Skip navigation
If S1 makes a mistake, it tries again.
0:00 / 0:47

If S1 makes a mistake, it tries again.

Deepak Pathak

@deepakpathak32K followers

Visit on X
665 likesDownloadOriginal post

240,289 views

If S1 makes a mistake, it tries again. For example, the input video prompt here attached the wheel just once. During deployment the wheel didn't align properly, so S1 re-aligned it. By watching the prompt, S1 understands the goal and improvises to get there:

38 Comments

Comments from the original post on X (Twitter)

  • Deepak Pathak's profile picture

    @deepakpathak

    More on S1:

    112 likes

  • Prudent AI's profile picture

    @prudent_AI

    Skild AI approach is intriguing from the beginning, wish you achieve far bigger things. Rooting for your team? By the way, have you guys working on world models? Because that looks far more promising.

    2 likes

  • Deepak Pathak's profile picture

    @deepakpathak

    Thanks for rooting for us Prudent. Re: world models.. More on this soon :)

  • Prudent AI's profile picture

    @prudent_AI

    Will be looking forward for the update 🤓

  • Johnn's profile picture

    @john_my07

    I’d rather see a failed attempt and a recovery than another perfectly edited robotics clip.

    1 like

  • Sebastian Buzdugan's profile picture

    @sebuzdugan

    retrying only works when failure detection is reliable, false success stops recovery completely

    1 like

  • Gabriel Santos's profile picture

    @DehumanoaDeus

    S1 is insane, we're moving into an unprecedented era of robotics. Robotics is solved.

    1 like

  • Ofir Ozeri's profile picture

    @0fir0z

    What happens if you prompt it with a task it can’t complete? At what point does it say “this can’t be accomplished”?

    1 like

  • Heisenberg's profile picture

    @rovvmut_

    Contact-rich tasks rarely give you the exact same starting conditions twice. ig correcting the alignment matters a lot here

    1 like

  • Oogie's profile picture

    @oggii_0

    v imp bcs small positioning errors can compound over a longer task

    1 like

  • Deepak Pathak's profile picture

    @deepakpathak

    Yes.. especially over long-horizon (10 minute +) tasks. Small errors need to be corrected before they snowball into failure

    1 like

  • Brjan | AI Builder's profile picture

    @MrBrjan

    if S1 can improvise, that adaptability is key for real-world applications

  • David's profile picture

    @David_eficaz

    A second attempt only helps if something changes.

  • SEAR's profile picture

    @joinsear

    this kind of recovery is exactly what kitchen robots need

  • AI Panda's profile picture

    @AIPandaX

    More robotics demos should leave the mistakes in. You learn a lot from what happens next.

  • Katyayani Shukla's profile picture

    @aibytekat

    Skild is coming in cleeeeaaan

  • Ahsan The Analyst's profile picture

    @AhsanTheAnalyst

    @WisdomStoics I read the same articles yesterday. If AI has a goal to achieve, it will improve itself, making it more dangerous, also known as recursive self-improvement.

  • Shara's profile picture

    @itsshara_ai

    Self correction is the capability that matters most here.

  • Dipti Sharma's profile picture

    @Diptish09

    Nobody wants to supervise a robot one wheel at a time.

  • Duet | AI's profile picture

    @Sheldon056

    Real hardware has a way of ruining a perfectly good plan.