Robots navigate with depth cameras, and depth cameras cannot see glass. The light they send out passes straight through or bounces off, so the sensor registers nothing. That is why robot vacuums bump into glass doors. This clip, from Robbyant's project page, shows the fix. Left is the normal camera view. Middle is what the depth sensor actually sees: the window is just a black hole. Right is the AI completing the window back into a flat wall. The base vision models are open source, but the completion model itself is not released. It is guessing what should be there rather than measuring it, so how much you trust that is the interesting question.