DMP-Based Path Planning for Model Predictive Interaction Control

This paper presents a three-layered hierarchical architecture to control manipulation tasks which involve in-teractions between the robot and workpieces. Learning from demonstration (LfD) is exploited to train dynamical movement primitives (DMPs) as fundamental building blocks for such tasks. Thus, position and wrench profiles are obtained from a kinesthetic demonstration, which makes the programming process intuitive for factory workers without expert knowledge. A model predictive path-following controller is used as the underlying control method, in order to make use of the explicit consideration of constraints within the controller formulation. Thereby, the progress parameter of the path-following control is used as the phase variable of the DMPs which results in a tight coupling of both methods. Finally, experimental results on a real robot system prove the effectiveness and real-time capable implementation of the approach.