Physics-informed sparse reinforcement learning for hybrid VTOL UAV control: HILS verification and tethered hover benchmarking

Hybrid Vertical Take-Off and Landing (VTOL) UAVs present significant challenges for control design, particularly due to nonlinear dynamics, actuator coupling, and complex hover-cruise transitions. Conventional controllers, such as PID and gain-scheduled MPC, often lack robustness to uncertainties, while recent reinforcement learning (RL) methods remain computationally demanding and offer limited interpretability, restricting their suitability for embedded platforms. To address these limitations, this paper introduces a physics-informed sparse RL framework that combines Sparse Identification of Nonlinear Dynamics (SINDy) with RL. Unlike black-box policies, the proposed approach constructs symbolic models of dynamics, reward, and policy, yielding interpretable controllers that are lightweight, data-efficient, and dynamically consistent. The framework is validated on a hybrid UAV with blended quadrotor-fixed-wing dynamics using high-fidelity simulations, hardware-in-the-loop testing, and tethered prototype bench experiments executed in real time on an NVIDIA Jetson platform (no free-flight experiments are reported in this work). Results demonstrate reliable trajectory tracking, robust mode transitions, and efficient real-time execution, confirming the potential of SINDy-RL as a scalable pathway for physics-aware autonomy in next-generation VTOL UAV systems.