Except how you don't.

Are you saying that it isn't possible to measure and then simulate a robot doing sophisticated manufacturing work?

Here is a video of a robot following a program that was created by algorithm and then tested in simulation, before manufacturing a real, usable product.

https://youtu.be/CL-I-RvMeAU

You will notice that it is fourteen years old.

Autodesk makes Fusion 360 a product specifically designed and built to run CNC machines, which includes accurate simulators. There are tens of thousands of videos of both the simulation and then the actual cutting of parts. Go and look!

Now why do you think that this process can't be used to grade Neural Network output?

Is it that you don't know that CNC tools are robots? Or that you think that the same techniques can't be applied to a six axis robot arm with a tool? Is that your issue?

https://youtu.be/Hze1D_B2ui4

Here is a three year old video of a robot arm using cold spray additive manufacturing to print a solid copper rocket nozzle bell housing. The tool paths were generated by algorithm and then simulated before the print.

Where do you think I said that robots would be able to repair themselves? Can you highlight that for me?

Why would replacing technicians that do repairs on equipment be more important than actually NN tools to reduce the cost of the design, manufacturing and assembly of parts?

It really seems like you are responding to an argument I didn't make.