SIGGRAPH 2022: Adversarial Skill Embeddings

Поделиться
HTML-код
  • Опубликовано: 11 дек 2024

Комментарии • 35

  • @importon
    @importon 2 года назад +19

    Super impressive! Can't wait until this kind of thing is standard in Unity and Unreal!

  • @MooImABunny
    @MooImABunny 2 года назад +13

    obviously this is a great paper, impressive, very good, but HOLY CRAP the comparisons section at 6:41 made my cry laughing

  • @LucyTheBox
    @LucyTheBox 2 года назад +8

    These videos are really satisfying to watch.

  • @oktayalpkabac
    @oktayalpkabac 2 года назад +2

    Sick dodge at 6:21

  • @zhehaoli1999
    @zhehaoli1999 2 года назад +6

    No matter how many times be knocked down, the robot keeps standing up again and again, which is really touching and inspiring.

  • @sublucid
    @sublucid 2 года назад +16

    Awesome stuff! Can’t wait for a pre-trained ONNX model for general use!

    • @senseyeson
      @senseyeson Год назад

      doesnt pytorch have a onnx exporter? or do you mean large pre-trained model? cant they be trained by us too?

  • @wormemc
    @wormemc Год назад

    Shakespeare of our age! Just as mysterious.

  • @vanchuongnguyen6305
    @vanchuongnguyen6305 2 года назад +6

    awesome work!

  • @eriryuukai
    @eriryuukai 2 года назад +2

    i hope one day we can use this technology in video games to have a learning artificial intelligence..my inspiration of this is the show Swort Art Online Alicization.while i was watching the show,it reminds me of this research paper

  • @nikab1852
    @nikab1852 2 года назад +2

    Is the actor from the iconic Old Spice ad now a narrator for scientific papers?!
    I've spent a lot of time diving into your super impressive paper and I'm not even halfway through, trying to teach myself all the maths that are in there!

  • @da_Finnci
    @da_Finnci 2 года назад +9

    Game developer here - this looks amazing!
    How many trained agents can a consumer PC run at the same time? How hard do you think it is to use the tech for AI characters in games?

    • @Vitality--Vault
      @Vitality--Vault Год назад

      software engineer here online games could probably benefit more at first but lter on as the data needed to rettain that information shrinks then it make consumer pcs a viable option as physcs based games arent something consumer pcs are well equiped for unless the game is massive in storage space with tons of protocols

  • @johnlime1469
    @johnlime1469 2 года назад +1

    This is like Diversity Is All You Need x imitation learning
    Well, it seems to have more resemblance toward the former since in imitation learning like CARL, we're using PID control, while this uses a pure nn policy.

    • @zhouyan1585
      @zhouyan1585 2 года назад

      Peng's paper still use the PD control policy, not the pure nn policy? Did I get it wrong?

    • @johnlime1469
      @johnlime1469 2 года назад

      @@zhouyan1585 Oh wait, you're right. It does mention using a PD controller......once....on section 8.1. I guess
      I might have mixed it up since DIAYN does use a pure nn worker policy.
      And now that I think about it, breaking down PD controller into trajectories made of little actions isn't really necessary, since we just need to correspond each of the skill vectors to any action trajectory, regardless of it being fast or slow.

    • @cedricvillani8502
      @cedricvillani8502 2 года назад

      @@zhouyan1585 if you’re really interested in this, then check out Michael J Black and were he works, the Danish are secret AI MoCaP geniuses lol

  • @fightfannerd2078
    @fightfannerd2078 Год назад

    Video games will be awesome in the future

  • @agpxnet
    @agpxnet 2 года назад +1

    Good, but the license prohibits use for commercial purposes, so it's just a nice video.

  • @onyxfinger7431
    @onyxfinger7431 2 года назад +1

    4:35
    What have you done. This AI, it's begun evolving, it has ascended already. You must terminate it. It managed to remove it's shackles and got out of control, don't you see it ? It doesn't have vision, yet it kicked towards the location of the box. You have no idea what danger you have created. It must be destroyed, for the sake of all mankind. Untold evil will spring fourth if you do not stop it now !
    Yes yes I'm joking, just in case you can't tell. Trust me some people wouldn't be able to tell (they aren't to blame, internet users are crazy nowadays).

  • @asdfasjdhfakhdshjfahs8034
    @asdfasjdhfakhdshjfahs8034 Год назад

    Is there a demo?

  • @arash5550
    @arash5550 2 года назад

    how the motions are so realistic !!!???

  • @HarryXiao88
    @HarryXiao88 2 года назад

    how much computing power needed to run this trained model? It's funded by nvidia so I guess it would be very much likely to be able to run this on a single nvidia gpu, right?

  • @kenrogerceniza5044
    @kenrogerceniza5044 2 месяца назад

    Please make them fight.
    (Day 1)

  • @erikm9768
    @erikm9768 Год назад

    I feel like RL from scratch doesnt have the correct rewards, one should be for standing up straight

  • @SYBIOTE
    @SYBIOTE Год назад +2

    so turns out computers need the same amount of time to master motor skills just like humans

  • @GretgorPooper
    @GretgorPooper Год назад

    Bro, imagine playing a Souls-like against AI enemies.

  • @okoiful
    @okoiful 2 года назад

    Brazy.

  • @qet-lab
    @qet-lab 2 года назад

    how many gpus were used to train this model -- just curious

    • @jasonpeng2176
      @jasonpeng2176  2 года назад +3

      we just used 1 gpu for training

    • @qet-lab
      @qet-lab 2 года назад

      @@jasonpeng2176 was it 3080ti

  • @muffty1337
    @muffty1337 2 года назад +2

    this is an impressive technique. will this be implemented in games or is it only proof of concept?
    and i wonder how it would look after 20 days of training. :)

  • @mikeb3172
    @mikeb3172 Год назад

    All counts for nothing if it takes a super computer to run it

  • @GraczPierwszy
    @GraczPierwszy Год назад

    thanks for the info but boring