Uncertainty-Aware LiDAR Localization Gets a Calibrated 'I Don't Know'

Robots and self-driving systems need to know where they are in 3D space, and most answer with a single confident number even when sensors or maps are degraded. A new preprint from Warsaw University of Technology proposes a different shape of answer: a per-point confidence envelope attached to every voxel, plus a principled way to check whether the network's confidence matches its actual error.

UQ-Loc, posted as arXiv 2608.06307 by Jacek Komorowski on 6 August 2026, extends the LightLoc scene coordinate regression architecture with a covariance head that predicts a full positive-definite matrix per voxel. Training swaps a regression loss for a Negative Log-Likelihood (NLL) loss augmented with a kNN-based spatial smoothness regulariser, so neighbouring voxels share uncertainty structure rather than guessing independently.

At inference, the paper uses a modified SC2-PCR solver with uncertainty-weighted seed scoring and a Mahalanobis-distance inlier test that rejects seed candidates whose geometry disagrees with the predicted covariance. Expected Calibration Error (ECE) is the principled yardstick for whether the network is "honest": predicted uncertainty must track empirical error at the rates claimed.

The target application is 6-DoF LiDAR localization without explicit map retrieval, and the paper frames itself as the first per-voxel anisotropic aleatoric uncertainty attachment to LiDAR scene coordinate regression.

Caveat: single-author preprint, no peer review, no independent benchmark confirmation. The TLDR wrapper is a discovery aggregator, not authority; anchor on the arXiv primary. No deployer, regulator, or incident attaches yet.