Uncertainty-Aware World Model for Aerial Image-Goal Navigation
UA-NWM scores UAV routes by separating goal mismatch from uncertainty it can explain.
The paper frames aerial image-goal navigation as conditional out-of-distribution detection. Its model represents plausible future states in an uncertainty subspace, then ignores the part of the prediction-goal gap that falls inside that subspace. The remaining residual is used to choose trajectories without sampling many possible futures. The authors report better performance than existing navigation world models, low inference latency, and validation in real UAV experiments. HF Daily Papers' note
The paper frames aerial image-goal navigation as conditional out-of-distribution detection. Its model represents plausible future states in an uncertainty subspace, then ignores the part of the prediction-goal gap that falls inside that subspace. The remaining residual is used to choose trajectories without sampling many possible futures. The authors report better performance than existing navigation world models, low inference latency, and validation in real UAV experiments. HF Daily Papers' note
score 4