Model Predictive Motion Control based on Generalized Dynamical Movement Primitives

In this work, experimental data is used to estimate the free parameters of dynamical systems intended to model motion profiles for a robotic system. The corresponding regression problem is formed as a constrained non-linear least squares problem. In our method, motions are generated via embedded optimization by combining dynamical movement primitives in a locally optimal way at each time step. Based on this concept, we introduce a model predictive control scheme which allows generalization over multiple encoded behaviors depending on the current position in the state space, while leveraging the ability to explicitly account for state constraints to the fulfillment of additional tasks such as obstacle avoidance. We present a numerical evaluation of our approach and a preliminary verification by generating grasping motions for the anthropomorphic Shadow Robot hand/arm platform.