IEEE Systems, Man and Cybernetics Magazine - January 2020 - 43

initial success of the so-called intelligent space and ubiquitous computing concepts in robotics [1], several elaborated
implementations emerged, demonstrating distributed architectures supporting robotic research [2]. Component-based
software frameworks, such as the Robotic Operating System (ROS) [3], [4] and RT-Middleware [4], are good bases for
considering scalable cloud services; however, the original
designs of these middlewares do not consider the clouddriven use case. Regardless, there are some promising
efforts built around utilizing ROS as the backbone for cloudpowered setups [5], [6]. Among the numerous workarounds,
Rosbridge [7], [8] and RobotWebTools (roslibjs) [9] have
been proven to be the most propitious.
Considering the cloud computing paradigms, various
goals can be defined within cloud-related robotics research.
On the one hand, the robotics analogy of the Software as a
Service (SaaS) concept can be defined as Skill as a Service. On the other hand, let us consider an inverted scenario where the physical robots represent the infrastructure
with minimal onboard functionality in which those robots
are animated by the high-level control manifested in complex cloud services (like Skynet in the Terminator movies
[10]). This conceptual framework is depicted in Figure 1.
This article is aimed at finding a good balance between
onboard and cloud execution. After the general discussions, a real-world industrial example is reviewed that
uses cloud resources to deal with the unstructured environment effectively.

approach that puts everything onboard to the entirely
cloud-based robots where the physical device is only
responsible for sensing and actuation. Let us refer to this
latter case as a puppet robot.
Virtual assistants, like Amazon's Alexa or the Google
Home devices, are precursors of the puppet robotics era.
These units currently lack motion capabilities, but it is
easy to imagine the next level when digital assistant technology and the Internet of Skills move into mechanically
capable humanoids.
Architectural Challenges and Key
Technologies
Very few, if any, technologies exist that were specially tailored for cloud robotics. Nevertheless, many existing frameworks and techniques contribute to the field. In this section,
I mention some of these for reference, focusing on their
future perspectives.
Before reviewing the technological pieces, I will define
two distinct sets of cloud robotics services (skills) and
consider their contextual conditions.
◆ Context-free services: These services neither need
much information about the current state nor the internal specifics of the target application. Typical
examples are text-to-speech, speech-to-text, or even

b
ra
lne
Vu isk
R
m
co )
fo G
In ., 5
.g
(e

Motion
Interpolation
Safety
SLAM
Functions
Servo
Loops
Motion
HMI Planning

Required Data Rate

Onboard

ilit

Cloud
Vision-Based
Object Recognition
Speech
Processing

Acceptable Latency
Figure 2. The tradeoff between real-time requirements

Cloud Robotics
All Functions Are
Onboard

Puppet
Robots

and cloud execution. HMI: human-machine interface.

Conventional
Robotics

Balance of Onboard and Cloud Execution
In system design, engineers must decide where to execute
a given computation considering the practical aspects of a
highly complex system. Without the claim to completeness, onboard battery capacity, CPU performance, and
memory and storage limitations are the primary features
opposed to network latency/throughput and requirements
of real-time operation. Figure 2 provides a qualitative overview inspired by the RoboEarth project [11], [12], a pioneer
of cloud robotics, which made significant advancements in
the elaboration of the so-called train one, train all knowledge-sharing concept.
In an ideal situation (such as unlimited bandwidth and
zero latency), the curve could meet the vertical axis. One
focus of 5G mobile communication is to provide suitable
qualities even for applications that include fastidious forcebased interactions [13], bringing the so-called tactile Internet concept into reality [14]. The current situation is not as
bright: safety-critical functionalities and real-time features
are still running onboard, and there is no chance of changing it quickly even after the debut of 5G in 2020. However, a
very useful set of noncritical features, such as speech processing, mapping, and path/motion planning can be cloud
sourced. This is an obvious manifestation of the Internet of
Skills paradigm [13].
As Figure 3 suggests, the onboard versus cloud-sourced
spectrum spreads from the conventional industrial

Only Sensing and
Actuation Onboard

Figure 3. The extremities of robot control implemen-

tations on the spectrum of onboard versus cloud
execution.

Ja nu a r y 2020

IEEE SYSTEMS, MAN, & CYBERNETICS MAGAZINE

IEEE Systems, Man and Cybernetics Magazine - January 2020

Table of Contents for the Digital Edition of IEEE Systems, Man and Cybernetics Magazine - January 2020