Uncertainty-Aware World Model for Aerial Image-Goal Navigation
Deyi Zhu, Haoyu Fan, Yinan Zhu, Weichen Zhang, Shilin Ma, Xinlei Chen, Yansong Tang
cs.CV
2026-08-06
A latent world model splits the gap to the goal into explainable and unexplainable parts and scores only the latter in one pass; it reaches 76% closed-loop success at 2.7 s/step.
Aerial image-goal navigation gives a UAV a goal image and asks it to fly to that location. World-model methods score candidate trajectories by predicting their future states and picking the one closest to the goal. In large outdoor scenes the future is highly uncertain (occlusion, long-horizon drift), so one or a few point predictions are inadequate; stochastic world models that output a distribution are expensive because they require iterative sampling.
UA-NWM is a latent world model that uses a frozen DINOv3 ViT-B/16 feature extractor. It recasts trajectory scoring as conditional out-of-distribution (OOD) detection. For a candidate trajectory's future, it represents all plausible futures with an uncertainty subspace S around the predicted mean, then decomposes the prediction-goal discrepancy into a part within the subspace (explainable by uncertainty) and the residual (unexplainable). Only the residual is used for scoring: if the goal lies inside the plausible future distribution, the residual is small and the score is high. This yields a score in a single forward pass without sampling multiple futures. HEP (hierarchical error partition) performs this decomposition at multiple scales; the subspace rank is best at 2, since a higher rank absorbs true trajectory error into the explainable part and weakens scoring, while the finest scale contributes the most.
| Metric | UA-NWM | RAE-NWM | NWM |
| Ranking ATE (32 cand.) | 1.09 | 1.50 | 1.37 |
| Closed-loop success rate | 76.0% | 70.0% | 63.0% |
| Time per step | 2.70 s | 202.5 s | 191.7 s |
Closed-loop navigation in AirSim (100 episodes): 76.0% success rate, 64.5% SPL, 2.70 s per step, versus RAE-NWM 70.0% / 57.8% / 202.5 s and MWM 69.0% / 56.0% / 66.6 s. The best policy-based baseline (NoMaD) reaches 56.0%. Real-world deployment on a custom quadrotor succeeded across five tasks.
The reframing, scoring as OOD detection, is the core idea: instead of demanding accurate future prediction, judge whether the goal falls within the plausible futures. A single forward pass yields a score, which is why the per-step latency drops by two orders of magnitude (2.7 s versus 67 to 202 s). For online planning, where each step runs several CEM rounds of candidate sampling, that is the difference between real-time and not.
Evaluation is mostly on the authors' own AirGoal-10k (AirSim urban scenes); breadth across other simulators or real environments is limited. Success is defined as stopping within 20 m of the target, a lenient threshold. Planning still uses CEM (3 iterations, 32 candidates), and 2.7 s per step is still slow for real-time control. Requiring a goal image is inherent to the image-goal navigation paradigm.