Otto and self-driving trucks -- what do they mean?

Topic: 

Today sees the un-stealthing of a new company called Otto which plans to build self-driving systems for long haul trucks. The company has been formed by a skilled team, including former members of Google's car team and people I know well. You can see their opening blog post

My entire focus on this blog, and the focus of most people in this space, has been on cars, particularly cars capable of unmanned operation and door-to-door service. Most of those not working on that have had their focus on highway cars and autopilots. The highway is a much simpler environment so much easier to engineer, but it operates at higher speeds so the cost of accidents is worse.

That goes doubly true for trucks that are fast, big and massive. At the same time, 99% of truck driving is actually very straightforward -- stay in a highway lane, usually the slow one, with no fancy moving about.

Some companies have done exploration of truck automation. Daimler/Freightliner has been testing trucks in Nevada. Volvo (trucks and cars together) has done truck and platooning experiments, notably the Sartre project some years ago. A recent group of European researchers did a truck demonstration in the Netherlands, leading up to the Declaration of Amsterdam which got government ministers to declare a plan to modify regulations to make self-driving systems legal in Europe. Local company Peloton has gone after the more tractable problem of two-truck platoons with a driver in each truck, aimed primarily at fuel savings and some safety increases.

Safety

While trucks are big and thus riskier to automate, they are also risky for humans to drive. Even though truck drivers are professionals who drive all day, there are still around 4,000 killed every year in the USA in truck accidents. More than half of those are truck drivers, but a large number of ordinary road users are also killed. Done well, self-driving trucks will reduce this toll. Just as with cars, companies will not release the systems until they believe they can match and beat the safety record of human drivers.

The Economics

Self-driving trucks don't change the way we move, but they will have a big economic effect on trucking. Driver pay accounts for about 25-35% of the cost of truck operation, but in fact early self-driving won't take away jobs because there is a serious shortage of truck drivers in the market -- companies can't hire enough of them at the wages they currently pay. It is claimed that there are 50,000 job openings unfilled at the present time. Truck driving is grueling work, sometimes mind-numbing, and it takes the long haul driver away from home and family for over a week on every long-haul run. It's not very exciting work, and it involves long days (11 hours is the legal limit) and a lot of eating and sleeping in truck stops or the cabin of the truck.

Average pay is about 36 cents/mile for a solo trucker on a common route. Alternately, loads that need to move fast are driven by a team of two. They split 50 cents/mile between them, but can drive 22 hours/day -- one driver sleeps in the back while the first one takes the wheel. You make less per mile per driver, but you are also paid for the miles you are sleeping or relaxing.

A likely first course is trucks that keep their solo driver who drives up to 11 hours -- probably less -- and have the software drive the rest. Nonstop team driving speed with just one person. Indeed, that person might be an owner-operator who is paying for the system as a businessperson, rather than a person losing a job to automation. The human would drive the more complex parts of the route (including heavy traffic) while the system can easily handle the long nights and sparse heartland interstate roads.

The economics get interesting when you can do things that are expensive for human drivers and teams. Aside from operating 22 or more hours/day at a lower cost, certain routes will become practical that were not economic with human drivers, opening up new routes and business models.

The Environment

Computer driven trucks will drive more regularly than humans, effectively driving in "hypermile" style as much as they can. That should save fuel. In addition, while I would not do it at first, the platooning experimented with by Peloton and Sartre does result in fuel savings. Also interesting is the ability to convert trucks to natural gas, which is domestic and burns cleaner (though it still emits CO2.) Automated trucks on fixed routes might be more willing to make this conversion.

Road wear

There is strong potential to reduce the damage to roads (and thus the cost of maintaining them, which is immense and seriously in arrears) thanks to the robotruck. That's because heavy trucks and big buses cause almost all the road wear today. A surprising rule of thumb is that road damage goes up with the 4th power of the weight per axle. As such an 80,000lb truck with 34,000lb on two sets of 2 axles and 6,000lb on the front axle does around 2,000 times the road damage of a typical car! An interesting solution is now possible. With fully self-driving trucks (or platoons with nobody in the rear vehicle) you can have two half-weight trucks, which would do 1/8th the damage. 4 1/4 weight trucks would do roughly 1/16th the damage. This is a bit more expensive in fuel and truck wear, though you can get back some of it with platooning. The platoon can space out further on bridges to avoid stressing them.

Right now, trucks pay a higher tax for their high weight, but it's not 2,000 times what the cars pay. Proper pricing of the externalities of road wear would cause smaller loads to be hauled and save the government a ton of money on road repair. (On the other hand, the standardized 40' shipping container is so common and useful that there are other challenges here.)

The ability to send smaller loads will enable truck shipping to many destinations where it's not that economic today. It will also make it easier for the truck to get a load "going back" because robots don't mind waiting very much. Today, if a truck driver takes a load to a minor location, they may have a long wait before there is a load going back home, and human beings can't wait quite so easily. (There is always a cost of keeping the tractor idle, but it's less than the human cost.)

Replacing jobs

Most self-driving teams are working on cars. They do not seek to replace cab drivers, they seek to replace amateur drivers (ie. you.) Of course, they will replace cab drivers, even if that is not the first goal. Cab driver is generally not thought of as a career or something people aspire to -- it's a low skill job people can do to quickly make modest wages. With Uber, people also like setting their own hours; it is a fantastic form of part-time work you can start and stop as you wish, minute by minute.

Truck driving is closer to a career, even though it is not a well-loved one for most. There are a bit over 3 million truck drivers in the USA. At first, this technology will not displace many workers, because of the driver shortage. For owner operators it will actually be a boon.

But it would be false to pretend that in the much longer term, this doesn't reduce the number of people working as career truck drivers. This is one of many of the jobs today that will be modified or replaced by automation in the years to come, and there is no stopping it.

In addition to the jobs, this article paints an more bleak picture concerning those who derive their income from truck drivers, such as Truck Stops, motels and even the towns around them.

Jobs vs. lives

All of this, however, must be put in contrast with the safety benefits. It will be hard to claim that society should work to preserve a job class which kills 4,000 people a year. Probably the only job which kills more is doctor, and they do that in the process of saving lives, not hauling cargo. While some jobs have higher death rates, I suspect only the military, police, doctors and drivers actually kill large numbers of people in the course of doing their jobs. In addition, those who kill while driving often suffer permanent mental scars.

While some will criticise the self-driving trucks as putting the public at risk, once they reach their safety goals, the self-driving trucks will be doing the reverse.

What about rail?

My first reaction looking into this was why don't we use more rail? Rail is efficient and takes very little human labour per car. To get that efficiency, however, you can't have service on demand from anywhere to anywhere -- you need to group many cars all taking the same trip. In addition, intermodal switches (getting off the train and onto a local truck) are still slow and expensive. In the future, a good mix of rail and on-demand robotrucks combined with automated intermodal switching will provide the most efficient and quick answer for cargo.

Note

I should disclose that while I have no formal relationship with Otto, an advisory role has been discussed which may introduce some bias.

You may also like the coverage in the New York Times by John Markoff.

Comments

I recently read about a scheme to introduce train-style overhead power lines over the slow lane of the autobahn in Germany to allow electric trucks to drive there. Apparently Siemens is also building something similar somewhere in the USA.

Combine self-driving capability with this turbine-electric hybrid:
https://nikolamotor.com/one
Also remote driving capability might be an intermediate step or adjunct to full autonomy. I recall reading about a mine in a remote area of Australia using mostly autonomous trucks to haul ore, and remote drivers supervise and can intervene in situations the robots can't handle. This saves the cost of housing drivers in the remote location. Similar to flying military drones all over the world from a base outside Las Vegas. Could truck owner/operators control their trucks from the comfort of their home? With higher efficiency and lower maintenance of a turbine-electric, a fleet of trucks could operate 24/7 by shifts of drivers in a central command center.

The data networks are not up to remote driving yet, unless your autonomous system is good enough to handle any outages in the data connection. That might not be impossible, since the answer to a data outage might be "pull over if you can't understand the road" since this should be rare.

For mines etc. you build your own data networks and can rely on them more. And it's also OK if a vehicle just stops during a data outage, it is not blocking other people. Of course if you pull over in a cellular dead zone and can't get a connection, you have to wait for somebody to arrive to get you out of it -- expensive unless very rare. One option would be auto aimed satellite as a backup (for use only when cell network is not running) but this is high latency and expensive. Not good for live driving but enough to get out of a dead zone.

intermodal switches (getting off the train and onto a local truck) are still slow and expensive

I would imagine that the same kind of tech going into automated driving could also be applied to this problem.

Actually, while this tech could put a truck into position to receive a shipping container. otherwise I think it would be quite different tech to make automated robots that can disconnect a container from a rail car, crane it over, place it on a truck trailer, secure it and then, if need be, hook up a tractor to haul it. And vice versa. Very different stuff, but useful.

Another issue rails have is that they pull off an average speed of 22mph, which means that even a solo driver doing 11 hours/day going door to door beats that speed.

Ok the tech needed to automate shipping container transfers is quite different than autonomous vehicles, but to my mind it's orders of magnitude easier. It's well-defined, contained by a private shipping yard with no public yahoos doing what normal drivers do. It's exactly the same kind of manufacturing automation that's been building our cars and electronics for years, just scaled up to large tonnage robots doing a relatively simple well-defined task. Why on earth wouldn't automated transfer to computer-scheduled rail (and back for the last miles) be a part of this kind of thing?

We should get that sort of transfer automated (the unions of course would fight that) but in the end you have to get the average speed of the trains up.

In fact, in spite of the fact that trains are very energy efficient, this low average speed dooms them against trucks that can drive 23 hours a day at 55mph, with no waiting time for cargo transfer, and no going out of your way to find a train going roughly where you want.

If the truck does 1300 miles in a day and the train does only 550, the train will be relegated only to cargo either not in a hurry, or so in need of the higher efficiency that it can't afford the truck -- which just got cheaper. In fact, rail is losing to trucks even though trucks are five times more expensive per ton-mile than rail. This tech will not improve things much for rail. You see rail hauling stuff like coal and grain that can't pay a lot for shipping and is in no hurry. For rail the path will be:

  • Load on a truck anyway, and truck drives some distance to rail terminal
  • Wait at rail terminal for a train big enough to be efficient on a trip to your target area
  • Wait for intermodal transfer
  • Travel in the train, including various switch-outs as it switches engines or switches to other trains
  • Intermodal switch to waiting truck (no wait this time)
  • Truck travel to customer

I would guess this all takes 3 times or more as long, and costs perhaps 1/3 of the truck (you needed some trucks here.)

Ships are awfully slow compared to cargo planes as well, though, yet there's still a huge sea freight industry.

You're focusing too much on single, isolated trip times. In reality it's a combined logistics problem. There's probably a lot more cargo where it's much less important that it arrives fast than that it arrives on a predictable schedule. In those situations a 66% savings on shipping costs is a big win. Automated trucks driving 23 hours a day are not going to fully negate this advantage, even at the current low average speed of freight trains.

BTW, are the numbers you cited U.S.-only or worldwide? Thanks to the interstate system and general neglect of rail, the U.S. of course has a much weaker rail infrastructure compared to cross-country roadways than most of the rest of the industrialized world.

Sea cargo is indeed much cheaper than trains (blows away trucks) but of course it primarily is used where there are not land routes, or there is no hurry at all.

The world has moved much more to just-in-time supply chains, which value that speed. People don't like needing to know how much to order weeks or months in advance, since even though trucks may be 5x as expensive, the cost of warehousing and keeping inventory and sometimes ordering the wrong amount far in advance seems to dominate. So most cargo is moved in trucks. (Of course planes are much more expensive but we still use them a lot.)

I believe average speeds are higher in Europe, but in part I suspect a feedback problem. Because rail freight is not in a hurry, they are not in a hurry to transfer it. Most rail cars switch trains several times I think, because only a few cars are going from A to B. And the trains usually switch engines when they switch lines, because that's how lines make their money.

It's about predictability. So much of what we buy and own in the U.S. (and Europe, and elsewhere) is made in China and elsewhere overseas, and most of it is transported by sea. Just-in-time manufacturers, and retailers like Walmart/Target/Amazon/etc. only need to be sure that the right products are available at the right time to supply their needs, in order to keep their inventory as low as possible. So it's all a big balancing act between every mode of transportation, and the costs or savings of one over the other. In other words, logistics. So cargo ships are going to keep sailing while air freight is flying the same routes, and trains are going to keep rolling while trucks are driving to the same destinations. None of them are "doomed," as you claim, to any of the others, due to unique efficiencies each has that others can't match, even with self-driving tech. In fact, you might as well claim that cross-country air freight is "doomed" to self-driving trucks, once the trucks can drive 24 hours without needing to stop, because the air freight will become so inefficient in comparison that no one will be able to afford to fly any planes.

But no, of course there will still be cargo that needs the higher speed of air freight, even at its higher cost. And of course there will still be cargo that doesn't need as much speed and is more efficient to ship by rail. Just as there's still sea freight even though we have cargo planes.

You are right, I didn't mean doomed as in going away, but I do think there will be a decline if trucks offer more for less money.

"Ships are awfully slow compared to cargo planes as well, though, yet there's still a huge sea freight industry."

Right, but that's because marine engines can burn tar. Even the biggest ship has fuel costs measured in single-digit dollars per ton. If air transport ever got that cheap, people would stop using ships (because it's a lot easier to build an airport than a seaport.)

Hello, Brad. I read your blog regularly, and I'm not clear on why robotrucks are not imminent, or indeed, already operating.

From everything I have read, Google and Tesla independently solved interstate driving quite early in their driverless car development, and Tesla is already offering a "driverless lite" full auto mode for interstate driving. Since long haul trucking is over 90% interstate based (that number may be wrong, but it's a large number greater than 50%, correct?) it would seem like these systems should be very robust and available right now.

Is the remaining technical hurdle managing the difficult dynamics of the larger truck chassis? Is it handling the logistics of route management to optimize efficiency? Does Otto plan to field trucks that can drive from one freight loading location all the way to a final destination, or does it assume a handoff to human control on inner-city surface streets?

Also, you said in your post that you envisioned the robot and the human taking turns driving. I'm not clear on why you'd EVER let the human drive, once the robot was qualified to operate autonomously. It's not like the robot would need a break. If anything, I'd let the robot do ALL the driving on the open road, and only put the human in charge during the non-interstate portions. Assuming this arrangement lasted more than a year or two (after which, the robots might be qualified for inner city driving too), it might be prudent to arrange handoff yards at truck stops, where waiting "ferry pilots" could switch in to "bring the rig in" from the road to the distribution center. Since this would involve low-speed driving, it would presumably be safe, and because the drivers only work locally with the arriving and departing trucks, they could be sleeping in their own beds every night. In other words, let the robots do what they're good at, and free up the humans to do what the robots can't (yet).

Am I missing something?

What you're missing is that highway driving is not "solved."

Humans have a fatal accident around every 100 million miles. Sadly, truck accidents with heavy vehicles at speed are probably more likely to be fatal than car accidents. You could have the trucker sleep in a "crash bed" but that does not do anything for a car they hit.

So anyway, nobody has that reliability record yet.

As to why you would mix human and software, I describe it above. The human drives the complex zones (urban, heavier traffic) and the software drives the simpler zones (night, low traffic, rural.)

Hey, Brad, thanks a lot for answering!

As to your first point [the rate of human accidents and the danger to humans outside of the truck] that just emphasizes to me that I'd rather have the robot in control! I'm fully willing to accept that the problem isn't solved, but I don't know (and do not lean yea or nay) on whether robots would currently have a better or a worse safety rate than humans. As you say, we're already sacrificing 4000 people on the road every year. The part I'm specifically unclear on is whether robots are out-performing humans on robot truck simulators right now. I understood the robots to be pretty good and reliable on interstate scenarios, and I'm not clear on why they'd be much worse in trucks than they are in cars. I'm willing to accept the possibility that trucks are harder to control, but I don't know either way.

Is the reliability record lacking just because the software hasn't accrued enough time controlling trucks under real world conditions yet?

If the mix of truck/human were really restricted to trucks controlling rural or empty interstate time and humans controlling urban zones, once again, I assert that it would be VERY cost attractive to push to have the trucks *empty* on interstate routes, and ONLY have humans invoked on urban portions of the route. Again, it seems to me that once you have certified the robots to operate without direct human intervention (if you are saying the humans have to be in direct access to control at all times, can you clarify that?), why even have humans in the cab? If a crazy scenario kicked off that required human control (a massive oil well fire right besides the interstate, as happened in Ramah Louisiana, in 2007, as a random example (http://www.techetoday.com/i-10-fire-prompts-tough-new-rules)), I think it would be reasonable for the vehicle to pull over and call for human assistance. Is the counterargument that the technology is still too unreliable to operate without a human available for immediate takeover?

It just seems wasteful to me to put a human in the cab and make him or her ride along so that they could take over for a few minutes of urban/high traffic driving _hours into_ a long haul route. But my logistics are poor, so correct me if I'm missing the big picture.

Some day you won't have the human inside, but in the early stages, you will need the human to do things like take the truck off the highway to the depot. You might have trucks pull into truck stops where humans wait to take them into town. But to start, you might certify your system only for the wide open and empty road, and so when the traffic gets thicker, the truck will want a human, which means it either stops for a long time (not actually allowed on the shoulder for non-emergency reasons in most places) or it has a human in it already. For now, transfers require leaving the highway.

When I say not solved, I mean that nobody can show they are safer than a human, and only Google is close to showing that. So no, you don't want the robot instead of the human quite yet.

While fully automated trucks would have a variety of advantages, they pose their own challenges (longer stopping distances = longer sensor detection and integration needs; more unpredictable humans in that radius; slower speed than cars means anticipating different merging behaviors; lane changes to make room for mergers only at exits with cars present, etc).

Until that is fully developed, much progress could be made from what you briefly mentioned of an upgraded Peloton. Any reason this was dismissed so lightly? What if there was a human driver with software-assist in the first truck. A series of smaller trucks could follow behind 100% automated in peloton formation. While capable of independent reactions for safety, merging, they have the comparatively predictable lead truck ahead, with its human / software pairing allowing for predicting human-car behavior and operation on undermapped roads. Just as importantly, this can be deployed much sooner than full automation, and with little need for strictly staying on thoroughly pre-mapped roadways because the lead truck can 'map' for the slave trucks while they are only responsible for driving their small part of the road between them and the slave or lead ahead of them.

Smaller trucks and narrower gaps between vehicles (due to faster reaction time; saves fuel too) are both wins on congestion and road wear but offer flexibility. Shouldn't this be easier to create?

In particular SARTRE which I have written about here. SARTRE also had cars after the trucks, which did not work (stone chips destroy the cars, and are not that great for the truck.)

Finding other convoy members also turns out to be hard. The right spacing is hard -- too close is dangerous, not too close has cars always moving in between the vehicles if they can fit. An accident could turn into something truly massive with multiple trucks crashing.

Add new comment