Reliability-Aware Intervention Policies for Interactive Gaussian SLAM

Abstract

Interactive Gaussian scene representations enable dense RGB-D capture, online reconstruction, and free-view navigation, but reliability may degrade under sensor corruption, unstable tracking, and long sequential map updates. This paper presents SplatGuard, a lightweight runtime-assurance layer for Gaussian-splatting SLAM that keeps the backbone unchanged while monitoring temporal tracking and mapping diagnostics. The system uses these diagnostics to select normal continuation, additional refinement, or temporary freezing of map updates. Evaluation includes 41 pilot runs, a 216-run sweep over 24 policies and 9 conditions, surrogate scoring of 12,000 candidates, and two 54-run validation stages, with 48 completed summaries in the current snapshot. Safeguard behavior is condition-dependent: refinement helps when tracking remains recoverable, while short freezing can prevent unreliable depth observations from entering the map; frequent intervention may instead increase runtime cost or trajectory error. The results frame reliability-aware Gaussian scene capture as an accuracy–runtime–intervention trade-off rather than a single-threshold policy problem.