Most of what we see about the upcoming self-driving car phenomenon comes from the industry. Press releases and contributed articles may disagree on timing and exact phase-in mechanisms, but they’re pretty unanimous in one respect: “This is happening.”
There are certain segments, of course, for whom self-driving cars will be money in the bank. Anyone who has to pay pesky humans out of their revenue stream can eliminate this cost and simply keep all the money. (Or improve the economics… more on that in a sec…) Hence truck-driving and Lyft-driving are viewed as careers that will disappear.
But what about cars for the rest of us? Will we like self-driving cars? Will we buy them? Will we need to buy a car at all? These questions are as yet unanswered, and speculation is all we can engage in by looking at unofficial comments about this pending radical shift.
And where might we find such commentary? Social media, of course, which has democratized messaging with something of an end-around, bypassing the officially controlled channels. I debated how to approach this angle – name and quote? The thing is, I do have something of a preference for observing without affecting, however, and my goal really isn’t to direct everyone’s eyes to any particular platform or thread. Rather, it’s to filter out some of the messages. So I decided to stay vague about where I’m seeing these things – even though that’s the opposite of citing sources.
Different platforms have different strengths. Twitter, for example, with its small message size, evanescence, and ease of automation, tends to be flooded with corporate messages. It’s not well suited to thoughtful conversation. (I know what you’re thinking… not going there.) Facebook and LinkedIn have nominally longer attention spans; they’re more conducive to an enlightening discussion.
My attention was drawn by notification emails I would get regarding new threads. And there was a decided alarmist tint to them – along the lines of, “How can we trust our lives to algorithms? What if they’re wrong?” (Not an actual quote, but paraphrasis.) So I dug in to see what horror scenarios were being created in the discussion.
Turns out that these alarms didn’t reflect the overall tone of that community (I’m not sure how the platform selects threads for emailing – by managers/owners or by algorithm). Yeah, there were a few such discussions, but not nearly what it seemed via email. So the first lesson was, “No, they’re not all freaking out.”
There were interesting things to dig out, however. And I should mention that these conversations involve people interested in the topic as well as practitioners from inside the industry, so you get some industry-informed content without the PR-managed messaging. Then again, because many of these voices don’t represent official positions, it’s entirely possible – even likely – that the ideas and concerns they present could go nowhere. One must also be cautious about extrapolating to the public at large, since these groups are self-selecting.
Still, those caveats notwithstanding, there are some nuggets. I took away four main themes from the discussions. Some are more obvious than others.
1. Should intermediate levels of automation be allowed?
Depending on whom you talk to, we may ease from unautomated to assisted to partially automated to fully automated. We looked at this before; I voiced the concern that, if there weren’t standards for handling and communicating automation hand-off, then people borrowing or renting unfamiliar cars might have a hard time figuring out what level of automation they were dealing with and when that automation would turn on or off.
But some are taking this discussion to a different level: there is a position that these intermediate automation levels should be used only internally for development– they shouldn’t be marketed until full automation is achieved.
The received wisdom is that the gradual release strategy will win out in order to get a jump on the return-on-investment aspect of this grand project. Car companies will be eager to commercialize, unless…
2. Are the car companies really interested in self-driving cars?
There are a couple of notions wrapped up in this question. The first is that of the automotive OEMs as constituting something of a cartel (no pun intended) – a closed club that’s being challenged by runny-nosed pipsqueak Silicon Valley upstarts (again, paraphrasis, with language strengthened to stir outrage and debate). Will they sit idly by while companies like Waymo threaten the cozy status quo? Or will they play along, acquiring technology and then slow-rolling its release?
Interestingly, Tesla straddles these two camps. It’s clearly in full-fledged production (even if targeting, historically, the moneyed class) – and yet it clearly represents a new, technology-driven approach to building cars.
So, Tesla aside, why wouldn’t the traditional OEMs be excited about all the new things technology will allow them to do? Well, for one, cost. Technology ain’t cheap – at least, to start with. So it goes in gradually, where clear value and up-selling is possible. (But, if they do too good a job, it may be mandated in future models, which they don’t like.) But there’s nothing new about this concern; it could apply just as easily to smart windshield wipers as it would to full automation.
No, there’s a more subtle possible concern. And, at this point, it bears repeating that this is not the OEMs saying they have a concern; it’s the voices of individuals who may have insights or may have conspiracy theories. Regardless, the ideas are interesting.
The issue here becomes clearer if we think about where our personal cars spend most of their time: parked. They are transportation, but, most of the time, for most of us, they’re not doing any transporting. Automated cars open up the possibility that we could forego buying personal cars and simply order up rides as we need them.
We can do this today, of course, by calling a cab or a ride-hailing service. In heavily urbanized areas like New York City, many people don’t own cars, and they use taxis all the time. In many other cities, however, taxis are few and expensive and slow to arrive – which explains the popularity of ride-hailing. But even as it gets easier to get a ride, personal cars typically stay home only when people go out to drink (so they don’t have to drive home) or when parking would be a nightmare. For everyday stuff, we use our own cars.
So, even if we use public transit and cars driven by others, most of us still own a car – parked in the driveway while we use these other modes. What if self-driving cars lowered the cost of a summoned vehicle (due to the lack of a compensated driver) to the point where they became ubiquitous and started to make it silly to own a car oneself?
At present, the math usually doesn’t support going carless (depending on your driving habits, of course). As one anecdote, when I was living in Santa Cruz, I started keeping track of my driving habits to see what might happen if I tried to go carless. My motivation was to see whether I could save money by eliminating the car, the fuel and the insurance (eliminating which, didn’t help a whole lot because I had some of the cheapest car insurance quotes for Michigan).
For those of you not familiar with that territory, Santa Cruz is almost attached to Silicon Valley – and yet it’s separated by a windy highway through the Santa Cruz Mountains. So it’s a tenuous part of the Bay Area, and yet many people don’t consider it part of the Bay Area. Point being, it’s almost part of a metro area, but not quite. A drive into San Jose – with no traffic (and there’s often traffic) – is an easy half hour or more.
I discovered that there were lots of local things I could do without a car. But, a few times a year, I’d want to go on a roadtrip somewhere. Say, 1500 miles of driving. Heck, I once drove to ESC Boston and back. In theory, I could rent a car for those rare occasions. But I quickly learned that the cost of renting a car a few times a year would completely wipe out any savings from not owning a car.
So the economic model would have to change pretty dramatically for going carless to pencil out outside a very few major cities. If that did happen, a tipping point might be crossed, leading to less individual car ownership.
For the auto industry, that would mean replacing the large number of cars sold per year, which mostly sit idle, with a smaller number of self-driving cars – in theory, just enough to keep them from sitting idle. That would be a significant drop is overall sales – something the OEMs wouldn’t want to rush to realize.
Of course, as with so many similar discussions, it bears noting that this analysis doesn’t apply well to rural areas. If it takes an hour to hail a ride, you’re going to want your own car. And if you need a vehicle to bring home a load of 2x4s, you’re probably not going to find that app on your phone; you’re going to own a pickup. My own guess is that car ownership would persist indefinitely outside metropolitan areas – but if that’s the entirety of the served market that OEMs can look forward to in the eventual future, it represents huge shrinkage.
So this motivates the theory that OEMs would like to a) maintain control over the future of the industry and b) slow-roll automation.
3. Ethical questions persist
As algorithms evolve, there’s a general feeling that they will always be at risk of facing a decision between two bad outcomes. The typical manifestation of this question is the so-called “trolley problem,” which evolves from similar decisions that a trolley operator might have to make if there are people ahead on the track, but also someone on the side track. So doing nothing or doing the only other possible thing – going onto the side track to avoid the people on the main track – means hurting one or more people.
With self-driving cars, the conundrum is more often expressed as tension between protecting the occupants of the car and protecting people outside the car. In this case, the choice might be to do nothing and mow down a crowd of people or to turn onto the only trajectory that includes no pedestrians– which takes you into a wall, which may injure the passengers.
This also gets to the notion of liability and how fault would be assessed (and monetized). And, as far as I can tell, there are no clean answers. It’s a perennial problem. It bears noting, however, that the trolley version has applied for as long as there have been trolleys, and it certainly hasn’t kept them from riding the rails.
4. Will we really give up our cars?
This final topic has come up in numerous places. Most of the self-driving car discussion focuses on the essential role of cars: transportation. But cars are more than just practical conveyance – especially in the US. Cars are synonymous with freedom. The ability to come and go exactly when desired, not when someone else’s schedule permits. The ability to go wherever you want – including dusty, bone-rattling, oil-pan-denting paths that are barely roads – and not be limited by some pre-determined route.
There are those who believe that we will never cede control of our ability to be where we want when we want to some other company. This is driven partly by culture, but also by what we’re used to. And it has potentially nothing to do with the economic calculations we looked at above. I could possibly go carless if the appropriate economics and convenience evolved out of a thriving, competitive market (real competition, that is, not the modern version of five brands all owned by one or two companies), but it would be a hard sell – and I’m not a cold-dead-hands type.
Will others relinquish their car keys under such conditions? If so, my guess is that it will take a couple of generations of changed expectations before we purge our identities of autophilia. And that might not be enough.
All four of these themes – and the many others intertwined or yet to be introduced – give no clear answers at the moment. But they do seem to be worthy of consideration as we rush headlong into the automotive future.
6 thoughts on “Self-Driving Cars: Unofficial Views”
What are your views on these and other possible concerns about self-driving cars?
Like most really bad public policy failures, this topic will certainly come down to “follow the money”. Most of the strong advocates, also have annual pay checks plus stock/performance incentives highly dependent on approval that are near, or more than, a million dollars. If they publicly state anything less than perfectly rosey for this technology, that pay check will vanish. The money is so great, key players purposefully, and openly, are willing to break, seriously bend, and outright ignore the law and oversights necessary to safely introduce this technology on public roads. One player has even established themselves as the compliance testing agent … what kind of oversight is that?
A decade ago I registered a team for the DARPA Grand Challenge, and we spent half a year working on the problem, before deciding there wasn’t a safe solution. I’ve followed this field closely since.
There are several false talking points that these industry people hide behind.
First false assertion is that self driving cars are safer than human drivers.
Possibly in a decade or two, but certainly NOT in the period of they they are advocating releasing these cars onto public streets.
The best possible cars that will be released, will have the sensor resolution of a nearly legally blind driver, and be using highly error prone “deep neural networks” that mimic the skill of a low IQ, rationally impaired minimal acceptable human driver.
There will be absolutely NO WAY to certify that a well trained “deep neural network” will recognize ALL simple hazard situations that a low quality human driver would. It will ONLY recognize a large portion of events, that are very close to training events, but never ALL, and never a significant portion of similar, and never events that are unlike the training events. Even poor, relatively unskilled, normal drivers, will have a better success rate.
The typical driver, and certainly skilled drivers, will perform significantly better on nearly all less common cases that are not in the training events for the neural network.
Expected sensor resolutions will be a small fraction of what the human eye and brain achieve. Because of this, where a human will see lights, warning lights and brake lights, these small lighted objects will become blurred, and non-detectable when they are at the intersection of four pixels … where the eye will clearly see and process them as clump of bright pixels.
What is a small child on/near the road, will be way too few pixels for the machine vision system to classify … where a typical human will clearly identify the object … maybe not a near legally blind person with permission to drive … but a normal driver.
Normal skilled drivers, and even many less skilled drivers, are aware of the vehicle operating points … sounds, and feeling of the road, and will tend to recognize these warning sign and prevent dangerous operations. Things like failing tires, failing wheel bearings, mushy brakes, and other human sensory inputs that are not even present in the most advanced self driving systems. Humans are strongly likely to significantly out perform neural networks on unexpected accident cases that can be fatal.
yes, the self driving cars will probably perform better than a fatigued driver in the normal cases … but the automation will most likely never even see/sense the many other operational failures that a fatigued human will. And many of those failures will unnecessarily kill when a robot car makes the wrong choice. Sure some fatigued drivers will be saved … but that is the class of accidents we already have high success rates with existing safety requirements … that can be improved more with advanced collision avoidance systems, and advanced monitoring of the human’s engagement levels on the road that indicate fatigue well before an accident. Advanced systems that are also in current high end cars.
The Tesla deaths were blamed on human faults … the system was not held accountable for the deaths they allowed with a driver that was not actively engaged in driving. The safety review said the system operated as designed, and therefore was not at fault.
In some ways it’s too bad these first Tesla deaths were just the driver/passenger. Had the Tesla failed to see humans in the roadway, and had killed a half dozen non-motorists … sure they could have again said the system operated as designed, and not at fault. But the public outrage against robot cars killing bystanders would have FORCED significantly higher operational testing standards to protect others that did not PURPOSEFULLY choose to be killed by a Tesla operator.
So far, other mfgrs are holding drivers to higher standards of engagement … including tracking eye movement and attention to the road.
I would like to propose that any executive team that wants to ship/sell/use self driving cars on public roads, should be willing to submit to a self driving darwin test, where a public sourced gauntlet of worst case tests are constructed for that specific platform.
The exec’s must agree to being occupants of the car, or external bystanders in unsafe locations, during this series of tests, including tests that have a high probability of the death of the exec’s.
If the exec’s are unwilling to submit to the test, then the product isn’t ready for public use and roads, where others (either riders or bystanders) may be killed by the robot cars.
If the exec’s car’s are not good enough, then we give the exec’s their Darwin award, and block the robot car from public streets.
It’s difficult to really identify where the customer pull for this technology comes from?
I will not want a self driving car until every employee at google/whoever is sending their children to school across town every morning in them.
At least when they have loved ones in the solution they will have “skin in the game” so to speak.
@lefty — I certainly agree … and at the same time I’m more concerned about the robot car driving over kids on the side of the road (side walk even) when the robot car vision loses it’s lane tracking due to water, blowing leaves, silt left on road from snow or runoff, hail, etc.
I’m pretty sure the regulators will let them kill a few high tech car owners, but the regulation and certifiable performance mandates will not get real until it starts killing innocent bystanders.
I’m pretty sure the first robot car that kills or cripples a cyclist will create a nation wide roar too.
Both children and cyclists have a small enough profile, to make the number of available pixels small enough that deep neural networks will have a very high probably of miss classification … especially when the shape, color, texture, and blending in with the background will cause extreme difficulty for accurate and safe deep neural network matching of these people/objects.
Cars in the lanes ahead are easy … kids and cyclists are not easy, and none of these matching algorithms will be near 100% accurate, in every possible configuration/location/clothing/etc. Thus the most critical object, may actually be outlier’s that were not properly trained for simply due to the lack of forethought in constructing the training data sets.
The problem, is that we are 10-20 years into the first 90% prototypes, with less than 10% of the work to solve the ENTIRE problem. History has shown that the maturity point in any complex product, is VERY far from the 10% point, where 90% of the goals are met … and in fact 90% of the work remains to complete the critical functionality of the last 10% of exception cases.
We are giving a lot of “kids” (young professionals with less than 10 years experience out of school), golden handcuffs (very high pay checks plus equity incentives) and asking them if they can finish the product in a few years … their lack of experience in “hard” problems is greatly magnified by their “educational” experience, where every “hard educational problem” is constructed by teachers to be finished inside a term/quarter/semester or two.
Will they risk their pay check and say “no — this is 20 years from completion” … I don’t think so … and I’ve watched a half dozen similar young teams grind way past first customer ship, still thinking they are just one or two bugs away from being done … another then, another one or two, with a horizon that they can not see past.
I’m pretty sure there are another few hundred senior engineers that have watched similar un-experienced expectation failures kill critical leading edge projects in their careers.
“Skin in the game” is a really important concept. It’s why I would not want to fly in a plane that doesn’t have a crew with skin in the game. If the crew is safely on the ground, operating the plane by remote control, then the decision calculus changes – and not necessarily to the benefit of the passengers.
The car situation isn’t completely parallel, but the high-level notion of “if you don’t get it right, it affects YOU” still holds, for the reasons you state.