IEEE Robotics & Automation Magazine - December 2018 - 107

the KUKA omniMove) can carry whole
airplanes or concrete bridge parts that
are still in production; flexible manufacturing concepts such as a matrix body
shop no longer belong in the domain of
science fiction. Nevertheless, the real
challenges of the field started when the
robots were let out of the factories.
The U.S. Defense Advanced Research
Projects Agency (DARPA) Grand Challenge series is one of the most notable
autonomous vehicle contests of the
2000s, representing the first long-distance competition for driverless cars
and facilitating robotic development of
unmanned military ground vehicles.
The first two challenges took place in
the Mojave Desert; the third, the
DARPA Urban Challenge, required
navigation, negotiation, and scenariohandling in urban environments as
well (at an abandoned Air Force base in
Southern California). Whereas these
vehicles were custom-designed for the
specific environment and tasks, countless best-practices and yet-unknown
challenges have been published on the
topics of perception, decision making,
and high-speed vehicle control, some
that are still used as a useful guide in
self-driving development. The top vehicles were equipped with multiple lidar
sensors and high-precision global positioning system receivers, using camera
vision as secondary information-an
approach that was taken on by many of
today's research teams as a heritage
from DARPA Grand Challenges.
However, Stanley, the 2005 winner
from Stanford University, also used
machine learning techniques for obstacle detection, forecasting the new era of
self-driving car development. This was
also the first public appearance of these
technologies as a potential megatrend,
because the best car of the 2004 challenge, the CMU Red Team, could
accomplish only 5% of the race route;
in 2005, five teams completed the route.
This focused significant attention on
the domain, so, after 2007, DARPA
gave up with this series, because it
had become too easy for state-of-theart systems.
Meanwhile, Google was the first to
invest big on this technology, beginning

with the Stanford team behind Stanley.
Google began operation in its X division in 2009 with retrofitted Toyota
Prius models; in 2011, the state of
Nevada gave permits for the first public
road testing. It was Google's own developed self-driving car that claimed the
first fully driverless ride on public
roads in 2015. Google has kept its confidence in lidar; in late 2016, the technology was spun off to Waymo, which
promises to start offering public service very soon.
With the increasing computational
efficiency in image data processing and
sophisticated algorithms, vision is slowly taking over the role of primary sensors in self-driving cars. Many of today's
advanced functionalities, such as lane
keeping or pedestrian detection, are
solved by the use of cameras. In production vehicles, these functionalities rely
on traditional computer vision algorithms, whereas research communities
and progressive technology companies are migrating to deep learni ng (DL) techniques using neural
network (NN) architectures.
To define current capabilities of
autonomous vehicles and provide a
harmonized classification system and
supporting definitions, the Society of
Automotive Engineers defined six
levels of autonomy within the scope
of the standard Taxonomy and Definitions for Terms Related to On-Road
Motor Vehicle Automated Driving
Systems [3]:
● L0: no automation
● L1-driver assistance: specific functions under control
● L2-partial automation: combined
function automation (e.g., adaptive
cruise control)
● L3-conditional automation: automation of all critical functions with limitations (limited self-driving); the
driver must be able to take control at
all times
● L4-high automation: the vehicle
performs all driving tasks under certain conditions; the driver may take
control
● L5-full automation: the vehicle performs all driving tasks under all conditions; driver may take control.

Due to its consistency for industrial
practice and steplike progression through
the levels, this definition is widely accepted and applied by both the academic and
industrial communities. Levels 0-2 refer
to advanced driver assistance systems
(ADAS), where the human driver monitors and performs the dynamic driving
task (decision
making and
maneuver execuThe U.S. Defense
tion), whereas
Advanced Research
levels 3-5 require
the system to
Projects Agency
monitor the enviGrand Challenge
ronment actively
and carry out
series is one of
these tasks.
the most notable
Although
strict boundarautonomous vehicle
ies cannot be
contests of the 2000s.
defined from
the technological point of view,
as a rule of thumb, levels 0-2 can be
solved by traditional sensor-processing
algorithms with currently available
hardware processors and, thus, can be
found in most of today's production
vehicles. Levels 3-5, on the other hand,
refer to variants of hands-off driving,
where the use of advanced machine
learning and DL techniques is essential.
The complex task of coordinating the
sensing-planning-control chain requires a general, integrated approach,
where the individual development of
ADAS functions is just a small portion
of the big picture.
The human factor introduces another
issue: due to the relatively high level of
automation, drivers tend to become distracted and bored and look for other
activities. The main problem is that
humans are not efficient in long, inactive
monitoring-type tasks, and drivers usually overtrust the system [4].
Components of Self-Driving
Development
The road to fully autonomous, L5 selfdriving cars leads through a structured
development process, where one needs
adequate tools, algorithms, and processing hardware to test and deploy
these systems:

december 2018

*

IEEE ROBOTICS & AUTOMATION MAGAZINE

*

107
IEEE Robotics & Automation Magazine - December 2018

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - December 2018
