A Bayesian Benchmarking of GBEES Applied to Outer Planet Orbiter Estimation

Benjamin L. Hanson, Todd A. Ely, Thomas R. Bewley and Aaron J. Rosengren · 2025

Read A Bayesian Benchmarking of GBEES Applied to Outer Planet Orbiter Estimation by Benjamin L. Hanson, Todd A. Ely, Thomas R. Bewley and Aaron J. Rosengren (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-08453.txt.

Introduction

Benjamin Hanson and co-authors grade a grid-based tracker for spacecraft in this 2025 paper. Future moon tours face long signal delays and wobbly orbits, which strain classic trackers. Their tracker carries the full odds cloud of a craft's state, not just a center guess.

The test case tracks a Saturn moon orbiter with sparse, curved measures. Scores cover raw error, odds-cloud match, and a weight health check borrowed from rival methods. The grid tracker beat the rival ensemble filters on those scores. It did so at a steep compute cost, the authors note. They frame it as a strong pick where accuracy trumps speed.

Catalog record

CREATORBenjamin L. Hanson, Todd A. Ely, Thomas R. Bewley and Aaron J. Rosengren
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE25 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-08453

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

A Bayesian Benchmarking of GBEES Applied to Outer Planet Orbiter Estimation Benjamin L. Hanson, Todd A. Ely, Thomas R. Bewley and Aaron J. Rosengren Abstract 0 0 footnotetext: Presented as Paper 25-194 at the AAS/AIAA Space Flight Mechanics Meeting, Kauai, Hi, Jan. 19-23, 2025 [ 1 ] . 1 Introduction N avigation has long adhered to an established assumption that is often underappreciated: spacecraft (S/C) state uncertainty remains Gaussian on the time scale of the measurement updates [ 2 ] . This assumption, while nontrivial, is operationally validated via frequent telecommunication with the S/C, dividing the nonlinear trajectory into rectilinear segments. Adequate measurement cadences are assured for near-Earth S/C, but deep-space missions pose advanced challenges. To thoroughly explore outer planet systems, mission designers may leverage unstable, periodic orbits, the dynamics of which are often highly chaotic [ 3 , 4 ] . Such chaoticity, dominated by third- and fourth-body perturbations, make linearizations accurate on much shorter time scales, thus requiring more frequent measurement updates and station-keeping maneuvers. However, in deep-space regimes, ground-based communication has significant round-trip light-time delays which may render this form of communication insufficient. Historically, deep-space S/C have operated in stable, two-body orbits. However, proposed missions to outer planet systems, such as the Enceladus mission, suggest that the risk of operating an S/C in a dynamically unstable environment is outweighed by the scientific potential. The Enceladus S/C is expected to operate in an unstable, near-rectilinear halo orbit and perform a flyby within 10 km of the surface to perform in situ analysis of the plumes of gas ejecting from the moon’s surface [ 5 ] . As such, it is likely that the instability of these conditions will cause the state uncertainty to become non-Gaussian faster than can be corrected by ground communication. In this analysis, we aim to compare a fundamentally different estimation filter, Grid-based Bayesian Estimation Exploiting Sparsity (GBEES) [ 6 ] , to a range of conventional and state-of-the-art filters. The application used for comparison is the propagation of the state uncertainty of a low-fidelity outer planet orbiter. This investigation is arranged into six sections. Section 2 explains the filtering process and provides rationale for the filters compared in this study. Section 3 presents a novel Bayesian approach to filter benchmarking, an alternative to the frequentist approach. Section 4 defines the nonlinear dynamics and measurement models used in the Saturn-Enceladus DPO example. The numerical results are presented in Section 5 . Finally, Section 6 discusses in depth the results of the investigation and comments on the efficacy of GBEES as an onboard navigation filter. 2 Overview of Sequential Single-update Filters The Sequential Single-update Filter (SSF) is the most widely used class of algorithms for S/C navigation and estimation. Sequential refers to the way measurements are incorporated: one at a time as they become available, in contrast to batch filters, which process all data simultaneously. Single-update denotes that each measurement is assimilated through a single application of Bayes’ theorem 1 1 1 The particle flow filter [ 7 ] , a modern version of the Lagrangian approach, incrementally performs the Bayesian update by flowing the M particles subject to a continuous parameter \lambda . Because this update step employs Bayes’ rule multiple times per measurement, a methodological advantage that is unique to the particle flow filter, the authors believe its inclusion in this work would remove the one-to-one nature of the comparison. . In this investigation, we label SSFs as either moment filters or ensemble filters, indicating the abstraction updated in the correction step (for clarification, moment filters in this context do not include those that propagate mixtures of moments or higher order moments). The evolution of a stochastic process \bm{X}(t)\in\mathbb{R}^{d} governed by a combination of deterministic and random effects can be described by the following stochastic differential equation: d\bm{X}(t)=\bm{f}(\bm{X}(t),t)dt+\bm{q}(\bm{X}(t),t)d\bm{W}(t), (1) where \bm{f}(\bm{X}(t),t) is the drift vector, \bm{q}(\bm{X}(t),t) is the diffusion vector, and d\bm{W}(t)=\bm{\xi}(t)dt is a Wiener process, meaning \bm{\xi}(t) is zero-mean, uncorrelated white noise. In continuous-time, the Fokker-Planck equation gives the evolution of the probability density function (PDF) p(\bm{x},t) of \bm{X}(t) in Equation ( 1 ) as follows: \displaystyle\frac{\partial p(\bm{x},t)}{\partial t}=-\sum\limits_{i=1}^{d}\frac{\partial f_{i}(\bm{x},t)p(\bm{x},t)}{\partial x_{i}}+\frac{1}{2}\sum\limits_{i=1}^{d}\sum\limits_{j=1}^{d}\frac{\partial^{2}Q_{ij}(\bm{x},t)\,p(\bm{x},t)}{\partial x_{i}\partial x_{j}} (2) where \bm{x}=(x_{1},\dots,x_{d}) is a realization of \bm{X}(t) , f_{i}(\bm{x},t) is the i^{\text{th}} component of the drift vector at realization \bm{x} and time t , and Q_{ij}(\bm{x},t) is the (i,j)^{\text{th}} element of the stochastic state disturbance matrix Q=\bm{q}(\bm{x},t)\bm{q}(\bm{x},t)^{T} . In general, Equation ( 2 ) is non-integrable, and p(\bm{x},t) cannot be described by a finite number of parameters. At measurement epochs t=t^{(k)} , the PDF is updated via Bayes’ theorem: \displaystyle p(\bm{x},t^{(k+)})=\frac{p(\bm{y}^{(k)}|\bm{x})\,p(\bm{x},t^{(k-)})}{C}, (3) where p(\bm{x},t^{(k+)}) is the posterior, p(\bm{y}^{(k)}|\bm{x}) is the measurement likelihood, p(\bm{x},t^{(k-)}) is the prior, and C is a normalization constant. Again, in general, the number of parameters necessary to represent p(\bm{x},t^{(k+)}) is not finite. Thus, the goal of the SSF is to approximate and propagate the full PDF with as few parameters as possible while incorporating information from measurements updates. We now list the filters GBEES is compared with in this analysis, as well as the rational for their inclusion. A large smooth bootstrap is used as a benchmark for both accuracy and efficiency. The Unscented Kalman Filter (UKF) [ 8 ] is compared as a moment filter control, as its nonlinear prediction step makes it more accurate than other moment filters. The ensemble filters compared are the Bootstrap Particle Filter (BPF) [ 9 ] and the Ensemble Gaussian Mixture Filter (EnGMF) [ 10 ] . The BPF is widely used, while the EnGMF and GBEES are more state-of-the-art. Recent efforts have focused on reducing the computational cost of determining the optimal bandwidth parameter for the EnGMF samples [ 11 , 12 ] . Similarly, GBEES has recently been significantly computationally optimized [ 13 ] ; one objective of this investigation is determining if GBEES may be feasibly used as an onboard S/C navigation algorithm. 3 Filter Benchmarking Criteria In this investigation, the performances of the selected SSFs are evaluated based on three criteria: accuracy is assessed on probability distribution similarity, degeneracy is assessed on an approximation of the effective sample size, and efficiency is assessed on computation time. Each SSF is compared to a truth implementation, discussed in Section 3.3 . 3.1 Accuracy metric The frequentist approach to filter evaluation assumes that many Monte Carlo (MC) runs of the filter may be performed to calculate an average accuracy and consistency. Practically, this may not be feasible due to time constraints. Instead, we perform a deterministic, Bayesian comparison of the filter-estimated distributions with the representative truth model. The Bhattacharyya coefficient ( BC ) [ 14 ] , a metric of the similarity between two probability distributions P and Q , is defined as \displaystyle\text{BC}(P,Q\,|\,\chi)=\sum_{\bm{x}\in\bm{\chi}}\sqrt{P(\bm{x})Q(\bm{x})}\quad\text{where}\quad\chi\subseteq\mathbb{R}^{d},\quad|\chi|<\infty,\quad\text{and}\quad 0\leq\text{BC}\leq 1, (4) where \chi is the domain and |\chi| is the cardinality. \text{BC}(P,Q\,|\,\chi)=0 indicates complete dissimilarity while \text{BC}(P,Q\,|\,\chi)=1 indicates identicality. For this analysis, P is the truth distribution and Q is the filter-estimated distribution, with \chi=\{\bm{x}_{i}\}^{M}_{i=1} , where \bm{x}_{i} are the members in the truth distribution and M is the sample size. Nearest-neighbor interpolation is utilized for calculating Q(\bm{x}) when \bm{x}\notin\text{support}(Q) . If the compared filter is a moment filter, then a normalized distribution Q is created by evaluating all points in

Read the complete text