IEEE Robotics & Automation Magazine - December 2022 - 45

user interface, such as classifying new objects or teaching
novel actions.
A possible solution for supervised learning by a human
tutor has been presented by Wyatt et al. [67]. However,
enabling robots to adapt to the environment requires new
techniques. On the one hand, as proposed by Wyatt et al. [67],
this technique could close knowledge gaps. On the other
hand, learning environment-adapted perception skills once
with a tutor will not avoid adjusting the perception within the
deployed environment.
The Lack of Commonsense Perception
Our third identified gap is the lack of commonsense perception
caused by specific recognition capabilities. Fed
by a high amount of training data, the artificial system
becomes an expert for narrow recognition tasks. The previous
arguments reveal several applications that hit the
performance for specialized tasks. Although the recognition
performance of a specific perception task can outperform
humans, the fundamental
recognition. Contrastively, robots benefit from scene-adjusted
sensor modalities and their flexible configuration. They can
overcome the limitations of human senses, such as through
ultrasonic or radar recognition techniques [113]. Robots can
be equipped with sensor modalities, such as radar, that
enable them to freely adapt to their environment. The flexibility
of the robot design allows the adaptation to the application.
Thus, robots can use sensors with the desired
amount, properties, and alignments. For instance, multiple
visual sensors could enable
robots to recognize the
scene in 360°. Herewith,
blind spots can be avoided,
which is especially important
for safe usage.
issue is the lack of
flexibility due to the required retraining or missing capability
to extend recognition. Therefore, even the latest
approaches achieve just low overall recognition performance
(see the " The Holistic Scope of the Latest Integrated
Approaches " section).
Humans have individual perception skills depending on
various influences, such as profession, culture, and age,
since humans learn from practice and experience [108].
However, in a society, there are perception skills simplifying
a commonsense understanding. This enables humans
to execute trivial tasks, such as knowing how to open a
door or how to use an elevator. Nowadays, trivial commonsense
perception skills, such as detecting a door and
using its handle to move through the door or detecting
the elevator buttons, are not default functionalities of
mobile robots. Thus, robot recognition is not making
sense of the whole scene. Similarly, the interpretation of
scene data indicates specific capabilities. There are a few
advanced approaches for interpreting a complex scene,
such as KnowRob [65] (see the " Knowledge Interpretation "
section). However, they provide narrow interpretations
independent of the available scene information,
whereas robots cannot instantly interpret an unseen
scene. Therefore, the knowledge interpretation may not
retrieve sufficient scene knowledge.
The Potentials of Robots
This section highlights the potentials in robot scene perception
going beyond the perceptual capabilities of humans.
Flexible Sensor Design
The first presented gap in robot scene perception (using few
or solely a single sensor modality) can be overcome by the
usage of flexible sensor design. Human perception is limited
to the recognized information of the senses and their range
and accuracy, e.g., fog or darkness negatively influences scene
Initial Perception
Capabilities
In contrast to the highlighted
gap of missing perceptual
learning, robots
can be deployed with prelearned
perception capabilities
and with initial scene knowledge. This trivial fact
allows reducing the setup time of the robot to a minimum.
For instance, mirroring the perception skills of an existing
robot in an environment enables a new robot to perceive the
scene equivalently. In contrast to robots, humans would have
to learn from scratch.
Our third identified gap is
the lack of commonsense
perception caused by
specific recognition
capabilities.
Cooperative Perception
Speaking the " same language " in terms of understanding and
exchanging information enables the sharing of perceptual
data independent of the robot, assistive system, or infrastructure.
Therefore, artificial systems have fewer restrictions than
humans as the number of collaborators and the communication
range and bandwidth are not hard restricted. The sharing
of scene information in real time with multiple artificial
agents enables the fusion of scene observations from various
perspectives. The data exchange enhances the scene knowledge
of every agent, which offers a comprehensive scene
perception from its scene overview. A popular research area
presents cooperative approaches that share navigation
data, e.g., to distribute the mapping [64] or for the reactive
path planning of multiple robots [114]. Moreover, possible
decentralized computation, e.g., cloud computing, can save
the resources of mobile robots [115]. Therefore, cooperative
perception is a key technique for reducing the setup time of
large environments and for providing the safe navigation
and better economy of larger robot fleets through its intercommunication
[114], [116].
Future Directions
We propose two comparatively nonpopular future directions
that contribute to deploying robots within our society. First,
we propose developing frameworks to combine multiple
DECEMBER 2022 * IEEE ROBOTICS & AUTOMATION MAGAZINE *
45
IEEE Robotics & Automation Magazine - December 2022

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - December 2022
