Many things that many people believe robots can do are in this category. The absolute state of the art is balancing (not walking) on 2 legs (which does allow you to move, but ...), and walking on 4 legs (it's not really walking like for example a cat does, but it's better than balancing)
But the control loop for grasping arbitrary 3d objects given a camera image of the object and the hand ... such a stupid thing is a completely unsolved problem, and it's we're not anywhere near solving it.
Here's how amazon sidesteps that problem : http://www.youtube.com/watch?v=6KRjuuEVEZs&noredirect=1
Is there some downside to this that I'm not seeing?
Most real world engineering problems boil down to long term reliability. Its just assumed you'll get your statics and dynamics correct and all that basic stuff, the actual brain power burned on the job is designing stuff that actually works 1M times and tolerates every component being 5% off the exact value or whatever. I think this would be a huge headache in the very long term reliability category.
Given that Amazon is no stranger to technology I suspect the answer is that it simply doesn't make financial sense. I suspect the capex on a standard Amazon shipping center is probably pretty low (lease a warehouse, put in computer systems, hire a crowd of temp workers).