A Learning-Based Shared Control Approach for Contact Tasks

This work presents a novel shared control architecture dedicated to teleoperated contact tasks. We use Learning from demonstration as a framework to learn a task model that encodes the desired motions, forces and stiffness profiles. Then, the learnt information is used by a Virtual Fixture (VF) to guide the human operator along a nominal task trajectory that captures the task dynamics, while simultaneously adapting the remote robot impedance. Furthermore, we provide haptic guidance in a human-aware manner. To that end, we propose a control law that eliminates time dependency and depends only on the current human state, inspired by the path and flow control formulations used in the exoskeleton literature (Duschau-Wicke et al. (2010), Martínez et al. (2019)). The proposed approach is validated in a user study where we test the guidance effect for the bilateral teleoperation of a drawing and a wiping task. The experimental results reveal a statistically significant improvement in several metrics, compared to teleoperation without guidance.

I. Introduction

Between full autonomy and direct teleoperation lying at the two ends of the spectrum, Shared Control (SC) has emerged as a viable solution, initially to exploit autonomy for overcoming large communication delays in teleoperation [3]. From thereon, SC became a fundamental concept in many scenarios where a human interacts with a robot, either in a collocated manner such as in cooperative manipulation tasks, or for remote interactions as in teleoperation [4]. The central idea in SC is that a human interacts with an Autonomous Agent (AA) that encodes prior knowledge, in order to achieve a desired task in a manner that attempts to combine the human cognitive abilities with the precision and repeatability of robotic execution, and with the aim to reduce the mental workload of the human and improve task execution.