look at the boston dynamics android vs the tesla optimus. creating a robot that can perform physical tasks is far from impossible. creating a fleet of robots to replace all the physical tasks done in a nationwide economy is economically infeasible.

the key distinction is which sector of labor is more scarce. if knowledge work in a digital world is scarce, there's more incentive (i.e. investment) to deploy software at scale. if robotics manufacturing in a physical world is scarce, there's more incentive to stockpile the limiting reagent (usually the chips & minerals).

in terms of timeframe, robotics intelligence is accelerating rapidly. look at RT-1, for example. it's clear that the public domain has already adopted the means to operate robotics, at a hobbyist level, on a global scale. there are a plethora of youtube videos demonstrating this. I suspect that RT will be iterated enough within 5 years to include the capability of self-reproduction. by that point it will not be economically infeasible for independent actors with enough capital to procure their own self-assembling factories with one ML model. the limiting factor will be the raw resources for manufacturing enough robotics to do the physical labor of a given niche.

within 10 years there will be a ML model that convincingly operates an android that instructs/outperforms humans in basic physical tasks. that's the trajectory we're on. it is likely that within 20 years, even specialized physical tasks will be done by android-operating models. before then, I suspect that there is a point of no return with androids playing televised sports. by then the form factor will be a nonissue for industrial purposes.