Title:
Learn To Be Stable: Imitation Learning with Dynamical Systems
Abstract:
Imitation learning, which is the capability of learning new skills by imitating other people’s actions, is a big driver in the development of sensory-motor system in human beings. In robotics, imitation learning arises as a prominent approach to rapidly acquire new skills without explicitly programming them. This talk presents data-driven approaches to learn stable robotic skills from human demonstrations. The key idea is to represent the motion as a parameterized dynamical system and to learn the parameters while preserving the stability of the system. We then discuss several useful properties of the dynamical system formulation and show how to exploit them for constrained and reactive motion planning. The last part of the talk presents on-going research on learning stable systems evolving on Riemannian manifolds.
