Uncertainty-Guided Belief World Models for Long-Horizon UAV Navigation under Partial Observability and Visual Degradation

Abstract Long-horizon UAV navigation under partial observability and visual degradation requires reliable decisions from uncertain current beliefs and increasingly uncertain imagined futures. We propose the Uncertainty-Guided Belief World Model (UG-BWM), which exposes posterior covariance to the actor and critic, propagates ensemble-based predictive uncertainty through latent rollouts, penalizes uncertain and collision-prone imagined transitions, and triggers candidate-sequence replanning through a hysteretic uncertainty rule. We evaluate UG-BWM in Gazebo Harmonic with PX4 SITL across six locked-test conditions, three route lengths, four visual-degradation types, and unseen structural shifts. UG-BWM achieves a success rate of 0.63, a collision rate of 0.15, and an SPL of 0.62, compared with 0.54, 0.20, and 0.54 for DreamerV3-CU. Controlled tests show that posterior confidence, the combined predictive uncertainty score, and the state-dependent imagined objective each contribute to performance. Predictive uncertainty remains positively associated with rollout error beyond the training horizon. During deployment, uncertainty-triggered replanning matches always-plan performance while activating the planner on 18\% of control steps, reducing planner use by 82\%. The results support uncertainty-explicit belief modeling and selective planning for UAV navigation under degraded sensing.