So sounds like we'll get level 4 autonomy in 2020. Once that happens the discussion around the nature of the approach to level 5 (which is probably asymptotic) will be a welcome change vs today where people are questioning if we'll ever have a consumer successfully use an autonomous vehicle.
Even if the algorithms more-or-less know how to drive, the best algorithms will eventually be chosen by natural selection. I wouldn't want to be along for the ride during this process.
I think if you look at evolution more closely you'll see that it is messy and is constantly creating dead-ends. Mammals were insignificant until an extinction event occurred. It's not like we were superior and displaced the dinosaurs.
We have lots of errors and junk DNA. We make all kinds of mistakes. Perhaps not fatal ones yet, we'll see how climate change plays out.
As for our perception and reflexes... I wouldn't brag that they're that good just because they haven't been bad enough to cause our extinction.
> I think if you look at evolution more closely you'll see that it is messy and is constantly creating dead-ends.
That's my point. The natural selection process for self-driving automobiles is just beginning, as they meet the challenge of the real world. I'm concerned that it is going to mirror the messiness that biological evolution involves.
Sure there is ~200,000 years of natural selection. That probably 10,000 generations of learning assuming 20 years.
But computers can run 10,000 simulations to learn in minutes. And obviously we are not going to create a autonomous driving ability in minutes but we can speed up the learning evolution significantly in machine world vs humans that have 10-50 years per evolutionary learning event.
My view is the hard thing about autonomous driving will be the change in style of accidents. Even if autonomous driving has half of human accidents we will probably see things that to the human eye seem really obvious so people will be 'if it can't predict that why should I trust it' type thing. But we won't see the many accidents a computer avoids that a human would make or finds obvious.
I'm positive about it but suspect we will see the roll out take many decades as it starts in cities with the right conditions, moves to trucking routes but will probably take some time for national metro coverage.
I was waiting for that :). Let's put it out at billions of generations for pre-human evolution.
Point stands. I computer will do millions of simulations in a day. And be done across a population far greater than historic evolution.
Doesn't take long to get billions of learning events. On top of which evolution pre-intelligence (as a very loose term) was somewhat random, where our systems are starting from a much more evolved position that does take advantage of existing evolution as much as humans do.
Not really- a single base-pair can change the entire tertiary structure of a protein. Which is something computers are not great at predicting. Depending on how you calculate it, 1 "cycle" of life can be worth trillions^2 or more cpu cycles. It's a meaningless comparison, as presented.
the 80/20 principle says it's going to be a long, long time before level 5 autonomy is reached. That's 100% in all conditions. In fact, I would say it's more theory than attainable.
There's a reason they're testing in Phoenix. It has almost no Inclement weather. Throw a little snow, mud, or sleet in there and it gets hairy fast. Level 2 is basically the 20% of effort in 80/20. It's exponential effort from then on out.
We shouldn't set a dramatically higher safety standard for autonomous drivers than human drivers.
When autonomous drivers have the same or fewer injuries per billion miles driven as human drivers, then for all practical purposes it doesn't matter if it's level 4 or level 5.
If so why go through all this trouble to creat autonomous vehicles? I, for one, am glad this was not the prevailing attitude when commercial airliner regulatory bodies were set up.
Yes, even if autonomous vehicles are on average exactly as safe as human drivers, they're still worth the effort. Here are some reasons:
1. They're cheaper. This sounds crude, until you realize that this makes things like buses and shared rides far more viable. Meaning millions of people have a better commute and collectively save many lifetimes worth of time. It can also mean less air pollution, due to less traffic congestion (this only applies if congestion pricing is done), which will also save lives.
2. They can always get better over time. Humans can't.
3. Fewer parking lots are needed, which makes cities far more pleasant to live in.
4. Disabled, elderly people and children have better mobility, and aren't dependent on others to drive them everywhere.
> 1. They're cheaper. This sounds crude, until you realize that this makes things like buses and shared rides far more viable. Meaning millions of people have a better commute and collectively save many lifetimes worth of time. It can also mean less air pollution, due to less traffic congestion (this only applies if congestion pricing is done), which will also save lives.
IMO autonomous vehicles are just going to further push public transit into something just for the poor. All the convenience and comfort and privacy of sitting in your own vehicle, but now you don't have to drive yourself! Why would you choose anything else?
Sure, but but buses would be cheaper to run if they didn't need a driver. My university runs a bus that you have to call and it drives directly to you and picks you up, I could imagine something like that being more common (for public transit)
But the buses wouldn't need windscreens or wipers or chairs or steering wheels or pedals or payment/coin handling, and the drivers wouldn't need uniforms or eye tests or retraining on routes or overtime or time and a half for holidays. Bus companies won't have to insure the busses for human drivers.
So, yes savings. Nyah on your "unions defending humans against exploitation are the worst thing ever" narrative.
That's why I mentioned congestion pricing. Or inverting the carpool/non-carpool lane ratios i.e. every lane is carpool-only except those designated as single-occupant-allowed, which makes driving by yourself expensive in terms of time.
Also, if you don't have to accommodate a human driver and controls, many safety features, bulky gasoline tank, and trunk space, automobile design could change radically. Self-driving taxicabs that look like small pods, or covered golf carts, could be used for individual travel, increasing space efficiency.
Zero deaths are ideal, but the point is that self driving cars don't have to clear a very high bar. The human driver alternative doesn't do very well, as outlined in your link.
the link show aggregate death for all kind of road vehicles and all kind of weathers. select for clement weather, exclude bikers, filter by city drivig so the comparison is in line with what waymo does and suddenly the bar is much much higher.
I bet most people think they're a better-than-average driver though. So if the cars are only a little better than the average person, you'll still have everyone assuming it's less safe for them.
That’s a straw man. Humans routinely drive in inclement weather. In Atlanta during the summer I85 will maintain 50 mph in rain so hard that you can’t see the lines on the road. Expecting self driving cars to handle those conditions isn’t setting a higher standard.
Self driving cars that can't operate in heavy snow or rain will be totally useless.
One of the lessons of engineering history that self-driving car proponents fail to grasp is that promising technology often fails to get traction because it can't close the final 5%. Supersonic aircraft, for example, have effectively been derailed by people on the ground near airports complaining about sonic booms. VTOL aircraft never replaced aircraft because of control problems. Flying wing aircraft never became popular for similar reasons.
Operating domain is the key concept with Level 4 and Level 5. Level 4 any time anywhere is not a 2020 goal, but Level 4 in specific locations and conditions is on the way. Same with Level 5. I wouldn’t be surprised to see a Waymo car without a steering wheel driving people around Phoenix in 2025, but a Chicago winter is a different story.
The difference here is that now the "public paying customers" and "full self-driving" customer groups are merged - so this means there's genuinely a paid full self-driving service launched as of today! A miracle
Calling this service public is misleading, especially since Waymo created the bifurcation between Waymo One (no NDA) and Early Riders (NDA). Both of these programs you have to apply for and be accepted, so neither is public. And Waymo has made no claim of merging these. The upcoming driverless rides Waymo refers to in this article could very well be for the Early Riders still under an NDA - which theoretically could already be happening based on their past announcements (although the abundance of cell phones strongly suggests this is not the case).
This announcement is just a repeat of what they've said before. If they actually do it this time then that's great, but they haven't actually said anything new here.
Wait, what?
Waymo does also some rides hailed via lyft, or are you saying that you have to apply also in lyft?
If this is not a public service with paying customers then I don’t know one...
https://www.google.co.uk/amp/s/techcrunch.com/2019/06/27/way...
I am unaware of anyone getting picked up in a self-driving car using Lyft (outside of Aptiv in Las Vegas and other testing that has nothing to do with Waymo). Although honestly I wouldn't be surprised if I was wrong on this one, and there have been people that got picked up in a Waymo using Lyft, but a quick search on twitter and google shows up nothing (but I spent less than a minute searching just now and welcome being proven wrong).
This just reinforces my point. Waymo can announce anything with vague terminology of things happening "soon" but we still don't have any verification of anything happening.
Lyft has been dispatching Waymo self-driving vans (with safety drivers) in Chandler AZ for a while now. I've opted out of Waymo for my occasional rides with Lyft, so I can't say how frequently they show up. But I can say there are lots of the vans running in the service area.
Not the GP, but presumably because opting in to being a beta tester isn't something they're interested in, especially since the risks involved are less 'my app doesnt work' and more 'my life may stop working' if something goes wrong.
Note that while that risk remains true of every transportation journey you take, there is a lot of press generated when an autonomous vehicle takes a life or causes injury, which plays a part in a lot of peoples decision making.
Also note recent articles about automatic stop detection failing >50% of the time. This may or may not be relevant to AVs, but I personally would definitely not want to be in an AV that hit a child. That (theoretical) child's life is worth more than pushing driverless cars into the world before they're potentially ready.
Waymo is still Waymo whether they're running rides on their own behalf, or under the Lyft name. Someone who's ridden Waymo in an early adopter program may make their decision on whether to ride Waymo under Lyft based on that experience.
Except in the previous article, driverless rides were already happening:
> since mid-October, the company has been operating its autonomous minivans on public roads in Arizona without a safety driver
The difference is that those were without any users inside. This new article is saying that the public offering, with real customers inside, will get rides with no one inside.
It is unfortunate that it took 2 years to get to the point where they felt safe putting people customers in the car, but it's very misleading to equate the two and use is as a way to imply they lied last time around.
> It is unfortunate that it took 2 years to get to the point where they felt safe putting people customers in the car
Except they're not at that point yet, and even when they are they will still have a Waymo employee on board "initially".
I don't think that anybody is accusing anybody of lying, just making the point that predictions of the marketing team aren't really something you would want to set your watch by.
> even when they are they will still have a Waymo employee on board "initially".
I think you're confused. They've been having customer rides with employee in the car for almost a year now. The email sent to customers today is saying that they will soon find rides with no employee inside.
And again, rides without employee inside have been happening also since 2017, just not with customers inside.
Maybe I am confused. I'm just going on what Waymo says.
The point is that "soon" is meaningless in this context and Waymo very clearly states that "a Waymo employee would likely be present in the vehicle initially."
Like I said, a lot of different misleading statements are being confused people in this thread.
1. Waymo has been having rides with no one in the car for 2 years now.
2. Waymo has been having rides with real customers inside for 1 year now.
3. Waymo has told said customers in #2 that they may soon see rides with no employee inside
I agree that the soon in #3 doesn't mean much, but at the same time, this wasn't a press call, it was a private email sent to users, so it's not like they're showing off.
But 1 and 2 have already been happening for a long time.
> it was a private email sent to users, so it's not like they're showing off.
Of course. Companies tell their customers about new products in the hope that they will keep that info secret.
I think the salient point about your list is that while 1 and 2 have been done separately for some time, the only timeline we have for them happening concurrently is "soon"
That's a mandatory system they need to be on the street [0]
> But those cars won’t be operating completely unmanned — at least for now. Under these regulations, driverless cars being tested on public roads must have a remote operator monitoring the car, ready to take over as needed. That remote operator — who will be overseeing the car from a location outside of the car — must also be able to communicate with law enforcement as well as the passengers in the event of an accident.
That being said, the way Waymo does it is different from other companies. As HN can understand, adding the ability to remote access a car is very very risky. Waymo on the other hand has the ability to suggest a path to the car, and the car will itself locally evaluate to see if the suggested path is safe before executing it. So it's not quite a "remote control" access.
How many cars can one employee be "inside of"? If that number is an order of magnitude higher than 1 the cost of these employees is probably relatively negligible.
If it's an order of magnitude larger than one, then it's probably a pointless exercise anyway. What mortal can honestly say they're attentively watching 10 streams at once? A car accident can happen in the blink of an eye, if you're dividing your attention between several streams it would be very easy to have a slow reaction time to a child jumping out into the street, for instance.
Waymo is so far ahead in this game. Tesla has the biggest online cheerleading section and an "autopilot" that behaves as a low-budget adaptive cruise control with lane keeping that sometimes works, and a pretty big body count. Uber has a dead pedestrian and a major lawsuit. Cruise has a lot of people on staff but virtually nothing to show for it. Lyft exists. Waymo is in production revenue service.
> Tesla has the biggest online cheerleading section and an "autopilot" that behaves as a low-budget adaptive cruise control with lane keeping that sometimes works, and a pretty big body count
I don't think you're being very accurate there. There are lots of videos of people commuting 50+ miles to and from work every day without touching anything while on the interstates, including interchanges. It's doing a lot more than just lane keeping.
I disagree. It’s completely fair. Handling basic tasks on nice freeways is exactly what adaptive cruise control + lane keeping is for. Last I checked, autopilot couldn’t even stop for a stop sign unless you count “shadow mode” claims.
Based on the description in that article, it doesn’t seem that the Tesla autonomously reacted to the stop sign, so much as it slowed to a stop before making a right turn based on its navigation system. The driver says he had to manually start AutoPilot after the stop for it to initiate the right turn:
> “The car always stops at the end of a highway exit since Navigate on Autopilot. At the stop, you just need to touch the accelerator to start Autopilot again without Navigate on Autopilot. I didn’t accelerate, you just need to touch the accelerator for less than a second and the car will do the rest at the intersection.”
Perhaps you should confirm what it's actually doing in the video? Tesla itself says that stop sign and traffic light handling is in shadow mode only [1].
It's trying to center in an extremely large lane. An easy fix would be to not have a huge concrete wall jutting into the middle of the freeway without lane markers.
I feel like people forget human drivers hit those same barriers with alarming frequency.
Comma ai's OpenPilot makes freeway driving really nice. I've gone miles without touching pedals or the wheel, it's definitely better than stock. Though I haven't tried AP on Tesla, and OpenPilot is basically just a really good LKA and ACC.
What other vehicle will change lanes by itself, and take on/off ramps through interchanges to take you where your navigation is routed, without you touching a single thing?
OK, sure. You have to touch the wheel or it turns off.
What I mean is, you are not providing input to the wheel, or pedals, or turn signal or driving the car in any way. You're just letting it know you have not fallen asleep.
None, because every other manufacturer is more concerned about not killing people than they are about releasing beta features for cars. The most surprising thing to me about autopilot is that Tesla hasn't been sued into oblivion for it.
Freeways are a very ideal environment - wide, clearly marked lanes, no traffic lights, no pedestrians/bikes, no random stops, no navigation ambiguity. Unless Tesla has self driving cars on city roads (which all of the others so), they can't really be part of the same conversation.
I used to think that, but now I agree with George Hotz. Tesla uses machine learning to map inputs to control outputs. Google is using it just to build a model of the world and then use program logic to decide the control action to take.
The problem with the latter approach is the long tail of real life variety doesn't lend itself to a fixed set of programmed rules. You'll always have more edge cases.
I think Waymos approach gets you to 90% faster, but then plateaus.
Not sure there is enough evidence in support of those claims about Tesla. Tesla doesn't collect anywhere near enough data from its customer fleet to support the kind of massive training being claimed. In contrast Waymo and Cruise vehicles are regularly in the depot where each car offloads terabytes of data. Who has the bigger training set?
1:49:30. I’m on mobile otherwise I would link to the exact spot.
Tesla has tens (hundreds?) of thousands of vehicles on the road right now that are continuously feeding training data back to Tesla. What evidence is there that this is insufficient, or even deficient? From what I can gather, Waymo has limited testing geofenced to Phoenix, Arizona. Tesla has cars all over the world with autopilot, all providing training data.
I don’t see a reason for your conviction that Waymo has an advantage here.
Hundreds of thousands. The 200k mark only in US was achieved more than 1 year ago, and it triggered the phase out of the incentives.
For the last couple of quarters Tesla has been delivering almost 100k vehicles so it won’t be that long to reach the 1 million mark globally.
I think it's the other way around: Direct mapping of inputs to outputs will work great for a subset of conditions, but the last N% is infeasible without explicit logical programming.
You'll never run out of edge cases if you have to explicitly program the logic. At some point you have to let the machine learning decide it. Maybe there is room for a hybrid approach, but I think anything relying on programmers having thought of every situation ahead of time is doomed to fail.
You're wrong about the requirements. It doesn't have to be perfect, just better than us, and that's easily doable for computers that don't get tired, distracted, or drunk.
I'm not sure how long it will take, but I'm sure we'll get there.
>Tesla uses machine learning to map inputs to control outputs. Google is using it just to build a model of the world and then use program logic to decide the control action to take.
I suspect that isn't true in so far that engineers will build what works. Building a complicated system like an autonomous car will necessarily involve a messy combination of a bunch of different strategies and with the result converging on a common solution.
While Tesla is far behind Waymo and probably never will catch up, especially because of CAPTCHA, I think your description is unduly harsh. Autopilot works much better than every other TACC out there and is the best thing on the market you can get.
Since it lets you fall asleep as long as your hand (or something) is on the wheel, I'm not sure it's the best thing you can get. Cadillac SuperCruise seems better from a driver awareness standpoint, given that this is "predictable abuse" otherwise.
I am not sure what you are talking about. Tesla autopilot regularly makes you either tug at the steering wheel or press one of the buttons on the steering wheel. Having your hand on it isn’t enough.
It doesn't let you fall asleep. It needs force on the wheel to prevent disabling itself, and just a weight attached isn't enough. It needs a change in force.
Honestly, if your only critiques are based on hearsay then maybe you shouldn't critique.
That said, seems like Tesla will still benefit from Waymo being the first out of the gate. While Tesla plays catchup in technology, Waymo will take all the early criticism, while ostensibly making driverless cars more palatable to mainstream society. Similar to how Lyft benefitted from Uber being the first to bully local governments into legalizing non-taxi app-driven car service.
I’m going to ignore the fact that you’re flatly incorrect about the rest of your posts regarding Tesla, since others have more ably corrected you than I could, and simply ask what the hell this even means.
A “production revenue service”? This sounds like meaningless babble. Waymo does not currently have anything available that you can buy, lease, subscribe to, or otherwise exchange money for. That is, in fact, the entire point of the article: “coming soon,” they say, with an additional clause on top of even that: “coming soon you may...”
So: what do you mean when you say “Waymo is in production revenue service”? Because that sounds like nonsense, to be honest.
Nah man. They have a ride hailing app and seevice in Arizona. The email was to customers of this "production revenue service" that some rides soon will not have a safety driver in the car.
Good to see this finally happening. I wonder how they're going to scale this out though -- AVs are still pretty expensive to build/operate (COGS/TeleOps/Maintenance etc)
Haven't seen a recent study/article around cost/ROI (not sure if these have gone down significantly) -- any recommendations?
Techcrunch took an email sent to the customers from reddit and wrote an article from it. If they wanted to fundraise, they would've made a whole press call for it, not sent an email to their handful of users in Phoenix.
I can have this major feature done for you, but I'm not sure if it can be done in 2 months, so I say 3 to be conservative.
OR
I can have many small sub-features that relate to this major feature done in 2 months, but I might know exactly which features make the cut in those 2 months.
"promise features or promise a release date, but never both"
I agree and it's such a golden rule that no one follows or cares about even the slightest. Features and deadlines are to SE like velocities and positions are to QM.
I have very mixed feelings about autonomous cars. One side of me is so excited about the prospect and finds it all amazing and futuristic. I love reading about how Tesla and Waymo approach it differently.
The other side of me is absolutely not ready to trust my life to a machine - even though I know I do that all the time in other ways in modern life.
I think it will be people's emotional evaluation that will matter more than anything else and it will be a bumpy road to acceptance.
Driverless cara will probably be safer statistically but they’ll simultaneously make errors a human wouldn’t, so some people are going to die in what will seem like really dumb ways. This will be hard to accept even though that already happens with human drivers. Probably because their mistakes will be easier to identify with and explain.
every now and then I do something while driving that, IMO, requires "human cognition". Not just image recognition, obstacle avoidance, rule following, etc.
Like, I'm cruising along at 45mph and a plastic bag is blown in front of my car. Yes, I see an obstacle and need to make a decision about what to do - but I do this as a human who knows what a plastic bag is, and how it is far safer to drive right over it than slam on the brakes. I know a plastic bag is not, say, a Pomeranian.
Another example: I'm driving, and from beyond my plane of vision, a basketball bounces onto the road up ahead. I can't see past a building, but I know I'm near a park where kids frequently play, and I know it's possible a kid comes running into the road to grab the ball. I proactively slow down. I guess you could say a leavel-5 autonomous car would see the ball and slow down anyway, but there's still that lack of cognition that concerns me.
We like to think of driving as a purely rule-based game that is simple enough to model and train against. And I believe in 99% of situations, it is. But in that 1%... the plastic bags and all that...
I think that sums it up really well. In those kind of edge cases like the basket ball, if an AI driver didn’t slow down and then killed a kid, we’d all be up in arms about It saying how obvious it was that a kid would run out, and we’d have stopped, and we’d be right.
But at the same time that AI driver will have stopped in time when 10 other kids jumped out without any warning because it has much, much better reflexes.
But it’ll be the first instance that will get all the airtime.
Nearly 1.25 million people die in road crashes each year.
Think about that.
Self driving cars aren't just a solution to nuisance of driving. They're a solution to one of the leading causes of DEATH and INJURY in the country.
Sure there are going to be edge cases where we as humans will be able to point at the car and say "damnit, we could have prevented that"!
But I sure hope when we do that, the car points right back at us and asks "And for the other 1.20 million lives I've saved this year? You would have killed them"
> Sure there are going to be edge cases where we as humans will be able to point at the car and say "damnit, we could have prevented that"!
This is a meaningless assertion unless you account for how often "edge cases" arise. If unpredictable pedestrians, unplanned road construction, etc., is an "edge case" the car basically won't work in DC or New York.
no, they will, at some point in the future, be a solution. until we seen serious miles in comparable conditions they are an unproven technology, that why I appreciate waymo approach of getting from here to there slowly and carefully compared to Tesla which has a dozen death already under the autopilot and hides them under a pile of pr and terms of services.
but for all the tech they throw at it until the fundamental issues of cognition and prediction are solved they will only be able to reduce a fraction of the accidents because they aren't replacing the most dangerous vehicles on the road and they only are better for a subset of cases, like inattentive or impaired/intoxicated driver.
>> Self driving cars aren't just a solution to nuisance of driving. They're a solution to one of the leading causes of DEATH and INJURY in the country.
If reducing DEATH and INJURY are the motivation for self-driving cars, we can achieve a larger reduction with technology that is available right now: public transport. And there is even a no-technology solution: a drastic reduction of road traffic that passes through residential areas.
Putting more cars on the road might or might not reduce DEATH and INJURY, given that those cars are going to be self-driving. But _removing_ cars from roads will _certainly_ reduce DEATH and INJURY- and also reduce the environmental damage from the cars' internal combustion engines and the infrastructure needed for those cars to circulate.
Considering all of the above, I'm really not convinced that the self-driving car industry is really that interested in reducing DEATH and INJURY. I think their main motive is to increase PROFIT.
I don't think anyone is debating that the #1 factor driving the innovation of self driving cars is profit.
That said, the benefits of self driving cars aren't precluded by the motives for developing them.
While I hear your points about removing cars from the road, and generally agree with them, there are definite benefits to cars which are arguably not replaced by things like public transit. There will always be a need for cars, so lets make them as safe as possible. If self driving is a step in that direction, sign me up.
Yes, except this presumes that AI would recognize the kid at the same time as a human driver. Or at least soon enough such that AI's quicker reflexes can compensate for any delay.
If AI can't recognize some obstacles at all, then I think it's safe to assume there may be some delay in obstacle recognition compared to humans.
You’re probably right. The point I’m making though is there’s going to be some things AI is better at and some things humans are better at and it’s going to be much harder to accept dead people between those points.
1) Everyone assumes the good will outweigh the bad. (And why proof of humans doing one task better than AI is somehow also taken as self-evident proof that the AI will be better at everything else.)
2) No one assumes that we, as humans, will ever be able to improve our driving skills in way that can keep up with AI advances. Someone could invent a magic driving pill. Which is crazy. I just don't assume that it's any crazier than a near-perfect driving AI.
> if an AI driver didn’t slow down and then killed a kid
An AI driver would have 360 degree coverage with much faster reaction time - think about the cases where a car brakes before the passenger even realises there is danger ahead.
The state of the art systems designed only to brake before hitting obstacles can't even do that. It's a pretty simple problem, as far as self-driving tasks go.
I don't think that bodes well at all for full autonomy - especially since at least one of those vendors (Tesla) is promising self-driving sooner than later.
> Not just image recognition, obstacle avoidance, rule following, etc.
I would argue that the examples you gave are just image recognition, etc. What is your brain doing if not image recognition and obstacle avoidance?
We're not there with today's technology but eventually our machines will perform better at those uniquely human tasks you mentioned. After all, we humans and our cognition is just the output of these giant meat computers inside our skulls pulling in sensory data and running pattern recognition algorithms.
The $65,000 question is whether or not today's technology is sufficient (however barbaric) to be acceptable for the job.
> I would argue that the examples you gave are just image recognition, etc. What is your brain doing if not image recognition and obstacle avoidance
Planning, anticipation, reading body language, etc? Is that person looking like he’s going to jaywalk? That’s not image recognition. Yes, once he steps into the road the self driving car will respond quicker. But the human driver might already be slowing down and watching for hints of initiating movement. Image recognition is just one tiny piece of it.
"Image recognition" is a small subset of "receiving visual input and processing it." We know how to do image recognition with machine learning. Planning and anticipation is a much harder problem.
Simply track all the pedestrians all the time, and you'll know in advance where the common crossing points are, and when any pedestrians are moving towards a road where your car is moving. Apple, Google, Facebook, have every advantage here.
Tesla track where every Tesla goes, all the time. Will Apple make a car before Tesla makes a 'smartphone'?
Actually there are powerful techniques to predict what a person will do based on current position, velocity, body language and other cues. They can be trained by trying to predict future actions from past state, and self driving cars apply these models to all objects - people, cars, other things on the road.
1) If I'm considering whether to brake for a plastic bag, I'm also going to check for anyone tailgating me, and take that into consideration. Including whether said tailgater is a geo prism or a Freightliner attached to a tanker of gasoline.
Slamming on the brakes for a plastic bag is not the universal safer choice.
2) Already mentioned this, but reaction/breaking time is meaningless without timely recognition.
> Slamming on the brakes for a plastic bag is not the universal safer choice.
...I didn't say it was. And you call me the one making assumptions.
> reaction/breaking time is meaningless without timely recognition.
If it can't quickly recognize large moving objects then it's not qualified to drive at all. I'm assuming it can do that because self-driving at all is part of the premise.
(Well, specifically the recognition is necessary for basic driving functionality, and being faster is a matter of adding compute power and using cameras at least as fast as the ones in an iphone.)
1) I'll rephrase: occasionally (and needlessly) stopping for a plastic bag (or any other false positive) should not be considered a massive improvement over current human abilities
2)
>>>you might preemptively slow down, but a properly designed system can brake a whole lot faster.
>If it can't quickly recognize large moving objects
I'm not questioning whether it can recognize quickly. Just whether it can ever be better/ faster than humans, taking into account cognitive ability.
It's whether human
(prediction-or-recognition time + reaction time + breaking time)
is more or less than
(AI recognition time + plus reaction time + breaking time)
If AI can react and brake 2 seconds faster than a human, but doesn't recognize the threat until 4 seconds after humans can, then AI drivers will kill more people than human drivers. Which won't be fine.
>being faster is a matter of adding compute power and using cameras
There are other comments about the limitations of relying only on image recognition. While compute power per might improve the AI recognition time, humans still have the advantage of being able to predict instead of react. This will continue to be a major limitation for AI.
> 1) I'll rephrase: occasionally (and needlessly) stopping for a plastic bag (or any other false positive) should not be considered a massive improvement over current human abilities
Okay, but I didn't say that either. I said it would be an improvement to the experience I have. I will trade an abrupt stop in front of a plastic bag for several hours of being able to do what I want while the car does the driving.
> If AI can react and brake 2 seconds faster than a human, but doesn't recognize the threat until 4 seconds after humans can, then AI drivers will kill more people than human drivers. Which won't be fine.
The AI processing time to recognize things in a frame needs to match the framerate, or the system is just broken. Maybe it waits for 3 frames to show the same thing before acting, that's fine. That's still faster than a human. Especially if you use a 240Hz camera, which is not very expensive. And you can and should put in enough computing power to run the 'reacting' system at the same frequency.
Then once it decides to brake, instead of waiting for a foot to move it can put the virtual brake pedal at any position in under a millisecond.
I can imagine a car so poorly designed that its reactions are slower than a human. But it's hard to imagine a car that's simultaneously certified for letting me take my eyes off the road, and is that broken.
So I still say that a self-driving car would be able to brake quicker and harder, and it should be fine. Even without the ability to predict.
>for several hours of being able to do what I want while the car
One word: Chauffeur
>The AI processing time to recognize things in a frame needs to match the framerate
If you're talking about framerates at all, you're still missing the point. I'm talking about human ability to predict and anticipate. The fact that a human can brake before there's anything in the frame to even detect. We've (humans) got a head start, so the speed of the AI isn't all that important.
> brake before there's anything in the frame to even detect
The scenario was slowing down a bit, but still needing to hit the brake in response to the child. So if the AI can brake faster it is extremely important.
Up to a point - ordinary seatbelts even with smart pre-tensioners aren't great at this, and it may increase the rate of whiplash injuries for the passengers who aren't expecting braking.
How many parks in the USA have speed zones around them that aren’t restricted to 15-25mph? It is stupid easy to slow down for a kid chase ball scenario when you are already going slow anyways, the only reason it is such a problem at all is because human drivers prefer speeding so much.
Tons of them? I don't think any of the parks around here in exurban maryland have any sort of special speed zone. My 100+ year old neighborhood doesn't even have its layout mapped correctly in Google Maps. (The road names in the app don't match the street signs and the wayfinding doesn't realize that the road in front of my house is segmented with barriers so you can't drive all the way down it.)
Even in that case, the city default speed limit (when not posted) is usually something low like 25 (Seattle is 20!). Yes, many human drivers ignore that, but self driving cars don’t have that luxury.
We really should do this like the Dutch: when something problematic is encountered, engineer and regulate the hell out of it. Not just for self driving cars, but it works for humans to some extent that the laws are actually enforced.
The flip side is every now and then a situation arises that requires machine reaction times or attentiveness.
Waymo cars have driven over 10 million miles so far without any fatal accidents. That's not nearly enough miles to compare against human drivers, but it'll be interesting to see who does better once they log some more miles.
While that's true, the autonomous car can also react so much faster than you can. It may be able to largely trade-off your level of contextual awareness for pure reaction speed in most cases.
Are you sure? Image recognition or radar does not make self driving cars necessarely faster than a cognitive model and eye input just becouse computers are way faster at math and IO.
You need a certain hysteresis/threshold to filter out false positives.
Its not just plastic bags and fringe error cases. As a driver you know that you sometimes have to break the rules. Say an empty truck is parking on one side of the road, blocking the lane entirely, to overtake him you need to make an illegal maneuver overtaking him or stand there holding up traffic for everyone. Or one step further, you have to make room for an ambulance,. I dont think there is a realistic scenario were all rules to drive in normal traffic can be sufficiently covered.
What you describe feels similar in certain ways to insulin pumps. They’ve definitely helped people live longer and healthier lives but still have cause injuries that otherwise wouldn’t have happened with human administered injections.
Maybe semantics, but the biggest problem isn't necessarily trusting the machine to do what's expected. The problem is the machine being able to predict what pedestrians and other drivers are going to do.
This isn't just a problem for driverless cars, it's become a real problem for people, too.
The rules of the road are there for a reason, people!
When you wave someone into traffic, or have your emergency flashers on while still moving in the rain, or wait too long at a busy 4-way stop, you're not being courteous or defensive or cautious, you're being unpredictable! (And most likely inefficient, as well.)
Worse, (and this is going to be driverless cars' Achilles heel, IMHO) is people pulling into or across traffic (or otherwise failing to yield) based only on the expectation that the vehicle with the right-of-way will brake and do whatever is required to avoid a wreck.
If they do it in front of a driverless car while trying to cross a busy highway, then people making this assumption are going to get killed.
Or it might go the other way and driverless cars will be push-overs on the road and brake for everyone all the time. There’s a lot of decisions to be made about how they should react to that kind of stuff.
Tesla’s last automation talk was interesting about that - they’re basically training the cars to act like a human driver would, looking for tiny signals that might warn of unexpected behaviour.
And it could have both problems. There's no reason I can think of to assume mutual exclusion.
>act like a human driver
Anything AI making this or a similar claim is automatically suspect to me.
>looking for tiny signals that might warn of unexpected behaviour
I suppose this is progress, except that it sort of implies that the AI wasn't already doing that. Which to me says the programmers are still having to anticipate too many possible scenarios.
I agree. The humans are the bugs thing is an excuse that the pre-programmed proponents keep leaning on. That hypothetical has the unintentional implicit frame; as if it's the non pre-programmed humans that are the problem, when clearly a "self driving" car should be able to handle that mundane situation (although it never will). Adding rules to abstract away the humans is the natural next step that their bad assumptions lead to.
I actually think driverless cars will do better here. They will likely be able to recognize cars deviating from their current path in O(hundreds of milliseconds) and avoid a collision, whereas if a car from the opposite direction lane starts drifting into mine seconds before we hit there's little I can do.
I mentioned this elsewhere, but I'll elaborate- I don't think the detection or reaction time is the problem, it's the concept of useful recognition (and prediction). Is that first movement/deviation (that an AI car can certainly measure sooner than I can see) going to be a swerve within their lane or is the car going to come all the way over? It's not the watching for the trigger, but the intuition of what the trigger is, well, triggering.
Humans have an uncanny ability in predicting the overall movement before any perceivable deviation. Of course I'm not saying it's magic; there's obviously some kind of signal to our subconscious that we're able to process and make useful. Perhaps AI can pick up on some of that.
But it's not necessarily the "first movement" of the swerve that is most useful. And the "signal" that humans pick up on may not even be something externally measurable by the car. Like knowing of a particular interstate exit that drivers always seem to almost miss, but swerve across lanes to catch at the last second.
There are also signals that are externally measurable, but have to be learned (seen on a previous trip, or warned by a friend). Like knowing about a bad pothole ahead, and knowing that the car in front of you is going to turn the corner, slam on the brakes and then swerve to avoid it.
These are the kinds of things that I suspect AI will struggle to keep up with in the foreseeable future. And humans rely on this type of intuition way more than I think most people realize.
Is this the 'disconnected lever' problem of AI, applied to vehicle driving?
Of course I'm not saying it's magic; there's obviously some kind of signal to our subconscious that we're able to process and make useful.
That signal is connected to some much larger dataset of years of human experience - things like what a distracted driver might do, what a tired driver might do, what an old unmaintained vehicle looks like, what hints at an out-of-state or out-of-country driver who might not be aware of local routes or problems, how people tend to behave differently at dawn/rush hour/day/night.
An AI car should be able to tell which is a busy junction pretty easily, but watching a car swerve in a way that might indicate the driver is not paying full attention will make you more cautious around that car for a long time up ahead.
Still, AI cars which don't tailgate, will get an ENORMOUS safety advantage over the people who drive 2ft behind me at highway speeds. That alone might be enough to overcome every other benefit. "AI car doesn't get involved in multi-car pileup", "AI car didn't get shunted at traffic lights", "AI car let me in at that junction" will keep happening and keep boosting people's confidence in them little by little.
This idea, but also extended to include the fact that as AI drives more, humans drive less, and might lose some of that knowledge. If AI is driving 80% if the time, but can't handle 20% of situations, then humans are doing the 20% most difficult driving with 80% less practice. Obviously this is over simplifying things, as lack of fatigue and other things will come into play. But I'm merely stating that both sides of the equation will likely change, not just one.
>Still, AI cars which don't tailgate...ENORMOUS safety advantage
As long as they do this in a way that doesn't create congestion, I agree. But I'm not yet convinced. Overly conservative driving can be just as, if not more dangerous than some aggressive driving.
A lot of the risk might be merely transferred, not reduced. Preventable congestion is the bigger problem, tailgating is usually just a symptom. (Not saying it's ok, just that it's preventable in other, more indirect ways.)
And many solutions to congestion require complicated risk-now vs risk-later evaluations (including some that you listed- id'ing a bad driver or a tire about to blow out). These require more decisive, if not aggressive, driver actions, like speeding for 20 seconds to get away from the bad driver or to clear the passing lane. Or accelerating aggressively on an on-ramp (and not have to stop at the end of it to wait for an opening, which is part of why on-ramp metering has become so popular).
Or if you're on the interstate and it's posted for 45mph, but everyone else is driving 80mph. It's probably safer for you to speed (at least 60+) than to drive 45mph.
That's not really compatible with AI's current approach of being overly conservative with everything.
All this to say, I don't doubt AI will make some aspects of driving safer, but I still have doubts about whether it will improve overall safety, at least in the foreseeable future. Driving and traffic systems are complex.
> The problem is the machine being able to predict what pedestrians and other drivers are going to do.
That's something machine learning can do pretty well, actually. There has been a lot of research on this topic, and not applied just to self driving - natural language generation, stock prediction, playing games with multiple agents, even content recommendation do future prediction based on past information.
>natural language generation, stock prediction, playing games with multiple agents, even content recommendation
I'm not familiar with the games, but from what I've seen with all the other things listed, that's not a technology that I'm anywhere near ready to trust my life with.
Because the overwhelming majority of traffic deaths are due to human errors (A) a computer could easily avoid.
It would of course introduce a class of errors (B) that humans would not have made, but those can be iteratively improved until they're substantially lower than the category they're replacing.
My point is that once B is substantially lower than A, the tech should be deployed. But I fear people will demand B = 0.
It's the presumption that this ^ is a given; that's what I have a problem with.
I think people assume B will continue to decrease in a linear fashion, (or maybe log/exp, but at least not hitting it's asymptote / limit until it's less than A).
But we're already seeing diminishing returns, and since A and B are both difficult to measure, I think presuming that the good will automatically outweigh the bad (at least in the foreseeable future) is nothing more than blind hope at this point.
There's also no reason to believe that it won't become more dangerous at some point.
Look at air travel. Automation made great improvements in safety without many downsides. Until maybe about 10 years ago, when automation related accidents started creeping up (yet usually blamed on pilot error). And then you get to the 737-max problem.
(I'm not claiming air travel is more dangerous, it's not. But mechanical reliability and materials testing / repair are doing far more for safety than automation is. Modern jet engines are an underappreciated marvel of science.)
Personally, I would accept a greater risk for the convenience of being driven by AI, and I don't think I'm alone.
I just think it's a bit of blind hope (and even some arrogance) to assume AI will be safer anytime soon.
When true autonomous cars are truly ready, you’re not going to have a choice. It’s going to quickly become prohibitively expensive to insure a human driver.
>> Driverless cara will probably be safer statistically but they’ll simultaneously make errors a human wouldn’t, so some people are going to die in what will seem like really dumb ways.
Statistically it would take tens or hundreds of years before we can know whether self-driving cars are safer than human-driven cars:
Key Findings
∙ Autonomous vehicles would have to be driven hundreds of millions of miles
and sometimes hundreds of billions of miles to demonstrate their reliability
in terms of fatalities and injuries.
∙ Under even aggressive testing assumptions, existing fleets would take tens
and sometimes hundreds of years to drive these miles — an impossible
proposition if the aim is to demonstrate their performance prior to
releasing them on the roads for consumer use.
∙ Therefore, at least for fatalities and injuries, test-driving alone cannot
provide sufficient evidence for demonstrating autonomous vehicle safety.
From:
Driving to Safety
How Many Miles of Driving Would It Take to Demonstrate Autonomous Vehicle Reliability?
People already die in VAST numbers on the roads. This is all a question of which deaths are reported/discussed. If some technology leads to death, it will undoubtedly get way more media coverage than if human error leads to death.
~50 people died when a compounding pharmacy had mold in its spinal injections. ~300 died in SK when a ferry sank. Over a hundred people died due to Boeing's software issues. Hell, look at Uber killing that single pedestrian in Arizona -- how many traffic fatalities occur during Uber rides overall? This is all since 2010 with "advanced" technologies. With scale of tech, comes fringe/rare events, which will be reported on.
The important question is: will less people die with Waymo cars than without them? All logic and reason point towards "less people will die," but evidence is needed.
The one thing we can be sure of: when people DO die with SDCs, it will be HEAVILY reported on, even though thousands die daily on the roads due to preventable, human error, which is rarely reported on.
I want to see the terms of service agreement. Does anyone know if you can sue for wrongful injury or death? Or do you waive your rights away to Google?
I'm still a little amazed that the AI hype train is still running. That's been the fusion power of CS for almost as long as fusion power has been 30 years out.
This afternoon I was driving up Alma in Palo Alto, near the Embarcadero underpass. There was a Waymo test van behind me, and another two cars in front of me. My thought was: "look at me, actually driving my car, like a sucker".
If you then proceeded to Castro and made a right, you can see the Waymos misread that intersection every damn time.
I've also seen Waymos make panicked Lane changes near the JCC many times, willing to come to a dead stop in the right lane and wait indefinitely for a break, even though it's an easy go-down-another-block-and-Uturn
Just a reminder that they're on their way.
We have no idea when they'll get here, but they just wanted to say that they haven't forgotten that we were promised driverless cars.
Just to remind us in case we forgot. Just wanted to make sure that we don't forget about waymo!
Nowadays I hope it takes a very long time. As Elon Musk pointed out, once cars can independently generate income as autonomous taxis, there's no reason to sell them to the public at all. Self-driving cars isn't the death of personal driving because they'll be too convenient; it's the death of personal driving because we won't be able to buy a car at a reasonable margin over manufacturing costs.
Why would companies not sell autonomous cars if customers exist who are willing to pay a reasonable margin over manufacturing costs? The fact that autonomous taxis are viable doesn’t preclude personal car ownership.
Because they would make far more money adding the car to their own fleet. Why sell a car for $40,000 when it could generate $300,000 in revenue over a decade? Tesla has already said they'll stop selling to the public after they have autonomous driving completely safe and working.
> ...when it could generate $300,000 in revenue over a decade?
If markets are competitive the profits will be lower. If markets are not competitive (not many companies can make or license self-driving AI) the rest will sell cars with steering wheels.
The variable that is actually important is demand. If lots of people want their own car, or want to drive their car, the market will provide those things. If most people are content to rent, ownership might become an unusual luxury.
Because more than 1 company exists on the planet. If you won’t sell a car for $40,000 because you’re afraid of cannibalizing your taxi business then someone else will.
If for some reason charging technology is unable to be improved to the point that it's comparable in time to filling up, and if for some reason the solution is to go back to the idea of battery swaps. I could see this being reason enough to keep the whole system as rentals.
If I'm an owner, I'm going to think twice about swapping out a part that might be 50% of my cars value.
There would need to be some serious quality assurances to make that work.
This is cool. One thing that always bothers me me with these announcements, you never hear of them testing in the rain or fog in London or Boston in February or string winds in Kabaaa.
I wonder when rental car buses at airports will adopt autonomous driving. On the surface it might seem ideal because it is repeatedly the same route, but the density, diversity and loose rules of vehicular and pedestrian traffic would make it a very challenging environment for machine learning, and I speculate whether or not that might accelerate building the "dense urban environment" corpus.
Driverless, but they didn't say autonomous (yet). I think this means the safety driver will be remote, teleoperating. It is the obvious next step, and if things go well they can start lowering the ratio of operator:car in the way that drone operations already do.
Will they be the current cars, or cars with the new next-gen hardware that are substantially safer? The cars that end up driving could signal a lot to the industry.
If Waymo starts driving without safety drivers en-masse, even on small fixed routes, it will signal certain takeover metrics have been hit. It will be interesting to see if Waymo uses existing cars (the ML and planners just got better) or if they use new hardware and signal a certain lidar improvement is necessary for achieving a viable takeover rate.
The signal won’t be clear but should be there. And then it’s up to the rest of the industry to show how fast they can react to the change.
I know this is a big task, but could someone give/link me an overview of the state of self-driving cars and the issues they currently face? I know its a hard problem but very bright minds have been doing this for a while, so I'd like to know what is the issue, since we've been told for a very long time that the technology is right on the horizon.
(throwaway account to protect friends at Waymo)
data points of 2:
1. took a ride, at one point a sudden violent sewering and breaking for no apparent reason - explanation from the backup driver - the truck in the next lane was too close. Then it couldn't take a left turn into the parking lot in front of Waymo HQ and was just stuck even though the maneuver was simplest to anyone. remote driver couldn't do it anything and backup driver had to do the 'normal driving'.
2: spoke to someone (again a waymo employee and her friend) just after the ride and the friend was visibly shaken and verbatim expression "worst ride of my life".
This is all around the MV campus and in the last couple of months. So Waymo's claim seems way off unless they are talking about absolutely fixed paths, much like a public transit and maybe, just even then, a big maybe.
220 comments
[ 2.5 ms ] story [ 250 ms ] threadWe have lots of errors and junk DNA. We make all kinds of mistakes. Perhaps not fatal ones yet, we'll see how climate change plays out.
As for our perception and reflexes... I wouldn't brag that they're that good just because they haven't been bad enough to cause our extinction.
That's my point. The natural selection process for self-driving automobiles is just beginning, as they meet the challenge of the real world. I'm concerned that it is going to mirror the messiness that biological evolution involves.
But computers can run 10,000 simulations to learn in minutes. And obviously we are not going to create a autonomous driving ability in minutes but we can speed up the learning evolution significantly in machine world vs humans that have 10-50 years per evolutionary learning event.
My view is the hard thing about autonomous driving will be the change in style of accidents. Even if autonomous driving has half of human accidents we will probably see things that to the human eye seem really obvious so people will be 'if it can't predict that why should I trust it' type thing. But we won't see the many accidents a computer avoids that a human would make or finds obvious.
I'm positive about it but suspect we will see the roll out take many decades as it starts in cities with the right conditions, moves to trucking routes but will probably take some time for national metro coverage.
Point stands. I computer will do millions of simulations in a day. And be done across a population far greater than historic evolution.
Doesn't take long to get billions of learning events. On top of which evolution pre-intelligence (as a very loose term) was somewhat random, where our systems are starting from a much more evolved position that does take advantage of existing evolution as much as humans do.
There's a reason they're testing in Phoenix. It has almost no Inclement weather. Throw a little snow, mud, or sleet in there and it gets hairy fast. Level 2 is basically the 20% of effort in 80/20. It's exponential effort from then on out.
When autonomous drivers have the same or fewer injuries per billion miles driven as human drivers, then for all practical purposes it doesn't matter if it's level 4 or level 5.
If so why go through all this trouble to creat autonomous vehicles? I, for one, am glad this was not the prevailing attitude when commercial airliner regulatory bodies were set up.
[0] https://www.asirt.org/safe-travel/road-safety-facts/
1. They're cheaper. This sounds crude, until you realize that this makes things like buses and shared rides far more viable. Meaning millions of people have a better commute and collectively save many lifetimes worth of time. It can also mean less air pollution, due to less traffic congestion (this only applies if congestion pricing is done), which will also save lives.
2. They can always get better over time. Humans can't.
3. Fewer parking lots are needed, which makes cities far more pleasant to live in.
4. Disabled, elderly people and children have better mobility, and aren't dependent on others to drive them everywhere.
IMO autonomous vehicles are just going to further push public transit into something just for the poor. All the convenience and comfort and privacy of sitting in your own vehicle, but now you don't have to drive yourself! Why would you choose anything else?
So, yes savings. Nyah on your "unions defending humans against exploitation are the worst thing ever" narrative.
Also, if you don't have to accommodate a human driver and controls, many safety features, bulky gasoline tank, and trunk space, automobile design could change radically. Self-driving taxicabs that look like small pods, or covered golf carts, could be used for individual travel, increasing space efficiency.
So we can sleep in the car. Or watch netflix. Or drink. Saving 1+ million per year while you know, very nice, is not the main reason for doing this.
I think the future of self driving vehicles will just be inclement weather delays like we have with flying.
One of the lessons of engineering history that self-driving car proponents fail to grasp is that promising technology often fails to get traction because it can't close the final 5%. Supersonic aircraft, for example, have effectively been derailed by people on the ground near airports complaining about sonic booms. VTOL aircraft never replaced aircraft because of control problems. Flying wing aircraft never became popular for similar reasons.
Also, the prospect of being accidentally killed in a horrible car crash is less poignant if the alternative is more time spent in Chandler.
https://www.theverge.com/2017/11/7/16615290/waymo-self-drivi...
Hopefully this time it's the real deal. I remain skeptical.
This announcement is just a repeat of what they've said before. If they actually do it this time then that's great, but they haven't actually said anything new here.
Sorry you felt that was disingenuous! You're right, it's not public in that anyone can take a ride, your account has to be approved.
This just reinforces my point. Waymo can announce anything with vague terminology of things happening "soon" but we still don't have any verification of anything happening.
Note that while that risk remains true of every transportation journey you take, there is a lot of press generated when an autonomous vehicle takes a life or causes injury, which plays a part in a lot of peoples decision making.
Also note recent articles about automatic stop detection failing >50% of the time. This may or may not be relevant to AVs, but I personally would definitely not want to be in an AV that hit a child. That (theoretical) child's life is worth more than pushing driverless cars into the world before they're potentially ready.
Waymo is still Waymo whether they're running rides on their own behalf, or under the Lyft name. Someone who's ridden Waymo in an early adopter program may make their decision on whether to ride Waymo under Lyft based on that experience.
Get it?
FTA:
> "are on the way"
> "soon you may experience"
Vague promises without any concrete dates attached make this seem like output from PR rather than engineering
> since mid-October, the company has been operating its autonomous minivans on public roads in Arizona without a safety driver
The difference is that those were without any users inside. This new article is saying that the public offering, with real customers inside, will get rides with no one inside.
It is unfortunate that it took 2 years to get to the point where they felt safe putting people customers in the car, but it's very misleading to equate the two and use is as a way to imply they lied last time around.
Except they're not at that point yet, and even when they are they will still have a Waymo employee on board "initially".
I don't think that anybody is accusing anybody of lying, just making the point that predictions of the marketing team aren't really something you would want to set your watch by.
I think you're confused. They've been having customer rides with employee in the car for almost a year now. The email sent to customers today is saying that they will soon find rides with no employee inside.
And again, rides without employee inside have been happening also since 2017, just not with customers inside.
The point is that "soon" is meaningless in this context and Waymo very clearly states that "a Waymo employee would likely be present in the vehicle initially."
1. Waymo has been having rides with no one in the car for 2 years now.
2. Waymo has been having rides with real customers inside for 1 year now.
3. Waymo has told said customers in #2 that they may soon see rides with no employee inside
I agree that the soon in #3 doesn't mean much, but at the same time, this wasn't a press call, it was a private email sent to users, so it's not like they're showing off.
But 1 and 2 have already been happening for a long time.
Of course. Companies tell their customers about new products in the hope that they will keep that info secret.
I think the salient point about your list is that while 1 and 2 have been done separately for some time, the only timeline we have for them happening concurrently is "soon"
https://www.wired.com/story/designated-driver-teleoperations...
> But those cars won’t be operating completely unmanned — at least for now. Under these regulations, driverless cars being tested on public roads must have a remote operator monitoring the car, ready to take over as needed. That remote operator — who will be overseeing the car from a location outside of the car — must also be able to communicate with law enforcement as well as the passengers in the event of an accident.
That being said, the way Waymo does it is different from other companies. As HN can understand, adding the ability to remote access a car is very very risky. Waymo on the other hand has the ability to suggest a path to the car, and the car will itself locally evaluate to see if the suggested path is safe before executing it. So it's not quite a "remote control" access.
[0] https://www.vox.com/2018/2/26/17053898/driverless-cars-self-...
If the goal is to have the car be able to ask for help when it's confused though, it would likely be fine.
Your belief that the GP thinks they lied may be rooted in your own biases than in her or his words.
I don't think you're being very accurate there. There are lots of videos of people commuting 50+ miles to and from work every day without touching anything while on the interstates, including interchanges. It's doing a lot more than just lane keeping.
Perhaps you should check more frequently? Here is a video of a Tesla on Autopilot stopping for a stop sign:
https://electrek.co/2019/05/29/tesla-autopilot-stop-sign-mak...
> “The car always stops at the end of a highway exit since Navigate on Autopilot. At the stop, you just need to touch the accelerator to start Autopilot again without Navigate on Autopilot. I didn’t accelerate, you just need to touch the accelerator for less than a second and the car will do the rest at the intersection.”
[1] https://www.teslarati.com/tesla-autopilot-fsd-traffic-light-...
Which cars have you drive with a superior cruise control implementation?
https://www.youtube.com/watch?v=5z8v9he74po
I feel like people forget human drivers hit those same barriers with alarming frequency.
That being said, Tesla's cruise control capabilities are best in class, but it's misleading marketing-speak to call it true self-driving.
What I mean is, you are not providing input to the wheel, or pedals, or turn signal or driving the car in any way. You're just letting it know you have not fallen asleep.
And, yes, thanks for asking, I have used it.
The problem with the latter approach is the long tail of real life variety doesn't lend itself to a fixed set of programmed rules. You'll always have more edge cases.
I think Waymos approach gets you to 90% faster, but then plateaus.
It is already full self driving in some cities, whereas my Tesla cannot even detect stop signs.
I love my Tesla but I harbor no illusions about the eventual victor of this self-driving race.
https://youtu.be/Ucp0TTmvqOE
1:49:30. I’m on mobile otherwise I would link to the exact spot.
Tesla has tens (hundreds?) of thousands of vehicles on the road right now that are continuously feeding training data back to Tesla. What evidence is there that this is insufficient, or even deficient? From what I can gather, Waymo has limited testing geofenced to Phoenix, Arizona. Tesla has cars all over the world with autopilot, all providing training data.
I don’t see a reason for your conviction that Waymo has an advantage here.
So you have a lot more Teslas sending a lot less data each. I'm not sure which company has the richer training set.
Are they?
- they could store data to be collected in bulk when the car is in the garage for maintenance or servicing or repair.
- Superchargers could include fast data transfer in their cable for cars using them.
- Superchargers could include a wireless data collection from cars sopped at or near them.
- Cars could connect to public wifi to upload data at any car park, any place, any time.
- Tesla could have vehicles driving around, like Google mapping cars, collecting data wirelessly from nearby Teslas.
I don't know if they do any of them, but none are unreasonable things for companies of 2019 to think about doing.
I honestly don't expect go-anywhere L5 in the next decade at least, and I definitely expect an engineered approach to hit L4 first.
I'm not sure how long it will take, but I'm sure we'll get there.
I suspect that isn't true in so far that engineers will build what works. Building a complicated system like an autonomous car will necessarily involve a messy combination of a bunch of different strategies and with the result converging on a common solution.
Honestly, if your only critiques are based on hearsay then maybe you shouldn't critique.
I’m going to ignore the fact that you’re flatly incorrect about the rest of your posts regarding Tesla, since others have more ably corrected you than I could, and simply ask what the hell this even means.
A “production revenue service”? This sounds like meaningless babble. Waymo does not currently have anything available that you can buy, lease, subscribe to, or otherwise exchange money for. That is, in fact, the entire point of the article: “coming soon,” they say, with an additional clause on top of even that: “coming soon you may...”
So: what do you mean when you say “Waymo is in production revenue service”? Because that sounds like nonsense, to be honest.
Haven't seen a recent study/article around cost/ROI (not sure if these have gone down significantly) -- any recommendations?
2) Check Alphabets/Googles bank balance. They have a few bucks they can invest for now.
1) Say what but don't say when.
2) Say when but don't say what.
Never, ever, ever bet tempted to say something specific will happen in a bounded time range.
Flying cars are on the way! That disease you hate, a cure is coming! Next year will be a big year for other breakthroughs not mentioned here.
If you can't keep promises, and don't want to bullshit, then talk more about what the product already does.
OR
I can have many small sub-features that relate to this major feature done in 2 months, but I might know exactly which features make the cut in those 2 months.
Seems congruous to me
I agree and it's such a golden rule that no one follows or cares about even the slightest. Features and deadlines are to SE like velocities and positions are to QM.
This idea more thoroughly described here: https://news.ycombinator.com/item?id=21212087
But then again he probably uses deadlines as both carrots and whips.
Also seems makes good PR and gets people excited, even when exaggeration is accounted for.
The other side of me is absolutely not ready to trust my life to a machine - even though I know I do that all the time in other ways in modern life.
I think it will be people's emotional evaluation that will matter more than anything else and it will be a bumpy road to acceptance.
Driverless cara will probably be safer statistically but they’ll simultaneously make errors a human wouldn’t, so some people are going to die in what will seem like really dumb ways. This will be hard to accept even though that already happens with human drivers. Probably because their mistakes will be easier to identify with and explain.
Like, I'm cruising along at 45mph and a plastic bag is blown in front of my car. Yes, I see an obstacle and need to make a decision about what to do - but I do this as a human who knows what a plastic bag is, and how it is far safer to drive right over it than slam on the brakes. I know a plastic bag is not, say, a Pomeranian.
Another example: I'm driving, and from beyond my plane of vision, a basketball bounces onto the road up ahead. I can't see past a building, but I know I'm near a park where kids frequently play, and I know it's possible a kid comes running into the road to grab the ball. I proactively slow down. I guess you could say a leavel-5 autonomous car would see the ball and slow down anyway, but there's still that lack of cognition that concerns me.
We like to think of driving as a purely rule-based game that is simple enough to model and train against. And I believe in 99% of situations, it is. But in that 1%... the plastic bags and all that...
But at the same time that AI driver will have stopped in time when 10 other kids jumped out without any warning because it has much, much better reflexes.
But it’ll be the first instance that will get all the airtime.
Think about that.
Self driving cars aren't just a solution to nuisance of driving. They're a solution to one of the leading causes of DEATH and INJURY in the country.
Sure there are going to be edge cases where we as humans will be able to point at the car and say "damnit, we could have prevented that"!
But I sure hope when we do that, the car points right back at us and asks "And for the other 1.20 million lives I've saved this year? You would have killed them"
This is a meaningless assertion unless you account for how often "edge cases" arise. If unpredictable pedestrians, unplanned road construction, etc., is an "edge case" the car basically won't work in DC or New York.
Sleepy truck drivers are a major safety issue by themselves, and highways are easier for SDCs to drive on.
no, they will, at some point in the future, be a solution. until we seen serious miles in comparable conditions they are an unproven technology, that why I appreciate waymo approach of getting from here to there slowly and carefully compared to Tesla which has a dozen death already under the autopilot and hides them under a pile of pr and terms of services.
but for all the tech they throw at it until the fundamental issues of cognition and prediction are solved they will only be able to reduce a fraction of the accidents because they aren't replacing the most dangerous vehicles on the road and they only are better for a subset of cases, like inattentive or impaired/intoxicated driver.
Sure, humans drive a lot. The idea is that self driving cars will also, drive a lot.
Nobody is saying humans are terrible drivers (well, some maybe are). But if there's something better...
If reducing DEATH and INJURY are the motivation for self-driving cars, we can achieve a larger reduction with technology that is available right now: public transport. And there is even a no-technology solution: a drastic reduction of road traffic that passes through residential areas.
Putting more cars on the road might or might not reduce DEATH and INJURY, given that those cars are going to be self-driving. But _removing_ cars from roads will _certainly_ reduce DEATH and INJURY- and also reduce the environmental damage from the cars' internal combustion engines and the infrastructure needed for those cars to circulate.
Considering all of the above, I'm really not convinced that the self-driving car industry is really that interested in reducing DEATH and INJURY. I think their main motive is to increase PROFIT.
That said, the benefits of self driving cars aren't precluded by the motives for developing them.
While I hear your points about removing cars from the road, and generally agree with them, there are definite benefits to cars which are arguably not replaced by things like public transit. There will always be a need for cars, so lets make them as safe as possible. If self driving is a step in that direction, sign me up.
If AI can't recognize some obstacles at all, then I think it's safe to assume there may be some delay in obstacle recognition compared to humans.
1) Everyone assumes the good will outweigh the bad. (And why proof of humans doing one task better than AI is somehow also taken as self-evident proof that the AI will be better at everything else.)
2) No one assumes that we, as humans, will ever be able to improve our driving skills in way that can keep up with AI advances. Someone could invent a magic driving pill. Which is crazy. I just don't assume that it's any crazier than a near-perfect driving AI.
An AI driver would have 360 degree coverage with much faster reaction time - think about the cases where a car brakes before the passenger even realises there is danger ahead.
> Using a child-sized dummy, vehicles only avoid running over the child 11% of the time
They're not safe.
Not the same thing at all.
I don't think that bodes well at all for full autonomy - especially since at least one of those vendors (Tesla) is promising self-driving sooner than later.
I would argue that the examples you gave are just image recognition, etc. What is your brain doing if not image recognition and obstacle avoidance?
We're not there with today's technology but eventually our machines will perform better at those uniquely human tasks you mentioned. After all, we humans and our cognition is just the output of these giant meat computers inside our skulls pulling in sensory data and running pattern recognition algorithms.
The $65,000 question is whether or not today's technology is sufficient (however barbaric) to be acceptable for the job.
Planning, anticipation, reading body language, etc? Is that person looking like he’s going to jaywalk? That’s not image recognition. Yes, once he steps into the road the self driving car will respond quicker. But the human driver might already be slowing down and watching for hints of initiating movement. Image recognition is just one tiny piece of it.
How do you do it? Telepathically? You're still just receiving visual input and processing it.
Knowing there is a location ahead without a crosswalk that kids constantly cross at.
Or knowing that the car in front of you is going too fast, implying that they're unaware of the railroad tracks they're about to jump.
You don't need telepathy to make useful predictions. And humans don't need to be accurate, they just need to be prepared to take a measured response.
That's not based on reasoning. Most people driving by probably don't even know. And it might be in the car's database.
> Or knowing that the car in front of you is going too fast, implying that they're unaware of the railroad tracks they're about to jump.
If I'm behind them, and I'm going much slower because I know the layout of the road... I'm not sure how knowing this matters?
Tesla track where every Tesla goes, all the time. Will Apple make a car before Tesla makes a 'smartphone'?
There are still a nearly endless number of things the car can't get from on-board sensors, but that people will know, or at least be able to learn.
For children you might preemptively slow down, but a properly designed system can brake a whole lot faster. Should be fine.
1) If I'm considering whether to brake for a plastic bag, I'm also going to check for anyone tailgating me, and take that into consideration. Including whether said tailgater is a geo prism or a Freightliner attached to a tanker of gasoline.
Slamming on the brakes for a plastic bag is not the universal safer choice.
2) Already mentioned this, but reaction/breaking time is meaningless without timely recognition.
...I didn't say it was. And you call me the one making assumptions.
> reaction/breaking time is meaningless without timely recognition.
If it can't quickly recognize large moving objects then it's not qualified to drive at all. I'm assuming it can do that because self-driving at all is part of the premise.
(Well, specifically the recognition is necessary for basic driving functionality, and being faster is a matter of adding compute power and using cameras at least as fast as the ones in an iphone.)
2) >>>you might preemptively slow down, but a properly designed system can brake a whole lot faster.
>If it can't quickly recognize large moving objects
I'm not questioning whether it can recognize quickly. Just whether it can ever be better/ faster than humans, taking into account cognitive ability.
It's whether human
(prediction-or-recognition time + reaction time + breaking time)
is more or less than
(AI recognition time + plus reaction time + breaking time)
If AI can react and brake 2 seconds faster than a human, but doesn't recognize the threat until 4 seconds after humans can, then AI drivers will kill more people than human drivers. Which won't be fine.
>being faster is a matter of adding compute power and using cameras
There are other comments about the limitations of relying only on image recognition. While compute power per might improve the AI recognition time, humans still have the advantage of being able to predict instead of react. This will continue to be a major limitation for AI.
Okay, but I didn't say that either. I said it would be an improvement to the experience I have. I will trade an abrupt stop in front of a plastic bag for several hours of being able to do what I want while the car does the driving.
> If AI can react and brake 2 seconds faster than a human, but doesn't recognize the threat until 4 seconds after humans can, then AI drivers will kill more people than human drivers. Which won't be fine.
The AI processing time to recognize things in a frame needs to match the framerate, or the system is just broken. Maybe it waits for 3 frames to show the same thing before acting, that's fine. That's still faster than a human. Especially if you use a 240Hz camera, which is not very expensive. And you can and should put in enough computing power to run the 'reacting' system at the same frequency.
Then once it decides to brake, instead of waiting for a foot to move it can put the virtual brake pedal at any position in under a millisecond.
I can imagine a car so poorly designed that its reactions are slower than a human. But it's hard to imagine a car that's simultaneously certified for letting me take my eyes off the road, and is that broken.
So I still say that a self-driving car would be able to brake quicker and harder, and it should be fine. Even without the ability to predict.
One word: Chauffeur
>The AI processing time to recognize things in a frame needs to match the framerate
If you're talking about framerates at all, you're still missing the point. I'm talking about human ability to predict and anticipate. The fact that a human can brake before there's anything in the frame to even detect. We've (humans) got a head start, so the speed of the AI isn't all that important.
But for people with normal incomes too!
> brake before there's anything in the frame to even detect
The scenario was slowing down a bit, but still needing to hit the brake in response to the child. So if the AI can brake faster it is extremely important.
Up to a point - ordinary seatbelts even with smart pre-tensioners aren't great at this, and it may increase the rate of whiplash injuries for the passengers who aren't expecting braking.
We really should do this like the Dutch: when something problematic is encountered, engineer and regulate the hell out of it. Not just for self driving cars, but it works for humans to some extent that the laws are actually enforced.
Waymo cars have driven over 10 million miles so far without any fatal accidents. That's not nearly enough miles to compare against human drivers, but it'll be interesting to see who does better once they log some more miles.
You need a certain hysteresis/threshold to filter out false positives.
https://www.cbc.ca/amp/1.4915491
This isn't just a problem for driverless cars, it's become a real problem for people, too.
The rules of the road are there for a reason, people!
When you wave someone into traffic, or have your emergency flashers on while still moving in the rain, or wait too long at a busy 4-way stop, you're not being courteous or defensive or cautious, you're being unpredictable! (And most likely inefficient, as well.)
Worse, (and this is going to be driverless cars' Achilles heel, IMHO) is people pulling into or across traffic (or otherwise failing to yield) based only on the expectation that the vehicle with the right-of-way will brake and do whatever is required to avoid a wreck.
If they do it in front of a driverless car while trying to cross a busy highway, then people making this assumption are going to get killed.
Tesla’s last automation talk was interesting about that - they’re basically training the cars to act like a human driver would, looking for tiny signals that might warn of unexpected behaviour.
And it could have both problems. There's no reason I can think of to assume mutual exclusion.
>act like a human driver
Anything AI making this or a similar claim is automatically suspect to me.
>looking for tiny signals that might warn of unexpected behaviour
I suppose this is progress, except that it sort of implies that the AI wasn't already doing that. Which to me says the programmers are still having to anticipate too many possible scenarios.
Humans have an uncanny ability in predicting the overall movement before any perceivable deviation. Of course I'm not saying it's magic; there's obviously some kind of signal to our subconscious that we're able to process and make useful. Perhaps AI can pick up on some of that.
But it's not necessarily the "first movement" of the swerve that is most useful. And the "signal" that humans pick up on may not even be something externally measurable by the car. Like knowing of a particular interstate exit that drivers always seem to almost miss, but swerve across lanes to catch at the last second.
There are also signals that are externally measurable, but have to be learned (seen on a previous trip, or warned by a friend). Like knowing about a bad pothole ahead, and knowing that the car in front of you is going to turn the corner, slam on the brakes and then swerve to avoid it.
These are the kinds of things that I suspect AI will struggle to keep up with in the foreseeable future. And humans rely on this type of intuition way more than I think most people realize.
Of course I'm not saying it's magic; there's obviously some kind of signal to our subconscious that we're able to process and make useful.
That signal is connected to some much larger dataset of years of human experience - things like what a distracted driver might do, what a tired driver might do, what an old unmaintained vehicle looks like, what hints at an out-of-state or out-of-country driver who might not be aware of local routes or problems, how people tend to behave differently at dawn/rush hour/day/night.
An AI car should be able to tell which is a busy junction pretty easily, but watching a car swerve in a way that might indicate the driver is not paying full attention will make you more cautious around that car for a long time up ahead.
Still, AI cars which don't tailgate, will get an ENORMOUS safety advantage over the people who drive 2ft behind me at highway speeds. That alone might be enough to overcome every other benefit. "AI car doesn't get involved in multi-car pileup", "AI car didn't get shunted at traffic lights", "AI car let me in at that junction" will keep happening and keep boosting people's confidence in them little by little.
Yes, it's good to know this has a name. Thanks.
This idea, but also extended to include the fact that as AI drives more, humans drive less, and might lose some of that knowledge. If AI is driving 80% if the time, but can't handle 20% of situations, then humans are doing the 20% most difficult driving with 80% less practice. Obviously this is over simplifying things, as lack of fatigue and other things will come into play. But I'm merely stating that both sides of the equation will likely change, not just one.
>Still, AI cars which don't tailgate...ENORMOUS safety advantage
As long as they do this in a way that doesn't create congestion, I agree. But I'm not yet convinced. Overly conservative driving can be just as, if not more dangerous than some aggressive driving.
A lot of the risk might be merely transferred, not reduced. Preventable congestion is the bigger problem, tailgating is usually just a symptom. (Not saying it's ok, just that it's preventable in other, more indirect ways.)
And many solutions to congestion require complicated risk-now vs risk-later evaluations (including some that you listed- id'ing a bad driver or a tire about to blow out). These require more decisive, if not aggressive, driver actions, like speeding for 20 seconds to get away from the bad driver or to clear the passing lane. Or accelerating aggressively on an on-ramp (and not have to stop at the end of it to wait for an opening, which is part of why on-ramp metering has become so popular).
Or if you're on the interstate and it's posted for 45mph, but everyone else is driving 80mph. It's probably safer for you to speed (at least 60+) than to drive 45mph.
That's not really compatible with AI's current approach of being overly conservative with everything.
All this to say, I don't doubt AI will make some aspects of driving safer, but I still have doubts about whether it will improve overall safety, at least in the foreseeable future. Driving and traffic systems are complex.
That's something machine learning can do pretty well, actually. There has been a lot of research on this topic, and not applied just to self driving - natural language generation, stock prediction, playing games with multiple agents, even content recommendation do future prediction based on past information.
I'm not familiar with the games, but from what I've seen with all the other things listed, that's not a technology that I'm anywhere near ready to trust my life with.
It's obviously the best choice to adapt a technology like that, but I'm not sure our society has the fortitude to do it.
Seriously. Why does everyone assume this? Because Tesla's marketing department says so?
It would of course introduce a class of errors (B) that humans would not have made, but those can be iteratively improved until they're substantially lower than the category they're replacing.
My point is that once B is substantially lower than A, the tech should be deployed. But I fear people will demand B = 0.
>once B is substantially lower than A
It's the presumption that this ^ is a given; that's what I have a problem with.
I think people assume B will continue to decrease in a linear fashion, (or maybe log/exp, but at least not hitting it's asymptote / limit until it's less than A).
But we're already seeing diminishing returns, and since A and B are both difficult to measure, I think presuming that the good will automatically outweigh the bad (at least in the foreseeable future) is nothing more than blind hope at this point.
There's also no reason to believe that it won't become more dangerous at some point.
Look at air travel. Automation made great improvements in safety without many downsides. Until maybe about 10 years ago, when automation related accidents started creeping up (yet usually blamed on pilot error). And then you get to the 737-max problem.
(I'm not claiming air travel is more dangerous, it's not. But mechanical reliability and materials testing / repair are doing far more for safety than automation is. Modern jet engines are an underappreciated marvel of science.)
Personally, I would accept a greater risk for the convenience of being driven by AI, and I don't think I'm alone.
I just think it's a bit of blind hope (and even some arrogance) to assume AI will be safer anytime soon.
Statistically it would take tens or hundreds of years before we can know whether self-driving cars are safer than human-driven cars:
From:Driving to Safety
How Many Miles of Driving Would It Take to Demonstrate Autonomous Vehicle Reliability?
https://www.rand.org/pubs/research_reports/RR1478.html
~50 people died when a compounding pharmacy had mold in its spinal injections. ~300 died in SK when a ferry sank. Over a hundred people died due to Boeing's software issues. Hell, look at Uber killing that single pedestrian in Arizona -- how many traffic fatalities occur during Uber rides overall? This is all since 2010 with "advanced" technologies. With scale of tech, comes fringe/rare events, which will be reported on.
The important question is: will less people die with Waymo cars than without them? All logic and reason point towards "less people will die," but evidence is needed.
The one thing we can be sure of: when people DO die with SDCs, it will be HEAVILY reported on, even though thousands die daily on the roads due to preventable, human error, which is rarely reported on.
I've also seen Waymos make panicked Lane changes near the JCC many times, willing to come to a dead stop in the right lane and wait indefinitely for a break, even though it's an easy go-down-another-block-and-Uturn
If markets are competitive the profits will be lower. If markets are not competitive (not many companies can make or license self-driving AI) the rest will sell cars with steering wheels.
The variable that is actually important is demand. If lots of people want their own car, or want to drive their car, the market will provide those things. If most people are content to rent, ownership might become an unusual luxury.
If for some reason charging technology is unable to be improved to the point that it's comparable in time to filling up, and if for some reason the solution is to go back to the idea of battery swaps. I could see this being reason enough to keep the whole system as rentals.
If I'm an owner, I'm going to think twice about swapping out a part that might be 50% of my cars value.
There would need to be some serious quality assurances to make that work.
New lidar was spotted a few months ago and more recently on the Jaguar cars and the CEO said it’s an order of magnitude better: https://thelastdriverlicenseholder.com/2019/04/21/new-waymo-... https://medium.com/waymo/waymo-iaa-frankfurt-2019-b3cca36d84...
One would imagine this early announcement comes in reaction to their recent “valuation” cut: https://www.cnbc.com/2019/09/27/waymo-valuation-cut-40percen...
If Waymo starts driving without safety drivers en-masse, even on small fixed routes, it will signal certain takeover metrics have been hit. It will be interesting to see if Waymo uses existing cars (the ML and planners just got better) or if they use new hardware and signal a certain lidar improvement is necessary for achieving a viable takeover rate.
The signal won’t be clear but should be there. And then it’s up to the rest of the industry to show how fast they can react to the change.