SEDS-II: Generating Stable, Reactive, and Robust Robot Motions with Smooth Regression Techniques

We consider Dynamical Systems (DS)-based modeling of a class of robot motions which ends at a given target point, i.e. the so-called point-to-point motions [2], [1], [3]. Typical examples of such motions are reaching out for an object, closing fingers in a particular grasping configuration, stepping motion, etc. When modeling robot point-to-point motions with DS, ensuring stability of the learned DS (from a set of demonstrations of the task) is a key requirement to provide a useful control policy. There are numerous nonlinear regression techniques to estimate nonlinear DS. Each of these techniques has its own pros and cons which make their use very task-dependent. However, the majority of these techniques cannot be used directly to model DS-based robot motions because they do not ensure stability of DS. In this paper we present an extension to our previous work Stable Estimator of Dynamical Systems (SEDS) [2] so as to ensure stability of DS-based motions independently of the choice of the regression technique. Therefore the new approach, called SEDS-II, allows adopting the most appropriate technique based on the requirements of the task at hand without compromising DS stability. SEDS-II also provides the possibility of online learning and using a combination of two or more regression methods, which could be helpful to satisfy the requirements of more advanced robot tasks.