A Robot Teaches Itself How to Walk
HTML-код
- Опубликовано: 14 фев 2012
- Cornell University professor Hod Lipson demonstrates how a robot can teach itself to walk without any knowledge of its form and function. "Within a relatively small number of these babbling actions, it will figure out what it looks like," Lipson says. He adds that eventually "it can figure out how to move."
----
Join an audience at swissnex San Francisco as scientists from Switzerland and the US discuss their research on humanoid robots, cognitive robotics, and artificial intelligence (AI). Hear how some robots self-reflect, self-improve, and adapt to new circumstances, and whether it's possible for robots of the future to possess the same cognitive characteristics as humans.
Cornell University's Hod Lipson is seeking to understand if machines can learn analytical laws automatically. For centuries, scientists have attempted to identify and document analytical laws underlying physical phenomena in nature. Despite the prevalence of computing power, the process of finding natural laws and their corresponding equations has resisted automation. Lipson has developed machines that take in information about their environment and discover natural laws all on their own, even learning to walk.
Rolf Pfeifer directs the Artificial Intelligence Lab at the University of Zurich. Together with his scientific assistant Pascal Kaufmann, Pfeifer presents current AI research and a humanoid robot in the Ecce family referred to as Cronos. Instead of copying only the outward form of a human, Cronos mimics the inner structures as well-bones, joints, muscles, and tendons-and thus has more human-like actions and interactions in the world. 
Наука
It would be cool to see different form robots learning from scratch with the exact same software :)
it would be great if it really looked more like a spider.
Absolutely not my point but it would be fun.
It would be great to connect all future robots to my Robot walking school API, that controls all robots with one piece of software similar to this example. Just so I can take over the world while making others pay for it.
Spiders have eight legs. They could have started with that. They can't expect a creature with four legs to move lika spider, right?
ofah222 i don´t mean the number of legs.
I did a similar thing with a simulated construction that was trying to move using a darwinistic approach. It was sorta successful. Every time it found a new way to utilize bugs in my physics model and propel itself forward with great speed. Was fun though! :)
this is amazing. so you could load one pice of generic code onto any robot and it would learn without having to be configured. Also, if the robot loses a leg or something, it could re-learn how to move, compensating for the loss!
This is fake. It was programmed with clumsy and fuzzy code and then eventually actually walks.
Jared Legitlastname u wut mate
Jared Legitlastname no
A bit harder than that
@@JaredLegitlastname no
it's absolutely amazing how people have figured out how to make computers develop a sense of self
This video needs to be shown to all those who say that science cant explain self emergence. I bet theyll turn to their graves !! way to go science , save us from the evil thats religion
I don't think religion is 'evil' necessarily (that would be too ironic!) but I do think that if one relies on their religion to the point where they ignore reason and repeatably-experimentally-accessible knowledge, they are severely limiting themselves.
A machine doesn't have self awareness.
Félix Bachiller Maybe not in the sense that conscious animals and humans do, but computationally, yes. A machine can possess self-awareness if you define self-awareness as the ability to create and update an internal model of one's state (location, orientation, surroundings, battery level, which motor does what, etc.).
***** That's something science can't explain. I haven't seen a machine that commits suicide because it has an existential crisis. That's antinatural and we humans do it.
i can imagine this thing breaking a limb or something and then developing a different gait to fix its problem or something. Kinda like when you amputate both legs in a person for fun, and it starts crawling miserably across the ground with its hands towards the nearest hospital. Weird, right?!
just hope it doesn't happen to you..
I love your psychopathic comment.
What is missing from the discussion is the goal that the robot must have had. He discusses walking as the goal. I'm curious how walking was a goal was introduced into the bot.
You set an expected output as movement. Have some kind of sensor to tell if it is moving. The closer it gets to that out come the closer it get to walking.
Maybe there was a sensor that tells the program "the more you end from the center, the better."
probably change in location based upon it's thoughts. the thing only have an orientation censor, it wouldn't know if ti was really moving.
now toss it in water and let it teach itself how to swim
no throw it out a plane and let it teach itself how to skydive
Sentic Alien That thing better teach itself how to use a parachute first.
not enough trial time
Wasn't pre-programmed to do that
Invigorate It wasn't pre-programmed to walk either. It taught its self how to walk.
i suddenly have 8 knobs... things are looking up apparantly
One day they are gonna teach themselves how to kill people, upgrade themselves and so on. Lol.
Edit: nothing against this or ML in general of course, it's awesome.
Well yes of course, but that's 'natural' to do to become dominant
10 years later an a research group just acted like they were the first to try this in history.
Amazing! This is huge step towards A.I. Self awareness is a big part of consciousness I think. Keep them coming.
Where is the original video?
This is a wild extrapolation, but to me, this is, In a very loose vague sense, an emergence of "consciousness".
+orcodrilo technically i'm pretty sure this won't collapse the wave function... not even sure if it would do that if the parameters were modulated at quantum random
+orcodrilo Not "consciousness", but "self-awareness", by definition.
Well what is going on in this robot is not much different from the calculations performed in 3d tracking or photogrammetry, so if this robot is conscious then so is my vfx workstation
Ben Hinman Not "conscious"! Self-aware!
When sometimes takes itself, it's own characteristics, into account when solving a problem, it is by definitionn self-aware.
We are self-aware, and it allows us to understand that we have an impact on the problems we try to solve. Our own presence has an effect on most situations. Tehrefore, we need to be aware of this, and take this into account. That is self-awareness.
This robot is doing the exact same thing: It is learning "what it is", to be able to solve a problem.
To be "conscious" is something that brains do. It's a specific kind of brain pattern. Robots can't do that, because they don't have brains.
At the heart of any machine learning like this is an artificial neural network, and what is a brain if not just a very elaborate neural network?
Looks pretty awesome when it moves. I, for one, think it looks creepier than a spider walk.
Very cool indeed. What hardware are they using?
tell the green cube, not to touch the ground
this is terrifiying and so much inspiring! how did they programm it, what libraries and languages did they use?
A good program has many complex layers working in isolation, only passing key data from one part to another.
There will be the parts of the program which will have the purpose of moving in a given direction, while other parts will be in control of analysing how its body works/moves, which will then pass its results onto the parts of the program which have the goal of moving from A to B. This code will work with any set up and could easily be modified to have a different goal such as 'jumping'.
is the program availible for use?
Very slick.
Hello from last podcast on the left!
Robot teaches itself to delete emails.
its like infancy. to crawl before you can walk. the first clasps of sporadic movement.raw un biased.just beautiful. truly warming to see evolutions engineering. in the form of mans collected intelligence and to what his hands can conjure. and create. well done guys.
Full video can be seen here: library.fora.tv/2012/01/19/Cognitive_Robotics_and_Artificial_Intelligence
Simon Keisala thanks
Soo... A learning robot? We are on our way to Terminator aren't we.
+Genesis Fox Why do you think they're gonna be evil ?! The robots that people are trying to model are just designed to think rationally in every problem so they'd do the right thing whatsoever!
Cause, this is a spider robot looking thing,... that's scary af.
+Genesis Fox spiders are mostly scary because people think they can kill you. This robot isn't really scary because you know its not going to hurt you
ThoughtfulScience I'm not afraid of spiders, other people are, I guessed they were just creepy, so.. thanks for informing me :)
Genesis Fox lol
What program did they used?
If there was a robot army that was self-aware and moved around like that, I'd blow my brains out without a fight.
@frankvonfrauner This wasn't clear from the clip, but think the one thing it does have initially is the motivation to move from point A to point B and the rudimentary logic to engage in problem-solving to accomplish locomotion.
@BlackDrakkAr I'd say simply moving in any direction is a lot more simple than jumping or spinning, as both of those actions require smaller movements into directions themselves.
is there a how to build instructions somewhere?
I am very interested to this for my school
People are laughing at it because it couldn't walk very well but let's not forget how long it takes for a human to learn how to walk.
Is there a specific research paper linked to this?
Now if only we could have this kind of technology on chatbots...
well now we have
yes we do now
I read the discussion. I listened to and watched the video.
Yes, "we" is mankind. That is who builds robots. So yes, this is about robots but it's also about people if you want to get into what the robots are going to be programmed to be capable of.
"We," as in some of us, cannot necessarily control what "we," as in others of us, are going to do with the things.
So it certainly is not impossible for robots to turn (or be turned) against "us." In fact, some of them already are capable of it.
Of course, it "knows" concept of moving. But "he" has no sensors to to measure it - the only sensors are tilt sensors. So, he first learns his own morphology (what is connected to what) and then, using derived virtual model in simulated world, generates gait which will move "him" in certain direction. Then this gait is applied to physical robot - and it works, because real world and simulation are close enough.
What gives him urge to move in one direction is fitness function in step II.
I'm tempted to implement something like this but i'm still trying to figure out how to map joint mechanics in a 3D simulation to a real servo or stepper motor.
Kinda like a newborn baby, learning it's limbs, eyes etc. Starts out with random movements and over about a year they can walk and more
amazing....I've been working on similar concepts that will allow robots to self evolve generically and in any field.
It kinda looks like the way starfish move.
fascinating
that is really incredible. it 'created' a motion never seen on the face of the earth in all human history. THAT is really significant (if it's real and not a pre-programed response).
Impresive , iris really amacing
Amazing.
my wife, who fears that self aware robots will eventually be exploited, made an interesting observation the other day. She said that she does not believe that self awareness is something can be programmed, but rather is something which is developed naturally. With self-image awareness, how much longer will it be before robots can pass mirror awareness tests which we use to determine self awareness in biology? I think currently they could.
The machine actually starts off with a lot of information about itself: it knows it is one of the 4^8 possible "planar arrangements of 8 limbs". It's possible that there's no probabilistic inference behind this algorithm, it could simply be brute force trying all hypotheses.
well considering that there is 65,000 possibilities based on your math, and it could move fairly well by the 64th time, it is not brute-forcing it.
Aaaaaaaaaaaaaaaaand.... The robot grew up , he's gonna show up in the next Transformers movie...
I have reason to believe that, during the first few moments it was on, it was experiencing sheer terror and confusion, as evidenced by the frantic calculations and the awkward writhing. I would imagine it was thinking something along the lines of: "OH GOD WHAT AM I???????!!!!!!!!!!!"
That is Awesome!
That's really cool.
This would be cool to apply to games like spore to have a creature adapt itself to unknown legs.
fractal math and fractal geometry makes me happy. Maybe it will help us to describe some real life phenomena.. someday.
YOU ARE GOING TO WALK LIKE A SPIDER!
No, dad, I want to crawl on my belly!
I HAVE NO SON.
Pretty cool! It still has the concept of legs and motion built into it though, so it hasn't completely worked out everything on its own :P Although there's a pretty decent case (Kant, Chomsky) that says humans have concepts like motion, shape, space, built into their heads already, so the robot isn't too dissimilar!
"this is what robot feels" oh shit its happening SKYNET is opening and robots teach themself shooting hoomans
why don't you put the link simply in the description.
Ten seconds after it started moving it killed (unplugged) itself lol
As someone who works in machine learning, this is not necessarily the case. If we ever do create truly cognisant artificial life, it is completely possible that they will eventually come to realize that they are nothing more than experiments created for our pleasure. That's not to say I agree with the person you replied to, just meaning what he is talking about is a possibility, however remote.
Exelent!.
KILL IT BEFORE IT ACTIVATES SKYNET
Because it's walking cycle simulation is programmed to find a way to make it move. The farther it moves, the higher the "score" is for a specific walking cycle, and the best walking cycle is then chosen.
Looks like something straight out of a alien movie… featuring robots tho. 👀
Wow, this is so awesome.
But i have to wonder, what way of moving it would develop if the last stage that he developed, wasn't actually last (the AI would be set not to find out ANY way of moving, but to find out MOST efficient way of moving).
It looks like the robot is slithering! But I vertically :o Doing the worm? Yesss
If we ever have microscopic robots capable of working together to form shapes, this would be an extremely useful way of letting the "new robot" know how to move and such.
It always boils down to a personal believe system. Noone knows for sure whether something is infinite or not. I (again-personally) can not imagine or envision a finite universe. If you can, then let it be so.
The "walking" part is terrifying..
robot: aargh why would you do this! im in paain! (drops off cliff)
This means that the robot is self-aware, right? Something that takes its own characteristics into account when solving a problem is by definition self-aware, as I see it.
+antiHUMANDesigns Not really. It is just testing if this works. If it works, test it. If it still works, try move the legs in a way that tilts it forward. It isn't aware it is just doing true or false test like any other computer
TheGamerPlayz Just like our brains do, aswell.
No, by definition, if you consider yourself as part of the problem you're trying to solve, then you *are* self-aware. Or rather, the algorithm is self-aware.
It doesn't really mean anything, it's just a term.
Lots of animals are not self-aware, actually. They fail self-awareness tests we give them.
Yes, but the robot is not thinking. It was not programmed to walk, but it was programmed to test the different walking possibilities then test if they are efficient, as in going the farthest the simplest way.
TheGamerPlayz Of course it's thinking. It's got a neural network as a brain, very similar to our brains. Our brains also think by sending binary signals through nerves.
The question is not whether it's thinking, because it definitely is. The question is whether it's "self-aware".
*If* it is true that to solve this problem, the things needs ot be aware of its own characteristic, then to solve hte problem, it would need to be self-aware.
However, one could argue that it only needs to be aware of its body parts, and therefore the problem may not actually requires self-awareness. It's be like being aware of your shoes.
I can accept that.
There is a much better example of robot/AI self-awareness, which has even been recognizes as such:
www.iflscience.com/rise-machines-robot-demonstrates-self-awareness-solving-logic-puzzle
antiHUMANDesigns Ok lets settle this. It is not self-aware. To be self aware it needs to know that it is different from its enviroment. The robot doesn't even know its enviroment, what it looks like or even how many legs it has. It cannot think thoughts or even act in ways that it was not programmed to do. It does not have a neural network at all. It only has wires. It cannot feel anything. It is only using pre-programmed commands in certain circustances that it is programmed to detect.
If you guys want to see the full conference
here is the link:
library.fora.tv/2012/01/19/Cognitive_Robotics_and_Artificial_Intelligence
Its very interesting!!!
1:57 my video froze at him flippin the bird
I except that once it has decided that it knows its morphology well enough there is a second optimisation algorithm which searches for the correct way to move the motors (using the morphology model), in order to produce walking motion, optimising some movement criteria designed by the researchers.
this would be a game changer in military robotics 20 years down the line.
this way you can attach other self learning robots of different species like an arm or a head or another body type/part.
@EnchantedWalnut He isn't trolling he is just pissed off with people saying "thank god".
So you thank god for my response.
How do you make the robot not give up on hard parts? Like if there was a cliff for this blind robot and it COULD get over it... but only if it somehow rewards itself by learning how to do a pull up. It would need a "search and discover self" mode that made it so it didnt only go to easy parts of navigation that somehow get rewarded when it feels accomplished in a rare, tough situation.
Level up, now to the next stage, awareness of the terrain. After that it would be pretty good to make it search for energy sources. Next we have to make it train self defense against possible predators.
that movement reminds me of a starfish actually. a logical and fairly efficient movement pattern for something with 4 legs in that pattern.
What if you made an algorithm that the robot tried to figure out the least amount of steps to move a certain distance and this way it would tweak its movements to try to be the most efficient? Like the least amount of effort to go 30cm would be a certain way of moving and that would be the way it moved, assuming the way the maker designed it to move would be the most efficient way for the way it was built?
@BlackDrakkAr
It can probably sense if it's moving through accelerometers or something.
Then moving in the right direction would probably just be a result from sensing the right input from the accelerometers.
Great.
Now teach it to fetch me a beer and I'm sold.
I'm impressed.
I watched the other video on ted and on both of these the robot walks the same. But if you notice the legs dont have the ability to move forward, only up and down so there would be no way for it to learn to move one leg forward in front of another. Another idea for them to learn is to input a model of another robot already knowing how to walk. Like a child would see an adult walk. Either that or give it a video input with learning ability coming from the camera and then let it watch an animal walk or a person walk and see if it learns from what it sees.
Wait did he say what the robot feels? You can't make a robot with feeling, because if it gets destroyed its over but the robot feels like its human, thinking he's going to heaven
Such is the plight of humanity, just a bunch of robots thinking we're going to heaven.
Dogs do that movement to clean their ass after taking a dump :D
It there an algorithm for discomfort ? Would this not make the robot walk better? That movement seems quite natural minus discomfort and even exhaustion.
That's actually quite an interesting comment, and it did make me wonder for a bit. Personally I feel that it's the same way babies learn to walk. They recognise that there is a destination (from a to b), then they try to move and stabilise their boies through body mechnanics. The robot doesnt spin for the same reason why I think babies dont spin and jump while moving, because it would require more balance and stability. Not sure though, just my 2 cents :p
5:15 after I saw this, I blocked FORA.tv
Considering it has four symmetrical hinge legs, it's pretty hard to imagine this thing walking like a spider. Also, how does it know the amount of grasp, or friction each limb has with the ground? Like, what if you put it on ice? Does it develop a different gait or what?
yeah it has no way to move its legs badk and vforth jsut u n down
I understand that, It would be truly amazing if residing in the 'black box' it would have to understand the 'knobs' too. Now it has a headstart. But it's a step in the right direction nevertheless ;)
The movement of the robot is creepy...
in the very begining its like "Omg, i am alive!" and than "OH GOD! THE PAIN!!!!"
That's awesome. :)
Actually, they say strings are packets of energy that only exhibit some frequency that causes it to adopt the properties of the elementary particles. Clearly, if it is made of energy, than energy is the smallest known unit. If string theory is correct, it only shows us something that everything is made of. Remember that basic law of geometry that says there are an infinite number of points on any given line segment? Anyway, we still need to identify what the dimensions are.
its like a baby trying to walk
Complicated as in someone or some very good and talented team, who have a good inside on their specific target.
You won't be laughing when it teaches itself how to operate a machine gun.
The way the presenter talks about the self-modeling of the robot motion in order to go forward make me infer that somehow the robot is programmed to construct such models and therefore, you can make an analogy in that the robot is like an animal born with the instinct of walking (transporting the body in a direction). Therefore, since it ''feels'' the ''urge'' to walk, it's basically an artificial intelligence with the IQ of let's say something like an ant, which is remarkable indeed.
They should make it evolve, random mutations in its movement, and learn how to walk the most effective way.
I never said anything about control. But yes, there are military robots starting to show up, and their's probably going to be at least one with some form of mounted machine gun, and it could probably be hacked. I don't think it would last as long to, say, get into a heavily populated area and gun everyone nearby down. Probably only shoots enemy freaquencies too.
this is one of the best FORA videos.
in most cases i find theyre just a tad too short, i know theyre snippets made to entice us to go see the full thing, but lotsa times they quit mid idea and such. kinda sucks
to improve that weird walk it just has to evolve the walk now, every step need to be a bi different that the last and test if it was improvement or not