There is a software authoring problem (which is where the ML bits are crucial).
If we had to program all robots like we had to with CNC machines, then programming them would be a high skill problem, even if we throw a lot of tools at it.
I can work my way through a Tormach, but is that really what I want to spend time with? The ultra low level specification of what I need done?
I'd love a pedal based training system with something like "Identify", "Orient", "Place", "Count", "Test" to teach it things in steps & get a program out of the demonstration (that donut computer vision project was amazing, because it showed you didn't really need ML to do these things).
Like we have people who are demonstration learners, I wish I could do something like that of going from many scenarios to a final one and have the robot to dissect every one of my actions into a flow-chart of its own.
My approach would be to manufacture custom arms for particular tasks and in principle 3d printing the arms is exactly what i'm getting at (e.g. that optimised physical design processes save on cost and improves performance much more than software + expensive externally manufactured arms). 3D printed arms with comparable repeat accuracy would be an excellent optimisation over buying v expensive Kuka products. Then you could start think about different mechanisms (compliant mech, soft parts etc) and control systems/software.
Kukas are not really just a couple of servos (e.g. encoders) and there are many examples from the 90s of self walking robots with little software too. There's good literature on "morphological computation" or Rolf Pfeifer's book How the Body Shapes the Way We Think: A New View of Intelligence.