One thing it doesn't do, that would be particularly powerful, is emulate the geometry of the human arm. If it did, and it if it was typically paired with a mirror image arm, you would be able to very easily train it to do most tasks that can be done by a human. Instead of "programming" it to, say, iron a shirt, you just get someone who is good at ironing shirts to do the task over and over while it copies them, gradually improving its abilities.
Combine that with designing it in a way that it can do a large part of making new robot arms, and the costs should drop and the capabilities should increase....dramatically.
Yes I assume there are cameras, and that it isn't copying exact moves but actually learning. I'm not suggesting it would figure out how to iron a shirt quickly, it would take thousands upon thousands of hours of watching/copying as someone is ironing. But hey we were worried about robots taking jobs...and there's a job! (and they are also producing ironed shirts!)
Cameras aren't particularly expensive, the ones made for cell phones can do fine and they are a couple bucks. The software is hard but machine learning is improving by leaps and bounds. And software may be expensive to develop, but it isn't expensive to make another copy of.
The problem isn't the hardware here, the problem is the software. Robots just aren't smart enough to perceive the world around them yet. People are working on it, but it's still very much in the research phase at universities. You should check out HERB at Carnegie Mellon (1), which is working on things just like this. A fun anecdote - all the students spent forever trying to get the robot to open a microwave (and keep in mind this is a highly controlled environment and they knew exactly where it was). Eventually they gave up and modified the microwave.
True. I bet if the hardware was cheap enough, the software would advance a lot quicker.
I mean, imagine the only really useful thing one of these things can do is make a copy of itself. And it might takes it a month running non-stop to do it, but the parts are cheap so it only costs a few hundred bucks for a new one. Would you want one of these things in your basement working away? I sure would. And I'd be pretty interested in tweaking its software and training data and whatnot to make it faster and better.
Sure it'll be a good while before it is folding laundry or picking weeds in the garden or whatever, but I think the pace of improvement would pick up quick once these things are available in mass.
I went to a seminar on collaborative robotics (industrial robots that work in close proximity with human workers) and one of the vendors had a robot arm that you programmed by simply moving it from position to position, telling it what technique to use to interpolate between them. You could also easily set up reference planes by probing the work area or workpiece, then constrict motion to an axis in that coordinate system, say for inserting parts into holes on an assembly. Not quite the "learning" system you describe, but easy to use and understand. I think one example application they showed was a robot that was programmed by this method to insert lifter rods in one side of a V8 engine block (which was hard to reach), while a human did the ones on the easy-to-see/reach side. It was pretty slick.
Combine that with designing it in a way that it can do a large part of making new robot arms, and the costs should drop and the capabilities should increase....dramatically.