Generalization of discrete Compliant Movement Primitives

Abstract

This paper addresses the problem of achieving high robot compliance while maintaining low tracking error without the use of dynamical models. The proposed approach uses programing by demonstration to learn new task related compliant movement. The presented Compliant Movement Primitives are a combination of 1) position trajectories, gained through human demonstration and encoded as Dynamical Movement Primitives and 2) corresponding torque trajectories encoded as a linear combination of radial basis functions. A set of example Compliant Movement Primitives is used with statistical generalization in order to execute previously unexplored tasks inside the training space. The proposed control approach and generalization was evaluated with a discrete pick-and-place task on a Kuka LWR robot. The evaluation showed a major decrease in tracking error compared to a classic feedback approach and no significant rise in tracking error while using generalized Compliant Movement Primitives.