DDP-WM: Disentangled Dynamics Prediction for Efficient World Models

Checkpoints of DDP-WM (ICML 2026). Code: https://github.com/HCPLab-SYSU/DDP-WM

Folder	Content
ddpwm_pusht, ddpwm_point_maze, ddpwm_wall, ddpwm_rope, ddpwm_granular	DDP-WM models (all three training stages in one checkpoint) with their training configs (hydra.yaml)
dinowm_rope, dinowm_granular	DINO-WM models we trained with the DINO-WM code; planning only uses their image decoders to render predicted features

For PushT, PointMaze and Wall, the DINO-WM checkpoints are the ones released by DINO-WM.

Usage

From the root of the code repository:

hf download yinkx3/ddp-wm --local-dir checkpoints/outputs
python plan.py --config-name plan_pusht.yaml

See the code repository for the datasets, the other environments and training.

Citation

@inproceedings{
yin2026ddpwm,
title={{DDP}-{WM}: Disentangled Dynamics Prediction for Efficient World Models},
author={Shicheng Yin and Kaixuan Yin and Weixing Chen and Yang Liu and Guanbin Li and Liang Lin},
booktitle={Forty-third International Conference on Machine Learning},
year={2026},
url={https://openreview.net/forum?id=gEc0iJCYxw}
}