Every state estimator, in one library, on one benchmark.

A library

One header per estimator, each templated on a small model concept (f, h, Q, R; Jacobians optional). No heap, no exceptions, fixed-size linear algebra with compile-time dimensions, Real = double or float, warning-free under -Wall -Wextra -Wpedantic -Werror on GCC and Clang, and it cross-compiles to a Cortex-M7. The battery model is one instance; a drone navigation model and an attitude model are others — the same Ekf<M> estimates all three.

A benchmark

Two truth plants — an enhanced equivalent-circuit cell and an electrochemical single-particle model built from LG M50 electrode data — over four missions, twelve fault scenarios and three noise seeds, plus a twelve-cell pack, a 120-mission ageing study and an instrumented flight. Cross-validated against FilterPy, PyBaMM and ahrs to floating-point precision.

The finding

On a cell whose model the estimator knows, the nominal problem is solved by any tuned closed-loop filter; the field is separated only by faults. On a cell whose model it does not know, every converged Kalman filter parks on a ~4 % bias with a clean innovation, while fading-memory and fixed-gain estimators hold under 0.7 %. The choice of estimator matters much less than whether it is told the truth about its model and its sensors.

The plain EKF fails the hardest scenario (MHE-RTI and the adaptive family reach 0.50 % where the EKF has 12.4 %) not because it is a poor filter but because it is an exact one for a model that is wrong.

Engineers

Drop one header into your project and run any filter on your own model in twenty lines.

Researchers

A reproducible, cross-validated benchmark and a 958-page report with one chapter per estimator.

Practitioners (BMS)

A reference estimator stack for an electric-aircraft battery-management system, with the evidence.