If you believe the hype now surrounding self-driving,
autonomous vehicles (AVs) you would think they will be widely available just
around the corner, certainly within the next few years. The automobile and tech
industries are abuzz with fantasies of AVs conquering the byways and changing
society from the tires up. The “go” green light shines brightly for them.
Sizeable financial bets have been placed on who among the AV
competitors may win this contest. Some think the traditional automakers can’t
succeed (because, uh, they’re “traditional”) and the high-tech Alphabet/Waymos
will win. No one has any real idea about the outcome, despite all the
breathless media attention from AV wannabes. All kinds of pronouncements regularly
erupt from the media; like this pro-AV story;
fewer critical reports like this one
also pop up.
But don’t believe the hyperventilation. At this point, I
vote to take the “auto” out of autonomous vehicles. I think the AV publicity
barrage’s singular goal is to raise the hundreds of millions of dollars needed
to further develop AV technologies rather than soon place actual,
fully-automated cars on the highways. Droves of fully-autonomous cars and
trucks will not appear on streets anytime soon.
Instead, I hope AVs will be taking many, many test drives on
circuitous routes in Arizona,
California, Michigan and elsewhere until they prove to the National Highway
Transportation Safety Administration, and more importantly to potential buyers,
they’re able to consistently move safely on our crowded byways. Autonomous
vehicles have a long, bumpy, non-GPS’d road to travel before that happens. To
be charitable, the wheeled techies’ acceleration to introduce software-driving
cars to the public is at best premature.
Autonomous vehicles face large challenges that advocates
mostly understate. Perhaps the largest is how unenthusiastic most drivers now
feel about using an AV. According to a Pew Research Center poll, 56% of Americans
are wary of driverless cars and would prefer not to ride in one. According to
the MIT Technology Review, “The
relationship between human and robot driver could be surprisingly
fraught.”
The recently-launched Las Vegas autonomous shuttle took just
an hour
before it got into an accident that displayed an overabundance of “A” (artificial)
and a lack of sufficient “I” (intelligence) – as in autonomous-driving’s touted
AI. Remember the accidents that killed Tesla drivers using Autopilot in Florida
and in China last year? What could possibly go wrong in the all too near future
(according to proponents), when herds of AVs start sharing the road with normal,
capricious human drivers? Among the obvious safety issues, these incidents have
raised important and unresolved liability concerns about injuries and
fatalities connected with AV accidents.
Nevertheless, folly or not, self-driving cars are the prospect
that nearly every major auto manufacturer now thinks it must bet on. Alongside
the billions of dollars being raised by AV proponents and manufacturers, a yellow
caution light should be blinking on AVs.
Safe, successful driving is a miracle
of neurobiological coordination that involves the flawless, simultaneous
execution of a billion tiny and subtle reflexes, as one expert conceded. It’s a
much more complicated task for a computer to properly drive an AV, than
teaching it to play chess or robotically assemble transmissions. A fully self-driving
car must correctly identify and label millions of objects, understand city
layouts and traffic laws and safely operate in a variety of road conditions. It
has to be taught to handle everyday driving hazards (high-speed merges) and
rarer incidents (objects in the road), as well as issues that would never
affect a human driver like a chunk of road debris that flies up and knocks out
a sensor.
To operate successfully, self-driving cars’ systems need to
approach driving not just a mechanical, reactive task but as a social act cooperating with all other
vehicles in the vicinity. Autonomous vehicles’ algorithms will need to understand
the norms that dictate driving’s acceptable customs. This is a tall order. As
anyone who has driven in different places knows, these norms and customs vary
considerably by locality. New York City drivers’ norms, which David Brooks characterizes
as “foreplay to genocide,” are very distinct from Seattle’s, “dawdling” drivers.
Thus, safe self-driving cars will need to recognize and
respond to how human drivers actually
behave on roads, not merely how they should from a legal/regulatory standpoint.
These technical and behavioral AI challenges facing AV systems’ development are
in part why most major automakers expect fully-AV cars to be available in no
less than 15 years from now, not the hyperbolic day after tomorrow vision of
Elon Musk.
There are six different “levels”
of automation for vehicle control systems, based on the system’s
sophistication; Level 0 through Level 5. A Level 0 vehicle has no automation,
like a 2005 Porsche 987 or a 2017 Chevy Malibu. Level 5 – Full Automation – rules
the other, driverless end of the automation scale. A Level-5 car will operate
without a human driver on any road and in any conditions that a human driver
could negotiate. After the driver enters the destination into the vehicle’s GPS-based
map, he/she can sit back and relax with no further active involvement needed
during the journey.
No AVs now operate anywhere close to Level 5, but Waymo—formerly
Google’s driverless-car project, now a separate division of Alphabet—is using a
fleet of 600 Chrysler Pacifica hybrids to develop its AV technology for future production,
as shown below. In this fully-automated level, no human control of a vehicle is
needed at all, at any time. Level 5 vehicles won't require any pedals, steering
wheels, or controls for a human to take charge. It’s the Jetsons car travelling on properly-augmented
roadways.
Waymo AV
test vehicle
Real-world vehicle systems are now available that can
operate at Levels 1 and 2. Level 1 is
Driver Assistance where the car can control either the steering or the speed,
but not both at the same time. The driver performs all other aspects and has
full responsibility for the vehicle at all times. An example of a Level 1
technology is Adaptive Cruise Control now available in some automobiles. Level 2 is Partial Automation, where the
car can steer, accelerate and break in certain circumstances. The driver
remains responsible for virtually all “tactical maneuvers” (e.g., responding to
traffic signals and changing lanes). Audi, Cadillac, Mercedes, Nissan, Tesla
and Volvo have such partial, semi-automation systems in certain cars now. Some
knowledgeable people believe Tesla’s revised Autopilot system can operate at certain times as a
Level 2 system.
At this point, there are no AVs that can operate at Level 3 (Conditional Automation), Level 4 (High Automation) or Level 5 (Full Automation). In a Level 4
car the driver might manage all driving duties on surface streets then become a
passenger as the car enters a highway. Ford expects to put Level 4 AVs –
actually a fleet
of autonomous, self-driving taxis (aka, robotaxis) – operating on city streets
by 2021. Ford says it hopes to have a "high-volume, fully autonomous"
car working commercially by offering a "ride-hailing or ride- sharing
service." This goal sounds like Ford is operating very close to the
bleeding edge with its three-year timeline. Waymo is making similar plans. I’m
betting that Lyft drivers need not worry in 2021.
Finally, there’s the crucial issue of an AV’s cost that usually
goes unmentioned, probably because technical and safety questions have
dominated discussions so far. Nevertheless, if AVs are to occupy space on
American roads, their purchase price will need to pass interested customers’
value proposition. That is going to be a challenge over the short-run, just
like it has been for electric vehicles.
The first-generation Google/Waymo AVs were Priuses laden
with $150,000 worth of lidar (LIght
Detection And Ranging) and radar sensors and related AV equipment. An
AV’s lidar sensor (the bulky attachment on the car’s roof, shown in the picture
above) is the single most expensive additional requirement. This sensor bounces
multiple laser beams off nearby objects all around the AV to create accurate,
real-time 3-D maps of their surroundings. The previous generation lidar sensor,
built by a leading manufacturer and used by Waymo that was never mass produced,
cost $75,000 to $80,000 per unit. Waymo now builds its own lidar and related AV
equipment. A second-generation (pre-production) lidar sensor uses 128 laser
beams – twice as many as previously – with an effective range of possibly 300m
(more than twice as long as before). Such a sensor could cost “thousands of
dollars” when it enters mass production, although it is not clear when that
will happen. Even if eventually newer, improved lidars’ cost can be reduced by 90% as the market expands –
an often-cited, but never documented statistic by AV techies –that would
translate to a still-pricy $7,500/unit, which represents slightly over 20% of
the average price
of an entire 2017 car.
These sizeable incremental costs of driverless Level 4 and 5
AVs are a likely the reason Ford, GM, Waymo and other AV technology players are
aiming first at commercially-owned AV vehicles, including Uber (who has heavily
invested in its own AV technology, and faces a lawsuit
with Waymo who claims Uber stole its technology), Lyft (who signed a $500
million partnership
with GM last year to use self-driving Chevy Bolt EVs, and recently also signed
an agreement
with Waymo) and perhaps even good ol’ taxi and rental-car fleets. Notice that at
the AV party, everyone is now dancing with practically everyone else.
A lidar manufacturer’s CEO stated
that some ride-sharing companies have told him, “if you had an autonomous car
that just worked, they’d be willing to purchase these cars for $300,000 to
$400,000 apiece and buy as many as you could possibly make.” He said, “that’s
because the total vehicle cost—including the price of sensors—is less important
than it is in a consumer-owned car. This steep investment in AVs could be
recouped quickly by keeping a vehicle on the road nearly 24 hours a day.”
The AV hype blows from many directions, including CEO
offices. His rationale is fiscal fantasy. Profitably operating such
super-expensive Lyft/Uber/taxi AVs by running them continuously would increase
the cars’ operations and maintenance costs and lead to lower operational
lifetimes. No vehicle, including steeply-priced ones with lidar, can
continuously run “nearly 24 hours a day” without costly mechanical
consequences. Even without such cost consequences, Lyft/Uber AVs’ high costs could
result in folks who take an AV Lyft having to pay a super-premium “AV all-hours
surge” price that I doubt would be popular, except as an extravagance. Thank
goodness for hedge fund execs’ expense accounts and high-school proms (with
parental indulgence)?
If AV manufacturers and stakeholders want their vehicles to
become our future vernacular means of transport, they will need to satisfy two
goals. First, the complex AV technology systems will need to be proven safe and
reliable. Given the existing, remarkably low accident rates for regular
automobiles (1.27 deaths and 78 injuries per 100 million miles driven[1] in
2016), AVs would need to be driven hundreds of millions of miles before being
able to demonstrate their relative safety and reliability. Waymo’s AVs have driven
far more than any other player, a bit over 630,000 miles. According to Nidhi
Kalra, a researcher at RAND, there is no practical means at present for testing
the safety of AVs before their use becomes common. There have simply not been
enough AV miles driven to calculate a defensible safety estimate.
However, she also believes that waiting for the “perfect AV”
may cost
lives. I am not suggesting that we wait for the “perfect” AV. She’s right;
perfection always takes a lot longer and hasn’t been achieved in a century by
human-driven cars. My caution about introducing AVs is founded on simply
requiring them to be comparably safe
and reliable to that of existing driver-controlled cars and trucks, not perfect.
AV safety depends on: lidar sensors’ accuracy and reliability, the AV code algorithms’
degree of optimization, precision and predictability and the entire system’s flawless
high performance.
The second, related goal is the cost of AV cars must be
reduced if their appeal will reach beyond rich aficionados and wealthy ride-sharing
firms. Cost reductions may likely occur as more AVs are built and their
production process becomes “scalable” (a favorite term of techies, including the
AV clan), just like they have for many other new technologies, from
photovoltaic solar panels to lithium-ion batteries. Li-ion batteries’ cost
dropped by a factor
of 4 over the past seven years. But, will such scalability and production
efficiency apply to AV technologies? AV proponents hope so.
Like other new automotive technologies from the past (e.g.,
air-conditioning, automatic brake systems), AV capabilities will first be sold in
higher-priced luxury cars, as companies already plan. In the next decade or so
if AV manufacturers want to sell to the far broader “regular” market of
individual buyers like Jane and Joe Van and capture larger production scale
economies, prices will need to be value-comparable with non-AV transport. This
value comparability would include some added premium based on the hopeful
higher worth of being a driverless car. I doubt the acceptable AV premium will
exceed 10% for non-luxury cars. Attaining these lower prices will pose a key challenge
for AV component manufacturers. Will they be able to sufficiently reduce their
sensors’ and AV algorithms’ unit prices at the same time as manufacturers
undertake costly redesign and miniaturization?
Discussions of AVs’ future rarely mention the ancillary
local, state and federal government expenditures needed to upgrade streets and
highways to be Level 4 or 5 AV-compatible. Forget about just filling pot holes,
also dump the Dots as part of AV infrastructure enhancement. One unexpected
change to California highways now being undertaken to help AVs is Caltrans’ on-going
removal of Botts Dots lane-markers. According to a Sacramento Bee story, “After more than a half
century of service, the Dots are expected to be relieved of duty because they
are a bad fit as a lane marker in the emerging world of AVs that rely on sensors
to ‘read’ and understand lane lines.” Such public spending will take money,
effort and time.
In a previous blog
about electric vehicles (EVs) I mentioned that despite nearly a decade of
marketing, incentives and subsidies, the far simpler-to-produce EVs’ national market
share remains dismal – less than 1%. Will herds of AVs be traveling on our
future streets and highways? Perhaps, and it’s unlikely to be anytime soon. As
the philosopher Yogi Berra once remarked, the future ain’t what it used to be.
Regarding AVs, my fingers are cautiously crossed that this technology will
serve, not master us.
[1]
Total motor vehicle deaths (or traffic fatalities) are far more often reported,
but are less meaningful. In 2016, 40,200 people died
in motor vehicle accidents according to the National Safety Council, a 6%
increase from 2015. The vehicle death rate
(deaths per 100 million miles driven) provides a more important, proper perspective
than the absolute number of fatalities. The 2016 level of fatalities occurred
as US driving rose to 3.2 trillion vehicle
miles travelled.
