Hacker News Reader

Having hauled a trailer, I'm not sure how an AI is going to deal with high wind or mountain passes. I think there is a naive understanding of long haul, but it is surprisingly hard to drive with a massive sail on your bed.

Ironically those kinds of things are the things AI deals with most easily. That's all math and physics. The hard part is understanding the environment and adapting the the unexpected.

That is the hard part, and the high wind on mountains create interesting and unexpected scenarios. Mountains are primarily a slow-going thing, but wind bursts create insane environments which are really only felt.

It's going to be a factory of edge cases for both the AI to handle directly and interpret of other vehicles.

There is also the off-highway challenges, you still have all the same requirements as full car automation in dense districts, except now you're in a long and wide truck managing a highly-destructive amount of tonnage. Sure the majority is on the highway, but your trailer has a destination, and your truck a depot, and it's often somewhere dense and busy. I think for full automation soon, it could only happen for specific routes and operations that have set themselves up on the highways with minimal challenges off-highway. Like grain hauling.

You could have the trailers dropped off at depots by bots and taken into town by drivers, but at that point you're building infrastructure specifically for highway hauling and still using local trucks for the last mile, you may as well have built a rail network.

The infrastructure around both long-haul and last-mile or LTL truck transport is already a thing, so in some circumstances the game doesn’t change that much under this model.

Train transport is unfortunately extraordinarily complicated in comparison. It’s a great option in some cases, but a totally non-viable option in many others.

It's true, the other upside to the road network is that it already exists, and so private companies can use it now not after a long period of lobbying to get rail introduced when it does make sense.

It would make more sense to put rfid tags into those rubber reflectors and put those on the freeways and have trucks follow those then do all of the work for full self driving, cheaper too. A human driver can take over for the last mile. The truck would read the rfid tags and each would have a database entry of the gps of each tag.

You still need full self driving I think. Car cuts you off, tire flat where and when does it pull off? Load gets unbalanced, wind shear throws you off line, road ragers, unexpected hazards, unmarked construction zones, weather and road condition changes.. you get it. I think the remote driver is going to be too slow to respond in the majority of cases where circumstances are unexpected. What about load too, not having enough drivers to respond to incidents at peak wouldn't really be acceptable.

I think it's useful to note that we haven't even fully automated trains yet, which should show that even when most variables are accounted for you still need a robust detection and decision engine at the helm.

If low latency, high bandwidth satellite internet becomes common couldn't there be a remote pool of drivers that monitor and take over when needed?

That's very interesting way of looking at it - we do not need to fully automate cars to put millions people out of work. If we just create automated assistants (similar to plane autopilots) that takes care of 80% of work we can then create remote driving centers for people (drone pilots) that take over when thinks go wrong. Add some specialized road patrols to handle the on ground problems (changing tires) and driving truck as a job will go from one of the most popular jobs to an outlier.

Love the idea, but this sounds exceptionally difficult to legislate for and regulate properly. Who is liable in case the remote operator makes a mistake?

What if the remote operator does not make a mistake, but the connection is not stable?

Do the remote operators need full HGV licences? Could they be outsourced to a third party? A third party in another country?

And what happens when air traffic controller make mistake? Or the terminal malfunctions? This somehow did not limit air travel boom. Those are really minor problems when there are thrillions to be made.

> And what happens when air traffic controller make mistake

Then the human pilots in control of the aeroplanes try not to crash into each other.

You don't even need satellite internet... Realistically power might be biggest issue. Otherwise you can just bury some fiber next or under the road and then use terrestrial solutions. Probably with decent beam forming directionally over the road you don't even need that many base stations.

Maybe they should just automate the long sectors and then a human jumps in the cab to drive the truck for the last few complicated miles. The humans would probably prefer that too (for the most part) since they could go home every night.

If they have to share roads with other drivers or the roads are unprotected from the elements, people, animals, debris, etc, you'd still need human-level intelligence to prevent harm or catastrophic loss.

That's exactly what will happen but who's going to pay a trucker 95k a year to park trucks 8 hours a day 5 days a week?

Walmart, apparently. They'll employ less truckers than now since each one will be able to handle more trucks, but they'll need some.

Last mile delivery looks to be about half that, around 50k/year.

If everything else was in place for this to work, I don't think you'd have to pay someone $95k to do it. To me, the $95k pay rate has a lot to do with the time on the road and away from home. Walmart might have to pay $95k for that now, but as autonomous driving takes on some of those miles, the human pay will likely go down (adjusted for inflation).

Airliners have been dealing with high winds for decades.

> I'm not sure how an AI is going to deal with high wind or mountain passes.

Better than humans, probably.

Airliners have it easier though. Hitting a weird wind pocket means a plane might drop 50 feet and everyone on board jolts awake and drinks spill. In a truck that’s a catastrophic crash. And no planes have zero pilots to take over..

My car's traction-control reacts faster than any human; often the dash light is the only indication I have that the car had lost traction for a fraction of a second and corrected for it. I don't see how automated systems would struggle with "sudden" wind changes that take humans half a second or more to react (which is eons for computers at modern clock speeds).

I expect it will be much like harbor pilots pulling ships into crowded ports: the tricky stuff will be handled by highly skilled humans, the easy stuff (long highway miles through cornfields ok flyover states) by AIs.

There will be a few exhaustively mapped and highly trafficked main routes where the truckers will carefully pull out of a warehouse and drive to the onramp, get out, and take a shuttle back to the warehouse. The truck will drive itself continuously for the next 30 hours/2000 miles, and pull itself into a rest stop outside the target city, where another trucker will drive it the remainder of the trip.

I wonder how hard and how much more expensive it would be to build a warehouse designed for automated trucks. Seems like it shouldn't be too much trouble to have closed off sections where only automated trucks are allowed for most of the time. And then they would feed to special access roads designed to be safe and easy to navigate.

What if there is an accident along the main route and a diversion is set up down side roads that haven't been exhaustively mapped?

I don’t think mountain passes are low-hanging-fruit. But high winds should be more doable, in fact with wind sensors on trucks you might be able to have a lot of stabilization that even beats a human. A human is skilled but reacting to stimuli from their perch in the cab.

My father was a truck driver and from the stories I've heard, yes I think it's a very difficult job in a lot of cases. eg. hauling doubles (two trailers linked together) during snow storms down steep inclines and having to fight the back of the trailer jack knifing as it swings one way, then the other, over and over until at the bottom of a slope.

I would imagine that would be exactly the sort of things that robots could actually handle much better than humans. You would load up the trucks and trailers with sensors so that the AI knows exactly what is happening, force wise on every part of the trailer with millisecond accuracy, including full data on things like friction under the wheels, and the road surface temperature and road conditions ahead. Add a detailed 3D map of the whole road and the robot driver should handle this fine. Robots will even have advantages over humans since they can do things like control exact levels of breaking, power and even steering to each wheel independently giving them far more manoeuvring options than what a human has.

I'm sure it would be possible, but I would assume controlling fish tailing tractor trailers in snow would be on the harder end of the spectrum of automating truck driver problems. It is a hard problem and there is a lot more risk to human life than automating a taxi in a city. Even if they perform better than real life drivers, it seems like we are trending towards requiring perfection and we might be a long way out to a scalable solution.

controlling fish tailing tractor trailers in snow would be on the harder end of the spectrum of automating truck driver problems

Yes and no. On the one hand it's hard problem because of all the variables involved, on the other hand it's easy because it's 'just' a physics problem. There is no trying to second guess the intentions of other humans acting weirdly, no trying to decipher hastily erected diversion signs, no trying to interpret what that guy standing in the road waving a sign at you actually wants you to do etc.

> There is no trying to second guess the intentions of other humans acting weirdly

It would still have to contend with humans even in the snow storm scenario.

> I'm not sure how an AI is going to deal with high wind or mountain passes.

More sensors? It shouldn't be too difficult to detect cross-wind or shear forces: I'm not in robotics, but I can think of several different ways to do this (sensing sideways forces applied by trailer to the coupling, or laser emitter/sensor pair or even a camera on the truck with high-contrast markers affixed onto the trailer

This is my sense of things as an outsider, so it's nice to hear an inside account that matches up with that. That said, I wonder if there's some room for human and AI cooperation here to help the driver, perhaps with the humans as safety drivers able to focus on other tasks or perhaps even nap while the truck is under ideal self driving conditions and then take over when there is a situation ahead that requires human attention? From everything Truckers say, it's an extremely stressful job due to a number of factors, and I wonder if there is some way to apply technology to actually make the job more tolerable.

I would _love_ to see people pursue this sort of line. I see people make really simple mistakes that could be easily corrected when they drive. For instance, I think a HUD that told you how far ahead of you, in seconds, the next car was would do wonders. New drivers actually measure this and keep proper distance, but in my experience, experienced drivers adopt the norm of the driving culture, which is too short. I have asked people who care to be safe drivers how far the car ahead of them was, after I counted, and gotten answers that were way off.

It's crazy to me people are so bullish on self driving cars. This is _exactly_ the sort of problem that humans are better at than computers - it's about exercising sound judgement in the face of novelty. I don't think humans are bad drivers, I think we have a toxic driving culture. I think if we found a way to professionalize driving, that didn't undermine the rights of employees by turning everyone into contractors, we could develop a healthy driving culture. Something closer to what exists in aviation.

It is my observation that driving is not something you do with your full attention. It is a means to an end. And that leads to "get there itis", aggressive driving, all sorts of shenanigans.

Couldn’t agree more. Seeing body bags at accidents on just about every commute in a large city (LV) between home/school really made that point for me. It started to feel like that city was engaging in some weird daily sacrifice since no one talks about “gee maybe we should drive more safely?” Or just ludicrously violent.

Road design plays a massive part in accident rates but it's often unacknowledged as a factor. Many roads in the USA combine high-speed design (wide lanes, multiple lanes, clear zones) with frequent turns, intersections, busy shopping plazas with entering and exiting traffic, and generally complex traffic flows. To be safe around turns and intersections, traffic should be slow, but everything about the road design is telling our subconscious mind "it's safe to go fast here", regardless of the speed limit.

> multiple lanes

I'll bite. Suppose you have enough peak traffic that multiple lanes are necessary to transport it even at low speeds. What do you do instead?

I think adding more lanes is a mental trap. It seems like when you have a lot of traffic, what you need to do is add more capacity to your road. But that doesn't end up working the way you'd expect, roads aren't like pipelines, because a road network is not only capacity for transportation but a subtle system of incentives. It's easy to add capacity and lose throughput because you change people's decision making and reap results you didn't anticipate.

So, I would say, invest in a bus system to make better use of the lanes you have. Invest in a subway system to give people safe, cheap alternatives to car ownership. Change zoning so that people don't have to travel so far to get where they want to go.

These are all incredibly expensive, a subway is certainly more expensive than adding a lane, but they have the advantage of actually addressing the problem. Transport really matters and we need to pony up for it, and we'll reap the benefits for generations.

> I think if we found a way to professionalize driving, that didn't undermine the rights of employees by turning everyone into contractors, we could develop a healthy driving culture.

Many (most?) truck drivers today that work as a part of a larger fleet have dash cams and in-cabin cameras that use AI/ML vision algorithms to evaluate how safely they drive.

Coca-Cola really doesn't want to get sued for tens of millions of dollars because one of their truck drivers hit a pedestrian while he was checking his cell phone.

Not most. Inside cameras are still controversial, but outside cameras are almost standard. It's mostly for avoiding lawsuits rather than anything else.

Also, all of the AI/ML is terrible (like almost everywhere else).

> But self-driving cars have been 2 years away for almost a decade now. I really think it's similar to nuclear fusion.

Literally last week it was being discussed that Waymo now offers full self driving in SF right now. Unless I’m missing something, the future you’re saying is akin to fusion has just happened.

https://news.ycombinator.com/item?id=30855624

If you look in to why Waymo started “full self driving” in Arizona and, now, San Francisco it’s because the cars cannot even drive in light rain and the entire fleet is garaged with even mild inclement weather. We are still lightyears away from actual self-driving cars that can manage variable weather and even slightly-unpredictable terrain.

> If you look in to why Waymo started “full self driving” in Arizona and, now, San Francisco it’s because the cars cannot even drive in light rain and the entire fleet is garaged with even mild inclement weather. We are still lightyears away from actual self-driving cars that can manage variable weather and even slightly-unpredictable terrain.

It rains in SF and there’s even fog… The fog’s name is Karl by the way. It’s also an old coastal city that doesn’t have a sane layout that you might expect from a test area like Arizona. SF is not an easy challenge and there are plenty of self-driving startups that have tried there and failed.

> It rains in SF

It hasn't rained regularly for years. We're in a bad drought, or as some other people put it, a new normal of very little rainfall.

Self-driving cars that "only" work on non-rainy days = self-driving cars available 90%+ of the time in sunbelt states.

I have no idea why people minimize level 4 self driving. It would absolutely be enough for most people in those geofenced areas/cities to no longer need a car, would massively drive down drunk-driving deaths, and would be a huge benefit for the blind, the elderly, the rural poor, children, etc.

Maybe it's not hapoening soon, but what they said was that it would happen there first.

You don't think it will happen in simpler environments before more complex ones?

I think the first examples, whenever that happens to be, will be on highways and more than that, only between special endpoints, IE yards with their own ramps. The vehicle never leaves the highway or the dedicated enpoints, no local ramps. possibly special areas of some rest stops for charging. They will be just a new form of port or train station.

He only commented that when self-driving vehicles exist, long-haul trucking will be one of the first jobs replaced. Not how long it would take.

I agree with him. IMO long-haul driving is the most ideal case for self-driving vehicles - the 'spherical cow' of physics. I imagine that they'll self-drive into depots right off the interstates, at which point humans will take over for the last leg.

It might be a bit circular, but I'd go a step further and say that the lack of self driving in long haul trucking is evidence that general self driving is at least a decade away.

Humans need to choose job families that will exist their entire working life (40-50 years). Even if the technology takes 30 years to develop, drivers will be 50 years old and forced into early retirement due to.

I suspect that driver shortages will get worse before they get better, b/c if we are "25 years out" from self driving long haul trucks, ppl will shy away from that job category so they don't get caught jobless before they are ready to retire.

Assuming you live somewhere that's somewhat affordable and self driving is a couple of decades off then 95k/year to follow the FIRE model.

Very very very very few people are both capable and interested in FIRE. Many software engineers could follow the FIRE model and retire before they are 40, but they choose to remain in the work force.

Even FIRE bloggers don't actually FIRE. They continue to support their blog or get bored and return to the work force.

Our brains are wired to "keep up with the Jone's".

What about SAE Level 2 self-driving à la the Comma Three: the driver is still there, but turning and braking is taken care of by the computer. Automate the drudgery to reduce mental fatigue.

The operator is just there to scan the conditions and make sure nothing is going off the rails.

Tow truck drivers are making a killing on youtube right now. Humans cant drive cars/trucks/properly; robots have no chance. Not with the current tech promoted as "AI".

I think I agree as I am not in that space my go to reason is that if it would be easy, we would have fully automated trains with no train drivers already.

Waymo is San Francisco is actually doing a fantastic job. With just what we have there we could do most driving, maybe even including last mile.