Abstract
While many visual odometry (VO), visual-inertial odometry (VIO), and SLAM systems achieve high accuracy, the majority of existing methods miss to assess risks at runtime. This paper presents SUPER (Sensitivity-based Uncertainty-aware PErformance and Risk assessment), a generic and explainable framework that propagates uncertainties via sensitivities for real-time risk assessment in VIO. The scientific novelty lies in the derivation of a real-time risk indicator that is backend-agnostic and exploits the Schur complement blocks of the Gauss-Newton normal matrix to propagate uncertainties. Practically, the Schur complement captures the sensitivity that reflects the influence of the uncertainty on the risk occurrence. Our framework estimates risks on the basis of the residual magnitudes, geometric conditioning, and short horizon temporal trends without requiring ground truth knowledge. Our framework enables to reliably predict trajectory degradation 50 frames ahead with an improvement of 20% to the baseline. In addition, SUPER initiates a stop or relocalization policy with 89.1% recall. The framework is backend agnostic and operates in real time with less than 0.2% additional CPU cost. Experiments show that SUPER provides consistent uncertainty estimates. A SLAM evaluation highlights the applicability to long horizon mapping.
Introduction
Modern Visual-Inertial Odometry (VIO) and SLAM systems are highly accurate but often lack "self-awareness"—the ability to detect when they are on the verge of failure. In safety-critical robotics, silent error accumulation due to environmental factors like motion blur or low texture can lead to catastrophic results before a system even realizes it has lost track.
This paper introduces SUPER, a real-time, backend-agnostic framework designed to predict impending tracking failures. By extracting uncertainty and geometric sensitivity directly from the optimizer's existing mathematical operations, SUPER provides reliable early warnings with negligible computational overhead (<0.2% CPU cost).
Methodology: Recycling the Optimizer's Math
The core innovation of SUPER is its ability to "recycle" the heavy mathematical lifting already performed by VIO/SLAM optimizers (such as Ceres, g2o, or GTSAM).
The framework operates in several key stages:
- Schur Complement Extraction: Instead of redundant covariance calculations, SUPER extracts Schur complement blocks from the Gauss-Newton normal matrix. This provides per-landmark uncertainty estimates while accounting for pose correlations at almost no extra cost.
- Geometric Conditioning: By analyzing the singular values of the projection Jacobian, the system measures sensitivity to viewing angles and triangulation baselines, identifying "poor geometry" that residuals alone might miss.
- Unified Risk Score: SUPER aggregates measurement noise (pixel uncertainty), misfit (reprojection residuals), and geometric conditioning into a single, normalized risk score.
- Temporal Dynamics: By monitoring the derivative of this risk score, the system distinguishes between transient sensor noise and dangerous, compounding trends.
Experimental Evaluation
SUPER was rigorously tested on standard benchmarks—EuRoC MAV, TUM-VI, and KITTI Odometry—under various synthetic corruptions like blur, pixel noise, and occlusions.
Key Results and Insights
- Extreme Efficiency: Operating at less than 0.2% additional CPU cost (~0.14 ms/frame), SUPER is significantly faster than filter-based or deep-learning alternatives.
- Predictive Power: It can predict trajectory degradation up to 50 frames (2.5 seconds) in advance, outperforming standard residual checks by 20%.
- Actionable Safety: When used as a safety trigger, SUPER achieved an 89.1% recall for catching SLAM failures with only an 8.4% false positive rate.
- Safety Standards: The risk score is designed to map directly to industrial safety standards like ISO 12100 and ISO 13849, facilitating its use in regulated environments.
Conclusion
SUPER addresses a critical gap in VIO/SLAM by providing an explainable, real-time risk indicator. While it may struggle in extreme cases where feature support vanishes entirely, its ability to provide early warnings makes it a powerful tool for autonomous systems. Future work will focus on handling spatially varying noise and deeper integration into long-horizon SLAM decision-making pipelines.
Summary updated