| Summary: | Robotic satellite operations are an integral component of future space missions, such as on-orbit servicing, in-space robotic assembly, and orbital debris mitigation. A key requirement shared among such space missions is the capability to carry out robust and autonomous close proximity operations between the involved agents. Scenarios involving unknown and uncooperative target objects require online estimation capabilities to acquire target information, such as inertial properties for target motion prediction, in order to enable the subsequent mission phases.
This inspection can be formulated as a simultaneous localization and mapping (SLAM) problem, where localization of the inspector with respect to the target imposes the necessity of acquiring a map against which to perform relative navigation. Current state of the art space inspection approaches are either prohibitively expensive for online operation, or compute partial solutions by segmenting the problem into incremental algorithms and batch approach formulations. The objective of this work is to alleviate the complexity issues that arise in the information fusion steps of the estimation process, and that prevent for a full solution of the problem to be computed incrementally and online.
For such resource constrained systems, state of the art approaches offer focused inference solutions to deal with the computational bottlenecks of complex SLAM problems. While the majority of these methods center on exploiting the conditional independence structure of the problem’s model, they operate directly on graphs instead of their underlying tree decompositions. A Bayes tree, which is the tree decomposition associated with a factor graph, blatantly exposes this sought conditional independence structure in the form of cliques and, furthermore, is the actual data structure used by the inference algorithms.
This work makes use of the readily available conditional independence property to explore focused inference approaches directly on the Bayes tree for resource constrained incremental smoothing and mapping. It elucidates the impact that graph-centered resource constrained methodologies have at inference time, and presents a simplified approach that unifies many inference strategies (e.g., filtering, fixed-lag smoothing, incremental smoothing and mapping) under simple clique- and tree-based operations. A proof of concept is presented to highlight the advantages and the versatility obtained by reasoning using the tree instead of a graph when trying to explore the performance/quality tradespace, with the added benefit of doing so while avoiding the need to modify the problem’s original model. The key insights obtained from this analysis are then leveraged for the development of novel factors to incorporate the estimation of the target object inertial properties to the SLAM formulation, obtaining a real-time and incremental solution to the space inspection problem.
|
|---|