Toward Physically Grounded JEPA World Models for Goal-Conditioned Robotic Planning
Paper • 2609.03565 • Published
This repository contains our GA-WM checkpoints accompanying Toward Physically Grounded JEPA World Models for Goal-Conditioned Robotic Planning.
| File | Task | Reported checkpoint success rate |
|---|---|---|
gawm_tworoom_object.ckpt |
TwoRoom | 100% |
gawm_reacher_object.ckpt |
Reacher | 86% |
gawm_pusht_object.ckpt |
PushT | 98% |
gawm_cube_object.ckpt |
Cube | 88% |
gawm_cube_bcinit_object.ckpt |
Cube, additional standalone BC-init model | 96% |
These checkpoint-specific results were supplied by the authors for this release. They have not been re-evaluated during upload. The paper abstract reports 87% for OGBench-Cube; the 88% above describes the supplied Cube checkpoint. The additional BC-init Cube checkpoint is listed separately.
from huggingface_hub import hf_hub_download
checkpoint_path = hf_hub_download(
repo_id="kikihuang/GA-WM",
filename="gawm_tworoom_object.ckpt",
)
The files are provided in their original .ckpt format. Downloading them does not install the model implementation or its dependencies.
Muyuan Liu, Yue Huang, Zheng Liang, and Xiang Gao. Toward Physically Grounded JEPA World Models for Goal-Conditioned Robotic Planning. arXiv:2609.03565, 2026.
@article{liu2026physicallygroundedjepa,
title={Toward Physically Grounded JEPA World Models for Goal-Conditioned Robotic Planning},
author={Liu, Muyuan and Huang, Yue and Liang, Zheng and Gao, Xiang},
journal={arXiv preprint arXiv:2609.03565},
year={2026},
url={https://arxiv.org/abs/2609.03565}
}