1.5 million parameters indeed is small compared to an LLM. But why would that be more efficient than old-school inverse kinematics? That’s either simple trigonometry or some moderately complicated matrix maths… Something you’d be able to do at an undergraduate level. And not that many CPU cycles.

Capturing the human movements just with a regular rgb camera and having thr robot mimick it, is impressive, though.

I would say it’s in general easier to control robots than specialists and experts, they usually have a mind of their own and sometimes motivations outside of the controllers requirements