What Breaks Monocular SLAM in Microgravity? An Initial Benchmark on Rotation-Dominant Astrobee ISS Sequences

Jungil Ham, Brian Coltin, Ryan Soussan, Pyojin Kim

Keywords: Microgravity Visual Navigation, Monocular SLAM, 3D Gaussian Splatting (3DGS)

TL;DR: Rotation-dominant microgravity motion exposes highly local failures in monocular SLAM on Astrobee ISS data, and dense 3DGS pipelines with stronger stabilization outperform learned-prior approaches.

Abstract:

We present a challenge-oriented benchmark for Astrobee free-flyer monocular SLAM inside the International Space Station (ISS), guided by a visual challenge taxonomy.
This initial study focuses on rotation-dominant subset of this taxonomy.
We evaluate monocular methods spanning the dense 3D Gaussian Splatting (3DGS) Simultaneous Localization and Mapping (SLAM) family and geometric foundation model based approaches on rotation-dominant monocular sequences, emphasizing dense monocular 3DGS pipelines for their explicit scene representation, which allows tracking, mapping, and rendering failures to be directly examined, and contrasting feed-forward geometric foundation models with tightly coupled SLAM systems built on top of them.
Our evaluations reveal that robustness is highly local and motion-conditioned; optimized 3DGS pipelines improve stability, whereas geometric foundation models require tight coupling with temporal tracking to be effective.
By exposing these algorithmic vulnerabilities, this work enables a more targeted approach to space robotics autonomy.

Submission Number: 37