Robotic Trajectory Tracking: Position- and Force-Control

This thesis employs a bottom-up approach to develop robust and adaptive learning algorithms for trajectory tracking: position and torque control. In a first phase, the focus is put on the following of a freeform surface in a discontinuous manner. Next to resulting switching constraints, disturbances and uncertainties, the case of unknown robot models is addressed. In a second phase, once contact has been established between surface and end effector and the freeform path is followed, a desired force is applied. In order to react to changing circumstances, the manipulator needs to show the features of an intelligent agent, i.e. it needs to learn and adapt its behaviour based on a combination of a constant interaction with its environment and preprogramed goals or preferences. The robotic manipulator mimics the human behaviour based on bio-inspired algorithms. In this way it is taken advantage of the know-how and experience of human operators as their knowledge is translated in robot skills. A selection of promising concepts is explored, developed and combined to extend the application areas of robotic manipulators from monotonous, basic tasks in stiff environments to complex constrained processes. Conventional concepts (Sliding Mode Control, PID) are combined with bio-inspired learning (BELBIC, reinforcement based learning) for robust and adaptive control. Independence of robot parameters is guaranteed through approximated robot functions using a Neural Network with online update laws and model-free algorithms. The performance of the concepts is evaluated through simulations and experiments. In complex freeform trajectory tracking applications, excellent absolute mean position errors (<0.3 rad) are achieved. Position and torque control are combined in a parallel concept with minimized absolute mean torque errors (<0.1 Nm).