A Programming by Demonstration Approach for Robotic Manipulation with Model Predictive Interaction Control

This paper presents a programming by demonstration (PBD) approach for model predictive interaction control (MPIC). Therein, the user teaches the manipulation task including interactions multiple times. Then, the individual demonstrations are aligned using multi-dimensional dynamic time warping and modeled to retrieve nominal paths for the Cartesian pose and the interaction wrench. Based on this model, the reference profiles are segmented by a classification algorithm and a suitable parameterization of a hybrid force/motion control scheme is determined by solving an optimization problem. Finally, the overall task is saved as a sequence of manipulation primitives (MPs) and passed to the model predictive controller which is based on a path-following formulation. The effectiveness of the approach is shown by an experimental validation on a 7 degrees-of-freedom robot.