Reinforcement-Learning Control of a Hybrid Airship Using a High-Fidelity Digital Twin

This study introduces a high-fidelity digital twin and a learning-based control framework for a custom hybrid airship that exploits buoyancy, airfoil-generated lift, and rotor propulsion. The digital twin combines six-degree-of-freedom rigid-body motion with hull aerodynamics, rotor thrust, and actuator dynamics, providing a comprehensive, reproducible environment for design exploration. Within this virtual testbed, a reinforcement-learning autopilot is trained to steer the airship from arbitrary initial conditions to specified waypoints. Comparative experiments demonstrate that the learned policy achieves a root-mean-square position-tracking error below 3.5 m across multiple unseen missions. These results underscore the promise of data-driven methods for next-generation hybrid-lift vehicles.