It’s got to get good enough for open source versions at some point though.

Eventually, but consider the power envelope for compute in an android is 10-100x lower than for a car*, while at the same time the range of tasks a general purpose bot has to be good at is a strict superset of driving a car, and that open source attempts at self driving cars are far behind the proprietary attempts, and that compute efficiency is only doubling every 2.5-3 years (Koomey’s Law), so expect something like 20 years before these are properly general purpose and also not remote controlled.

Still plenty of value from special purpose and from remote control, but that's the timeline for solving both at the same time.

* even with the compute being external, the global electricity supply is presently a few hundred watts/capita and already being used for all the other things we want electricity for, hence all current anger about data centers making electricity too expensive, but renewables could tilt this to 8 billion robots with 1kW compute each in as little as 10 years if we brush aside all the decarbonisation efforts and keep burning petrol in cars and gas in stoves etc., otherwise more like 15 years for that.

That would shock me. These things are going to have a TPM