LeCun Team H JEPA Study Raises Maze Success From 18 to 73 Percent

The short version

Researchers built a layered AI world model called H JEPA that plans in steps from big goals down to small moves. On one hard maze task, success rose from 18 percent to 73 percent while using less planning compute. Turing Award winner Yann LeCun co-authored the work on arXiv in October 2026.

What the researchers did

Wancong Zhang, Basile Terver, Michael Rabbat, Yann LeCun, and Randall Balestriero posted preprint, not yet peer-reviewed work on arXiv on 8 October 2026. They train a stack of action-conditioned JEPA models. Each layer learns its own latent space and predicts farther into the future than the layer below. Planning runs top down: the top layer picks progress toward the goal, and each layer turns its forecast into subgoals for the next layer down.

They tested four simulated navigation and manipulation settings. They also used real robot video from the DROID dataset with inverse-dynamics labels to see if hierarchy helps offline planning on real hardware footage.

What they found

On Visual AntMaze, a flat single-layer JEPA reached 18 percent success. A three-level H JEPA reached 73 percent success with less planner compute. Ablations showed gains came both from splitting time scales and from richer goal codes at higher layers. On DROID videos, hierarchy improved offline planning fidelity at lower compute than flat models.

Why it matters for you

If you build agents, game AI, or warehouse robots, long horizon planning is often the bottleneck. This recipe offers a concrete way to let fast noise live in lower layers while slow, task-relevant state stays upstairs. That mirrors how people plan a trip as “get to the airport” before “turn left.”

What to keep in mind

Results are mostly in simulation plus offline real-video tests, not a full live robot trial report in the preprint. Preprint, not yet peer-reviewed. Gains depend on data having factors that change at different speeds.

Think of planning like a road trip. You first choose the city. Then you pick the highway exit. Only then do you watch for lane changes. H JEPA tries to copy that order inside the model. Each layer has its own compact code for the scene. A compact code is a short list of numbers the network uses instead of raw pixels. That saves work because pixels change every frame.

The paper tests four simulation worlds. Simulation means the robot exists only in software. That lets the team run thousands of tries cheaply. They compare a flat model with one timeline against a stack with three timelines. The flat model reached eighteen percent success on Visual AntMaze. The three layer stack reached seventy three percent. They also report using less planner work for those gains. Planner work is like the search a GPS does before suggesting a turn.

Real robot video adds a harder check. Video from DROID shows arms and tables in real homes and labs. Offline planning means the team scores plans without sending commands to hardware during the test. Hierarchy still helped fidelity there. Fidelity means how closely the plan matches what a human expert would expect. None of this replaces safety review on a live factory floor.

Product teams watch world model papers for planning APIs. Layers that plan far ahead may cut cloud search bills. Simulation wins are step one. Live robot proof comes later. Meta and university authors mix industry and academic views. Track peer review on arXiv. Do not treat maze scores as warehouse KPIs yet.

Teams share papers so others can repeat the work. A preprint is an early public draft. Peer review may change claims later. Numbers in tables can move after feedback. Always read the linked paper for exact values. This desk note tracks when we last checked sources. We remove claims if sources disagree. If you build on this work, cite the primary paper. If you spot an error, tell the publisher through the site contact path.

Key takeaways

H JEPA uses several latent layers, each predicting further ahead in its own space.
Top-down planning turns high forecasts into subgoals for lower planners.
Visual AntMaze success moved from 18% to 73% with a three-level stack.
Real DROID clips showed better offline plans with less compute when hierarchy was used.
LeCun’s team ties the idea to JEPA world models used in broader AI research.

Photo: Pexels

Sources

Primary — arXiv, H-JEPA: End-to-End Learning of Hierarchical World Models for Visual Planning (8 October 2026). Methods, maze success rates, and DROID extension.