Citation: Teodorescu, C.S.; Groves,
K.; Lennox, B. Learning-Based Shared
Control Using Gaussian Processes for
Obstacle Avoidance in Teleoperated
Robots. Robotics 2022, 11, 102.
https://doi.org/10.3390/
robotics11050102
Academic Editors: Shuai Li, Dechao
Chen, Mohammed Aquil Mirza,
Vasilios N. Katsikis, Dunhui Xiao and
Predrag Stanimirovi´c
Received: 27 July 2022
Accepted: 15 September 2022
Published: 21 September 2022
Publisher’s Note: MDPI stays neutral
with regard to jurisdictional claims in
published maps and institutional affil-
iations.
Copyright: © 2022 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
4.0/).
Article
Learning-Based Shared Control Using Gaussian Processes for
Obstacle Avoidance in Teleoperated Robots
Catalin Stefan Teodorescu * , Keir Groves and Barry Lennox
Depterment of Electrical and Electronic Engineering, The University of Manchester, Manchester M13 9PL, UK
* Correspondence: s.teodorescu@manchester.ac.uk; Tel.: +44-7-308-010180
Abstract:
Physically inspired models of the stochastic nature of the human-robot-environment inter-
action are generally difficult to derive from first principles, thus alternative data-driven approaches
are an attractive option. In this article, Gaussian process regression is used to model a safe stop
maneuver for a teleoperated robot. In the proposed approach, a limited number of discrete experi-
mental training data points are acquired to fit (or learn) a Gaussian process model, which is then used
to predict the evolution of the process over a desired continuous range (or domain). A confidence
measure for those predictions is used as a tuning parameter in a shared control algorithm, and it
is demonstrated that it can be used to assist a human operator by providing (low-level) obstacle
avoidance when they utilize the robot to carry out safety-critical tasks that involve remote navigation
using the robot. The algorithm is personalized in the sense that it can be tuned to match the specific
driving style of the person that is teleoperating the robot over a specific terrain. Experimental results
demonstrate that with the proposed shared controller enabled, the human operator is able to more
easily maneuver the robot in environments with (potentially dangerous) static obstacles, thus keeping
the robot safe and preserving the original state of the surroundings. The future evolution of this work
will be to apply this shared controller to mobile robots that are being deployed to inspect hazardous
nuclear environments, ensuring that they operate with increased safety.
Keywords:
Gaussian process regression; semi-autonomous vehicle; shared control; obstacle avoid-
ance; nuclear robotics
1. Introduction
Gaussian process regression (GPR), also referred to as Kriging [
1
] in geostatistics, is a
generic supervised learning method that is designed to solve regression problems. GPR
models can be used to predict the effect variables of a partially observed physical process
by building a continuous map with artificial data. Such maps have been used to model
spatial phenomena, including temperatures [
2
], magnetic field intensity [
3
], radiation
field [
4
], spatial localization of minerals in mining geology [
1
], and epidemic growth [
5
].
Key strengths of GPR modeling include the ability to use repeated observations, collected
at the same location, within the training data [
4
]; acceptance of both sparse and highly
clustered data [
4
]; data acquisition sampled at variable time and/or space [
1
]; ability to
learn an effective representation of nonlinear dynamics even using small data sets [
6
]; and
predictions consist of two types of information, including the mean function (corresponding
to the best estimation in average) and an associated confidence (expressed using the
standard deviation). Although often neglected when applying GPR, in this article, we make
explicit use of the confidence metrics.
Combining GPR with control systems is an emerging topic that has had increasing
interest in recent years [
7
,
8
]. This is a versatile method that has numerous applications
in robotics. From a broader perspective, in our work, we seek to design bespoke control
algorithms that are improved by integrating the stochastic nature of a real process (or plant)
into their design. In this respect, GPR has been demonstrated to be a highly effective
approach to modeling, and it is the technique utilized in this work.
Robotics 2022, 11, 102. https://doi.org/10.3390/robotics11050102 https://www.mdpi.com/journal/robotics
