---
title: UA-NWM Aerial Navigation
emoji: 🚁
colorFrom: indigo
colorTo: red
sdk: gradio
sdk_version: 6.22.0
app_file: app.py
short_description: Uncertainty-aware world model for aerial navigation
python_version: "3.12"
startup_duration_timeout: 30m
---

# UA-NWM: Uncertainty-Aware World Model for Aerial Image-Goal Navigation

This demo showcases **UA-NWM**, an uncertainty-aware navigation world model for aerial
image-goal navigation from the [paper](https://arxiv.org/abs/2608.05597).

## How it works

Given context frames (aerial observations) and 32 candidate navigation trajectories, the
world model:

1. Encodes images with a frozen DINOv3 ViT-B/16 encoder
2. Compresses DINO features into latent state tokens
3. Autoregressively rolls out each candidate's action sequence
4. Predicts future DINO features and scores them against the goal image using **HEP**
   (Hierarchical Error Projection)

**HEP** decomposes the prediction-goal discrepancy into:
- **|e_parallel|** — explained by the predicted uncertainty subspace (low-rank)
- **|e_perp|** — the orthogonal residual that cannot be explained (the HEP cost)

The demo visualizes the trajectory ranking, per-candidate scores, and the HEP uncertainty
heatmaps for the selected candidate.

## References

- [Paper (arXiv)](https://arxiv.org/abs/2608.05597)
- [Project Page](https://duryi.github.io/UA-NWM-Project-Page/)
- [GitHub](https://github.com/DurYi/UA-NWM)
- [Checkpoints](https://huggingface.co/DurYi/UA-NWM-Checkpoints)
- [Dataset](https://huggingface.co/datasets/DurYi/AirGoal-10k)