| --- |
| license: mit |
| --- |
| <br> |
| <p align="center"> |
|
|
|
|
| <h1 align="center">🌍 Towards <ins>Uni</ins>fied <ins>W</ins>orld <ins>M</ins>odels for Visual Navigation via Memory-Augmented Planning and Foresight</h1> |
| <p align="center"> |
| <a>Yifei Dong<sup>1,*</sup>,</a> |
| <a>Fengyi Wu<sup>1,*</sup>,</a> |
| <a>Guangyu Chen<sup>1,*</sup>,</a> |
| <a>Lingdong Kong<sup>2</sup>,</a> |
| <a>Xu Zhu<sup>1</sup>,</a> |
| <a>Qiyu Hu<sup>1</sup>,</a> |
| <a>Yuxuan Zhou<sup>1</sup>,</a> |
| <a>Jingdong Sun<sup>3</sup>,</a> |
| <a>Jun-Yan He<sup>1</sup>,</a> |
| <a>Qi Dai<sup>4</sup>,</a> |
| <a>Alexander G. Hauptmann<sup>5</sup>,</a> |
| <a>Zhi-Qi Cheng<sup>1,†</sup></a> |
| <br> |
| <sup>1</sup>UW, <sup>2</sup>NUS, <sup>3</sup>Apple, <sup>4</sup>Microsoft Research, <sup>5</sup>CMU |
| </p> |
| |
| <p align="center"> |
| <a href="https://arxiv.org/abs/2510.08713" target="_blank"> |
| <img src="https://img.shields.io/badge/ArXiv-2510.08713-red?logo=arxiv&logoColor=white"> |
| </a> |
| <a href="https://github.com/F1y1113/UniWM" target="_blank"> |
| <img src="https://img.shields.io/badge/Project-UniWM-blue?logo=github&logoColor=white"> |
| </a> |
| <a href="https://huggingface.co/datasets/fly1113/UniWM_Dataset" target="_blank"> |
| <img src="https://img.shields.io/badge/Dataset-UniWM_Dataset-yellow?logo=huggingface&logoColor=white"> |
| </a> |
| <a href="https://github.com/F1y1113/UniWM" target="_blank"> |
| <img src="https://img.shields.io/badge/License-MIT-green?logo=open-source-initiative&logoColor=white"> |
| </a> |
| </p> |
| |
| |
| <p align="center"> |
| <img src="assists/comparison.png" alt="task" width="660"/> |
| </p> |
| |
| **UniWM** introduce a unified, memory-augmented world model paradigm integrating egocentric visual foresight and planning within a single multimodal autoregressive backbone. Unlike modular frameworks, UniWM explicitly grounds action decisions in visually imagined outcomes, ensuring tight alignment between visualization and planning. A hierarchical memory mechanism further integrates detailed short-term perceptual cues with longer-term trajectory context, enabling stable, coherent reasoning over extended horizons. |
| |
| You are also welcome to explore our previous work, including [**GOViG**](https://github.com/F1y1113/GoViG), which introduces a new task that we leverage multimodal LLM reasoning to generate navigation instructions directly from egocentric visual observations of the initial and goal states and [**HA-VLN**](https://github.com/F1y1113/HA-VLN/), where we introduce HA-VLN 2.0, a unified benchmark coupling discrete (DE) and continuous (CE) navigation paradigms with explicit social-awareness constraints. |
| |
| |
| ### Data |
| |
| We host the UniWM dataset on Hugging Face: [`fly1113/UniWM_Dataset`](https://huggingface.co/datasets/fly1113/UniWM_Dataset). |
| |
| - [`go_stanford`](https://huggingface.co/datasets/fly1113/UniWM_Dataset/resolve/main/go_stanford.tar), [`recon`](https://huggingface.co/datasets/fly1113/UniWM_Dataset/resolve/main/recon.tar), [`sacson`](https://huggingface.co/datasets/fly1113/UniWM_Dataset/resolve/main/sacson.tar), [`scand`](https://huggingface.co/datasets/fly1113/UniWM_Dataset/resolve/main/scand.tar) used for both training and evaluation. |
| - [`tartandrive`](https://huggingface.co/datasets/fly1113/UniWM_Dataset/resolve/main/tartandrive.tar) reserved for unseen evaluation only. |
| - [`1XHumanoid`](https://huggingface.co/datasets/fly1113/UniWM_Dataset/resolve/main/1XHumanoid.tar) reserved for humanoid navigation. |
| |
| The directory structure will look like: |
| |
| ``` |
| data/ |
| ├── go_stanford/ |
| │ ├── traj_0000/ |
| │ │ ├── 0.jpg |
| │ │ ├── 1.jpg |
| │ │ ├── ... |
| │ │ ├── n.jpg |
| │ │ └── traj_data.pkl |
| │ ├── traj_0001/ |
| │ └── ... |
| └── ... |
| ``` |
| |
| Each `traj_xxxx/` folder contains a sequence of egocentric frames (`0.jpg`, `1.jpg`, ..., `n.jpg`) and a `traj_data.pkl` file storing the per-step metadata (e.g., actions, poses) for that trajectory. The other splits follow the same layout. |
| |
| |
| ## Contributing |
| |
| We welcome contributions to this project! Please contact yfeidong@uw.edu or fyiwu@uw.edu. |
| |
| ## Acknowledgement |
| |
| We would like to thank [ReCon](https://arxiv.org/abs/2104.05859), [Go Stanford](https://arxiv.org/abs/1803.03254), [SACSon](https://arxiv.org/abs/2306.01874), [SCAND](https://arxiv.org/abs/2203.15041) and [1XHumanoid](https://github.com/1x-technologies/1xgpt) for their publicly available datasets. |
| |
| ## 🌟 Citation |
| |
| If you find this repository or our paper useful, please consider **starring** this repository and **citing** our paper: |
| ```bibtex |
| @misc{dong2026unifiedworldmodelsvisual, |
| title={Towards Unified World Models for Visual Navigation via Memory-Augmented Planning and Foresight}, |
| author={Yifei Dong and Fengyi Wu and Guangyu Chen and Lingdong Kong and Xu Zhu and Qiyu Hu and Yuxuan Zhou and Jingdong Sun and Jun-Yan He and Qi Dai and Alexander G. Hauptmann and Zhi-Qi Cheng}, |
| year={2026}, |
| eprint={2510.08713}, |
| archivePrefix={arXiv}, |
| primaryClass={cs.AI}, |
| url={https://arxiv.org/abs/2510.08713}, |
| } |
| ``` |
| |
