Robots learning from robots: A proof of concept study for co-manipulation tasks

Abstract

In this paper we study the concept of robots learning from collaboration with skilled robots. The advantage of this concept is that the human involvement is reduced, while the skill can be propagated faster among the robots performing similar collaborative tasks or the ones being executed in hostile environments. The expert robot initially obtains the skill through the observation of, and physical collaboration with the human. We present a novel approach to how a novice robot can learn the specifics of the co-manipulation task from the physical interaction with an expert robot. The method consists of a multi-stage learning process that can gradually learn the appropriate motion and impedance behaviour under given task conditions. The trajectories are encoded with Dynamical Movement Primitives and learnt by Locally Weighted Regression, while their phase is estimated by adaptive oscillators. The learnt trajectories are replicated by a hybrid force/impedance controller. To validate the proposed approach we performed experiments on two robots learning and executing a challenging co-manipulation task.