Skip to content
RobothecaBeta
Log in

H-JEPA

  • AI model
  • Humanoids and robot learning

H-JEPA is a training and planning codebase for hierarchical action-conditioned world models that plan visual tasks top-down in learned latent spaces. It covers simulated and DROID robot data.

Source on GitHub: github.com/kevinghst/H-JEPA (opens in a new tab). Facts read at commit 97054ec (opens in a new tab).

Licenses

Code
MIT (opens in a new tab)

Read by Robotheca from LICENSE (opens in a new tab) (MIT) at commit 97054ec (opens in a new tab), on 16 Oct 2026.

Commercial use
Yes

What it does

  • Learns a stack of action-conditioned JEPA world models end-to-end, with each level predicting farther ahead in its own latent space.
  • Plans top-down: the top level scores progress toward the goal and each level's predictions serve as subgoals for the level below.
  • Supports Visual AntMaze, FourRoom Distractors, OGBench Cube, Push-T, and offline planning on DROID.
  • Includes scripts to reproduce the paper's depth comparison, cost-ladder tables, and DROID planning results.
  • Provides training commands, standalone planning evaluation, and pretrained checkpoints for download.
  • Built on LeWM and stable-worldmodel.

Written by Robotheca from the project's README (opens in a new tab) at commit 97054ec.

Getting started

Difficulty
Advanced
Hardware it needs, as the project states it
Training DROID models uses 2 GPUs with batch 128 per GPU, but no GPU model or memory amount is stated.

Maintenance

Status
Active: a commit in the last 90 days
Newest commit
10 Oct 2026 on main
Latest release
None published on GitHub

Checked on GitHub on 16 Oct 2026.

Links

Robots that can run it

No robot on Robotheca is linked to it yet. Links come from a robot's robot.yaml, from its dependency files read as text, or from Robotheca's editors.

Robotheca lists AI models as a reference: nothing is hosted or run here, and every fact links to where it was read.