跳过导航
Traditional tokenization methods for robotic actions struggle with high-frequency, dexterous tasks…
0:00 / 1:41

Traditional tokenization methods for robotic actions struggle with high-frequency, dexterous tasks…

The Humanoid Hub

@TheHumanoidHub11万 位关注者

在 X 上查看
297 个赞下载原帖

26,198 次观看

Traditional tokenization methods for robotic actions struggle with high-frequency, dexterous tasks due to redundancy and inefficiency. Inspired by JPEG compression, Physical Intelligence has developed a compressed action representation that accelerates VLA model training 5x.

7 条评论

评论来自 X (Twitter) 上的原帖

  • The Humanoid Hub 的头像

    @TheHumanoidHub

    More details in this thread:

    17 个赞

  • UserInterface 的头像

    @SeketAanru

    Unveiling the Future of Prompt Engineering for Better AI Interactions #tech

    1,508 个赞

  • navuud 的头像

    @navuud

    How does one build a laundry robot like this at home 😁

    2 个赞

  • Brian Bellia 的头像

    @BrianBellia

    So, does this mean the end of teleoperation as a means of training humanoids? I hope it spells the end of teleoperation, period. At this stage, it should be autonomous or nothing - except in rare cases like Optimus catching a ball.

    2 个赞

  • The Humanoid Hub 的头像

    @TheHumanoidHub

    The DROID dataset they using for training is generated with teleop

    1 个赞

  • Disarm.AGI.UBI 的头像

    @PageLyndon

    Give'm some practice. They will get better.

    1 个赞

  • Rethynk AI 的头像

    @rethynkai

    Really like the way it put the 2nd on the top of the first. After intellectual capital, machines are getting better understanding of physical environment.

    1 个赞