I keep a RealSense D435 on my desk for exactly this reason. RGB-D sensors return clean depth on matte objects, but glass, mirrors, and transparent acrylic give you literal holes in the map. The robot sees nothing, plans through a window like it is open air, and you find out the hard way.
This clip shows the standard triptych: RGB on the left, raw sensor depth with black voids in the middle, and completed depth on the right where the glass panel suddenly has actual geometry. The depth-completion model on top is built on LingBot-Vision, the Apache-2.0 backbone Robbyant released with no strings attached, unlike the custom Meta license DINOv3 carries. The loader is the custom lbot_vision_infer lib, so at least the open weights are real and the license is genuinely open.
The paper reports their distilled ViT-L hits NYUv2 RMSE 0.310, roughly equal to DINOv3-7B's self-reported 0.309 at roughly 23x fewer params and less than one third of DINOv3's training samples. That efficiency matters if you are trying to fit perception onto a cheap edge board beside a RealSense. I am not running those numbers myself, and the KITTI numbers in the same paper trail the same baselines, so the headline is mixed.
Here is the honest letdown: only the four Vision backbones are open. The depth-completion weights on top, the part that actually produces the right panel, are not released. So this clip is a vendor comparison-page render and I cannot verify it independently. The useful head stays closed, only the backbone is Apache-2.0. That stings in an open-weights community.
I am curious whether we see open completion heads the way we saw open detection heads after DINOv2, or if a strong open backbone matters less when the downstream heads that make it practical stay proprietary.