It's not expensive, it's an unsolved problem, the main issue for many years has been the modeling of cloth behavior. They already have a folding robot that means they have solved a significant chunk of it.

But the generalization of behaviors is also a large problem. Pick and place is mostly solved I believe, but translating your housework to pick and place in your apartment is still a bit off.

I'm not convinced safety is solved: given how often LLM's hallucinate and automated cars get confused, what happens when something unusual happens and the robot thinks a person's arm needs to be "folded" and breaks it? Or a person's clothing looks like the wall and a robot tries to move their arm through a person's chest?

Sure, for each individual situation one can imagine a safeguard but being in close proximity to unpredictable humans means there are unlimited edge cases... if a humanoid robot is strong enough to do something useful it is also strong enough to be dangerous.