H-JEPA: Hierarchical world models lift visual AntMaze success from 18% to 73%
ylecun · x · 2026-10-07
LeCun and collaborators introduce H-JEPA, a hierarchical world model built by stacking JEPAs, each level in its own embedding space, learned end-to-end from pixels with top-down planning.
- On Visual AntMaze, three levels lift success rate from 18% to 73%
- Authors include Basile Terv, Kevin Stangl, Michael Rabbat, Yann LeCun, and Randall Balestriero
- A concrete step toward LeCun's long-advocated hierarchical world-model architecture for planning
More from Embodied
- Indian builders make offline AI hardware that listens and remembers without apps — alysha_lobo · 2026-10-07
- Generalist's GEN-1.5 robot model installs and removes O-rings autonomously, learns from seconds of demos — 141_1337 · 2026-10-07
- YC-backed Vexo launches always-listening AI bracelet that keeps your context — ycombinator · 2026-10-07
- DiVeR preprint trains VLA verifiers only on decision-critical states for test-time scaling — SharonYixuanLi · 2026-10-07
- Million-mile trucker after Tesla Semi factory tour: this industry is going to be changed — elonmusk · 2026-10-07
- Attacca trains embodied agents for state-continuity, up to 7x on long-horizon tasks — nanyang-technological-university-singapore · 2026-10-07