The CEO the article is about wasn't exactly at the frontwards edge of technology, he was an automotive industry expert.
I'm concerned about the numbers, but I'm hopeful about change in leadership. Especially given the ties with Google, possessing an insane, astronomical amount of street-based data. Surely they can scrape together a L2 driving agent given a few years?
Seems like an "L2 car" is what the market wants and can get, though. Elon may not say what you want him to say, but he's selling these things like hotcakes.
Broadly: that's seems like senseless pedantry to me. Tesla owners, even numerate serious ones who understand the problem area, know when their car is "driving itself" and when it isn't. Quoting industry jargon to explain to people how the product isn't akshooally doing what it clearly is isn't going to win many arguments.
A long drive slightly less fatiguing is better than a short drive in a limited area of operation.
This, I think, is why Tesla will eventually succeed. Every Tesla sold is gathering data about its environment and phoning home - we already know that. So whether you have Autopilot or not, Tesla gets data for its Full Self-Driving Capability; when you swerve to avoid a possum at 12:47 a.m. because you've been working all day and night and into the morning and you really need some Whataburger, or, if you must... In-N-Out..., some engineer somewhere at Tesla can look at that event, see why you swerved out of lane, recognize it was a possum, and then jot that down into his or her notebook...
twwooo... forrrtty-seeeven... ayyy emmm... driver swwwerved... to avoid... possum..., integrate possum-avoidance behavior into Autopilot for next release...
That would be valid if Tesla is really collecting all of that data. But the people who have done analysis of what the cars sends back say that what is actually sent is very limited statistics on times, locations, and general characteristics of disengagements (speed, direction, list of detected targets, etc). This is doubtlessly useful but it's not something you can just drop into training a neural network or even be sure exactly what the issue was.
They can write query strings for very specific situations they're interested in, too, and receive video frames.
On the other hand, Waymo is gathering vision data on every drive, and ground truth LIDAR data.
It's a win. But-- if it's just driving freeway miles and lanekeeping and doing basic collision avoidance-- there's a lot of cars doing that and Tesla is in first place but not strongly differentiated from GM, etc. Other cars can provide you this "less fatiguing road trip" experience.
If you want it to actually drive and be able to remove your attention, well, something like this https://www.youtube.com/watch?v=uClWlVCwHsI makes the present limits rather clear.
The fact something that exists is selling better than something that doesn't exist doesn't really prove much :)
Pretty much every car vendor already has L2 autonomy [1] - Google's not going to extract tens of billions of dollars by selling the industry something they already have.
> The fact something that exists is selling better than something that doesn't exist doesn't really prove much :)
On the contrary, it proves that the other thing doesn't exist. Which is sort of important to analysis.
I know that's glib. My point was the pontification about what "self driving autonomy" is or should be or might be or is supposed to be or (my personal peeve) "is supposed to be called" is essentially meaningless when we could be discussing what is actually happening in a vibrant and interesting market.
I get it. People think about this stuff in the abstract. But it's not an abstract problem anymore, and I think arguments like yours are very much missing the forest for the trees.
Roll back the clock to, say, 1985 and let's imagine we're talking about network technology. Elon is selling a working TCP/IP stack while you're yelling about the fact that akshooally it doesn't conform to the proper OSI layers and so can't ever be a successful network. How did history work out for the OSI pedants? Same principle.
Elon is selling a "Full Self Driving car" that is not full self driving at all by Tesla's own admission. It isn't doing the thing it is advertised to do.
Calling an L2 car "full self driving" is incredibly irresponsible. I personally hope Musk faces consequences when yet another Tesla kills someone because of false advertising.
> Calling an L2 car "full self driving" is incredibly irresponsible
I repeat: this is an argument about semantics, further obscured by the use of jargon, and it does nothing to inform a discussion about what autonomy features are being offered on existing cars, nor about what the market wants, nor even about whether or not those features are safe or effective. It's just words.
If you want to have a discussion about whether or not Tesla's self driving feature is safe or not, you need to come with evidence about safety. You can't just yell about "L2", no one cares anymore.
Could you be more pedantic? I'm not even the one who brought up L2 in the first place. I used the same jargon as the person I was responding to.
HN is a technical forum, if you're going to have a discussion about autonomous vehicles on a technical forum then technical terms should be fair game. If I'm having this discussion elsewhere then I'll explain it differently.
But fine, if you want to be pedantic: Calling a car that cannot drive itself without human intervention a "fully self driving car" is incredibly irresponsible.
I don't think it's comparable to Theranos. He took over a company w/a heavily (but still in a pre launch) product in its "beta" phase and was chosen for an obvious reason (get the tech into cars). He doesn't have the technical chops to understand the tech enough to know whether it's ready for prime-time or not aside from looking at metrics some of his engineering subordinates gave him.
Elizabeth Holmes didn't have a product which functioned like she claimed at all (nothing compared to the tech Waymo has already developed)... She assumed w/enough effort her ideas would eventually come to fruition, so she lied about the progress her company had made so gullible investors would fork over $$. Waymo has non-tech people trying to sell a tech which is still in development, they've made an (incorrect) assumption about when it'll be ready, as I doubt anyone really thinks it's flat out impossible... just really difficult.
>as I doubt anyone really thinks it's flat out impossible... just really difficult.
You see a few people saying we need AGI but yeah. I've definitely been a skeptic but mostly in the vein of it needing decades, not single digit years. And the transition to even a true L4 system is going to be hard.
That said, we seen a lot of progress and even over the past ten years, there's been a lot of progress on assistive driving system as well as field work on actual self-driving.
I suspect that machine-assisted driving will continue to get better, albeit with issues associated with humans increasingly not really paying attention. But I expect those wanting robo-Ubers will continue to be disappointed for quite some time. (And I'm not sure how much it will matter because they will have actual Ubers.)
As to your point, I agree with you. I guess the comparison was meant in quite a pointed way that went over the heads of a lot of commenters here.
I was meaning to just draw attention to the fact that juxtaposing the numbers 82000 -> 600ish is quite a fantastical leap.
From what I've read about the older Waymo days, I think the leadership ideas were a lot different than what they became under Krafcik. It's hard to imagine from where I'm sitting, such a stark pivot from inside the company.
I think Krafcik probably meant as well in the claims, they need X # of vehicles and so that implies that there must be X # of agents ready to pilot vehicles or that many vehicles even ready for installation and operation/testing.
Of course they were very much further than Theranos, I guess it's just the sheer magnitude of the claims in Krafcik's place.
All this to say: I don't know anything about cars, self-driving cars, electric cars, rental cars, etc. I do know bullshit when I see it, and Krafcik was selling it like someone with bills to pay.
Makes me somewhat curious what state Levandowski is in. I recall that he was pardoned, and filed bankruptcy after the ruling that he owed Google $179M and probably owed lawyers quite a lot as well. I'm curious if the bankruptcy shed all of the debt, and where he ended up financially after that.
I also wonder how much all the costly talent departures hurt Waymo. Lots of brain drain, each one also taking lots of equity, etc, on the way out.
I would surmise that the talent knew better than anyone how over-hyped and overly optimistic full self-driving was (is), and what huge challenges still remain. Talent take the best deal for themselves while the getting is good. By the time the hype cycle is over, they're on either a tropical island or in some different field.
I’d be curious to hear the full quote and context on the 600 number. It sounds like that is talking about operational taxis, while the 82k number is all vehicles. It seems like that >90% of waymo’s vehicles would not be on the roads but in testing/design/etc. the articles point still stands, but I don’t think the number comparison seems totally fair.
You don't build 82K for prototypes. You build hundreds to a few thousand for prototypes as you're scaling production.
82K is way too many to build when your design isn't finalized. It makes much more sense as a production goal when they hit mass production, or the number they expect to be able to sell.
That's an ML problem. Building out a factory to produce 100K cars is expensive: ask Tesla how much they've spent to hit their production goals, how many engineer/tech-years, and how much in supply chain effort it is.
You just don't build that many prototypes, period, even if you're a trillion dollar company.
You only build that many if they're saleable as real products so you can get market feedback that yes you're going down the right path.
82,000 doesn't sound like much, but for a prototype build with yield rates that are undoubtedly shitty, non-final hardware, and required automotive-spec that is completely nuts. There's a reason why it's dropped to 600 so far and it's because they haven't finalized the hardware and they realize they aren't ready for mass production yet.
Honestly, even 600 prototypes sounds like a lot to me for something as complicated and expensive as a car, but I don't work in that field so I can't say what normal is. It's about right for beyond-the-prototype-and-trying-to-get-to-a-manufacturable-product in consumer or industrial hardware, depending on the target market size.
The justification for why they don't need an auto industry CEO makes sense. But do co-CEOs make any more sense? I never get a good vibe from that arrangement.
If I were one of the "co-CEOs", I'd read that as "Google doesn't trust either one of us enough to do this alone". Co ownership of say, a project might be okay, but co-ownership of a specific leader title is silly to me, especially the top spot.
Around the time Krafcik was hired Google/Alphabet was dealing with scandals and fallout from egomaniacal leaders like Tony Fadell, Andy Rubin, Anthony Levandowski, and that guy they put in charge of Verily.
Larry and Sergei realized they weren't very good judges of character, and so they went in the opposite direction and put the milquetoast John Krafcik in charge. Their hope was that Krafcik would be better at forging auto industry partnerships.
I'm not sure what Waymo would look like today if a tech exec had been in charge, but I doubt the CEO would have had much of an impact one way or the other on the hard problems associated with self-driving.
A brief read over his career seems to show he was pretty successful. I guess you could argue "right place/time", but his time at Hyundai and TrueCar looks like it went well.
Certainly, I wasn't trying to suggest Krafcik was spineless or lacking competence, just low key and not high on coke.
The biggest hazard Waymo faced the past 5 years was some kind of Media witch hunt based on a sensational accident of some sort, and that didn't happen on Krafcik's watch.
His job was to do a political highwire act and manage the interests of his team, Larry and Sergei and Ruth, regulators and politicians across multiple levels of government, partners across several industries, competitors as well as the media and public expectations. Somehow nothing catastrophically shitty happened.
Self driving cars are still a science experiment, and that's not on Krafcik. He killed it. At the front end Waymo is as ready to scale and commercialize as it can be at this stage.
Yes they matter very much sometimes. A CEO like Elon with a very large depth of understanding of this problem space is going to be able to allocate resources much better than someone without any AI experience. There are many many many dead ends in this space and you need to be able to experiment and pivot quickly which is what Tesla is doing. Waymo is stuck in the mud.
The stuff that scares you may be the beginings of the technology thats actually needed to get to full L5. We just don’t know yet so anyone that is certain on the internet I would be skeptical of. The bet Elon is making is on pure vision which is a much different approach than Waymo and Cruise are taking.
It might be. But I find it difficult to believe that having fewer sensors makes things easier.
I also just don't think the Tesla cameras have enough angular resolution to perform a lot of scenarios safely, like detecting a much faster vehicle in a short merge.
(Note Tesla relies upon the front-facing radars quite a bit, too, so it's not really pure vision).
Elon said they’re not going to need the radars going forward. I guess that makes sense if vision gets good enough but they’re going to run into issues like you said if camera placements and poor visibility. They should have stuck cameras facing the left/right from the furthest corners of the car so it can see around corners without inching into the street.
But who knows, it seems like they’ve made good progress as is so well see. I don’t understand the people that think they’re never going to upgrade the camera suite though, cybertruck and the semi already have a new camera layout.
> I don’t understand the people that think they’re never going to upgrade the camera suite though
Well, they've committed an awful lot to current FSD customers that I think will be difficult to deliver upon with the current camera suite.
I think LIDAR is a huge advantage. There's the direct benefit: you really know exactly where things are. But secondarily, you have ground truth near-perfect range data to use in training your vision system.
The next beta is not going to use radar. Fewer sensors is much better for Tesla since they need aerodynamics and less power consumption for more vehicle range. Remember they are making a fully integrated product from the start, not bolt-on technology.
We don’t yet know the limits of the pure vision approach. Elon is very confident in it and I want to see where it goes. Your comments about resolution may or may not matter.
I'll believe it when I see it, no matter what recent statements have been.
> Fewer sensors is much better for Tesla since they need aerodynamics and less power consumption for more vehicle range.
Automotive radars use like 10W. The least your car is using while driving is like 5000W. This may not be the best place to save power-- getting heat pumps on more vehicles and better ones than the present one would make a far bigger difference.
> Your comments about resolution may or may not matter.
If you can't see a car that will hit you when you commit to merging on a short merge, it is a slight problem.
Tesla's cars can drive everywhere (although not always perfectly), even in unfamiliar situations, while Waymo can drive only on a very small geofenced area with restricted routes within the area. Waymo mispresents their performance by having selected only those routes that they can drive in. It's the oldest trick in the book of cheating. Waymo is nowhere near level 4 on all roads.
Tesla would kill you half the time making this unprotected left turn. Waymo can make unprotected left turns and have never killed or injured anyone or even bent metal in this situation. :P It's not a particularly challenging left turn.
And the goal is to kill you less than .0001% of the time making left turns.
Yes, Waymo tries to avoid them in route planning, but it still takes them if necessary.
A) It's further along than anyone else, and B) it's a critical milestone. People are willing to pay for the experience and the company is able to provide a service for money without drivers to those not pledged to silence.
Yes, profitability isn't here yet, nor have they recovered development costs to date. But that is not what was asserted. Instead, they are to (limited) market, which is a milestone.
I agree, it's strange to axe the type of person they'd need to rollout the tech, unless the tech itself just isn't going to be ready soon at all vs. giving him a different role? I wonder how long the project of "L4/L5 autonomous vehicles on today's roads" would have to meander on for before the Alphabet leadership decides to axe it or radically change to scope would/will be?
Maybe it's enough of a pet project of the senior leadership to get cut an extraordinary amount of slack. But even given continuing incremental advance it still seems a very long way from a viable product that they could sell.
I strongly suspect we will soon enough find out that Waymo hit a wall in self-driving capability possible with ML methods and got rid of its in-car safety drivers only by moving the actual humans to a central facility where they monitor and assist the vehicles remotely. And that plus the extreme maintenance required to maintain the cars’ sensors is going to keep Waymo and any other such company from being able to roll out an economically viable taxi service at any meaningful scale.
If autonomous driving is approaching a sophistication that is practically useful on a large scale shouldn't we first see it widely deployed for delivery of packages and food and the like before we see it being used to move much more valuable charge like people?
They'd still be on the road with other people and the actual delivery of a package requires quite a bit of "common sense" absent some sort of drop-off zone with a beacon.
I agree but some problems disappear too. For example, occupants of cars get frustrated when they are stuck for quite some time on a complicated turn that a human would accomplish quickly. Packages don't care either if the driving is not smooth and the vehicle responds in favour and with extra care of other vehicles around it. Not to mention routes that might pick up and deliver multiple simultaneous loads. It just seems that would be an easier first step for the tech that should be profitable. But these are just questions I have. I'm very unsure about what the reality is.
As someone else said, that doesn't solve any problems because you still have to avoid crashing into other cars, cyclists and pedestrians.
I think a more likely thing is that we see it for lorries on motorways (sorry, err.. trucks on freeways) where it's a much easier problem, there are no cyclists or pedestrians, you can have them drive in the middle of the night when there's no traffic and there are convenient service stations you can stop at and have human drivers take over.
Like, why pay for a lorry driver to take all of your containers from Felixstowe to London when you can just have automatic lorries deliver them to a depot just outside London?
I haven't heard of anyone actually trying to do this though. Maybe not sexy enough. Not a big enough total addressable market or some such bollocks.
If problem was cost of building and maintaining hardware they would go after that problem by iterating on hardware much faster. That is not the immediate problem. What they need to solve first is to reach to a level of confidence with "happy ride ratio" in which they don't need to hire too many remote assistant drivers to scale up. After all, this technology wants to remove human drivers from the taxi business. What's the point of paying often more expensive remote human drivers plus super expensive gear to do a low margin business like taxi service?
It's a lot worse than that. The reason Waymon is in Arizona is that it has almost no weather. Remote assistance isn't enough to rescue a ride in weather.
Uber bought out almost the entire CS department of CMU (Pittsburgh) yet they run their cars in Arizona.
It's kinda an interesting problem-- the technology readiness re: weather.
Right now a lot of self driving, like Waymo, is aimed at taxi fleets, because the equipment is expensive, so amortizing the cost over many driving hours per day makes fiscal sense.
But-- a taxi fleet that requires a massive scale up / instant appearance of a human workforce when it's rainy is a non-starter.
On the other hand, personal cars would be useful with an "only sunny" autonomous mode.
To a discussion upthread, it sometimes seems as if many of the folks working on full autonomy decided to deliberately turn their difficulty up to 11. What use case shall we tackle first? Highway driving in good weather? "No that's boring. How about largely-urban taxis in all weather types?"
The problem is, it's not clear that camera-only and relatively modest processing can even handle fully autonomous highway driving in good weather. So, if you need a very expensive sensor package, the only large market available is taxis.
>if you need a very expensive sensor package, the only large market available is taxis.
But absolutely the only benefit it brings to a taxi is reduced labor costs. So if the capital cost is so high that you can't sell an $80-$100K self-driving car to moderately wealthy people, are you really going to be able to undercut Uber drivers? (And labor costs don't go to zero anyway as someone needs to clean and maintain the vehicles.)
>camera-only
I'd just point out that even Toyota uses some sort of radar for its dynamic cruise control today.
> But absolutely the only benefit it brings to a taxi is reduced labor costs. So if the capital cost is so high that you can't sell an $80-$100K self-driving car to moderately wealthy people, are you really going to be able to undercut Uber drivers?
Say you can do it for $100k, and it looks wonky. That's a very small car market. But it's potentially a very large autonomous taxi market.
> I'd just point out that even Toyota uses some sort of radar for its dynamic cruise control today.
Yes, I'm sorry, I didn't mean "camera-only". Even low end cars have radars for collision avoidance. I meant "with cameras and front-facing RADAR, but no LIDAR / other expensive sensors".
Though-- cameras need to be good enough to perceive everything radars miss, and once they're reliable enough at that for L4, the incremental value provided by a radar may be limited.
I don't think so. I've watched a lot of Waymo videos and it's pretty clear when the car needs remote assistance help. Mostly parking lot situations.
It's still likely a bit too much, but I think they don't need much further progress.
The bigger problem is that they still have absolutely no confidence in any weather worse than when it's barely sprinkling... then the safety drivers come back out.
There are so many weather scenarios that are obstacles we as humans barely overcome. I think about it like the meme ”Almost at my destination, better turn off my radio so I can see where I’m going”. That is we as humans the best drivers in the universe so far as we know have to limit some sensory intake so we can focus that every elsewhere to make even simple calculations. Asking a vehicle sensor suite to handle white out snow or fog with halos around lights or to ignore its lidar in super heavy rain just seems like the equivalent of blaring the radio when you’re trying to get across 8 lanes of superhighway traffic and it’s your first day in a left hand drive vehicle.
Twenty years ago I handled situations in sensor fusion where some of the sensors go to shit in specific circumstances, and it wasn't blazing new ground. The state of the art and the tools to deal with this have grown massively since then.
I’m an automotive EE, so I encourage you to come fix all the problems. Because imo we’re 5 years out from where people think we are today. Maybe the problem isn’t practical, it’s just we just haven’t dumped enough brain power into it.
I think there's massive problems, but the one you described -- attentional saturation -- is not the way sensing systems behave. That is a human issue. Trying to anthropomorphize the computer solving the problem is not a path to understanding the problem space.
As a 1:1 exact problem, no of course. As an analog to challenges, sure.
There is also the “who is more important” problem? A self-interst version of the trolly problem. When asked this, a Mercedes executive explained that the driver was the most important person to the vehicle. So given the option of hit this biker or swerve into oncoming traffic, they would chose the biker. So, somewhere in software there is a decision to always protect the driver, interesting problem that we have no good answer for. Back to my point though, I’m skeptical we’re even close to having all the data to make the decision properly.
> As a 1:1 exact problem, no of course. As an analog to challenges, sure.
I disagree.
> So given the option of hit this biker or swerve into oncoming traffic, they would chose the biker.
You act like this is really complicated, but given that self-driving cars generally are quite conservative in their dynamics and where they choose to go, this is pretty much the default behavior. Swerving into oncoming traffic is just not in their DNA right now, even in the rare circumstances where it's advantageous. Right now there's some work on aggressive dynamics for difficult situations, but this will tend to be underused not overused (just like it is by humans).
Doing the default, conservative thing is generally the self-protective one.
>Doing the default, conservative thing is generally the self-protective one.
Yes, given that scenario, the "correct" response is almost certainly to lay on your brakes not to get into a head-on collision with another moving vehicle. And while one can construct trolley-type scenarios, braking hard is generally the right choice unless you think you can safely swerve.
Waymo ride service is not fully autonomous. There are remote assistant drivers that take over the situation if the system comes to a halt. I am guessing that the number of times they have to take over the cars remotely vs. number of times rides are fully autonomous in Arizona doesn't make it a good business to scale up. It makes sense to go back and focus on technology instead of scaling up a platform that might be just a loss leader.
Interesting, I didn't know that. I wonder what the real-world lag is for the remote driver, and how much it matters. Sounds dangerous at face value, like the lag could create a chain of events that ends badly. Even a "stop button" could have unintended consequences.
My understanding is that it's not live remote driving. When the system doesn't have enough confidence in an environment, it asks for remote assistance to label objects and manually add enough data so that it can resume route planning.
There is also a cruise automation video showing a drive through SF where this same type of situation comes up. I believe both systems rely on HD maps and the route is planned based on the vehicles position relative to the very precise maps. If a large enough obstacle is in the way of the planned route it will disengage to the safety driver to plan a new route around the obstacle.
I don't know anything about their finances but please consider the car is about $200k with current sensors. Just like Uber and Lyft, the load is not evenly distributed. You need a lot more remote assistant drivers than most people imagine is needed.
Those prices should come down, though. Lidar is fairly exotic at the moment, which contributes to those prices. Once more applications are found, I would expect those to drop. One interesting application is for motion detection. Lidar gives a detailed enough image to determine whether an object is a person or not, paving the way for motion detection systems that only trigger when humans move, as opposed to when anything moves.
The lifetime of the sensors is also important. If the sensors outlive the car, you can amortize a big part of that $200k by just putting old sensors on a new car. If the sensors are shorter lived than the car, the total cost of the car over its lifetime will be much higher.
Level 4 autonomy is still a far off sci-fi fantasy at this point. The real future for self driving is high quality level 3 technology like Autopilot and openpilot that will solve the 90% case of people wanting highway autonomy and stop-and-go traffic automation. That stuff is already here, and works incredibly well. Once it trickles down to the economy price point, no one is ever going to want to buy a car without it. But robo taxis are not happening until we solve artificial general intelligence.
Humans already drive distracted (phone calls, texting, eating, makeup), no automation required. If the automation improves the accident and death rate, it’s still a win.
One solution to bridge the gap between 'mostly' self driving and 'totally' self driving is to increase the safety systems.
Currently, safety systems for the people in the car are very good. But for people outside the car, they are very bad. If you can make it such that the car is very safe for all the people in a possible accident, then maybe the AI problem won't be such a problem.
Granted, that's a hard problem too, but we're pretty good with the theorems and modeling that goes into safety. It's more of an economics problem, not a design one.
Agreed, but the "safer for the outside" systems aren't (in my opinion) automation; they're regulations and policy to lower vehicle mass (more mass + velocity = more potential harm) and to be more antagonistic to vehicular traffic to force it to behave with greater care [1] [2] (regardless of whom it in control of the vehicle) around vulnerable stakeholders such as pedestrians and bicyclists.
Hmm, that could be a potentially interesting model in and of itself. Rather than presuming near-term viable full autonomy, use the equivalent of drone pilots in remote locations.
If the tech/network were reliable enough to make that work at scale, the service would be much more efficient than Lyft/Uber: minimized cost of labor whether the rider is in New York or India, an abundance of available drivers or riders to match with at any hour of the day, dramatically reduced odds of physical violence between drivers and riders, and elimination of social pressure for drivers and riders to talk to each other. The main (debatable) downside I see is the major up front capital expenditure of rolling out and maintaining a global fleet of remote/autonomous vehicles, whereas Lyft/Uber can rely on drivers providing pre-existing vehicles and eating all the negative externalities of ride sharing.
If you start with the assumption of humans as the primary drivers rather than assistants, then any amount of the work you're able to automate (both in the present and increasingly over time) is just gravy.
Waymo is far more advanced than anyone else, but not cheap enough to undercut the pay of an Uber driver ($9/hour?). There's still huge value if the Waymo Driver (their name for the self-driving system) can be added to private vehicles as a upgrade. A car that could pilot itself on road trips while you sleep overnight would be a game-changer for weekend getaways. Imagine leaving your Chicago office after work on Friday and waking up in Denver to go skiing Saturday morning with all your gear, no rental car, and no airport hassle.
The problem is that, unlike Uber, they need to pay the cost to maintain their fleet of vehicles.
I’m guessing they aren’t comfortable enough with their tech to deploy it everywhere. What happens if it starts snowing on your trip to Denver, you’re asleep, you’re in traffic, and there isn’t sufficient internet access for a safety driver to intervene in an emergency?
I'd assume it would have failsafe upon failsafe. E.g. in your example it would pull over at the earliest safe location while waking up the driver because the remote backup was unavailable.
Add some autoracks to the back of the train and let people bring their cars with them. (Amtrak already has such a service between DC and Florida; I've always thought a Chicago-Denver trip would make a lot of sense as well).
The "future" of automonous travel is going to be large geofenced shuttles with no other cars on the road (think large college campuses, corporate campuses, amusement parks, etc.)
I’ve seen this in action already and it works quite well. A company called Optimus Ride is shuttling people between an apartment building and a local dc metro station on a mostly private road.
I am surprised Waymo did not try to target this market first as its a lot lower risk and easier to pull off.
>I am surprised Waymo did not try to target this market first as its a lot lower risk and easier to pull off.
Probably the same reason Google kills projects if they don't become a giant market leading success quickly. At Google/Alphabet's current scale a moderate win is not worth pursuing.
I wouldn't bet on it but I also wouldn't be completely shocked in Alphabet ends up pulling the plug. Can they create a product they can sell that provides full autonomy under a useful subset of conditions within the next decade? If their honest answer is no then I have to think people within Alphabet are going to start asking hard questions.
Geofenced limited access highways also seems like a fairly promising area.
It's amazing how many people bought into the hype that there would be widespread self-driving taxi services by the end of last decade and that kids growing up today would never need to learn to drive. Even a relative skeptic like Rodney Brooks is looking to be overly optimistic at this point. https://rodneybrooks.com/my-dated-predictions/
>I’ve seen this in action already and it works quite well. A company called Optimus Ride is shuttling people between an apartment building and a local dc metro station on a mostly private road.
The real question here though is "Is this safer and cheaper than just paying a bus driver $12/hr?". My suspicion is no. And when you don't have the scale effects of millions of applications, it simply doesn't make sense.
Because the hardware and software needed to run that system reliably costs more than hiring a bus driver or two. The autonomous system might, in theory, eventually be cheaper to operate, but it isnt now.
Bird poop on sensors? Dirt? Grime? Rain? Snow? How sensitive are these sensors? Half the rental cars I get have broken blindspot or rearview camera sensors. How often will L3-L5 sensors need to be maintained/replaced in order for software to work?
Poorly marked road? Messed up road signs? Erratic drivers? Construction cones? Construction workers? Potholes? Road vibration? Road kill?
Just because it can navigate a sunny California parking lot doesn't mean any of these systems are ready for primetime. The most common job in most American states is truck driver. If there's basically 0% market penetration by L3-L5 systems in the space as of 2021, I'm not holding out any hope that L3-L5 systems will be commonplace in people's garages in the next 15-20 years.
There's also the option of only allowing automated cars on the road. This obviously won't happen on a national level in any country any time soon, but I could see it being on a city-by-city basis.
The biggest issue with autonomous travel is that it needs a 0% failure rate. Anything worse than that is not good enough, even if it's safer than non-autonomous driving by an order of magnitude.
I remember the articles from 2018, where “up to” was buried in the details if said at all. I work in automotive and am a perpetual autonomous vehicle skeptic. All the comments (HN) about how ”it’s happening” and all the downvotes I got saying ”it’s not”.
Here is how you can tell when it’s for real... when the focus is on long haul fixed A to B routes. Trucks becoming trains. When you can learn over and over on a fixed set for all the scenarios of road and weather and light and obstacle on one path you can then expand it.
All this focus on “everywhere” development is the exact opposite of minimum viable product and it’s only for headlines and grants imo. Obviously we’ll eventually get there, but, I’m personally betting that human occupant drones “take off” faster.
It seems so obvious to me that I think I must be missing something that the clear narrow use case of limited access highway driving (probably in a subset of weather conditions) doesn't get more discussion. Long highway drives are a thing for many people. They're boring and easy to zone out on. It would be huge for both safety and comfort. And it seems much easier to implement given the lack of pedestrians, cyclists, and unprotected left hand turns.
>It seems so obvious to me that I think I must be missing something that the clear narrow use case of limited access highway driving (probably in a subset of weather conditions) doesn't get more discussion
Welcome to my brain every single time I read this.
It makes no sense. Trucks and busses specifically have a ton more room for sensor packages, different height options, no one cares if that extends to or interrupts visible surfaces, and it makes sense to still have local past mile human drivers. It’s all there now.
The issues are that certain things available in a city aren’t there on a highway. Paint lines may be faded, over shot, missing, things are moving faster, there less reference points (think farm country), etc. But these are FAR better IMO than trying to drive in urban areas and running over people.
What metric are you using for saying they are the market leader? Autopilot has been around for longer and is deployed to more cars and works in more situations than super cruise which still requires mapped roads. The only advantage I can see for super cruise is that they have a driver monitoring that doesnt require pressure on the wheel.
Market leader in terms of working functionality in products that consumers can buy today. Unbiased reviews have consistently shown that Tesla is inferior.
Inner city taxi service you can build with a relatively small number of cars on a per-city basis.
Doing highway routes as a taxi service will need more cars and/or pre-booking, otherwise you end up with a situation where no cars are in the stating city. Within a city you can redistribute relatively quickly.
This leaves an assistant mode in a car owned by an individual. There you see more and more advanced assisting systems, like distance checkers and lane keeping things and those are improving ...
For a truck it's only valuable in some situations. Generally the driver isn't only there to drive the truck, but also to load/unload and protect the goods. Also one could wonder how well people would load the truck (especially one going to different locations) if there is no driver risking his life and therefore doing at least basic safety ...
Yes, you can't have a taxi service if it isn't end to end.
So this would be a feature in a car you own. And many manufacturers are heading there incrementally with assistive driving features. Key questions are at which point you can say "Doze off if you like" (And is there a point before that where drivers will doze off anyway because the system is good enough that they stop paying attention.)
I predict that in 10-15 years, no major car company will offer self driving features (besides auto breaking and maybe "beep when out of lane"). When you get hit by random Joe in normal car, he has mortgage and $5000 in the bank. It makes no sense to sue. But when you get hit by self driving car, there's company behind it, you can sue for millions. Everybody and their dog will be doing it till they stop offering self driving features or they go bankrupt, whichever comes first.
Liability is certainly one issue that will have to be dealt with. Outside of drug side effects, there are very few products sold to consumers that, if used and maintained properly, will still sometimes randomly kill people and we shrug our collective shoulders because "stuff happens."
I imagine that before that point insurance costs for non self driving cars would be so expensive that random Joe won’t be able to afford a “normal” car.
If insurance is a profitable product today (it is), why would it skyrocket in price just because there are safer cars on the road (presumably paying lower rates)?
It's the people who are currently the highest risks and paying the most who would be the first to switch to a self-driving car if they can afford it. The people with DWIs, etc.
Therefore the pool of people who drive for themselves would be getting less risky and their insurance premiums would go down.
Also, consider that right now, we have a minority of people who drive classic cars, which haven't got safety equipment, and their insurance is cheaper, not more expensive. Same with motorcycles, which have much higher accident rates than modern cars. So clearly a vehicle being more dangerous doesn't mean it's uninsurable or more expensive.
I'd really like to know if you or anyone can articulate what specific logic leads to your assumption, because people have been saying it for years. Obviously it's something people like to say. I feel like it might have something to do with a misguided analogy to health insurance.
That’s assuming that they have a truly L4 system. I don’t think it’s that simple. Danger can go from 0 to 100 fast in a car. There are so many corner cases that it isn’t hard to imagine a few stacking up and serious accidents happening.
That Phoenix, with its very clear weather, is the only market with their taxi service is damning.
144 comments
[ 4.3 ms ] story [ 64.5 ms ] threadThe CEO the article is about wasn't exactly at the frontwards edge of technology, he was an automotive industry expert.
I'm concerned about the numbers, but I'm hopeful about change in leadership. Especially given the ties with Google, possessing an insane, astronomical amount of street-based data. Surely they can scrape together a L2 driving agent given a few years?
An L2 car isn't a self driving car no matter what Elon says.
Broadly: that's seems like senseless pedantry to me. Tesla owners, even numerate serious ones who understand the problem area, know when their car is "driving itself" and when it isn't. Quoting industry jargon to explain to people how the product isn't akshooally doing what it clearly is isn't going to win many arguments.
People are bad at that kind of supervisory task-- they see it work a couple of times and stop paying attention.
And it doesn't get to me taking a nap in the car and ending up at Disneyland-- at best it makes a long drive slightly less fatiguing.
This, I think, is why Tesla will eventually succeed. Every Tesla sold is gathering data about its environment and phoning home - we already know that. So whether you have Autopilot or not, Tesla gets data for its Full Self-Driving Capability; when you swerve to avoid a possum at 12:47 a.m. because you've been working all day and night and into the morning and you really need some Whataburger, or, if you must... In-N-Out..., some engineer somewhere at Tesla can look at that event, see why you swerved out of lane, recognize it was a possum, and then jot that down into his or her notebook...
twwooo... forrrtty-seeeven... ayyy emmm... driver swwwerved... to avoid... possum..., integrate possum-avoidance behavior into Autopilot for next release...
They can write query strings for very specific situations they're interested in, too, and receive video frames.
On the other hand, Waymo is gathering vision data on every drive, and ground truth LIDAR data.
...which sounds like kind of a killer feature to me? (Like, it really is -- I may be within days of pulling the trigger on a Model Y).
Can you walk back and explain how making long drives more fun is somehow a disappointment?
If you want it to actually drive and be able to remove your attention, well, something like this https://www.youtube.com/watch?v=uClWlVCwHsI makes the present limits rather clear.
Pretty much every car vendor already has L2 autonomy [1] - Google's not going to extract tens of billions of dollars by selling the industry something they already have.
[1] https://en.wikipedia.org/wiki/Lane_centering#Sample_of_level...
On the contrary, it proves that the other thing doesn't exist. Which is sort of important to analysis.
I know that's glib. My point was the pontification about what "self driving autonomy" is or should be or might be or is supposed to be or (my personal peeve) "is supposed to be called" is essentially meaningless when we could be discussing what is actually happening in a vibrant and interesting market.
I get it. People think about this stuff in the abstract. But it's not an abstract problem anymore, and I think arguments like yours are very much missing the forest for the trees.
Roll back the clock to, say, 1985 and let's imagine we're talking about network technology. Elon is selling a working TCP/IP stack while you're yelling about the fact that akshooally it doesn't conform to the proper OSI layers and so can't ever be a successful network. How did history work out for the OSI pedants? Same principle.
https://www.caranddriver.com/news/a35785277/tesla-fsd-califo...
https://www.roadandtrack.com/news/a35878363/teslas-full-self...
Calling an L2 car "full self driving" is incredibly irresponsible. I personally hope Musk faces consequences when yet another Tesla kills someone because of false advertising.
I repeat: this is an argument about semantics, further obscured by the use of jargon, and it does nothing to inform a discussion about what autonomy features are being offered on existing cars, nor about what the market wants, nor even about whether or not those features are safe or effective. It's just words.
If you want to have a discussion about whether or not Tesla's self driving feature is safe or not, you need to come with evidence about safety. You can't just yell about "L2", no one cares anymore.
HN is a technical forum, if you're going to have a discussion about autonomous vehicles on a technical forum then technical terms should be fair game. If I'm having this discussion elsewhere then I'll explain it differently.
But fine, if you want to be pedantic: Calling a car that cannot drive itself without human intervention a "fully self driving car" is incredibly irresponsible.
Elizabeth Holmes didn't have a product which functioned like she claimed at all (nothing compared to the tech Waymo has already developed)... She assumed w/enough effort her ideas would eventually come to fruition, so she lied about the progress her company had made so gullible investors would fork over $$. Waymo has non-tech people trying to sell a tech which is still in development, they've made an (incorrect) assumption about when it'll be ready, as I doubt anyone really thinks it's flat out impossible... just really difficult.
You see a few people saying we need AGI but yeah. I've definitely been a skeptic but mostly in the vein of it needing decades, not single digit years. And the transition to even a true L4 system is going to be hard.
That said, we seen a lot of progress and even over the past ten years, there's been a lot of progress on assistive driving system as well as field work on actual self-driving.
I suspect that machine-assisted driving will continue to get better, albeit with issues associated with humans increasingly not really paying attention. But I expect those wanting robo-Ubers will continue to be disappointed for quite some time. (And I'm not sure how much it will matter because they will have actual Ubers.)
I like the similarity of your username to mine.
As to your point, I agree with you. I guess the comparison was meant in quite a pointed way that went over the heads of a lot of commenters here.
I was meaning to just draw attention to the fact that juxtaposing the numbers 82000 -> 600ish is quite a fantastical leap.
From what I've read about the older Waymo days, I think the leadership ideas were a lot different than what they became under Krafcik. It's hard to imagine from where I'm sitting, such a stark pivot from inside the company.
I think Krafcik probably meant as well in the claims, they need X # of vehicles and so that implies that there must be X # of agents ready to pilot vehicles or that many vehicles even ready for installation and operation/testing.
Of course they were very much further than Theranos, I guess it's just the sheer magnitude of the claims in Krafcik's place.
All this to say: I don't know anything about cars, self-driving cars, electric cars, rental cars, etc. I do know bullshit when I see it, and Krafcik was selling it like someone with bills to pay.
I also wonder how much all the costly talent departures hurt Waymo. Lots of brain drain, each one also taking lots of equity, etc, on the way out.
82K is way too many to build when your design isn't finalized. It makes much more sense as a production goal when they hit mass production, or the number they expect to be able to sell.
Tesla is only now getting enough data to get from level 2 to level 3 after generating 15M+ miles of data per day for years.
You just don't build that many prototypes, period, even if you're a trillion dollar company.
You only build that many if they're saleable as real products so you can get market feedback that yes you're going down the right path.
82,000 doesn't sound like much, but for a prototype build with yield rates that are undoubtedly shitty, non-final hardware, and required automotive-spec that is completely nuts. There's a reason why it's dropped to 600 so far and it's because they haven't finalized the hardware and they realize they aren't ready for mass production yet.
Honestly, even 600 prototypes sounds like a lot to me for something as complicated and expensive as a car, but I don't work in that field so I can't say what normal is. It's about right for beyond-the-prototype-and-trying-to-get-to-a-manufacturable-product in consumer or industrial hardware, depending on the target market size.
Larry and Sergei realized they weren't very good judges of character, and so they went in the opposite direction and put the milquetoast John Krafcik in charge. Their hope was that Krafcik would be better at forging auto industry partnerships.
I'm not sure what Waymo would look like today if a tech exec had been in charge, but I doubt the CEO would have had much of an impact one way or the other on the hard problems associated with self-driving.
A brief read over his career seems to show he was pretty successful. I guess you could argue "right place/time", but his time at Hyundai and TrueCar looks like it went well.
The biggest hazard Waymo faced the past 5 years was some kind of Media witch hunt based on a sensational accident of some sort, and that didn't happen on Krafcik's watch.
His job was to do a political highwire act and manage the interests of his team, Larry and Sergei and Ruth, regulators and politicians across multiple levels of government, partners across several industries, competitors as well as the media and public expectations. Somehow nothing catastrophically shitty happened.
Self driving cars are still a science experiment, and that's not on Krafcik. He killed it. At the front end Waymo is as ready to scale and commercialize as it can be at this stage.
Tesla ... Tesla is nowhere near that, and is just doing stuff that scares me.
I also just don't think the Tesla cameras have enough angular resolution to perform a lot of scenarios safely, like detecting a much faster vehicle in a short merge.
(Note Tesla relies upon the front-facing radars quite a bit, too, so it's not really pure vision).
But who knows, it seems like they’ve made good progress as is so well see. I don’t understand the people that think they’re never going to upgrade the camera suite though, cybertruck and the semi already have a new camera layout.
Well, they've committed an awful lot to current FSD customers that I think will be difficult to deliver upon with the current camera suite.
I think LIDAR is a huge advantage. There's the direct benefit: you really know exactly where things are. But secondarily, you have ground truth near-perfect range data to use in training your vision system.
We don’t yet know the limits of the pure vision approach. Elon is very confident in it and I want to see where it goes. Your comments about resolution may or may not matter.
I'll believe it when I see it, no matter what recent statements have been.
> Fewer sensors is much better for Tesla since they need aerodynamics and less power consumption for more vehicle range.
Automotive radars use like 10W. The least your car is using while driving is like 5000W. This may not be the best place to save power-- getting heat pumps on more vehicles and better ones than the present one would make a far bigger difference.
> Your comments about resolution may or may not matter.
If you can't see a car that will hit you when you commit to merging on a short merge, it is a slight problem.
Waymo is going to get canned or pivot to a new concept, like bus rapid transit on closed loop systems.
This tech has to execute in the six nines or more safety margin, and it won't. You have to solve AGI first.
You can tell by the exits that this org isn't doing well.
Tesla won't solve this problem either, but they've got a product that sells.
That's a completely safe bet. US fleet turnover is >15 years. If it were 100% capable and relatively cheap today, you'd barely hit that milestone.
> Waymo is going to get canned or pivot to a new concept, like bus rapid transit on closed loop systems.
I don't think so.
If driverless cars are as revolutionary as proponents claim, fleet turnover would happen much faster.
Waymo hasn't been able to expand their drivable area in years. This is their coverage map: https://i.redd.it/yc80awv2o3z51.png
Tesla would kill you half the time making this unprotected left turn. Waymo can make unprotected left turns and have never killed or injured anyone or even bent metal in this situation. :P It's not a particularly challenging left turn.
And the goal is to kill you less than .0001% of the time making left turns.
Yes, Waymo tries to avoid them in route planning, but it still takes them if necessary.
Because Tesla's are able to drive everywhere, there's way more ways for it to fail.
Seriously?
At this point, the total revenue they've collected so far still hasn't covered Levandowski's $125M bonus, not to mention anything else.
Yes, profitability isn't here yet, nor have they recovered development costs to date. But that is not what was asserted. Instead, they are to (limited) market, which is a milestone.
It's that their revenues are negligible compare to their expenses. Their "revenue-generating service" is a PR stunt.
https://news.ycombinator.com/item?id=26674924
I think a more likely thing is that we see it for lorries on motorways (sorry, err.. trucks on freeways) where it's a much easier problem, there are no cyclists or pedestrians, you can have them drive in the middle of the night when there's no traffic and there are convenient service stations you can stop at and have human drivers take over.
Like, why pay for a lorry driver to take all of your containers from Felixstowe to London when you can just have automatic lorries deliver them to a depot just outside London?
I haven't heard of anyone actually trying to do this though. Maybe not sexy enough. Not a big enough total addressable market or some such bollocks.
Uber bought out almost the entire CS department of CMU (Pittsburgh) yet they run their cars in Arizona.
Right now a lot of self driving, like Waymo, is aimed at taxi fleets, because the equipment is expensive, so amortizing the cost over many driving hours per day makes fiscal sense.
But-- a taxi fleet that requires a massive scale up / instant appearance of a human workforce when it's rainy is a non-starter.
On the other hand, personal cars would be useful with an "only sunny" autonomous mode.
But absolutely the only benefit it brings to a taxi is reduced labor costs. So if the capital cost is so high that you can't sell an $80-$100K self-driving car to moderately wealthy people, are you really going to be able to undercut Uber drivers? (And labor costs don't go to zero anyway as someone needs to clean and maintain the vehicles.)
>camera-only
I'd just point out that even Toyota uses some sort of radar for its dynamic cruise control today.
Say you can do it for $100k, and it looks wonky. That's a very small car market. But it's potentially a very large autonomous taxi market.
> I'd just point out that even Toyota uses some sort of radar for its dynamic cruise control today.
Yes, I'm sorry, I didn't mean "camera-only". Even low end cars have radars for collision avoidance. I meant "with cameras and front-facing RADAR, but no LIDAR / other expensive sensors".
Though-- cameras need to be good enough to perceive everything radars miss, and once they're reliable enough at that for L4, the incremental value provided by a radar may be limited.
Not really. They hired a few faculty and dozens of other staff from a robotics institute.
https://www.post-gazette.com/business/tech-news/2016/03/07/A...
It's still likely a bit too much, but I think they don't need much further progress.
The bigger problem is that they still have absolutely no confidence in any weather worse than when it's barely sprinkling... then the safety drivers come back out.
There is also the “who is more important” problem? A self-interst version of the trolly problem. When asked this, a Mercedes executive explained that the driver was the most important person to the vehicle. So given the option of hit this biker or swerve into oncoming traffic, they would chose the biker. So, somewhere in software there is a decision to always protect the driver, interesting problem that we have no good answer for. Back to my point though, I’m skeptical we’re even close to having all the data to make the decision properly.
I disagree.
> So given the option of hit this biker or swerve into oncoming traffic, they would chose the biker.
You act like this is really complicated, but given that self-driving cars generally are quite conservative in their dynamics and where they choose to go, this is pretty much the default behavior. Swerving into oncoming traffic is just not in their DNA right now, even in the rare circumstances where it's advantageous. Right now there's some work on aggressive dynamics for difficult situations, but this will tend to be underused not overused (just like it is by humans).
Doing the default, conservative thing is generally the self-protective one.
Yes, given that scenario, the "correct" response is almost certainly to lay on your brakes not to get into a head-on collision with another moving vehicle. And while one can construct trolley-type scenarios, braking hard is generally the right choice unless you think you can safely swerve.
You can see a little of it in action in this ride: https://youtu.be/D1sZnbORfAE?t=862
Even in remote assistance mode, the car already has at least a short path planned any time it's moving.
It’s like xkcd #1879[0] IRL.
[0]: https://xkcd.com/1897/
The lifetime of the sensors is also important. If the sensors outlive the car, you can amortize a big part of that $200k by just putting old sensors on a new car. If the sensors are shorter lived than the car, the total cost of the car over its lifetime will be much higher.
Which brings the ques is L3 really better than whatever L Toyota is at?
Currently, safety systems for the people in the car are very good. But for people outside the car, they are very bad. If you can make it such that the car is very safe for all the people in a possible accident, then maybe the AI problem won't be such a problem.
Granted, that's a hard problem too, but we're pretty good with the theorems and modeling that goes into safety. It's more of an economics problem, not a design one.
[1] https://en.wikipedia.org/wiki/Traffic_calming
[2] https://www.nyc.gov/html/dot/html/pedestrians/traffic-calmin...
If the tech/network were reliable enough to make that work at scale, the service would be much more efficient than Lyft/Uber: minimized cost of labor whether the rider is in New York or India, an abundance of available drivers or riders to match with at any hour of the day, dramatically reduced odds of physical violence between drivers and riders, and elimination of social pressure for drivers and riders to talk to each other. The main (debatable) downside I see is the major up front capital expenditure of rolling out and maintaining a global fleet of remote/autonomous vehicles, whereas Lyft/Uber can rely on drivers providing pre-existing vehicles and eating all the negative externalities of ride sharing.
If you start with the assumption of humans as the primary drivers rather than assistants, then any amount of the work you're able to automate (both in the present and increasingly over time) is just gravy.
I’m guessing they aren’t comfortable enough with their tech to deploy it everywhere. What happens if it starts snowing on your trip to Denver, you’re asleep, you’re in traffic, and there isn’t sufficient internet access for a safety driver to intervene in an emergency?
The "future" of automonous travel is going to be large geofenced shuttles with no other cars on the road (think large college campuses, corporate campuses, amusement parks, etc.)
I am surprised Waymo did not try to target this market first as its a lot lower risk and easier to pull off.
Probably the same reason Google kills projects if they don't become a giant market leading success quickly. At Google/Alphabet's current scale a moderate win is not worth pursuing.
It's amazing how many people bought into the hype that there would be widespread self-driving taxi services by the end of last decade and that kids growing up today would never need to learn to drive. Even a relative skeptic like Rodney Brooks is looking to be overly optimistic at this point. https://rodneybrooks.com/my-dated-predictions/
The real question here though is "Is this safer and cheaper than just paying a bus driver $12/hr?". My suspicion is no. And when you don't have the scale effects of millions of applications, it simply doesn't make sense.
The narrative around FSD has changed very quickly this year. By the end of the year, Tesla is going to be neck deep in lawsuits.
Poorly marked road? Messed up road signs? Erratic drivers? Construction cones? Construction workers? Potholes? Road vibration? Road kill?
Just because it can navigate a sunny California parking lot doesn't mean any of these systems are ready for primetime. The most common job in most American states is truck driver. If there's basically 0% market penetration by L3-L5 systems in the space as of 2021, I'm not holding out any hope that L3-L5 systems will be commonplace in people's garages in the next 15-20 years.
Bird poop: https://cdn.vox-cdn.com/uploads/chorus_asset/file/8529507/_9...
Construction: https://youtu.be/s1mAWg23phU?t=484
Erratic, unexpected crap: https://www.youtube.com/watch?v=weXDUc5Osto
Even high speed rail is highly automated ( due to the fact that there's little a human can do better when at 320km/h).
The biggest issue with autonomous travel is that it needs a 0% failure rate. Anything worse than that is not good enough, even if it's safer than non-autonomous driving by an order of magnitude.
Here is how you can tell when it’s for real... when the focus is on long haul fixed A to B routes. Trucks becoming trains. When you can learn over and over on a fixed set for all the scenarios of road and weather and light and obstacle on one path you can then expand it.
All this focus on “everywhere” development is the exact opposite of minimum viable product and it’s only for headlines and grants imo. Obviously we’ll eventually get there, but, I’m personally betting that human occupant drones “take off” faster.
Welcome to my brain every single time I read this.
It makes no sense. Trucks and busses specifically have a ton more room for sensor packages, different height options, no one cares if that extends to or interrupts visible surfaces, and it makes sense to still have local past mile human drivers. It’s all there now.
The issues are that certain things available in a city aren’t there on a highway. Paint lines may be faded, over shot, missing, things are moving faster, there less reference points (think farm country), etc. But these are FAR better IMO than trying to drive in urban areas and running over people.
https://www.cadillac.com/world-of-cadillac/innovation/super-...
https://www.motortrend.com/cars/cadillac/escalade/2021/cadil...
https://www.consumerreports.org/car-safety/cadillac-super-cr...
Doing highway routes as a taxi service will need more cars and/or pre-booking, otherwise you end up with a situation where no cars are in the stating city. Within a city you can redistribute relatively quickly.
This leaves an assistant mode in a car owned by an individual. There you see more and more advanced assisting systems, like distance checkers and lane keeping things and those are improving ...
For a truck it's only valuable in some situations. Generally the driver isn't only there to drive the truck, but also to load/unload and protect the goods. Also one could wonder how well people would load the truck (especially one going to different locations) if there is no driver risking his life and therefore doing at least basic safety ...
So this would be a feature in a car you own. And many manufacturers are heading there incrementally with assistive driving features. Key questions are at which point you can say "Doze off if you like" (And is there a point before that where drivers will doze off anyway because the system is good enough that they stop paying attention.)
Therefore the pool of people who drive for themselves would be getting less risky and their insurance premiums would go down.
Also, consider that right now, we have a minority of people who drive classic cars, which haven't got safety equipment, and their insurance is cheaper, not more expensive. Same with motorcycles, which have much higher accident rates than modern cars. So clearly a vehicle being more dangerous doesn't mean it's uninsurable or more expensive.
I'd really like to know if you or anyone can articulate what specific logic leads to your assumption, because people have been saying it for years. Obviously it's something people like to say. I feel like it might have something to do with a misguided analogy to health insurance.
0. https://ir.tesla.com/press-release/tesla-q1-2021-vehicle-pro...
That Phoenix, with its very clear weather, is the only market with their taxi service is damning.