LeCun's AMI Unveils H-JEPA: Hierarchical World Models Lift Visual AntMaze From 18% to 73%
arankomatsuzaki · x · 2026-10-06
- H-JEPA (NYU/AMI, with Yann LeCun, Randall Balestriero et al.) learns a hierarchy of JEPA world models end-to-end for long-horizon visual planning.
- Method: each higher level predicts further ahead in its own latent space; planning runs top-down — the top level plans toward the goal, its predictions become subgoals for lower levels, and the lowest level emits primitive actions. Higher levels discard fast, unpredictable detail while keeping slow, task-relevant state.
- Results: a three-level hierarchy raises Visual AntMaze success from 18% to 73% with less planning compute; beats flat JEPA across four navigation/manipulation environments and improves offline planning fidelity on real DROID robot videos with inverse-dynamics supervision. Paper and code are public.
Related event: LeCun Team Unveils H-JEPA Hierarchical World Model(3 posts)→
More from Embodied
- Steam Frame lens impressions: bigger sweet spot and pleasant image, but glare persists — smtabatabaie · 2026-10-06
- Dev builds Warhammer Mechanicus-style servitor status display for Hermes agent on M5Stack — toddhooper · 2026-10-06
- Autonomous Lamp: $499 open-source desk companion robot ships with Jev reasoning classifier — dee_hw · 2026-10-06
- DIY e-paper fridge newspaper rotates personal info every 30 minutes — luismbat · 2026-10-06
- Tesla appears to be gaining ground on Waymo once data is aligned, scale outside Austin unclear — binarybits · 2026-10-06
- Wayve CEO Alex Kendall: Tesla Proves Self-Driving Demand, But Every Brand Needs to Compete — alexgkendall · 2026-10-06