Glad I didn't get into trucking

[NichG]

On the ninja-tech thing, that's exactly what this stuff is. With all of these new AI results, what you're seeing is that there was a traditional methodology used in making software and computer technology, but there was also a side-branch of research starting around the '60s which went out of fashion because it was very black-box and hard to prove things about. Since some time around 2006, that technology basically became very, very good for a variety of reasons - some fundamental insights, the ability to exploit much larger amounts of computational power, the availability of extremely large databases to train on compared to the sort of toy problems that researchers had used in the past, etc.

Most companies, regardless of their area of expertise, were not using this technology. So right now, you have a lot of people who are experts with this new technology who are, without particular domain expertise, making solutions that wildly outperform the stuff that was developed via very focused and domain-specific research. So essentially, things that people assumed were true about certain kinds of technologies and certain kinds of problems are quickly being proved to be artifacts of the specific approaches that were adopted, rather than fundamental limitations. As a result, a lot of newcomers are entering the field and showing really high-performance results right off the bat.

Basically the point is, this is a particularly bad time to assume that a given piece of technology you hold in hand represents the actual limits of state-of-the-art R&D projects.

[Vinyadan]

I read an interview to one of Scania's bosses. They are pretty set on this course, and mean that automation is the way to better safety and lower consumption. From what he said it didn't look like he foresaw loss of jobs, but he admitted that it would take some time for the technology to become ubiquitous, mainly because European trucks right now are about 7 years old on average, with areas reaching 11 years. They do aim to driverless trucks, but the costs they openly hope to sink are time and fuel. He also said that laws are a problem and that thet have people studying them, and that, otherwise, they may destroy the whole project.
It was, of course, a very PR interview, but interesting nonetheless.

[pendell]

Well, there's an article in technology review which suggests that we may see self-driving trucks long before we see the cars.

Well before self-driving cars hit the road, fleets of robotic long-haul rigs may be shipping goods across the country.

A startup called Otto is the latest company working on automated truck driving. And Otto’s team includes some engineers from the self-driving team at Google, as well as from Tesla, Apple, and Cruise Automation, who have joined the company to develop technology that turns a conventional truck into a partially automated vehicle.

Trucking could be a smart way to commercialize automated driving technology. Truck drivers spend a great deal of time driving on open highways, which is a lot easier for an automated driving system to handle. It’s also possible for automated trucks to drive in tightly-packed, aerodynamic platoons. And for trucking companies, small efficiencies can add up to big savings.

Wouldn't want to be a human driver in the way of one of those "tightly-packed, aerodynamic platoons", but you can't have everything. Sounds a bit like a train, but without the tracks.

And as Grey Wolf says, it looks like they're going to be concentrating on US-style driving, where you may be driving for literal days from depot to depot in someplace like Texas or Wyoming. Last-mile delivery or lots of local stops, which JustSomeGuy describes, isn't something they'll be doing any time soon.

If JustSomeGuys' driving experience is typical, then it sounds like there will be a lot less need for this kind of automation in the UK than there would be on the continent or over here in the US.

Respectfully,

Brian P.

[Grey_Wolf_c]

Well, there's an article in technology review which suggests that we may see self-driving trucks long before we see the cars.



Wouldn't want to be a human driver in the way of one of those "tightly-packed, aerodynamic platoons", but you can't have everything. Sounds a bit like a train, but without the tracks.

And as Grey Wolf says, it looks like they're going to be concentrating on US-style driving, where you may be driving for literal days from depot to depot in someplace like Texas or Wyoming. Last-mile delivery or lots of local stops, which JustSomeGuy describes, isn't something they'll be doing any time soon.

If JustSomeGuys' driving experience is typical, then it sounds like there will be a lot less need for this kind of automation than there would be on the continent or over here in the US.

Respectfully,

Brian P.

So, I found a nice page on the topic by the US government (damn they collect statistics about everything, don't they?), and a separate one with a bit of detail on the meaning of short and long haul. It's about ten years old, but the numbers should still be in the right ballpark. In short, there is a lot of short haul traffic, but mostly for cheap things that are everywhere (gravel is the example they give). They represent 80% of all haul weight. However, long haul is 80% of the ton-mile traffic, and the long haulers are paid about 30-ish% more than the short haulers (for obvious reasons like being on the road for weeks at a time). Bottom line: there is a lot of money to be saved/made from automating long haul.

Grey Wolf

[Bohandas]

Has anyone seen the Simpsons episode where Homer uncovers a conspiracy where it turns put the trucks are secretly already self-driving and the truckers are getting paid for nothing?

[J-H]

I remember that.

I'm in sales in the transportation/trucking industry. My company specializes in white-glove services; we move mostly un-packaged, sensitive items, like new and used server racks (fully loaded, they can be up to 3000#), medical equipment, copiers, telecom equipment, motorcycles, artwork, and some other things. We do both LTL (Less than TruckLoad) and direct partial/full TL shipment.

Our drivers are typically getting a helper (from local warehouse, or labor ready if they're a long-haul driver in from out of town) and inventorying, padding, and loading everything themselves. That part just can't get replaced by a robot. Flatbed drivers (strap & tarp everything down, re-check straps en route) are going to be harder to replace for similar reasons. The first to go would be the warehouse-to-warehouse dry vans of palletized freight that require no special handling, and that are being picked up from and delivered to industrial areas with large yard spaces to work in.

Driver recruiting is a big issue. The industry as a whole has a shortage of 80,000+ drivers in America, and that's projected to go up. The average age of truck drivers is somewhere over 50 - almost as bad as farming. (1) Who wants to go into trucking and get paid $50k to $100k per year (depending on how you manage money and your truck lease as an owner-op) to be away from home weeks at a time? Plus (2) Who also wants the extra burden of doing actual work loading and unloading, and can do it carefully and well?

We've taken a couple of 5% price increases over the last three years, and if we don't hit the expected recession dip next year, I expect another one. The industry needs to charge more, so it can pay more, to make it more attractive to hire drivers.

[Bohandas]

The average age of truck drivers is somewhere over 50 - almost as bad as farming.

With any luck we'll soon be automating that too

(1) Who wants to go into trucking and get paid $50k to $100k per year (depending on how you manage money and your truck lease as an owner-op) to be away from home weeks at a time? Plus (2) Who also wants the extra burden of doing actual work loading and unloading, and can do it carefully and well?

People who have seen Big Trouble in Little China too many times and developed a warped impression of the industry

[JustSomeGuy]

Well, i can't speak for your lot, but personally i can't move for evil cults prophecising and kidnapping; it really messes up my mandatory rest periods!

[Lord Torath]

Here's an interesting article about how we assess the risk of self-driving cars (and other things):
Slate: Future Tense: The Terrifying Technological Unknown

[Alent]

So... this happened: Tesla's Autopilot being investigated by the government in a fatal crash.

From compositing several articles together, it looks like an 18 wheeler made an unsafe left turn across traffic at speed on a divided highway with no traffic controls (no light or warning lights or stop signs) in Florida to get into a gas station parking lot, and the Tesla and driver both failed to react to it. There's no information about speeds, but from the size of the highway the prevailing speed of traffic would've been either 55 or 65 MPH.

Tesla indicates that there was visual impairment from bright light blinding the optical sensor and possibly the driver. (who appears to have just been distracted)

It's like a perfect storm of everything that could go wrong all lined up together.

[Vinyadan]

I recently read of the chance of autopilots incorporating a risk-evaluating element, with the freedom to choose to take an action which may kill the driver, if this can save other people. Any opinion on this?

[NichG]

I'm not really a fan of that kind of approach - I tend to trust the algorithm over the programmer, and this feels like feeding in some human preconceptions for sake of a press sound-bite ('X company - we sell ethical vehicles!').

To be clear, risk evaluation is important, but I think trying to program in some kind of hard-coded decision standard on the basis of a human ethical judgement process is a mistake.

[Grinner]

...Tesla indicates that there was visual impairment from bright light blinding the optical sensor and possibly the driver. (who appears to have just been distracted)

It's like a perfect storm of everything that could go wrong all lined up together.

That's a bummer, although it does highlight the fact that relying on humans as an emergency failsafe is a bad idea.

I recently read of the chance of autopilots incorporating a risk-evaluating element, with the freedom to choose to take an action which may kill the driver, if this can save other people. Any opinion on this?

I also think it's a mistake but for different reasons.

If we accept the notion that humans are, barring certain circumstances, generally self-interested agents, the marketing department would have a pretty hard time selling a car that is programmed to endanger its occupants when the software believes that might be the lesser of two evils. Doubly so when their competitors offer vehicles without that "feature".

[Grey_Wolf_c]

That's a bummer, although it does highlight the fact that relying on humans as an emergency failsafe is a bad idea.

Google has already said that they have abandoned this approach, because a human simply can't react fast enough when the computer would need him to.

GW

[georgie_leech]

Google has already said that they have abandoned this approach, because a human simply can't react fast enough when the computer would need him to.

GW

To have any sort of chance of doing so, the 'driver' would need to constantly be focused on the road and in position to adjust immediately. At which point, why are we bothering with the self-driving part in the first place?

[Grey_Wolf_c]

To have any sort of chance of doing so, the 'driver' would need to constantly be focused on the road and in position to adjust immediately. At which point, why are we bothering with the self-driving part in the first place?

Indeed. The idea, though, at the start was that the car could predict human need far enough in advance (say, 20 seconds - say, it detects a lane closure coming up that it can't navigate), giving the human enough time to put down the coffee and newspaper and start paying attention. But there are too many situations in which the car does not have 20 seconds' notice.

GW

[Bohandas]

In fact our emotional nature gives us some unique advantages, unlike a computer we can be nervous about a situation that we aren't comfortable with and behave more cautiously. We can see certain behaviors and recognize that they are problematic, in ways that it would be very difficult for the computer to do.

Generally a nervous state of mind isn't in itself helpful for anything. It's a side-effect of being able to recognize potential danger.

[Bohandas]

I recently read of the chance of autopilots incorporating a risk-evaluating element, with the freedom to choose to take an action which may kill the driver, if this can save other people. Any opinion on this?

I'm against it.

The article I read that mentioned this used the example of hitting a truck to avoid hitting a motorcyclist. In addition to the fact that a product should be designed with the consumer in mind, this is a particularly bad case because people shouldn't be given special consideration just because they've chosen to make some kind of dumbass statement.

It brings to mind the asinine Laws of Robotics from Asimov's I Robot stories. All the problems derived from the superflous first and third laws, whereas only the second law ("...obey...") was necessary (and even then only in regard to the bot's owner).

[Grey_Wolf_c]

It brings to mind the asinine Laws of Robotics from Asimov's I Robot stories. All the problems derived from the superflous first and third laws, whereas only the second law ("...obey...") was necessary (and even then only in regard to the bot's owner).

That is quite wrong. "Liar!" is, in fact, a story about the problem of having a robot that obeys, but has an incomplete first law. "Escape!" has a similar situation, in which they downplay the danger to humans so much that the computer disregards law 1. "Runaround" does involve law 3, but only to show the problems with law 2. And so on.

Grey Wolf

[NichG]

I think also there's a comparison to the kinds of rules like doctors' 'do no harm' and lawyers' client confidentiality rights and things like that. At the point that you're trusting someone with your life, even if there could be some greater good served by them betraying your interests, in general we find it unethical. And that's basically what you're doing by giving over control of your vehicle, even if in a much smaller way.

[Grinner]

Generally a nervous state of mind isn't in itself helpful for anything. It's a side-effect of being able to recognize potential danger.

I have to disagree there. Surely it serves some purpose? I'd have to look for studies on the subject, but I would think that being nervous is a fairly commonplace thing. That is, if it were harmful, it might be reasonable to expect that the process of evolution would have resulted in its elimination from the human population, and while humans have been devolving since the dawn of civilization, the dawn of civilization wasn't that long ago from an evolutionary perspective.

[Grey_Wolf_c]

I have to disagree there. Surely it serves some purpose? I'd have to look for studies on the subject, but I would think that being nervous is a fairly commonplace thing. That is, if it were harmful, it might be reasonable to expect that the process of evolution would have resulted in its elimination from the human population, and while humans have been devolving since the dawn of civilization, the dawn of civilization wasn't that long ago from an evolutionary perspective.

According to QI, there was a study done on a bunch of US soldiers. They were in a transport plane, and they were told the plane was going to crash, and that they had to fill in a form making decisions about their last will & testament. They were terrible at it.

Trying to find the details, but google is not being helpful.

GW

[Grinner]

According to QI, there was a study done on a bunch of US soldiers. They were in a transport plane, and they were told the plane was going to crash, and that they had to fill in a form making decisions about their last will & testament. They were terrible at it.

Trying to find the details, but google is not being helpful.

GW

Yes, but filling out documents isn't exactly a skill humans are evolved for. If I had to speculate, I'd say nervousness is more useful for maintaining focus on survival, maybe feeding into the flight/fight/freeze response.

[Grey_Wolf_c]

Yes, but filling out documents isn't exactly a skill humans are evolved for.

And driving is?

GW

[Grinner]

And driving is?

GW

...Not wholly relevant to this discussion either?

I should point out that I never argued that nervousness is relevant to driving, merely that it's more than a side-effect of noticing danger.

Although now that we're on the subject...I'd say nervousness is actually highly beneficial to driving. When I totaled a car some years back, I recall not feeling much. The other car moved in front of mine from a blindspot, and I just hit the brakes as hard as I could. I had no time to be nervous. Therefore, I don't think nervousness affected my decision-making ability. That said, when someone's tailgating me, I do get nervous, and that may help me remain vigilant.

[Grey_Wolf_c]

...Not wholly relevant to this discussion either?

The hell? The topic literally is "is a nervous human driver better than a not-nervous AI driver?", at least as far as I was concerned. I must conclude we are talking about two different things entirely.

GW

[NichG]

Nervousness and similar things are useful for calculating changes in task priority. Its directly relevant with respect to this recent crash. In the crash, what happened was that the sky was very bright, and a white object passed in front of the car that was indistinguishable by the algorithm from the sky, so the car drove straight into that object.

If you have a human driver who is managing to drive, but is finding themselves squinting to see things, almost missing stuff against the sky, etc, probably that human will grow nervous about driving in this situation. That might induce them to change their long-term plans (pull over, stop for lunch until the sun has moved a bit; drive more slowly; etc). So the nervousness may diminish their capability at the immediate scale of the task they're performing, but it can induce them to change judgments they had already made in static form that decide what task they should be doing.

The soldiers in the plane couldn't say 'screw this, we're not going to fill out forms, we're getting out of this plane before it crashes'. So they would have been constantly thinking about how to solve the bigger problem 'we're all going to die!', because that nervousness is telling them 'you have something bigger to worry about than filling out forms'.

But that said, the car AI probably can't tell its driver 'no, we're stopping here for 30 minutes, I don't feel up to this right now'. Though, perhaps thats something it needs to be able to do.

[Kish]

It doesn't require an emotional reaction to make the assessment, "I should not attempt to drive under these conditions," whether well or badly.

It does require making that assessment. The car that crashed was apparently not programmed to assess "are my optical sensors reporting accurately enough to be safe"; it was programmed to drive its course regardless.

[Porthos]

It's like a perfect storm of everything that could go wrong all lined up together.

Piggybacking off of this, 538 had a good look at this and made the very salient point that No Technology — Not Even Tesla’s Autopilot — Can Be Completely Safe

Before people roll their eyes at the Three Mile Island comparison, it was more about the interaction of the humans and technology and how things can break down in panic situations:

The crash reminded me of “Normal Accidents,” a 1984 book by Yale sociologist Charles Perrow. The book grew out of Perrow’s work on the President’s Commission on the Accident at Three Mile Island. In that case, disaster happened both despite and because of a complex chain of safety systems.

The underlying problem was small. No big deal. Just a pump that broke in the non-nuclear part of the power plant. To be on the safe side, though, the entire system went into automatic shutdown. There was a system to stop fission in the nuclear reactor. That worked. There was a system to reduce pressure in the reactor as the fission reaction slowed. That worked, too. There was a system to close the pressure-relief valve once the reactor was back at safe levels. All the computers told the human operators that was working, as well. But it wasn’t, and there was no way to double-check the computer’s report. There was no reason to even assume the computer might be wrong. So steam kept escaping, and the reactor kept losing coolant. Meanwhile, the system that was supposed to add more coolant automatically in an emergency had just been tested to make sure it was working — and somehow, during that test, the workers had failed to reset it properly.

People designed Three Mile Island to be safe. They designed systems that would automatically solve a whole host of foreseeable problems. But nobody expected that those systems might, under just the right set of unforeseen circumstances, crash headlong into one another, creating a disaster nobody could have predicted. And when that happened, the people in control of the systems couldn’t wrap their heads around what was going on fast enough to stop it.

Perrow’s book argues that the people who fail to catch systems as they fail aren’t the bad guys and they aren’t being stupid. Instead, they’re faced with a problem that Perrow calls “incomprehensibility” — in a high-stress situation, your brain reverts to known facts and practiced plans. You respond as you have been taught to respond. But in these “normal” accidents, the problem is almost never the thing you’ve planned for.

Normal accidents are a part of our relationship with technology. They are going to be a part of our relationship with driverless cars. That doesn’t mean driverless cars are bad. Again, so far statistics show they’re safer than humans. But complex systems will never be safe. You can’t engineer away the risk. And that fact needs to be part of the conversation.

I bolded this bit at the end becuase it is absolutely true. You can make things safer, but there is no way to make anything remotely as complex as driverless cars 100% safe, and the sooner everyone wraps their heads around that and accepts it, the better for everyone involved.

[Grinner]

It doesn't require an emotional reaction to make the assessment, "I should not attempt to drive under these conditions," whether well or badly.

It does require making that assessment. The car that crashed was apparently not programmed to assess "are my optical sensors reporting accurately enough to be safe"; it was programmed to drive its course regardless.

If you believe Marvin Minsky, that's exactly what an emotional reaction is.