Re:Do large language models understand us?
by thaumasiotes on Saturday December 18, 2021 @04:32AM
(#29603145)
Manipulation involves a few things:
1. Recognize objects in the world, like a keyhole or whatever it is you want to take out of a bin.
2. Recognize your own position -- how close is your hand to the keyhole?
3. Adjust your position until it matches your goal.
My impression of things is that most of our difficulty comes from task 3. It's not so much that we have trouble recognizing where we are, where we want to be, or how to move from where we are to where we want to be. But we have difficulty defining movements that don't cause damage to nearby (or target) objects.
Your human hand has many features that allow it to grip an object securely without damaging the object. And you also have a decent sense for how to move your hand to an object without damaging the object -- or your hand -- though this problem is far from being solved in humans too. Ever stub your toe?
1. Recognize objects in the world, like a keyhole or whatever it is you want to take out of a bin.
2. Recognize your own position -- how close is your hand to the keyhole?
3. Adjust your position until it matches your goal.
My impression of things is that most of our difficulty comes from task 3. It's not so much that we have trouble recognizing where we are, where we want to be, or how to move from where we are to where we want to be. But we have difficulty defining movements that don't cause damage to nearby (or target) objects.
Your human hand has many features that allow it to grip an object securely without damaging the object. And you also have a decent sense for how to move your hand to an object without damaging the object -- or your hand -- though this problem is far from being solved in humans too. Ever stub your toe?