Waymo Tests Autonomous Trucks in Texas | Hacker News Reader

Waymo Tests Autonomous Trucks in Texas

ttnews.com

110 pointsapsec11293 comments

Curious what the founder of Starsky Robotics is up to. The site is still up and there's even a blog post after the post about them shutting down: https://www.starsky.io/ https://medium.com/starsky-robotics-blog/the-poor-roi-of-aut...

Here's the post entitled "The End of Starsky Robotics" https://medium.com/starsky-robotics-blog/the-end-of-starsky-...

That's a terrific post; thanks for linking. It seems like Starsky did everything right— they focused on a highly safe system for only highway driving, and an elaborate teleop setup for covering the last mile and exceptional conditions. Basically they just hit a wall as far as VC timelines and expectations.

But I wonder if Waymo can basically follow the same playbook and do it long enough to make it work courtesy of Google's pocketbook.

CEO of Starsky here. My assumption is that they won't - Waymo has a lot of inertia around trying to build full autonomy, and it working in a corner of Phoenix seems to not have convinced that that it's way too hard.

The winner in this space is almost certainly a company that hasn't been founded yet, doing an updated version of Starsky's approach...which is why I've been trying to explain Starsky's approach in detail through blogposts.

60 Minutes continues to show repeats of their glowing feature on them, how they're going to be the next big thing and how they're set to make hundreds of thousands of truckers unemployed.

I haven't watched the full thing again to see if they mention the company has since closed down or that basically everything that was promised in the report isn't coming true anytime soon, or ever.

Website is still up because it's too sentimental to bring it down. I also still have a ridiculous amount of Starsky swag in my closet.

The driver is such a small cost in all of this. Contrary to what the media says, drivers are really not making that much for the amount of time they spend in the truck and on the road.

It would be good if an autonomous system could be more of a supporting technology that works with the driver allowing him/her to rest while it takes over on those long interstate segments of a trip. It would definitely save time and money for everyone. Best of all you still have the driver to complete the difficult segments or deal with bad weather and break downs.

> Contrary to what the media says, drivers are really not making that much for the amount of time they spend in the truck and on the road.

Uhh, what media are you watching that are saying that truck drivers are raking in the dough?

VHS tapes of the Saturday afternoon movie matinee on the local UHF channel, showing ads from the Truck Driving Institue.

Here is just one example. If you're familiar with Google, you should be able to pull up some more articles on the subject.

https://www.washingtonpost.com/news/wonk/wp/2018/05/28/ameri...

On the other hand [1] says "Walmart has upped the ante by raising driver salaries to $87,500 [...] The median annual wage for most tractor-trailer truck drivers across the U.S. is $44,500, [...] only 10 percent of truckers earn above $64,000."

Needless to say, whether those count as "raking in the dough" depends on whether you consider a night sleeping in a truck stop as work time or leisure time.

[1] https://www.cnbc.com/2019/01/28/walmart-is-hiring-hundreds-o...

They're like 43% of costs in trucking though.

It's not just the drivers salary, it's finding drivers (up till covid, trucking companies were getting desperate and jacking up rates here) and the rules about hours behind the wheel. Self driving trucks don't have drug testing, physicals, better job offers, and mandatory rest periods

If a truck has two drivers on board, can one sleep while the other drives?

If so, it should be pretty straightforward - even with current rules - to have the human sleep (or watch movies) while the computer does the long-distance runs, then have the human do the last mile and exception-handling locally (not talking about split-second emergency takeover, just "the truck has stopped because it can't figure out what to do").

The problem with highway driving is that there often isn't time to safely stop and pull over if it doesn't know what to do going at highway speeds.

> If a truck has two drivers on board, can one sleep while the other drives?

There is a difference between "can" and "does it help re: times" - and the answer at least for Europe is no (https://de.wikipedia.org/wiki/Lenk-_und_Ruhezeiten). The "down time" has to be in a stationary vehicle or outside of the vehicle (e.g. hotel) in order to count.

The trucking companies were desperate, but apparently not desperate enough to significantly raise wages or improve working conditions.

uh..not this tangent again.

this isn't a helpful comment

What is "this" tangent again? I might be going off on a tangent here but, I bet it's this novel idea that companies should raise their compensation when they have trouble finding people to fill their open positions.

> companies should raise their compensation when they have trouble finding people to fill their open positions.

So you are anti H1B ? another tangent :D

Another inflammatory strawman comment, not this again...

exactly!

They can't. Trucking is an undifferentiated / commoditized service in a fragmented market. If one company increases pay to drivers, they can't increase rates to customers because they'll just shift their demand to someone cheaper.

As a results rates can't rise very quickly, which means that trucking companies can't really afford to pay drivers any better (obviously with some and slight exceptions, and only talking about US long haul trucking).

And yet shipping prices rose quickly when there was a spike in diesel fuel costs several years ago.

Trucking companies could solve some of their work environment related employee retention problems for free. For example by firing toxic managers, and treating drivers with a little more respect.

It's not that they can't find new drivers, it's that the business model specifically relies on new drivers to make money so they want as many as they can get. The trucking companies finance the purchase of trucks by these freshly trained "independent" drivers, who may own their truck but also agree to a contract to only take jobs from the trucking company that trained them and provided financing. Many of these drivers make almost no money despite delivering loads, and eventually return the truck and get out of their contract, having made nothing or even lost money.

Somewhere around 1.7 million people do that job full time. They average $45k/year. That's $76.5 billion per year before any of the other costs of employing them besides salary.

Automation allows for scaling, isn’t that the point? So that trucks can be simply purchased, dispatched, or retired to alter capacities rather than manually handling individual drivers.

Planet Money had an interesting episode recently of some of the sketchier ways that drivers are exploited. Worth a listen if you're interested - https://www.npr.org/2020/08/10/901110994/big-rigged

It's true that drivers don't earn much in the end when it comes to $ / year—but the "truck" does cost quite a bit: fuel, maintenance, insurance, equipment loans, etc.

The driver is still making ~50% of the overall cost of a truck (whether that amounts to a livable wage is a whole 'nother discussion).

https://www.freightwaves.com/news/understanding-total-operat... has a good breakdown of the cost structure for carriers.

Agree with what you said. However, I believe drivers have a high indirect cost. In fact, in Europe truck drivers can drive only 8h/day maximum with mandated breaks every so often. This means that your truck has, at best, 1/3 efficiency unless you cram 2 or 3 drivers in the truck that can rotate.

Drivers still need to sleep. Can't an autonomous solution run all day and night and increase capacity by ~30%?

If driving pays badly we should still automate it for the sake of reducing "busywork".

The driver severely limits the amount of work the truck can do. A driverless truck could be working pretty much 24x7, which is 168 hours a week. A human driver can only do a maximum of about 60 hours/wk. Therefore each robot truck can do more than twice as much work, saving significantly on capital costs.

Could a truck actually legally drive 168h per week in the US?

I heard that trucks in Germany need exceptional reasons (like transporting medicine) to drive on Sundays, and in Switzerland, vehicles above 3.5t aren't allowed to drive between 2200 and 0500 hours at all.

I've never heard of any regulations in America on when a truck can be on the road, and have seen them on the interstate every time I've driven on one, regardless of time or date.

I strongly believe that focus on drivers is a red herring born from inaccurate sci-fi.

Drivers are necessary anyway, because humans can adapt to more situations. Roadblock? The driver (of the autonomous vehicle) can call dispatch to warn the next vehicles, reroute, and get the blockage handled. Bandits? They can detect and assess the danger, call police, and de-escalate the situation.

But truly autonomous vehicles can drive while the driver sleeps. That alone means reduced latency for the transport of goods over large distances. Even better, drivers can handle multiple vehicles following each other, which means increased bandwidth.

My theory is that the self-driving car industry, similarly to Theranos, at some point switched from a fake-it-till-you-make-it mode to full blown faking it. This was either some time around 2016 when Waymo made a video of driving a blind man around Austin [0] or maybe 2015 when Elon pronounced that self-driving is a "solved problem." [1]

[0] https://youtu.be/uHbMt6WDhQ8

[1] https://youtu.be/xQhb3C2hQoE?t=3510

It could be repeating virtual reality’s steps. There was some hype in the 90s and some attempts that led nowhere. It took 30 years for a real breakthrough to happen.

What was the breakthrough?

Things like smallish screens getting good and cheap enough, leading to the Oculus Rift and HTC Vive.

Or in the case of the Quest, having powerful and reasonably priced SoC's.

I'd say the breakthrough term was not the ideal word choice - it's more that technological advances passed a threshold:

For VR, a lot was advances in (affordable) processing power (latency causes significant issues in VR), and screen (manufacturing) technology (i.e. pricing improvements).

For self-driving, it's again processing power (with the focus on the ratio between computation power and power consumption/size), but also the advances in technologies like Lidar, as well as advances in theoretical research (and better understanding) of machine learning.

You’re in for a treat if you haven’t heard of the oculus quest :)

I presume Waymo has been knocking on every trucking company door in the phone book, trying to get some sort of pilot up, before opting down this rather costly path.

I think the primary fallacy in the theory is confusing companies like Waymo with companies like Tesla, and mingling their (in)capabilities.

Both have a very different approach in trying to get to self-driving, and when combining what both can't do, it would seem nobody can do anything. And whilst both throw out their fair share of "demo videos" (=advertising), they have different track records for what they can (and actually do), and what they can't (but claim to do).

You just named the two companies closest to autonomy, IMO. Waymo has the lead in actual self driving, Tesla has the lead in actually commercializing features. I'm confident these two companies will reach mostly autonomous driving before many of the other fakers.

I named these 2 companies because the parent referenced them in their argument. But I agree that they are the ones that are the ahead in their individual approaches, though I think Tesla is not really that near to autonomous driving. They have a lot of features in development that their test drivers (=consumers) are using, and that gives the impression that they got that already (Elon even claimed Tesla has FSD on highways in early 2019), but the step to "the driver doesn't always have to pay attention" is gigantic and far away - and on that scale, Tesla's lead to other manufacturers (who don't do that kind of marketing or testing on consumers) is relatively small.

They're still doing better than a lot of the wannabe-unicorns that pursue self-driving or Lidar though.

Tesla has certainly got the furthest forward in releasing features - but I fear they've gone a long way down a dead end by putting all their eggs in the we-don't-need-LIDAR basket.

And that decision means their Level 1 autonomy is nowhere near good enough to form a foundation for Level 4 autonomy.

Except we've seen Waymo doing fully driverless rides with customers in the Phoenix area.

They've obviously gone a lot slower than planned though.

Waymo is doing fine but they need more time (a lot more).

Tesla, on the other hand, is completely faking it. They have nothing to show for yet.

Nothing to show except thousands of vehicles collecting data and doing partially autonomous driving, you mean. That's way more progress than other auto manufacturers, most of whom are going to be riding coat tails of a tech company to autonomy.

Talking about data collection is a slippery slope. Surprisingly little of a Tesla ride gets transmitted, so whilst it seems like "A huge fleet of cars out there is collecting driving data", only a tiny amount of it is used to train. And so far, Tesla has a driver assist feature, and a lot of promises that get postponed every time the deadline is up.

Their current set of feature outshines most car manufacturers, yes - but you have to remember: Tesla seems that far ahead because - unlike most other manufacturers - they push their beta-features to the production fleet.

[Edit]: I should note: I don't subscribe to the sentiment that Tesla is "completely faking it" - I think they very much oversell their capabilities, and are nowhere near the point where the driver can stop paying attention (which to me is a core component of autonomy).

Certainly miles/time for adding training data and to refine driving rules is important. But there's also a naive belief out there that piling on the mileage somehow directly leads to "better driving" and that's mostly just not the case.

That doesn't mean they're anywhere close to self-driving. Anyone can start collecting data and do partially autonomous self-driving with a modest investment - this is a solved problem.

California's DMV won't let autonomous vehicles over 10,000 lbs. on the road with just a testing permit. CA's testing permit for autonomous vehicles is much like a learner's permit - no passengers for hire, no heavy vehicles. Which makes sense.

If you're interested in trucking, the podcast 'Over The Road' is worth a listen. The host Long-Haul Paul has a voice to die for, and the oncoming change to autonomous trucks plays a part in some episodes. I think there's 12 or so episodes in total

https://www.overtheroad.fm/

I never expected that Google would enter trucking business.

It's a multi-billion dollar market that they could, potentially, completely dominate using a technology solution. That's exactly the industry Google goes after.

Would you like to be the last car in a high way traffic jam when you see one of these coming closer in your rear view mirror?

Sure, if it hits me I know I'm getting a multimillion dollar settlement.

At least your funeral will be paid for then.

It's not an Uber self driving vehicle. It will probably have at least some safety features.

Lidar-based systems are pretty good at detecting stationary objects. Humans on the other hand crash into traffic jams pretty regularly.

So in that situation I'd prefer that to be an automated truck.

https://www.logicallyfallacious.com/logicalfallacies/Relativ...

I'm expecting to see waymo autonomous vehicle fatalities in the near future!

You mean like this? https://www.11alive.com/article/traffic/fatal-wreck-75-north...

Even a simple level 2 system with driver monitoring like comma.ai in big rigs would save tens of thousands of lives per year.

I'm not sure tens of thousands of people die in accidents with big rigs every year (IIHS says around 5,000)[0].

Additionally, many big rig accidents could not be avoided even with perfect driving - they are caused by things like weather, other drivers doing stupid things, sudden stops in traffic, etc. AI does not really improve your stopping distance in a big rig.

0. https://www.iihs.org/topics/fatality-statistics/detail/large...

> they are caused by things like weather, other drivers doing stupid things, sudden stops in traffic, etc. AI does not really improve your stopping distance in a big rig.

AI doesn't need to improve your stopping distance, just your propensity to follow by more than your stopping distance. Even simple adaptive cruise control can do that.

> AI doesn't need to improve your stopping distance, just your propensity to follow by more than your stopping distance. Even simple adaptive cruise control can do that.

Do you write AI software for self driving vehicles?

> AI does not really improve your stopping distance in a big rig.

It doesn't improve the braking distance, but it should be able to improve the stopping distance if it can react faster.

Hmm, my general point is that there is a lot of energy in a semi running at highway speed. AI could help, but that energy doesn't just go away.

An autonomous truck could drive twice as conservatively as a regular truck driver or even just pull over in bad weather and still deliver its load faster than a human driver. A human has to stop for 10 hours every 11 hours, while an AI truck and just sail on through. Tandem drivers might be competitive on time, but now you're paying two humans!

> Tandem drivers might be competitive on time, but now you're paying two humans!

But you're only paying them for the time they're driving - so this is actually incredibly efficient.

That's not how labor laws work in many countries - if a driver is sleeping in the car, and not in his own bed at home that still counts as working time.

But they reach the destination in half the time.

I actually looked into this, you're half right. Truckers don't get paid by the hour, rather they're paid by the mile (or km). Each member of the pairing gets paid for half of the total distance of the journey. However, tandem teams are in high demand as customers are willing to pay a significant premium to get their goods delivered quickly. This has resulted in team drivers earning more for a unit of travel than solo operators.

That article on efficiency comes to mind: currently businesses have some slack built in on their deliveries because they can only arrive so quickly.

If AI trucks are the norm, and all that slack is eliminated for efficiency sake because the "AI truck can always get here in 13 hours, no stopping", then that increased efficiency only raises the urgency that it arrives on time, thereby eliminating the time savings on delivery.

It can actually improve stopping. Theoretically you can pack some systems to disintegrate the truck in case of an emergency, you don't have a driver inside to save.

The one big rig accident that I know of personally was road rage - a car cut off a truck at at 3AM on the entry into Phoenix, and the rig got even by running the car off the road.

Now, that is entirely anecdotal, and I have no idea how common such things are, but these type of things would at least be avoided by self driving.

I wonder if Self Driving Vehicles will retaliate by reporting bad drivers to the police instead of traditional road rage.

That is a completely unjustified statement. The prevailing rate of large truck fatalities in the US in 2017 was 1.6 per 100M miles which is ~1 per 60M miles [1]. The comma.ai website markets that they have driven 25M miles, so there is insufficient data to conclude that such a system would improve safety at this time.

[1] https://www.fmcsa.dot.gov/safety/data-and-statistics/large-t...

>there is insufficient data to conclude that such a system would improve safety at this time.

The hype behind self-driving is getting out of control. Sure, from a technological viewpoint it appears that, one day, computers should be able to drive better than humans.

But none of this has been proven yet.

I'm actually curious whether there are any trustworthy studies showing that driving automation helps safety, for any type of automation. I can buy the claim that emergency braking helps (although not as much as one would expect [0]). But is there good data that shows that cruise control helps? Lane keeping?

[0] https://www.theverge.com/2019/10/4/20898773/aaa-study-automa...

That AAA test uses a disintegrating dummy. So it looks really bad - the pedestrian gets smashed to pieces. But notice the car stops in most of these sequences almost immediately afterwards. That's because the vehicle was already braking, significantly reducing the impact energy, but it was hard to see because the AAA chose to shoot it in the most dramatic way.

The article does mention the main things Americans could do if they actually wanted to reduce the number of pedestrian deaths, which they don't - stop buying SUVs and other oversize vehicles that disproportionately kill pedestrians.

I suspect that most Americans care more about not killing themselves than in preventing overall number of deaths.

The chances of surviving in an oversize vehicle are disproportionately higher than in a tiny car (what I have heard referred to as an "aluminum covered electric bike").

yes, it's an arms race/prisoners dilemma, which is why the government needs to step in and align incentives. Personally I think a danger tax (similar to gas guzzler taxes) makes sense.

The pedestrian dummies in the AAA tests are standard test dummies. They disintegrate so that the test driver does not get hurt in the case of an impact. The video looks really bad not because the AAA staged it this way, but because it's really bad when a vehicle driving at 20 mph in broad daylight hits a pedestrian that's stationary or moves slowly along a line.

Emergency braking and automated cruise control appear to help a little but lane keep assist has no safety benefit.

https://www.iihs.org/news/detail/benefits-from-advanced-driv...

You can have fatalities and still have several orders of magnitude fewer fatalities than non-AV.

This is true, but ignores two things:

1. People will (over)react to autonomous deaths. Every AV death "matters" more.

2. Getting the death rate below non-AV is difficult to test. Accidents are rare. Are you going to let these things loose on the public for hundreds of millions of miles before determining that the death rate might not be better yet?

A quick search shows in January 2020 that Waymo has already driven 20 million miles. So I guess that’s the plan?

fatalities probably track pretty closely with the overall number of injury-causing accidents, of which there are over 100k per year in this country where large trucks or buses are involved. So, hypothetically, rolling out at 1% and closely monitoring all accident types seems like a fine plan imo. At least one type of fatality will be eliminated straight away --- no truck drivers will be dying :-)

1. And any accident of a properly-used/maintained vehicle will be viewed as some company's fault because of design/software flaw.

2. There will doubtless be accidents that will be quite different from the sorts of accidents humans have. And those will be viewed as proof that these cars aren't safe.

Reply on news.ycombinator.com