Cooperative Model Predictive Control for Coordination of Orbital Robot Fleets

This thesis develops a distributed Model Predictive Control architecture for formation control of free-floating robotic platforms, deployed and tested on ESA's REACSA and MANTIS testbeds. Each agent solves a local receding-horizon quadratic program that holds a commanded standoff distance from a shared target while keeping clear of its neighbors, with the non-convex distance constraints replaced by separating hyperplanes and softened by penalized slacks, so that formation keeping and collision avoidance need no central coordinator. A world-frame reformulation removes the attitude from the optimization and keeps the problem linear time-invariant; a receding-horizon dynamic program maps the continuous wrench onto REACSA's on/off thrusters under their dwell-time limits; and an augmented-state Kalman filter estimates the arena's residual floor tilt online and feeds it forward. The controller supersedes an earlier potential-function design, settling roughly twice as fast and an order of magnitude more accurately across a matched simulation grid. A 221-run design of experiments fixed three deployment requirements before any hardware ran. The architecture was validated in a Gazebo campaign mirroring the real test cases and deployed as a ROS 2 package during a two-week campaign at the ESA Orbital Robotics Laboratory, yielding 229 recordings. On hardware, a single platform holds the commanded standoff to 0.009 m (MANTIS) and 0.029 m (REACSA), five chasers acquire formation from over a metre of initial error in 10 to 25 s, the formation is dragged rather than broken when the target is displaced by hand, and no run produced inter-agent contact. Holding implementation, weights, geometry and metrics fixed and varying only the plant, simulation predicts hardware to within a factor of 1.6 wherever at most one platform is physical and is optimistic by a factor of nine with two, a gap localized to the residual error of an individual real platform.