Mixture of Attractors: A novel Movement Primitive Representation for Learning Motor Skills from Demonstrations

Simon Manschitz, Michael Gienger, Jens Kober, Jan Peters

Abstract:

In this paper, we introduce Mixture of Attractors, a novel movement primitive representation which allows for learning complex object-relative movements. The movement primitive representation inherently supports multiple coordinate frames, enabling the system to generalize a skill to unseen object positions and orientations. In contrast to most other approaches, a skill is learned by solving a convex optimization problem Therefore, the quality of the skill does not depend on a good initial estimate of parameters. The resulting movements are automatically smooth and can be of arbitrary shape. The approach is evaluated and compared to other movement primitive representations on
data from the Omniglot handwriting data set and on real demonstrations of a handwriting task. The evaluations show that the presented approach outperforms other state-of-the-art concepts in terms of generalization capabilities and accuracy.