Automation of robotics in an uncontrolled physical environment is so, so hard to program for.

This is why we got advanced language, music, art, speaking models first. There's SO much of that data (they stole and pillaged every scrap, but the fact is that it exists) that models can be trained on.

There isn't vast stores of raw sensor data that maps physical environments to servo, hydraulic, pneumatics correlated with optical, heat, tactile feedback etc.

If you think about how uncoordinated toddlers are, and how many years it takes to map all the inputs humans receive and learn how that translates to our environment so we can move around without much thought, you'll kind of get what I mean. That information just hasn't been collected on a large scale in a way that machine learning can train on it.

This is the underlying challenge of autonomous robotics. We'll get there, but for now we still need humans controlling pretty much everything that doesn't have an ultra-controlled space like factory floors.

I was in a grocery store the other day and they had some sign on a Roomba looking thing that was struggling to navigate the produce section. I can't imagine how hard it would be to create a machine that can reliably and safely mine ores autonomously.