September 23, 2016

How much? How soon? A little realism about automated vehicles

I was asked this week to comment on two visions/proposals for future automated transportation systems, from two different organizations. You can find some of my reactions in coverage by KUOW and the Washington Post, but I thought it would be good to elaborate on my thoughts here and talk about what these two proposals get right and where they fall short.

The first proposal came from Madrona Venture Group, a leading Seattle VC firm, and calls for converting I-5 and provincial highway 99 between Seattle and Vancouver BC into a corridor reserved strictly for autonomous vehicles:

An autonomous vehicle plan for I-5 could initially allow autonomous vehicles to share the HOV lanes.  Over time, with more and more autonomous vehicles on the road, this would evolve into HOV lanes being exclusively for autonomous vehicles.  The final step as autonomous vehicles largely replace existing vehicles would be to exclude non-autonomous vehicles from I-5 except for certain defined times when highways are not congested such as most of weekends and 8:00 p.m. to 4:00 a.m. on weekdays.  The first phase of this plan could begin to be implemented immediately and the final phase could occur in ten to fifteen years.

Let’s take this proposal piece by piece. First, in the near term, should automated vehicles be given the special benefit of HOV lane access? There might be a case for this, but it is not a sure thing. Automated vehicles may end up being safer than manually driven cars – though enhancing safety is best thought of as an opportunity to be exploited, not an inherent characteristic of the technology. And automation may allow for increased operational efficiency, increased capacity, or other effects that reduce the external costs of driving, and would justify this special benefit. But if automated vehicle technology turns out to be even half as wonderful as its proponents claim, adoption will happen on its own, and extra incentives like HOV lane access will be superfluous.

In the long run, should there be an outright ban on manually driven vehicles on the freeway? Curiously, the authors don’t really make the case for why such a heavy-handed policy move would even be required. Nevertheless, there are arguments for it in the (very) long term: mainly, automated vehicles may be able to increase lane capacity and speeds, allow for decreased lane widths, and enhance safety beyond levels achievable by manually driven vehicles, and they will likely do these things better in an environment where all the other vehicles are also automated. But banning manually driven vehicles makes little sense when there are only a few automated vehicles on the road.

This brings me to the more important point, which is that the timeline proposed – 10 to 15 years – is unrealistic and would lead to highly inequitable outcomes. First of all, we are talking about a technology that is commercially available in precisely zero vehicle models today. Tesla’s Autopilot is a Level 2 system, meant to assist, not replace, the driver in controlling the vehicle. Second, while Madrona notes the rapid adoption of smartphones and cell phones, technological change in automobiles tends to be slower. Analysis (paywalled, but also found in Chapter 7 of this report) by Stanford’s Stephen Zoepf has shown that new automotive technologies these days take about 10 years from initial market introduction until they hit their peak growth rate. Moreover, that maximum rate of growth is usually less than 10% per year, and almost never more than 20% per year. A plausible, if optimistic, adoption curve is shown below, where adoption refers to the percentage of new vehicles equipped with a feature.

But even if automation somehow can be deployed more quickly than any other automotive technology in history (bear in mind that many of Zoepf’s technologies are things like keyless entry, which are vastly cheaper and less complex than automation), there is still the issue of fleet turnover time. 60% of cars in the US survive 15 years or longer. As a rough estimate, 15 years from now, we will still have 1.5 million cars on Washington state roads that were manufactured in 2016 or earlier. Even if every car sold starting next year were automated, Madrona’s proposal would be effectively shutting out from travel on I-5 the individuals and families that rely on these 1.5 million vehicles. In all likelihood, these will be the people with the fewest alternatives to driving, not the sort of people for whom seaplane travel from Seattle to Vancouver (the Madrona team’s other solution) is a realistic option. Exacerbating this problem is the fact that there is not really a good alternative to I-5/BC-99 between Seattle and Vancouver:

Nonetheless, Madrona assures us that this change would “benefit all who use this corridor.” This might be correct, if you limit your analysis to only those people who are still permitted to use the corridor, and ignore those who have been shut out.

The Madrona proposal also dismisses the risks of automation somewhat blithely:

As with most beneficial innovations, there are risks.  For example, autonomous vehicles will make it possible for people incapable of driving because of age or infirmities to use vehicles thereby increasing the total number of vehicle miles traveled.  Of course, providing a means for these people to travel or visit friends and doctors is itself a social benefit.  Such usage will also be offset by increases in ride sharing in private autos, Uber-type services, and mini-buses and buses which would reduce the number of vehicles on the road.

They are right that increased mobility for the elderly, disabled, and other underserved groups is almost certainly a net social good, even if it leads to an increase in travel demand and associated externalities. However, our own research has shown that the bulk of increased travel would probably come from increased amounts of travel by those who are already able to drive. Personally, as someone originally from Vancouver who lives in Seattle, I would make the trip up and down I-5 a lot more often if my car could do the driving for me.

Madrona’s basis for dismissing this concern is simply an assertion, without supporting evidence, that automation will increase ride sharing. Ask yourself this: how do you feel about getting in a car and sharing a ride with a stranger, using a service such as UberPool or Lyft Line, when there is also a driver present? Now take away the driver, so it’s just you and some rando in the car. How does that sound now? While I agree that greater use of mobility services, and ideally sharing rides, is key to realizing many of the benefits of automation, it is dangerous to assume that this will simply happen on its own.

The second study is from the Rocky Mountain Institute (RMI), and is more of a vision of the potential benefits of shared, automated, electric vehicles (SAEVs). RMI considers the case of shared vehicles, but does not assume sharing of rides. In these kinds of analyses, it’s important to distinguish between what could happen, and what we think will happen. At the outset, the RMI authors use words like “could” and “possible” and “plausible”, and rightly note the importance of shifting policies to ensure that benefits are realized. If I’m reading their report correctly, their goal is to illustrate how SAEVs offer a large business opportunity and the potential for a greener transportation system – not the guarantee of one.

 

RMI makes the case that, firstly, automation will make TNC services (Transportation Network Companies, e.g. Uber) cost competitive with private vehicle ownership, and secondly, that EVs will be the most cost effective powertrain technology in high-mileage TNC fleet operations:

The unstated caveat to this analysis is that TNC services are great in the city, but don’t work so well for longer trips outside the metropolitan area. Going to the cottage or the state park for the weekend? Hiking or skiing for the day? That TNC vehicle has to either sit around all day waiting for you, or make a long deadhead (empty trip) back to the city where it can pick up another fare. Either way, you’re going to be paying for more than just the one-way trip. Similarly, EVs are not as well suited to longer trips where they may need a lengthy recharge along the way. The risk here is that as long as people have some need for trips that can’t cost effectively be met by an SAEV, they will continue to own private vehicles. And if those privately owned vehicles are automated, watch out for people to travel in them more, due to the reduced burden of driving.

RMI is absolutely right that fleet applications, where vehicles would rack up miles much faster than in private household operations, present a stronger business case for EVs. The figure below shows the discounted present value of lifetime fuel cost savings for an EV versus a gasoline vehicle, as a function of annual miles traveled. Even assuming that vehicle lifetime (in miles) remains the same, compressing the service life into a shorter period of years means that future fuel savings are discounted less heavily.

Finally, RMI seems to take a more realistic view than Madrona about the timeline for growth of AVs:

One thing that is unclear in the above figure, and I was unable to determine from the report, is whether their “number of vehicles deployed” is a cumulative number or an annual number. If it’s cumulative, then their fast growth scenario suggests sales of about 1 million SAEVs in 2024. Shifting the adoption curve above to a market entry date of 2021, sales of 1 million SAEVs (~7% of US light duty vehicle sales) in 2024 is not implausible. If 2 million + is an annual sales estimate for 2024, then sorry, but I’m not buying it.

Now, where does RMI go off the rails? First, there’s this figure, which suggests that new vehicle sales are set to plummet:

This is a common fallacy, that fewer vehicles in circulation means fewer vehicles sold. But if vehicles in TNC fleet operations are accumulating miles faster than privately owned vehicles, they will need to be replaced that much more quickly too, and overall car sales should remain more or less unchanged. Digging into this, RMI appears to have assumed a useful life of 280,000 miles for a conventional vehicle, but 490,000 miles for SAEVs. The justification for this assumption is not articulated in the report, except for an oblique reference to “optimal value,” but both of these numbers are much longer than the lifetimes of vehicles today, which are generally less than 200,000 miles.

Another issue with the RMI analysis is that they dismiss the risk of automated vehicles making congestion worse, arguing that “Studies also show that autonomous vehicles could increase vehicle throughput by at least three times, meaning even double VMT from autonomous vehicles would still have reduced congestion versus today.” They don’t bother to say which studies they are referring to, but one that I’m familiar with suggests a doubling of capacity, not “at least three times.” Moreover, this doubling of capacity is realized only when all vehicles are compatible. During the decades-long transition period when automated and manually driven vehicles are sharing the road, the increase in travel demand from AVs could well outweigh any increases in effective capacity. This is not a sure thing of course, but we should not be so quick to dismiss the risks.

These two studies make a useful contribution to the discussion. In the very long term – decades from now – it will probably make sense to exclude manually driven vehicles from some roads. And the convergence of new mobility services, electrification, and automation present an opportunity to remake our transportation system to be more socially, economically, and environmentally sustainable. But as with many others, the authors of these studies seem to have gotten carried away by hype a little bit, which leads to some potentially misleading results and unhelpful recommendations.