Hierarchical world models, learned end-to-end from pixels! Excited to introduce H-JEPA: stacked JEPAs, each level in its own embedding space, planning top-down. Visual AntMaze: 3 levels lift success from 18% to 73% 👇 w/ @kevinghstz, Michael Rabbat, @ylecun, @randall_balestr
@Basile Terver
1 grabs
No comments yet
Say something. You’re first.