Reinforcement Learning-Based Control for Multi-Drone Systems

This study develops a reinforcement learning (RL)-based control framework for multi-drone systems, employing Multi-Agent Deep Deterministic Policy Gradient (MADDPG) and Multi-Agent Twin Delayed Deep Deterministic Policy Gradient (MATD3) algorithms for target-reaching and obstacle-avoidance tasks. Implemented in a Gazebo simulation environment with PX4-Autopilot integration, the system adopts a modular architecture for flexibility. Training occurs in a simplified 2D environment to optimize computational efficiency. Key findings include: MADDPG minimizes steps to reach targets but exhibits inconsistent formation stability due to overestimation bias. MATD3 enhances obstacle avoidance with faster convergence and improved stability. In the Gazebo environment, three drones maintain formation and avoid obstacles by switching between MADDPG and MATD3 based on a 2-meter obstacle proximity threshold. This framework demonstrates the feasibility of RL for multi-drone coordination, with future work targeting 3D environments and stochastic dynamic factors.