SNGR: Selective Non-Gaussian Refinement for Ambiguous SLAM Factor Graphs

Researchers Anushka Kulkarni and Sarthak Dubey have introduced Selective Non-Gaussian Refinement (SNGR), a novel SLAM framework that addresses a critical weakness in modern simultaneous localization and mapping systems: the reliance on Gaussian approximations, which often fail in ambiguous environments like those with wrong data associations or multimodal posteriors. SNGR builds on the popular iSAM2 incremental smoothing and mapping algorithm, adding a detection mechanism that uses the condition number of joint marginal covariances to identify windows where Gaussian approximations are likely inaccurate. Once identified, SNGR selectively applies nested sampling to refine these regions using the full nonlinear factor graph likelihood, while a gating mechanism prevents performance degradation in inherently multimodal cases.

In experiments on range-only SLAM with deliberately incorrect data associations, SNGR achieved high-precision failure detection and consistent local likelihood improvements compared to standard iSAM2. Crucially, it did so with significantly lower computational cost than exhaustive non-Gaussian inference methods, making it practical for real-time robotics applications. The results highlight both the promise of selective refinement for improving approximate SLAM posteriors and the limitations that remain in handling complex multimodal scenarios. This work represents a practical step toward more robust SLAM systems that can operate reliably in ambiguous real-world environments.