Uncertainty-Aware Vision-Based Autonomous Aerial Refueling

Abstract

Probe-and-drogue refueling (PDR) requires accurate terminal alignment of a receiver probe with an uninstrumented, dynamically moving drogue despite uncertainty in both relative-state estimation and closed-loop motion. Existing vision-based docking approaches provide relative-state estimates, but generally do not quantify estimation and operational uncertainty in a way that informs the control authority required for robust docking performance. We present an uncertainty-aware vision-based PDR docking architecture that fuses keypoint-based drogue measurements, differential GPS, and avionics in an extended Kalman filter. Split conformal calibration characterizes near-contact error in the fused probe-to-drogue displacement estimate, while a covariance-based operational uncertainty quantification designs position-dispersion limits along with corrective-command variance budgets under modeled handoff variability and disturbances. The resulting uncertainty statistics define geometric and command margins for a range-conditioned deterministic predictive controller. In high-fidelity FlightGear/JSBSim simulations with four controller variants and 50 runs per variant, the combined-margin controller achieved 45/50 centered basket-plane crossings versus 34/50 for nominal MPC. Median planar crossing error decreased from 0.231 to 0.175~$\mathrm{m}$ and the empirical 90th percentile from 0.403 to 0.286~$\mathrm{m}$, while the roll-command RMS increased from 0.22$^\circ$ to 0.48$^\circ$.