Performance Evaluation Method for UAV AI Pilot System with Uncertainty Injection

Aiming at the challenges in evaluating the performance boundaries of reinforcement learning algorithms, this paper proposes a performance evaluation method for intelligent decision-making algorithms based on uncertainty injection. The method systematically identifies sources of uncertainty in both input states and output actions. By employing a controlled variable approach, uncertainties of varying magnitudes are injected into the system. Based on the fault propagation pathway of “uncertainty factors - AI pilot system - aircraft model,” a three-tiered parameter monitoring framework is constructed, encompassing “sensor measurement parameters, algorithm process parameters, and vehicle macroscopic parameters.” This enables an in-depth analysis of the formation mechanisms underlying abnormal behaviors in UAV AI pilot system, clarifying the complete fault propagation chain from input uncertainty to output anomalies. Simulation results demonstrate that the proposed method can effectively evaluate the performance boundaries of UAV AI pilot system.