Announcement
UniAction proposes one world model for several kinds of action
In an X post, @zizhang_li describes mixing object forces, camera movements, hand motions and robot-arm actions in one generation.
TLDR
@zizhang_li introduced UniAction in an Oct. 26 X post as a framework for conditioning one world model on forces, camera movements, hand motions and robot-arm actions, including combinations in a single generation. The author claims jointly trained UniAction outperforms action-specific models, but the post provides no benchmark or evaluation conditions for that comparison.
Combined views
6K
4 Sources, first seen ago
