---
license: mit
library_name: pytorch
pipeline_tag: robotics
tags:
- world-models
- jepa
- planning
- robotics
datasets:
- jepa-world-models/h-jepa
---

# H-JEPA checkpoints

[![arXiv](https://img.shields.io/badge/arXiv-2610.06805-b5212f?logo=arxiv)](https://arxiv.org/abs/2610.06805)
[![Website](https://img.shields.io/badge/Website-h--jepa.com-blue)](https://h-jepa.com/)
[![Code](https://img.shields.io/badge/GitHub-kevinghst/H--JEPA-black?logo=github)](https://github.com/kevinghst/H-JEPA)
[![Dataset](https://img.shields.io/badge/🤗%20Dataset-jepa--world--models/h--jepa-ffc107)](https://huggingface.co/datasets/jepa-world-models/h-jepa)

Trained models of **H-JEPA: End-to-End Learning of Hierarchical World Models for Visual Planning**
(Wancong Zhang\*, Basile Terver\*, Mike Rabbat, Yann LeCun†, Randall Balestriero†). The
[code release](https://github.com/kevinghst/H-JEPA) trains, evaluates and reproduces the paper's
planning results with them.

## Contents

| Folder | Models | Seeds | Checkpoints | Size |
|---|---|---|---|---|
| `ant/` (Visual AntMaze) | LeWM, H-JEPA 2/3/4 levels, HWM 2/3/4 levels | 42, 43, 44 | 21 | 2.6 GB |
| `fourroom/` (FourRoom Distractors) | same 7 models | 42, 43, 44 | 21 | 1.7 GB |
| `cube/` (OGBench Cube) | same 7 models | 42, 43, 44 | 21 | 2.7 GB |
| `pusht/` (Push-T) | same 7 models | 42, 43, 44 | 21 | 2.7 GB |
| `droid/` (DROID, 5 fps) | LeWM + IDM, HWM 2 levels, H-JEPA 2 levels | 1, 1000, 10000 | 9 | 2.7 GB |

Each run is a folder `<env>/<env>_<model>/seed<seed>/` with `<env>_<model>_object.ckpt` (the trained
model) and `normalizer.pt` (action/proprio normalization), the layout the training code writes.
`<model>` is `lewm`, `hjepa_l<n>` or `hwm_l<n>`, with `<n>` the number of levels. The DROID
checkpoints are the epoch-100 models.

## Download

Into `$HJEPA_HOME/ckpts`, where the code reads checkpoints:

```bash
# the 4 simulation environments (84 checkpoints)
hf download jepa-world-models/h-jepa --local-dir $HJEPA_HOME/ckpts --exclude "droid/*"
# DROID (9 checkpoints)
hf download jepa-world-models/h-jepa --local-dir $HJEPA_HOME/ckpts --include "droid/*"
# one environment or one seed: --include "cube/*", --include "*/seed42/*"
```

## Usage

The `_object.ckpt` files are pickled `HJEPA` modules, so loading them needs the code release on the
Python path (run from its `h_jepa/` directory):

```python
import torch

model = torch.load("cube/cube_hjepa_l3/seed42/cube_hjepa_l3_object.ckpt", weights_only=False)
```

To reproduce the paper's planning results, pass a checkpoint as `policy=` to `eval.py` or run the
evaluation scripts, as described in the code release README (simulation: section 4, DROID: section 5).
The evaluation tasks and datasets are in
[`jepa-world-models/h-jepa`](https://huggingface.co/datasets/jepa-world-models/h-jepa) (datasets).

## Citation

```bibtex
@article{zhang2026hjepa,
  title   = {{H-JEPA}: End-to-End Learning of Hierarchical World Models for Visual Planning},
  author  = {Zhang, Wancong and Terver, Basile and Rabbat, Mike and LeCun, Yann and Balestriero, Randall},
  journal = {arXiv preprint arXiv:2610.06805},
  year    = {2026}
}
```