Highly tactile jobs and those in unstandardized environments (plumbers, electricians, etc) are famously difficult for machines to do, and they’ve made little to no progress in that department.

It's true that general-purpose robots have made little progress, compared to general-purpose AI, but the reasons for that might just be financial. AI has become a self-justifying busines model whereas generalist robots are still academic projects, kind of how deep learning was a decade+ ago. Maybe making tactile-aware robots that can navigate random environments will end up being subsumed by ML algorithms using cheapo hardware, kind of like this:

https://mobile-aloha.github.io/

AFAIK this is the current for-profit SOTA robot, which goes for $16,000:

https://www.unitree.com/images/Unitree%20G1%20EN%201080p%20%...

My guess is that it's just an expensive toy, only useful to robotics researchers. But I can imagine in 10 years the big LLM+vision models could somehow pilot them, perhaps very slowly compared to a human, but as long as they can finish the job at 0.1% of a human's wage, it might not really matter how much slower they are.