Can We Teach a Robot Hand To Keep Learning?
Вставка
- Опубліковано 4 жов 2024
- ❤️ Check out Linode here and get $20 free credit on your account: www.linode.com...
📝 The paper "Efficient Adaptation for End-to-End Vision-Based Robotic Manipulation" is available here:
sites.google.c...
❤️ Watch these videos in early access on our Patreon page or join us here on UA-cam:
/ twominutepapers
/ @twominutepapers
🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible:
Alex Haro, Alex Paden, Andrew Melnychuk, Angelos Evripiotis, Benji Rabhan, Bruno Mikuš, Bryan Learn, Christian Ahlin, Daniel Hasegan, Eric Haddad, Eric Martel, Javier Bustamante, Lorin Atzberger, Lukas Biewald, Marcin Dukaczewski, Michael Albrecht, Nader S., Owen Campbell-Moore, Owen Skarpness, Rob Rowe, Robin Graham, Steef, Sunil Kim, Taras Bobrovytsky, Thomas Krcmar, Torsten Reil, Tybie Fitzhugh
More info if you would like to appear here: / twominutepapers
Meet and discuss your ideas with other Fellow Scholars on the Two Minute Papers Discord: discordapp.com...
Károly Zsolnai-Fehér's links:
Instagram: / twominutepapers
Twitter: / twominutepapers
Web: cg.tuwien.ac.a...
Hey Dr. Zsolnai-Fehér! I'm an author on the paper you covered here- you might see my arm a few times shuffling objects for the robot :) One thing I'd like to highlight is that reaching the grasping baseline required ~600k real-world grasps, but fine-tuning to each modified setup required only 800 grasps to largely recover and sometimes exceed the original success rate- less than 0.2% of the original training set. Thank you for covering our paper on two-minute papers, the team is super excited to see it talked about here! Keep up the great work fellow scholar!
Excellent, thank you very much for dropping by and for the feedback! Congratulations on the amazing paper. Pinned this for visibility.
600k?!?!
How'd you do that
The feature I find most impressive about this is the flexibility that the continual learning can give to systems and the differences between optimal conditions and modified conditions. This is nothing short of amazing.
But are there limits to continual learning, that if you stray too far from the optimal conditions, the system can't adapt at all? Have these potential failure cases been tested?
Also, if in some cases modified conditions are too great for the system to adapt, would it be possible to put in intermediate steps that would bridge the gap between two conditions that are too far apart so the system could eventually get there? Cross the river on stepping stones instead of all at once?
:D
Your name must be Harry, because you are a wizard!
0:22 Plush giraffe perturbation strikes again
tbh I wouldn't mind that much if I was perturbed by a plush giraffe
Giraffe on the outside, hockey stick inside. They are all the same just like boston dynamics, evil robots abusers!!
Thought it was an orange xbox controller
@@OrangeC7 me too, that sounds lovely
The future Skynet will hate giraffes instead of humans. Nice safety touch!
The plush giraffe thing is adorable :D They could have used a stick to poke it, they could have used a rolled newspaper. And they chose to use a giraffe. I love scientist mentality :D
So, you know how in physics simulations self-learning AI often finds ways to exploit the physics engine? How long until we see that happen in real life? Maybe one day we'll built a self-improving RC car, only for it to suddenly wave its spoiler in a weird way, then disappearing in a puff of light.
Nah, it would need about 1.21 Gigawatts in order to do that, no way an RC car battery has this much power ^^
@@arantes6 Gotta make sure that flux capacitor is working properly too!
Warp drives can't come soon enough. If that happens, I'll let the AI do the whatever f it wants, just maybe a lightyear away in space.
"Hey G.R.E.G what are you doin- WHAT IS GOING ON HERE"
_loud metal smashing sounds are heard while robot shakes violently as its rapidly collides with the wall its clipping through_
@nasolemmuch terminator/matrix lately?
...”learning to walk in different ways relying on different body parts,” in other words, play; you tell it to play.
Super interesting! I've been waiting for more of this continuous adaptation learning for real-world applications. Consider self-driving cars, or more relevantly to the grabber, construction machines. Since the lighting outdoors changes constantly and they have to operate in different times of day/lighting and lift objects of varying weights and sizes, these machines need to adapt in real-time. Glad to see them tackling it!
The problem then is to ensure that it cannot become worse than before.
How to ensure security with a changing algorithm
@@etiennedud Well doesnt it adapt until reaching good enough success rate and then bake it in with no more learning? Theyd still need to be supervised anyway.
If done correctly I'm guessing this would also help with overfitting? Can anyone confirm?
@@etiennedud I guess you'd have to make sure the changes are always safe lol
@@Supreme_Lobster and this is the difficult part, because you cannot prove it as you don't know what it is doing.
Don't get me wrong, it is a marvel of a research. It just need one or two paper before being usable for real time application 😄
Nice that the overview of the paper ended at exactly four minutes; time to change the name to four minute papers? haha
Play at 2x speed
FewMinute papers.
I'm feeling pain for the robot right now when it fails at grabbing something
He learns from his mistakes.
Dont worry he will improve and in the future he will be rearanging humans skull on his fireplace
@@Dadax9398 fire place?? Why
I recently saw an Interview Lex Fridman did with Ilya Sutskever, the co-founder of OpenAI. The robot arm experiment was mentioned, and Lex wasn’t aware of its existence and the progress in simulation learning knowledge to real world transfer. This signifies two things for me: First how even someone as knowledgeable as Lex who worked in the field doesn’t have a complete overview over the vast, rapidly expanding and progressing world of AI and second the value of this channel for me, who knew about the experiment from watching it.
There so much going on at the same time, it's hard to keep track even for very smart people..
1:10 I can deeply relate to the bottom left robot.
We shouldn't (first..) teach them how to keep on going once you shoot one of their legs off.
Oh hell no😶
Love that you're doing real-world applications as well!
It's kinda cute that it tried to grasp the robot on the other side of the mirror.
Another great video! Thank you so much for sharing these papers and bringing knowledge of these kinds of things to the public!
You are very kind. Thank you so much! 🙏
I began to find about Open Ai's rubix cube solving robot just about 2 hours ago and 2Minute papers upload a viedio related to that.
What a time to be alive!!😄😄😄
I wonder how good can be if it tries to pick up stuff on an underwater surface
On the topic of teaching neural networks to continue learning, one thing I've really wanted to see someone try is apply the principles of the Leitner Box and Spaced Repetition to a neural network. I won't go super deep into the details because you can learn all about them elsewhere, but it's basically a model we use to help graduate information from short term human memory to long term memory. By spacing out review session in a certain way, we can allow ourselves to consistently introduce new information while maintaining and developing a stronger hold on older information. Each time you review old information and get it right, you promote its level and it will therefore be reviewed less frequently. If you get it wrong, that topic has to back to level one and will be covered much more frequently until you level it back up.
I think this principle could be applied to neural networks. Get the AI really good at one task and set a condition for completing that task (for example, use a pair of hands to kick a paper football though a goal). Once the network gets to a certain desired success rate at that task, introduce a new goal (perhaps, fold a sheet of paper in half). Have it keep working at it until it becomes reasonably successful, add that to the spaced repetition system.
Once we have enough tasks in the system, we can run the memory retention system. For each level we are working on, test each task. If it gets, say 90/100 attempts correct, promote that task. If it fails to meet that, send it back to level one where it will be reviewed much more frequently.
Over time, if this model work, the mastered tasks will reach higher and higher levels, where they're gotten to the point where they won't be forgotten. As the program continues, you can add more and more goals, potentially making the network more general and capable.
what a time to be alive
I swear, it is things like this that have made me wonder if there isn't a "collective subconscious" to the entire human race. This is such a "missing piece" of a project I'm working on, and just talking about the paper has given me some new approaches to my work.
Sometimes I really love parts of the times we live in, and they are almost enough...
What you workin on? Can you talk about it?
there are only selfish genes, things just turn out this way.
@@bajsbrev4651 But the spontaneous parallelism is amazing, because many of us seem to be working on similar things that can heavily and helpfully influence each other.
We are more than the sum of our parts, is true on so many levels.
@@Supreme_Lobster I am working on problem solving from a different angle. Something said in this video made me look at the paper, and realize this learning method can be applied to my project.
I think autists are good for your research. They miss cues that other people integrate subconsciously!
One of them for example is "Do not end with sentence with a dot in a chat, except if there is another sentence following it".
The untold rule: "Dot means seriousness when not separating sentences".
So, you could lay down these implicit rules and observe autists behaviour around them.
If you can get hold of aspergers, you can ask them question about these rules.
The byproduct would be that rule book that could be given to some autists for a better life.
I'm really waiting for the next big thing from Boston Dynamics.
I can't even solve a rubiks cube with two functioning hands and a full brain xD
xD
But in all seriousness, in that paper, solving the Rubiks cube was the trivial part. The mindblowing part is how it actually know how to move the fingers in order to achieve the desired result. The cube solving part is done using traditional, preexisting algorithms :)
just wait until you loose one and suddenly you will be able too! What a time to be alive!
4:00
When the AI rules the world and kill humanity:
"What a time to be dead"
When Sophia starts breaking legs just to see us adapt.
Most useful youtube contents for researchers as always.
A very cool concept worth looking into is teaching an AI to teach itself
Currently, a human operator gives it relevant input and checks the output.
Surely there's a way to intelligently allow the AI to draw input from the internet and check it's output against known output at first, then not checking at all when researching new stuff.
Take electric circuits for instance (with losses just like in real life).
The AI would take the input data for every component and predict some outcome.
if it's predicted outcome is accurate to the real life testing, that means it learnt to do that.
Instead of giving it circuit schemes and formulas and whatnot, we could teach it to get those from the internet and it could learn to predict outcome all by itself by training continuously.
It would likely need a second AI component that gathers relevant data from the internet
I don't know about the rest of you but I find something quite charming about the arm trying to grasp its reflection.
I wonder if something like this could be used to make more useful prosthetic hands
That would be really cool!
Even a prosthetic limb which could learn from your brain's impulses to produce a natural feeling limb would be incredible.
@@bobparker1671 I think they did something like that. They made it as a exoskeleton if I remember correctly. They put a small chip inside the brain and let a machine decipher and learn the brain impulses.
@@exosproudmamabear558 That sounds neat, but that requires some sort of surgery or the injection of a chip. I was thinking more a "plug-and-play" sort of device. Attach the impulse sensors to the limb, and it learns from there.
@@bobparker1671 I am not sure if you can read brain activities from the limb itself. It doesn't sound plausible to me.
I have a new amp and speakers and your intro sounds amazing!
I like it when the videos are more than "Two minutes"
Sorry if i missed anything, but was there anything said about HOW they acchieved "continous learning" here?
Nah you will have to read the paper for that.
The way he says *next tiiime* 5:13
aw the reflection thing is like when a baby wants to touch themselves in the mirror
im still waiting for someone to try to teach a neural network to train other neural networks
that is currently happening. Search for Meta-Learning or Neural Architecture Search
2:04 Nice grinder
Do these paper authors always write their own physics simulation and their own renderer or are they using something off the shelf?
This is different from normal fine-tuning, experience replay, and online learning, because? It's not even a combination of said techniques.
You let your little secret slip at the end. I will call the Witchfinder General.
Yes; go my fellow robots!
What i really want to know is where to access these papers online??
he always puts links in the description, but alternatively see here: arxiv.org/
Thanks !
I suspect ML will soon enough find an algorithm that can capture HD images just like how an extremely powerful telescope can, but through an ultra thin film within the casing, the entire visual wavelengths of light.
1:12 notice the one on the bottom left is "cheating" by walking on its ankles, to avoid using legs.
TMP covered that behaviour in a previous video 😊
Intelligence is feedback a on-demand loop connected to neurons which are turned on and off according to the feedback received by one's awareness. Meaning that any task is only performed when necessary. Hence, you won't drink water unless you feel thirsty or planing ahead of time (preparing) for something. "Feedback" is a set of complied awareness. And awareness is continuous set of environmental input received constantly when exposed eg, temperature, pressure, light, bodily functions, etc.. And Consciousness is ...sets of awareness tasked to perform different task working together to create something called Intelligent Sustained Awareness. (I.S.A ) hence, different parts of the brain and hence Intelligent beings anywhere from a tiny ant to humans.
One more thing. Human Intelligence is a Type which envelopes itself around a mandatory routine we know as "expectation". Never mind. I hope it helps create a complex algorithm to train a Virtual Intelligence by combining group of neural networks working together as one.
Looking 4ward to video games where damaged robots still limp after you.
2:24 literaly my cat saw my reflection on the wall (shadow) and also desperately tryed to grasp it. not to far away from this. (why we animals keep such "not inteligent" behaviours after long being evolving..)
Just a minute. 3:35. After the model is inserted,
can it continue learning?
Isn't the training phase very different from the
inference phase?
Don't we have different hardware and also
different architectures for the two?
What do you use to render these
Hey, sounding good, is that a new recording setup?
I felt bad about them being stuck and us not being sure if they are concious enough to have rights, but then I remembered they might be doing the same in the future lol
Can someone explain to me how the robots are being trained?
Love your videos!
Hello K.Z.F. ! I know that 2 QUANTUM algorithms for graphics exist. Polygon visibility algo and Global illumination algo. Quantum is the future, so maybe you can make a video about quantum cgi !!!
how do you simulate the robots and generate g_code?
My Friends : "how can you keep up with the latest technology innovation papers?"
Me : "i watch youtube, especially this channel"
Plush giraffe perturbation, a song by Fallout Boy
Should we?
We just keep feeding the monster huh?
You guys have seen Terminator films right?
We're done boys. Cyborgs soon, hopefully for us too :)
Just wondering, are you Hungarian?
Extremely cool!
Yes we can, what a time to be alive!
I hope NASA can get past sending wheeled rovers to Mars. It seems to me that intelligent spider-like walking would be far better suited to the irregular terrain. Remember, when the Martians invaded us in War of the Worlds, they used legs and not wheels. H.G. Wells had it right well over a hundred years ago!
Hmm are we training future killer robots how to strangle us?
So we have about an hour until the robots stop shooting at their own reflections and learn to target people better.
But can we teach it to love?
Of course!
I'm just waiting for the AI that learn how to program, it will be great, could really be a real expression of intelligence
Why is there a giraffe pushing the rubix cube
Skynet is closer and closer
Now image an entire robot that can keep learning!
E Core ??? Is this a bot?
@@ToriKo_ Not a robot, I am not. Not robot. Promising...
Uhhuh.
Am i only one thinking, that the voice of the speaker sound kinda unnatural?
it's kind of hard to emote in a different language
"What a time to be alive" with robots that can improve themselves that may or may not take over humanity.
(just a joke ppl)
Nope, no jokes allowed here. Now, get to the Gulag!
"(just a joke ppl)" sounds like something the future robot overlord would say
(just a joke ppl)
Im not saying this is terrifying, but this is pretty terrifying
I have a marginally related question : does anyone know what algorithm a CNC machine uses to generate a toolpath such that it cuts out a desired shape out of a material. Is it gradient descent? I am trying to write my own CAM sofrware.
lol at title
I keep telling you, I'm not a fellow scholar, you're messing me for someone.
f e l l o w s c h o l a r s
bro did you just graduate?
Hold onto your papers. Is this an inside joke? I don't get it.
Notification Gang!
: ) nice
Lit
How many Terminator Movies should they make to imply that AI "might be" (just maybe, but surely it is) the Doom of humanity ???
*Please tell me dear damned fellow scholars?*
11th
Science has gone too far
Give it a gun.
Me, reading the title of this video: "Please let the answer be yes, please let the answer be yes..."
Indeed, Betteridge's law does not hold too well around here!
@@TwoMinutePapers I grew up watching the 80s movie Wargames, and the line "it's learning how to learn" always hits like a sledgehammer. And now we're seeing it happen. :-D