Пропустить навигацию
JoyIn unveiled Aether: a 4B-param energy-based model for robots, trained on ~200 hrs of human video…
0:00 / 0:27

JoyIn unveiled Aether: a 4B-param energy-based model for robots, trained on ~200 hrs of human video…

The Humanoid Hub

@TheHumanoidHub116 тыс. подписчиков

Открыть в X
108 отметок «Нравится»СкачатьОригинальный пост

12 885 просмотров

JoyIn unveiled Aether: a 4B-param energy-based model for robots, trained on ~200 hrs of human video, zero real-robot data. Not VLA (perception+language to action) or world models. It's an "action-energy" model: actions emerge from minimizing a learned energy function over perception, memory and dynamics. No planner mapping instructions to trajectories. Claims >200 Hz control and orders-of-magnitude less data/compute than VLA. In this autonomous demo, one robot gets stuck because it's carrying too many items, and the other steps in, adapting to its peer's limitation.

6 комментариев

Комментарии из оригинального поста в X (Twitter)

  • Фото профиля The Humanoid Hub

    @TheHumanoidHub

    The architecture

    7 отметок «Нравится»

  • Фото профиля WAKE

    @buildwithwake

    About 200 hours of human video and zero real-robot data is a strong claim about where skill lives. If actions come from that footage alone, the scarce work is still episode quality: task, scene, and clean cuts before any robot body is in the loop.

  • Фото профиля Bidhan Roy

    @bidhan

    200 hz means the energy search has a few milliseconds at most i am probably missing something

  • Фото профиля Nghẹo | DOPE

    @Quaocchoo

    Interesting approach for learning robot behavior from human video alone

  • Фото профиля Teacher in the loop 🤖

    @teachgineer

    액션-에너지 모델은 처음 들어보네요 인간 영상만으로 어디까지 가능할지도 궁금합니다

  • Фото профиля Armstrong Too | Photo + Film + AI

    @arms_too

    The action-energy framing is more interesting than the label. For robots, the useful question is whether the learned objective stays stable when perception is noisy and the scene changes. Recovery behavior will matter as much as the demo.