Cities have a "low-speed shuttle" problem, and I'm partly responsible for it. Now I'm in a mission to help cities get autonomous mobility right.

“The best way to predict your future is to create it.” ― Abraham Lincoln

It's 2011. #Hastags don't exist. Smartphones are a luxury device. And autonomous vehicles are science fiction. But on a sunny day of May 2011, in La Rochelle, a mid-size city in France, our research team at Inria launched the world's first autonomous vehicle service on a public road. I was in charge of coordinating the team that put this service together. This was not our first try: we had done at smaller scale it in the UK in 2007, and in Norway and Finland in 2009, and other tries had been done between 2001 and 2004 by the Cybermove project. But this time it was bigger : a two months pilot, with a local operating team that we had to train and supervise remotely from Paris. And it was a huge success. The local and national French press covered extensively the news (only in French, here, here or here). But in 2011, this remained rather confidential. Crunchbase didn't covered it. Techcrunch didn't heard about it. Still, we got additional research funding to do the same all over Europe. So from 2012, we headed to 12 European cities to test "low speed experimental shuttles". Yes, we were still doing applied research. We requested several robotics companies to propose a fleet of at least 6 vehicles and at least 60 seats. The companies simply divided 60 seats by 6 vehicles, and designed 10 passengers' shuttles. That's where today's unofficial "10 passenger shuttles" standard comes, and, except startups using commercial golf vehicles (here and here), even newer companies (or Toyota) still design low-speed shuttles to have roughly that capacity.

Why did we deployed the fleet at low speed? We had been working on this field since the early 1990's, slowly building trust in our research with cities, research funding organizations, national safety authorities the press, etc., so we had to be extremely cautious to avoid that the crash of a prototype destroyed this trust. Therefore, before the fleet arrived to a city, we worked between 12 and 15 months with the local authorities to determine a simple place to test the service, where traffic wasn't huge and where the technical complexity was relatively low. If necessary, we asked the city to redesign the infrastructure, as shown in the picture below (where the city of Trikala removed a lane of roadside parking to create a dedicated lane for the autonomous minibuses). And a thorough risk assessment determined the speed that always allowed the vehicles to stop safely. Since we were mostly in areas with a high pedestrian density, the resulting speed was obviously low.

By the time we finished this research project, it was 2016. Something like this didn't remain confidential anymore. There was a huge buzz around this: we got coverage from Euronews, National Geographic, National Spanish TV... The buzz started creating its effects. Startups got created based on the concept of "low speed shuttles". Cities started piloting them to "understand the technology". Around that time, big G and big U were announcing big plans for autonomous services, so billions started flooding the "low speed shuttle" economy. But it's easier to get excited with videos than with research publications, and based on what has happened since 2016, it seems no one bothered to read our research's results, and our research's results showed that, despite enthusiastic users and stakeholders, low speed shuttles, in the form we designed them for our research projects, are not financially sustainable (see 1).

Indeed. Low speed shuttles cost a lot, somewhere between 200 000€ and 400 000€ apiece, which is absolutely normal: they are hand crafted vehicles, produced in small volumes, with high tech sensors which cost a lot too; companies making them need to hire highly skilled and specialized engineers, a scarce resource in today's digital economy; in the absence of safety assured technology, they must embark a "safety steward" to ensure safety, a cost of around 36000€/year per vehicle. And most of all, they drive, well, slowly. So how does all this add up in a successful business case? Well, it doesn't. At a low average speed, say 12 km/h, and a reasonable trip length of 3 km, a low speed shuttle can do, at most, four 15-minutes trips. With a 10 passenger capacity, and supposing the shuttle is always full (it isn't), that means 40 passengers per hour. Over 2 morning and 2 evening peak hours, plus an average occupancy of 30% over the rest of the day (10 hours), a low-speed shuttle will transport a maximum of 280 passengers per day (80+80+(10x40x30%)), generating 168€ of revenue per day. That's 36 960€ per year, barely enough to cover the cost of the on-board operator, and enough to repay a 300 000€ shuttle in... 312 years (43 years if the on-board operator costs just 30 000€)! And, to the best of my knowledge, all these shuttles experiments are free for riders, so no low speed shuttle service has even produced any revenues.

So if shuttles are such a financial disaster, who's been paying for all these pilots? Public subsidies and VC's private funds. The latter is not a problem, as VC's diversify their investments so that some other company repays the losses. But the former is a true concern, because this money isn't going to make shuttles more efficient, it's only being used for public authorities to learn how low-speed shuttles (again, still a research concept) work. That's a laudable goal, but it can be accomplished at much lower cost. Yet, hype and FOMO are two strong economic and political forces. So since 2016 we started seeing low speed shuttles "pilots" pop up in all the wrong places, as a recent study confirmed: parks, pedestrian areas, dense city centres, business and technology parks... What's wrong with that? Competing for space with pedestrians, the shuttles' speed gets even lower, and a distance that a human driver would normally cover in 3 to 5 minutes, can take three times as much in a low-speed shuttle: people will use it once or twice for curiosity, but it will never make them leave their car. In dense city centres, where mass transit is and will always be the most efficient transport mode, low speed shuttles not only take up as much space as cars, but they create more congestion. Basically, in those places and conditions, low speed shuttles provide little to no service at all. Therefore, cities end up concluding that, with this configuration, low speed shuttles are useless, and end up rolling back their experiments.

So today, do I think autonomous vehicle technology is not useful? Not at all. I still think autonomous vehicles are the only hope for cities to eliminate car dependence, improve public transport and reduce congestion, pollution (with electric AV's) and crashes. But the commercial development of low speed shuttles since 2016 to date has been done based on an unfinished research concept, not on a real product fit for real needs! Cities have been spending their limited research and innovation budget in something that, by definition, will never demonstrate to be profitable or efficient, because when we did these research projects, we were looking to define the technical conditions for their deployment, not (yet) for their financial sustainability.

After seeing the research concept of low speed shuttles spread for the past four years, disappointing to the point of being qualified as the the new dot-com bubble, I feel partly responsible for this situation. With the fall of the autonomous vehicles hype, and the risk that intermediate, research-level results lead cities to discard automated vehicle technology, I want to help cities steer autonomous vehicles in the right direction and towards the right outcome: getting people out of their -private, individual- cars. In fact, urban mobility is approaching a crossroad. Partial automation technologies are already on car showrooms, and if the right measures are not taken today, cities risk suffering the consequences of the hellish mobility scenario of privately owned autonomous vehicles depicted by Robin Chase in 2014:

If single-occupancy vehicles are the bane of our congested highways and cities right now, imagine the congestion when we pour in unfettered zero-occupancy vehicles.

What needs to be done then to make autonomous vehicles financially sustainable and useful for urban mobility? First of all, the burden of dealing with the present real-world constraints to deploy autonomous vehicles must be on the autonomous vehicle industry, not on cities. I'll start by debunking a few fallacies that the autonomous vehicle hype has imposed over cities, and which are deviating their attention, and their funds, from the real issues they should be preparing for.

• A low speed and an on-board attendant as a guarantee of safety: As I explained above, during our last research project (2012-2016) we deployed 12 automated shuttles in 6 European cities. The shuttles drove without a single safety incident for more than 28000 km. Today, this mileage may seems ridiculous, in front of the race to autonomous driving miles, but there is a key difference: we were carrying passengers, from the public, from day 1. More than 60000. And this was done with 2012 technologies, for the first time at this scale, with a budget that Waymo or Uber now spend in less than a month. So how come low-speed shuttles equipped with more recent technologies have been involved in several crashes, or caused passenger injuries, and that a behemoth as Uber, who employs hundreds of extremely brilliant engineers, was at the origin of the world's first pedestrian fatality caused by an autonomous vehicle in 2018? This is difficult to explain, but I can tell that our safety record was not achieved by chance, but thanks to the 12-15 months of careful preparatory work with the cities, before the minibuses even took their first spin, and not on the assumption that the technology alone would solve everything. Which is the key takeaway for cities? That when there is a clear vision, a good analysis is worth more than the biggest budget or the latest technology. Concerning safety, the takeaway is that technology alone cannot guarantee safety, if that technology is not encompassed in a global safety approach.
• Adapting the infrastructure: This fallacy comes from the hype-led idea that autonomous vehicle software will operate based on vision and "deep learning" or "machine learning", and therefore, without perfectly drawn and maintained lanes, road signs, etc., autonomous vehicles wouldn't be able to operate. Anyone who has seen machine vision in action knows how unreliable it is. Tesla uses vision to follow lanes, but as this video shows, using vision-based lane marking following can be a really bad idea. Computer vision goes against the one thing cities need when deploying a fleet of autonomous vehicles for commercial service: dependability. It's up to the industry to provide technologies that cope with real world constraints in a reliable manner, and neither road markings nor vision are reliably enough to depend upon. Vision can be useful in certain cases, but certainly not to be in charge of a two-ton machine. Fortunately, there are several other kinds of sensors and technologies that can cope with vision's deficiencies, and guide autonomous vehicles in a more reliable manner. So, until autonomous vehicles technologies are not stabilized, cities should definitively not waste their time and money "adapting the infrastructure" for machine vision. The infrastructure will probably have to be adapted after autonomous vehicles are deployed, in order to improve their efficiency, but investing before their arrival is wasteful.
• Build a digital infrastructure: This fallacy says that until 5G is available, or until all roads are equipped with sensors, autonomous vehicles will not be able to operate. Nothing is further from the truth. 5G is only necessary to make remote operation of "autonomous vehicles". But if you do remote operation, it means the vehicle and the automation technology was not designed to cope with the road environment. This equals having a vehicle with a driver, which eliminates any economic benefit of automation. Again, it's up to the industry to cope with the real world, and an autonomous vehicle should be able to operate with the existing cellular technology, eventually updating once 5G is globally deployed. Until then, cities should not bear the burden of solving the "chicken and egg" problem for 5G or roadside sensor deployment.
• Huge investments: Since the autonomous vehicle industry has sunk billions (some account U$35 billion) with little to no results, and has kept delaying their commercial arrival further and further, there is the impression that any autonomous vehicle deployment requires billions, solely on the industry's willingness. It does not. Why? Companies developing autonomous vehicle technology, especially the car industry, don't have the same objectives than cities. The car industry needs to keep selling cars to individuals. That's what they have always done, and in the short term, there's no reason for that to change. With that goal in mind, the car industry obviously needs an autonomous technology that works throughout the world on launch day (the so-called level 5), and this definitively requires billions and decades in research and development. But do cities need that? Not at all. Cities just need autonomous vehicles that serve their citizens in a reliable way. And that's a much closer and attainable objective.
• Cities must wait for autonomous vehicles to progressively reach SAE level 4 (or 5): The first version of SAE's automated driving levels standard was published in 2014. These levels are numbered from 0 (your 1980's car with no automation at all) to 5 (the 2073 model that will, maybe, drive anywhere), and the fact they are numbered sequentially made people think autonomous vehicle technology has to evolve sequentially from 0 to 5. This document became a de-facto reference in a world where autonomous vehicles where still (for most people) a distant science fiction concept, especially for journalists who didn't have any technical reference. But in 2011, when we started our autonomous vehicle service in La Rochelle, we were doing already a level 4 system. In fact, technically speaking, it's easier to go straight to level 4 (a vehicle capable of driving and stopping safely in a limited set of conditions), than to follow the SAE ladder, for a simple reason: up to level 3, automated driving systems have a fundamental problem, the fact that a person must be ready to serve as "fallback of the Dynamic Driving Task (DDT)". This obscure term means that a person must be ready to take over the driving in case the computer fails. This is an unsolvable problem. If the system performs fine 95% of the time, it will cause driver accommodation and, inevitably, the "driver" will distract, being unavailable for the 5% (or more) cases in which the technology fails. Therefore, there is no way to create a reliable level 3 system and under, simply because human reaction is impossible to rely upon, whereas the system will be more reliable if it is already designed with the assumption that there is no human fallback.
• Think of any other fallacy? Let me know in comments and I'll try to debunk it.

With the above myths clarified, city officials should focus on demanding to the AV industry the performance that will make autonomous vehicles bring real benefits to their local mobility. The first and most important step in this direction is to understand is that low speed shuttles were and still are the by-product of a research project, not the final version of a transportation mode we expect to eliminate car ownership in cities. To date, to deploy these research-grade low speed shuttles, cities had to find places and use-cases that coped with their technical limitations (low speed, on board safety steward), leading to -expectably- poor transportation results and unsustainable economics. To raise the bar for the market, it is therefore crucial that city authorities demand that "final version"'s performance. What's that level of performance? Take the places where ridership doesn't cover the public transport's costs, areas where public transport stations are more than 1,5 km (1 mile) away from houses or offices, areas where riders have to wait more than 10 minutes to be served. Those are the areas that cities should target to deploy autonomous vehicle services. I voluntarily say "areas" and not "lines", because autonomous vehicles today can serve entire areas, because lines will reproduce the same weaknesses of bus lines today.

How do we plan to help cities on this endeavour? First, we have built a very basic simulator that will let cities test the impact of an autonomous vehicle on their public transport service in the areas meeting the above mentioned conditions (expect bugs, it is at the very first version). Secondly, we have transformed our expertise on a piece of software that will allow high-speed and unaccompanied operation of autonomous vehicles at a higher level of safety than low speed shuttles available on the market today, and that doesn't require cities to heavily adapt their infrastructure.

Does your city or metropolitan area have a use-case that would require a performance level unmet by today's research-grade low speed shuttles? Do you want to ask us something? or just chat? Then email us at info at autokab dot com, we'll be happy to help.

Sources

A. Alessandrini, R. Alfonsi, P. delle Site, V. Gatta, E. Marcucci, Economic assessment of driverless buses for urban applications: a case study from the European project Citymobil2, XVIII Conference of the Italian Association of Transport Economics and Logistics (SIET), Genova 4th-5th July 2016.