ETD

• manipulator tasks
• model predictive control (MPC
• non-linear optimization
• obstacle-aware
• shrinking horizon

Robotic manipulation in dynamic and collaborative environments requires the execution of motions within prescribed time constraints while simultaneously guaranteeing exact terminal states and collision avoidance. Although Model Predictive Control (MPC) provides an ideal framework for online reactivity, its application to high-degree-of-freedom manipulators introduces severe computational limitations. Furthermore, standard Receding Horizon approaches struggle to enforce strict temporal constraints on the terminal state.
To overcome these limitations, this thesis proposes a novel three-layer control architecture based on a Shrinking Horizon MPC (SH-MPC). Formulated in joint space using the jerk as control variable, the shrinking horizon mechanism dynamically reduces the prediction window to guarantee target attainment at the exact prescribed time with desired terminal velocities. Obstacle avoidance, for both static and dynamic obstacles, together with self-collision prevention, are natively integrated into the Optimal Control Problem (OCP) through a capsule-based geometric modeling. To bridge the gap between the computational responsibility of predictive optimization (~30 Hz) and the stringent stability requirements of the real robot, the architecture introduces a 1 kHz integration node. By analytically propagating the system dynamics forward from the optimal jerk computed by the solver, this node provides continuous references to the local controller (Computed Torque Control), thereby ensuring motion fluency and closed-loop stability.
The architecture is validated in high-fidelity simulation (ROS/Gazebo) and real experiments on a 7-DOF Franka Emika Panda manipulator, demonstrating real-time performance with average solving times below 30 ms.