All is based in “robots can’t see outside text” = robots can’t see reality, but I think they can be programmed to see “outside text” and take data from their “own eyes”.
They are not programmed, they learn from piles of text. That is literally how large language models work.
1