Factor Graph-Based Active SLAM for Spacecraft Proximity Operations

TL;DR

This work tackles autonomous relative navigation for a chaser near a target using a monocular camera by casting the problem as smoothing-based SLAM on a factor graph. It introduces an information-theoretic belief-space planner that directly derives a reward from the SLAM graph, evaluating candidate camera-pointing actions via graph augmentation and the posterior information gain $\mathcal{Q}(\beta(X_{k+L})) = -\frac{n'}{2}\log(2\pi e) + \frac{1}{2}\log \frac{|\Lambda_{k+L}|}{|\Lambda_k|}$, where the belief $\beta(X)$ is Gaussian with information matrix $\Lambda=\Sigma^{-1}$. The contributions include the first application of information-theoretic BSP to spacecraft proximity SLAM and a demonstration that active attitude planning improves uncertainty and map coverage compared with passive strategies in simulated HST proximity scenarios. The approach offers a principled, integrated mechanism to couple sensing, estimation, and control, enabling more reliable autonomous ISAM operations with limited onboard sensing.

Abstract

We investigate a scenario where a chaser spacecraft or satellite equipped with a monocular camera navigates in close proximity to a target spacecraft. The satellite's primary objective is to construct a representation of the operational environment and localize itself within it, utilizing the available image data. We frame the joint task of state trajectory and map estimation as an instance of smoothing-based simultaneous localization and mapping (SLAM), where the underlying structure of the problem is represented as a factor graph. Rather than considering estimation and planning as separate tasks, we propose to control the camera observations to actively reduce the uncertainty of the estimation variables, the spacecraft state, and the map landmarks. This is accomplished by adopting an information-theoretic metric to reason about the impact of candidate actions on the evolution of the belief state. Numerical simulations indicate that the proposed method successfully captures the interplay between planning and estimation, hence yielding reduced uncertainty and higher accuracy when compared to commonly adopted passive sensing strategies.

Figures (7)

• Figure 1: Schematic of the graph augmentation operations. $\mathcal{F}_k$ is the factor graph modeling the current belief $\beta(X_k)$, from which the posterior $\beta^m(X_{k+L})$ can be obtained by "augmenting" $\mathcal{F}_k$ with $\mathcal{F}^m_{k+1:k+L}$. Squares and circles represent poses and map variables, respectively, edges encode probabilistic constraints (factors) arising from observations. Future poses and factors in color.
• Figure 2: Rendered images of the HST as seen from the chaser onboard camera during the reconnaissance phase (Figs. \ref{['fig:hst_rgb_1']}-\ref{['fig:hst_rgb_4']}) with the ORB keypoints used in the map initialization step (green dots in Figs. \ref{['fig:hst_gray_1']}-\ref{['fig:hst_gray_4']}).
• Figure 3: SLAM results along trajectory $\tau^1$, $\tau^2$, $\tau^*$ for $L=L_1$. rgb frames represent the SLAM estimate while the Matlabstd color triplet is used for the camera pose ground truth.
• Figure 4: Ground truth (colored in Matlabstd triplet), estimated camera trajectory (in rgb) and initialized map $\hat{M}_k$ (green dots) during the reconnaissance phase around the HST.
• Figure 5: Performance of pose trajectory estimation for time horizons $L_1$, $L_2$. The values of the errors are scaled up by a factor of $50$ for visualization purposes.