UA-NWM

Uncertainty-
Aware
World Modelfor Aerial
Image-Goal
Navigation

UA-NWM scores candidate navigation trajectories by asking whether the goal image lies within a conditioned future-state distribution represented by an uncertainty subspace.

Deyi Zhu*, Haoyu Fan*, Yinan Zhu, Weichen Zhang, Shilin Ma, Xinlei Chen, Yansong Tang†
Tsinghua Shenzhen International Graduate School, Tsinghua University
Paper Code Dataset Demo BibTeX
Future-state uncertainty is pronounced in large outdoor scenes, making point-to-point trajectory scoring unreliable. A point-to-distribution scoring mechanism is needed to handle this uncertainty.

Summary

Aerial image-goal navigation requires a UAV to reach a target location specified by a goal image. Existing world-model rankers usually compare the goal to one or a few predicted futures, which is fragile when outdoor scenes admit multiple plausible future observations. UA-NWM formulates trajectory scoring as conditional out-of-distribution detection: a trajectory is good when the goal discrepancy is explainable by the predicted uncertainty subspace, and bad when a large portion of the discrepancy remains orthogonal to that subspace.

Comparison of UA-NWM with deterministic and stochastic navigation world models

Video demos

Closed-loop UAV navigation demonstrations

Two demos show UA-NWM performing real-world UAV image-goal navigation.

Demo 1

Visualization of a navigation task under clear visibility.

Demo 2

Visualization of a navigation task under occluded visibility.

Core idea

Score only what the uncertainty subspace cannot explain

UA-NWM predicts a deterministic future feature map μ and an uncertainty subspace S. The goal discrepancy e is hierarchically decomposed into e∥ and e⊥; only the unexplained residual e⊥ is used as the trajectory cost.

The Hierarchical Error Projection (HEP) module decomposes the prediction-goal discrepancy through a coarse-to-fine uncertainty subspace.

Conditional OOD scoring

A goal is treated as in-distribution when its discrepancy from the prediction lies primarily in plausible future-variation directions.

One forward pass

The model conducts score calculation in a single forward pass, avoiding stochastic future sampling and keeping trajectory scoring efficient.

DINO latent prediction

The backbone predicts dense DINOv3 features with compact latent tokens and an action-conditioned causal Transformer.

Benchmark

AirGoal-10k

AirGoal-10k is designed for large-scale 3D outdoor UAV image-goal navigation. It contains 9,000 training trajectories, 1,000 validation trajectories, and 1,000 test trajectories across diverse urban aerial navigation scenes.

Hugging Face Baidu Netdisk
AirGoal-10k trajectory distribution

Experiments

State-of-the-art ranking, planning, and online navigation

Offline ranking

1.09 ATE / 0.30 RPE at 32 candidates with 8.47 ms per frame.

Standalone planning

1.22 ATE / 0.33 RPE with CEM planning, outperforming deterministic and stochastic world-model baselines.

Online AirSim

76.0% SR and 64.5% SPL in closed-loop navigation.

Qualitative comparison of future predictions
Qualitative latent visualization: UA-NWM preserves global scene structure while modeling plausible local variation.

Deployment

Real-world UAV experiments

UA-NWM is deployed locally on a battery-powered MacBook Air connected to a custom-built quadrotor. At each step, it plans an 8-waypoint trajectory with CEM, executes the first waypoint, and replans from the latest observation. The system demonstrates zero-shot sim-to-real transfer without real-world fine-tuning.

UA-NWM real-world deployment system overview

Training & inference

Training & inference pipeline

The deterministic backbone predicts DINO latent futures, while HEP is trained on top of the frozen backbone and used at inference time to score unexplained residuals.

UA-NWM backbone training and inference pipeline

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},
}