New Convergence Metric Tells LiDAR SLAM When It Is Truly Lost

Robots that navigate tunnels, wind turbine towers, and other repetitive environments face a deceptively simple problem: when everything around them looks the same, their laser-based maps stop telling them where they are. A new study published in Autonomous Intelligent Systems introduces a detection method called the Optimization Convergence Efficiency (OCE) metric, which decides whether a LiDAR-based simultaneous localization and mapping (SLAM) system is genuinely failing or merely operating in a space that looks, to conventional diagnostics, as though it should fail. The work, led by Junrui Zhang and colleagues with corresponding author Runjie Shen, addresses one of the most persistent failure modes of laser navigation in geometrically self-similar scenes.

LiDAR SLAM systems build maps and estimate position by repeatedly aligning incoming laser scans with a reference map, a process known as point cloud registration. In feature-rich surroundings, the geometry of walls, corners, and objects constrains the alignment problem from every direction, and the optimizer that solves it converges confidently. In long straight tunnels, open fields, or the smooth interior of a cylindrical tower, however, the cost function that the optimizer minimizes develops a flat valley along one or more directions in the state space, typically the axis of the tunnel or cylinder. Along such a direction the measurements simply do not determine a unique answer, and the estimated pose can drift without bound, degrading navigation or crashing the entire system.

The dominant way to detect this condition has been to examine the eigenvalues of the Hessian matrix, the second-order description of the cost function’s curvature at the current linearization point. Small eigenvalues flag directions in which the measurements provide insufficient information, an approach established in the seminal framework of Zhang and colleagues and refined by many subsequent systems, including methods that separately treat translational and rotational degeneracy and others that apply statistical tests or adaptive clustering to the eigenvalue spectrum. The logic is sound in principle, but it carries a practical weakness: the Hessian is evaluated statically, at a single linearization point, and real measurements are noisy.

That weakness matters most in what the authors call semi-degenerate environments. Inside a wind turbine tower, the dominant cylindrical geometry suggests degeneracy, yet internal ladders, cable trays, welded seams, and mounting brackets supply a scattering of genuine constraints along the weak direction. Those sparse features can be enough for the optimizer to keep producing a valid pose estimate, but their contribution to the Hessian spectrum is small compared with the overwhelming repetitive geometry. Meanwhile, in a truly featureless tunnel, sensor noise, discretization, and surface irregularities inflate the smallest eigenvalues above zero, blurring the spectral distinction between the two cases. The result is false alarms: detectors that declare degeneracy and discard usable LiDAR estimates when the pose is actually still good.

The paper illustrates the problem with a striking concrete example. Comparing representative scans from a fully degenerate tunnel and a semi-degenerate tower, the ratio of the smallest to largest eigenvalue of the translational Hessian block was actually slightly higher in the tunnel, 0.048 versus 0.036, the opposite of what a threshold-based detector needs. A single cutoff on this ratio cannot reliably separate the two scenes, because noise-driven spectral components in the feature-poor tunnel mask the difference between absent constraints and small but real ones.

OCE takes a different tack: instead of asking whether the geometry looks degenerate, it asks whether the optimizer actually converges. During each LiDAR frame, the system’s Levenberg-Marquardt solver performs a series of evaluations, and at each one OCE extracts the component of the gradient along the monitored state direction, in this work the axial translation direction of the cylinder or tunnel. The metric is the relative contraction of that directional gradient between consecutive evaluations, expressed as a percentage: a large positive value means the first-order response shrank substantially after the previous update, a value near zero means weak contraction, and a negative value means the gradient grew after relinearization. The per-evaluation scores are averaged across the frame and compared with a detection threshold.

The reasoning behind this design rests on how noise and genuine structure behave differently over successive iterations. Noise-induced gradient components can produce nonzero state updates, but they are not tied to persistent geometric structures, so after successive updates and relinearizations their responses fluctuate, contract weakly, or grow again. Genuine features, by contrast, impose state-dependent residual constraints that progressively reduce the residuals as the optimizer works, producing a consistent contraction of the directional gradient. By watching the temporal evolution of that gradient rather than its instantaneous magnitude, OCE distinguishes coherent convergence from noise-driven churn. Notably, the method adds no modification to the optimizer itself; it reuses gradient quantities the solver already computes, and in the experiments its overhead averaged 25.31 microseconds per frame against a 100-millisecond processing interval, roughly 0.025 percent of the frame budget.

The quantitative evaluation used a controlled Gazebo simulation in which a simulated UAV flew through a cylindrical environment with three regions: feature-rich, sparsely featured, and nearly featureless along the axis. Crucially, the ground truth for detection was not the geometric region itself but whether the SLAM output remained usable, judged by a short-window relative-displacement error against the simulator’s ground-truth positions, with a failure threshold calibrated from a feature-rich interval. At the nominal operating threshold, OCE detected 55 of 56 failure-labeled frames, a recall of 98.21 percent, while raising only eight false alarms among 344 usable frames, a false-positive rate of 2.33 percent. The conventional eigenvalue-ratio detector caught all 56 failures but also fired 170 false alarms, a false-positive rate of 49.42 percent. OCE thus cut false alarms by roughly 95.3 percent while missing only one additional failure frame, and its first matched alarm came 0.135 seconds after the reference failure onset.

The comparison extended to other baseline families, and the pattern held. A normalized statistical spectral criterion in the style of Switch-SLAM remained in the alarm state throughout the sequence with published thresholds and still produced a 48.26 percent false-positive rate after recalibration; its continuous score reached an AUROC of 0.810 against 0.985 for OCE. An adaptive DBSCAN-based condition-number detector, hampered by initialization requirements, covered only 60 percent of the sequence and detected no failure event. A residual chi-square consistency test derived from the registration objective fared differently depending on its formulation, with one variant alarming throughout and the other missing the failure entirely, an AUROC of 0.500, indicating that residual statistical consistency and geometric degeneracy probe different aspects of the estimation problem. Sensitivity analyses also showed that performance stayed high across a broad range of the OCE threshold, roughly 36 to 67, so the method does not depend on precise tuning.

Two real-world UAV flights, conducted with a Livox Mid-360 LiDAR on an onboard computer, put the metric to practical test. In a wind turbine tower, the UAV ascended to about 5.4 meters and returned, and the OCE score stayed above threshold for most of the flight, flagging degeneracy only in the uppermost section where ladder and wall features ran out; the eigenvalue ratio, by contrast, dropped low throughout the ascent and descent, which would have triggered persistent false warnings. In a tunnel whose physical start-to-end distance exceeded 65 meters, the estimated axial position oscillated wildly and covered only a small fraction of the true travel in the featureless initial section, a clear localization failure that OCE correctly flagged, while the eigenvalue ratio spiked above 0.25 at lateral tunnel intersections and fell when the unambiguous end wall came into view, tracking point distribution rather than actual observability. The authors note that OCE’s numerical behavior may depend on the registration backend and that its threshold should be calibrated per system, and their complete source code is slated for release on GitHub after a follow-up paper is accepted. For robots working in the long, smooth, self-similar spaces where laser navigation is hardest, the message is that the question is not whether the geometry looks degenerate, but whether the optimizer can still find its way home.

Subject of Research: A degeneracy detection metric for LiDAR SLAM based on optimization convergence behavior

Article Title: OCE: an optimization convergence metric for LiDAR SLAM degeneracy detection

Article References: Zhang, J., Zhu, M., Sun, C., Guo, F., Zhan, Q., & Shen, R. (2026). OCE: an optimization convergence metric for LiDAR SLAM degeneracy detection. Autonomous Intelligent Systems, 6(1), Article 21. https://doi.org/10.1007/s43684-026-00144-1

Image Credits: AI Generated

DOI: 10.1007/s43684-026-00144-1

Keywords: LiDAR, SLAM, degeneracy detection, optimization convergence, Levenberg-Marquardt, Hessian eigenvalues, UAV navigation, tunnel localization, wind turbine tower, GICP, point cloud registration, autonomous systems