IEEE Robotics & Automation Magazine - December 2018 - 110

tasks related to perception and localization based on visual data. These codes
are released under a license such that
both academic and commercial use are
supported. For
advanced algoSociety demands an
rithms utilizing
understanding of the NNs of any
type, countless
operation principles
frameworks
and consequences of are available,
providing prean increased number coded network
architectures as
of self-driving
building blocks,
vehicles on the road.
widely used
solvers and cost
functions, full
architecture examples, and references
to publicly available training data sets.
These frameworks (among the most
popular are Caffe, Theano, Torch, and
TensorFlow) generally support fully
connected, convolutional, and even
recurrent NN architectures.
Besides open-source software tools,
open standards are also emerging in
optimized hardware development.
Today, most NN tool kits, frameworks, and inference engines use proprietary formats to describe the trained
network architecture and parameters.

As a consequence, many proprietary
importers need not be constructed to
enable a trained network to be executed across multiple hardware platforms. The Khronos Neural Network
Exchange Format [20], originally initiated by AImotive, is designed to
simplify creating a network and running the trained network on other
tool kits or inference engines. The
standard was released in August 2018,
and it is expected to reduce deployment friction and create a bridge for
cross-platform DL tools, engines,
and applications.
Safety
Safety has been an intensively re searched topic for automated vehicles,
as the projected complexity of self-driving capabilities extends to traditional
safety assessment methods. Current
validation and verification tools include
safety standards applied to specific
components or driver assistance features of the car, assuming that vehicle
control is overseen by the human driver. Systems, such as collision mitigation
systems (CMSs) or lane-keep assist,
used to be tested and verified according
to established pipelines or processes,
such as the automotive standard

V-model or International Organization
for Standardization (ISO) 26262
[6]. However, in the case of CMS,
false-detection filtering and avoiding unintended braking are properties
addressed by the newly developed
Road Vehicles-Safety of the Intended
Functionality standard [7]. Highly automated systems will require full assessment of their performance and a safe
development pipeline to verify their
readiness for the cluttered human environment [8]. Simulation as an automotive tool for test and development is
recognized and discussed in detail in
some recent works.
One of the most important statements of the Vienna Convention (ratified in 1968 by 74 countries, excepting
the United States) is that the human
driver is liable for accidents caused
with the vehicle, taking full responsibility. This is primarily due to the lack
of existing test procedures for L2+ selfdriving systems.
Figure 1 presents a safety development workflow that rethinks traditional automotive product development
techniques regarding system design,
software and hardware design, and integration. Laws and standards serve as
boundary conditions/inputs for the

Laws and
Standards

Freeride Scenarios
Features

System and
Safety

Functional
Requirements

Functional Safety

Consistency
Technical Safety
Technical
Solutions

Development

Hardware and
Software
Realization

Fault-Injection
Scenarios

Verification Scenarios

Figure 1. An alternative safety-development workflow that better fits the extended requirements of highly autonomous cars.

110

IEEE ROBOTICS & AUTOMATION MAGAZINE

december 2018

IEEE Robotics & Automation Magazine - December 2018

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