Gaussian Process Regression-augmented Nonlinear Model Predictive Control for Quadrotor Object Grasping

Abstract

Grasping objects using an unmanned aerial vehicle (UAV) equipped with an onboard manipulator is far more flexible compared to ground mobile robots, since the UAV has a larger operation space. However, it still faces a great challenge, particularly when the dynamics of the UAV is not known exactly or the system is impacted by unknown external disturbances. To maintain a stable flight and grasp an object in the air precisely, a reliable control strategy is necessary. In this paper, we present a control framework based on nonlinear model predictive control (NMPC) combined with an augmented dynamics model employing Gaussian processes (GP) as a nonparametric regression model. Throughout the real-world experimental results, our proposed control framework eliminates at least 38% more of the trajectory tracking error in comparison to the NMPC with the nominal dynamics model alone, and it ensures a stable and dependable flight performance to grasp an object near the ground.