AVs After Alphabet

Word on the street is that Alphabet (née Google) is looking for some revenue from its Automated Vehicle unit, and absent that, might be getting bored with the whole endeavor. Google has gotten bored with things before, and one can imagine it doesn’t view itself as having an infinite pocket to pay for interesting things. There have been reports of this before.

One of the great questions is how long do you tolerate losses in an operation before you cut your losses and convince yourself there will never be gains. Many projects are undertaken (ventures are funded) with great promise.

Yet technological change often happens one bankruptcy at a time, where first movers borrow heavily to invest in new systems with unproven demand. This creates second mover advantages. This is a danger for initial massive investment in self-driving cars in the near future, especially as there doesn’t seem to be any demand from the public currently. (Recognizing that as Henry Ford never said (but almost certainly felt) “If I had asked people what they wanted, they would have said faster horses.”)

In the autonomous vehicle market, suppose someone claims there will be gains in excess of losses, and someone else, with access to a checkbook, believes them. Now we all understand Expected Value depends both on the Potential and the Probability of that Potential being realized. In practice both of those are uncertain and dynamic. Today’s potential in today’s market conditions differ from tomorrow’s. Today’s probabilities also differ from tomorrow’s.* At what point do you recognize the potential as lower than claimed, and the probability as lower still? When, in Bayesian terms, do you update your priors?

In Alphabet/Google’s case, I think the potential of being the first mover is lower because there are now many more people serious about AVs since Google kick-started the latest wave of development of self-driving cars, following upon the DARPA Urban Challenge. In contrast, the likelihood of success might be higher, since there is no more proof-of-concept.

Unless Alphabet imagines itself actually manufacturing cars (probably in a contract facility),  one assumes either it hopes to get revenue from (a) more eyeballs on Google-ad-powered screens (phones, dashboard, and especially heads-up displays)  in a world of AVs, (b) cutting a licensing deal with manufacturers, or (c) providing mobility services directly a la Uber-type of business.

But in a sense they have been overtaken. For services, all major automakers have now partnered with some other company to explore ride-hailing. Toyota and Volkswagen are the latest to announce this (following historically on similar patterns where automakers invested in rental car companies (e.g. Hertz by GM 1925-53 and Ford 1987-2005, Avis by GM 1989 – 96, Dollar Thrifty by Chrysler 1990-97). Daimler owns car-sharing service company Car2Go and Moovel and is certainly looking at this market. Google rival Apple invested in Chinese hailer Didi.

For autonomous technology, Tesla has cars on the road already.While Tesla Autopilot is not the equivalent of Google cars, as it still requires driver attention, it is also a real product accumulating real experience at an accelerating rate, as shown in the graph, with two orders of magnitude more distance traveled and seemingly fewer incidents per distance traveled (I am going on press-reports of Google and Tesla crashes).

Trends in Autonomous Vehicle Usage, note this is a Log scale.
Trends in Autonomous Vehicle Usage, note this is a Log scale.

If the logic of machine learning is right, it will get better and better over time. There have apparently yet to be serious crashes with 100 million miles of Autopilot experience (though there have been a few issues, there have also been claims of lives saved), so perhaps this is the right technological trajectory.

This is the incrementalist approach, leaving the steering wheel and brakes facing the driver, which counters the more radical Google approach of removing (almost) all control from the driver, and trusting the machine completely. The risk has always been the transfer, when the car tells the driver to take over if the driver isn’t ready. But if that issue is small compared to the general safety benefit of letting the machine do the steering, accelerating, and braking, it is a risk worth taking.

So will Google close its unit and set its code and data free,  let it wither away as staff start up spinoffs like Otto, or sell it outright to a manufacturer if they choose not to pursue it with gusto? Or, perhaps the rush to dominate the initial market for AVs will lead to a unsustainable bubble that hits everybody’s valuation (not just Alphabet’s).

I expect there is value to the unit collectively above and beyond the value of the individuals (though I have no personal knowledge), so one hopes they keep the team together, under their ownership, or another patient patron.


* In more typical infrastructure terms, when someone comes and says the Northstar Line will have a Benefit Cost ratio of 2, but it turns out to have a B/C ratio of 0.15, how long do you keep pouring good money after bad? At what point do you recognize the potential as lower than claimed, and the probability as lower still? Sacramento, for instance, is considering shuttering a disappointing LRT line.

Riding in a Tesla with AutoPilot (2015)

In the Fall of 2015, the electric vehicle maker Tesla remotely upgraded its most recent model year cars (about 50,000 vehicles) with “Auto-Pilot”, making them semi-autonomous (according the NHTSA scale, late Level 2, early Level 3). Elon Musk, the CEO of Tesla, says he expects fully autonomous vehicles within 3 years (i.e. by 2018). I got to take a test ride in one of these vehicles from a friend with a Tesla.

Tesla Model S
Tesla Model S

Upgraded Teslas are able to function in hands-off mode some of the time. They use adaptive cruise control to follow the vehicle in front at a desired speed constrained by a fixed following distance and use lane markings to stay in lane. They change lanes automatically at the request of the driver (who must hit the turn signal).

Tesla Model S User Interface
Tesla Model S User Interface

How it works

As of Fall 2015, none of these functions can be safely performed in a Tesla running “Auto-Pilot” in the absence of driver observation and monitoring. In fact the vehicle requires the driver to periodically return hands to the steering wheel. Rules for automated vehicles are still taking shape. Clearly this is “beta”, and intended for limited access roadways, not city streets, though Tesla drivers do use it on local roads as well as freeways. Here are a few of the issues:

  • Stopping: The vehicles do not yet automatically stop at traffic lights or stop signs, though it is assumed that engineers are working on and testing those functionalities, which may already be in the hands of testers.
  • Following traffic: When following a vehicle in city traffic, the Auto-Pilot may induce the car to run the red if the car in front ran the red (or made a right turn) instead of stopping at the light.
  • Lane marking issues: Ambiguities in lane markings (for instance at freeway merges and diverges, or as a result of road construction or restriping) still create difficulties for the vehicle in Auto-Pilot mode. During the drive, the vehicle would pull toward the exit by following lane markings. Drivers have reported “increasingly less tendency to try to take exits and overall it is clearly improving and needing less driver intervention each week.”
  • Curves: First person observations are that vehicles still over-react on curves (following the average of the inside and outside curve, rather than a fixed distance from the inside curve). Elon Musk has tweeted that slowing for curves is coming, and some Tesla drivers are reporting that their vehicles have been updated. Changes like this are part of the brilliant learning system Tesla has deployed.
  • Merging: The give-way game between merging vehicles and an on-road Tesla cannot yet be safely conducted in the absence of driver intervention. As we drove in the right lane, a Mercedes approached from an on-ramp and neither decelerated to come in behind us, nor accelerated to pass us. Our vehicle stayed at a constant speed. The Mercedes would either sideswipe us or run off the road. The driver manually intervened and accelerated (which Teslas do quite well; I can’t wait for Plaid mode, since Ludicrous mode is injurious enough if you are not braced).

Comparison to Google

The manual intervention thus requires drivers pay attention. Thus far, it doesn’t seem like drivers are being lulled to unawareness with autopilot mode on cars, but lulling is a risk if drivers trust too much. This is the advantage of Google’s all-in approach, where the driver can’t retake control even if they want to. Nevertheless, Auto-Pilot has saved lives already, see the video at this link, where an ill-timed U-turn across traffic which would have otherwise resulted in a crash was prevented).

Teslas do not presently drive independently via a map from origin to destination the way Google’s test cars do. There is no obviously linkage between satellite navigation and mapping and the control function. Teslas appear to be map-independent, and controls are through on-vehicle sensors.

The car still smells new despite being nearly a year old. I believe the car’s filters “Bioweapons Defense Mode” has something to do with that. Tesla also still retains some pluckiness and personality, despite having a market capitalization of $27B.

The vehicles are constantly learning, however, using driver interventions as expert trainers, so many of these problems will resolve themselves. None of these should be taken to mean cars won’t be automated; they will be, as a series of technical hurdles to be overcome, and interesting ambiguities and tacit knowledge on the part of drivers must be made explicit before we can hand our fates to our machines.


See video of the ride.

The race is on to figure out what self-driving cars should look like | WaPo

Matt McFarland writes in the Washington Post  “The race is on to figure out what self-driving cars should look like

David Levinson, a civil engineering professor at the University of Minnesota, argued in a recent paper that we’ll see a Cambrian explosion of new vehicle forms that are designed for specific tasks.

“The fleet will have greater variety, with the right size vehicle assigned to a particular job. Today there is a car-size arms race: people buy larger cars, which are perceived to be safer for the occupant, and taller cars, which allow the driver to see in front of the car immediately in front of them,” Levinson said. “Both of these advantages are largely obviated with autonomous vehicles. The car-size arms race ends.”

The road as an Ecosystem in the 21st Century

Today in the Tech World, there is discussion of “platforms” and “ecosystems”. When we hear talk about Apple vs. Google, it is as much about the Apple ecosystem, particularly that around iOS, the operating system for the iPhone, vs. Android OS. The Operating Systems enable both device-based and cloud-based software services. I can buy apps that work in either eco-system, but not both (without purchasing twice). I can buy peripherals that work on one or the other, but generally not both. This mobile telephone ecosystem logic follows and is much larger than the previous decades’ PC operating system ecosystems.

Roads are a different form of economic ecosystem, and perhaps the original one. There is the ecosystem for building roads, and there is an ecosystem for those using roads. Carriers as well as private vehicles are the users. But they have a set of roadside services (energy (hay, gasoline), shelter (inns (hello Jesus), motels, and hotels), and sustenance (food)) as well as many others that are less frequently used (tollbooths, money changing, black smithing, wheel wright, vehicle repair, and so on) that are configured a particular way for users of the road ecosystem.

While the types of vehicles using roads, as well as the materials with which they are made has changed over time, the platform of the road as a place on which to hang a series of road-serving businesses is long-standing, and unlikely to disappear even as roads change with the next technological shift in vehicles.

Without roads (dirt, gravel, block, rail, asphalt, or concrete), there would not be much economy. Certainly off-road vehicles and their passengers and drivers of various kinds would still require services, but the much higher cost of travel would significantly reduce the total economic impact. Secondary economic impacts on things like manufacturing, agriculture, and non-transportation services which do depend on transportation thus depend on this eco-system as well.

There is a fascinating series of books by John Jakle and colleagues describing the emergence of the first order 20th Century Road Ecosystem: Fast Food, Motels, Gas Stations, and so on. What happens in the 21st Century with Vehicle Electrification and Automation?

We can certainly speculate that charging stations ultimately replace gas stations. Even more, vehicles may be charged in motion from the roadway.

Food production and delivery may also change in ways that are difficult to foresee. We can speculate that with automated vehicles, food may come to us in motion, rather than us stopping at the side of the road. While this synchronization, resembling the in-air refueling of Air Force One, seems far out, with full information and automated drivers, it may be quite trivial. This may or may not be a net improvement in food quality.

Why stay at a hotel when your car can move you forward in space and time while you sleep?

How else will the Road Ecosystem Change in the 21st Century?

What if car driving is like playing chess


JS Writes in with an intriguing idea:

“What if car driving is like playing chess?  Self-driving cars may be possible and even valuable but the safest most efficient driving may be the combination of the computer and the person/people.  What if one Uber “driver” could drive 10 cars at once, or a team of 3 Uber drivers could drive 100 cars?”

And then sends in the following from the EconTalk podcast …
From Econtalk: Tyler Cowen on Inequality, the Future, and Average is Over 11:01


Russ: So let’s talk about what you’ve learned as a chess fan. And you write at some length. At first I was rather taken aback by this, but I grew to find it quite fascinating. You write at some length about the role of machines in chess tournaments, and particularly in freestyle. Talk about that and why it’s a nice potential template for future human interaction.

Guest: Freestyle is a form of chess where a human teams up with a computer. So, if you play human-and-computer against computer, for the most part human-and-computer, if it’s a practiced human, will beat the computer. Even though computers per se are much stronger than humans at chess, it’s the team that’s stronger than either one. And I think this is a good metaphor for a lot of what our job market future will look like. So there’s a big chunk of the book that looks rather closely at freestyle chess and tries to see what we can learn from it.

Russ: The thing I found most provocative about that is that the best freestyle teams do not necessarily have the best human players. In fact that could be something of a handicap.

Guest: That’s right. The really good human players are too tempted to override the computer and substitute in their own judgment. The best freestyle teams, they are quite epistemically modest, the human or humans involved. And what they are really good at is asking questions. So they’ll run two or three different computer programs and then just check on where do those programs disagree. And then they’ll probe more on those points. And that’s what the humans do well that the computers, at least not yet, aren’t able to copy. So it’s knowing what questions to ask that has become the important human skill in this freestyle endeavor.

I still think we will need to turn it all over to the computers, and the sooner the better. Human intervention will need to be so real-time that it is likely to be worse than the algorithm, and the lags in communication are sufficient to be debilitating. But the history of self-driving cars has yet to be written.

Autonomous vehicles: The legal and policy road ahead

CTS Catalyst summarizes some of the discussion from the recent Conference Autonomous vehicles: The legal and policy road ahead


… David Levinson hypothesized some possible directions:

  • Autonomous vehicles enable more car sharing. Instead of the sunk cost of car ownership, people pay the marginal cost per trip—and thus make fewer trips.

  • Shared cars can be right-sized for any given trip, so fewer large cars are needed. Increased safety reassures people about driving smaller cars.

  • Smaller cars travel closely together on narrower lanes, so capacity increases.

  • As networks get faster, people choose to travel farther. Cities decentralize and more megacities and “placeless places” develop.

  • At the same time, inner cities get denser, as less space is needed for parking and garages.

  • With lower labor costs, transit becomes more cost-effective.

  • Driverless trucks lower delivery costs. Combined with drones, robotics, and online shopping, retail shopping declines.

Driverless Cars on KFAI

I talk about Driverless cars on the KFAI Morning Blend (2014-12-18). The link is here. My interview starts at 1:31:30.

Autonomy Island

Ricardo Montalban and Herve Villechaize Fantasy Island (1977)
Ricardo Montalban and Herve Villechaize Fantasy Island (1977)

“Ze Car, Ze Car.”

“My dear guests, I am Mr. Roarke, your host. Welcome to Autonomy Island.”

Yes, here on Autonomy Island, all of the cars are autonomous. Your adventure will be to ride and drive in a place without fear of a human running you over.

When will an automaker (or collective of automakers, or government, or Google) buy all the cars on an island (and perhaps rent the government), replace them with new autonomous vehicles, and see what happens … to safety, to travel behavior, etc?

This is the kind of real world laboratory experiment that would be highly useful to understand the implications, the unintended side effects, the bugs and so on of robotic cars.

For instance, take the US Virgin Islands. St. Croix has a population of about 50,000 people. If it follows general US patterns, it has about 33,000 light vehicles. For about $1B [Less than the cost of a single NFL stadium] all of the cars could be replaced with autonomous vehicles at about $33,000 each. [This might be a stretch, but that would be a typical mass production cost.]

The USVI collectively has between 10 and 20 auto fatalities annually. At a $9.1 million value of life, that is at least $91M per year. In 11 years, the experiment would pay for itself if in fact it eliminates fatal crashes the way autonomous vehicles are expected to, leave aside any other potential benefits.

The advantages of an island are that it is a closed system, it can be fully mapped, no one can drive on or off. The advantages of a real island with real people are the ability to see how these interactions might actually occur in use.

Autonomous vehicles interacting with only autonomous vehicles should be much easier to design than autonomous vehicles in mixed traffic, as the environment is less variable. People, animals, weather, and so on are still potential confounding factors, but should be simpler to manage than a person in a car.

Robot cars will change your life — maybe | MinnPost

Bill Lindeke at MinnPost covers driverless cars: Robot cars will change your life — maybe. My quotes below …

David Levinson is an engineering professor at the University of Minnesota who attended this month’s conference. He explained to me that the key for autonomous vehicles is that they can react far more quickly and precisely to their surroundings.

“We could go down from 33,000 to a few hundred deaths per year by car,” Levinson told me. “In mixed environments, speeds can be regulated so that cars go much slower. People might be more wiling to travel at slower speeds in neighborhoods when they don’t have to stop at stupid traffic lights. And we won’t have the option to be more aggressive, like we can right now.”

Unlike many harried urban drivers today, robot cars would always stop for a child crossing the street or give plenty of room to a bicyclist. Saving tens of thousands of lives, while making cities safe again, is an inspiring vision.

Impeccably driven robot cars would also greatly expand our road capacity. Compared to mistake-prone humans, over twice as many robot cars might fit onto a lane of highway, which could make traffic jams (and freeway expansions) obsolete.

At the same time, the ability to daydream while driving is appealing to stressed-out commuters. As Levinson explained to me, robot cars might lead to even more driving than we see today.

“Autonomous cars will be faster on average, and as a result they’ll increase the distance people are willing to travel, “ Levinson told me. “They will also reduce the cognitive burden of drivers, and so people will be willing to spend more time driving. Both things would lead to further suburbanization.”