Researchers from Hugging Face have published a paper detailing a new approach to aerial image-goal navigation, where an unmanned aerial vehicle (UAV) must reach a location specified by a goal image. The work, titled "Uncertainty-Aware World Model for Aerial Image-Goal Navigation," addresses a key limitation in existing world-model-based methods.
Current methods rank candidate trajectories using predicted future states, but they typically rely on only one or a few point predictions. This proves inadequate for large-scale outdoor environments, where future-state uncertainty is substantial. To overcome this, the team proposes the Uncertainty-Aware Navigation World Model (UA-NWM), an efficient latent world model that treats trajectory scoring as a conditional out-of-distribution (OOD) detection problem.
UA-NWM represents plausible futures within an uncertainty subspace and decomposes the discrepancy between prediction and goal into two components: one that is explainable by uncertainty, and another that is unexplainable. Only the unexplainable residual is used for scoring, enabling robust trajectory selection without the need for multiple future samples.
Extensive experiments show that UA-NWM consistently outperforms existing navigation world models while maintaining low inference latency. Real-world UAV deployments further validate its practical applicability. The project page, code, dataset, and model checkpoints are publicly available.