Wind Estimation Using Gaussian Process Regression for Safe Drone Control

Drones are becoming increasingly important in our society. They are used for different purposes ranging from search and rescue to package delivery and entertainment. Reliability is an important aspect to consider for these real-world applications. One of the most significant challenges to drone stability and control is wind disturbances, which can affect flight accuracy, safety, and efficiency. Conventional control systems are often incapable of adapting to such external forces, leading to the risk of collisions and even system failures. This research aims to enhance the stability and robustness of the control system by integrating the traditional Model Predictive Control (MPC) algorithm with Gaussian Processes for real-time online wind predictions. In this thesis, I present a modified version of the conventional Gaussian Process Regression (GP), which uses past wind information in a sliding window fashion to estimate future disturbances. This probabilistic approach enables the construction of a data-driven model for the wind dynamics, capable of providing, apart from the simple predictions, also a confidence interval around such estimates. The non-parametric nature of the GP framework allows to easily include new pieces of information and remove less significant ones during the online phase. This also enables the model to capture complex nonlinear dynamics without the need to explicitly consider an underlying model for the wind. The predictions are included inside the MPC of the drone, which is then used to adjust the drone trajectory and control action based on the estimates given by the model. By incorporating the model's predictions, the drone dynamics are improved to account for external forces, thereby enhancing control effectiveness. Furthermore, the use of the uncertainty resulting from these estimates becomes a key component when considering the obstacle avoidance constraints: not only the system is able to properly avoid the obstacles, but the resulting uncertainty renders the controller more conservative when the model prediction is more uncertain, leveraging the use of chance-constrained optimization. This study validates using Gaussian Processes to model and predict wind disturbances for quadrotor navigation. This new control formulation is compared with the traditional MPC algorithm and an improved version that considers a constant baseline for wind disturbances. The algorithm’s effectiveness is evaluated in terms of real-time feasibility and tested against wind fields consisting of real-world collected data. The proposed algorithm is shown to perform well in windy and cluttered environments, being able to handle complex wind fields while avoiding excessive conservativeness.