Hybrid PBVS-IBVS Model Predictive Visual Servoing : For Autonomous Docking of a Free-Flying Robot

Sammanfattning

Interest in the development of autonomous rendezvous and docking (ARD) has been growing since the 1960s, as it is an important procedure for refueling spacecraft and transferring resources. Achieving centimeter-scale accuracy in ARD commonly relies on Visual Servoing (VS). Two common classical VS methods are Image-Based Visual Servoing (IBVS) and Position-Based Visual Servoing (PBVS) utilize a velocity-based control law that is straightforward, yet they suffer from depth or visibility sensitivity and local minima. This thesis presents a hybrid VS with Model Predictive Control (MPC) docking pipeline for a free-flyer that requires only an RGB-D camera and a Computer Aided Design (CAD) model of a docking station. The pose of the station is initialized by Globally optimal-ICP (GO-ICP) and refined with Generalized Iterative Closest Point (GICP). This estimation seeds a trajectory-tracking MPC that orients the free-flyer to maximize the visibility of geometric point features. A dynamic weighting strategy in the MPC cost is then used to control the soft-switching between PBVS and IBVS. Finally, the feature dynamics is described to the MPC to minimize the image errors while respecting system constraints. We evaluated the method in Robot Operating System (ROS) with Gazebo on a planar 3-degree-of-freedom (DOF) free-flyer and validated further at the KTH Space Robotics Laboratory (SRL), reporting docking success rate, steady-state pose error, and computation time. The results indicated reliable docking without AR tags, using only geometric fiducials, and demonstrated that the PBVS-IBVS transition improves visibility and mitigates local-minimum failures.