Autonomous Vehicle Engineering - May 2021 - 3

The Navigator
Sensor Monoculture is a Terrible Idea
Monocultures in nature often lead to catastrophic ends. Infamous examples of this include
the Irish potato famine or the extinction of
bananas from fungus. The same is true in the
technology world. In the early 2000s, the dominance of the Microsoft Windows platform led
to a scourge of cybersecurity issues as hackers
attacked its vulnerabilities.
Biodiversity in both natural and digital ecosystems leads to greater resilience. It's is one of the
reasons automated driving systems (ADS) cannot
rely on any single sensing modality.
Almost everyone working on ADS agrees that
multiple sensor modalities, typically
at least three modes, are required
to generate a robust environment
model of the vehicle's surroundings. Cameras, radar and lidar
have become the baseline sensing
package for ADS, but most of the
leading companies in the sector are
adding more. Cruise and Zoox have
both incorporated infrared thermal
imaging, while many developers
are now including microphones.
High-definition maps are almost
universal in the AV space and C-V2X
is commonplace on Chinese AVs.
The need for sensor diversity
isn't new. More than two decades ago
when I was working on traction and
stability control systems - extremely
primitive precursors to current ADS efforts - this
lesson was learned early on. Vehicles are dynamic
and constantly moving and rotating along multiple
axes. Among other things, it became clear that
we needed multiple sensor modes to understand
what was happening when cornering on a banked
turn. Three-axis accelerometers provided a false
perception without also factoring in yaw rate and
vehicle speed. We had to create virtual signals from
different sensors to verify the direct measurements
and understand the ground truth.

While we used those sensors to measure
vehicle motions, today's ADS sensor suite is
measuring what other road users are doing.
However, the fundamental problem of verifying
what the system thinks it is seeing from any one
sensor remains.
As usual there is an outlier in this debate. Elon
Musk is insistent on trying to do Tesla's ADS on
the cheap, using cameras only. During a recent
exchange on Twitter to explain why his company
wasn't using the radar sensor as part of the " full
self-driving " system, Musk responded: " When radar
and vision disagree, which one do you believe?
Vision has much more precision, so
better to double down on vision than
do sensor fusion. "
The problem with this statement,
as any competent engineer that
remembers their lessons in significant
figures knows, is that precision and
accuracy are two distinctly different
things. Just because you have a
lot of pixels doesn't mean you can
accurately measure the range or
velocity of an object. If you have
only two signals and they disagree,
you cannot assume either is correct.
You must hand back control to the
human driver. This is why other ADS
developers use at least three diverse
signals. If at least two agree, you can
work from that.
Musk is also wrong in assuming that sensor
fusion is necessary. While most ADS use some
degree of fusion, Mobileye only fuses radar and
lidar. It uses that data to compare against the
cameras. But they also run the camera data through
two distinct algorithms to compare results.
Based on the videos that are regularly posted
of the Tesla FSD beta in action as well as comments
by company lawyers to California regulators, it is
clear that the 'camera monoculture' approach is
highly unlikely to be successful. ■

Almost
everyone
working on
ADS agrees that
multiple sensor
modalities
are required
to generate
a robust
environment
model of
the vehicle's
surroundings.

AUTONOMOUS VEHICLE ENGINEERING

Sam Abuelsamid
Principal Analyst,
eMobility
Guidehouse Insights
sam.abuelsamid@
guidehouse.com

May 2021 3

Autonomous Vehicle Engineering - May 2021

Table of Contents for the Digital Edition of Autonomous Vehicle Engineering - May 2021