Vision-based Approach and Landing with Distributed Sensing and Hierarchical Mixture of Experts

Advanced Air Mobility (AAM) operations face significant navigation challenges during approach and landing, particularly in dense urban environments where GPS may be degraded or unavailable. Prior work has developed an onboard Vision-based Approach and Landing System (VALS) and a ground-based Vision Distributed Sensing (VIDIS) network. VALS provides onboard state estimation by fusing Coplanar Pose from Orthography and Scaling with Iterations (COPOSIT) measurements with inertial data in an Extended Kalman Filter (EKF), while VIDIS leverages distributed ground-based cameras and bundle adjustment to estimate the incoming aircraft state. At longer ranges, when onboard detection of landing-zone features is limited, VIDIS serves as the primary navigation source; as the aircraft approaches the landing zone and onboard feature observability improves, VALS becomes increasingly dominant. This paper introduces a Hierarchical Mixture of Experts (HME) framework that dynamically combines VALS and VIDIS estimates based on their relative performance, enabling resilient navigation under both nominal and off-nominal conditions. Simulation results, including scenarios with injected navigation anomalies, demonstrate that the proposed HME architecture maintains continuous and accurate state estimation by adaptively prioritizing the most reliable navigation source. These results highlight the potential of combined airborne and ground-based navigation to enhance safety, robustness, and operational readiness for future AAM vertiport operations.