LiDAR-based robot maps can leave glass doors and walls marked as open space. If a system puts an estimated pane in the wrong place, the map can mark a clear route as blocked. Carnegie Mellon researchers report that GlassGuard, a new robot-mapping system, found more visible glass and added fewer false obstacles than two tested methods in recorded building runs.
Glass is difficult to map because laser scans may pass through it or bounce away, while a camera image alone cannot say how far away a pane is. A guessed wall placed at the wrong depth can turn a clear passage into an occupied area on the robot’s map.
GlassGuard uses a smaller, glass-focused version of SAM3, an image-segmentation model, to outline panes in camera images. It matches each outline with nearby LiDAR points to estimate the pane’s location and angle in 3D. The camera outline helps reject a plane pointing the wrong way, while LiDAR measurements anchor where it sits. As the robot sees a pane from new angles, the system can merge estimates or remove planes that no longer line up with the image.
The team tested the method on a wheeled robot across nine recorded scenes in six indoor and outdoor environments, including daytime and nighttime conditions. The routes covered more than an hour and 2.1 kilometers. For the quantitative comparison, each method was replayed against the recorded sensor data and its map was checked against manually annotated ground truth.
The researchers divided maps into one-meter 3D cells, or voxels. In the pinhole-camera comparison, GlassGuard marked 82.1% of visible glass cells within the study’s evaluation range at least once in the recorded data. MonoGlass3D and GlassRecon marked 61.0% and 44.4%, respectively, with the same camera view. GlassGuard also averaged 16.7 map cells wrongly marked as obstacles per frame, compared with 290.2 for MonoGlass3D and 85.4 for GlassRecon. Its 360-degree-camera variant reached 85.2% on the same at-least-once coverage measure, but used a different camera input. These map scores do not measure collisions or how often a robot completed a route.
The paper shows planner paths qualitatively but does not report a controlled navigation-success or collision-rate test. GlassGuard averaged 0.74 seconds per frame on an RTX 5070 Ti laptop, while the test robot’s LiDAR ran at 10 hertz; the paper does not show that the method updated at the sensor’s full rate. The Oct. 1 preprint has been submitted to IEEE Robotics and Automation Letters and is not yet peer reviewed. Its public code repository does not include the 2.7-gigabyte model weights.
Sources
- GlassGuard paper, arXiv:2610.02110v1
- GlassGuard code repository
- Illustrative glass-partition photograph by MaarsLivingWalls: Wikimedia Commons file page, licensed under CC BY-SA 3.0. The photograph does not depict the GlassGuard trial.
