README.md

metadata

license: apache-2.0
pipeline_tag: robotics
tags:
  - world-model
  - aerial-navigation
  - image-goal-navigation
  - uncertainty-estimation
  - dino
  - raev2

UA-NWM Checkpoints

This repository provides the released checkpoints for UA-NWM: Uncertainty-Aware World Model for Aerial Image-Goal Navigation.

Project page: https://duryi.github.io/UA-NWM-Project-Page/
Code: https://github.com/DurYi/UA-NWM
Paper: https://arxiv.org/abs/2608.05597

Files

ua-nwm/
└── ua-nwm_best.pt

raev2-dinov3b-k1-airgoal10k-ft/
└── step_0010000.pt

• ua-nwm/ua-nwm_best.pt: the released UA-NWM checkpoint. It contains the shared deterministic backbone and the HEP module. Both inference strategies in the codebase use this same checkpoint: hep for uncertainty-aware scoring and dinov3_cosine for the deterministic baseline.
• raev2-dinov3b-k1-airgoal10k-ft/step_0010000.pt: the fine-tuned RAEv2 RGB decoder used only for DINO latent visualization.

The DINOv3 ViT-B/16 encoder is not redistributed here. Please download the official Meta checkpoint from facebook/dinov3-vitb16-pretrain-lvd1689m.

Download

From the root of the UA-NWM code repository:

hf download DurYi/UA-NWM-Checkpoints \
  ua-nwm/ua-nwm_best.pt \
  raev2-dinov3b-k1-airgoal10k-ft/step_0010000.pt \
  --local-dir pretrained

After downloading, the expected local layout is:

pretrained/
├── ua-nwm/
│   └── ua-nwm_best.pt
└── raev2-dinov3b-k1-airgoal10k-ft/
    └── step_0010000.pt

You still need to place the DINOv3 encoder at:

pretrained/dinov3-vitb16/

Usage

See the main repository for environment setup, dataset preparation, evaluation, visualization, and training instructions.

Citation

@misc{zhu2026uanwm,
      title={Uncertainty-Aware World Model for Aerial Image-Goal Navigation},
      author={Deyi Zhu and Haoyu Fan and Yinan Zhu and Weichen Zhang and Shilin Ma and Xinlei Chen and Yansong Tang},
      year={2026},
      eprint={2608.05597},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2608.05597},
}