Vision based on visible light doesn't suffer from the problem in the same way as echolocation or lidar, since it is not dependant on observing signals emitted by the observer. I guess very smooth surfaces might act as mirrors, which will probably bring it's own set of difficulties for machines. Anecdotally, I can say my own stereo vision doesn't have big difficulties with most smooth surfaces though :)

I know I have slammed into transparent surfaces because I couldn't see them.

It probably gets worse the higher you go, because there is less grassy people adding marks to them, and less "people are hitting this too much, we'd better reduce that glass size". Birds are famous for getting them wrong.

But bats also use vision based on visible light. I haven't read the article yet, but do they explain about why the bat wouldn't see the building with its eyes?

"researchers can rule out the possibility that the bats were visually confused, because the experiments were done under infrared light, which bats can't see, and so they would have been relying entirely on echolocation."

It might be nighttime.

Yes, lidar bounces off surfaces at shallow angles to give false readings. Also, the long wavelength of IR means that lots of flattish surfaces are good mirrors. In mobile robotics, glass panelled corridors are perceived very badly.[1]

Passive depth from stereo usually has big holes where no features are detected[2]. That's why Kinect and the like project their own IR texture into the scene. Outdoors ambient IR overwhelms the projection unless you use powerful lasers.

The upshot of this is that robots are vulnerable to banging into stuff, like bats.

[1] I don't know for sure but it's likely that the right hand side of this map I found online has glass walls causing bad range data: https://i.pinimg.com/originals/e0/87/04/e08704b6dc925b575c2f...

[2] If it doesn't, it's because the holes have been filled by interpolation - guessing what is there rather than observing it directly.