IMHO there is a big difference between FSD and bot: FSD is designed to be used on public roads and has to deal with lots and lots of crazy edge cases that come along once in 10 years or so. To make it safe for those, the training requires tons of driving data because such cases have to be included.
For the bots, on the other hand, the most likely (and still possibly very profitable) use case is working in a well defined factory environment, where they are only trained to do one job as seen in the video. That requires massively less training data. Also, as you said such jobs usually don´t put anyones life at threat when things go wrong, so you can do much more trial and error which reduces the time to get the bots to work even more. You don´t need one bot that can work on lots of different jobs without new training, while FSD has to do traffic circles, unprotected lefts, blinking red lights, emergency vehicles, construction areas, rain, snow... you get the idea - in one single piece of software.
Regulatory problems should be much easier to as the bots will likely not be working im public places accessible to anyone (at least in the beginning).
I agree that's one (and probably initial use case of the bot), however if they become consumer products they will have to navigate the real world, peoples houses, outside world etc. However I think that is years away so they have plenty time to train the bot for that.