactor—from noisy, incomplete sensor data. Traditionally, this has been the domain of recursive filters like the Kalman Filter, which update the current state estimate using only the most recent measurement and the previous estimate.
However, recursive filters often struggle when systems are highly nonlinear or governed by strict physical constraints, such as the fact that a chemical concentration cannot be negative. Moving Horizon Estimation (MHE) addresses these challenges by reframing estimation as a constrained optimization problem solved over a sliding window of recent history.
The Intuition of the Sliding Window
The core idea behind MHE is to look at a fixed-size "buffer" of the most recent observations rather than just the current moment. As Rao et al. describe, MHE reformulates state estimation as a quadratic program solved over this sliding window. By considering a trajectory of data points rather than a single snapshot, the estimator can reconcile multiple measurements over time, which Vukov and Diehl note improves robustness against noise and environmental fluctuations.
This windowed approach provides a middle ground between two extremes:
- Recursive Filtering: Computationally cheap but "forgetful," making it sensitive to outliers and poor at handling constraints.
- Full-Batch Optimization: Highly accurate as it uses all historical data, but computationally infeasible for real-time systems because the problem size grows indefinitely.
By keeping the window size fixed, MHE bounds the computational complexity, making it practical for real-time deployment in demanding applications like wind turbine control (Vukov and Diehl) and autonomous racing (ETH Zurich researchers).
Handling Physical Constraints
One of the primary advantages of MHE is its ability to explicitly incorporate inequality constraints. In many engineering fields, we have "prior knowledge" that certain states are physically impossible. For instance, in retinal laser surgery, tissue temperatures must stay within realistic bounds to ensure safety. Kleyman et al. demonstrate that MHE offers superior stability in these scenarios by enforcing these physical constraints directly within the optimization.
Similarly, Jang et al. argue that MHE is often better suited for complex chemical processes than stochastic filters because it can handle non-negativity constraints that recursive filters might otherwise violate. By treating estimation as a mathematical program, MHE ensures the resulting state trajectory is not just statistically likely, but physically possible.
Robustness Through Temporal Redundancy
MHE leverages the physics of the system to "fill in the gaps" when sensors fail. Because the estimator optimizes over a window of states, it maintains a consistent trajectory even during temporary signal dropouts. The authors of the Chronos platform show that this mechanism allows a miniature car to navigate accurately through extended sensor failures by relying on its internal physics model and the history of states stored in the sliding window.
This temporal redundancy also makes MHE highly effective for smoothing erratic data. In the context of heavy vehicle telemetry, Khoche found that MHE provides superior velocity smoothing compared to standard filters, effectively eliminating the speed spikes and outliers often found in manual human annotations.
Balancing the Past and Present
To make a windowed approach mathematically sound, the estimator must account for the data it has discarded (the "pre-window" history). This is typically handled via an arrival cost, which summarizes the information from all measurements prior to the current window. Wolff et al. explain that the MHE objective function must balance the reliability of this summarized initial state against the fitting error of the new observations within the window.
While calculating the perfect arrival cost is difficult for nonlinear systems, Mishra et al. propose terminal cost approximations that simplify this process, eliminating the need for parallel filters and ensuring the estimator remains stable and computationally efficient.
Diverse Applications in Complex Dynamics
The flexibility of MHE has led to its adoption across various fields where dynamics are fast and constraints are critical:
- Power Systems: Lygeros et al. use MHE to track the internal states of synchronous generators, achieving accuracy that static snapshot methods cannot match.
- Robotics and SLAM: For skid-steering robots, Zuo et al. employ a sliding-window estimator to fuse sensor measurements and model complex kinematics online. Similarly, Mateus 0001 et al. use a moving horizon observer to improve the resilience of 3D reconstructions against sensor noise.
- Aerial Tracking: Tong et al. combine MHE with trajectory planning to pinpoint moving targets from micro-UAVs, mitigating the cumulative errors inherent in real-time monocular vision.
By shifting from recursive updates to windowed optimization, MHE provides a robust framework for state estimation that respects the physical realities of the systems it tracks.
Key Foundational Work
Constrained linear state estimation - a moving horizon approach (2001) — contributed 2 concept elements to this direction.
A real-time MHE and NMPC scheme for wind turbine control (2013) — contributed 1 concept elements to this direction.