ParcelDrone: A Modular and Graph Neural Network-Based Decentralized Approach to Aerial Parcel Delivery

This paper introduces a decentralized control framework for modular aerial delivery systems, leveraging a Graph Neural Network-based controller to enable robust and adaptive flight across various parcel configurations. The system comprises homogeneous propeller modules that are flexibly attached around a parcel, with each module relying only on local state and 1-hop neighbor information communicated over a robust wired network. The proposed controller is trained offline to emulate the thrust differentials of an optimal centralized controller, enabling stable flight without requiring global state estimation or centralized coordination. Experimental evaluations on two parcel types (0.7 kg and 1.7 kg) with quadrotor, hexacopter, and octocopter configurations demonstrate reliable performance: hover tests achieve roll, pitch, yaw, and altitude standard deviations below 0.3 degrees, 0.3 degrees, 0.2 degrees, and 0.02 m, respectively; L-shaped trajectory tracking shows velocity errors below 0.1 m/s and path-following errors under 0.05 m. The predicted thrust norms exhibit mean squared error deviations of only 5%-10% relative to the optimal controller across all configurations, underscoring the high fidelity of the learned control policy. The proposed framework provides a scalable and robust solution for decentralized aerial delivery, addressing practical challenges in center-of-mass variation and asymmetric payload arrangements.