IEEE Technology and Society Magazine - September 2020 - 15

But first, a bit of history: In his
account of the battle for automobile
safety in the Twentieth Century, historian Lee Vinsel describes the SAE
levels as an attempt at standardization [5]. The J3106 "Taxonomy and
Definitions for Terms Related to
Driving Automation Systems for OnRoad Motor Vehicles" was first published by the Society of Automotive
Engineers in 2014. By the time SAE
published its standard, there were
already two competing frameworks
for increasingly automated vehicles.
In 2013, the National Highway Traffic Safety Administration had drawn
up five levels for self-driving vehicles. NHTSA had called Level Zero
"no automation" and Level Four "full
automation," and focused primarily
on the role of the human operator in
carrying out "safety-critical control
functions" [6].
An expert group of the German
Federal Highway Research Institute
(BASt) had in 2013 described its
own levels for automated driving (as
opposed to automated vehicles),
from "driver only" to "full automation," with the latter involving an
automated detection of "system limits" and return to a minimal risk condition. In other words, there was no
limitless "full autonomy," in which
the system could operate competently in any environment such as a
licensed human driver can [7]. The
members of the SAE's task force,
after reviewing the BASt document
chose to largely adopt the BASt formulation with a few key changes [8].
The SAE made two interventions.
First, they added a sixth level, "Level
Five," above BASt's "Level Four."
SAE called their Level Four "High
Automation," and added" Full Automation" above it to disambiguate
the conditions under which such a
vehicle could operate. This also had
the impact of dividing NHTSA's Level
Four, which described a vehicle
capable of carrying out all safetySEPTEMBER 2020

Design Domain" (ODD), but othercritical tasks without oversight, in
wise responds to the same questwo. In an SAE Level Four system,
tion: under what conditions can the
the entire driving task is carried
system operate?
out autonomously, but with some
limitations on the environment in which the vehicle is
expected to operate. These
limitations are not specified
by the taxonomy, but include
Implying that progress means
such things as geofencing,
more autonomy does little to
controlled infrastructure, or
even weather-dependent
inform public decision-making
operation. A Level Five system by contrast is expected
about how automation provides
to perform under "all roadmeaningful benefits.
way and environmental conditions that can be managed
by a human driver" [9].
Looking further back, frameSecond, the SAE added more defworks for levels of automation in
initions and explanatory content.
transport did not begin with NHTSA,
They were careful to explain that it
SAE, or BASt. Some of the first
is the task of driving, not the vehicle
examples of the modern, numbered
itself, that is being automated. They
levels format come from the late
applied a precise definition of the
1990s, when Endsley and Kaber
driving task, which involved londeveloped a ten-level model for
gitudinal and lateral control funcautomation roles that was intended
tions, as well as monitoring of the
"to be applicable to a wide array of
environment and fallback perfordynamic process and automated
mance, although they equivocated
system control domains, specifically
on the "minimal risk condition" to
advanced manufacturing, teleoperawhich systems would be expected
tions, air traffic control, and aircraft
to retreat in the event of failure.
piloting" [10]. The authors drew on
The SAE approach gave the levels a
earlier attempts, such as Thomas
technologically-centered, and less
Sheridan's work defining roles of
ambiguous, set of descriptions than
humans and machines descripNHTSA had provided.
tively, rather than numerically.
The SAE levels have not been a
Sheridan proposed that machines
static document, though the overextend, relieve, back-up, or replace;
all structure of the levels has not
and human supervisors trust, comchanged substantially. Through
mand, plan, monitor, and intervene
two revisions, in 2016 and 2018,
[11]. The original role of Endley's
the document has tripled in length,
numerical taxonomy was to provide
integrated more descriptive examan ordering between more human
ples, especially around ambiguities
control on one side, and more comand edge cases in the levels, and
puter control on the other. It was
switched to increasingly specific,
not intended to encourage others to
technical language to describe all
aim only for high automation, but to
aspects of the taxonomy. "System
think about the options for appropricapability," which described which
"driving modes" could be handled
ate partitioning of tasks to achieve
by a system at a given level, has
a more reliable and functional jointbeen replaced with "Operational
human-machine system. In a recent
∕

IEEE TECHNOLOGY AND SOCIETY MAGAZINE

IEEE Technology and Society Magazine - September 2020

Table of Contents for the Digital Edition of IEEE Technology and Society Magazine - September 2020