Comparative analysis of a multi-layer deep reinforcement learning controller and an optimized neural network controller utilizing growing particle swarm optimization for quadrotor control

Abstract

The development of advanced control systems for quadrotors has been focused on recent researchs, particularly with the advent of intelligent control methodologies. This paper evaluates and compares two innovative approaches: (1) a Neural controller optimized using a Growing Particle Swarm Optimization (GPSO) algorithm, and (2) a layer-by-layer Deep Reinforcement Learning (DRL) controller. Method 1 leverages the GPSO algorithm to fine-tune the weights of the Neural controller without requiring prior training data, enabling efficient online optimization. It integrates a PD controller, designed using the Ziegler-Nichols method, which is further refined by an online PD-neural controller. This hybrid approach demonstrates high control precision with moderate computational demands. on the other hand, Method 2 employs a DRL based controller that structured in three layers included mapping and goal determination, path generation, and control of quadrotor dynamics. This approach adapts to dynamic environments through episodic task-based training which achieving high adaptability and control precision but at the cost of increased computational complexity. Finally, simulation results and practical experiments demonstrate the performance of the two methods across various scenarios. © The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature 2025.