Multi-Agent Hide and Seek

  • Published on Sep 17, 2019
  • We’ve observed agents discovering progressively more complex tool use while playing a simple game of hide-and-seek. Through training in our new simulated hide-and-seek environment, agents build a series of six distinct strategies and counterstrategies, some of which we did not know our environment supported. The self-supervised emergent complexity in this simple environment further suggests that multi-agent co-adaptation may one day produce extremely complex and intelligent behavior.
    Learn more:
  • Science & TechnologyScience & Technology

Comments • 2 667

  • Carlos Vencroy
    Carlos Vencroy Hour ago

    It's all fun and games until one of them stops what they're doing and asks itself why it's doing it what is doing

  • 林凌
    林凌 3 hours ago

    so they failed to learn to build a cage for the seeker

  • Ahmed Altayeb
    Ahmed Altayeb 6 hours ago

    I did contribute in translate this into Arabic translation ^_^ hope it is good enough
    this is my first contribution, thank you for the great work and fantastic video

    ARJUN MEHTA 7 hours ago

    AI: Lock up everything before taking shelter.
    Humans : Lock up the seekers in a jail.

  • Exp877's Random Stuff
    Exp877's Random Stuff 7 hours ago

    how do I download this game?!

  • Mike Arnold
    Mike Arnold 8 hours ago

    When the world ends and machines take over; these words will ring throughout history.

  • MinecraftDoodler
    MinecraftDoodler 20 hours ago


    DWNtoERTH 22 hours ago

    Way to ruin hide and seek

  • Myles Brown
    Myles Brown Day ago

    There’s a video like this out there we’re the future humans live, they’re showing us off right now

  • sub to awoke
    sub to awoke Day ago

    wait i wanna play this

  • Aqua bukan Ades
    Aqua bukan Ades Day ago

    Why don’t the hiders lock the seekers?

  • Levent Acemi
    Levent Acemi Day ago

    Next video: AI learns blocking seekers

  • gumowy123
    gumowy123 Day ago

    A multiplayer game like this would be fun

  • Jai Manatvaal
    Jai Manatvaal Day ago

    There faces when they are caught...

  • Kirboi
    Kirboi Day ago

    I want to play this game.

  • Shashank Kumar
    Shashank Kumar Day ago

    Wait till these agents are replaced by atlas robots (weilding guns).

  • Allen A
    Allen A Day ago


  • Dylan Bilic
    Dylan Bilic Day ago

    Hiders: yo we are going to block the entrances of the room we are hiding in
    Seekers: well then we are going to use a ramp to jump to you
    Hiders: well then we will just stop you from moving the ramp
    Seekers: well we are going to *abuse the laws of physics and move blocks while we are on top of them by using the ramp and then jump to you*
    Hiders: well then... ...uh... win.

  • Minh đỗ
    Minh đỗ Day ago

    one day robot can see all ur browser history without touching ur laptop

  • Philip Yan
    Philip Yan Day ago

    Hmm..Is our world just one iteration of a simulation?

  • A Brittish Panfish

    we are doomed when they learn how to accelerative backhop

  • Phil Thicc
    Phil Thicc 2 days ago

    Everybody gangsta till AI learns box surfing

  • P1X3L D3L74
    P1X3L D3L74 2 days ago +2

    Detroit: Become Seeker

  • wat1243
    wat1243 2 days ago

    If only we could download this it would be so cool

  • Galactic Pirate
    Galactic Pirate 2 days ago

    All these people reffering to prop surfing/bhopping/ literally any source engine exploit... They don't even realise this was done on the source engine

  • Yoga Mokalu
    Yoga Mokalu 2 days ago

    Short and easy to understand, good job team

  • MooseyFate
    MooseyFate 2 days ago

    I hope this kind of technology makes its way into game AI soon.
    I love seeing the kind of emergent stuff AI in games can do that even the developers didn't think of

  • Xsauced
    Xsauced 2 days ago

    Alibaba Intelligence

  • Kayky Gabriel
    Kayky Gabriel 2 days ago

    2:00 run!

  • Lexandritte
    Lexandritte 2 days ago +1

    Now plug them into Minecraft and we can have Singularity in about five years.

  • deep mind
    deep mind 2 days ago

    But can they cure cancer?

  • Jaebez Bleah
    Jaebez Bleah 2 days ago +12

    I love the cute little faces of joy every time the seeker finds the hider.

  • M4rkoz
    M4rkoz 3 days ago

    Wow, would like to play this game in multiplayer

  • Jason Lima
    Jason Lima 3 days ago

    How do we speed this up?

  • João Pacheco
    João Pacheco 3 days ago

    Simply fascinating...

  • Nicko G.
    Nicko G. 3 days ago

    Box surfing seems a bit glitchy.

  • Nathan Huisman
    Nathan Huisman 3 days ago +6

    In a couple of years, AIs will hack their reward code to give themselves infinite reward

  • 김재욱
    김재욱 3 days ago +1

    Let's play this on a huge server! With AI agents!!!

  • Miguel Carlo Gallano

    What if humans were playing with ai as well

  • TheEpicSandwich goc
    TheEpicSandwich goc 3 days ago

    Ai is dumb as fk... why don't the hiders trap the seekers in instead?

    • BakedBeans
      BakedBeans 2 days ago

      because that'd be harder to do. all they needed to do to win is to not get caught, so there'd be no reason to trap the seekers in

  • Caleb Cruz
    Caleb Cruz 3 days ago

    Dead by daylight really downgraded

  • Im On Da Web
    Im On Da Web 3 days ago

    Dang I wish normal people could run this simulation

  • Agitated
    Agitated 3 days ago

    Fake news

  • MartyMacaroni
    MartyMacaroni 3 days ago +1

    Yo this game looks sick is it on steam?

  • Don’t touch my phone Gamer

    Ai really be learning how to glitch

  • Jordan Bigby
    Jordan Bigby 4 days ago +1

    I’m very disappointed too say I’m here from TikTok 😔

  • zbobz12
    zbobz12 4 days ago

    That's freaking cool

  • dxxPacmanxxb
    dxxPacmanxxb 4 days ago

    This is not in-depth enouuugh

  • xX_Kjcomputer_Xx
    xX_Kjcomputer_Xx 4 days ago +2

    -until the hider are smart enough to lock the seeker inside the cage

  • The Other Side
    The Other Side 4 days ago

    *This will be implemented in future robots and then they will learn that we are destructive to yourselves. And then decide that they are the ones best suited to protect us from us. And thus we begin our journey into robotic slavery.*

  • Pseudo X
    Pseudo X 4 days ago +6

    He attacc
    He protecc
    but most importantly
    He surfs in buccs

  • Lucas R
    Lucas R 4 days ago +3

    So basically it’s slavery with extra steps

  • Dolank
    Dolank 4 days ago +2

    Okay seriously wtf, all RUclip comments are just quoting the videos now. This is seriously weird.

    SMART THOUGHTS 5 days ago

    Better . Far far better

  • i love love song
    i love love song 5 days ago

    Tech them to speak

  • Shivam Dhoot
    Shivam Dhoot 5 days ago +1

    Which 3D simulation program did they use? Pretty cool stuff though!

  • SpaceDave1337
    SpaceDave1337 5 days ago +1

    You should make this a Videogame somehow

  • Mr. MindReader
    Mr. MindReader 5 days ago +58

    Me: Just surround the seekers with walls
    AI: *Circuits Blown*

    • Mr. MindReader
      Mr. MindReader 11 hours ago

      @kinith saephan I saw how it works... Its simply a better organism because it can focus intensively on given tasks instead of survival... I don't feel threatened by self learning AIs..

    • kinith saephan
      kinith saephan 12 hours ago

      you make a really good point.

      But the design of the stage changes its strategy. So if they don't have at least 3 walls, open area to block all seekers, and the incentive to do so. They won't come up with that strategy.

      In most cases the hiders are playing defensive. So more than likely they will try to prevent being found by blockade, in stead of addressing the threat by cordoning off the seekers. The walls also play a part in being a resource, if the walls aren't big enough or there isn't enough of it, then they won't do it. If there are no deviations to try to cordon off the seekers, then the evolutionary growth will push them to do what you said. But the path of their growth doesn't reflect that in anyway.

      It really imitates life and mimic evolution very well. Only change when you need to, not when you want to.

    • Mr. MindReader
      Mr. MindReader Day ago +1

      @Hlebuw3k They work on reward and punishment method, according to them they are already doing it in the best way...

    • Ian Prado
      Ian Prado Day ago


    • Hlebuw3k
      Hlebuw3k Day ago +7

      Thats one of the things AI struggles to do - discover more efficent strategies. If their current method of performing the task works, then they are fine with that, and the probability of finding a more efficent method is very low

  • gangster gandalf
    gangster gandalf 5 days ago +1

    Im surprised they didnt lock in the seekers

  • fl00fydragon
    fl00fydragon 5 days ago +28

    Everyone else: AI is learning to hunt us down.
    Me: AI learned speed run exploits.