Waymo seems to be the only question mark in my mind making me doubt my bet that we're a decade or more away from usable self-driving cars.
If Waymo has truly solved the problem while everybody else is just trying to catch up and/or bluffing, I guess I wouldn't be that surprised.
Edited to add: The reason is that they've apparently ordered 10's of thousands of these cars. Can they really use that many only in good-weather places with perfect roads? It seems unlikely that they are just that stupid to spend that much money without being confident that their plan will work. One last option would be that they intend to put 10's of thousands of them on the road just to exponentially speed up the data collection that they've decided is necessary to truly get the system up to par.
I tend to agree. Waymo ordered 62,000 Chrysler Pacificas with delivery slated for late 2018. Base retail on the hybrid model
is $40,000USD. Even if they got an unbelievably generous 50% discount, that means they're $1.24B USD serious, excluding the cost of the additional autonomous driving equipment. So this will either be one of the biggest debacles in Google/Alphabet's history, or they really are ready to go.
Waymo has also ordered 20k Jaguar I-paces. Rumour is they have invested about 11 billion in R&d so far, and I'm crudely estimating the cost of their 80k upcoming robotaxis to be around 12 billion. Alphabet is committed.
Waymo estimates about 50 trips per day will be served per vehicle. At that rate, 80k robotaxis is enough to displace most of the taxi/rideshare services in the southwestern United States, though I'm not expecting to see all of them on the roads until 2022 or so. Waymo has a ton of work ahead of them. Mapping and validating in the ~100 square mile area of Chandler has taken them about 18 months so far, and supposedly they are on the verge of a commercial launch. Every city they hope to deploy in has differing signage, differing driving habits, and many funky intersections and road anomalies that need specific attention.
In this scenario there's no moat for Lyft/Uber either. If Waymo can meet demand, they'll be able to undercut ride sharing prices so deeply that most consumers will switch overnight. Banks will be falling over themselves to provide the massive credit they'll need to bankroll the expansion.
I see comments like this constantly. But does anyone ever do even a basic spreadsheet to explain the unit economics of this argument?
Low skilled people who can drive are pretty cheap and plentiful. Even if self-driving tech is flawless (spoiler: it's not and won't be soon) it still only replaces part of their responsibilities. Someone will still have to clean the cars for example. Presumably there will be some required level of human monitoring, etc.
Conversely, capital and highly reliable technology isn't free. You can calculate pretty easily the rate at which it's profitable to substitute technology for labor, this is a trade-off we've literally been making for centuries.
One one side you have plentiful cheap low skilled labor. On the other you have lasers, fast computers, graphics cards, cameras, and the associated programming inspection and maintenance costs.
Why do we think the latter side is going to be cheaper in anything remotely like the near term?
That's not true at all. There are a lot more use cases for self driving technology than taxi services. In fact I'm continuously perplexed as to why this specific use case, which is obviously the most challenging and likely to be the last to reach the market, is always first in these discussions.
There are very good reasons as to why Waymo and all the other major players are focused on robotaxis.
A truck still needs to be able to drive in cities, and if you figure out how to automate all the other thing a trucker does you can replace 3 million of them. Big whoop. Robotaxis stand to challenge the economics of personal vehicle ownership, the potential market is many orders of magnitude larger.
Why can't they just have normal cars for sale that also drive themselves to start with? Like I have a car and recline and read a book on a commute most of the time. If snow gets crazy or there's very strange construction or a flooded road or a flat tire we go back to manual driving.
That's plausible and incremental. No cleaning problems, no homeless people living in them, no need for remote monitoring, a far simpler liability and insurance situation, and so on.
My assumption is we'll start talking seriously about robotaxis on the day this scenario is practical and battle tested. The idea of starting with taxis seems ludicrous in comparison.
> That's plausible and incremental. No cleaning problems, no homeless people living in them, no need for remote monitoring, a far simpler liability and insurance situation, and so on.
These are all solved problems, I mean when was the last time you saw a homeless person camped out in a Zip Car? BMW's ReachNow vehicles are all over Seattle and they are clean, insured, fueled/charged and exactly 0 people live in them.
I live in Brooklyn, where ReachNow launched with much excitement and pulled out within a year or so due to problems with maintenance, damage to cars, and difficulty with the technology.
In suburban environments, robotaxis are likely to be a key firm of public transit rather than an alternative (rideshare as public transit if already a thing in some suburban environments.)
First, I've read over and over that the single largest cost in transportation is human labor (the driver). Low-skilled and low-paid they might be, but if you can replace them with a low-paid person behind a monitor that can handle 8-10 trips at once, you've saved a substantial portion of the labor cost.
Yes, that cost will be offset by the cost of the extra hardware, but that can be amortized over tens (possibly hundreds) of thousands of miles. Likewise, the investment in software (programmers) will eat into profits, but once again, you can spread that cost out over hundreds of thousands or millions of cars. Besides, an awful lot of the latter is a sunk cost at this point.
> One one side you have plentiful cheap low skilled labor. On the other you have lasers, fast computers, graphics cards, cameras, and the associated programming inspection and maintenance costs.
Why do we think the latter side is going to be cheaper in anything remotely like the near term?
Because, other than inspection and maintenance costs, everything on your list has been getting much cheaper really quickly for a long time.
I don't disagree that some may be over-estimating how much cost can pragmatically be cut, but I also think you may be glossing over some of the costs and inefficiencies that come with operating a large heterogenous operation comprised of individuals and personal or rented vehicles.
A company deploying large mostly-homogenous autonomous fleets may be able to benefit from:
- we probably don't need to tip AVs
- economies of scale on obtaining vehicles, fuel, maintenance, and cleaning
- tuning maintenance/cleaning schedules across a fleet towards keeping vehicles in service for longer
- may be able to perform cleaning/maintenance during off-hours, vs owners who'll often have to trade
working hours to do these tasks
- autonomous systems may drive (and be tuned towards) in ways that also help preserve long-term vehicle
value and minimize costs
- lower insurance, legal, and PR costs if they outperform human drivers, don't molest/murder
passengers, etc.
- minimizing costs around acquiring and managing a human workforce
- there are probably many small ways to optimize the positioning and functioning of an AV fleet that
just won't work with a large contractor fleet
That said, the potential for a lot of these savings depends on current prices actually reflecting these components. It may very well be more expensive to perform some of these activities, no matter how efficiently, than to exploitively externalize their costs on drivers and riders.
I haven't looked for work on the price/demand relationship here, but it's also worth noting that price reductions may not just purely be a matter of cutting costs and passing those on at a static ridership and profit level. There are likely to be some price points that profoundly change user behavior
To give an example, my favorite coffee shop is about 7 minutes (+wait) from my apartment by car either way, and around 18-26 minutes (+wait) by bus (shorter there, longer back). Without a ride pass and with tip, it's probably around $7-8 to Uber this one way, vs. $1.25 for the bus.
At this price, I'll usually only Uber if heat or rain make getting to and from the bus miserable. I have a ride pass atm that knocks this down to around $6-7 with tip, which makes me marginally more likely to take Uber, but it isn't my default. I took the bus this past Sunday morning, planning to get some open source work done, but I'd forgotten my laptop wasn't in my bag. The time/sweat cost of the bus round trip and the money cost of the Uber round trip were high enough that I just sat and read a book instead.
I'm not sure exactly where, but somewhere between ~$2-5 total, I'd probably default to taking Uber both ways. Down at the low end of that range, I even would've gone back for the forgotten laptop.
> "One one side you have plentiful cheap low skilled labor. On the other you have lasers, fast computers, graphics cards, cameras, and the associated programming inspection and maintenance costs.
Why do we think the latter side is going to be cheaper in anything remotely like the near term?"
That's a weird thought to have. Humans need on the order of ~15K a year in the US. On the other hand, even an ultra high end SDC computer costs only around ~12K in production and much, much less to actually make since Waymo will no doubt use their own TPUs. Lasers are cheap, compact and efficient fiber lasers. Cameras are mass produced in immense volume and are dirt cheap for very high quality. The only other factor is software development costs which is mostly a fixed cost.
Uber currently pays these "low skilled people" ~75% of the fare which leaves a lot of headroom to lower prices and still pay for vehicles and cleaning.
Rental car companies show how few people it takes to maintain a very large fleet of vehicles and it will take even fewer if they can drive themselves (an autonomous car can go through a wash on its own, charge itself, show up at a detailing station on demand, etc).
Does that number reflect an analysis of the cost of the actual human work of driving the car only, or is that just the number for average actual remittances from Uber to the driver?
If so it includes assumption of capital risks, auto depreciation, gas, tire wear, broken glass, vandalism and cleaning, traffic tickets, maintenance, towing, and every other expense of physically delivering the ride.
Getting rid of the human driver only saves real money on driving labor. The payments to drivers have much more to them.
People are expensive. Very expensive. Covering 24h/day requires 4.7 regular week workers, which in the UK will cost (for workers over 25) a minimum of £80k. That's just raw salary and basic taxes, other things push it up (as well as the fact that you can't get people to work exact hours consistently so well and people are I'll, etc). So per "baseload" taxi, that's likely £100k/year minimum in human costs in the UK. If the tech has only a few year lifespan you've still got north of a quarter of a million to spend on the self driving things alone.
The number of hours per day the car needs cleaning will be drastically lower than the hours it needs driving.
> Every city they hope to deploy in has differing signage, differing driving habits, and many funky intersections and road anomalies that need specific attention.
The very fact that those things do need specific attention tells me they don't have true autonomy yet; they just have a way to fake it using detailed maps of known areas.
I don't see how an autonomous vehicle can handle all the corner cases of real-world driving (construction zones, etc.) until it can also work reliably using only available sensor data, so that detailed mapping is no longer necessary.
The escalation of data collection will be immense though, once they deploy the fleet. Even if they have to have more human involvement than desired currently, I think the first few years will show growth they could have have dreamed of in the beginning.
Waymo might also tell a different story: They've realized that a fundamental breakthrough in self-driving cars is unlikely to happen in the next decade and therefore there's no point in waiting any longer to perfect the technology - and instead rely on remote controlling of the cars for difficult situations. They can still gradually improve the self-driving systems but from a business perspective it might make sense to capture as much market share as possible right now.
It doesn't work that way. Waymo's remote control does not even support direct manual remote control. It just lets the remote operator give some direction to the onboard systems.
I am giving an analogy. Street car on digital track is the first part of it. The remote control aspects are there as well. It is not "direct" remote control. It is like a drone. You give a direction and the onboard computers execute on it. Direct remote control is not possible because of latency. But it also seems that the commercial launches of these vehicles might be dependent on 5G networks being available (lower latency, better reliability) in which case the remote control capabilities will be greatly enhanced before launch.
There's another startup, for self-driving trucks, which does have direct remote control. Their idea is that the trucks are self-driving for long-haul, and then the final non-Interstate part is remote controlled.[1]
In all of these deal announcements, the key words are “up to”. That’s very different from a confirmed 62k order, and in all announcements related to that agreement timing of delivery was unconfirmed [0].
[0] https://mobile.nytimes.com/2018/05/31/business/waymo-chrysle...
At the scale of Alphabet, $1.25B about 1% of of 2017 revenue and about 5% of operating income. [1] Alphabet isn't betting the future on it. Relative to the automotive industry sixty thousand vehicles in 2018 compares to about 100,000,000 vehicles worldwide per year. [2] Compared to the number of cars in my driveway and the balance of my checking account, the numbers are big. At industrial scale, they are small. If all 60k vehicles were registered in California, it would look like rounding error on the 36,000,000 vehicles registered there. [3]
Alphabet spent $12.5 billion on Motorola Mobility in 2011 and sold it to Lenovo for $3 billion three years later. The lost $9+ billion? Nobody really noticed. For Alphabet, $1 billion is mad money, not the rent. Self driving cars are a moonshot.
"According to the filing, Google and Motorola began discussions about Motorola's patent portfolio in early July, as well as the "intellectual property litigation and the potential impact of such litigation on the Android ecosystem".
Although the two companies discussed the possibility of an acquisition after the initial contact by Mr. Rubin, it was only after Motorola pushed back on the idea of patent sale that the acquisition talks picked up steam.
The turning point came during a meeting on July 6. At the meeting, Motorola CEO Sanjay Jha discussed the protection of the Android ecosystem with Google senior vice president Nikesh Arora, and during that talk Jha told Arora that "it could be problematic for Motorola Mobility to continue to exist as a stand-alone entity if it sold a large portion of its patent portfolio".
In connection with these discussions, the two companies signed a confidentiality and non-disclosure agreement that allowed Google to do due diligence on the company's patent portfolio."
You are missing a lot of details. Google purchased and then sold the cable hardware business for $2.4B to Arris. Then there is the $6.6B tax credit and the selling the factories and also they kept the AP group.
On the other hand, irrespective of how confident they are that things will pan out, corporations like Alphabet and Fiat Chrysler have no problem whatsoever with issuing press releases saying that Fiat Chrysler is willing to sell up to 62k Pacificas to Alphabet within the next few years under an agreed framework if Alphabet decides it's made enough progress to actually spend 1% of its annual revenue on more cars.
The Alphabet-scale comparisons are fair wrt to whether this is a "big" bet or not, but trying to minimize this via comparison to total world sales or California registrations less so (given that a primary promise of transit AVs is that they may replace many privately-owned vehicles).
While some have projected how many traditional vehicles they can displace, I don't think anyone really knows how that'll play out. I doubt any of these companies will go making bets that obviously rival the scale of traditional vehicle sales/ownership at the time they're made, especially before they get a sense of how reliable AV transit affects the decisions individuals make about car ownership as they face major maintenance, repairs, and replacement.
I've read one projection that says each could replace 10 traditional vehicles. If that worked out, it seems like a bet on 10m vehicles, enough to replace all 100m sold worldwide, would be astonishingly bold. If you assumed an initial scope of the entire US, the equivalent bold bet would be 1.7m vehicles. If you reduced the scope to California, that bet would be about 200k vehicles. If we guess that the difference between "astonishingly bold" and "bold" is an order of magnitude, these would be reduced to 1m, 170k, and 20k...
No. The annual production of Pacificas is only 65K vehicles or so. Google ordered those vehicles in the same way that airplanes are bought. The "62,000" is over many years, and is purely speculative, and cancellable/modifiable as needed.
Google is buying a couple thousand Pacificas in the near future. That's it. It's more or less a commitment to keep buying a couple thousand a year for the next however-many years, in exchange for a discount.
It is NOT the case that Google is going to open a big box and find 62,000 vehicles inside in late 2018.
All journalists covering the self-driving car beat are basically required to ignore Waymo so they can continue writing “When will cars drive themselves? Experts say nobody knows” articles. If you ignore Waymo it’s a race; if you include Waymo it’s not.
This article makes the further error of assuming “AI” is required for self-driving cars, rather than just a shitload of code.
>> “When will cars drive themselves? Experts say nobody knows” articles.
Nobody knows because it has yet to happen. We have all seen demos and studies but that doesn't convince anyone not involved in the tech. Cars will be capable of driving themselves when I see one, sans humans, ask to merge in front of me on the highway. It will happen when I see one navigate the lineup to the ferry I take every week. It will happen when I see one accurately obey hand signals from a human directing traffic. Not youtube vids. It will happen when I see it with my own eyes. Just like electric cars, we will have to see them on the roads playing the game we all play every day. Until that tipping point, driverless remains a myth that happens in a faraway place.
So put them on the road. Let's give it a try. Maybe it works. Maybe it doesn't and people die and they sue. That's been part of driving since day one. Roll the dice. I'm sick of the endless debates.
> So put them on the road. Let's give it a try. Maybe it works. Maybe it doesn't and people die and they sue. That's been part of driving since day one. Roll the dice. I'm sick of the endless debates.
Well I'd certainly hope that there is a high level of confidence that they don't kill people BEFORE they're put on the road with other people.
All we need is that they don't kill more people than average human drivers do per passenger-mile. Assuming ride-sharing we could even let them be a little more deadly than human drivers and still save lives.
Don't. Cost isn't stopping people today from ridesharing. Carpooling won't become cool just because there is no driver, in fact I could see it becoming less cool as you will be much more alone with strangers. Many people today would be uncomfortable on a bus at night without a third party present (ie the driver, or at least many other strangers).
Think about all the extra traffic from autodriver cars doing jobs that today are too expensive. And the double-commute possibility. Your car takes you to work, drives itself home to park/recharge, then drives back to pick you up at the end of the day. Any 2-way trip could become 4-way, doubling the per-trip traffic. Or, rather than pay for parking while you shop downtown, you ask your car to drive around in circles until you are ready to go home.
Carpooling has already become much more practical with Lyft Line, etc. No driver would mean even cheaper rides. Also, there's no reason your car couldn't park itself at work...or rent itself out (via Lyft/Uber) while you're at work. Assuming we get to that level of tech.
If cars which require no effort whatsoever to operate and cost significantly less than taxis are on the market, it's a safe bet that being slightly less dangerous per passenger mile will kill far more people than are currently killed on the roads due to more, longer car journeys. That's before we even get into the fact a car which is slightly less likely to cause a fatal accident than the average human driver on an accidents per mile basis is a lot more likely to cause a fatal accident than the average sober, careful and experienced human driver, the complications of estimating fatality rates from small test drive samples and the potential for new bugs to cause massive accident spikes.
Of course if people make more journeys, i.e. become more mobile, you could argue that the increased number of deaths should be balanced against life-hours of drivers freed up while also improving quality of life of passengers at the same time, i.e. it might still be QALY-neutral or even positive.
> than the average sober, careful and experienced human driver
But ultimately we do accept the fatality rates caused by all drivers and not just the good ones as a cost of having individual transportation. If we didn't only teetotalers would be allowed to operate cars to remove one risk factor. And we also accept the risk of inexperienced drivers because we have to teach them eventually, the same reasoning can be applied to autonomous cars (just substitute teaching with data gathering)
The total distance driven also does not just depend on technology alone but on urban planning. AVs may change that dramatically over time.
True, there are QALY arguments for even unreliable SDVs (and also for other things like abolishing speed limits or allowing people to continue to operate older vehicles currently deemed unroadworthy if they can't afford better), but they're dependent on how you adjust for quality and at best orthogonal to standard safety arguments.
> But ultimately we do accept the fatility rates caused by all drivers and not just the good ones as a cost of having individual transportation. If we didn't only teetotalers would be allowed to operate cars to remove one risk factor. And we also accept the risk of inexperienced drivers because we have to teach them eventually, the same reasoning can be applied to autonomous cars (just substitute teaching with data gathering)
We don't accept the fatality rates caused by all drivers; we confiscate the licenses of or jail bad drivers on a regular basis, and we certainly don't permit drivers being taught to operate commercially. Driver accountability for accidents is part of the social contract that permits individual transportation despite its cost; an corollary of that is that the accident rate is higher than we believe it should be, and that some degree of negligence is involved in the current level of accidents which it would seem unreasonable to exonerate corporations from if they were to replicate them using software.
And on a purely practical basis, if you estimate SDV safety based on narrowly beating the mean accident rate across all drivers in tests, and then early commercial adoption tends to be replacing experienced commercial drivers who drive the most miles with the fewest incidents first, you're certainly going to push that accident rate up in the short run even without additional miles driven.
> And on a purely practical basis, if you estimate SDV safety based on narrowly beating the mean accident rate across all drivers in tests, and then early commercial adoption tends to be replacing experienced commercial drivers who drive the most miles with the fewest incidents first, you're certainly going to push that accident rate up in the short run even without additional miles driven.
Uber could be skewing statistics in similar ways.
Anyway, this is all handwavy speculation. What I'm trying to get at is that the overton window for AV behavior may be surprisingly large, especially if they exhibit their worst behavior in the learning years when the fleets are still small. For some metrics possibly even larger than that for humans.
No, that's a wrong benchmark. We shouldn't compare driverless cars with the average human driver. We should compare them with the average human driver assisted by the best technology (which is say waymo's driverless software running passively to prevent accidents).
You’re incorrectly assuming that a solution to the problem of self-driving a car is a superset of the problem of passive accident avoidance.
If the software is in active control pretty much everything that goes into that is different from software that is deciding if and when to emergency brake.
To put it another way, a self-driving system will become safer than human drivers, not because it knows how/when to emergency brake really well, but because it won’t have to emergency brake as frequently as a human does.
Sure, I understand that a driverless system might prevent accidents by strictly maintaining safe distances. But the benchmark still can't be unaided human, rather a human aided by the best assistive technology that can be built (say by the same company).
That's an irrelevant benchmark except for people who pay premium to have the safest ride in the world.
Everyone else performs a tradeoff along multiple axes including price, convenience and safety. Not needing a driver, freeing up time and lowering prices increase utility which could even justify a slight increase in risk.
But even if we don't want to increase risk then a self-driving system still only has to be as good as humans on average to result in no net-increase of deaths per passenger mile while still increasing overall utility.
And more simply, it is silly to expect the maximum possible value to be the minimum requirement.
Let's agree to disagree. But if we allow driverless cars on the road (with no backup), it should come after emergency braking, lane keeping assist, side collision assist etc have become mandatory. Just like airbags are mandatory now, and we enforce fuel efficiency on new cars but not existing cars.
Just to add. In a world where airbags are not mandatory,(autopilot+airbag) driverless technology would be safer than the average human (without airbags) on the road. So, you would be arguing for legalizing driverless cars completely overlooking that the gains are not because of driverless-ness but because of airbags. This holds true now too, almost all gains by driverless-ness can be had by emergency braking, lane keeping assists, side collision warning, and pedestrian detection etc.
And people will die anyways. It would be the biggest miracle on earth I think if those things ended up being perfect (at least, the ones that work.. not Uber, obviously).
So it's all a numbers game. What's acceptable, and what we can improve on.
I had a famous cancer researcher at a world famous cancer treatment hospital tell me (paraphrasing) “we won’t ever see the kinds of breakthroughs we saw in the past because we were reckless back then, foolish even. It was the Wild West.”
I’m not saying I want the highways to be a Wild West for self driving cars, but it’s worth admitting it would probably accelerate the pace of innovation.
Living around the Mountain View area, this is one of those moments where I'm like "Oh, the future is here, it's just not evenly distributed."
I've been driving next to these cars for the past 7 years. A number of my friends & former coworkers have ridden in them; a couple of them work on them. They've stopped for me at pedestrian crosswalks. I've gotten stuck behind them, driving 25mph on El Camino. I've waved them on to make a left in front of me (I'm still not entirely sure how they do that, but I think that if they detect that the driver at the intersection is stopped and has not gone within any reasonable time period, they creep out and eventually go if no movement is detected). I saw one last week attempt to switch lanes, find that the car behind it had sped up and it was no longer safe to make the lane change, turn off its blinker, pull ahead in its own lane, and then make the lane change safely several cars up.
They are on the road, and they work pretty well. I trust them a lot more than cars that say "Student Driver Onboard", or even the average Californian driver.
Nobody has self-driving cars, not even Waymo. Every "self-driving car" has "disengagements" (a euphemism for "it stopped working") periodically every few thousand miles. That's simply unacceptable and needs to be sorted out. If a human driver was performing a panic stop for no reason or hitting a pedestrian or another car every 3 months or so we would take away their license.
Waymo is doing better than almost everyone else, a testament to their skill and their rigor, but they are still a long distance away from truly autonomous vehicles. The temptation is to say "well, they are so close, I'm sure they'll finish up the niggling details", but the reality is that the stuff they are failing at are some of the hardest remaining problems and will probably require at least as much work as they've put in already, if not significantly more, to tackle. I think we'll end up with self-driving cars eventually but probably not on a time scale of within the next 5-10 years, more likely it'll take that long or longer for the technology to reach maturity and then yet another decade plus for it to start seeing practical application.
It's the classic problem of truly appreciating magnitude and scale. Humans are amazing at it within a very narrow realm, and then terrible beyond that narrow range.
This is the sad realization I had about brain-computer interfaces towards the end of my research. 99.9% is NOT good enough when the stakes are life and death and the factors within events that the environment throws at you are basically independent and continuous.
At least with self driving cars we can enforce some hard boundaries. In BCIs implementing boundaries is tantamount to denying the user of their free will.
I repeat this again and again, and am always surprised that nobody else likes to raise the unfair comparisons.
I am willing to assume that waymo/Tesla etc have found the holy Grail, and made a driverless car safer than the average current human driver on the road. I'm not convinced that we should still allow these driverless cars on the road with no human backups. Not because of expectations or liability, but because the benchmark is wrong. We should compare them to a human assisted by the same driverless technology being used passively to prevent accidents. And it's not clear if a driverless car with no human backup is safer than a human assisted by the same technology.
Is part of Waymo's perceived success in that they are more conservative with deploying their technology? I.e. it is tested at lower speed, on specific routes, in specific weather etc.?
I don't really see Waymo as being far ahead at this point. Uber pretty clearly stole Waymo's work. So unless Waymo had a major breakthrough since then that wasn't related to the basically everything Uber got then everyone is pretty much on the same page: promising it's almost here but in reality still about 10 years out or more.
Uber stole a small portion of Waymos work several years ago. They supposedly stole some lidar designs. There’s also code and personnel that Uber doesn’t have. In my mind, Waymo is miles ahead of Uber.
I'm sure Waymo will make a lot of progress with their massive investment but using SV money as an indicator for a project's ultimate success seems unwise at this point.
I think the real question is whether the big investors in self-driving cars will make progress creating virtual rails for their vehicles, e.g. roads where we can keep pedestrians out.
While I agree they are closest of all the competitors with published numbers they are still very far off from human drivers and just looking at the linked articles within the original article do make it clear they have a long way to go. the key difference is the level of caution Waymo is exercising.
So how do we get a determination of which disengagement events would have possibly lead to a crash? Plus I am still not convinced these numbers are even going to remotely similar over roads not heavily mapped and imaged. I think the hope many have for AV is that is allows for driving in the worst of conditions but in the end we will have AV for limited access or very tightly monitored road ways only. Which in itself is not a bad idea, rush hours could much safer and efficient with HOV/Express lanes redone to explicitly support AV.
We certainly aren't going to be able to get people to drive more lawfully when most think they are the better driver when compared to others.
> Can they really use that many only in good-weather places with perfect roads?
I would hazard a guess that yes, that's pretty much the Southwestern US year round, if you include Southern CA. There are enough metro areas, and enough new sprawling development, to put those vehicles to work.
>Can they really use that many only in good-weather places with perfect roads?
Waymo supposedly processes their lidar signals to detect other cars with reasonable accuracy even in snow/rain(https://youtu.be/UrJ4-AUL4U0?t=9m41s). They only operate in pre-mapped areas so they should be able to follow lanes just fine even if the actual markings are obscured (by weather or by poor maintenance).
This kind of approach (mapping everything beforehand and using lidar for perception) is used by other players as well, and seems the most likely to work for short term level 4. Trying to do everything without the benefits of lidar and high-res maps (what e.g. Tesla are doing) is a more general solution, and cheaper hardware-wise, but will take longer to get right.
Thanks for posting that Waymo vid, that's the best processing of Lidar in adverse conditions that I've seen. I think the million dollar question is how far Waymo (or others) can push noise reduction with Lidar. The video seems impressive in isolation, but if you compare it to what Nvidia was doing with cameras two years ago (object recognition in snow/fog that appears to outperform humans[1]), it quickly becomes apparent that Lidar is still nowhere close. But maybe they can get it good enough to bridge us between specific and general techniques.
Personally, I'm long on the camera-based approach because I think reliable autonomy will require a generalized vision solution, and once you've achieved that, the Lidar is redundant.
Read the contract if you can. I guarantee that they have not put down a billion dollars for all those cars. This isn't like ordering from Amazon. This is a complicated delivery contract. It will have stages. Money will be owed according to a schedule. If things don't work out, there will be many places where delivery can be delayed or the order canceled. Waymo has to have a plan to pay for this in the long run, but they aren't today putting down all the cash.
If you are buying 10,000 cars, what you are essentially doing is taking on that car manufacturer as a partner. You are betting that you will need the cars, and they are bettering that you will be around to pay for them. Everything remains speculative.
This has been hanging out there for a while it's just that it never got much mainstream press. I wonder what it is specifically about the concept of "AI" that makes it so prone to the magical thinking that drives these boom-bust cycles.
"I tell adult audiences not to expect it in their lifetimes. And I say the same thing to students"
"Merely dealing with lighting conditions, weather conditions, and traffic conditions is immensely complicated. The software requirements are extremely daunting. Nobody even has the ability to verify and validate the software. I estimate that the challenge of fully automated cars is 10 orders of magnitude more complicated than [fully automated] commercial aviation."
- June 2015, Steve Shladover, transportation researcher at the University of California, Berkeley
"While I enthusiastically support the research, development, and testing of self-driving cars, as human limitations and the propensity for distraction are real threats on the road, I am decidedly less optimistic about what I perceive to be a rush to field systems that are absolutely not ready for widespread deployment, and certainly not ready for humans to be completely taken out of the driver’s seat."
- March 2016, Mary Cummings, director of the Humans and Autonomy Laboratory at Duke
"With autonomous cars, you see these videos from Google and Uber showing a car driving around, but people have not taken it past 80 percent. It's one of those problems where it's easy to get to the first 80 percent, but it's incredibly difficult to solve the last 20 percent. If you have a good GPS, nicely marked roads like in California, and nice weather without snow or rain, it's actually not that hard. But guess what? To solve the real problem, for you or me to buy a car that can drive autonomously from point A to point B—it's not even close. There are fundamental problems that need to be solved."
- September 2016, Herman Herman, director of the Carnegie-Mellon University Robotics Institute
I'd argue that it's because (true/hard) AI represents, in many ways, the "final frontier" of scientific research. Not in a literal sense -- they'll always be something else to study. But the notion of creating a thinking machine is, in many ways, the ultimate goal of studying nature and the world around us: to allow us to understand the processes and phenomenon at work well-enough to be able to capture/manipulate/recreate them at will.
This isn't new though -- "thinking machines" have captivated audiences for millenia. See the Antikythera mechanism [1], the Turk [2], and so on. The ability to replicate the capabilities of our own brains is intoxicating -- it finally signals that we, homo sapiens, understand our own minds well-enough to recreate their abilities at will.
>the "final frontier" of scientific research. Not in a literal sense
Real AI would literally be the final frontier, because at that point, all the scientists could quit their jobs. The AI could replace them.
But real AI is a lot different than what everyone is selling as AI these days, which is essentially a hyper-glorified generalization of linear regression.
The technology actually seems perfect for some kind of hybrid between road and rail, if we could build a city from scratch you could make something better than either, with train-like capacity alongside road-like routing.
Although I suppose the difficulty limits the market enough to make investment in such a scheme uncompetitive. Perhaps this is what Musk is hoping to get around that with his tunneling idea.
I'd vote for the other way around. Combine the flexibility of the car with the freedom of riding a train.
Let people drive manually to predetermined locations, let's name them a "station". Where the "AI" takes over and lines up cars on predetermined routes, let's name them "tracks". The AI will control the car on the "track" up until the nearest "station" to the destination. From there the driver will manually continue to her destination.
I'm glad to read these because I've been feeling like and been treated like I'm on crazy pills for thinking we aren't even remotely close to autonomous vehicles.
I think it is because AI, like intelligence, is undefined in a concrete sense. When someone creates a new algorithm that solves a problem of identification or logic we call it AI. So when people say a thing will be better with better AI you could say it would be better with more effective problem solving. What thing in life couldn’t be made better with more effective problem solving..?
Right now we have fully autonomous vehicles operating within a roughly 100 square mile geo-fenced area of Chandler Arizona. They are overseen by remote operates who intervene when the vehicles get hung up. The vehicles stop and start at predetermined pick up points. But within those constraints, they work.
Waymo has been developing the technology for 9 years, and have accumulated nearly 8 million real world test miles in that time.For the last 18 months (or so) they have focused their testing in the Chandler area, but are not yet ready for a commercial launch.
For every new area they head into they need to solve specific problems at specific intersections. The vehicles are not yet validated for a full range of weather conditions. There is a huge amount of preliminary work that goes into preparing an area for commercial operations.
Developing an autonomous vehicle that can go anywhere as well as a human might just be 10 orders of magnitude than for commercial aviation, but what Waymo has is enough to disrupt.
'It may be a hundred years before a computer beats humans at Go — maybe even longer,'' Piet Hut, an astrophysicist at the Institute for Advanced Study in New Jersey, told the NYT in 1997. ''If a reasonably intelligent person learned to play Go, in a few months he could beat all existing computer programs. You don't have to be a Kasparov.''
If you care to, you can dig up plenty of examples of people underestimating, overestimating, or correctly estimating the future capabilities of AI. Just because some guy in New Jersey underestimated it once doesn't mean we should accept AI hype uncritically, especially since much of that hype is driven by moneyed interests.
People are just bad at estimating in general. They tend to assume things will go in a straight line. You can see this most often when the curve is exponential, but asymptotic curves will throw people off too.
"With a further small percentage increase in cost we could reach the level of the baby Newton and better. We could then educate it and teach it its own construction and ask it to design a far more economical and larger machine. At this stage there would unquestionably be an explosive development in science, and it would be possible to let the machines tackle all the most difficult problems of science… For what it is worth, my guess of when all this will come to pass is 1978, and the cost of $10(8.7 ± 1.0)."
- Irving Good
"Within a generation, I am convinced, few compartments of intellect will remain outside the machine’s realm - the problem of creating ‘artificial intelligence’ will be substantially solved"
The crashes that the article claims came from edge-cases that the engineers didn't account for weren't, in fact, edge-cases. The Tesla crashes are due to the way Tesla's system works, it doesn't recognize stationary hazards. Similarly Uber's system failed to brake for a hazard that it could clearly see. That's not an edge-case, it doesn't have to know that the object is a woman pushing a bicycle. It's a solid object moving steadily into the car's trajectory.
but the problem is that you need to properly resolve base cases before starting on edge cases. The fact that base case scenarios are not properly solved hints at major problems in execution.
I think you have that flipped. 20% of code covers 80%. As you try to get to 100% coverage of all possible scenarios the complexity gets in some cases exponentially harder as all the low hanging fruit is picked
I like to think of it as an order of magnitude increase for each 9. Ex. If getting to 90% coverage takes 10 hours, getting to 99% will take 100, getting to 99.9% will take 1000, etc.
Citing the Uber case is particularly egregious. Have we all forgotten the NTSB report? It was only a few weeks ago. The Uber car tried to brake, but Uber had deliberately disabled the brakes for a more comfortable ride.
Which basically makes the argument, to provide "a more comfortable ride" they had to disable a critical safety system that didn't do its job good enough to be used all the time. There were so many false positives that it made the car a road hazard due to random hard breaking.
Rodney Brooks gave an interview that was recently reposted on ArsTechnica. He has so much real world experience productizing AI and robotics that no one else really compares. He is not pessimistic just realistic about self driving cars. And he has great reasons why true autonomy is going to take a long time.
Rodney Brooks has a history of skepticism about AI (said in 2009 that he doesn't think AI is even possible), as does Gary Marcus (who's been anti-deep learning since 2011).
"AI is hype" hit pieces always quote them.
I assume you are trolling, but in case someone misses your user name: To say that the director of the MIT Artificial Intelligence lab for ten years and the founder of the a successful AI workstation company and the most successful robotics company in the world "has a history of skepticism about AI" is idiotic.
True. But if you listen to his interview it is clear that he is neither pessimistic nor optimistic. He recognizes that unanticipated breakthroughs are likely. His point is that there are too many unsolved problems for a quick development of useful driverless cars. He goes through the issues in some detail, but to offer one: real world drivers need to break the law all the time in order to not get stuck in traffic. How do you decide when it is OK to break the law? Who will accept the liability associated with a car that sometimes breaks the law. Etc.
not trolling. I don't think you know Brooks well enough, i've been following him since the 90s. Rodney Brooks is famous for his subsumption architecture and disregard of statistical methods. Because of the failures of Marvin Minsky GOFAI, the previous director of MIT AI, he went in the opposite direction completely and became highly skeptical of AI while serving as MIT AI head. This was the AI winter days of the 90s.
Why create a name "totaldick" to make a single post if not to troll? My response would be: You don't head the most prestigious AI lab in world with pure skepticism.
I think you have elaborated on your true objection, his past skepticism about statistical methods. I don't think that equates to a blanket skepticism. I also refer you to the talk in my original post. He points out that there are breakthroughs and pitfalls that are unanticipated. He outlines specific pitfalls to autonomous driving that he feels no one has adequately addressed.
This is my only account, just signed up recently, and the name was just spur of the moment.
Sorry if it comes across as trolling, that wasn't my intention.
I just know a bit about Rodney Brooks, am a fan of his work, so I gave my 2 cents.
I believe that you're not trolling but could you please email us (hn@ycombinator.com) with a clearly-non-trolling username that we can replace "totaldick" with? A name like that amounts to subtly trolling other users with every post you make.
As long as software is composed of primitive language derived statements and "functional" logic (IF, WHEN, WHILE, DO, GOTO, etc...) AI will remain a fantasy.
This is hard to parse. All software that can be represented as if and goto (and some sort of assignment i suppose)?
when, while and do are just if and goto.
alpha go is all just if and goto. I'm not sure what you're looking for or alluding to.
Is there some other operation that isn't just if and goto? because i think any function you come up with, i can just make a big table of inputs that result in specific outputs.
Heck, decrement and jump if zero is enough. I'm not convinced at all that specific operations are limiting us somehow. Could you elaborate?
I believe what he's saying is that we don't "think" in if/then statements. So true AI will never be intelligent-like if you judge intelligence as thinking like us.
Ummm.... no. You might be able to make the case that a neuron can be simulated with a very complex set of if-then statements, but that assumes that we can know everything about the state of a dozen inputs, which we can't. At any rate, the brain is not a computer:
https://aeon.co/essays/your-brain-does-not-process-informati...
i like that this essay puts the complexity of the brain into perspective, but the distinction between a computer and an organism is kind of arbitrary. brains are physical systems that can be modeled mathematically. if you can create a mathematical model of a thing, you can execute that model on a computer. a brain isn't a classical computer, but that doesn't mean that a computer can't simulate a brain.
we currently lack the ability to recreate a brain-like entity, but the subtext that i am reading here is that the complexity of the brain is such that accurately modelling a brain in mathematical terms is impossible. the "brain-as-computer" model may not be accurate, but everything that exists can be expressed in mathematical (and therefore compute-able) terms.
i doubt that cyberbrains will run on anything that we recognize as a general-purpose cpu. gpu micro-architecture is already a significantly more efficient option for performing nn computations. as our grasp on this stuff improves, more specific silicon is being developed to make it even more efficient.
I don't think it's arbitrary. You even distinguished in this comment between a brain and a simulation of a brain. Steel is a physical system that can be modeled mathematically. Accurately simulating steel doesn't make steel.
Agreed. I think radioactive decay is an even better example of a well modeled physical system that defies simulation. A simulation of an ounce of decaying uranium won't tell you which atoms will decay in which order in a physical chunk. Ergo, somethings defy simulation.
This is not very complex and accurately models neurotransmission. What's missing here is the vastness of connections from a particular neuron towards thousands or more other neurons, but the inherent function is definitely not complex.
* weight being either 1 if you adopt a continous modelization (multiple serial input provide multiple serial output) or a float if you prefer the discrete modelization (sum of input to sum of output)
All software can compile down into mathematical expressions, logical branches and conditionals. That doesn't mean AI is limited to "Hello World" and FizzBuzz.
I work on self-driving cars. Imho, companies need to collaborate on certain aspects like perception instead of competing. The vehicle is basically blind beyond the range of the sensors or if you miss a detection. Collaboration on these issues will lead to faster progress instead of competing.
This is an example where companies could put people over profits and share what they’ve learned to keep things safe. I don’t see a lot of research papers (I see none) coming out of Waymo, Tesla, Uber, or Cruise.
This is another form of collaboration. You could publish your innovations in some technical forum w/o code and it'll automatically drive the industry forward since SOTA results will be available.
Right now, no one really knows the performance of the Waymo system. Millions of miles doesn't mean a lot. I can do billions of miles in a parking lot with no cars or pedestrians. Even millions of miles on road doesn't say anything about the performance of the components.
Is there a business model for providing detection data to other fleets in real time? This is a stretch of the word "collaboration" but it could possibly prevent deaths while leaving the IP in the hands of the company.
Why would I share technological innovations with another company?
Because nobody has a real market until some of these problems are solved significantly better than current state-of-the-art?
Having some baseline collaboration on common components and safety systems could plausibly move the entire industry towards viability much faster, both on a safety and tech trajectory and for regulatory oversight. It's not a crazy idea.
Isn't it ethically preferrable for Waymo to push their own, better, safer system, as far as they can, rather than to improve competitor systems in a lesser way by sharing their secret sauce that competitors can only partially implement?
Because they pay you for a license, might be one reason. Maybe some system (possibly backed by legislation) makes it easy for them to share technology without some companies being freeloaders and others being suckers.
Competition obviously is a powerful motivator, but sharing tech can reduce costs for everyone, at least in theory.
I would support laws that said, in effect, you can't compete on safety. Meaning, if one self-driving car[1] company comes up with the software equivalent of the seatbelt or the airbag I think they should not get exclusive right to implement it. They should have to share that with the other companies. I don't see how doing otherwise wouldn't put safety above profit.
So how do you compete? I don't know. Along other dimensions, I guess.
There is another aspect of this that I want to bring up: auto-autos should share data with each other and with the surrounding traffic infrastructure in real-time, for safety and for dynamic traffic-shaping. They should be able to cooperatively track pedestrians and non-automated vehicles. The cars and the roads and the signals and something like Waze should all be integrated and cooperating for maximum safety and efficient throughput with low latency. And they should all share experience (training data) of normal and exceptional events across the whole fleet (regardless of manufacturer.) Optimizing across the whole thing will be, uh, optimal. From this POV, non-cooperative behavior (due to the profit motive or just people being people) by any single actor will be seen as a bad-faith move and the network can be expected to route around it one way or another.
Fair enough. I'm sure some would be willing to do that, but likely not many. I do think it is going to be very interesting to work out some of the scenarios if we do get to Johnny Cab style self driving cars.
By what mechanism though? I could see a big griefing opportunity if there is some automated system that allows law enforcement to signal the vehicle to pull over.
Check out waymo's videos; the car detects the flashing lights + reflective fabric on the police officer's uniform, as well as the shoulder and arm gestures. Griefing is as much an issue as is police impersonation right now, that is a very rare occurence that will be punished very severely when occurring.
Have a look[1]. Observe how they all kind of look like each other regardless of the country: shoulder pads, head cover, belt, uniform colour, formal attire, etc. Once the model knows the baseline, that is "a uniformed individual waving characteristically in front of the car", don't you think it would be quite easy to teach it with a mere few hundred pictures how local police officers look like in the current region?
They don't look remotely similar to me, especially considering that black, blue, navy, grey and khaki (and reflective yellow for people on roads) are not unpopular colours for clothes worn by people who are not police officers or otherwise involved in traffic direction, many of whom may have cause to move their arms in ways comparable to signals. Humans are just a lot better at gauging intention and even simple stuff like parsing the word "police" at an oblique angle
Frankly even with local training, you'd still think giving law enforcement devices to stop, restart and redirect vehicles was a minimum requirement.
All of these agents follow a similar pattern of clothing (as I said, combination of similar garnments and colours) and behaviour (placement on the road, gesturing with authority, directly facing the car, etc.). Machine learning algorithms are especially good at recognizing patterns and storing their abstracted form, so it should come as no surprise that understanding what a police officer looks like in abstract is not the main issue of self driving.
> Humans are just a lot better at gauging intention and even simple stuff like parsing the word "police" at an oblique angle
Google seems to understand these intentions well enough, and at a much higher level than mere word parsing. This video is from 2015: https://youtu.be/tiwVMrTLUWg?t=9m5s
You got the car understanding all that happens at a complex intersection at 9'05, understanding what a police car looks like at 9'35, then detecting and reacting to a schoolbus and then parsing a police officer gestures right at the 10' mark. I'd say chances are these are pretty solved situations 3 years later. You can even see some creatures from their "zoo" of patterns for cars & people at 10'35.
> "When a Waymo car hears sirens, it will automatically pull over, yield, and stop. For example, when a number of vehicles are moving towards the scene of an accident on a highway and ambulances and other emergency vehicles are headed toward it, driverless cars will move aside and give way. Using audio sensors, the cars can detect exactly which direction the sirens are coming from and move out of the way."
> All of these agents follow a similar pattern of clothing (as I said, combination of similar garnments and colours) and behaviour (placement on the road, gesturing with authority, directly facing the car, etc.)
Repeating an assertion does not make it cease to be false. A very small handful of pictures you linked to shows a wide variety of coats, vests and shirts of many different colours, all of which heavily overlap with general garment types and colours used in everyday clothing which tend to indicate police only with small and greatly varying trim detailing (and sometimes hats). And is the clothing and trim designed to convey authority? isn't the sort of abstract pattern recognition computers do better than humans, or even at all well. Sure, you could certainly create specific police uniform training sets for every jurisdiction and possibly even cut down false positives in other jurisdictions by geofencing them (so you don't get people stopping in California for commuters wearing the distinctive er... blue shirts and black trousers of the Hong Kong police) but it's a non-trivial undertaking even if there are bigger problems for SDVs to tackle
More importantly, unlike humans evolved to have an intimate understanding of human mannerisms, machine learning has no concept of "gesturing with authority", beyond whether moving human shapes fit very specific patterns within its calibration parameters, and police officers often don't have scope to place themselves in a particular position in order to get the car to understand them.
> Google seems to understand these intentions well enough, and at a much higher level than mere word parsing.
The video shows examples of predicting possible directions of travel of moving road users based on maps and movements (i.e. its fundamental driving model) and a shot of it recognising two arm gestures in an idealised front on positions. Neither fall under the scope of being able to understand how the traffic policemen intends to clear the blocked intersection from his shouts and gesticulations at you and various other vehicles. Humans also don't need to be signalled to go again if the black-jacketed man they've stopped for was actually trying to hail the taxi behind them.
Instead of recognizing the lights or something you could create a communication of some sort where the officer's vehicle provides a key that the car can then authenticate that it's a real police officer making the pull over request.
Heck you could even check with a police department system that the officer is on duty and in that area so a stolen police car or key couldn't be used.
"By working with the police department in Chandler, AZ, for example, Waymo has been able to train its cars to yield, pull over, or stop when it hears or sees sirens." [1]
The only reason this even looks like a problem is that humans don't follow the law.
In most jurisdictions if there's an emergency vehicle running flashers and/or siren behind you on the same side of the road you're supposed to slow, pull over, stop, then proceed only when the vehicle passes you.
If the emergency vehicle parks behind you then you have effectively been pulled over.
Pulling over safely is practically the first thing a self-driving car team works on. It is very easy to detect flashers and sirens. So... no problem.
Well, except that self-driving vehicles may accidentally be pulled over by firetrucks.
So if the autonomous vehicle hears the sirens, does it start to pull over, or wait until it can also detect flashers? What about a loud stereo playing a siren sound or a car with flashers installed on it? I'm just wondering, because although I know this sounds like a pretty simple problem, I don't think that it is.
When you hear a siren you're to become alert for the possibility of flashers behind you. You are not supposed to just start pulling over because it may not be on the same road as you and pulling over may block traffic. Imagine if every car in a crowded downtown pulled over every time the drivers could hear a siren. Instant gridlock.
So a siren played loudly on the radio will not cause a rule-following car to pull over, regardless of who or what is driving it. I presume you're talking about siren sounds in music, but...
Installing flashers that produce the same pattern as official police flashers and/or playing a continual siren is impersonating an officer. That's a serious crime.
Fake flashers will probably cause self-driving cars to pull over but they'll also cause humans to pull over. Not a self-driving car problem.
In the future police flashers can have a cryptographic element emdedded in the flash pattern that a self-driving car or a human-driven car with sensors can authenticate but that's an enhancement over the current situation.
It's one of the best-defined situations in traffic because it's precisely the fallback behavior. Detect emergency vehicle behind you, pull over as far as is safe, wait. If a self-driving car can't do that then it's hardly self-driving at all.
The real trouble, which you haven't brought up, is when the traffic is already at a standstill and an emergency vehicle approaches. The soft and gradual scooch-scooch movements that we have to do to clear enough space for the emergency vehicle are a much tougher case.
You can pull over a self-driving car. You just may have difficulty getting past it to an emergency.
The story of computation in general — “The Rise of the Machines” — has been more about humans becoming more like machines than the other way around. The question is not “will computers ever be able to drive as well as a human?”. It’s “Will humans remake their driving systems to fit a sufficiently intelligent computer-driven model?” I believe the answer, as the US ages and falls below replacement rate, is yes. We are already fully invested in the age of benevolent machines, and there is no turning back.
Cadillac super cruise(GM) is only capable of handling itself on a limited set of highways. Tesla's nearly anywhere. Mercedes is considered subpar compared to GM's system in nearly all independent reviews.
I read this as Tesla is also an imitator. There system is not autonomous driving, it is just a poor imitation of autonomous driving. Tesla, GM, etc. are all just advanced cruise control.
All due respect, have you used any of the Autopilot clones? There are endless debates about what's better on paper, but everybody I know who's actually used them says Autopilot is much more mature. Comma.ai seems to be right on their tail, but it's an aftermarket system requiring a fair bit of DIY (so it's not very accessible yet for most end users). Tesla was first to market with auto steering, and seems to have the most mature offering, so framing them as the market leader seems pretty accurate to me.
But the other companies aren't imitators even if what you're saying holds true. They have products that did, before or at the same time, similar functions. Did they actually copy Tesla once the functionality was released on the Model S? The answer is 'no'.
> They have products that did, before or at the same time, similar functions.
I don't think that's accurate. Autosteer is the fundamental innovation of Autopilot, and no other manufacturer released a system that "did, before or at the same time, similar functions" as Autosteer. Other manufacturers have since released systems that technically check the box of turning the wheel. But the value of a system isn't how it's described on paper; it's how useful it is on real roads. I haven't seen anyone release a system that's competitive with Tesla's except for Comma.ai (which is a very special case as they are not an OEM and only sell raw hardware).
"Imitators" wouldn't be my word of choice (it feels needlessly pejorative), but until someone actually innovates beyond what Tesla has done, I think it's fair to frame the conversation as the gold standard vs everything else.
As someone who has used Autopilot and many of its "clones" as well as worked with several OEMs on these features, Tesla's offering is not more mature, it's more irresponsible. Autopilot is a combination of ADAS features that many other OEMs have had for a while and deliberately chosen not to tie together due to safety concerns. These concerns have been validated multiple times now by Autopilot errors. The "clones" are deliberately less good because the manufacturers offering them care more about safety and reliability than about marketing beta products.
How mature a system is and whether it's morally responsible (or imperative) to release it are different questions.
> The "clones" are deliberately less good because the manufacturers offering them care more about safety and reliability than about marketing beta products.
When I say "maturity," I'm not talking about holding back features. I'm saying that Tesla's system works better in situations where other manufacturers offer similar features. For example, how many other manufacturer's TACC will slow down (below the set limit) when a turn is coming up and then accelerate through the turn as a human is taught to do? How many autosteering systems will sense an adjacent car drifting from its lane and "creep" away from it to avoid a possible collision? These are the sorts of mature behaviors that put Autopilot ahead of the pack and make it much more real-world useful than its competitors.
On irresponsibility, I passionately disagree. But rather than rehash a bunch of old discussions I'll just link to prior comments explaining why I feel the way I do, and we can debate it further in those comments if you're interested.
Tesla seems to have focused on getting lane following to behave well. That gives the user the impression that the system is competent. Obstacle detection, though, is not reliable, as has become painfully clear from crash reports.
Wasn't the recent reproduction of the merge lane issue [1] showing that lane following is also not competent? I'm personally not sure any of the individual features of autopilot are really all that reliable.
Ok, we have different definitions of maturity. To me, it means your system's core functionality works reliably in the use cases it is designed for. Nice-to-haves like speed smoothing are secondary.
Thanks for the links to your other comments. I don't think there's fruitful debate to be had here - we're just of different opinions. I think autopilot encourages bad driving behavior in drivers who would otherwise be more alert, and if I've read your comments correctly you think the responsibility is nevertheless still on the drivers and autopilot probably helps more people than it harms. I'd love to revisit this if we ever get apples to apples accident data showing comparisons between autopilot-enabled vehicles and comparable vehicles from other manufacturers in the same price and age range (thus with corresponding modern and expensive ADAS systems).
> Ok, we have different definitions of maturity. To me, it means your system's core functionality works reliably in the use cases it is designed for. Nice-to-haves like speed smoothing are secondary.
I agree with this definition, but I see it more as a sliding scale than a checkbox. I think it comes down to whether you look at each system in isolation or whether you put them all on a single spectrum. I'm imagining human driving at one end of a spectrum and fully automated driving at the other, and looking at what percent of driving conditions each system provides a safety or utility improvement in (so you can see why features like speed smoothing and reacting to nearby cars quickly become relevant to my thinking). But I can see why, judging each system on its own, one would say that Autosteer is less mature at what it aims to do than simpler and more constrained systems are at what they aim to do.
You summarized my take on Autopilot as accurately as one can in a single sentence (thank you for that btw). Like you, I look forward to the day when data is available to resolve competing hypotheses like these. Elon pledged on the last earnings call to release quarterly Autopilot data reports; I'm hoping those will include Autopilot vs non-Autopilot driving usage and their respective highway accident rates, broken down by feature (TACC, Autosteer, TACC + Autosteer, etc). I think that'd be the best way to assess Autopilot safety, as it would control for all other factors. It'd be great to have apples-to-apples numbers from all manufacturers, but I think that will take longer and probably some coordination from NHTSA.
Braking before a turn ( or trailing-off in the turn ) is only necessary when the approach speed is higher than the entry speed.
For a computer-driven car with accurate mapping, and hence foresight, it is more efficient to plan the approach so that no braking is necessary. After all, reducing cost requires efficiency rather than speed. Likewise on a long cycle ride I'll try to match the approach and entry speeds by easing-off rather than throwing-away calories by braking.
How much better than human drivers do self-driving cars have to be to succeed? If you take a metric like fatalities or damage per mile I'm guessing we ask our robots to be much better than humans, even though it doesn't really make sense. Over time the differential will come down but a better intermediate strategy would be to keep humans away from robot cars until collision avoidance techniques improve substantially.
But the thing is most safety features can be applied to a regular car and we'd see a significant decrease in accidents. Adaptive cruise control and pre-collision breaking haves existed as premium packages. The biggest added safety feature I can think of self driving cars is sleeping at the wheel, CDC estimates up to 6000 fatal crashes a year but I wonder how much modern safety features would reduce it if it was required on all cars.
FWIW, I've heard the "100x safer" number tossed out at conferences on ADAS. I.e. an autonomous vehicle needs to crash 100x less frequently, kill 100x less people, etc. for the public to truly accept it as better than human drivers.
Logically, this doesn't make any sense, but taking control away from people leads to some psychological hurdles.
I had this argument the other day with someone else.
You can't make a robot that goes out in public and kills random people. You just can't do that. Okay?
Q: "What if the robot looks just like a car and there's a person inside the robot watching TV? Can it kill random people then?"
Still no.
Q: "But people do that all the time! Tens of thousands of people die or are maimed by traffic collisions every year! My robot can't join the fray?"
Still no. It's insane and horrible that we set up a death and mayhem lottery and that we force everyone-- children, old people, pregnant ladies, folks in wheelchairs, ev-er-ee-one -- to play it whether they want to or not. That's a bad thing we shouldn't do. But it still doesn't make it okay for you to make a robot that goes out in public and kills random people.
If you make a robot, send it into the world, and it kills someone, you are a murderer and you should go to jail, even if your robot looks like a car and there's a person riding inside it.
- - - - -
We could make a self-driving golf cart that never went fast enough to be able to injure people and it could take my mom (who has dementia and can't drive or ride the bus on her own anymore) to the doctor and back safely. We could build that today, with existing technology, there's a market for it, and it wouldn't kill anyone.
Self-driving cars are a fetish that distracts from solving real problems! (One day that won't be true, let's not kill too many more people until then, hey?)
- - - - -
Humans drive really well. Like really really crazy good well. When I first realized how well people drive it made me consider that Guardian Angels might be a real thing. But then I learned more about how the motor cortex worked and some of the uncanniness faded.
We should be so lucky as to have robots that drive as well as we do.
In the meantime, the industry should concentrate on incrementally automating traffic and stop trying to bite off more than current technology can chew.
What we're seeing now is hubris driven by ego and greed. And it's killing people.
The driverless technology has to be better than an average human assisted by the same technology. In over words, the human has to add no value in safety, and then can be allowed to not drive.
Notice it is not the average human on the road right now, but the average human assisted by the same magical driverless technology that is developed.
As someone who worked at a ridesharing co, I'm also really curious about the ops side of things, like how will passengers behave when there's no driver in the car? Will people try to have sex, do drugs, make a mess? What happens when someone vomits in the car or is very drunk (actually a very common problem in ridesharing)?
They change the interiors of the car to accommodate. There is a reason the public transportation doesn't feel like my Ford Fusion. It is made for lots of different butts to be in and if a mess is caused they come in and spray it out. A bus for example can be cleaned with a pressure washer (sometimes inside and out.) Seats, headboards, etc. etc. can all be put on by snaps or simple screws. Anything stolen can easily be monitored and the last occupant can be charged for it. There are a lot of ways to get this done and make it extremely cheap and durable.
This is also my main concern.
It's the reason (non autonomous) car sharing is failing in France: the people behave like pigs and the cars are disgusting.
One solution would be to add censors to check after each ride if the interior of the car has been degraded.
You have a credit card on file for anyone who gets in, it seems pretty easy to police. Even the most simple method of a button on the app to report a problem with the car when it arrives would work fine.
All commercial AV services will put cameras inside the car to ensure that you don't misbehave. You'll have no more right to privacy in an AV than you do today in taxis when the cabbie looks at you in the rear view mirror. Less, probably.
They'll successfully promote the practice as offering greater security and warning you when you leave something behind.
> “Rather than building AI to solve the pogo stick problem, we should partner with the government to ask people to be lawful and considerate,” he said. “Safety isn’t just about the quality of the AI technology.”
Wow. What a goal post move. I don't think this is how successful technology spreads. Governments usually accomodate tech because it proves itself in the market, not vice versa.
Also, good luck marketing a product which only works when everyone is lawful and considerate, and otherwise has a considerable chance of killing occupants or pedestrians.
Not in the same way, there is accountability when a driver or knife wielder uses them dangerously. He's arguing for putting the accountability on pedestrians, which would be the equivalent of putting accountability on a murder victim. "Hey, it's not the murderers fault, the victim shouldnt walk around in bad neighborhoods."
The expectation with cars is not that everyone is lawful and considerate, but that everyone has a decent chance of being held accountable for their actions.
Quite the opposite. Non-autonomous cars are a product that works despite virtually nobody keeping the rules completely, and most drivers being inconsiderate.
This statement stood out to me too. If this were truly the way forward, just make the cars inherently unsafe for pedestrians. Then we will all learn that these vehicles mean certain death. That has generally worked well with poisons like cyanide. Don't drink cyanide, it will kill you. We can say the same for autonomous cars, don't step into the road, the AI cars will kill you.
To be clear, I am not in favor of building our urban transportation networks with an automotive mode focus. I feel that non-motorized transportation and safety needs to be the main focus, even if that means auto modes become inconvenient.
> Drive.AI founder Andrew Ng, a former Baidu executive and one of the industry’s most prominent boosters, argues the problem is less about building a perfect driving system than training bystanders to anticipate self-driving behavior. In other words, we can make roads safe for the cars instead of the other way around.
I'm not going to give up on crosswalks just so you can make money on your shmancy new cars.
Ng didn't say anything about giving up on crosswalks. He mentioned "bouncing on a pogo stick in the middle of a highway" could be a bad idea but I guess that's true with human drivers too.
And yet, the Uber car hit someone crossing the road, walking a bicycle, because it couldn't slot her into either "pedestrian" or "bicycle". What about someone using a walker, bent over and shuffling? It's not going to recognize that person either. The only way to make the "road safe for cars" (of that sort) is to remove all the huge variety of actual road-users from the road.
"Seriously fuck that guy" is low-grade internet sludge. We don't need that here. If you post it, others will post more, and we will get all sludgey. That's point #1.
Point #2 is more subtle and more important. You picked an uncharitable interpretation in order to get snippy at it. That's a self-referential activity, not the thoughtful engagement we're looking for. This is why the site guidelines say:
"Please respond to the strongest plausible interpretation of what someone says, not a weaker one that's easier to criticize."
That applies to articles and quotes as much as to fellow commenters. If people just rail against weak versions of what others say, the only ones interested in reading it will be others in the same mood. Everyone else will get bored and leave, which in the long (or maybe not so long) run amounts to the death of this community. That's a downward spiral we really don't want, so please make an effort not to do that.
Constantly impressed by the time you spend explaining your moderation suggestions dang, thank you for all the work you put in to make this place what it is :)
OK, thanks -- I'll be more careful with language, and next time will back up my point better. (I don't think I was picking an uncharitable interpretation, and elsewhere in this thread I clarify that.)
"Fighting Traffic - The Dawn of the Motor Age in the American City" By Peter D. Norton
> Summary
> The fight for the future of the city street between pedestrians, street railways, and promoters of the automobile between 1915 and 1930.
> Before the advent of the automobile, users of city streets were diverse and included children at play and pedestrians at large. By 1930, most streets were primarily a motor thoroughfares where children did not belong and where pedestrians were condemned as “jaywalkers.” In Fighting Traffic, Peter Norton argues that to accommodate automobiles, the American city required not only a physical change but also a social one: before the city could be reconstructed for the sake of motorists, its streets had to be socially reconstructed as places where motorists belonged. It was not an evolution, he writes, but a bloody and sometimes violent revolution. Norton describes how street users struggled to define and redefine what streets were for. He examines developments in the crucial transitional years from the 1910s to the 1930s, uncovering a broad anti-automobile campaign that reviled motorists as “road hogs” or “speed demons” and cars as “juggernauts” or “death cars.” He considers the perspectives of all users—pedestrians, police (who had to become “traffic cops”), street railways, downtown businesses, traffic engineers (who often saw cars as the problem, not the solution), and automobile promoters. He finds that pedestrians and parents campaigned in moral terms, fighting for “justice.” Cities and downtown businesses tried to regulate traffic in the name of “efficiency.” Automotive interest groups, meanwhile, legitimized their claim to the streets by invoking “freedom”—a rhetorical stance of particular power in the United States. Fighting Traffic offers a new look at both the origins of the automotive city in America and how social groups shape technological change.
There is still accountability when a driver causes harm to someone on the road, pedestrian or otherwise. He's suggesting that AVs shouldn't be accountable when they harm pedestrians under abnormal circumstances. I think that's a lot different than simply making roads amenable to drivers.
In regards to my initial point, government made regulations around cars after there was already an overwhelming market demand for them.
That argument would create an insane precedent. By that standard, it'd be OK for air-to-ground missiles to be wildly inaccurate. As long as you design the missile to be really noisy in flight, you can assume that people will hear them and take cover in bomb shelters before it lands. Collateral damage avoided. Problem solved.
-A decision was made to do nothing based on ambiguity in perception, and the emergency braking was turned off because it got too many false alarms from the sensor”
So instead of solving this problem it was decided to go forward;abominable.
At least the truth came out.
This article talks about issues with deep learning than admits it's not used on these projects. Making most of the article not relevant to self driving cars despite being about them.
One major aspect a lot of comments aren't capturing is that autonomous vehicles create a platform that goes beyond moving people around. Forget about the passengers during peak demand, that's already a given -- what about all that cargo that can be moved either with passengers or during off-peak demand hours?
As I keep pointing out, the way you start to do automatic driving is by first profiling the terrain to see where you can go. If it's not flat road, you don't go there. Doesn't matter why it's not flat. Then try to classify other road objects and predict their behavior. This only matters for moving objects.
Waymo gets this, as we know from Urmson's talk at SXSW a few years back.
Most of the DARPA Grand Challege vehicles got this, because they had to drive off-road, where you have to profile terrain or else.
Tesla does not get this. Cruise may or may not get this. Uber - well, Uber's system detected the pedestrian and ran into her anyway, which should end with someone in jail.
There's this mindset that you just throw deep learning at camera images and automatic driving comes out. Musk claimed that. It didn't work. We don't hear much from Tesla about self-driving any more. Udacity's self-driving course is also deep learning based.
As for how much testing is required, read the California DMV accident reports.[1] This gives you a sense of what the real-world problems are. 25 minor accidents so far this year. Mostly Cruise. The most common problem, especially with Waymo, is being rear-ended while cautiously entering an intersection with limited visibility. Their system will start forward to get a better view, then detect cross traffic and stop. What may help there is some convention such as rapidly flashing the brake lights when a sudden stop is likely and there's a vehicle close behind.
On the LIDAR front, Continental's flash LIDAR is already working well enough that drone makers are buying it. Continental is ready to produce that thing in volume, but they need volume orders from automakers before the price comes down.
I don't know what's up with Uber, and why they crashed into any object in the middle of a freeway with no other cars nearby, but Deep Learning is part of everyone's systems, including Waymo. Terrain analysis is great, for static terrain, but once you need to classify what objects are, or predict where they are going to go, that gets a lot less feasible as the only approach.
> The most common problem, especially with Waymo, is being rear-ended while cautiously entering an intersection with limited visibility. Their system will start forward to get a better view, then detect cross traffic and stop
This actually highlights a really hard problem that (perfect) self driving will need to solve: theory of mind.
Not only does the AI need to empathize and predict how typical humans react, they also need to be easy for humans to empathize with. Humans typically use themselves as a template for a reasonable range of reactions. So when the capabilities don't match up (like needing to scoot up to process how to cross where a typical human wouldn't need to), it will violate their expectation of you. So in this case being cautious is considered a crazy and erratic behavior to a human. (e.g., "what were you thinking?! why did you stop!? shouldn't you've seen that the road was empty?!")
Yeah, this is a good point. Most of the HN crowd understands that ML-powered autonomous vehicles will have crazy failure modes that no human would ever do. Proponents of self-driving cars claim that the reverse is also true and that, on average, self-driving cars will be safer on average than humans – and that might be true! But it still matters that the humans remaining on the road can't predict what these things will do and our theory of mind won't work: even bad drivers are usually bad in predictable ways, but a software system that makes unexpected mistakes humans can't understand has second-order dangers on the road besides the ones it creates itself.
The only time I've ever been rear-ended was that way — pulling out of a parking lot where the street was at a slightly higher level, so I couldn't see around the parked cars on either side of the exit. I started to pull out rather quickly, but then was forced to hit the brakes, and the person behind me — well, I don't know where they were looking.
I guess that's not quite the same scenario, though, since the person behind me couldn't see the cross traffic either; they were just cueing off my own behavior to conclude that it must be clear.
Anyway, you're right: good driving isn't just predicting what other people are going to do; it's also making sure your own behavior gives them the correct idea about your intentions.
> Not only does the AI need to empathize and predict how typical humans react, they also need to be easy for humans to empathize with. Humans typically use themselves as a template for a reasonable range of reactions.
Hm, this past season of Westworld kinda touched on how this very particular thought can easily go wrong.
Trying to avoid spoilers, one of the programmers commented on how the teeny tiniest tweak to parameters created Hosts that would [motions to what they were watching, a normally-benign personality killing townspeople for fun]. The awake Host with him commented "it's gone insane". His response was something along the lines of, "in the range of possible actions, what humans call 'sane' only exists in a very narrow band".
> A Cruise autonomous vehicle ("Cruise AV"), operating in autonomous mode, was traveling westbound on El Camino Del Mar between 32nd Ave and Legion of Honor Drive. The Cruise AV was struck by a golf ball from a nearby golf course causing damage to the Cruise AV's front driver windshield. There were no injuries and police were not called.
> There's this mindset that you just throw deep learning at camera images and automatic driving comes out. Musk claimed that.
Shouldn't there be serious repercussions? But seems like he can fuck up majorly and people will still line up to blow him...he does have a BILF thing going for him.
I know that the actual system is currently terrible in many places, wherein the recycle and trash are mixed together and re-sorted by humans. Ignoring that for the moment...
There are some practical reasons for recycle bins, or at least for separating out paper. Paper can get damaged if mixed with liquids or other foods/compostable material, and harder or not worth recycling. Keeping it separate from the very beginning increases the ability and efficiency of recycling => less dead trees.
What is the most optimistic benefit of a driverless car?
If we put aside creation of unemployment, putting limits on what human drivers will be able to do, and similar dystopian outlooks, what are the benefits for the masses?
People who cannot drive will be able to use cars as people who can. Is this good? Are interesting places on the planet going to have the same fate as Internet did, when it transformed from elite audiences in 90s to tragedy of commons today?
> argues the problem is less about building a perfect driving system than training bystanders to anticipate self-driving behavior
Can we stop listening to him now?
I'm joking, he's an expert of course, but seriously that's a pretty dumb view. It's exactly the sort of thing I expect from a SV "disruptor" type.
> “Rather than building AI to solve the pogo stick problem, we should partner with the government to ask people to be lawful and considerate,” he said. “Safety isn’t just about the quality of the AI technology.”
Just crazy.
These people just really do not seem to care if they kill people. They are engaging in foolish behavior rushing ahead with a technology that's not ready and thinking they can just paper over the gaps with laws and patches.
Well thankfully the Google CEO is not some super salesman like Elon Musk who is trying to emulate Steve Jobs and failing hard with Tesla.
I would bet the farm on Waymo, they realized it takes a long time to build a level 4 autonomous vehicle, so no doubt they were being poached by Uber which all failed.
Likely the problem is one of numbers. The sheer amount of data, and the traffic data Google has access to, all of this adds to a superior safety experience.
Meanwhile a level 2 masquerading as level 3 autopilot should be up for some serious scrutiny, especially if it's true that there appears to be some issue with the autopilot on the Tesla.
The Q isn’t when complete autonomy arrives, but how much autonomy we will add before the cost of adding more is vastly greater than having on-call humans take over. At that point the ROI of the research will fall as all you gain in going from 99.05 to 99.10 % autonomy is a few employees across a whole fleet of vehicles.
I agree with the pessimists: driving is 99% a mindless activity that even a mediocre AI will handle soon. But the last percent requires not just a better AI but a human. At least as long as it drives with other people in an environment designed for people.
QUESTION:
Is this really and AI problem or is it a sensor problem.
I mean, if there’s a big blob of something solid in the road in front of you, it doesn’t really matter what it is does it. You still have to not run into it.
It seems that because these poor AIs are relying on LIDAR and even just radar and cameras with Tesla, that they aren’t confident enough in that data to stop if they sense a big blob of something in the road.
If you gave any of these AIs a really high quality 3D rendering of the area around them I would presume they’d do a damn fine job of navigating it and not running into anything. At that point it’s just a computer game.
So either we need better sensors or better algorithms for extracting better 3D data from the existing ones.
It’s not like these fatalities are caused by particularly crazy problems. It’s not that a kid ran behind an ice cream van and the AI couldn’t predict it would run into the road like a human would. It’s just that they make dumb decisions a learner driver wouldn’t make because they can’t see the road properly.
More I read about AV, more I get the idea that, these AVs need to think to some extent like a human driver. My notion is that, these AV research focus blindly on making individual cars think too much. Am curious what happens if the thinking and data-collection is distributed instead, like swarm intelligence, each AV nearby can broadcast what it sees and where it is(extends the vision of each car in a given location beyond it's own sensors); front car suddenly breaks/had an accident/roadblock/switched to human intervention => all following cars will negotiate if they should bypass(reporting car will be parking) or wait (reporting car will use another lane); a car is on a blind turn => oncoming traffic from both sides can negotiate if this car should attempt the turn because others are far enough/slow enough to let it turn safely or wait until they are gone(high speed so not enough time for turning car before that one approaches); a road is very confusing so all cars following should switch to/request human intervention with ample time ... etc and given how cellular phones switch nicely between different towers and areas, we can use similar tech to allow all nearby cars to communicate ... that is communicate, coordinate and decide. Now this have some issues like privacy, drivers with bad intention can feed bad data to cause mass confusion, legitimacy of data received etc ... which can probably be an interesting place to borrow some ideas from blockchain to verify legitimacy etc ... I think instead of focusing too much on ml, try to incorporate other techs which solve minor but similar issues is a good idea.
Disclaimer: I do not claim to have full/partial/enough understanding of any AV technology/blockchain/distributed computing/swarm intelligence etc... so my comment is more or less an opinion based on what I think may be interesting, which may or may not be already thought by and discarded for not-enough viability by the AV research.
Talking about self-driving cars, I always interested in the legal aspect of these cars. What's the regulation, what are the driver's (or passenger?) responsibilities, and what are the automakers' responsibilities. I have just read a nice information related to this matter at https://www.lemberglaw.com/self-driving-autonomous-car-accid.... It's always interesting to talk about this future technology.
255 comments
[ 3.5 ms ] story [ 275 ms ] threadIf Waymo has truly solved the problem while everybody else is just trying to catch up and/or bluffing, I guess I wouldn't be that surprised.
Edited to add: The reason is that they've apparently ordered 10's of thousands of these cars. Can they really use that many only in good-weather places with perfect roads? It seems unlikely that they are just that stupid to spend that much money without being confident that their plan will work. One last option would be that they intend to put 10's of thousands of them on the road just to exponentially speed up the data collection that they've decided is necessary to truly get the system up to par.
Waymo estimates about 50 trips per day will be served per vehicle. At that rate, 80k robotaxis is enough to displace most of the taxi/rideshare services in the southwestern United States, though I'm not expecting to see all of them on the roads until 2022 or so. Waymo has a ton of work ahead of them. Mapping and validating in the ~100 square mile area of Chandler has taken them about 18 months so far, and supposedly they are on the verge of a commercial launch. Every city they hope to deploy in has differing signage, differing driving habits, and many funky intersections and road anomalies that need specific attention.
Why?
I see comments like this constantly. But does anyone ever do even a basic spreadsheet to explain the unit economics of this argument?
Low skilled people who can drive are pretty cheap and plentiful. Even if self-driving tech is flawless (spoiler: it's not and won't be soon) it still only replaces part of their responsibilities. Someone will still have to clean the cars for example. Presumably there will be some required level of human monitoring, etc.
Conversely, capital and highly reliable technology isn't free. You can calculate pretty easily the rate at which it's profitable to substitute technology for labor, this is a trade-off we've literally been making for centuries.
One one side you have plentiful cheap low skilled labor. On the other you have lasers, fast computers, graphics cards, cameras, and the associated programming inspection and maintenance costs.
Why do we think the latter side is going to be cheaper in anything remotely like the near term?
A truck still needs to be able to drive in cities, and if you figure out how to automate all the other thing a trucker does you can replace 3 million of them. Big whoop. Robotaxis stand to challenge the economics of personal vehicle ownership, the potential market is many orders of magnitude larger.
That's plausible and incremental. No cleaning problems, no homeless people living in them, no need for remote monitoring, a far simpler liability and insurance situation, and so on.
My assumption is we'll start talking seriously about robotaxis on the day this scenario is practical and battle tested. The idea of starting with taxis seems ludicrous in comparison.
These are all solved problems, I mean when was the last time you saw a homeless person camped out in a Zip Car? BMW's ReachNow vehicles are all over Seattle and they are clean, insured, fueled/charged and exactly 0 people live in them.
I have yet to hear a compelling reason why I would give up my car, and if I did why public transit isn't a better option.
Yes, that cost will be offset by the cost of the extra hardware, but that can be amortized over tens (possibly hundreds) of thousands of miles. Likewise, the investment in software (programmers) will eat into profits, but once again, you can spread that cost out over hundreds of thousands or millions of cars. Besides, an awful lot of the latter is a sunk cost at this point.
Because, other than inspection and maintenance costs, everything on your list has been getting much cheaper really quickly for a long time.
A company deploying large mostly-homogenous autonomous fleets may be able to benefit from:
That said, the potential for a lot of these savings depends on current prices actually reflecting these components. It may very well be more expensive to perform some of these activities, no matter how efficiently, than to exploitively externalize their costs on drivers and riders.To give an example, my favorite coffee shop is about 7 minutes (+wait) from my apartment by car either way, and around 18-26 minutes (+wait) by bus (shorter there, longer back). Without a ride pass and with tip, it's probably around $7-8 to Uber this one way, vs. $1.25 for the bus.
At this price, I'll usually only Uber if heat or rain make getting to and from the bus miserable. I have a ride pass atm that knocks this down to around $6-7 with tip, which makes me marginally more likely to take Uber, but it isn't my default. I took the bus this past Sunday morning, planning to get some open source work done, but I'd forgotten my laptop wasn't in my bag. The time/sweat cost of the bus round trip and the money cost of the Uber round trip were high enough that I just sat and read a book instead.
I'm not sure exactly where, but somewhere between ~$2-5 total, I'd probably default to taking Uber both ways. Down at the low end of that range, I even would've gone back for the forgotten laptop.
That's a weird thought to have. Humans need on the order of ~15K a year in the US. On the other hand, even an ultra high end SDC computer costs only around ~12K in production and much, much less to actually make since Waymo will no doubt use their own TPUs. Lasers are cheap, compact and efficient fiber lasers. Cameras are mass produced in immense volume and are dirt cheap for very high quality. The only other factor is software development costs which is mostly a fixed cost.
Rental car companies show how few people it takes to maintain a very large fleet of vehicles and it will take even fewer if they can drive themselves (an autonomous car can go through a wash on its own, charge itself, show up at a detailing station on demand, etc).
Does that number reflect an analysis of the cost of the actual human work of driving the car only, or is that just the number for average actual remittances from Uber to the driver?
If so it includes assumption of capital risks, auto depreciation, gas, tire wear, broken glass, vandalism and cleaning, traffic tickets, maintenance, towing, and every other expense of physically delivering the ride.
Getting rid of the human driver only saves real money on driving labor. The payments to drivers have much more to them.
The number of hours per day the car needs cleaning will be drastically lower than the hours it needs driving.
The very fact that those things do need specific attention tells me they don't have true autonomy yet; they just have a way to fake it using detailed maps of known areas.
I don't see how an autonomous vehicle can handle all the corner cases of real-world driving (construction zones, etc.) until it can also work reliably using only available sensor data, so that detailed mapping is no longer necessary.
Which must be worrisome for competition.
[1] https://phantom.auto/
[1]: https://en.wikipedia.org/wiki/Alphabet_Inc.
[2]: https://en.wikipedia.org/wiki/Automotive_industry#By_year
[3] https://www.dmv.ca.gov/portal/wcm/connect/5aa16cd3-39a5-402f...
Specifically, this section:
"According to the filing, Google and Motorola began discussions about Motorola's patent portfolio in early July, as well as the "intellectual property litigation and the potential impact of such litigation on the Android ecosystem".
Although the two companies discussed the possibility of an acquisition after the initial contact by Mr. Rubin, it was only after Motorola pushed back on the idea of patent sale that the acquisition talks picked up steam.
The turning point came during a meeting on July 6. At the meeting, Motorola CEO Sanjay Jha discussed the protection of the Android ecosystem with Google senior vice president Nikesh Arora, and during that talk Jha told Arora that "it could be problematic for Motorola Mobility to continue to exist as a stand-alone entity if it sold a large portion of its patent portfolio".
In connection with these discussions, the two companies signed a confidentiality and non-disclosure agreement that allowed Google to do due diligence on the company's patent portfolio."
https://www.forbes.com/sites/timworstall/2013/04/29/why-goog...
But the key was it allowed Google to do cross license deals that avoided paying a competitor patent license fees which is double bad.
Paying money and giving your competitor money.
Moto was a brilliant deal by Google.
While some have projected how many traditional vehicles they can displace, I don't think anyone really knows how that'll play out. I doubt any of these companies will go making bets that obviously rival the scale of traditional vehicle sales/ownership at the time they're made, especially before they get a sense of how reliable AV transit affects the decisions individuals make about car ownership as they face major maintenance, repairs, and replacement.
I've read one projection that says each could replace 10 traditional vehicles. If that worked out, it seems like a bet on 10m vehicles, enough to replace all 100m sold worldwide, would be astonishingly bold. If you assumed an initial scope of the entire US, the equivalent bold bet would be 1.7m vehicles. If you reduced the scope to California, that bet would be about 200k vehicles. If we guess that the difference between "astonishingly bold" and "bold" is an order of magnitude, these would be reduced to 1m, 170k, and 20k...
Google is buying a couple thousand Pacificas in the near future. That's it. It's more or less a commitment to keep buying a couple thousand a year for the next however-many years, in exchange for a discount.
It is NOT the case that Google is going to open a big box and find 62,000 vehicles inside in late 2018.
This article makes the further error of assuming “AI” is required for self-driving cars, rather than just a shitload of code.
Nobody knows because it has yet to happen. We have all seen demos and studies but that doesn't convince anyone not involved in the tech. Cars will be capable of driving themselves when I see one, sans humans, ask to merge in front of me on the highway. It will happen when I see one navigate the lineup to the ferry I take every week. It will happen when I see one accurately obey hand signals from a human directing traffic. Not youtube vids. It will happen when I see it with my own eyes. Just like electric cars, we will have to see them on the roads playing the game we all play every day. Until that tipping point, driverless remains a myth that happens in a faraway place.
So put them on the road. Let's give it a try. Maybe it works. Maybe it doesn't and people die and they sue. That's been part of driving since day one. Roll the dice. I'm sick of the endless debates.
Well I'd certainly hope that there is a high level of confidence that they don't kill people BEFORE they're put on the road with other people.
Don't. Cost isn't stopping people today from ridesharing. Carpooling won't become cool just because there is no driver, in fact I could see it becoming less cool as you will be much more alone with strangers. Many people today would be uncomfortable on a bus at night without a third party present (ie the driver, or at least many other strangers).
Think about all the extra traffic from autodriver cars doing jobs that today are too expensive. And the double-commute possibility. Your car takes you to work, drives itself home to park/recharge, then drives back to pick you up at the end of the day. Any 2-way trip could become 4-way, doubling the per-trip traffic. Or, rather than pay for parking while you shop downtown, you ask your car to drive around in circles until you are ready to go home.
> than the average sober, careful and experienced human driver
But ultimately we do accept the fatality rates caused by all drivers and not just the good ones as a cost of having individual transportation. If we didn't only teetotalers would be allowed to operate cars to remove one risk factor. And we also accept the risk of inexperienced drivers because we have to teach them eventually, the same reasoning can be applied to autonomous cars (just substitute teaching with data gathering)
The total distance driven also does not just depend on technology alone but on urban planning. AVs may change that dramatically over time.
> But ultimately we do accept the fatility rates caused by all drivers and not just the good ones as a cost of having individual transportation. If we didn't only teetotalers would be allowed to operate cars to remove one risk factor. And we also accept the risk of inexperienced drivers because we have to teach them eventually, the same reasoning can be applied to autonomous cars (just substitute teaching with data gathering)
We don't accept the fatality rates caused by all drivers; we confiscate the licenses of or jail bad drivers on a regular basis, and we certainly don't permit drivers being taught to operate commercially. Driver accountability for accidents is part of the social contract that permits individual transportation despite its cost; an corollary of that is that the accident rate is higher than we believe it should be, and that some degree of negligence is involved in the current level of accidents which it would seem unreasonable to exonerate corporations from if they were to replicate them using software. And on a purely practical basis, if you estimate SDV safety based on narrowly beating the mean accident rate across all drivers in tests, and then early commercial adoption tends to be replacing experienced commercial drivers who drive the most miles with the fewest incidents first, you're certainly going to push that accident rate up in the short run even without additional miles driven.
Uber could be skewing statistics in similar ways.
Anyway, this is all handwavy speculation. What I'm trying to get at is that the overton window for AV behavior may be surprisingly large, especially if they exhibit their worst behavior in the learning years when the fleets are still small. For some metrics possibly even larger than that for humans.
If the software is in active control pretty much everything that goes into that is different from software that is deciding if and when to emergency brake.
To put it another way, a self-driving system will become safer than human drivers, not because it knows how/when to emergency brake really well, but because it won’t have to emergency brake as frequently as a human does.
Everyone else performs a tradeoff along multiple axes including price, convenience and safety. Not needing a driver, freeing up time and lowering prices increase utility which could even justify a slight increase in risk. But even if we don't want to increase risk then a self-driving system still only has to be as good as humans on average to result in no net-increase of deaths per passenger mile while still increasing overall utility.
And more simply, it is silly to expect the maximum possible value to be the minimum requirement.
So it's all a numbers game. What's acceptable, and what we can improve on.
I’m not saying I want the highways to be a Wild West for self driving cars, but it’s worth admitting it would probably accelerate the pace of innovation.
Jesus... Let's kill people because you're sick of the debate on whether self driving cars will work.
I've been driving next to these cars for the past 7 years. A number of my friends & former coworkers have ridden in them; a couple of them work on them. They've stopped for me at pedestrian crosswalks. I've gotten stuck behind them, driving 25mph on El Camino. I've waved them on to make a left in front of me (I'm still not entirely sure how they do that, but I think that if they detect that the driver at the intersection is stopped and has not gone within any reasonable time period, they creep out and eventually go if no movement is detected). I saw one last week attempt to switch lanes, find that the car behind it had sped up and it was no longer safe to make the lane change, turn off its blinker, pull ahead in its own lane, and then make the lane change safely several cars up.
They are on the road, and they work pretty well. I trust them a lot more than cars that say "Student Driver Onboard", or even the average Californian driver.
Waymo is doing better than almost everyone else, a testament to their skill and their rigor, but they are still a long distance away from truly autonomous vehicles. The temptation is to say "well, they are so close, I'm sure they'll finish up the niggling details", but the reality is that the stuff they are failing at are some of the hardest remaining problems and will probably require at least as much work as they've put in already, if not significantly more, to tackle. I think we'll end up with self-driving cars eventually but probably not on a time scale of within the next 5-10 years, more likely it'll take that long or longer for the technology to reach maturity and then yet another decade plus for it to start seeing practical application.
This is the sad realization I had about brain-computer interfaces towards the end of my research. 99.9% is NOT good enough when the stakes are life and death and the factors within events that the environment throws at you are basically independent and continuous.
At least with self driving cars we can enforce some hard boundaries. In BCIs implementing boundaries is tantamount to denying the user of their free will.
see
>A select group of Arizonans has been shuttling around the Phoenix area in self-driving cars for the past year, providing insight into the future of the technology. http://ktar.com/story/2105890/waymo-early-rider-program-prov...
I am willing to assume that waymo/Tesla etc have found the holy Grail, and made a driverless car safer than the average current human driver on the road. I'm not convinced that we should still allow these driverless cars on the road with no human backups. Not because of expectations or liability, but because the benchmark is wrong. We should compare them to a human assisted by the same driverless technology being used passively to prevent accidents. And it's not clear if a driverless car with no human backup is safer than a human assisted by the same technology.
I'd say many millions of miles, actually :)
https://www.digitaltrends.com/cars/waymo-7-million-test-mile...
I think the real question is whether the big investors in self-driving cars will make progress creating virtual rails for their vehicles, e.g. roads where we can keep pedestrians out.
Aren't major roads pedestrian free in the US? Motorways and major roads are in the UK.
I've walked and cycled along multilane A-roads. It's not fun but it's legal.
So how do we get a determination of which disengagement events would have possibly lead to a crash? Plus I am still not convinced these numbers are even going to remotely similar over roads not heavily mapped and imaged. I think the hope many have for AV is that is allows for driving in the worst of conditions but in the end we will have AV for limited access or very tightly monitored road ways only. Which in itself is not a bad idea, rush hours could much safer and efficient with HOV/Express lanes redone to explicitly support AV.
We certainly aren't going to be able to get people to drive more lawfully when most think they are the better driver when compared to others.
I would hazard a guess that yes, that's pretty much the Southwestern US year round, if you include Southern CA. There are enough metro areas, and enough new sprawling development, to put those vehicles to work.
Waymo supposedly processes their lidar signals to detect other cars with reasonable accuracy even in snow/rain(https://youtu.be/UrJ4-AUL4U0?t=9m41s). They only operate in pre-mapped areas so they should be able to follow lanes just fine even if the actual markings are obscured (by weather or by poor maintenance).
This kind of approach (mapping everything beforehand and using lidar for perception) is used by other players as well, and seems the most likely to work for short term level 4. Trying to do everything without the benefits of lidar and high-res maps (what e.g. Tesla are doing) is a more general solution, and cheaper hardware-wise, but will take longer to get right.
Personally, I'm long on the camera-based approach because I think reliable autonomy will require a generalized vision solution, and once you've achieved that, the Lidar is redundant.
1: https://youtu.be/HJ58dbd5g8g?t=7m20s
Motorola?
If you are buying 10,000 cars, what you are essentially doing is taking on that car manufacturer as a partner. You are betting that you will need the cars, and they are bettering that you will be around to pay for them. Everything remains speculative.
"I tell adult audiences not to expect it in their lifetimes. And I say the same thing to students"
"Merely dealing with lighting conditions, weather conditions, and traffic conditions is immensely complicated. The software requirements are extremely daunting. Nobody even has the ability to verify and validate the software. I estimate that the challenge of fully automated cars is 10 orders of magnitude more complicated than [fully automated] commercial aviation."
- June 2015, Steve Shladover, transportation researcher at the University of California, Berkeley
https://www.automobilemag.com/news/the-hurdles-facing-autono...
"While I enthusiastically support the research, development, and testing of self-driving cars, as human limitations and the propensity for distraction are real threats on the road, I am decidedly less optimistic about what I perceive to be a rush to field systems that are absolutely not ready for widespread deployment, and certainly not ready for humans to be completely taken out of the driver’s seat."
- March 2016, Mary Cummings, director of the Humans and Autonomy Laboratory at Duke
https://www.commerce.senate.gov/public/_cache/files/c85cb4ef...
"With autonomous cars, you see these videos from Google and Uber showing a car driving around, but people have not taken it past 80 percent. It's one of those problems where it's easy to get to the first 80 percent, but it's incredibly difficult to solve the last 20 percent. If you have a good GPS, nicely marked roads like in California, and nice weather without snow or rain, it's actually not that hard. But guess what? To solve the real problem, for you or me to buy a car that can drive autonomously from point A to point B—it's not even close. There are fundamental problems that need to be solved."
- September 2016, Herman Herman, director of the Carnegie-Mellon University Robotics Institute
https://motherboard.vice.com/en_us/article/d7y49y/robotics-l...
This isn't new though -- "thinking machines" have captivated audiences for millenia. See the Antikythera mechanism [1], the Turk [2], and so on. The ability to replicate the capabilities of our own brains is intoxicating -- it finally signals that we, homo sapiens, understand our own minds well-enough to recreate their abilities at will.
[1] https://en.wikipedia.org/wiki/Antikythera_mechanism [2] https://en.wikipedia.org/wiki/The_Turk
Real AI would literally be the final frontier, because at that point, all the scientists could quit their jobs. The AI could replace them.
But real AI is a lot different than what everyone is selling as AI these days, which is essentially a hyper-glorified generalization of linear regression.
The autonomous cars only drive when it's 80 degrees out, sunny and on predetermined paths? Still welcomed.
Although I suppose the difficulty limits the market enough to make investment in such a scheme uncompetitive. Perhaps this is what Musk is hoping to get around that with his tunneling idea.
Let people drive manually to predetermined locations, let's name them a "station". Where the "AI" takes over and lines up cars on predetermined routes, let's name them "tracks". The AI will control the car on the "track" up until the nearest "station" to the destination. From there the driver will manually continue to her destination.
Waymo has been developing the technology for 9 years, and have accumulated nearly 8 million real world test miles in that time.For the last 18 months (or so) they have focused their testing in the Chandler area, but are not yet ready for a commercial launch.
For every new area they head into they need to solve specific problems at specific intersections. The vehicles are not yet validated for a full range of weather conditions. There is a huge amount of preliminary work that goes into preparing an area for commercial operations.
Developing an autonomous vehicle that can go anywhere as well as a human might just be 10 orders of magnitude than for commercial aviation, but what Waymo has is enough to disrupt.
- Irving Good
"Within a generation, I am convinced, few compartments of intellect will remain outside the machine’s realm - the problem of creating ‘artificial intelligence’ will be substantially solved"
- Marvin Minsky, 1967
Isn't there a saying that 80% of the code is written to handle 20% of the cases?
Do you know that for certain? Or is it just that Uber were too lazy to cope with any complications that might arise if Volvo's system braked?
There's certainly people who work with particular tech that are pessimistic or optimistic about it.
I think you have elaborated on your true objection, his past skepticism about statistical methods. I don't think that equates to a blanket skepticism. I also refer you to the talk in my original post. He points out that there are breakthroughs and pitfalls that are unanticipated. He outlines specific pitfalls to autonomous driving that he feels no one has adequately addressed.
when, while and do are just if and goto.
alpha go is all just if and goto. I'm not sure what you're looking for or alluding to.
Is there some other operation that isn't just if and goto? because i think any function you come up with, i can just make a big table of inputs that result in specific outputs.
Heck, decrement and jump if zero is enough. I'm not convinced at all that specific operations are limiting us somehow. Could you elaborate?
we currently lack the ability to recreate a brain-like entity, but the subtext that i am reading here is that the complexity of the brain is such that accurately modelling a brain in mathematical terms is impossible. the "brain-as-computer" model may not be accurate, but everything that exists can be expressed in mathematical (and therefore compute-able) terms.
i doubt that cyberbrains will run on anything that we recognize as a general-purpose cpu. gpu micro-architecture is already a significantly more efficient option for performing nn computations. as our grasp on this stuff improves, more specific silicon is being developed to make it even more efficient.
This is not very complex and accurately models neurotransmission. What's missing here is the vastness of connections from a particular neuron towards thousands or more other neurons, but the inherent function is definitely not complex.
* weight being either 1 if you adopt a continous modelization (multiple serial input provide multiple serial output) or a float if you prefer the discrete modelization (sum of input to sum of output)
I also work on self-driving cars. Why would I share technological innovations with another company?
This is an example where companies could put people over profits and share what they’ve learned to keep things safe. I don’t see a lot of research papers (I see none) coming out of Waymo, Tesla, Uber, or Cruise.
Right now, no one really knows the performance of the Waymo system. Millions of miles doesn't mean a lot. I can do billions of miles in a parking lot with no cars or pedestrians. Even millions of miles on road doesn't say anything about the performance of the components.
https://priceonomics.com/volvo-gave-away-the-most-important-...
Having some baseline collaboration on common components and safety systems could plausibly move the entire industry towards viability much faster, both on a safety and tech trajectory and for regulatory oversight. It's not a crazy idea.
Competition obviously is a powerful motivator, but sharing tech can reduce costs for everyone, at least in theory.
So how do you compete? I don't know. Along other dimensions, I guess.
There is another aspect of this that I want to bring up: auto-autos should share data with each other and with the surrounding traffic infrastructure in real-time, for safety and for dynamic traffic-shaping. They should be able to cooperatively track pedestrians and non-automated vehicles. The cars and the roads and the signals and something like Waze should all be integrated and cooperating for maximum safety and efficient throughput with low latency. And they should all share experience (training data) of normal and exceptional events across the whole fleet (regardless of manufacturer.) Optimizing across the whole thing will be, uh, optimal. From this POV, non-cooperative behavior (due to the profit motive or just people being people) by any single actor will be seen as a bad-faith move and the network can be expected to route around it one way or another.
[1] Please, let's call them "auto-autos".
Does anyone know how self driving cars will react to attempts by law enforcement officers to pull over the vehicle?
1: https://www.quora.com/What-are-the-different-police-uniforms...
Frankly even with local training, you'd still think giving law enforcement devices to stop, restart and redirect vehicles was a minimum requirement.
All of these agents follow a similar pattern of clothing (as I said, combination of similar garnments and colours) and behaviour (placement on the road, gesturing with authority, directly facing the car, etc.). Machine learning algorithms are especially good at recognizing patterns and storing their abstracted form, so it should come as no surprise that understanding what a police officer looks like in abstract is not the main issue of self driving.
> Humans are just a lot better at gauging intention and even simple stuff like parsing the word "police" at an oblique angle
Google seems to understand these intentions well enough, and at a much higher level than mere word parsing. This video is from 2015: https://youtu.be/tiwVMrTLUWg?t=9m5s
You got the car understanding all that happens at a complex intersection at 9'05, understanding what a police car looks like at 9'35, then detecting and reacting to a schoolbus and then parsing a police officer gestures right at the 10' mark. I'd say chances are these are pretty solved situations 3 years later. You can even see some creatures from their "zoo" of patterns for cars & people at 10'35.
Additionally, here is an article from last year: https://www.ibtimes.co.uk/googles-waymo-teaching-police-us-w...
> "When a Waymo car hears sirens, it will automatically pull over, yield, and stop. For example, when a number of vehicles are moving towards the scene of an accident on a highway and ambulances and other emergency vehicles are headed toward it, driverless cars will move aside and give way. Using audio sensors, the cars can detect exactly which direction the sirens are coming from and move out of the way."
Repeating an assertion does not make it cease to be false. A very small handful of pictures you linked to shows a wide variety of coats, vests and shirts of many different colours, all of which heavily overlap with general garment types and colours used in everyday clothing which tend to indicate police only with small and greatly varying trim detailing (and sometimes hats). And is the clothing and trim designed to convey authority? isn't the sort of abstract pattern recognition computers do better than humans, or even at all well. Sure, you could certainly create specific police uniform training sets for every jurisdiction and possibly even cut down false positives in other jurisdictions by geofencing them (so you don't get people stopping in California for commuters wearing the distinctive er... blue shirts and black trousers of the Hong Kong police) but it's a non-trivial undertaking even if there are bigger problems for SDVs to tackle
More importantly, unlike humans evolved to have an intimate understanding of human mannerisms, machine learning has no concept of "gesturing with authority", beyond whether moving human shapes fit very specific patterns within its calibration parameters, and police officers often don't have scope to place themselves in a particular position in order to get the car to understand them.
> Google seems to understand these intentions well enough, and at a much higher level than mere word parsing.
The video shows examples of predicting possible directions of travel of moving road users based on maps and movements (i.e. its fundamental driving model) and a shot of it recognising two arm gestures in an idealised front on positions. Neither fall under the scope of being able to understand how the traffic policemen intends to clear the blocked intersection from his shouts and gesticulations at you and various other vehicles. Humans also don't need to be signalled to go again if the black-jacketed man they've stopped for was actually trying to hail the taxi behind them.
Instead of recognizing the lights or something you could create a communication of some sort where the officer's vehicle provides a key that the car can then authenticate that it's a real police officer making the pull over request.
Heck you could even check with a police department system that the officer is on duty and in that area so a stolen police car or key couldn't be used.
[1] https://www.recode.net/2017/10/15/16472896/alphabet-waymo-se...
In most jurisdictions if there's an emergency vehicle running flashers and/or siren behind you on the same side of the road you're supposed to slow, pull over, stop, then proceed only when the vehicle passes you.
If the emergency vehicle parks behind you then you have effectively been pulled over.
Pulling over safely is practically the first thing a self-driving car team works on. It is very easy to detect flashers and sirens. So... no problem.
Well, except that self-driving vehicles may accidentally be pulled over by firetrucks.
When you hear a siren you're to become alert for the possibility of flashers behind you. You are not supposed to just start pulling over because it may not be on the same road as you and pulling over may block traffic. Imagine if every car in a crowded downtown pulled over every time the drivers could hear a siren. Instant gridlock.
So a siren played loudly on the radio will not cause a rule-following car to pull over, regardless of who or what is driving it. I presume you're talking about siren sounds in music, but...
Installing flashers that produce the same pattern as official police flashers and/or playing a continual siren is impersonating an officer. That's a serious crime.
Fake flashers will probably cause self-driving cars to pull over but they'll also cause humans to pull over. Not a self-driving car problem.
In the future police flashers can have a cryptographic element emdedded in the flash pattern that a self-driving car or a human-driven car with sensors can authenticate but that's an enhancement over the current situation.
It's one of the best-defined situations in traffic because it's precisely the fallback behavior. Detect emergency vehicle behind you, pull over as far as is safe, wait. If a self-driving car can't do that then it's hardly self-driving at all.
The real trouble, which you haven't brought up, is when the traffic is already at a standstill and an emergency vehicle approaches. The soft and gradual scooch-scooch movements that we have to do to clear enough space for the emergency vehicle are a much tougher case.
You can pull over a self-driving car. You just may have difficulty getting past it to an emergency.
"Tesla and a host of other imitators already sell a limited form of Autopilot"
Imitators? You mean GM, Mercedes and VAG that have had these systems for longer than Tesla are just bad copies of the glorious Tesla autopilot?
Unreal.
I don't think that's accurate. Autosteer is the fundamental innovation of Autopilot, and no other manufacturer released a system that "did, before or at the same time, similar functions" as Autosteer. Other manufacturers have since released systems that technically check the box of turning the wheel. But the value of a system isn't how it's described on paper; it's how useful it is on real roads. I haven't seen anyone release a system that's competitive with Tesla's except for Comma.ai (which is a very special case as they are not an OEM and only sell raw hardware).
"Imitators" wouldn't be my word of choice (it feels needlessly pejorative), but until someone actually innovates beyond what Tesla has done, I think it's fair to frame the conversation as the gold standard vs everything else.
> The "clones" are deliberately less good because the manufacturers offering them care more about safety and reliability than about marketing beta products.
When I say "maturity," I'm not talking about holding back features. I'm saying that Tesla's system works better in situations where other manufacturers offer similar features. For example, how many other manufacturer's TACC will slow down (below the set limit) when a turn is coming up and then accelerate through the turn as a human is taught to do? How many autosteering systems will sense an adjacent car drifting from its lane and "creep" away from it to avoid a possible collision? These are the sorts of mature behaviors that put Autopilot ahead of the pack and make it much more real-world useful than its competitors.
On irresponsibility, I passionately disagree. But rather than rehash a bunch of old discussions I'll just link to prior comments explaining why I feel the way I do, and we can debate it further in those comments if you're interested.
https://news.ycombinator.com/item?id=17302164
https://news.ycombinator.com/item?id=17233845
https://news.ycombinator.com/item?id=16773903
https://news.ycombinator.com/item?id=17236608
[1] https://www.youtube.com/watch?v=6QCF8tVqM3I
Thanks for the links to your other comments. I don't think there's fruitful debate to be had here - we're just of different opinions. I think autopilot encourages bad driving behavior in drivers who would otherwise be more alert, and if I've read your comments correctly you think the responsibility is nevertheless still on the drivers and autopilot probably helps more people than it harms. I'd love to revisit this if we ever get apples to apples accident data showing comparisons between autopilot-enabled vehicles and comparable vehicles from other manufacturers in the same price and age range (thus with corresponding modern and expensive ADAS systems).
I agree with this definition, but I see it more as a sliding scale than a checkbox. I think it comes down to whether you look at each system in isolation or whether you put them all on a single spectrum. I'm imagining human driving at one end of a spectrum and fully automated driving at the other, and looking at what percent of driving conditions each system provides a safety or utility improvement in (so you can see why features like speed smoothing and reacting to nearby cars quickly become relevant to my thinking). But I can see why, judging each system on its own, one would say that Autosteer is less mature at what it aims to do than simpler and more constrained systems are at what they aim to do.
You summarized my take on Autopilot as accurately as one can in a single sentence (thank you for that btw). Like you, I look forward to the day when data is available to resolve competing hypotheses like these. Elon pledged on the last earnings call to release quarterly Autopilot data reports; I'm hoping those will include Autopilot vs non-Autopilot driving usage and their respective highway accident rates, broken down by feature (TACC, Autosteer, TACC + Autosteer, etc). I think that'd be the best way to assess Autopilot safety, as it would control for all other factors. It'd be great to have apples-to-apples numbers from all manufacturers, but I think that will take longer and probably some coordination from NHTSA.
For a computer-driven car with accurate mapping, and hence foresight, it is more efficient to plan the approach so that no braking is necessary. After all, reducing cost requires efficiency rather than speed. Likewise on a long cycle ride I'll try to match the approach and entry speeds by easing-off rather than throwing-away calories by braking.
Logically, this doesn't make any sense, but taking control away from people leads to some psychological hurdles.
You can't make a robot that goes out in public and kills random people. You just can't do that. Okay?
Q: "What if the robot looks just like a car and there's a person inside the robot watching TV? Can it kill random people then?"
Still no.
Q: "But people do that all the time! Tens of thousands of people die or are maimed by traffic collisions every year! My robot can't join the fray?"
Still no. It's insane and horrible that we set up a death and mayhem lottery and that we force everyone-- children, old people, pregnant ladies, folks in wheelchairs, ev-er-ee-one -- to play it whether they want to or not. That's a bad thing we shouldn't do. But it still doesn't make it okay for you to make a robot that goes out in public and kills random people.
If you make a robot, send it into the world, and it kills someone, you are a murderer and you should go to jail, even if your robot looks like a car and there's a person riding inside it.
- - - - -
We could make a self-driving golf cart that never went fast enough to be able to injure people and it could take my mom (who has dementia and can't drive or ride the bus on her own anymore) to the doctor and back safely. We could build that today, with existing technology, there's a market for it, and it wouldn't kill anyone.
Self-driving cars are a fetish that distracts from solving real problems! (One day that won't be true, let's not kill too many more people until then, hey?)
- - - - -
Humans drive really well. Like really really crazy good well. When I first realized how well people drive it made me consider that Guardian Angels might be a real thing. But then I learned more about how the motor cortex worked and some of the uncanniness faded.
We should be so lucky as to have robots that drive as well as we do.
In the meantime, the industry should concentrate on incrementally automating traffic and stop trying to bite off more than current technology can chew.
What we're seeing now is hubris driven by ego and greed. And it's killing people.
Notice it is not the average human on the road right now, but the average human assisted by the same magical driverless technology that is developed.
One solution would be to add censors to check after each ride if the interior of the car has been degraded.
They'll successfully promote the practice as offering greater security and warning you when you leave something behind.
Wow. What a goal post move. I don't think this is how successful technology spreads. Governments usually accomodate tech because it proves itself in the market, not vice versa.
To be clear, I am not in favor of building our urban transportation networks with an automotive mode focus. I feel that non-motorized transportation and safety needs to be the main focus, even if that means auto modes become inconvenient.
> Drive.AI founder Andrew Ng, a former Baidu executive and one of the industry’s most prominent boosters, argues the problem is less about building a perfect driving system than training bystanders to anticipate self-driving behavior. In other words, we can make roads safe for the cars instead of the other way around.
I'm not going to give up on crosswalks just so you can make money on your shmancy new cars.
Downward-spiraling is the internet default, so effort is needed to avoid it.
https://news.ycombinator.com/newsguidelines.html
The guy in the article seems to be implying that pedestrians don't have the right to be in the road. I think it's pretty OK to be dismissive of that.
Point #2 is more subtle and more important. You picked an uncharitable interpretation in order to get snippy at it. That's a self-referential activity, not the thoughtful engagement we're looking for. This is why the site guidelines say:
"Please respond to the strongest plausible interpretation of what someone says, not a weaker one that's easier to criticize."
That applies to articles and quotes as much as to fellow commenters. If people just rail against weak versions of what others say, the only ones interested in reading it will be others in the same mood. Everyone else will get bored and leave, which in the long (or maybe not so long) run amounts to the death of this community. That's a downward spiral we really don't want, so please make an effort not to do that.
In short, please follow the rules: https://news.ycombinator.com/newsguidelines.html
"Fighting Traffic - The Dawn of the Motor Age in the American City" By Peter D. Norton
> Summary
> The fight for the future of the city street between pedestrians, street railways, and promoters of the automobile between 1915 and 1930.
> Before the advent of the automobile, users of city streets were diverse and included children at play and pedestrians at large. By 1930, most streets were primarily a motor thoroughfares where children did not belong and where pedestrians were condemned as “jaywalkers.” In Fighting Traffic, Peter Norton argues that to accommodate automobiles, the American city required not only a physical change but also a social one: before the city could be reconstructed for the sake of motorists, its streets had to be socially reconstructed as places where motorists belonged. It was not an evolution, he writes, but a bloody and sometimes violent revolution. Norton describes how street users struggled to define and redefine what streets were for. He examines developments in the crucial transitional years from the 1910s to the 1930s, uncovering a broad anti-automobile campaign that reviled motorists as “road hogs” or “speed demons” and cars as “juggernauts” or “death cars.” He considers the perspectives of all users—pedestrians, police (who had to become “traffic cops”), street railways, downtown businesses, traffic engineers (who often saw cars as the problem, not the solution), and automobile promoters. He finds that pedestrians and parents campaigned in moral terms, fighting for “justice.” Cities and downtown businesses tried to regulate traffic in the name of “efficiency.” Automotive interest groups, meanwhile, legitimized their claim to the streets by invoking “freedom”—a rhetorical stance of particular power in the United States. Fighting Traffic offers a new look at both the origins of the automotive city in America and how social groups shape technological change.
~ https://mitpress.mit.edu/books/fighting-traffic
In regards to my initial point, government made regulations around cars after there was already an overwhelming market demand for them.
So instead of solving this problem it was decided to go forward;abominable. At least the truth came out.
As I keep pointing out, the way you start to do automatic driving is by first profiling the terrain to see where you can go. If it's not flat road, you don't go there. Doesn't matter why it's not flat. Then try to classify other road objects and predict their behavior. This only matters for moving objects.
Waymo gets this, as we know from Urmson's talk at SXSW a few years back. Most of the DARPA Grand Challege vehicles got this, because they had to drive off-road, where you have to profile terrain or else.
Tesla does not get this. Cruise may or may not get this. Uber - well, Uber's system detected the pedestrian and ran into her anyway, which should end with someone in jail.
There's this mindset that you just throw deep learning at camera images and automatic driving comes out. Musk claimed that. It didn't work. We don't hear much from Tesla about self-driving any more. Udacity's self-driving course is also deep learning based.
As for how much testing is required, read the California DMV accident reports.[1] This gives you a sense of what the real-world problems are. 25 minor accidents so far this year. Mostly Cruise. The most common problem, especially with Waymo, is being rear-ended while cautiously entering an intersection with limited visibility. Their system will start forward to get a better view, then detect cross traffic and stop. What may help there is some convention such as rapidly flashing the brake lights when a sudden stop is likely and there's a vehicle close behind.
On the LIDAR front, Continental's flash LIDAR is already working well enough that drone makers are buying it. Continental is ready to produce that thing in volume, but they need volume orders from automakers before the price comes down.
[1] https://www.dmv.ca.gov/portal/dmv/detail/vr/autonomous/auton...
This actually highlights a really hard problem that (perfect) self driving will need to solve: theory of mind.
https://en.wikipedia.org/wiki/Theory_of_mind
Not only does the AI need to empathize and predict how typical humans react, they also need to be easy for humans to empathize with. Humans typically use themselves as a template for a reasonable range of reactions. So when the capabilities don't match up (like needing to scoot up to process how to cross where a typical human wouldn't need to), it will violate their expectation of you. So in this case being cautious is considered a crazy and erratic behavior to a human. (e.g., "what were you thinking?! why did you stop!? shouldn't you've seen that the road was empty?!")
I guess that's not quite the same scenario, though, since the person behind me couldn't see the cross traffic either; they were just cueing off my own behavior to conclude that it must be clear.
Anyway, you're right: good driving isn't just predicting what other people are going to do; it's also making sure your own behavior gives them the correct idea about your intentions.
Hm, this past season of Westworld kinda touched on how this very particular thought can easily go wrong.
Trying to avoid spoilers, one of the programmers commented on how the teeny tiniest tweak to parameters created Hosts that would [motions to what they were watching, a normally-benign personality killing townspeople for fun]. The awake Host with him commented "it's gone insane". His response was something along the lines of, "in the range of possible actions, what humans call 'sane' only exists in a very narrow band".
First link I clicked (Cruise "accident" on June 22, 2018) is pretty fantastic.
https://www.dmv.ca.gov/portal/wcm/connect/81d8865f-4c99-4342...
> A Cruise autonomous vehicle ("Cruise AV"), operating in autonomous mode, was traveling westbound on El Camino Del Mar between 32nd Ave and Legion of Honor Drive. The Cruise AV was struck by a golf ball from a nearby golf course causing damage to the Cruise AV's front driver windshield. There were no injuries and police were not called.
Shouldn't there be serious repercussions? But seems like he can fuck up majorly and people will still line up to blow him...he does have a BILF thing going for him.
When we can get rid of recycle bins, because we've finally figured out how to get our robots to sort our trash for us!
There are some practical reasons for recycle bins, or at least for separating out paper. Paper can get damaged if mixed with liquids or other foods/compostable material, and harder or not worth recycling. Keeping it separate from the very beginning increases the ability and efficiency of recycling => less dead trees.
> ...sophisticated trash-sorting robots are now turning up at recycling plants across the nation.
> The robots — most of which have come online only within the past year — are just as accurate as human workers and up to twice as fast.
https://www.nbcnews.com/mach/science/how-robots-are-reshapin...
The future is coming on strong.
If we put aside creation of unemployment, putting limits on what human drivers will be able to do, and similar dystopian outlooks, what are the benefits for the masses?
People who cannot drive will be able to use cars as people who can. Is this good? Are interesting places on the planet going to have the same fate as Internet did, when it transformed from elite audiences in 90s to tragedy of commons today?
Can we stop listening to him now?
I'm joking, he's an expert of course, but seriously that's a pretty dumb view. It's exactly the sort of thing I expect from a SV "disruptor" type.
> “Rather than building AI to solve the pogo stick problem, we should partner with the government to ask people to be lawful and considerate,” he said. “Safety isn’t just about the quality of the AI technology.”
Just crazy.
These people just really do not seem to care if they kill people. They are engaging in foolish behavior rushing ahead with a technology that's not ready and thinking they can just paper over the gaps with laws and patches.
I would bet the farm on Waymo, they realized it takes a long time to build a level 4 autonomous vehicle, so no doubt they were being poached by Uber which all failed.
Likely the problem is one of numbers. The sheer amount of data, and the traffic data Google has access to, all of this adds to a superior safety experience.
Meanwhile a level 2 masquerading as level 3 autopilot should be up for some serious scrutiny, especially if it's true that there appears to be some issue with the autopilot on the Tesla.
I agree with the pessimists: driving is 99% a mindless activity that even a mediocre AI will handle soon. But the last percent requires not just a better AI but a human. At least as long as it drives with other people in an environment designed for people.
I mean, if there’s a big blob of something solid in the road in front of you, it doesn’t really matter what it is does it. You still have to not run into it.
It seems that because these poor AIs are relying on LIDAR and even just radar and cameras with Tesla, that they aren’t confident enough in that data to stop if they sense a big blob of something in the road.
If you gave any of these AIs a really high quality 3D rendering of the area around them I would presume they’d do a damn fine job of navigating it and not running into anything. At that point it’s just a computer game.
So either we need better sensors or better algorithms for extracting better 3D data from the existing ones.
It’s not like these fatalities are caused by particularly crazy problems. It’s not that a kid ran behind an ice cream van and the AI couldn’t predict it would run into the road like a human would. It’s just that they make dumb decisions a learner driver wouldn’t make because they can’t see the road properly.
Disclaimer: I do not claim to have full/partial/enough understanding of any AV technology/blockchain/distributed computing/swarm intelligence etc... so my comment is more or less an opinion based on what I think may be interesting, which may or may not be already thought by and discarded for not-enough viability by the AV research.