IEEE Robotics & Automation Magazine - June 2020 - 171

guidance and human-robot-team cooperation [20].
Wearable haptic systems can also be
an effective and unobtrusive solution to
reproduce a wider range of haptic cues
because they can be comfortably worn at
different body locations and stimulate the
skin locally by conveying to it different
types of touch stimulations (see [21]).
Regarding vision, in recent years, augmented reality (AR) has gained increased
attention, and different commercial systems are now available for AR and virtual
reality, e.g., Oculus Rift, Microsoft Hololens, and Google Tango. In AR, components of the digital world may be
superimposed upon or composed with
the real world and used in teleoperation
[22]. The composite scene can be displayed to the user, e.g., through a headmounted display, to improve situational
awareness in HRCs.
Shared Intelligence
Task-allocation schemes have been
developed to share the work between
human and robot depending on different factors such as capabilities, execution time, performance, and so on. A
hierarchical framework for task allocation was developed that assigns a full
sequence of atomic tasks based on the
capabilities of each agent [23]. Similarly, in [24], complex tasks are split
into basic subtasks and attributed
depending on their skills. The decision-making algorithms are often
based on a multiple-criteria approach
using a cost function, such as in [25],
and use Markov decisions to determine the best execution plan [26].
Other allocation algorithms have been
proposed to extend the task-assignment scheme to the multihuman-multirobot context [27]. The authors in
[26] and [27] also took ergonomics
into consideration when developing a
task-allocation scheme.
On the other hand, the principle of
path-planning algorithms is to generate
suitable trajectories for the robotic arm
using, e.g., cubic interpolation functions, spline functions, or coverage path
planning. Those solutions are often
only valid for static environments,
making them ill-suited for collaborative

Figure 2. Wearable and lightweight sensing and feedback mechanisms can improve the
usability and performance of HRC systems.

scenarios. Traditionally, the online
method for constrained handling is to
use closed-form laws such as potential
field methods [30] or antiwindup strategies [31]. The former is able to avoid collisions by generating a field of repulsive
forces around the obstacle but is typically
unable to deal with actuator saturation.
Model predictive control (MPC) is a general purpose control solution able used
to handle both state and input constraints in real time. This method is
based on the idea of solving at each time
instant a constrained optimal control
problem over a receding horizon. The
main disadvantage of this method is its
high computational load; although
recent advancements in computational
power have made it feasible to implement MPC on robots, MPC schemes are
not commonly used in mechatronic
applications due to their high computational cost and need for a precise model.
Therefore, methods such as the explicit
reference governor [32] can enforce both
state and input constraints without having to solve an online optimization problem so it can be computed real time.
When applying task-allocation
schemes, more criteria than the relative
performance can be considered. These
criteria may refer to reliability, the number of personnel, workload, or safety
[33]. Regardless of which criteria are
used to apply a function-allocation
scheme, work designers have to be
aware of the fact that the automation of
functions may introduce new work tasks
for the operator that are not directly
related to any single function [34].

HRC Flexibility
Cox Jr. [35] defines manufacturing flexibility as "the quickness and ease with
which plants can respond to changes in
market conditions." Hence, HRC flexibility is needed at two layers: to adapt to
the aforementioned manufacturing flexibility typical for Industry 4.0 (due,
e.g., to the variety of part shapes
and weights, each
The principle of
with a small batch
path-planning
size) and to adapt
to worker intenalgorithms is to
tions and comgenerate suitable
mands (which
may vary from
trajectories for
one person to the
the robotic arm
other). Several
specific manufacusing, e.g., cubic
turing applicainterpolation
tions are reported
in the research
functions, spline
literature, where
functions, or
c o b ot s h a v e
ad dressed the
coverage path
c o l laborative
planning.
assembly of a
homokinetic
mechanical joint
[36] and of cellular phones [37] among
several others. As of yet, all the aforementioned research works target specific applications, and it is rare to see a
cobot capable of addressing multiple
and diverse factory tasks. Ideally, such
cobots should be mobile, dexterous,
bimanual, and easily reprogrammable. A
platform developed with such flexibility
in mind is the mobile cobot Bimanual

JUNE 2020

IEEE ROBOTICS & AUTOMATION MAGAZINE

171

IEEE Robotics & Automation Magazine - June 2020

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - June 2020