AnalysisAI ModelsOctober 6, 2026

H-JEPA hierarchical world model lifts Visual AntMaze success from 18% to 73%

Read original source →arxiv.org

H-JEPA stacks action-conditioned JEPAs so each level predicts farther ahead in its own learned latent space, planning top-down with subgoals. On Visual AntMaze a three-level hierarchy raised success from 18% to 73% using less planner compute; with inverse-dynamics supervision it extends to real-robot DROID videos.

People · Yann LeCun

How this story unfolded

3 days · 2 reports · 3 community posts · from Oct 6

  1. Oct 6
  2. Oct 7
  3. Oct 9

More stories today

Open the live feed