Cell networks are even longer lead time and more capital intensive than autonomous driving ;) even if we only consider the fcc + local permitting time, it already makes this option difficult.
Author here. At the mean time between failures needed to exceed human performance, the uptime of the network connection quickly becomes a limiting factor. It’s possible to pick only routes that have great cell coverage but this limits commercial viability.
Satellite internet is fast becoming good enough and cheap enough.
As an additional safeguard, you can make your trucks go into 'safety' mode when connection becomes spotty or when too many operators become too busy with other trucks.
'Safety mode' could mean slowing down the trucks or even stopping some of them. And in general, letting the autonomous systems err on the side of caution more often.
I believe the MTBF argument still applies to Starlink.
Regarding the minimal risk condition / fallback behavior, a central point of the article was that slowing or stopping are almost always unacceptable on freeways because of the speeds involved
A centralized control like that coupled with a stop as a safe default would be a very juicy attack target. Want to cripple a city / country or create some chaos? Just jam or hack the control system.
Remote does not mean just sitting in a room far away. It could be simply one truck with a lead driver and few automated ones following it closely. Road trains are an option. Even just two autonomous trailers to a lead truck will reduce the driver costs by 2/3rds.
Unlike autonomous cars where everyone needs to go wherever, inter city trucks have fixed well known routes with predicable volume of cargo that can be easily chained together.
Most high end sensors, especially lidars, are targeted at L4 applications. Otherwise the price cannot be justified. It’s a safe bet that sensor makers are including AV developers in their market research.
For lidar, the range is also limited by power limits + physics, which cannot be overcome by increasing money/power/device size. Some dependencies on semiconductor manufacturing tech or better signal processing might be possible to solve with more money.
LiDAR transmit power is practically only limited by eye safety. And retinas are fixed size while the aperture of your transceiver isn't. Get a big lens and you can transmit a lot of power and collect a lot of reflected photos.
The cab, sensors, and compute are also expensive, not to mention other variable costs like staff for remote assistance, maintenance, and first/last mile
Sensors and compute will get rapidly cheaper over time, because they are on the cost curve of electronics. (And they are getting cheaper even if autonomous trucking has only small production numbers, just like batteries will continue to get better, even if nobody builds electric cars: the whole electronics industry is working on those technologies regardless.)
The cab can be a lot cheaper, if you don't have to keep a human comfortable and safe inside. Also keep in mind that an autonomous truck can drive 24 hours a day, and doesn't need to take regular breaks throughout the day nor sleep. (They will need maintenance, but probably not more than a manned truck.)
You are right that the costs are real. Things like first/last mile (or loading and unloading) would probably need a major reshake of the industry, if the truck doesn't bring its own labour, in the form of a driver, with it.
I agree achieving human safety equivalent is the minimum bar. Ex: We can all agree that if your system is below human safety, it is definitely unacceptable.
But issue is that human safety is kind of long tailed. Eg more than a third of all fatal crashes are DUI. I’m guessing if you take out high risk behaviours then the rate will be an order of magnitude lower. What we care about is not being better then an average driver (which basically is bad due to high risk individuals), but better than median driver.
I think better automated collision avoidance as well as automated “you’re driving erratic take a rest in a safe ___location, we can help get you there” are clear wins. But forced autopilot is definitely not there yet and would require a lot of improvement over average driver (because I don’t want to increase my personal risk in order to decrease risk of high risk individual).
Yes but this misses it, autopilot being better than a drunk driver but worse than the median driver isn't enough unless they only use it when they were going to drive drunk which is kinda a political nonstarter. Otherwise the overwhelmingly sober usage of the feature makes it a net loss in driver quality weighted by miles driven.
The problem is that it is better than average human performance but it needs to be better than specific human performance for any human to be incentivized to switch.
For now what you could do is to demand that anybody that has a DUI or other such item on their record to mandatory only be allowed to be in vehicles that have self driving if the manufacturer is willing to assume liability. And if that doesn't happen then they might as well take a regular cab.
Yeah, I should clarify that my response is in anticipation of the comment sections I often see on hackernews about self driving cars, rather than arguments from the self driving car companies themselves. Waymo won't say "we're slightly better than people, let us on your roadways" but I feel like every time I see something about self driving on Hackernews there's a handful of commenters taking the hyper-utilitarian viewpoint of "they're better than people, we all need to let the robots do driving for us" which will never convince anyone outside these comment sections.
They don't operate in snow now, but everybody was like "This is great, but it doesn't work in the rain!" this time last year. I've now taken many flawless rides in outright downpours in SF. It might take a year or two, but I have no doubt snow will be something that is a non-issue in the near future.
Sensing tech has to improve to get through the noise. As far as I'm aware, you can't see a yellow line or white line on a highway if it's covered in dirty snow.
We could also rework our roadways to include better sensing design and tech (passive or active!), but we are a ways out still from willingness to pay for that.
Do what humans do and drive without lines while trying to stay to one side of the road. This doesn’t seem like a major problem to self driving snow cars. More like, how do you deal with other drivers who are RWD with summer tires and are fish tailing all over the place.
If the sensors are up for it... Mountain roads covered in snow pack are a white blur, especially in certain light. Best bets for actually knowing where the road is might be previous tracks. I've dipped a wheel in a ditch more than once, not from carelessness, but snow drift obscuring the road... A road that I drive 500 times a year
I don’t drive at all when it snows as a human. Our whole city basically shuts down then, and we are pretty far north as far as cities go.
Even in places that expect snow, cars are moving much more slowly and cautiously than normal. It almost seems like self driving cars would do much better in those situations given the speeds and caution involved.
That's the smart move. I've been quite astounded at family members without a driving license pushing family members with driving licenses to drive when they thought it wasn't safe. If it snows and it isn't 911 level urgent then simply stay put.
What is the point you're trying to make? Waymo isn't making claims about their driving performance in snow. And they're not comparing their performance to using a basket of human performance in all weather conditions, if you read the linked blog post they're specifically comparing to human driving performance in the regions in California and Arizona they're testing in.
Honestly, I never want a self-driving vehicle of any sort. I like driving, and if my going on about 8 years since even a speeding ticket is anything to go by, I'm pretty good at it. Driving is genuinely a leisure activity for me (only complicated by other, shit drivers) both in the usual way, and in the track-day way. So you could say I'm a car guy for sure.
I am all for autonomous vehicles for other people. And you can call that elitist and I frankly couldn't even argue with you, but god damn, so many people have utterly no business being in control of a car. I travel quite a bit both for work and leisure, almost always by motorway, and some of the shit I see is just mind-bending. So much inattentive driving, so much pointless risk-taking, as they say it's time for Wisconsin's favorite game: Why's that car being driven badly? Old, stupid, drunk, or all three?
And like, there's not really an answer for a lot of this. As much as I love them, cars suck on a society scale. They're basically a poor tax as even the most low-income people in my neck of the woods must have them to get around, there's a bus network but it's shit and it has very low ridership, and that's all my city has apart from taxis. And, old people use them for the same reason, even if they're damn well aware that they shouldn't really be driving anymore, if they don't have someone to get their groceries n such, what are they supposed to do?
So yeah, self-driving cars are awesome. In theory. But then I see video of someone using the lane-guidance on a Tesla where the car just sees... who knows, something, and suddenly jerks to the right, right at pedestrians. Or we get the stories of autonomous cars just shutting down and refusing to move, even for emergency vehicles. Still others, they end up parked on someone's legs, and the stupid support system can't allow a remote operator to move the vehicle off said leg.
At this point, as much as I hope we get it, I also kind of recognize that truly autonomous cars simply might not be viable yet. I think a much better idea is to just de-carify our societies. I don't want them gone entirely, but people need more options than JUST buying or leasing these enormous, polluting, dangerous machines that they do not want to learn to operate well. And like, I don't even judge people for that necessarily? I love them, to be clear, but that's something that's true for me and not necessarily everyone else, and if you don't care about something, you're not going to put your best effort into it. It becomes a chore. A necessary step to doing what you actually want to do, and as a result, people don't try and they just suck at it.
Edit: I'd much prefer and think it's more practical to create cities where cars are an option, not a requirement. If you WANT one, you're welcome to have one. And maybe it drives itself, or maybe you drive it. With those options being on the table, we can then raise the bar for driving tests which we badly need to do, there needs to be a higher skill floor for driving and more stringent requirements, especially as people age. Maybe once you get to a certain point in your life, as most of us probably will, you just can't have a manual car anymore. Then you can buy one that drives itself, or just join people on a robust public transit system.
This doesn't need to be hard, but it's made hard because tons of entities involved have a financial stake in keeping us dependent on cars. I think that's why self-driving cars are getting so much traction and funding, because then they get to sell you a new, more expensive car, instead of actually solving the problems of mass transit. And that just sucks.
Thanks! Pretty alert pedestrian handling in that 1st video. Including one threatening to dash in between cars, one actually doing so, and one (apparently) making to cross and then not crossing. Also traffic cones are now explicitly shown on the monitor.
Author here. the shorter reaction times you mentioned are collected under ideal conditions, like the person knows they are being tested and only needs to push a button or whatever. In driving, the reaction time is end to end, including perception, decision-making, and actuation (moving your foot, pressing pedal all the way, shifting, etc.)
Also recommend checking out the citation. It is an accepted value used in American highway design.
"checksums of file and photo data are used to help Apple de-duplicate and optimize your iCloud and device storage"
This is likely describing content-addressable storage. It is the underpinning of many iCloud services that store user files / blobs. It is also a commonly used pattern in backend services generally.
The problem is that a stream cipher is going to have some per-object uniqueness (a salt, IV, etc.), so by design even if you feed it related input blocks you will get different output blocks. This is, of course, antithetical to deduplication: so you need to check/store the hash of the input before it goes through the cipher.
The presentation about ZFS' native encryption[1] covers many of these sorts of trade-offs necessary to do full-disk encryption at scale.
I guess this post is a little bit different, because VMs are designed to be portable across different hosts. Even hypervisor software without live migration still lets you freeze the VM’s state to a file which can be copied to a new host. However, an already running process is not designed to be portable in the same way.
