Abstract:
Unmanned Aerial Vehicles (UAVs), particularly quadrotors, have become highly versatile platforms for various applications
and missions. In this study, the employment of Multi-Agent Reinforcement Learning (MARL) in quadrotor control systems is investigated,
expanding its conventional usage beyond multi-UAV path planning and obstacle avoidance tasks. While traditional single-agent control
techniques face limitations in effectively managing the coupled dynamics associated with attitude control, especially when exposed to
complex scenarios and trajectories, this paper presents a novel method to enhance the adaptability and generalization capabilities of
Reinforcement Learning (RL) low-level control agents in quadrotors. We propose a framework consisting of collaborative MARL to
control the Roll, Pitch, and Yaw of the quadrotor, aiming to stabilize the system and efficiently track various predefined trajectories.
Along with the overall system architecture of the MARL-based attitude control system, we elucidate the training framework, collaborative
interactions among agents, neural network structures, and reward functions implemented. While experimental validation is pending,
theoretical analyses and simulations illustrate the envisioned benefits of employing MARL for quadrotor control in terms of stability,
responsiveness, and adaptability. Central to our approach is the employment of multiple actor-critic algorithms within the proposed
control architecture, and through a comparative study, we evaluate the performance of the advocated technique against a single-agent
RL controller and established linear and nonlinear methodologies, including Proportional-Integral-Derivative (PID) and Backstepping
control, highlighting the advantages of collaborative intelligence in enhancing quadrotor control in complex environments.
