From particles to features: Uncertainty-guided geometric reasoning for robust SLAM

Simultaneous Localization and Mapping (SLAM) systems must commit to feature correspondences and loop closure decisions under uncertainty, yet incorrect commitments cascade into catastrophic mapping failures. Traditional particle filters address state uncertainty but leave feature-level processing deterministic, creating an asymmetry that limits robustness. We present Particle-Guided Feature Fusion (PGFF), a framework that inverts the classical role of particles: rather than sampling robot poses, particles propagate through feature space to compute importance weights that inform geometric evidence accumulation. This paradigm shift enables three innovations: (1) geometric surprise priors that identify informative features through local eigenvalue analysis and information-theoretic divergence, (2) adaptive feature weighting where particles vote on correspondence reliability based on spatial and geometric consistency, and (3) multi-hypothesis loop closure management that defers commitment until sufficient evidence accumulates. We integrate PGFF into a tightly-coupled LiDAR-inertial odometry system and evaluate on the NCLT long-term autonomy benchmark and the VBR dataset, both featuring dynamic objects, geometric degeneracy, and perceptual aliasing. PGFF reduces absolute trajectory error by 23.5% compared to FAST-LIO2 and 29.8% compared to LIO-SAM, with gains evident against recent degeneracy-aware methods as well. Combining PGFF with graduated non-convexity (GNC) yields further additive improvements, confirming that pre-correspondence and post-correspondence robustness address complementary failure modes. PGFF achieves 10.1% faster IEKF convergence with only 0.7% total runtime overhead. Multi-hypothesis loop closure reduces false positives by 58% across multiple place recognition front-ends while maintaining 94% recall. These results demonstrate that uncertainty-guided feature selection and evidence accumulation, not just state estimation, is essential for reliable long-term autonomy, suggesting the boundary between filtering and perception should be fundamentally reconsidered. Our implementation is available at https://github.com/wencabrel/PGFF-SLAM