Equivariant symmetries for inertial navigation systems

Alessandro Fornasier, Yixiao Ge, Pieter van Goor, Robert E. Mahony, Stephan WeissAutomatica 2025 · DOI

TL;DR

This paper unifies various modern Inertial Navigation System (INS) filters under a single framework based on "equivariant filters" and symmetry groups. By exploring different symmetry structures, the authors demonstrate that the choice of symmetry group is the primary differentiator between these advanced filters, offering a new perspective on their design and performance.

What Problem Does This Paper Address?

Inertial Navigation Systems (INS) are crucial for estimating an object's position, orientation, and velocity using accelerometers and gyroscopes. For decades, the Extended Kalman Filter (EKF) and its variants have been the go-to solution. However, recent advancements have leveraged the underlying geometric structure of INS problems, particularly Lie groups, to develop more sophisticated filters that offer better performance and theoretical guarantees. The challenge has been to understand the relationships between these different geometric filters and to provide a unified framework for their design and analysis. This paper addresses this by showing that many of these modern filters can be systematically derived from a single principle: applying the equivariant filter design methodology to different symmetry groups that describe the INS problem.

Methods

The core of this paper is the application of the Equivariant Filter (EqF) design methodology. The authors systematically explore various symmetry groups that can act on the state space of an INS. For each symmetry group, they derive a corresponding EqF. This approach allows them to demonstrate that several existing advanced INS filters, such as the Multiplicative EKF (MEKF), Imperfect-Invariant EKF (IEKF), and others, can be interpreted as specific instances of the EqF framework applied to different symmetry choices. The paper also introduces two novel symmetry groups, the Direct Position Equivariant Filter (DP-EqF) and Semi-Direct Bias Equivariant Filter (SD-EqF), expanding the landscape of possible filter designs. A key aspect is the linearization error analysis of the error dynamics for each filter, which helps to understand their performance characteristics.

Datasets

The paper uses simulated data from a UAV flight scenario and real-world data from the INSANE dataset [5] for experimental evaluation. The INSANE dataset is described as involving a quadcopter with IMU and RTK-GNSS measurements.

Key Insights and Ideas

  • Unifying Framework: The paper provides a powerful unifying perspective, showing that the differences between many modern geometric INS filters boil down to the choice of symmetry group used in the EqF framework.
  • Symmetry as the Differentiator: The authors argue that the symmetry structure is the only fundamental difference between these advanced INS filters, simplifying their comparison and design.
  • Novel Symmetries: The introduction of DP-EqF and SD-EqF expands the set of potential symmetries for INS, offering new avenues for filter design.
  • Linearization Error and Performance: The analysis of linearization error reveals that filters with semi-direct geometric coupling between navigation and bias states (like TG-EqF) tend to perform better, especially in transient phases, by shifting linearization errors to the bias states.
  • Exact Navigation State Linearization: The TG-EqF is highlighted as the only filter that achieves exact linearization of the navigation states, with linearization errors confined to the bias states.

Key Results

  • Comparative Performance: Experimental evaluations (both simulated and real-world) show that filters derived from semi-direct bias symmetries (TG-EqF, DP-EqF, SD-EqF) generally outperform the classical MEKF and other IEKF variants, particularly in transient responses.
  • TG-EqF Superiority: The TG-EqF consistently demonstrates superior performance in terms of lower filter error, better consistency (ANEES close to 1), and faster transient response, making it a recommended choice for INS problems.
  • Reformulation of Measurements: The paper shows how global position measurements can be reformulated as body-frame residual measurements, enabling their use with equivariant filters and ensuring third-order linearization error.

Limitations and Future Work

  • Observability: The paper acknowledges that performance can vary in scenarios with unobservable states (e.g., straight-line flight, hovering), and the presented observations are primarily for "fully observable" cases.
  • Virtual States: Some proposed symmetries (like GTG) require the introduction of virtual states to maintain the desired group structure, which the authors note might be questionable in practice. Future work could explore symmetries that avoid such over-parametrization while still leveraging semi-direct structures.
  • Real-world Scenarios: While experiments were conducted, the paper suggests further investigation into semi-direct product symmetries in scenarios involving difficult alignment.

Summary updated

BibTeX
@article{fornasier2025equivariant,
  author = {Alessandro Fornasier and Yixiao Ge and Pieter van Goor and Robert E. Mahony and Stephan Weiss},
  title = {{Equivariant symmetries for inertial navigation systems}},
  journal = {Automatica},
  year = {2025},
  doi = {10.1016/j.automatica.2025.112495},
}
Figure