GlassGuard: Verified Glass Plane Mapping for Robot Navigation

· Editorial Team estimated
navigation lidar glass-reconstruction slam mapping mobile-robots perception

Transparent and specular surfaces break LiDAR navigation: laser returns can pass straight through glass, so collision boundaries are simply absent from the map. Prior work tries to reconstruct the missing surfaces, but inaccurate placement creates the opposite failure — it contaminates traversable free space. GlassGuard is a navigation-oriented framework that reconstructs planar architectural glass from complementary visual and LiDAR evidence while treating both success criteria explicitly: glass coverage and free-space contamination. A foundation vision model supplies glass-instance masks, structural 3D cues generate metric plane hypotheses, and depth-free 2D projective geometry checks their orientation before they enter a consolidated global map. Evaluated in nine building-scale scenes with over an hour and 2.1 km of real-world robot traversal, the panoramic version reaches 85% of total glass coverage; under identical pinhole inputs GlassGuard reaches 82% coverage versus at most 61% for the evaluated baselines, while producing 5-17x fewer false voxels per frame. Qualitative examples with a navigation planner show the reconstructed planes blocking paths through glass while leaving traversable routes open.

Paper · arXiv:2610.02110

Glass is a strange adversary for a laser scanner: it is fully visible to the camera and effectively invisible to the rangefinder. Reconstruct it carelessly and you trade one failure for another, walling off the very space the robot is supposed to drive through.

Background

Transparent and specular surfaces are a persistent failure mode for LiDAR-based SLAM and navigation. Because laser returns may pass through glass, the collision boundary simply never appears in the map, and a planner can command the robot straight at a window. The obvious fix — reconstruct the missing surfaces — introduces a second, subtler failure: if the reconstructed plane is placed slightly wrong, it lands in free space and makes traversable area look blocked. Prior work has largely optimized for coverage, reconstructing surfaces without measuring the damage those surfaces do to the free-space map. The paper’s central move is to treat this as a dual requirement rather than a single objective.

Core Innovation

GlassGuard is a navigation-oriented framework for reconstructing planar architectural glass from complementary visual and LiDAR evidence. A foundation vision model produces glass-instance masks; structural 3D cues turn those into metric plane hypotheses; and depth-free 2D projective geometry checks the orientation of each hypothesis before it is admitted. The dual success criterion of glass coverage and free-space contamination is applied throughout, both when verifying individual proposals and when building the consolidated global map, so a candidate plane has to justify where it is placed rather than merely existing. The result is a mapping pipeline that adds the boundaries a planner needs without carving up the space it must cross.

Results

  • Evaluated in nine building-scale scenes spanning diverse glass structures, spatial scales and lighting conditions.
  • Over one hour and 2.1 km of real-world robot traversal.
  • The panoramic version reaches 85% of total glass coverage.
  • Under identical pinhole inputs, GlassGuard reaches 82% total coverage versus at most 61% for the evaluated baselines, with 5-17x fewer false voxels per frame.
  • Qualitative navigation-planner examples show reconstructed planes blocking paths through glass while leaving traversable routes open.

Limitations

The strongest numbers are coverage and false-voxel rates, which are proxy metrics for the property that actually matters — whether a downstream navigation stack drives better and safer. The paper illustrates this with qualitative planner examples rather than a quantitative mission-level evaluation, so the translation from better mapping to fewer glass-related navigation failures is shown rather than measured. The evaluation is planar architectural glass in office-like or building-scale settings; curved glazing, glass furniture, mirrors and heavily cluttered reflective interiors are different problems that the plane-based formulation may not cover. Coverage of 85% also implies that 15% of glass stays missed in the panoramic case, and the abstract does not characterize where those misses concentrate — precisely the kind of residual that a safety argument would need to bound.

Industry Implications

For mobile robots deployed in human spaces — offices, hospitals, retail, airports — glass is a classic cause of embarrassing and occasionally dangerous navigation failures, and this work is directly actionable because it targets the deployed stack rather than a benchmark. The dual-criterion framing is the transferable lesson: mapping systems that add obstacles to fix false-negative collisions should be judged on false positives too, because a phantom wall is its own operational failure. The reliance on a foundation vision model for instance masks suggests the pipeline improves as those models do, which is a favorable cost curve for adopters. Teams should still treat the coverage and false-voxel metrics as necessary but not sufficient, and instrument their own fleets for glass-related interventions before declaring the problem solved.