GlassGuard: Verified Glass Plane Mapping for Robot Navigation
GlassGuard is a preprint system that reconstructs planar architectural glass for LiDAR navigation by combining a pruned SAM3 vision model, LiDAR-derived structural pillars, depth-free orientation checks, and persistent map management. On nine real-world scenes and 2.1 km of traversal, its panoramic version covers 85.2% of glass with 15.3 false voxels per frame, while the pinhole version covers 82.1% versus at most 61.0% for baselines and carries 5–17× fewer false voxels per frame. It matters because it blocks collision paths through glass without contaminating traversable space, a dual requirement for safe robots in glass-walled buildings.
Can a real-world robot map planar architectural glass well enough to block collisions through glass without falsely occupying navigable free space?
LiDAR passes through or reflects unpredictably off transparent and specular glass, so glass boundaries can be missing from navigation maps. Existing reconstruction methods can fill in glass but often misplace planes, blocking walkable space and making the map unusable. GlassGuard targets both missing glass and spill into free space simultaneously.
Earlier work used physical cues such as return intensity, reflection patterns, or sonar, which require certain returns or extra hardware. 2D segmentation methods like SAM3 detect glass but lack metric 3D placement. Learning-based methods such as MonoGlass3D and GW-Depth infer depth or plane parameters directly from RGB but are ambiguous because glass has no stable appearance; pipeline methods like GlassRecon align a foundation-model depth prior to RGB-D using local RANSAC but place surfaces from a single frame without orientation or depth checks or later revision.
GlassGuard separates 2D detection, 3D hypothesis construction, and verification. It uses a structurally pruned Slim SAM3 for glass-instance masks; associates masks with surrounding LiDAR to extract vertical and horizontal structural pillars and build bounded plane candidates; verifies yaw using depth-free interpretation-plane geometry from 2D mask edges; then consolidates hypotheses globally with plane merging, floor-evidence eviction, and multi-view reprojection. The pinhole and 360-degree variants run onboard at about 0.74 s per frame.
On nine recorded building-scale scenes across six indoor and outdoor environments, totaling more than one hour and 2.1 km of robot traversal with a Mecanum robot, Livox Mid-360 LiDAR, and Ricoh Theta 360 camera, the pinhole GlassGuard achieves 82.1% ever coverage of line-of-sight glass versus 61.0% for MonoGlass3D and 44.4% for GlassRecon; panoramic GlassGuard reaches 85.2%. At the current-map level on a 1.0 m grid, GG-360 covers 89.1% of collision-critical 0–2 m and 91.5% of 2–5 m demanded voxel-frames, versus GlassRecon 83.1% and 73.4% and MonoGlass3D 98.1% and 95.5%, though MonoGlass3D carries 290.2 false voxels per frame. Concurrent false occupancy is 16.7 false voxels per frame for GG-pin, 15.3 for GG-360, 85.4 for GlassRecon, and 290.2 for MonoGlass3D; under matched pinhole input the baselines carry 5–17× more false occupancy. Ablations on four scenes show removing the angle gate raises concurrent spill from 19.4 to 34.0 voxels per frame and accumulated spill from 269 to 607. Slim SAM3 detection: IoU 0.748 on GDD and 0.671 on GSD-S versus original SAM3 0.815 and 0.713; recall on GSD-S is 0.918; VRAM is reduced by about one-third.
The authors state the system assumes planar near-vertical glass and does not handle curved or strongly tilted panes; inference at about 0.74 s per frame limits update rate. The evaluation is on one robot with one LiDAR and camera combination, using pre-recorded rosbags replayed rather than live navigation, and navigation planner results are qualitative rather than a quantitative navigation-success metric. No code release is reported in the paper. The work is a preprint and has not been peer-reviewed.
This could be integrated into autonomous mobile robot navigation stacks for cleaning, security, delivery, and inspection robots operating in glass-heavy offices, retail, hospitals, airports, and modern buildings. The likely window is 1–3 years, provided inference is accelerated for real-time or near-real-time use, the method is validated on more robots, sensors, and glass types, and code or a commercial version is available. Before then, fleet operators could use it for replay or semi-autonomous mapping.
The full text is not republished here because the paper's license does not allow it. Read the original on arXiv.