Finally, someone is spending enough on robotic manipulation to make real progress. $2 billion and the cooperation of Fanuc should yield some progress.
It took Boston Dynamics about 15 years and $125 million to get to Atlas. There's still no market. This business requires a sugar daddy. BD had DARPA, then Google, and now Softbank. This is a very expensive area in which to work.
Robot manipulation in unstructured situations works very badly. Watch these videos:
DARPA-funded robot manipulation work from 2012.[1]
DARPA-funded robot manipulation work from 1973.[2]
See the improvement? No?
It's not clear that just throwing deep learning at the problem will work. It's been tried. Even general bin-picking still doesn't work.[3] The author of that article, who is from Fanuc, says maybe by 2020. Lots of special cases work already, but if all your parts are the same simple shape, a bowl feeder is simpler. Picking from a bin full of partly entangled parts is beyond what robots can do today.
Someone got bin picking to work for an irregular sheet metal part by going in with a magnet, picking up something, anything, then weighing it. If it's underweight, they missed, and they try again. If it's overweight, they got more than one part, so they drop it and try a different location in the bin. If they got one part, they now have it out where they can look at it and orient it. Tricks like that make production lines work.
Pure geometry isn't enough, and pure machine learning isn't enough. But together, they might work. Good to see this being funded.
Here's the 2017 Amazon picking challenge.[4] Still not there, but better than last year.
Take a look at RightHand Robotics. They’re shipping a bin packing arm that achieves millions of cycles working with entangled objects in recent demonstrations. I think this task is by and large solved. https://www.righthandrobotics.com/blog
That looks largely similar to Ocado's solution that gets trotted out at trade shows. Kinect V2 coupled with a suction-based system. The difference here is the gripper. Their rep said that they found that suction works reliably and provides good grasp detection (e.g. loss of suction).
It's worth pointing out that this capability solves the problem of moving an arbitrary object from A to B. That's really not that hard with a suction cup. All you have to do - and this is how Ocado solve it, from my reading of the presentation - is locate a mostly-planar surface in your point cloud and suck it. All that changes in this case is the gripper closes to hold the part in. Bing - it works if you have no idea what the part is, beyond some basic constraints like size or weight. You don't need to worry about 'correct grasps' or anything like that.
What it doesn't solve is picking up a part precisely from a bin/tray and in a known orientation. That second problem crops up in manufacturing all the time.
>It took Boston Dynamics about 15 years and $125 million to get to Atlas. There's still no market. This business requires a sugar daddy. BD had DARPA, then Google, and now Softbank. This is a very expensive area in which to work.
That's actually crazy, I had no idea the amount of money involved in this was so small. Investors will drop billions on the latest IRC clone, why aren't they putting more into robotics?
There's definitely a lot of progress being made! When I was paying more attention to the field last year, Google's research team working on robotics had pushed out a really interesting paper [1] and some more recently [2].
Every time I see another PR release about an awesome use of AI, I get a little depressed. Maybe learning Java 8 streams, or Python 3 , or Scala 3, or anything other than Andrew NG's AI course and my old linear algebra notes is a waste of my time...
And then I ask Alexa a question - something a bit more complicated than the weather or what's the best Prime deal. Then I'm relieved.
Until I see something more impressive than a massive if-then-else, I'm not losing any sleep over my lack of interest in this latest fad.
AI as a bunch of if else statements is a very 80 centrics view of current AI.
My computer vision professor has us read a current paper every week. He really really dislikes the use of deep learning and has said in class he tries to look for papers that don't use it.
But nearly all of the current papers are using it and he admits every week that the papers he presents are blowing old research out the water using it. Throwing a bunch of GPUs and image data at a problem is surprisingly effective for a lot of problems that were previously extremely difficult. He begrudgingly admits this every week, but he does continue to find papers that get slightly better results by combining deep learning with traditional computer vision techniques. I think he is right in that there probably is a peak quickly approaching where data/augmented data with zero previous knowledge will peak with the current black box approach, but cutting edge currently is definitely not a lot of if-else statements.
Well, what's disappointing from a vision standpoint is that we are not making much progress understanding vision. Instead we are making progress in solving machine vision.
I think your professor isn't looking for results so much as a good explanation for the vision problem, a "unified theory" of vision that you could, in principle, code up from theory alone.
That is my impression of current AI research as an almost layman at the field. Using traditional proven methods to get an approximation and then deep learning and/or other machine learning algorithms (and trained data) to refine those approximations.
Seem to be a very "organic" kind of AI, even with the flaws of human minds. Thinking about it can be spooky given how much it evolved in just a few years.
i agree there is definitely a gap between what is possible in the near vs distant future.
but i personally would embrace it rather than fear it. and i'm not a tech guy, i work in a call center selling insurance lol.
the day that i can say "Mr. Robot, make me a pizza, brew me some cold beer, and tell your robot friends to deliver me a couch for the game tonight" is the day that i am pretty damn satisfied with the state of technology
if everything in the world is manufactured and fulfilled by robots, and i never have to leave my home to do anything, i would be thoroughly pleased
I have a friend. He was hired in an industrial complex, doing a lot of hard work, moving stuff, building, most of the time outside, in cold. One day he got fired, not his choice. After some time he found a nice job, this time in an office. He was working 100% from the computer, managing some stuff, and sitting all day. In the first month he was amazed what he missed, the calmness, it was always warm, no physical labour, etc.
I met him after 3 months there. He was the most depressed person I ever met. What happened? Lack of activity.
It could free up time for other more enjoyable activities, though. For instance, I could spend my time playing my drums instead of mowing the lawn, if a robot did it for me.
I think about that what Musk said about Robots. He said, 'In the future, we will need a strobe scope to see robots'.
Automation scares me more than my state requiring me to buy insurance for my jalopy? (Insurance that seems like the price was colluded, but I will never be able to prove it. And that's second to the clever way the Insurance industry brainwashed us, and our legislatures.)
I'm a chit job guy.
I'm usually the guy with two lousy jobs, and the lousy jobs are slowly being eliminated by machines. Or, being filled by immigrants who wouldn't think about rocking the boat at their new career. People just happy to be alive? I get it.
I get it, but this used to be America. We knew we were a bit spoiled, but very greatful we were born here too.
People are clamoring over this great current economy.
Yes--I see a lot of tech guys driving new cars, and very satisfied in life. Well except the rent? I just see so many people with jobs that are a few years of being irrelevant. And yes--many will be tech. On a personal note, I know three homeless people. Two were former Programmers.
The chit jobs, including construction, will always be there, but they are paying less, and less, while expecting more. Yes--Mike Rowe I'll move to the middle of the country for a six month welding gig, usually tied to some natural resource that's hot.
The chit jobs I'm looking at are really getting bad.
Uber always poops up. Yea--I'm going to buy a Uber approved vechicle, and drive at their whim. It's like they pride thenselfs over screwing over a potential employee?
What I'm trying to say, is it seems like the wealthy are doing their best to eliminate guys like me, and with robots they look clever? Not greedy, but clever--
They didn't like unions, so move the whole racket overseas.
They still want to rid humans from the equation; build better robots. Something our Forefathers wouldn't have even hallucinated about when writing that document (The Constitution) that kinda served us well, or some of us?
Now--in order to make things more efficient--let the robots do the job. And get rid of that employee, along with his retirement account, and his pesky health care demands!
The movie Norma Rae, starring Sally Field, couldn't be filmed today. I was watching it the other night, and while I loved it, I just thought about currently empty (I'm assuming) Cotton Factory sitting empty in the Midwest. In the movie, the machines were the robots. The machines weren't the enemy though, it was the greedy owners.
The owners finally found a way to rid themselfs of those pesky unions by moving.
If machines/robots get so good, the greedy owners will be able to eliminate all human help. The world will be there marketplace. Every country will have the 1 percent, and there spawn who will buy, and the few who have very protected jobs, like doctors. (Doctors are a bad example. They are only protected here.)
Guys like me will dead, or like zombies stumbling around the buildings while the robots do their job. (I'm sorry about this essay. I know it's convoluted.)
I watch tv, and see the wealthy yelling things at a machine, and the items showing up on doorstep. Where will people be the money to pay for the connivence?
It's gonna all be rosy for the the wealthy, but guys like me will be delivering the packages, until a robot takes that job.
(Sorry about the discombobulated essay. I'm just seeing happy people, and way, way too many misserable/Homlessness people.)
It's a demo. Crafted to show off the best side they possibly could. Who knows how well it actually works in practice when google isn't cherry picking the best results to show.
I wonder when HN will start having HN-trained bots posting. There's lots of smart discussions by smart people here; surely high-quality grist for an AI mill?
I worked with them a while back, perhaps they were just not very interested in the project but I found PFI to pretty poor overall... disorganized, lacking and unwilling to obtain domain knowledge etc...
Will be interesting to see if anything actually useful comes out of this, this article seems mostly a PR piece without any concrete results.
Preferred Infrastructure. They have this weird naming system which I don’t fully understand there’s Preferred Infrastructure and Preffered Networks. They seem to be essentially the same company, but use two names...
From what I've understood, the founders started Preferred Infrastructure and then spun off Preferred Networks when deep learning started getting really hot. The two founders in the article moved to PFN and PFI kinda stopped doing things in general.
> Nishikawa spoke at his Tokyo headquarters, a drab collection of meeting rooms in an old office building more fitting of a down-on-its-luck insurance company.
Man, the reporter spared no words in this one. Reads like the beginning of a noir novel.
This is a very interesting team. In an interview, one of the co-founders explained that they do not intend to develop their own products, but want to cultivate deep relationships with companies in a variety of verticals and develop expertise that way.
A number of San Francisco VCs flatly rejected them because of this strategy, but they seem to be doing well.
That's if it learns from liberal news sources. If it read only conservative news then the robots would pull up their bootstraps, work hard and be damn happy with their paychecks.
Completely OT but I always find it fascinating how liberal differs so much between languages. In my country liberals are the ones actively working against unions and who embrace capitalism.
You might be confusing liberalism with neo-liberalism. Yes the term is being appropriated by the right-wing establishment to make it seem like they're promoting freedom.
Not really, the concept of Liberal as capitalists against social reforms proposed by unions goes back at least to the 19th century:
"What, then, did and do Liberals (for the most part) understand by this freedom of the individual, or individual liberty, and why have they always made it such a strong point in their political faith? The answer is, they meant by individual liberty, first and foremost, the liberty of private property as such, to be uncontrolled in its operations by aught else than the will of the individual possessing it. What was cared for was not so much the liberty of the individual as the liberty of private property. The liberty of the individual as such was secondary."
That part is the same. The big difference is that the political ranges between countries barely overlap so the people you call liberals would be conservatives elsewhere is more of an indication of these labels being relatives rather than absolutes. America doesn't really have a 'left' by European standards.
I realize this is in jest, but consider this. The only high level intelligence we are aware of (human) has all of these ... bugs. Like getting bored, or selfish, or lazy, or malicious. There is no example of high level consciousness that we know of without those things. What if we create truly autonomous AI and it suffers from depression, or existential angst? Science fiction has focused on the evil AI problem, but what about the lazy AI problem?
Those bugs aren't unique to humanity. You can find those behaviors In many of the more intelligent animal species as well.
Dogs are smart enough to get bored, or jealous.
Parrots seem to be a living nightmare because they are smart enough to develop all sorts of interesting traits. How do you deal with a depressed parrot?
At first glance, it doesn't look like they have much more secret sauce than just applying the state of the art deep learning techniques. The most promising AI companies are the ones advancing the field.
This company is behind chainer, a deep learning framework that partly inspired pytorch [0]. In fact, its was one of the first frameworks to come out with define-by-run (aka eager execution). The also developed CuPy (think numpy on gpu) [1]
The most promising ones (from the perspective of business) translate research into consumer ready products, not pushing the theoretical bounds of the field.
This. R&D is great, and gets us on HN excited, but it takes a Steve Jobs to come along a few years later to produce a product which creates a market which stimulates and funds the next round. Business success is important for technology development. Voice assistants are probably close, but I feel like we’re on the verge of a big category-defining concept/tech/product in the AI field. The basic technology is there, plenty of people can identify needs, but I don’t think we’ve hit the ‘iPhone’ moment yet.
Japanese are good at implementing stuffs. But don't trust Japanese in software development especially deep learning research. Education quality here has been decreasing for decades.
Why the downvotes? This is a serious issue. Developer salaries are low in Japan and the field does not attract a lot of smart people. Traditional engineering is respected, while developers are (still) seen as the "computer guys in the basement".
The downvotes are probably because it's phrased like a race-essentialist statement, a type of statement that many people are very wary of (rightly, imo!). As you point out, there's a good chance it wasn't intended that way.
It took Boston Dynamics about 15 years and $125 million to get to Atlas. There's still no market. This business requires a sugar daddy. BD had DARPA, then Google, and now Softbank. This is a very expensive area in which to work.
Robot manipulation in unstructured situations works very badly. Watch these videos:
DARPA-funded robot manipulation work from 2012.[1]
DARPA-funded robot manipulation work from 1973.[2]
See the improvement? No?
It's not clear that just throwing deep learning at the problem will work. It's been tried. Even general bin-picking still doesn't work.[3] The author of that article, who is from Fanuc, says maybe by 2020. Lots of special cases work already, but if all your parts are the same simple shape, a bowl feeder is simpler. Picking from a bin full of partly entangled parts is beyond what robots can do today.
Someone got bin picking to work for an irregular sheet metal part by going in with a magnet, picking up something, anything, then weighing it. If it's underweight, they missed, and they try again. If it's overweight, they got more than one part, so they drop it and try a different location in the bin. If they got one part, they now have it out where they can look at it and orient it. Tricks like that make production lines work.
Pure geometry isn't enough, and pure machine learning isn't enough. But together, they might work. Good to see this being funded.
Here's the 2017 Amazon picking challenge.[4] Still not there, but better than last year.
[1] https://www.youtube.com/watch?v=jeABMoYJGEU [2] https://archive.org/details/sailfilm_pump [3] https://www.robotics.org/content-detail.cfm/Industrial-Robot... [4] https://www.youtube.com/watch?v=1QqQLq5hsN4