Robots that localize by matching laser scans fail in long tunnels and corridors. Every scan looks like the last one along the tunnel axis, so the scan matcher cannot tell how far the robot moved in that direction. Current systems detect this by looking at the curvature of the matching error: if the error surface is flat along some direction, they stop trusting the scan there and fall back on inertial or wheel odometry. The decision is almost always a fixed threshold on an eigenvalue. That threshold is tuned on data that arrives at a steady 10–20 Hz. In real deployments, the scans that matter often arrive irregularly. Returns vanish in large voids and dust, frames get skipped under compute load, and keyframe logic drops uninformative scans. These irregularities cluster in exactly the places where geometry is degenerate.
This note proposes a simple change. Whether to trust a weakly constrained direction should depend on how long that direction has gone without a reliable constraint. Over that time, inertial uncertainty grows roughly with the cube of elapsed time. A scan that is too weak to trust 100 ms after the last good fix may be the best information available 3 s later. We propose a gap-aware localizability test inside a continuous-time trajectory estimator. It replaces the fixed eigenvalue threshold with a per-direction comparison between scan information and inertially propagated uncertainty, and it adds a time-scaled consistency check that catches "slip," where the matcher confidently reports a wrong motion. The central empirical prediction is falsifiable. A fixed threshold tuned at the native frame rate becomes miscalibrated once sampling turns irregular. The gap-aware rule, set once, should not need retuning.
Core Claim
Degeneracy detection, meaning the detection of directions in which the sensor data do not pin down the robot's pose, is currently treated as a property of a single scan. We claim it is a property of a scan relative to an interval: the time since each direction was last well constrained. Irregular temporal sampling is therefore not a nuisance layered on top of degeneracy. It sets the threshold.
To make this concrete, take the scan-matching Hessian $H = J^T J$, where $J$ is the Jacobian of the matching residuals with respect to pose. Each eigenvector $v$ of $H$ has an eigenvalue $\lambda_v$, which measures the information the scan carries along $v$. Let $t_v$ be the last time direction $v$ received an accepted constraint, and $\Delta t = t - t_v$. The variance of the inertial prediction along $v$ is
$$\sigma_v^2(\Delta t) \approx v^T P(t_v)\, v + \tfrac{1}{3} q_a \Delta t^3,$$
where $P(t_v)$ is the pose covariance at the last fix and $q_a$ is the accelerometer noise density. The $\Delta t^3$ term comes from double-integrating white acceleration noise into position.
We then use two quantities. The first is a relative-information score $r_v = \lambda_v \sigma_v^2(\Delta t)$, which compares how much the scan says against how much the robot has forgotten. The second is a slip statistic, $\delta_v^2 / (\sigma_v^2(\Delta t) + \lambda_v^{-1})$. Here $\delta_v$ is the component, along $v$, of the difference between the scan-matched and inertially predicted pose. A direction's scan residuals enter the estimator with a soft weight that rises with $r_v$ and is cut when the slip statistic exceeds a chi-square bound.
Both constants are set once, from sensor noise and a confidence level, rather than per terrain. This also gives a continuous treatment of the transitions between "localizable" and "non-localizable" that the field notes are missing.
Relation to Prior Work
Several papers supply the parts we reuse. X-ICP projects registration into the Hessian eigenspace and sorts directions into localizable, partially localizable and non-localizable. It zeroes updates in the last group and defers to odometry. We keep its eigenspace decomposition and coordinate invariance, but replace its tuned per-step classification with the gap-dependent score above. LIO-MARS supplies the trajectory model: non-uniform B-splines, smooth curves through pose space whose control knots sit at the actual, jittery scan times. It also propagates motion uncertainty with the unscented transform, which gives us $\sigma_v^2$ along the spline. LIO-MARS fixes ill-conditioning caused by irregular timing, but it does not detect geometric degeneracy at all. A tunnel leaves its control points unconstrained along the axis no matter where the knots sit. For numerically stable propagation of covariance across long gaps, we borrow the square-root (Cholesky) filtering recursions of NCDSSM.
Several papers motivate the gap. LODESTAR monitors the Jacobian condition number in a discrete Kalman filter and anchors to past well-conditioned poses. It assumes fixed frame steps, and its authors report that anchors run out during prolonged voids. LION computes a condition number per scan inside a fixed-lag smoother and hands off to a supervisor when it degrades. The event-camera tracker of Comparing Representations in Tracking for Event Camera-Based SLAM switches representations when the minimum eigenvalue falls below a fixed $\lambda_{th}$. In all three, the decision ignores elapsed time, and that omission is what we target.
From the time-series side, Neural Jump ODEs prove optimal filtering under irregular sampling, but only when observation times are independent of the state. Degeneracy-driven dropouts break that assumption, so the theorem does not license ignoring timing here. Continuous-Time Visual-Inertial Odometry for Event Cameras shows that spline trajectories evaluated at native timestamps beat binned discrete processing, but it does no degeneracy monitoring. We deliberately do not use a learned neural ODE. The inertial dynamics are known, and the open problem is the trust decision, not the dynamics model.
Research Program
We will build the estimator on a LIO-MARS-style non-uniform spline with IMU factors and point-to-plane scan residuals. At each scan we eigendecompose the registration Hessian. For each eigen-direction we look up the propagated variance since that direction's last accepted constraint, compute the score and slip test, and reweight the scan residuals projected onto that direction.
Because eigenvectors rotate between scans, the "last constrained time" is not stored per direction label. Instead we keep the full information matrix accumulated from accepted constraints and propagate it through the spline's motion. The effective gap along any query direction $v$ is the time since that direction's accumulated information last exceeded a fixed fraction of its current value. This keeps the method coordinate-invariant.
The second contribution is an evaluation protocol. Degeneracy-aware odometry is currently benchmarked only at native, regular frame rates. We will add irregular sampling regimes and run every baseline under the same regimes. The first is uniform random scan dropping. The second is state-dependent dropping, where the drop probability rises with the native minimum eigenvalue's deficit, imitating zero-return and compute-skip behaviour. The third uses the real dropouts recorded in SubT-MRS sequences with obscurants.
Baselines
The naive reference is point-to-plane ICP with IMU prediction and no degeneracy handling. On the SubT-MRS long corridor, GenZ-ICP reports this at more than 32 m APE (absolute pose error, the distance between estimated and true trajectory) and KISS-ICP at 6.83 m. It shows how bad ignoring degeneracy is, and we rerun it under every sampling regime.
The strongest existing method on our main benchmark is LODESTAR. It reports 1.386 m APE on the SubT-MRS long corridor (against 3.113 m for its plain error-state Kalman baseline) and 0.671 m on unstructured caverns, at native frame rate. It is the strongest competitor because it already combines degeneracy detection with temporal anchoring to past poses, which is the closest thing to "remembering" constraints over time.
GenZ-ICP is a second strong competitor. It reports 1.69 m APE on the same corridor and 0.51% relative error on KITTI. It sidesteps thresholds entirely by blending point-to-point residuals into point-to-plane residuals. X-ICP is the reference for eigenspace thresholding. It reports 0.27 m final position error over the 521 m Seemühle mine tunnel, but that data may not be publicly available, so we will compare with it on SubT-MRS using a reimplementation inside our spline back end. That reimplementation isolates the decision rule from the estimator.
For fairness, every method uses the same scans, the same IMU and the same dropped-scan masks. Each threshold-based baseline gets the same tuning budget: a grid search over its threshold on one held-out SubT-MRS sequence at native rate. That tuned value is then frozen across regimes, which mirrors deployment. In a separate "oracle" condition, baselines are retuned for every regime, which gives them an advantage we do not get.
Beating LODESTAR means matching its native-rate APE on the long corridor (within its reported 1.386 m) while degrading less than all baselines as dropping becomes state-dependent. Without equal native-rate accuracy, robustness gains would not count.
Three ablations isolate the key ingredient. The first uses the same spline estimator with a fixed eigenvalue threshold. The second uses the score $r_v$ without the slip test. The third uses the slip test with the gap term $\Delta t$ removed.
Why not just…?
Why not just feed the full Hessian into a Kalman filter or smoother and let the gain handle weak directions? This is the most serious threat. A correct Kalman update already weights a scan by its information relative to prior uncertainty, which is close to our score. The reason degeneracy-aware methods exist is that slip produces biased corrections that the information matrix does not describe: the naive point-to-plane result above is information-weighted and still drifts beyond 32 m. Our slip test is the part a plain filter lacks. To settle this, we will include an "information-only" estimator with no detection and no slip test. If it matches our method under irregular sampling, the contribution reduces to the protocol, and we will say so.
Why not just retune the threshold, or scale it by $\Delta t$ by hand? A per-regime retuned threshold is our oracle baseline. A heuristic threshold that grows linearly with $\Delta t$ is a cheap ablation. If the hand-scaled threshold performs as well, the principled $\Delta t^3$ form adds little. The experiments report this directly rather than assume it.
Why not just drop degenerate frames or dead-reckon through them? The file's evidence says this fails over long stretches. LION reports 9 m hallway error for systems that slip or fall back naively, against under 1 m for observability-aware estimation, and 19–25% wheel-inertial drift in DARPA tunnels against its 1.5–8.3%. Frame dropping also enlarges gaps, which is exactly the regime we study. Its X-ICP comparison point (6.37 m and 24.17 m for discarding approaches against 0.27 m) shows that partially constrained directions must be kept.
Why not just a neural continuous-time model such as a jump ODE or multi-time attention? These learn dynamics, and the dynamics here are already known physics. Their theory assumes uninformative timing, which NJ-ODE itself states and which degeneracy violates. None of them separates observable from unobservable directions in a scan. We see no reason they would solve the trust decision, and we do not include them.
Proposed Experiments
The decisive check is calibration transfer. We tune the fixed-threshold ablation and the X-ICP-style rule at native rate on a SubT-MRS corridor, then measure APE as state-dependent dropping increases on held-out corridor and cavern sequences. The prediction is that the fixed threshold's error rises sharply with dropping while the gap-aware rule's error rises more slowly, without retuning. If the curves track each other, the core claim is wrong.
A second check targets transitions. On segments where a robot enters and leaves a corridor, we will plot per-direction weights and error against time. The prediction is that the gap-aware rule readmits weak lateral and axial constraints sooner after long gaps than the fixed threshold does.
Third, on SubT-MRS sequences with real obscurant dropouts, we report APE and the dataset's velocity-AUC metric.
Finally, two checks confirm the method does no harm. On KITTI and Newer College, including the high-rotation "06 Spin" sequence where LIO-MARS reports 5.73 cm point-to-plane error, we test that the method matches non-degenerate accuracy. We also report per-scan runtime against LODESTAR's roughly 28 ms per frame, to show the eigendecomposition and propagation stay real-time.
References
- X-ICP: Localizability-Aware LiDAR Registration for Robust Localization in Extreme Environments
- LIO-MARS: Non-uniform Continuous-time Trajectories for Real-time LiDAR-Inertial-Odometry
- Neural Continuous-Discrete State Space Models for Irregularly-Sampled Time Series
- LODESTAR: Degeneracy-Aware LiDAR-Inertial Odometry With Adaptive Schmidt-Kalman Filter and Data Exploitation
- LION: Lidar-Inertial Observability-Aware Navigator for Vision-Denied Environments
- Comparing Representations in Tracking for Event Camera-Based SLAM
- Neural Jump Ordinary Differential Equations: Consistent Continuous-Time Prediction and Filtering
- Continuous-Time Visual-Inertial Odometry for Event Cameras
- SubT-MRS Dataset: Pushing SLAM Towards All-weather Environments
- GenZ-ICP: Generalizable and Degeneracy-Robust LiDAR Odometry Using an Adaptive Weighting