by kevinchen on 1/12/24, 8:18 PM with 427 comments
by beedeebeedee on 1/14/24, 11:50 PM
(I used to work for an autonomous truck company, and when you factor in the cost of roads in addition to the development of the trucks, it makes absolutely no sense to do autonomous trucking when you could do trains. As a culture, we've been brainwashed not to fund trains. We collectively spend billions and billions on roads but would not dare spend money to build more tracks. It is shocking and ludicrous, but that's what happens when you suck up a century of propaganda from the fossil fuel and automotive industry).
by aga98mtl on 1/14/24, 10:57 PM
by karaterobot on 1/14/24, 7:46 PM
> Therefore, all safety-critical decisions must be made by the onboard computer alone.
Why is the requirement that all safety-critical decisions must be made on board, versus the seemingly-simplifying assumption that only some or most decisions would have to be made on board, because a remote operator or backend service could be available a lot of the time? It doesn't seem unreasonable to me to have a single operator remotely monitoring multiple vehicles that are autonomous under ideal conditions (driving along a straight road in good weather) and then taking over when necessary. Let's say you would only use such a system for major routes with solid satellite visibility, not last-mile routes hauling heavy equipment on a dirt road in the boonies, or something like that. Maybe this wouldn't work, but it's not obviously ridiculous to me, so I wonder why he just starts out saying the truck most be fully autonomous with no human ever in the loop.
by mxwsn on 1/14/24, 5:42 PM
This analysis seems really suspect to me. Any clarification would be appreciated.
by zbrozek on 1/14/24, 5:40 PM
There are a lot of differences between passenger vehicles and trucks. The physical dynamics of articulated vehicles, the mission profile, and social dynamics come to mind. How does a robotruck place cones or flares while it awaits rescue?
Personally, I expect autonomous trucking to be a force-multiplier for humans who were formerly drivers. Such trucks will have sleeper cabs and the human will be there to maintain the vehicle and handle the long tail of tasks (filling tires, cleaning, refueling, repairs, rigging, whatever). You'll get 24-hour operation out of a single human employee because they'll be able to sleep and do other things most of the time. Maybe they'll work a second job as a remote call-center operator.
by ajsnigrutin on 1/14/24, 11:32 PM
This is a video from our firefighters here in slovenia... Humans are shitty drivers, sure, but how the hell would a computer react to a chaos like this?
by michaelgiba on 1/14/24, 5:30 PM
by Edman274 on 1/14/24, 5:49 PM
Also, I feel like there's a lot of talking past one another in these conversations because one person will say "Let's see an autonomous truck shipping hazmat to Pittsburgh in February with freeway lanes shut down" and another person might say "that's a rare instance" but I really don't feel like society will accept anything other than trucks / vehicles that are able to operate under all conditions, with greater safety than the safest human driver. We tolerate human failures but to use them as the benchmark for autonomous systems would be perceived as unethical, because autonomous systems are deliberately designed and any failures by them would be seen as an intentional oversights and errors, and no one at Waymo or Tesla or where ever is ever going to be charged with vehicular manslaughter for an autonomous vehicle error. We'd demand a way higher standard because these companies don't really have any skin in the game, except for financial penalties which we now understand is not a deterrent for anything. My observations are only moderately related but I'm anticipating the same well-trod talking points coming up and want to address them.
by prof-dr-ir on 1/14/24, 5:34 PM
It would be interesting to hear what a company like Aurora and their backers were thinking back in 2020. It seems unlikely that they did not come up with these arguments themselves, so what were the counterarguments?
by sva_ on 1/14/24, 6:56 PM
It seems like there are a lot of 'default routes' that trucks take, and in the beginning, it'd probably just be a small number of highway routes that could get clearance in such a way, and then scale up.
by hinkley on 1/14/24, 6:12 PM
Growing up we had a neighbor who drove UPS trucks and was rated for double trailers, and that earned him extra money. I gotta assume triples make even more. Doubling the payload per driver mile by making it so any trucker can manage doubles and the old double drivers can manage triples seems doable.
The thing with doubles and triples though is that they are shorter trailers. And it’s probably less likely to pack a triple trailer to the gills than a single, so are they getting 70% as much payload per trailer? If so that puts triples around 2x the practical capacity of a single.
by awinter-py on 1/14/24, 11:20 PM
48k miles in the interstate = 80k km, 400k posts, 4 billion dollars (and you can start regionally)
could also run semi-autonomous convoys where a robot truck follows a human truck, extending sensor range
by RockyMcNuts on 1/14/24, 6:21 PM
by ramboldio on 1/14/24, 7:26 PM
E.g. dedicated lanes for self driving or hand-off stations where human drivers can take over etc etc.
Amid the massive potential upsides, any reasonable government would invest in measures that turn out to effectivley improve self-drivability of the respective national grid.
by mchusma on 1/15/24, 3:16 PM
Trucking and truck routes do have a strong incentive to switch eventually. But the examples given of Cruise and Waymo are losing tons of money.
Trucking companies would need to rework their business model around these hubs. So they won’t invest in this yet, not until it’s closer.
And for parties trying to make self driving work, it makes sense to try it on smaller, less expensive vehicles first. It’s the number of miles driven that drives learning rates, and if your cost per mile driven is lower, you get more learning per dollar.
And achieving true autonomy also needs a variety of scenarios, not just freeway. So being freeway only isn’t quite the advantage you would want. A million miles of city and highway driving helps you learn more than a million miles of highway driving. I think regulators know this intuitively too, and are more prone to approve.
by AtlasBarfed on 1/14/24, 5:29 PM
Anyway, the speed doesn't have to be 50mph. Do 40mph on late night trips. More EV range/gas efficient anyway.
The deceleration would likely be helped by an EV drivetrain where regen and braking can contribute to stopping. Likewise the high-torque EV motors can accelerate a semi far more manageably.
Emergency/Fault stops should likely be handled by convergent infrastructure, or only using routes with sufficient shoulders.
How many scene understandings can be done with on-demand manual takeover? Highways generally have some of the best cell/data networks, and again, convergent evolution of infrastructure.
Sensor distance can probably be improved with the concept of a "scout car". Likely automated trucks would drive in formation (to draft to get better efficiency), a lead car can scout and provide additional effective distance. This car doesn't even need to be a full size car, it could be a smaller teardrop-optimized drone. And if the drone crashes or faults, the following train automatically pulls to the side.
Trucking routes can benefit from neural nets tailored to repetitive navigations of the route. No handling general queries, the tractors (or the control system utilized) isn't driving "a general truck". It is driving from Minneapolis to Chicago.
I think I disagree with the article. It assumes too fast for the vehicle, too general of routes.
Oh, my final disagreement may be the total size of the vehicle. Once automated vehicles can drive at distance ... do you need a massive tractor trailer, or can you instead use more capable (and more cheaply mass produced) vans or smaller form factors? The industry is what it is in current state and a trailer ~= a container, so I guess I see why that wouldn't be a go.
by 11thEarlOfMar on 1/14/24, 5:32 PM
Many cars monitor rotational speed of the tires in order to determine whether a wheel is slipping. A Model 3 can detect whether a tire is new or old based on tread depth, something I learned the hard way.
If trucks were outfitted with rotational encoders on all tires, 'slipperiness' could be monitored periodically by, for example, tapping brakes and checking whether all wheels continue to rotate at the same speed. In emergency braking or maneuvering, the rotational speed of all tires can be input to the recovery algorithm to perform braking. If the brakes are independently controlled, this can be performed per wheel.
More generally, trucks run in fleets and on routes. Similar to aircraft reporting turbulence so the next craft can adjust, trucks can report road conditions so that approaching trucks can adjust their speed or pull over completely if conditions are so bad.
by silvestrov on 1/14/24, 5:29 PM
Driving a car in inner Copenhagen is a stressful situation due to the insane number of cyclist you have to watch out for, see https://youtu.be/FaySp9i2zMA?t=113
by gwern on 1/14/24, 6:36 PM
by intrasight on 1/14/24, 5:10 PM
And an arbitrary one, in my opinion. We will be racking up millions of driven miles by hundreds of autonomous trucks before that bar is reached. We will have drivers in the trucks. We will have connections to remote operators.
by Havoc on 1/14/24, 6:34 PM
If anything driverless helps dramatically on this front. If you can run trucks 24/7 without stopping then you can cut their speed say 33% or whatever and still come out on top.
You can also adjust the cargo to reduce stopping distance - use autonomous truckers for things that are large but not heavy.
Or hell even just add more wheels for more stopping traction. Sure gets you more wear & tear but that's an infinitely easier problem than trying to teach an AI to recognise a kid running into the road in a busy urban steet or navigate construction site or whatever other surprises come up.
Author makes some valid points - stopping distance vs sensor range - but to me the conclusion that trucking on highways is hardest doesn't follow at all.
by jeffparsons on 1/14/24, 9:50 PM
Let's get a little bit crazy and imagine something very different from what we have today, but still only requiring a single driver. One truck cab at the front. Individual trailers have their own wheel motors, their own steering, and a small battery for short range manoeuvring (e.g. at a loading dock/port or). Trailers follow each other at a distance that allows for emergency braking, but close enough that they can be connected by high voltage cables for receiving power from a "battery car". Both the cab and individual trailers constantly communicate, and the moment any participant suspects that all is not right, the whole train tries to slow down and pull over.
If one trailer has a blowout or other issue, it can be left behind to be picked up by another convoy.
At a port or loading dock, the convoy disassembles and individual trailers are driven by remote control, either directly by a human or eventually by the "site AI", so that large articulated vehicles never have to enter the area.
I'm imagining ~12 trailer convoys to balance efficiency gains with disruption to local traffic.
by ahigherugliness on 1/14/24, 8:09 PM
What problems am I overlooking that are not easily solvable?
by theptip on 1/15/24, 2:54 AM
This approach would help data collection (since you have a ground truth from the drivers) and would also mean you have a mechanic on hand to fix the out-of-cab issues that inevitably come up.
by samtho on 1/14/24, 5:24 PM
They claim that stoping distance and non-normal driving conditions are an issue but these are precisely the same problems that AVs meant for ride share face but with a difference performance profile. The computer doesn’t care if it’s hauling 20 tons or 2 tons, what matters is we give it the correct performance profile and install adequate equipment to operate the vehicle safely.
The only real argument I see here is the limp mode case where it detects an error and must pull off. I can see the actual freeway being a problem here but then we just make sure the routes they initially travel have been checked prior to deployment - it’s a finite road that can be mapped and tested against. This can be further mitigated by having redundant computer systems and redundant environment sensors than it can fall back on to be able to safely pull off the side of the road or even make it to its next stop. Computers can make faster, more-informed decisions about nearly every driving situation and the challenges with operating heavy machinery are well understood.
The reason autonomous trucking has not been developed is not because of R&D, it all has to do with scale and what the funding is chasing. Running a truck is expensive, especially with a large load. Financing a project would have a lot of money dumped into the just the construction or retrofitting of a truck and it’s operation for roughly the same computer as the smaller vehicle - just with a different performance profile. When the vehicle you are testing has a kerb weight of 1100lbs and is electric, your maintenance costs plummet to almost nothing with fuel being electricity. We don’t have the same luxury with trucking. Additionally, Uber, Lyft, et al, are bullish on this technology because they would rather collect fares and pay a fixed cost for vehicles than having to deal with drivers that want silly things like compensation. The gig companies are also looking down a barrel of a loaded, regularly-shotgun that could drastically cut into their revenue. The dollars they are investing into AVs only care about ride share now.
by oatmeal1 on 1/14/24, 7:42 PM
Except in the case of Cruise.
MRC could be handled somewhat successfully for trucks on the highway on sections of highway with good cellular data connection. You could have someone making minimal money play a driving game on the computer all day to keep them paying attention. If an autonomous truck encounters an unusual situation, their screen switches over to a camera feed for the truck, and they can take over the controls of the truck to get it past the obstacle.
Not sure what fraction of highway has sufficient connection to give them a sufficient live feed. Probably a lot though.
by lettergram on 1/15/24, 4:16 AM
This is a weird way to think of it, I don't actually think of it as statistics. The failure cases will be different (between human and AI). Further, we want to be able to sue those responsible (not some nameless company).
Final thought, I wonder if this is the metric, when we'll start seeing insurance remove humans or increase the premiums. Basically, exponentially driving everyone into AI driving
by creer on 1/14/24, 5:40 PM
- Sensing distance is a good one for sure. At the same time, it's for now a technical issue. So that a viable temporary progress-making solution might be to have an additional vehicle in the autonomous system: a sensing car is in front, and one or more semi-trucks follow. Then the autonomous system far out-senses a human driver. And we get progress even if the cost-proposition becomes a little harder.
But otherwise:
- Nominal stopping distance as shown include 2.5 seconds of reaction time - fair for a human driver perhaps but hopefully the machine gains a lot of space there.
- "Stopping in lane" is as insane in the city as it is on a freeway. We can't seriously consider that a long term solution in the city? Less risk temporarily okay but not a serious solution. (But fair point on this one: if all else fails it is, for now, an option in the city).
- Even for humans, driving on the freeway is not a question of perfect judgement. It is about quickly but calmly picking a reasonable solution. Including, in doubt, moving to a safer lane and slowing down while measuring the risk of rear-ending. Nobody asks a human to completely eliminate the risk of rear-ending. That sounds within the range of what software can do?
- Jack-knifing and other complex behaviors are probably far better suited to computing solutions - which can sense far more measurements than the humans, independently actuate more control points, and compute actually more or less complete solutions when the human is seat-of-the-pants -ing in real time (with dubious real life record too - see brutal and fatal). Simple no, but all that hard to beat the human sounds absurd (anyone with more input on this?) See for example the drift-parallel-parking stunts achieved with cars. We may not yet be at drift-parallel-parking semi-trucks but anyone here doubts that this is achievable?
- Ease and quantity of samples might be a better point. Waymo really got on it and accumulated a lot of driving data. But then also, most of that data by now is city driving.
- Will trucks happen before cars? Well, cars are already there so... already answered.
by ZeroGravitas on 1/14/24, 6:54 PM
Air resistance means slower speeds are more efficient (you don't need to worry about human driver needing paid per hour or limited by regulations).
Less other traffic at night, but (presumably) sensors can be designed to work well in these situations.
If you split a truck in 3 it can travel together as a pack for aero yet reduce stopping times as each part brakes seperately.
You could probably even section off an entire lane in certain areas.
Some of these possibly needed EVs to mature more before they become practical.
by samstave on 1/14/24, 6:19 PM
A train that will take a smaller number of cars ONLY between point A and B. Then at each end offload to a truck.
Not the same as current rail frieght, I am saying a dedicated track between say port of oakland and port of long beach - with some intermediary way points along the way - it just has two lines - there, and back with a dedicated automated train that simply goes back and forth - with gantry ports the train drives beneath at certain locations to lift off cars quickly for that node.
We have done a poor job of leveraging rail in a future-proof manner...
by iancmceachern on 1/14/24, 8:00 PM
Edited - typo
by gumballindie on 1/14/24, 6:01 PM
The trick in driving is not straight lines and clear weather. It’s rainy, snowy, foggy, conditions that make it challenging. Edges cases are the real difficulty, as with everything.
I suppose when your knowledge of the world is limited by a few multiplayer notepad interactions everything is simple.
by light_hue_1 on 1/15/24, 2:36 AM
The reason autonomous trucking is exciting is platooning. You have one truck with a person in it at the head of a long chain of trucks.
This avoids all of the issues mentioned and it's the near-term future of trucking. Unfortunately, this will eliminate many long-haul trucking jobs.
by WhatsTheBigIdea on 1/14/24, 5:46 PM
It seems the basic argument is that sensor range is not always long enough to fully encompass stopping distance, but the data presented in the article shows that this is only the case at 70mph.
At 50mph, it seems that sensor tech is well within requirements in all published conditions.
Perhaps an autonomous truck could simply slow down when grade or surface conditions reduce its stopping distance?
While it is nice to do 70mph all the time, slowing down when you can't see or stop well is exactly what is expected of human drivers.
by jillesvangurp on 1/14/24, 8:27 PM
There are a few other possible reasons; such as electrical trucks still lagging a bit behind electrical cars in terms of volume production and a few other things. Electrical trucks are a lot easier to drive for humans because they don't have a lot of complicated gears, hydraulics, and other things to worry about. That makes them easier to control autonomously as well.
Other than being bigger, the job of autonomously controlling a big truck isn't that different from that of controlling a car. If you can do one, you should be able to do the other. Ballpark it's the same kind of problem. Same roads, obstacles, problems, traffic situations, challenges, etc. Same everything, basically. I don't think the speed or the size of the vehicles is that much of a factor for how hard this is to do.
As for safety, the barrier is actually pretty low. There os a high number of accidents that involve truck drivers asleep or unfocused behind the wheel because they've been driving for way too long. Or getting distracted by their phones. Happens more in the US than in Europe because some differences in rules related to mandatory breaks. But it happens on both sides and it's a bit of a challenge.
We're measuring with very different standards for humans and autonomous vehicles. With autonomous driving it's all about hypothetical things that may or may not go wrong and therefore we should wait until it's so perfect that it can never happen, 100% guaranteed. With human drivers, it's just an endless stream of never ending accidents, fatalities, and misery where we just go "ah well, that's just life". Never mind all the obvious issues with aging truck drivers that are severely out of shape due to life style issues associated with trucking after they've been on the road for ten hours. This attitude is more than a little bit irrational.
by aurizon on 1/14/24, 6:50 PM
I can see that this would allow integration with road trains, as methods evolve.
I saw several posts that spoke about the ground level interaction of bicycle traffic and car traffic being dangerous to bicycle traffic. Years ago we had elevated railways, that came to be unsightly, but functional ways to mix subways(elways) - it made me think that elevated bike-ways might work to reduce bike/car interactions. Bikes are so much lighter in weight that a suspended bike 'mono-road' would work well, it could even be roofed against weather. It has been done in a few places, and seems to work. https://www.google.com/search?q=bicycle+elevated+path&sca_es...
by lvl102 on 1/14/24, 6:08 PM
by acd10j on 1/15/24, 5:01 AM
by mcmcmc on 1/14/24, 6:48 PM
by mschuster91 on 1/14/24, 6:15 PM
The question is... isn't the problem being approached from an entirely bad premise in the first place? Why would it even make sense to develop autonomous trucks for that part of any freight's journey? Laying train tracks might incur a large initial expense since the tracks, unlike highways, aren't there anyway, but a train is way more energy efficient as only the first locomotive has to deal with air resistance and steel-on-steel is less friction than tires.
by ChuckMcM on 1/14/24, 7:37 PM
by aserafini on 1/14/24, 6:09 PM
Is this really correct? What about gatik.ai, they certainly appear to have driverless trucks on the road do they not?
by whiplash451 on 1/14/24, 8:21 PM
by cycomanic on 1/15/24, 2:01 AM
That said the conclusion from the article is really that self driving is hard and we likely will not see it for many years.
by jjtheblunt on 1/14/24, 7:01 PM
by ramboldio on 1/14/24, 7:28 PM
by shahbaby on 1/14/24, 9:17 PM
by stevage on 1/13/24, 1:06 AM
by CSMastermind on 1/14/24, 7:48 PM
It turns out we already have general intelligence (producible in about 16 years) available on demand. And so, the market is already a good indication of which jobs are easy and hard to automate for general intelligence.
by Log_out_ on 1/15/24, 6:53 AM
by uniTrial on 1/15/24, 10:27 PM
by jdksmdbtbdnmsm on 1/14/24, 9:56 PM
by skywhopper on 1/14/24, 8:08 PM
by bradley13 on 1/14/24, 5:36 PM
Stop the truck using a couple of vehicles, cut open the lock on the rear doors, and loot. Pick an isolated stretch of toad, and the pirates will be gone before anyone can respond.
by johndhi on 1/14/24, 5:22 PM