A Meta-reinforcement Learning Framework for Adaptive Quadrotor UAV Attitude Control

This paper presents a meta-reinforcement learning approach for quadrotor UAV attitude control, addressing the limitations of traditional control methods in complex, dynamic, and uncertain environments. Our method integrates the advantages of context learning, Gaussian sampling, and multi-task training to rapidly adapt to diverse aircraft parameters and environmental conditions. Specifically, we utilize historical information quadruples as contextual data to accurately estimate current task characteristics through real-time observation and learning. In the latent space, we employ Gaussian sampling to effectively handle unknown UAV parameters. To validate the method’s efficacy, we conducted a series of experiments in the Bullet simulator, including single attitude tracking, resultant acceleration control, and continuous trajectory tracking. Experimental results demonstrate that our approach significantly outperforms traditional PID controllers and domain randomization-based SAC methods in terms of stability, rapid response, and trajectory tracking precision. Through multi-task training, our controller has developed the ability to swiftly adapt to novel situations. This research provides a robust and flexible solution for adaptive control of quadrotor UAVs, showing promise for more efficient and stable flight control in complex and dynamically changing environments.