Its always funny to realize how "easy" beating human-intelligence is (Chess AI, Go AI, even Mathematical Proofs), but how hard beating human-simple behaviors are.
IE: Stand up on two legs, walk forward, and move these 10 objects from this bin to that bin without breaking any of those objects. Or fold these clothes (surprisingly difficult to get a robot to do that).
We can build super-advanced Chess and Go AIs that beat out the grand summation of theoretical human knowledge (with Chess theory going back hundreds of years, and Go theory going back maybe thousands!!), but we still can't move a variety of objects from one bin to another successfully and repeatably.
-----------
Another funny example: AIs (or various algorithms at least) have solved the object camera tracking problem, but still basically fail at depth perception and figuring out if something is "still" or "moving" based on sight alone.
A fun way to think about this is by analyzing the 'training-time' for these tasks.
The things we think are easy, we've been trained to do through millions of years of evolution. Running with two feet through a forest is pretty easy for us. Not so much so for computers.
The things that we think are mentally taxing, we've only been doing for a few thousand years or so.
>The things that we think are mentally taxing, we've only been doing for a few thousand years or so.
It's kind of fascinatingly serendipitous in the sense that computers/AI are super quick at performing the tasks that we humans find extremely taxing (complex mathematical computation, visual simulation) while struggling with tasks that comes naturally to us humans (gripping objects, running with two feet--as you described) owing to evolution.
I wonder how much of that is because of constituent 'building block' materials that make up the basic structure of homo sapiens and computers -- carbon and silicon, respectively.
This phenomenon has a name: Moravec's Paradox [1]. If a chimp or a four year old can do it, a computer probably has a hard time with it.
An interesting counter-example to the trend is theorem proving. A symbolic task that wasn't present in the ancestral environment, but humans crush computers at it.
> Its always funny to realize how "easy" beating human-intelligence is (Chess AI, Go AI, even Mathematical Proofs), but how hard beating human-simple behaviors are.
This is the baseless myth that won't die. We are nowhere near rivaling humans in mathematical proving. Even after a human has proved a theorem, it can be an astronomic amount of work to even explain the proof to a machine after the fact (i.e. to construct a machine-checkable version of the proof), even when using suites of sophisticated software tools developed solely to facilitate this process. Progress has been glacially slow since the field began in the 1950s. Deep learning has had zero impact.
We will have robots that can grasp competently way before we have machines that can rival humans in mathematical proving.
True: automated mathematical proving is strictly limited to certain domains right now. First order logic (and other higher forms of logic) are outside the scope of machines, although machines can somewhat solve logic puzzled correctly stated in first-order form (its a semi-decidable problem, so "somewhat solve" is basically the best we can ever hope for).
But move down to boolean logic and symbolic computation... the stuff that translates statements in Verilog or VHDL into pure logic statements for small LUTs in FPGAs or logical NAND gates of hardware synthesis??
Oh yeah, machines are way better than humans. People pretty much rely upon automatic synthesis for circuit design and boolean logic these days. Proof is in the pudding.
Perhaps it isn't the same kind of math you were thinking about. But Boolean algebra is still math, logic, and proofs. Of course, in practice, humans use these tools to build larger structures... but I'd argue that the machine handles a lot of the rote proof stuff (optimally figuring out the minimum number of NAND gates to represent a certain truth table).
It is very telling to me that very large animals that we would consider much less intelligent than a human (elephants, sperm whales) still have much larger brains than us. All that grey matter is needed for something! (not that all of it is going towards locomotion).
>> Its always funny to realize how "easy" beating human-intelligence is (Chess AI, Go AI, even Mathematical Proofs), but how hard beating human-simple behaviors are.
Perhaps the lesson to learn is that we are not as good as we think at board games and maths, etc.
That's not a very good article on gripping and manipulation. Nature was once a serious scientific publication, not PR Newswire.
The state of the art is still bad, and it's been bad for a long time. Watch these two videos: 1960s: [1] 2012: [2] Note how little progress there has been.
Here's the actual winner of the 2018 Amazon picking challenge.[3] The system mentioned in the article did not win.
Throwing deep learning at the problem helps, but not all that much. It's amazing how hard this is.
Christ... It's not a "brand extension" (whatever that means); it's a blog post intended for the general public, not researchers. If you want research, look at the journal archives.
A "brand extension" means you take your famous name with a good reputation and paste it on something else to boost the new thing. This risks the reputation of the famous brand if the new thing is a dud. AdWeek has a list of embarrassing brand extension flops.[1]
Even if it's not a total flop, it can degrade the brand. Holiday Inn Express and Hilton's low-end product lines are examples. The overall result was to move the brand downscale.
Great, thanks. Anyway, I assure you that the journal Nature will not suffer any loss of status by having blog posts summarizing recent research for laypeople.
There's been progress, just not from those people. Festo, the big German automation company, built a better gripper using a silicone semiliquid rather than ground coffee beans. Festo has many different end effectors for industrial automation.[2] That's more useful in this space than a one-idea company.
My pet theory is that this is going to be way easier when tactile/pressure sensors drop their cost/resolution by orders of magnitude.
Animals basically have a feedback loop that works a little bit with proprioception / position measures and way more with expectations of tactile sense (to see this, reflect on how you move in the dark: not thinking about your joint angles, but about that not to kick!).
As a hobbyist, for example, there is no way I can have a "video" of pressure maps on an artificial skin, so I can only hack my way with a one-dimensional finger-mounted pressure sensor. How on earth can I program a robot to use his hand only by looking at it? I would magine that most researchers are in a similar way restricted to work with subpar feedback loops.
While it's true that animals, including humans have a long list of tricks built into their bodies to increase their chances of a correct grasp, to enable force application in the first place, and in some cases to deliberately damage whatever they're grasping (thus wounding and/or killing prey), I think you'll find that the dexterity required to jump a cat body at all is very, very hard.
Moving a cat body, at 130 kph, on an angle that will take the front claw (they always use the same one) across the throat of a wildly jumping around animal (with enough force to allow the claw to rip it open) while keeping that same cat body away from any hooves (one hit from even deer hooves will at the very least end the hunt of any puma, and can do damage up to killing it) is ... it's not just ridiculously hard. It's almost absurd what sort of control you'd need.
I'm also not very sure how tactile sensors help, even if they could not just register contacts but air speeds as well, that does not seem like it would make the problem much easier.
Plus cats are a very "general" body. The number of ways it can move, because it lacks a fully connected inner bone structure (so almost any cat joint can move remarkably far in almost any direction). I'm pretty sure this adds insult to injury and makes it a lot harder yet again.
Well, it has taken years but seems like Boston Dynamics and similar companies have moderately agile walking robots. Yet somehow these don't seem to be appearing anywhere.
Now we hear there's progress in gripping.
The thing about both these steps is walking and grasping are extremely challenging tasks to robots but even more challenging is the easy integration of these and other movements that humans achieve when doing "simple tasks" and especially cooperating in simple tasks - in the "unstructured environments" referenced in the article.
It's hard to find practical applications for walking. The vast majority of environments in which mobile robots can be useful also have smooth floors. Legs have higher cost, weight, failure rate, and power consumption than wheels.
Better gripping, though, has immediate practical applications in manufacturing and distribution. We're much more likely to see that in use in the near future.
Next month I'll be going on a weeklong backpacking+camping trip. I don't believe in the notion that modern camping is "roughing" it (tents and shoes and cooking gear are essential tech). So on that note, I'd love to have a low-cost solar-rechargable "pack mule" robot that could carry more gear. (I'll survive without it, of course, but it would be nice.) Wheels on a robot wouldn't work where I'm going, but legs would.
That's never going to happen for the same reason there aren't any photosynthesizing animals : you just don't get the required power output.
A human walking uses about 80 Watts [1]. Assuming no losses that requires a 1/3 m2 solar panel (a perfect one, constantly in full sunlight) to just move during the day. So count on needing 2/3 m2 at the very least.
And while robotics have advanced, there is no robot I know of that uses less than 300 or so watts.
I think you're right though that there is a market for very large, very heavy vehicles with legs, for where slopes are either too steep, or the region just doesn't have roads. Such vehicles exist, but humans just can't control 4 legged vehicles. It quickly ends in disaster.
Yup, the military has this exact same use case, I think that's where all the funding for Boston dynamics big dog (and probably others) came from in the first place.
Most buildings have to have affordances for people in wheelchairs and robots can use those. Hopefully in the future there will be standard robot/elevator interfaces but in the meantime we can make due - I spent a lot of time at my last job integrating elevators with our delivery robots.
Humans have an entire mini-brain - the cerebellum - for offloading motor functions. Our brains don't need to learn to walk; instead our brains must learn to train the cerebellum to walk.
>"Some researchers are using machine learning to empower robots to independently identify and work out how to grab objects. Others are improving the hardware..."
Software is the main bottleneck in ML at the moment, imo. That and data. The hardware problem (computation), which the author discusses briefly, has been more or less solved in the industry as newer software require less miniaturization of ICs and peripheral components[0] -- this despite Moore's law effect wearing off. [1]
It's all about striking a fine balance between the hardware's raw computational capability and designing compatible software nimble enough to adapt rapidly to colossal influx of data.
Software is the main bottleneck in ML at the moment, imo. That and data. The hardware problem, which the author discusses briefly, has been more or less solved in the industry
I'm only a casual long time observer, but I think there's an opportunity here for haptic teleoperation. Data could be gathered from haptic teleoperators in much the same way that Tesla is gathering data for self-driving AI. Back in the 90's and early 2000's, Sarcos corp's website used to have a crappy realmedia video of a full-arm haptic rig operating an over-sized hydraulic arm casually holding an anvil like a beer mug. So evidently, we've been able to do full-body haptics and human-level agility manipulators for quite awhile. (Research in that stuff actually started in the 1970's!) On-orbit teleoperation from the ground could have interesting applications.
Interesting how we have designed the world for humans and now the optimal way for many robots to work is to be designed like humans. I always hoped for some more optimal design, maybe even something that can morph.
Well, it's natural for humans to design the world for humans (though a good part of it now is designed for humans inside machines, ie cars).
Note, however, that most robots today actually operate in spaces designed for machines - ie, factories. Getting robots to operate in the "unstructured spaces" the article mentions is extremely difficult and hasn't been cost effective.
A large amount of gains in productivity have been achieved by substituting machines for humans, with "robots" just being the most programmable of machines. But very few of those gains involve direct substitution of machine movement for human movement - rather it's involved imposing structure on the whole productive environment so the rigid motion of a machine has a predictable result.
Also of note, the societal backlash of replacing humans with robots has been pretty massive. So it's not just cost-effectiveness that's being balanced, but care with the workforce not to implement too many replacements too fast. The places where we've seen massive substitution are all industries where humans were in physical danger from the work being done. Risky manufacturing lines, etc.
I feel like that's starting to shift a bit where folks are realizing that more automation by robotics isn't actually reducing the potential workforce, it's just shifting it. Again though, it's up to the employers and manufacturing companies to really drive that point home.
The places where we've seen massive substitution are all industries where humans were in physical danger from the work being done. Risky manufacturing lines, etc.
That doesn't really reflect the situation. Machines have been substituted for humans since at least Eli Whitney's Cotton Gin. Substituting machines for humans has had nothing to do with providing safety, rather it has been driven by the desire for increased productivity and thereby profits. Certainly, this increased has provided massive benefit to society along with various drawbacks. That's literally the history of the "industrial revolution".
No no, you misunderstand me. I'm not saying it's FOR human safety...it's for profit. But the pushback on "losing jobs to machines" is less when the end-result is replacing human jobs that have a high degree of danger associated with them. Which is why, throughout history as well, we've seen larger adoption of automation via machines in those areas. The friction to do that replacement is lower, so we see it more often.
But the pushback on "losing jobs to machines" is less when the end-result is replacing human jobs that have a high degree of danger associated with them.
Well, OK, except for the long history of industrial automation that doesn't bear this out at all. There has resistance to automation from the English Luddites to Detroit Black Workers Union and beyond and arguments that automation is only OK safety is involved have not entered the picture that entire period.
Basically, there's never been a situation where public complaints stopped automation. It has proceeded as fast as technology allowed to the present.
EP Thompson is the reference for early automation and the luddites but it's hard to give a reference for a negative.
If you have any reference for situation where public sentiment limited automation, I'd love to see them. Because it sounds like you are mistakenly thinking the world actually works according to the vague headlines one see sees periodically.
> Here we demonstrate a completely different approach to a universal gripper. Individual fingers are replaced by a single mass of granular material that, when pressed onto a target object, flows around it and conforms to its shape. Upon application of a vacuum the granular material contracts and hardens quickly to pinch and hold the object without requiring sensory feedback. We find that volume changes of less than 0.5% suffice to grip objects reliably and hold them with forces exceeding many times their weight.
That's a cool appendage, but there are still a lot of problems in grasping / planning with materials like fabrics, right? e.g. this doesn't solve the notoriously hard laundry folding problem
Fingernails and Fingerprints. I can't fold laundry after I cut my talons, and I can't fold laundry when the friction is absent.
Training a robot to fold laundry would be like training an astronaut to fold laundry in Apollo era gloves (which btw, I have read were so bad, the lunar walkers lost fingernails inside 'em)
“When you pick something like a pen up off a table, the first thing you touch is the table.”
I stopped reading for a good two minutes of experimentation.
A robot being able to complain that something is very difficult for it (using more energy for something than it needs to for example) would be really cool!
IE: Stand up on two legs, walk forward, and move these 10 objects from this bin to that bin without breaking any of those objects. Or fold these clothes (surprisingly difficult to get a robot to do that).
We can build super-advanced Chess and Go AIs that beat out the grand summation of theoretical human knowledge (with Chess theory going back hundreds of years, and Go theory going back maybe thousands!!), but we still can't move a variety of objects from one bin to another successfully and repeatably.
-----------
Another funny example: AIs (or various algorithms at least) have solved the object camera tracking problem, but still basically fail at depth perception and figuring out if something is "still" or "moving" based on sight alone.