Deep Reinforcement Learning-Based Swing Suppression Control for Unmanned Multirotor Transportation System

Unmanned multirotors have demonstrated excellent flexibility in complex real-world scenarios, making them promising for aerial payload transportation. However, it is challenging to achieve stable control of the unmanned multirotor transportation system due to its double under-actuated property. Traditional control methods are limited by model complexity and parameter tuning difficulty, resulting in deficient control capabilities. In response to these shortcomings, an end-to-end controller based on deep reinforcement learning is proposed in this article, ensuring that the payload arrives at the target while suppressing the process swing. Based on the Soft Actor Critic (SAC) algorithm and the design of state-action space, the efficiency of the controller is enhanced. By comparing with PID and Fuzzy PID control methods, the proposed approach presents superior payload swing suppression capacity with comparable position control performance.