Uncertainty-Aware Adaptive Multi-Objective UAV Path Planning via Fuzzy Priority Adaptation and Preference-Conditioned Reinforcement Learning

Conventional multi-objective UAV path planning assigns fixed relative priorities to competing objectives—path length, energy consumption, mission time, and safety—even though the conditions that determine what "optimal" should mean change during flight. This paper proposes a closed-loop planning framework in which measured environmental uncertainty and operational conditions dynamically determine objective priorities. An online uncertainty estimator aggregates localization, sensing, wind, and obstacle-prediction uncertainty from signals available to a real avionics stack, and a risk assessment stage translates these, together with the operational context, into collision and energy risk. An interpretable fuzzy inference engine then maps uncertainty, risk, battery state, and obstacle density to a time-varying preference vector over the mission objectives. The preference vector steers a single preference-conditioned deep reinforcement learning policy, trained with vector-valued rewards over the full preference space, so priorities can be adapted at every decision step without retraining. Simulation experiments under progressively escalating uncertainty show that fixed-priority planners degrade substantially as uncertainty increases, whereas the proposed adaptive mechanism preserves mission success and reduces collisions; ablation studies attribute this gain to per-step priority adaptation itself, rather than to the learning architecture or the fuzzy machinery. The framework is further demonstrated on a realistic beyond-visual-line-of-sight medical-delivery mission across the Saronic Gulf, using representative Greek airspace geofences and real wind forecasts, where the policy trained in the abstract environment transfers without retraining. The results support adaptive objective prioritization as a practical, interpretable layer between situation assessment and learned planners.