Learning Task Priorities from Demonstrations

Abstract

As humanoid robots become increasingly popular, learning and control algorithms must take into account the new constraints and challenges inherent to these platforms, if we aim to fully exploit their potential. One of the most prominent of such aspects is their bimanual structure. Most research on learning bimanual skills has focused on the coordination between end-effectors, exploiting operational space formulations. However, motion patterns in bimanual scenarios are not exclusive to operational space, also occurring at the joint level. Moreover, bimanual operation offers the possibility to carry out more than one manipulation task at the same time, which in turn introduces the problem of task prioritization. Here we address the aforementioned problems from a robot Learning from Demonstration perspective. In particular, we present an extension of the Task-Parameterized Gaussian Mixture Model (TP-GMM) employing operators that allow for tackling such problems. The presented approach – despite the focus on bimanual operation – can be applied in any scenario where the prioritization between potentially conflicting tasks needs to be learned. We evaluate the proposed framework in: (i) two different bimanual tasks with the COMAN and WALK-MAN humanoids that either require the consideration of operational and configuration space movements or the prioritization of tasks and (ii) a loco-manipulation scenario with the Centauro robot in simulation, where the priority of the floating base position needs to be learned, showing that the approach can be exploited in generic task prioritization scenarios.