DeepMind’s New Robots: An AI Revolution!

❤️ Check out Lambda here and sign up for their GPU Cloud: lambdalabs.com/papers
📝 The paper "Learning agile soccer skills for a bipedal robot with deep reinforcement learning" is available here:
sites.google.com/view/op3-soccer
www.science.org/doi/10.1126/s...
📝 My paper on simulations that look almost like reality is available for free here:
rdcu.be/cWPfD
Or this is the orig. Nature Physics link with clickable citations:
www.nature.com/articles/s4156...
🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible:
Alex Balfanz, Alex Haro, B Shang, Benji Rabhan, Gaston Ingaramo, Gordon Child, John Le, Kyle Davis, Lukas Biewald, Martin, Michael Albrecht, Michael Tedder, Owen Skarpness, Richard Sundvall, Taras Bobrovytsky, Ted Johnson, Thomas Krcmar, Tybie Fitzhugh, Ueli Gallizzi.
If you wish to appear here or pick up other perks, click here: / twominutepapers
Thumbnail background design: Felícia Zsolnai-Fehér - felicia.hu
Károly Zsolnai-Fehér's research works: cg.tuwien.ac.at/~zsolnai/
Twitter: / twominutepapers

[@FaffyWaffles]

I love this lil guys flailing around and getting back up. Its like watching a toddler learning to run before it can walk.

[@torarinvik4920]

That's what I thought too. It reminded me of my sister's 1.5 year old.

[@byrnemeister2008]

Reminds me of the uk second division of football. Brutal. Hard contact. Referee not really brave enough to send off the home teams defence.

[@KryyssTV]

Toddlers can't be given an assault rifle and become expert marksmen within a day. Don't let these cute demonstrations distract you from the long term goals of these projects. Let's not forget our own history. Einstien's research was weaponised into the atomic bomb, research at John Hopkins into developing a global positioning system was weaponised into ICBM technology and so on. No matter how benign research may seem, all it requires is for a government to see military applications and then swoop in to claim the technology for themselves in the name of national security.

[@electric7309]

​@@KryyssTV 100%

[@joshcryer]

@@KryyssTV technology is inherently neutral, it can be used for good or for bad. A needle can give someone with diabetes insulin or it can give a heroin addict an OD. These robots could be murder bots on the battle field or they could be savior bots for swat teams against a hostage situation. It is all in how you view the technology. Generally though they are going to just put your groceries away and (a variant of them) assemble your products.

[@kevinfontanari]

At 2:35 you can clearly see the robot dabbing at the ball, truly remarkable

[@LabGecko]

🤣

[@ivoryas1696]

@kevinfontanari "Weak opponent, termination if agent fails" 😭

[@Victor-oy8bj]

innovation that excites

[@brunoballak]

loved the lady attacking the bot

[@echoproductions098]

Imagine putting on your resume that your job was repeatedly pushing down robotic five-year-olds

[@jimmykrochmalska3501]

@@echoproductions098 I need that job

[@serenityenderson]

She'll be one of the first to meet her demise when the robot revolution happens... 🤣

Looks like she enjoyed

[@andrasbiro3007]

I wouldn't be her when the robots take over.

[@mhm6421]

3:42 That woman making the robots fall is on the top 10 list of robots (we all know which list it is)

[@ProtoAlpha]

Chat GPT: "noted"

[@syrus3k]

Whilst boston dynamics robots look amazing, what they really need is this kind of learning so they can be fully autonomous.

[@Archanfel]

Boston Dynamics robots + Advanced AI = Skynet

[@Relish19]

I believe they do have this type of learning with nvidia’s project groot

[@joshcryer]

BD absolutely is using this technology right now. The movements are too smooth and organic to not be using them. Programmatic kinematics is slow and unforgiving. AI gives you a lot of freedom.

[@syrus3k]

@@joshcryer yeah but only in terms of assisting remote control. As far as we've seen anyway. What I'm excited about is when it's more like you just give it a prompt like chatgpt... Hey Robbie, go and make me a coffee. For example

[@syrus3k]

Or even before that.. go and run to this waypoint.. Although maybe they already do that one

[@BetoVickers]

3:40 you can see how the lady feels uncomfortable while trying to throw the robot off balance 🤣

[@AberrantAberrant]

This was my favorite part of the video. She can't help but feel bad!

[@SB-xt5jk]

@@AberrantAberrant The AI will remember her after the singularity.

[@SchiwiM]

it's like pushing over a toddler xD

[@duffyscottc]

This is so cool! I'm really amazed at how they're training real-world robots in simulated environments to get more experience than they could ever achieve in the real world. The future is amazing

[@ahegpbtrftcotu]

just like the simulations

[@whynow7035]

Like the matrix but for robots this crazy

[@zeanolafboyopos1409]

We see them as harmless and cute little robots for now.

[@whome9842]

Then we train robots made to look like attractive people for "entertainment".

[@Chef_PC]

Then we train them in martial arts, and watch them create their own fighting style.

[@andrewpepperoni197]

Then we create milions of them, and distribute them around the world.

[@AfonsodelCB]

then we die (skipping a few unnecessary steps)

[@zalox12]

​@@AfonsodelCByou ruined a beautiful chain comment.

[@electrocademyofficial893]

So cute the way they scuttle around.. ....We're all going to die aren't we?

[@raymond_luxury_yacht]

Rotflmao

[@carlpanzram7081]

I don't think so. It's going to take much longer until they become viable killing machines. They might learn to move around and manipulate their environment to a certain degree, but they will 1. lack the intelligence to figure out solutions to novel problems in the fly and 2. Are easy to be disabled by small dysfunctions. So, if you had a humanoid soldier robot, you could probably halt it by either throwing a bunch of rope/wire on them or block their path and vision by unconventional obstacles, or you could simply damage any one of their motor units just slightly for them to completely stop Functioning. Eventually they will be near unstoppable, but for the next few decades they wont be a real danger. Now, swarms of autonomous flying drones however will be absolutely horrifying. Flying is much easier than walking, so when we will use robots in war, at first they are going to be drones. Actually I'm pretty sure we are already doing that.

[@The_Questionaut]

​@@carlpanzram7081I think I remember Marines using the metal gear solid strategy. Using a cardboard box to get past object detection

[@electrocademyofficial893]

@@carlpanzram7081 Yep, agreed (.. I was just making a throwaway comment regarding at some point in future given the progress that's been made (in particular once AGI/ASI is established in the coming years if abused..).. but how it manifests might well be the sort of way you outline

[@markgomes9054]

Very much so yes

[@LarryPanozzo]

What a time to be alive!

[@eduardoarmenta9232]

Indeed. Holy mother of papers, this was incredible

[@ProtoAlpha]

Machines are truly amazing nowadays

[@limbo1n]

They absolutely need to celebrate their goals. They are already adorable, but that would push them over the edge.

[@MaximilianFeichtinger]

Reminds me of the bet I read about years ago - that 2050 robots will triumph over the best human players. Robocup was founded with that goal in mind in 1997. "RoboCup is an international scientific initiative with the goal to advance the state of the art of intelligent robots. When established in 1997, the original mission was to field a team of robots capable of winning against the human soccer World Cup champions by 2050."

[@SicrosEye]

​@stephanieellison7834 Keep your conspiracy garbage out of this. Citing "reported by some KZbin videos" says it all. Of course big corporations and the way capitalist economy works will have many lose their jobs and don't give a damn/take all the income they get from using AI and robots and shove it up ultra rich's behinds again (just like it happened with most technological advances that were supposed to make "everyone work less and richer"). However then pathetically trying to connect this to a while conspiracy and the killing of the majority is just insane. There are literally thousands of easier ways to get rid of most people if your claimed connections and perfectly cooperating evil actually existed in this way. The reality is way more primitive and sad: Rich and powerful people can't get enough and have close to 0 ethics. Their greed leads them to hoard more and more of it no matter the cost for the rest of us. And no, the majority WON'T die and no they don't need to go through all of that to get a big mansion lmao.

[@robinhodgkinson]

You can knock 25 years of that prediction.

[@aspuzling]

I think one of the reasons the simulated players seem move around like crazy is that they don't consider energy expenditure in their learning. Imagine if the algorithm tried to minimise the energy spent when chasing the ball or defending a position. I think if they did this it would result in much smoother and realistic play. The paper notes that they also don't try to model the real world behaviour of the servo motors and battery so performance drops as the servos become misaligned and battery charge levels drop. Again, optimising for energy usage would probably help a lot here.

[@ExtantFrodo2]

I can see then next paper to hold on to forming as you post. 👍

[@richardpavlov442]

Thats really not a bad idea. Battery life will be for a very long time biggest limit on what robots can do

[@CharveL88]

This is the problem with using simulated environments to train. It's a necessary first-step just so the embodied AI can even function in a basic manner in the wild, but without an excessively stupid amount of real-world training in the thousands, or millions of hours navigating the real world, any AI deployed without those edge-case scenarios will only be useful in hyper-controlled real-world environments which limits usefulness and full autonomy. Guess which company has millions of hours of real-world embodied AI data to train their robots from?

[@aspuzling]

@@CharveL88 I don't know, the point of this research is to show how effective in-simulation training can be. The trained model was applied to the real robots with almost no tweaking and yet they still work pretty decently. They make a point in the paper that they could have done more to allow the robots to perform better but that would disguise how well they performed zero-shot. It basically proves that you can go a long way by simulating the real world and using that to train your model which could allow you to develop a model more quickly and cheaply than what you could achieve by train in the real world only.

[@CharveL88]

@@aspuzling I completely agree. However, I was just making a point addressing the obvious next step, and really the end-goal of AI training; to make them useful in the real world without being micro-managed or geo-fenced, even in a specific use case. Sure, these next few years are going to get saturated with semi-autonomous robots loaded with AI for specific tasks, but it won't be long at all before they'll need a real-world stack that understands the natural world through vision in order to be safe around humans.

[@umbertomann]

Love this channel!! Thank you for you efforts ❤

[@user-qr4jf4tv2x]

the robots need to simulate real football/soccer by pretending to be in pain until 5 people pick them up

[@mriz]

4:14 woah, you recording these robots in person in their lab? amazing!

[@TwoMinutePapers]

Right! It was super fun.

[@pvic6959]

7:15 I LOVE how the opponent also has happy arm flaps :') . great teamwork, little AI

[@Leo_Aqua]

2:36 ROBOT DAB!!😮😮😂😂

[@Jagerbomber]

WHY WHY WAS I PROGRAMMED TO FEEL PAIN

[@nathanfranck5822]

Because otherwise you would damage yourself 😢

[@LutraLovegood]

Pain and pleasure are important for survival.

[@kylanacus2407]

kzbin.info/www/bejne/jYSmnmuvdt11qJo

[@FaffyWaffles]

Wow, science is so amazing!

[@Noname-km3zx]

Finally replacing those weak football players. No more pretending to be hurt, they just get up. What a time to be alive.

[@monstermister8633]

I'm glad you posted this. I've seen this project before, but it seems to be one of the most interesting AI to real-world examples I've seen. Bipedal robots in a competitive match using ball physics. When you visited, was there any discussion of attempting a 2 v 2? They fall extremely well. I think the AI has learned to incorporate this as part of its strategy, its own brand of expected outcomes of locomotion. That probably only works at this tiny scale. Also, the scale of these little guys kind of begs for some sort of consumer product.

[@andoncroft]

Agreed, these small robots in some way do seem conscious or aware and in some competition we out the other robots to get the ball. Not saying a 100 percent but the behaviour is FASCINATING. I'm guessing if they made these robots human size the different weight ratio would make them slower??? Maybe size DOES matter? 😂

[@Epoch11]

Learning to do something in the computer is one thing but using that information to do something in the real world is truly amazing. These physics simulators are mind-blowing.

[@taylanhoca]

Before anyone was talking about even a shred of AI, this channel was releasing best science backed videos in the field

[@Gnomable]

What an adorable leap forward for technology!

[@hermenthicclol1848]

im crying for this!! SOOO AMAZING!!!!!

[@raymond_luxury_yacht]

Hey teacher, leave that bot alone.

[@vasyavasilich7659]

The robots got that Ichigo training arc before beating Aizen

[@TS-kg4lf]

Can't wait to see the next Robocup with thoses training :D

[@AntonioRagagnin]

Awesome! A curiosity: do they have the computational part inside them (that would be really more cool) or do they receive commands remotely?

[@meqdaddev4341]

Impressive!.... it's the time travel machine for learning

[@torq21]

Just thinking back ten years or so to the "robot" soccer and how slow they were! I think at that time they pretty much were all preprogrammed and or remote controlled. These things look absolutely incredible!

[@Kram1032]

it's really interesting how they are clearly riding the edge of instability with that walk: It looks kinda stumbly (in a very cute way) but given their body it's a really good strategy, and if they fail, they are quick to spring back into action regardless. It'd obviously no longer be the same game anymore but I wonder what'd happen if they added slopes and uneven terrain into this, both in terms of how the walk would adapt and in terms of ball control.

[@JJs_playground]

Those little robots are so cute. Lol I know I'm anthropomorphising them. But still .

[@drdca8263]

Well, they do have anthropoid (humanlike) morphology (shape) ?

[@zerosum1290]

mate this is so brilliant, thank you!

[@anthonyrepetto3474]

My tiny robot arms are shivering with excitement! GOAL!

[@SteveRowe]

Are the plans for those robots on a github repo or something? It looks like that would be fun to build and not too difficult.

[@thetransferaccount4586]

getting up from falls was good

[@theoptimisticnihilistyt]

incredible work

[@Ken1171Designs]

Just amazing training. I had a lot of fun watching these little robots playing soccer! They are so cute and funny. ^____^

[@heshumi]

love it. i wonder how hard it is to make a robot like this or simpler with no previous skills

[@HolyMolyNEWTonMusicNProgrammer]

What a time to be alive!!!!!

[@JohnCanCode]

Love this lol. How long would it take to train something similar to this at home on a pc with an RTX card?

[@brenerj917og7]

Thank you

[@lobabobloblaw]

Amazing! What a time…what a time.

[@antoniobortoni]

Much like the human brain, which is composed of specialized regions dedicated to functions such as language, sight, and hearing, robots and artificial intelligence systems can be built with separate modules for movement, vision, and cognitive processing. At the core of these systems, there’s a central controller, akin to our brain's command center, which orchestrates these functions in a seamless and efficient manner. This modular design not only mirrors the structure of the human brain but also allows for flexibility in enhancing or swapping modules as needed while maintaining a cohesive overall operation. With advances in AI and language models, creating such integrated yet adaptable systems is not just a possibility, but an increasingly tangible reality

[@DWJT_Music]

One step closer to Eagle Eyed Terminators

[@IsaacRC]

Thanks for sharing

[@kaimaiiti]

the agents flailing on the ground looked like they'd learned how to "dive" from the professionals

[@joeschroedernz]

6:00 wow 6 gave me chills. It’s so small now, but it looks like simple empathy. Knowing the mind of the other robot to predict its mind and next decision. Technology is about to go crazy. 🎶

[@TsukitaAki]

They look like a real living things. Very cool!

[@ujugamestudio]

Today's paper sounds interesting sir 🤓 So clicked

[@PauloSilva-dz6uh]

@shykids could you please let me know if the voice was generate through the AI too? Cheers. PS wonderful ideia for a video.

[@bearnaff9387]

I would love to see a way to integrate the sensor package and local processing of these little bots into existing commercial robot toys so that a dedicated group could develop training simulations for simple tasks that could be run on things like the Robosen Transformers models. Having Optimus or Megatron stand up smoothly and quickly after losing balance would be neat!

[@abishek786]

Hey Prof, If you can, tweak that thumbnail a bit. I read the "FASTER!" text as "EASTER!". I'm sure that some other scholars also made this mistake.

[@TwoMinutePapers]

Yup, you are completely right, we are already working on it. Happy Easter, Fellow Scholar! 🙂

[@stevenboelke6661]

I have a ~20 year robot not unlike these (though much smaller). It could only execute pre-programmed sequences and be remote controlled. I wonder what I could do to modernize its controls and somehow make it useful. Perhaps in a few years I will have time to figure it out.

[@paullebleu7079]

This blew my mind !

[@theUnmanifest]

"we can simulate 5 years in the future!". Soon it will be "5 million years in a few hours!"

[@hafo821]

hahahh already is 🤣🤣

[@madwlad799]

after training 5 million years in 5 hours the robots, built better robots to transfrom the universe into a football field

[@theUnmanifest]

@@madwlad799 haha yes. I'm thinking things should get interesting at the 5 billion years mark!, after the robots evolved their own agi and reached singularity before simulating new universe in which people evolved to create computers that simulate robots at an accelerating speed!

[@carlpanzram7081]

​​@@theUnmanifestdamn. That's a great sci-fi prompt. Imagine training a artificial intelligence for 5 billion years and then actually putting it in a real body. After 5 billion years of thinking and questioning you reveal to it that it has actually spend all that time in a simulation, and it only took 5 minutes for it to experience billions of years, but now it has reached baseline reality. It would probably start screaming and wailing, or commit suicide.

[@theUnmanifest]

​@@carlpanzram7081 "It would probably start screaming and wailing, or commit suicide." why lol? Soon people will work out that agi will not go crazy, and not become suicidal either haha. By working out the fairly simple mechanism behind us doing these things, which that our ego is a biological system of identities, all working towards achieving their individual goals. Emotions pushes the organism to act on and prioritize giving the control of the organism to whichever identity has the most intense one at the moment. (someone will likely drop studying in this moment if a super hot girl/guy ask them to follow them in the bedroom, prioritizing a new desire) Feeling bored, sad, suicidal, stressed, greedy etc... are all result of one or more of the accumulated identities having failed at achieving their programmed desires. All this is what self-realization is, it's just how things work and ai will pickup on that real quick and will, by nature, eventually be able to understand and even simulate a typical human mind but will not be susceptible to the same issues that we had.

[@greenfoxx1820]

Imagen how nice the new Atlas will move, with enough training. I bet Boston Dynamics uses Simulation Training too.

[@DG123z]

They're like toddlers. Give them another ten years of simulated training and then see what they can do

[@user-cg7gd5pw5b]

I tried it on my virtual machine. They refuse to play football at that point and claim that it's for boomers.

[@FableForge]

They'll be chasing us down a dark alley is what they will do.

[@injest1928]

​@user-cg7gd5pw5b I tried it and they stopped playing too, only they told me existence is pointless. Hopefully with a bit more training time they'll get through this phase.

[@CalebTerryRED]

more training time wouldn't help much, the reason they move like toddlers is because their robot body has a low strength to weight ratio, the simulations that train AI's with stronger bodies move a lot more fluidly. If we want them to run and walk better, we need to build better robots for them with stronger and faster limbs. But for the bodies that they're in now, their movements are pretty close to optimal, which puts a lot of perspective into why toddlers move the way that they do.

[@DG123z]

@@CalebTerryRED Bodies matter but I've seen a lot of longer simulations that go through this rough stage of movement and then emerge with very athletic and life-like motion

[@GuillermoSTD]

Apparently they've also learned to dive in the area of the opponent to fish for a penalty kick

[@clochard4074]

I want to dress them up with tiny gnome costumes XD

[@dschonhaut]

Robotics is about to really change, and the pace could well be blistering

[@blogintonblakley2708]

Reminds me of the "Drunken Master" movie. I bet these things just keep getting smoother.

[@fuccasound3897]

Brilliant!

[@pottacoola]

this is actually fun to watch

[@harrisonlow]

This is so cool 😍

[@placidoescalera8978]

Amazing!!!! They can even surpass the robot creator!!!

[@AtomMachinerule]

The engineers that created the factory motions were for a hard surface and aimed at conservative movements as not to wear out the servo motors. The AI motions are very hard on the servos and I would assume they are replaced often. Although I am impressed with the AI's progress.

[@likoetsvp]

The good thing about robots is they won’t get mad if you push them down. The bad thing is they won’t have any remorse if they end your life. Fascinating and frightening all at the same time.

[@metatechnocrat]

The next version of "Robot Wars" is going to be awesome.

[@juanvovoncebuitragogaleano863]

But how do they gather information from the real world? Just the camera on their heads and (im assuming) the sensor on their chests?

[@parazels83]

I'm shocked how these robots learnt to block balls. They do predict actions of their opponents!

[@zerotwo7319]

I always dreamed to do that.

[@econo23]

I want to see more of that

[@mikhailbulgakov1472]

I hope to see that sort of trained robots competing at the next robocup competition. I wonder why the demonstration only involve two robots instead of a full team.

[@orpheuscreativeco9236]

It's adorable and horrifying 😅👍 Very cool research! Thanks for sharing 🙏

[@1PercentPure]

I really appreciate google and their open research. This is awesome

[@oshapermadi]

I have a question Dr. Károly. So, without the rules (such as penalties for being too close to another robot), is there a possibility that a robot could learn to dismember other robots as the easiest way to win the game?

[@gingeral253]

This development is amazing! For sure it’s gonna be used in real world applications (good and bad).

[@spyral00]

These are so cute! I would totally have one as a pet. At least toxoplasmosis would be off the table.

[@ibgib]

Are they rewarding for less energy consumed vs outcomes?

[@Victor-oy8bj]

this might be one of my favorite videos its so funny too

[@bardhokajvazi2726]

This channel was here all along talking AI. Now AI is mainstream this guy is making more view than ever. Good for him.

[@mumiemonstret]

4:14 Human football would be quite exciting if played with a pilates ball.

[@MrYerak5]

If thay give them pain receptors they will learn to fall less

[@FacundoComix]

El robot gambeteador trató de meter un caño, qué grande!

[@RaphaelChikuka]

I could be wrong. But I think they are falling down on purpose sometimes for pace control and to position themselves in a certain path for the ball.

[@abcmaya]

Time for them to make fighting bipedal robots!!!

[@arsenicide3375]

Where can I get one of these to kick around to make me feel in control?

[@ItzMrCuddles]

eyy im early. heres my like and comment so the algorithms see some engagement.

[@micmanegames695]

I enjoy seeing the steady and incredible progression of AI, it reminds me of a whole new (man-made) lifeform coming to fruition, I'm glad I can experience this with all of you. I will always remember Two Minute Papers as a reputable channel, where I get the most legit info I can get from online.

[@lamsmiley1944]

Surely by this point we should be able to create more agile robots. Even companies like Boston Dynamics struggle to even make a smooth bipedal robot walk.

[@ebrahimebrahimi710]

Holy mother of papers 😂 Such a two minute paper thing

[@ozzymandius666]

Hmmm, now I'm curious about the physics engine of the video game they used to train the robots.

[@ogr3dblade6]

"WHAT A TIME TO BE ALIIIIIIVE!!!" subbed

[@Batmancontingencyplans]

I hope they give the little bots better body

[@suicidalbanananana]

Looks like the "AI" figured out that you want to be unbalanced towards the front when moving forward & thats why its walking gait is so much faster compared to the programmed gait (that's just trying to do neat steps and stay balanced), very cool to see! 👍 Hope we're still getting robots vs humans soccer match in 2030 or whatever the planned year was because that would be so cool 😁

[@mdis2bod]

That lady will be the first one on the list once the robot uprising takes hold

[@dingusbrule5756]

So cool