Extending Dynamic Movement Primitives towards High-Performance Robot Motion

Dynamic Movement Primitives (DMPs) provide a means for parameterizing point-to-point motion. They have become very popular in robotic imitation and reinforcement learning due to their linearity in the parameters describing motion, their inherent complexity reduction, and the ability to scale both in space and time. However, if DMPs are used to describe a motion that has been demonstrated by humans, the encoded trajectory is typically far from being time-optimal. In this paper, we extend the DMP framework towards time (sub) optimal execution of the path encoded in a DMP, bridging one of the gaps between the DMP framework and industrial applications. Time-optimality is in fact a key goal for minimizing cycle times and thereby maximizing throughput. The proposed approach is applied to simulations of a planar two degree of freedom manipulator and is experimentally verified on an ABB YuMi, a 7 degree-of-freedom manipulator.