Geometric and Model Priors in Motion Primitives

Kurzfassung

When learning probabilistic policies from human demonstrations, data-efficient learning is a key requirement. Often, only few demonstrations or even only probabilistic via points are available for movement modeling. Probabilistic machine learning techniques such as Kernelized Movement Primitives (KMPs), Linear Quadratic Tracking (LQT) or Nadaraya-Watson kernel regression allow for modeling a rich set of motions with specific priors using scarce data. Traditionally, these methods are however only defined for Euclidean data. We show an extension to manifolds commonly used in robotics, allowing us to model full poses or other Riemannian manifolds. Each method induces distinct priors on the modeled primitives, resulting in different characteristics of the generated motions as seen in the evaluation.