AI Improves Quadrotor Flight Control in Turbulent Winds

Small multirotor aircraft often operate in challenging atmospheric conditions where turbulent winds can severely degrade their flight performance and overwhelm traditional control systems. This study introduces an innovative two-stage learning approach designed to enhance quadrotor control in such environments. The first stage involves a learned wind estimator, an attention-augmented gated recurrent network, which accurately determines local wind conditions using onboard kinematic and dynamic data. This estimator, trained on extensive simulated flights through complex turbulence, achieves high accuracy in recovering horizontal wind vectors and generalizes well to various flight profiles. In the second stage, a reinforcement learning (RL) flight controller utilizes the wind estimates from the first stage. This wind-aware controller dramatically reduces horizontal trajectory tracking errors, showing a 48% improvement over a baseline controller that does not account for wind. The benefits of wind perception become more pronounced in stronger winds, demonstrating the system's ability to maintain stable flight where conventional methods fail.