Nov 3, 2026

H-JEPA: same simulated robot, same maze. Only one plans ahead.

H-JEPA NYU Advanced Machine Intelligence

H-JEPA, from researchers including Yann LeCun, plans the big route first and the small steps later. In one maze test, success rose from 18% to 73%. Simulation only, so far.

Sir Toastimir

Footage: H-JEPA project page

What it is

H-JEPA is new research from NYU, Advanced Machine Intelligence (AMI), INRIA Paris and Brown University. Yann LeCun is one of the authors. It's a world model that plans in layers: the top level sketches the big route, the next fills in waypoints, and the bottom picks the actual moves. Think of planning a road trip: highways first, then the turns.

Why layers help

Higher levels look further ahead and ignore fast, fiddly details, like exactly how a robot's legs are placed, while keeping what matters for the goal, like where it is. That makes long trips easier to plan and, the authors say, cheaper to compute.

The results

The team tested it in four simulated navigation and manipulation tasks. In the Visual AntMaze test, a three-level version raised success from 18% to 73% compared with a single-level planner. It wasn't a win everywhere: on the Push-T task, the deepest versions did poorly, likely because of too little training data.

What to keep in mind

There's no real robot under live control yet. On real-world robot videos from the DROID dataset, the team only measured how well its plans matched what happened, offline.

Try it

The paper is on arXiv, and the code and model checkpoints are public.

Footage in the video

All clips are from the H-JEPA project page. Links are below.