The state of the art of robotic manipulation is really cool: robotic hands are able to figure out which object is which as in this video, the best points to grasp and pressure to apply to hold and move objects. Just need some group to combine it all and demonstrate a real world application of it!

Obligatory video dump because this stuff is just so fascinating:

https://www.youtube.com/watch?v=mIEbU7GfRhQ

https://www.youtube.com/watch?v=DPl_d7lbL84

https://www.youtube.com/watch?v=tLNyXAE7mLM

https://www.youtube.com/watch?v=ZhsEKTo7V04

Conversely to your comment, I find that -if this paper really represents the state-of-the-art- then the state-of-the-art is extremely limited.

The authors of the paper report as their major contribution a technique that must be applied to each specific object separately and their experiments cover 47 objects in 3 "distinct classes" (shoes, hats, mugs and a mix of objects from diverse other classes):

We’ve also shown that self-supervised dense visual descriptor learning can be applied to a wide variety of potentially non-rigid objects and classes (47 objects so far, including 3 distinct classes), can be learned quickly (approximately 20 minutes), and enables new manipulation tasks.

I'm guessing that by "real world application" you mean an application where a robot hand is required to manipulate arbitrary objects of unconstrained shape. Even given the relatively short training time reported in the paper (though with a pre-trained CNN model to begin with) this is not something that can be reasonably expected by this kind of system. There are potentially billions of different object shapes in the real world!

For real world applications a giant leap forward is required: a kind of AI system with the ability to generalise well from few examples. This seems to still be far out of the reach of the current generation of deep neural net- driven vision systems.

"Conversely to your comment, I find that -if this paper really represents the state-of-the-art- then the state-of-the-art is extremely limited."

The state of the art in robotic manipulation is extremely limited (the state of the art in robotic locomotion is also extremely limited, but somewhat easier to paper over - simplified special cases get you farther in locomotion than in manipulation). IMHO, the layer of robotic control systems above servo control, but below spatial reasoning is an enormous hole in the field's research portfolio. We've got pretty decent machine vision at this point -- still not up to human standards, but improving rapidly. We've got excellent servo control laws. We've got good planning and scheduling techniques. What we don't have is techniques for robustly and rapidly performing dexterous motion planning, particularly where the contact forces with the environment are an important part of the problem.

I was at Robotics-Science and Systems this year and heard a talk by Sergey Levine, one of the handful of researchers addressing this topic. He put up a picture of IBM Deep Blue playing Garry Kasparov. He pointed out that there was no robot in the picture -- there was a human moving Deep Blue's pieces around on the board. The point was that we are now in a strange place in robotics: the hardest part of playing a grandmaster--level chess match isn't figuring out the moves. It's reaching out and picking up the pieces.

Thank you for the useful insight. I'm not an expert in robotics, far from.

Like you say, machine vision has progressed, but it's still very far from human vision. For one thing, most of the progress is in object classification, so any problem that can't be reformulated as a classification problem is basically out of the question. Then, there's the issue of poor generalisation between classes- for any class of object that we want a system to recognise, we must train the system anew, at great cost- except for pre-training of course (which basically means someone else already did the hard work).

...and evenif we completely solved the machine vision problem, the motion planning under cintact problem would remain.

At some point the world will wake up and realize that this problem is approximately as hard as the machine vision problem, and will devote resources to solving it. But that awaeness hasn't happened yet.

If automation is being used, the amount of objects being dealt with is probably minimal to begin with.

One thing that is overlooked is that it isn't just a learning advantage that people have over a robotic arm, there is way more input as well. We have sensitivity to pressure, weight and sound - even so people don't generalize lots of physical objects from only a few examples either.

>> We have sensitivity to pressure, weight and sound - even so people don't generalize lots of physical objects from only a few examples either.

I dare say we do. For example, once a toddler has figured out how to grasp and manipulate objects, he or she can subsequently grasp and manpiulate any object that they are physically able to grasp and manipulate, regardless of shape, colour, size, texture etc.

For example, if I know how to grab a Barbie doll, I know how to grab a Teddy bear, a firetruck, my mother's skirt, a spoon, etc, etc with the same level of skill.

I must be misunderstanding your comment because this ability to generalise from even a single example is a hallmark of human intelligence and an especially striking example of its differences to machine learning systems.

The associated video: https://www.youtube.com/watch?v=L5UW1VapKNE

We've come not as far as you'd expect since SHRDLU (1968): https://en.wikipedia.org/wiki/SHRDLU

This is incredible! The temporal consistency is amazing. Watch the video.