Waymo's collision avoidance testing
blog.waymo.com
blog.waymo.com
Consider the hypothetical: a million self-driving cars on the road that, collectively, will have 1/10th of the fatal accidents that human drivers would have[0]. But, the ones they do have are accidents a human driver would almost certainly have avoided.
Is this something we would accept?
My guess is that no, we wouldn't. Because the accidents avoided don't make the news, but the accidents that occur- especially ones that you say "my god, how did it screw that up?" will make the news, and our perception would be that they are more dangerous.
Until Waymo's cars are better than most humans in every single situation, they won't be able to win over the public perception war.
[0]I'm making those numbers up. I acknowledge that. But it's a hypothetical so give me some leeway on this!
Continuing with your hypothetical, even though we’d be 90% safer as a collective, the safety of the individual feels compromised: the risk of an accident is non-uniform when involving humans (depending on e.g. age, experience, safety, alertness, etc.), but becomes uniform (or at least more uniform) with an algorithm in charge.
That’s a tough thing for people to buy into.
"Until Waymo's cars reduce any individual's chance of an accident."
So for example, suppose a Waymo car is better than humans overall, but tends to do worse than humans when there's a small bump on the road. And suppose that all humans (in a given regulator's area, e.g., California) tend to encounter such bumps at roughly the same rate (per mile driven) over their lifetime. In that case, it's probably going to be acceptable, since every individual is better off.
I don't know, maybe this is not impactful / obvious enough for people to care about?
What certainly is obvious is that the safest drivers are much safer than an average driver (does anyone know of a study that estimates this ratio?). Therefore, at the very least, the threshold for Waymo should be not the average accident rate, but the accident rate for the safest drivers.
Exactly. I have had zero accidents in 20 years; I'm not interested in a car that will lower the overall accident rate if it increases mine.
Who is "them" and who allocates the groups?
We can't even get any kind of car control skill testing or training to be part of a driver license issuance.
Then you've not been following the space. The one that immediately comes into mind is the Tesla that slammed into the side of the semi truck because it was painted blue like the sky.
If you follow the events with self-driving accidents, most of them are nonsensical crashes that no human would ever have done.
I got out of a ticket once because I didn't see a "no through traffic" sign against a bright sunset. No chance that same cop gives a self-driving car a pass, nor should he.
It’s also possible that they were so far behind Waymo in the journey to FSD that they weren’t yet at a point where multiple sensors would make a significant difference.
Just my interpretation, but it felt to me like a hail mary because they were fully committed to being the first mover. Waiting around for LiDAR prices to come down would have meant that Waymo would have beat them.
Unlike Waymo who doesn't care about selling cars to people to drive daily. No one going to care that the taxi they are taking looks ugly as long as it get them to the place they are going for cheaper. The cost is also a less of a factor due to them being able to produce an income by charging people to ride in them.
Why not? There are production cars with lidar now that isn't the big spinning thing on top of the car.
The thing they need to improve and are doing so rapidly is actual trajectory policy calculations.
And that’s not going to get better with more sensors seeing the same things it already sees.
As someone who drives a Tesla with the FSD beta, the vehicle has been getting progressively better since 2018.
It’s drives smoother and brakes more predictably dive they stopped using the front radar.
As someone who owned a 2019 Tesla and who owns a 2023 Tesla, the older car had better autopilot. It was starting to degrade in 2020. The new one is worse. Phantom braking was quite rare in 2019. The very first night I drove home the new car, it phantom braked on a lonely, empty stretch of I-5.
I want the radar back.
Essentially that trying to utilize multiple sensors cripples any progress (given that resources will never be infinite).
For example, if your lidar sensor is 99.999% percent sure there is an obstacle in front of you, surely it's helpful to take that information into account, even if it is a tiny bit uncertain/noisy.
More data in any situation where bandwidth is already maximized will lead to entropy and noise. However, if the capabilities were there to process all of that data in low latency scenarios with headroom to spare, surely adding additional sensors and data points would lead to a more complete model of spatial awareness for the car.
That's all hypothetical and reliant on ignoring the realities of running a business and tech development lol
Fusion is hard. As hard as getting humans to agree. Been there. In both situations.
And of course you can concentrate on improving one echo chamber, err single sensor. But you can never come past its fundamental limitations.
Point stands that the ones being put in the cars aren't.
The eye is an incredible piece of machinery, specifically for dynamic processing like this.
And agreed that eyes are impressive.
Someone else said it here on HN, and I think they're absolutely right -- Tesla is all about vertical integration, and this is preventing them from excelling at anything other than saving pennies. A good part of why the new EV competition is doing everything better is they didn't roll their own tech. They bought packaged solutions from companies that only do one thing, but do it well.
I have even earned the dubious honor of having been pulled over for suspected drunk driving because I used AP on a trip home from my MIL's house after Thanksgiving dinner. I was amused by how much trouble AP was having with mild corners on I-5, the state trooper behind me was less amused. Well, at least until he established that I was sober, then we had a good laugh about autopilot and he gave my kids sticker badges. A happy ending for everyone but I didn't activate AP again that night.
Maybe the problem was that it was clear and dry? Ya think it would have been better with some road spray on the cameras?
Subsequent to that, I ended up abandoning AP altogether because it was so jarring when it would decide an overpass looked like something it should slow down for. My wife suggested that she'd be happier if I just used my foot for the accelerator. For some reason we still ended up buying another one, but after the phantom braking on day one (empty freeway, overpass, apparently this is very much now "a thing"), I guess I'll just skip AP before even starting. I wonder if I could ask Tesla to deactivate AP altogether so I could have traditional cruise control?
It’s not FSD, but if you use it as a driver assist, it takes a giant amount of burden off of driving in traffic. Even on a busy highway that’s going somewhat fast, it’ll just follow the car in front of you. And truly, it’s never once been dangerous.
Only catch is that it can’t “see” very long distance — if you’re going 70mph and there’s a stopped car in your lane, it probably won’t stop (at least, I wouldn’t test it!). But again, it’s driver-assist and not FSD, so you’re still expected to notice that. And within its limitations, it is really exceptional.
This is like saying we need to get to the moon using steam power and ONLY THEN we can improve by using rockets. What if it's completely infeasible to solely use vision for FSD?
Even humans use audio processing to augment the vision.
https://www.forbes.com/sites/bradtempleton/2022/12/12/tesla-...
And in fact some regulators fully understand the tradeoffs and will prefer autonomy for the better good of the public. An example of this is the Boeing 737 Max, those crashes wouldn't have happened if there were no autopilot systems. But regulators are not suggesting that all autonomous systems on planes be turned off because of the safety and financial advantages of keeping them in place even though they are obviously not perfect.
Bad example. MCAS was an obvious case of a criminal corporate behavior, not a tradeoff between overall safety vs. technical perfection.
What if we would make cars and rules that are adapted to AI cars and ignore human drivers? e.g. Ban human drivers from some roads, allow AI cars with designs that exploit AI advantages (e.g. much better reaction time) but do not require or even allow human backup (enabling us to put the passengers in a secured shell), etc. I suspect we could than reach a 1/20th rate today.
The Vertrauensgrundsatz does not account for distracted, tired and inebriated drivers.
Given how reckless some drivers are, I'd imagine the distribution of crashes among drivers is very skewed. I don't know numbers, but I'd expect to see something like 80% of car crashes involving (or objectively caused by) 20% of the drivers, with most car crashes involving people who have been involved in numerous crashes, and a large minority of drivers having been involved in zero car crashes.
There is a lot of variation in drivers. Young ones with good reflexes and senses but bad on experience, old ones with worse reflexes and senses but better on experience. Ones that stop at stop signs, ones that blow through them if they think they can get away with it. Cities like New Orleans where drivers are much more aggressive because the cops don't care, etc...
It matters who is driving what cars. If you're taking safe drivers and having them drive FSD cars, and comparing the average of that against the average of all non-FSD driven cars (which includes a lot of really shitty drivers in the average), then the net safety impact of this could be nothing. Introducing FSD cars would have no impact to overall car safety, and might even have a negative impact, if the FSD software is worse than the good drivers who are now not driving manually-steered cars but are driving FSD cars instead.
I present, "Canada's Worst Driver," a series that I watched quite a bit when I had a cable subscription. Besides being hilarious, it should make you think a time or three before crossing the street. https://www.youtube.com/playlist?list=PLHB4WuM7Y_eP-PFv5F6To...
Basically, maybe there’s better ways than standard vision (which for computers is still kinda basic compared to a person) to solve the pedestrian hitting problem.
The cars are clearly the danger to everyone who isn’t in a car, why would the burden be on everyone else, all the time?
I really don't want to be fair though, because I think your point is totally correct.
Because reality.
Since the advent of motor vehicles, I would argue risk has generally scaled with the efficiency and power of vehicles.
At the minimum, for 70 years we have known that danger to both pedestrians and safe drivers is ever present and growing. The reality is not about who SHOULD deal with this problem, but how to most effectively arrive at a functional solution which actually reduces risk for drivers and pedestrians alike.
As well, it's worth noting most accidents occur in areas where you drive most often and you're most familiar. And the increase in traffic accidents and fatalities over the last 10+ years has been due to driver distractions, ie cell phones etc, more than anything by a lot.
Yes there are reckless drivers out there doing stupid shit. However, that's always been the case. Now it's just easier than ever to capture them on camera and for them to go faster than ever for cheap. However, even back in the 70's, crash safety technology was laughably behind anything we have so it goes both ways imo.
It would be more remarkable if most accidents occurred in areas where you rarely go.
No, but there's probably a good reason it's suspended. It's not that you forgot the rules. It's just that you're not following them.
Presumably by paying their child support.
I don't know if it works as intended, and it wouldn't surprise me if people have horror stories about clueless rulings by clueless judges, but obviously the intent is that the license is suspended because the person is intentionally not paying child support.
> No, but there's probably a good reason it's suspended.
Not paying the renewal fee on time is a big reason. That's a rule, indeed, but nothing related to driving or safety.
There have been suggestions that pedestrians could make cities interesting, since a place like New York would have every pedestrian no longer afraid to just walk into traffic. The cars will just wait until a fully clear moment that may never come
As you say, if you know the cars will stop and let you go, this power will corrupt.
Extend this thinking to bicyclists, skateboarders, delivery riders, motorcyclists, car enthusiasts driving themselves - all able to simply cut off a self driving car knowing it will give way.
But there is more adversarial out there.
AI input hacking to get known responses. Cars disabled because of gang tagging on sensors.
Car enthusiasts or privacy enthusiasts (privacy shouldn't require enthusiasm but here we are...) updating car firmware/software. People with full physical access!
And we haven't touched on the ethical stuff which is a world of discussion in itself.
Self driving cars by evolution in current societal rules etc will take a LONG time.
> Extend this thinking to ... - all able to simply cut off a self driving car knowing it will give way.
But this is already the case; if I jump into slow moving traffic in a (western) city centre the cars will let me through, possibly even without honking. The only things to fear are distracted drivers (which is the same with the AI, because I can't know whether the AI sees me) and the occasional madman, but those are rare.
What's more interesting as you point out are situations where other drivers will take advantage of the AI's adherence to rules and safety. That's going to be especially fun in cities like Napoli, where things like driving through red signal are completely normal. It's probably not going to work there at all in its current state.
Would stepping out in front of a car be illegal?
https://laist.com/news/transportation/understanding-new-cali...
If "person walks in front of car" event is rare, then the automatic car stop saving a life or minimising injury is nice. Saved a person from a bad day for a minor inconvenience for the car people.
If a NY city worth of people do it without regard, then cars will be gridlocked.
My guess is:
- Assuming self driving cars that that stop for pedestrians are common
- then laws against deliberately walking in front will exist
- to judge the dispute of "deliberate" there will be more cameras, because why wouldn't there be...
London already has near suicidal pedestrians.
And I say that with respect: it's the opposite of the usual tragedy of the commons. Every pedestrian in central London who just crosses the street without looking places themselves in appreciable danger, but contributes a tiny bit to training London drivers to be extra careful.
(Going off on a tangent: tipping is an interesting opposite.
If everyone in the US stopped tipping tomorrow, economic equilibrium would shift so that service sector people would have comparatively higher nominal wages and roughly the same overall income. That's how it is in Japan or Australia or Singapore (or the US before the 20th century), where tipping is either less common or severely frowned upon.
Because eg prospective waitstaff in the US expect to be tipped, they accept lower wages. Similarly, a single patron not tipping in the US gets a lot of social awkwardness without shifting the equilibrium much. A single waiter in the US not accepting tips loses out on income and also gets social awkwardness, also without shifting the equilibrium much.)
Next round of CAPTCHA: click the squares with blind pedestrians
think about it, this is how these things are being trained--by people frantically trying to click stupid images just to get to the webpage they wanted.
About comparable to the risk of falling. Lower risk than suicide. Or death by opioid overdose. And of course, the most dangerous thing most Americans do, by a huge huge huge margin, is overeat and lounge on the couch.
Relatedly, another "problem" with "self-driving" cars is that we want all of the convenience and ease of use, without adjusting liability and ownership considerations. Consider, if Waymo gets to the point where they have a self driving car that you have to have a subscription to use, do you own the car? Are you the liable for any accidents it has?
To lean in on that hypothetical. I'd imagine a lot of families will use self driving cars to send kids to school. Is effectively a bus that terminates at your house. Who is liable for a mistake if the operation of it is completely remote?
I'd imagine the company would take on liability, as long as humans can't drive the vehicle or they aren't driving when the accident occurred. Mercedes already got the ball rolling on this [1].
[1] https://www.roadandtrack.com/news/a39481699/what-happens-if-...
Aside from some niche cases in very dense cities, is this generally true anywhere? I've visited a lot of cities with various levels of public transit, and I can't think of many where it was faster. More convenient sometimes, sure, cheaper, yep, but faster? Not often.
I was leaning into it being a social thing. If you care about reducing fatalities and such, public transit has been a thing for a long time.
As far as convenience goes self-driving cars should be extremely convenient once we have them, no? I agree with you that public transport makes sense in very densely populated cities but most places in the world are actually wasting money on lots of public infrastructure that makes little rational sense, simply because it's a popular demand.
If anything, legislation and social acceptance has moved faster than the technology. That's the opposite what many of us observing this space expected 10 years ago.
At this point I'm starting to have doubts about whether the full dream of self-driving cars will even be realized within my lifetime.
I'm impressed that they're actually offering driverless rides on SF streets, but my point stands. The cars operate only on the slowest streets at the quietest hours. Any problem they encounter is handled by remote operators.
I'm not outright dismissive of self-driving cars. I truly want them to exist. I don't even own a car and dislike being behind the wheel. I just don't buy into infinite hype pushed by a revolving door of charlatans.
Also I do have a modern VR headset and celebrate the technology. But, to make a similar comparison, the metaverse "ready player one" vision is not within our lifetimes.
If you're into self driving tech, Cruise especially is regularly ridiculed for their vehicles blocking traffic. Take this latest example: https://twitter.com/as611/status/1597144790767788032
But more importantly in case you're not aware it's not true self-driving at all, they're geofenced and both Waymo as well as Cruise have been shown to have issues with even minor changes like a construction site. The tech isn't there, we don't really have full self-driving and if we actually put a million "self-driving" cars on the road as the above user suggested, it would be utter mayhem with current tech. If this will change in the future is to be seen and from what I've observed progress has already slowed as per ninety–ninety rule. It's easy to have some autonomy functions. It's hard to have millions of fully automated cars. So far, we're nowhere near it.
This level of sophistication makes me think it will not "frequently deadlock".
The modern Level 2 car is already today at below 1 fatality per billion vehicle miles travelled (VMT). The autonomous car then needs to be below 0.1 fatalities per billion VMT. Meaning that if you have 1 million vehicles of your make deployed, they each need to have driven 30-40 000 miles autonomously before you have enough statistics!
That means proving the safety of a Level 4/5 autonomous system is extremely expensive and slow, and requires significant public adoption before it's proven to be safe. The consequence is that, assuming proven safety is necessary before public adoption, it becomes impossible to prove safety.
Another point is that OTA upgrades for autonomy become entirely pointless, as you'll be polluting your statistics if you change the code more frequently than every ~3 years.
Does this statistic count fatalities to people outside the car? I ask because (in the US, at least) car safety ratings don't take those into consideration.
Which ratings? The NHTSA does include non-occupant fatalities in their statistics. E.g. https://www.nhtsa.gov/press-releases/2020-traffic-crash-data...
> For the first time ever, [the proposed rating program] includes technology recommendations not only for drivers and passengers but for road users outside the vehicle, like pedestrians.
News story on this: https://www.theverge.com/2022/3/3/22960262/nhtsa-ncap-five-s...
The incentives are misaligned. Tesla should want to make the software better not just to attract some more customers, but because if they screw it up they're going to be on the hook.
How many times do you think this has happened?
I think I'm relatively safe from cars on the sidewalk. Yet with fsd cars I'm not so sure anymore.
Not even true. I've been hit by a car (with human driver) on a sidewalk. Driveways cross sidewalks, and drivers seem particularly inattentive when crossing them. When you're walking on the sidewalk, do you ever walk somewhere that's more than one block away? If so, you're going to have to cross a street anyway.
My partial answer is, making an extremely reliable system is hard. If someone wrote a deadly bug even only happen at a very corner case, it still can kill people. And it's quite hard to prove there isn't such bugs.
You only need 1 or 2 drivers for a huge train carrying a lot of ppl/stuff. Optimizing away the driving barely reduces the cost of operating the train as a whole.
With cars, everyone drives themselves. There’s a lot more driving happening and so if you can automate that away, you create more value.
> it's a hypothetical so give me some leeway on this!
IMO you should not base (and broadcast) your opinions about safety on hypothetical statistics. I don't even believe it's true that overall statistics show self-driving is safer than humans. IIRC prior reports showed that companies were selectively picking statistics about safety.
Moreover, other players also have cars covered in cameras. If anyone thought vision only was the best path forward because they couldn't figure out sensor fusion, they'd already have done it and saved the BOM cost.
For what it's worth, my experience has been that cameras are one of the more problematic sensor systems overall. Vendor software is garbage, any particular tuning is finicky in extreme conditions, you have to clean the damn things, cameras streams take up lots of bandwidth, etc.
If this becomes true then society would have an even bigger problem with organ donations - available supply would plummet. Some of the largest sources of organ donations are from car crash victims.
Surely we’re in agreement that this is a good thing, right? That a shortage of organ donors due to the donors not dying is a good thing. Sometimes it feels like people are suggesting otherwise and I can’t fathom the logic.
> But, the ones they do have are accidents a human driver would almost certainly have avoided.
I suspect most human-driver accidents are also accidents that (other) human drivers almost certainly would have avoided.
That's scant consolation for all the people dying in traffic accidents each day, of course.
Airplanes are not fully autonomous, even in the instrument flight rules system which is highly standardized. You don't have non-standard or non-predictable things. In instrument conditions, only instrument planes and pilots exist. While portions of segments are automated, transitions between segments and phases of flight aren't. It's ripe for automation, yet isn't fully automated.
The automobile environment has more objects thus more density of complexity, more non-standard and non-predictable actors like non-autonomous vehicles along with bicycles, mopeds, pedestrians, etc.
The current situation as basically the opposite: waymo's cars are better than humans in almost zero situations. It's hard to gain the my trust when your car can barely drive in a drizzle.
> Is this something we would accept?
No, and that's good.
Human drivers behave mostly like humans, even the worst ones. We all have millenia of evolution fine-tuned to recognize human behavior based on the most subtle of cues. So human bad driving is recognizable and thus far more avoidable.
AI drivers are effectively an alies species who make errors that make no sense whatsoever to a human mind, thus they behave, for all practical purposes, apparently completely randomly.
They already have robotaxis in SF and are expanding into Arizona and Texas by the end of this year.
- only after 10pm - used such an odd slice of the city I not only can't get picked up, it doesn't GO any where I go.
I would love to use either of these programs both for the novelty and because I think Autonomous driving is great, but I literally can't use the program I do have access to.
Totally agree w/ you rn thou right now it's basically a neat party trick I have to go out of my way to show off instead of a super useful service.
It's been a month and a half since they've covered almost the entire city except FiDi/Union Square (and Twin Peaks)[1]. That's not a trivial omission, but do you really never take rides that start or end in any other area? In particular, that entire area is a fairly close walk from the only dense transit line in the city.
I do appreciate that my car can do full distance control and assist if I am drifting, but it doesn't control itself, so I can never disengage. Personally I feel that this is wonderful and should be the limit. Anything past that should just be fully autonomous. Otherwise you're asking for trouble.
I can't imagine trying to focus on supervising an AI pilot without that kind of feedback.
An airplane autopilot is a dumb device, in that it does execute _exactly_ the plan you tell it to, and it is up to the pilot to at all times decide whether the current plan still makes sense or needs to be altered. So the pilot makes the strategic decisions, and leaves most of the physical tasks of flying to the autopilot.
I find myself using my M3 w/FSD in exactly the same way, as that I put on autosteer pretty much immediately when I'm out of the driveway, but I constantly nudge it into the lane that I want it to be in (by using the turn signal) or push the accelerator when I think it is taking too long pondering a turn. So i leave the physical driving (keep lane and distance) to the car but manage the car to always go exactly where I want it.
I have no trouble staying alert this way when doing medium long drives. Long highway drives where autopilot is so good that it requires no manual interaction is where the trouble starts and I find it hard to keep paying attention.
This is where in an airplane you have a copilot and can discuss strategic things like overnight stops, fuel stops, etc... Maybe Tesla needs a built-in chatbot to make me do that :)
I get what you're saying about the physical/strategic split, but my perception is that automobile autopilots are simply not good enough to be trusted with the "physical" stuff in a hands-off way like an airplane's can. And that's mostly because driving a car in a straight line down a highway is way harder than flying a plane in a straight line in an empty sky.
Exactly!!!!!!
The trouble is when your attention wanes and you don't know exactly when you should be paying attention. What I found is the small amount of automation my car does is wonderful to keep my exhaustion down. When I drove for 4 hours in my 2012 prius, it was a chore, i had to do ALL the mental math myself. When I drive my 2023 CX-50 I have to pay attention at least 30% less, and that 30% is a massive difference, it feels like 90%. And any time my attention wanes, the car starts complaining because I am drifting or not putting enough immediate response to the road. It becomes a "pay attention quickly" wake-up call that happens within seconds of attention waning.
The worry is when the car doesn't snap you back into that attention mode, and you just trust it, right up to the problem.
These augmented systems will probably reduce the risk of accidents so greatly that the value proposition for Level 5 driving systems just won't be there.
I've driven a lot for decades and frankly enjoy driving. I drove from TPA to SLC via PHX and back for fun. But I will pay $500/mo level-5 subscription for a comfortable car that drives itself.
Volvo was talking about level 3 back in 2017, but they gave up.[1] Level 3 means that the system may ask the driver to take over, but if the driver does not do so, the system must get the vehicle to a safe condition. Preferably pulled over out of traffic, but as least stopped without hitting anything. The driver is not required to watch the road.
The serious players are trying to get to level 4, where the driver is not expected to take over but the set of roads you can use is limited.
e.g. let's say you have a level 2 system which starts auto-evading, suddenly steering without user input, the user is likely to reflexively try counter-steering in response.
The value proposition of L5 systems is also the not driving part.
Intuitively I would have agreed with you, except Tesla has been doing it for years and their cars are statistically safer by every metric (fatalities, indicents, etc.).
Compared to what population? Older, tech-savvy people buying 60k-120k usd vehicles? And do you mean in full self-driving mode Beta, which Tesla won't allow you to use unless you have a track record of save driving?
Even if a hypothetical, future Tesla FSD sometimes crashes in ways that could be prevented had the driver payed attention, it could still be statistically safer than a fully human driver (ie the number of FSD crashes even if left unattended < the number of crashes by humans driving).
To clarify, I'm not talking about the current state of FSD, I'm talking about a hypothetical, future Level 3.
You're free to not use it. But if Tesla FSD is safer than the average driver (even if that's because the average driver is on their phone) then, going back to your statement, it is more useful than dangerous for the average driver.
Only if they don't change they name. They can call it an upgraded RealDrive™ QuantumSense™ feature that no longer requires having your hands on the wheel.
LIDR isnt an option.
This is a strong claim. I thought there was a decent body of evidence that suggested most drivers make much less money than they think, when you take depreciation/repairs/etc. into account?
[0]https://www.statista.com/statistics/193018/number-of-us-cras...
I also think that the mapping and routing component matters a lot less than how good your collision and realtime avoidance systems are. And in that arena, Tesla is an unmitigated disaster.
Yet, I follow DirtyTesla's YouTube channel and I think FSD is quite impressive compared to any other self driving software I've seen.
Would you mind to direct me to similar videos from Waymo for example in similar situations? I can't find anything even remotely as good as what Tesla is doing now.
I'm not a fanboy nor do I possess any TSLA actions (or even a car for that matter), I'm just interested in the field and until now I thought Tesla was the most promising tech (it seems I'm wrong, but I really like to see it!).
These are easily searchable which leads me to question your sincerity in feinting ignorance.
As I understand it, Waymo can’t drive on unmapped roads, and therefore there are no comparable videos of Waymo actually doing that.
You chose a Waymo video from their marketing channel, and a newspaper hit piece. And then questioned sincerity of OP…
>I see a lot of bad things on Tesla FSD and I totally get the critics...I think FSD is quite impressive compared to any other self driving software I've seen. > >Would you mind to direct me to similar videos from Waymo for example in similar situations? I can't find anything even remotely as good as what Tesla is doing now.
I can't believe one can make an honest argument that Tesla is ahead of Waymo on FSD
I agree that Waymo could well be far ahead of Tesla, but there isn't enough information in the public domain to say this with confidence. We don't have the ability to make a proper comparative assessment.
Tesla has not proven any reliable driverless operation, anywhere. They have removed hardware from their cars (radar, uss) and have not shown any meaningful progress in the past ~5 years nor any willingness to change from their “vision only, big data” strategy.
If things continue on the current trajectory Waymo will likely be operating in all major US cities and metros in a few years while Tesla’s self driving offering will probably be forcibly renamed by regulation and end in a class action lawsuit.
Basically, Waymo has proven N and N+1 capability, meanwhile Tesla has yet to prove N, and has lied to consumers and actually reduced their chances at achieving N due to cost cutting measures.
I guess they did now that they can make their own cheap lidars and adding them back in 2023.
It sounds like they might actually be including a new radar system in January. Nothing official yet though from what I've seen.
https://electrek.co/2022/12/06/tesla-radar-car-next-month-se...
Really? I’ve had some casual interest in the progress of FSD beta and the past six months alone has seen dramatic improvements to numerous adversarial situations.
FSD beta is currently able to drive with confidence on unmarked roads at night in the rain, with only basic maps for wayfinding. This has been demonstrated by customers in their own cars, driving roads which haven’t been vetted by the developers.
I’m sure Waymo and other systems can do this too, but I haven’t seen it demonstrated.
And how many years did that take. Are they adding profitable operations in new cities year after year, with every year adding new cities?
As far as I can see they simply lose billion and billions of $ for no real success in actually having a product.
Tesla is actually using the technology to improve its Level 2 systems and make money with it.
> meanwhile Tesla has yet to prove N
First of all, Waymo has not proven N, because they don't make money on any of these things.
Tesla at least try to drive in N+10000000 other cases and navigate many of them without seeing them first.
If you have to go one by one threw every single city in the world its not clear to me that this is a better approach then solving a more general problem.
> reduced their chances at achieving N due to cost cutting measures
Tesla just made 3 billion $ of profit in a quarter. What cost cutting? They are currently doing major investments in upgraded sensors suits, upgraded data-centers and overall their team is still growing.
How much did Waymo make again?
The opposite is also true, just because Waymo does not make money does not reflect the capability of their self driving systems. Saying "Waymo has not proven N, because they don't make money on any of these things." doesn't make any sense, and is not even true.
I can go to downtown Phoenix right now and request (and pay for) a fully self-driving ride from point A to point B. Teslas can not reliably complete any self driving route without any disengagements.
We are discussing who is closer to realizing a fully self-driving system, not who runs a better business.
That not what I said, what it means it has more staying power.
> doesn't make any sense, and is not even true
What data are you basing this on? Are any of their operations profitable? If they are, why are they not expanding those operations to make more profit?
> I can go to downtown Phoenix right now and request (and pay for) a fully self-driving ride from point A to point B. Teslas can not reliably complete any self driving route without any disengagements.
Yeah, but Tesla didn't spend 7 years only focusing on Phoenix, so using that as a comparison is just dishonest.
You are stacking the field in favor of what you want the winner to be.
Here is my proposal for a more fair test on who is actually 'winning' self driving:
"You take a car, and put it on any random road anywhere in the world, how well can it navigate to any other random road in the world"
How well does Waymo do on that test? I would guess worse then Tesla.
That test is much closer to what it actually takes to really claim that self driving is 'solved'.
In my opinion neither are close to this and both will burn many more billions and many more years before getting there. So to just confidently claim Waymo is way ahead is nonsense in my opinion.
Tesla currently works not at all.
It's not valid to compare Waymo's current capability unfavorably to a version of Tesla's capability that only exists in someone's head.
I would bet on Waymo working on a dirt road before Tesla does.
Literally a Tesla with FSD working well on a dirt road: https://www.youtube.com/watch?v=wv1l6aTnB_I
I do find it mildly disturbing that in that video, the driver points out the car making fully blind turns where it cannot see that there's nothing it would hit.
Think about a section of lightly used suburban road. The amount of work that went into making it involved was immense. A crew of road workers using expensive machines and large amounts of material were required to make it, and are required in it's maintenance. Don't forget the surveyors and engineers who made a highly detailed map and plans in the first place! (Though that map format isn't useful to self driving cars).
Also consider the sheer number of cars that drive that patch in a day. One car every few minutes adds up over hours, days, months, years.
So, yeah they have to drive a mapping car down the street a bunch of times to expand their coverage area. However this is insignificant compared to the effort that goes into our transportation infrastructure already.
Besides, most miles driven are spent on highways and other town-connecting roads. To the average consumer, self-driving cars are way more interesting for commuting or long-distance travel than they are for a 5-minute drive to the supermarket.
That means that Tesla could make an HD map covering 95% of miles driven in the USA within a week with their fleet of users. And next week they could make an updated version of the same map.
So, making and updating an HD map isn't an issue.
I then hired a dude with a LIDAR scanner and did it properly. The difference in quality/accuracy is like 120x80 ASF video files in 1996 compared to 4K footage today.
Anyone who thinks you can build "HD" virtual worlds using the crappy cameras on a Tesla needs their heads examining. Maybe with thousands of passes and some epic compute and signal processing, but why bother? Just LIDAR it.
My Tesla can't even decide if a traffic light is a single traffic light or not on a sunny summer's day from a distance of twenty feet. Almost every time it is either dark or humid or winter (road grime) it tells me one or more cameras are obscured. But only after I've already started driving, obviously. This supposedly cutting edge AI driving machine frequently thinks I'm leaving the carriageway on UK B-roads (it's almost dangerous) and is significantly less reliable at distance cruise control and lane-assist than my Skodă. (I presume VW just quietly bought a black box from Bosch or whoever to do this.)
Tesla are barking up the wrong tree IMO. At this point the camera-only stance feels like a religious thing, not based on sanity or the real practical world. I imagine that someone came to Elon and said "reconciling conflicting radar and camera signals is hard" and he applied his considerable genius and issued an edict to "let's not do that then!" like it would magically make all the actual hard problems go away.
Heck, Teslas can't even seem to reliably parallel park themselves, frequently getting stuck halfway, or hitting kerbs. If they can't solve that highly constrained problem, I'm hardly going to trust taking my eyes off it at 70mph.
That may be a smokescreen. Tesla collects a lot of data from their cars. What they do not have are these supposedly superfluous high resolution maps. If Tesla's camera-sourced data proves to be insufficient, that will have been a very bad gamble, in addition to whether camera data is sufficient for real time decisions.
When they pay off, bold gambles make businessmen look smart. That's why nearly all business hagiographies are the product of survivorship bias. Just like your buddy who won in Vegas.
We will see this risk-taking play out in Starship and Starlink, too.
If you were to let FSD beta just drive by itself enough time, and intentionally never intervene, it would eventually crash, no doubt about it. Before v10.69 it was hard to get a 20 minute drive with no interventions (unless it was mostly straight roads).
Comparing disengagement and driver intervention data might be useful to compare Tesla FSD vs others.
"In its first report on its autonomous vehicle operations in Phoenix, Arizona, Waymo said that it was involved in 18 crashes" (in 6.1 million miles).
For comparison, in the latest Tesla safety report:
"we recorded one crash for every 4.31 million miles driven", which is roughly what they had 2 years ago.
So Tesla is objectively safer. And Teslas are diven anywhere, whereas Waymo only drove in Phoenix, which is extremely sunny and dry.
Plus the Tesla owner will save the his/her $80k Tesla rather than prove FSD can actually handle or fail a situation.
My impression of Tesla is mostly shaped from (1) nonparticipation in the research community, (2) a very early "mission accomplished" declaration by calling their cars fully self driving, and (3) a longterm refusal to use LIDAR.
I don't consider Tesla a player in self-driving (edit: self-driving research), but I don't think Tesla does either. There's no reason for them to try to "win" the space.
From Tesla's side, it makes more sense to continue on their current tack: Applying results from existing research. I think Tesla's strategy is to be the highest bidder when it comes time for Waymo (or Didi, etc) to sell their tech.
> I don't consider Tesla a player in self-driving, but I don't think Tesla does either
I think that's the distinction driving the confusion here.
EDIT: Sorry, and to clarify, I meant "not a player in self-driving research. I also do not think anyone has any vehicles we should call "self driving".
(I'll keep the rest of my pre-edit clarification below.)
To clarify further:
Tesla's offerings come from applying and engineering existing published research. That takes work, and they're making some money from that.
To the extent that "fully self driving" is an achievable goal, it makes no sense to expect Tesla to make the advances that get us there, when (1) they aren't doing that, and (2) they don't need to do that to make money.
To make this even more clear, let's make it concrete with one plausible future: In 2032, Waymo (or Didi, whoever) achieves true 'level 4' fully-self driving with proprietary technology. Their tech is seen in trucks, busses, taxis, as well as being equipped to a few thousand private vehicles. The safety stats are superhuman, and insuring such a vehicle is cheap.
In this future, Tesla Motors would like to enter into an exclusive partnership to integrate this technology into the cars they manufacture.
Google has tried and failed at commercializing similar technology in other verticals (like building environmental automation from their DC tech). The reason is actual incumbents (rightly) see the value of their position while Google comes at it as “our AI is the value, you just make dumb things”.
I expect the automakers to ship mediocre stacks that are put together by existing players like Bosch.
As an ex-Googler I would be floored if Waymo actually lands a sell-into deal with an automaker. A fully vertical taxi service is their path today because they tried and failed to sign any partnerships.
august 2020, tesla friendly blog: "Waymo Has Partnerships With Fiat Chrysler, Jaguar, Nissan, Renault, Volvo, & Magna"
https://cleantechnica.com/2020/08/10/waymo-has-partnerships-...
I see no possibility where existing large automakers will let Google (Waymo) own the FSD stack (and thus own the data and huge piece of value chain) and be relegated to a mere maker of a dumb car.
And if GM is not going to buy Waymo tech, Tesla will definitely not.
So, I still have difficulty reading your comment and extracting a meaningful point from it. Any reply to you would be regurgitating things that have already been said.
More importantly, I've been trying to interpret your comments as made in good faith, but I don't think you're arguing with the contiguity of the thread in mind. It really seems like you're here to defend Tesla, and that "Tesla might buy Waymo tech" might sound like an insult to Tesla? (To this extent, I'm worried this might actually be another proxy Musk argument, which I really have no interest in.)
With all this, out of courtesy, I don't get HN notifications and I won't be navigating back to this comment thread to check for further replies.
R&D for commercialization is a very difficult and very profitable task! But one would not expect this to advance the field of autonomous vehicles, nor would it require participating in the research.
(Also, I thought Waymo had partnerships, but I'm not close to Google. I don't know a lot about Waymo as a business, only as a research entity.)
A huge amount of R&D is never published. Eg: lithography tech at ASML.
Technically speaking: yes, they were the first ones, back in 2017.
Their car might've been more likely to crash then actually end up at the desired destination, but they did release first with a pretty hilariously bad product.
They doesn't necessarily mean that they themselves believe their own marketing.
AFAIK, they aren't working on self-driving or trying to advance research there, so it doesn't make sense to compare them to Waymo either.
But an open-source driver-assist upgrade package is interesting, but doesn't overlap with my experience. Sorry I don't have anything more meaningful to say!
[1]https://electrek.co/2022/12/14/tesla-full-self-driving-data-...
Do self-driving cars model, learn from, or both, from other cars? Say there's a new obstruction (traffic cones around maintenance crew). When deciding what to do, eg left or right, does the self-driver observe what the proceeding cars did?
Do self-drivers remember prior decisions? On the daily commute, there's a speed bump, pot hole, or whatever. Does the car anticipate the remembered road feature? Like maybe "Oh, last time I saw this pot hole, I had to swerve right. So this time I'm going tack right a little earlier."
Sorry for noob questions. Imagining a ubiquitous self-driving future, I keep thinking of boids and uncoordinated collective action, like flocking and murmurings. Wouldn't it be cool if cars did similar stuff?
Now, 2 billion dollars and 7 years later, they are still only in a handful of small regions with limited numbers of vehicles.
That tells me there is still some fundamental issue that is hard to solve. I wonder why they aren't more transparent and tell us what that issue is that they've been battling for 7 years?
Imagine if every time you started your car, a robotic voice said "Welcome to your Ford Pickup XYZ model. Please ensure your seatbelts are fastened. If you are too hot, you can adjust the climate with the climate controls. If you want to lower the windows, please don't put your arms out. etc etc. Have a nice ride today in your Ford(tm) Pickup(tm).".
We're a long way from the ideal of "show up at the airport 5 mins before, hop on a plane, and hand cash to the pilot for your ride".
Also known as the private jet model. I've experienced it, through a bit of lucky employment, but I remain too poorly funded to make it my normal method of air travel. I do wish I could afford it.
I seems like the hardest 90% of the work is the last 10%.
I don't see how an automated car with no driver can deal with the "are they crossing or not?" question that you get every few minutes while driving in a city. Both because body language is a hard problem to get right, and because there's a lot of non-verbal communication that a driverless car doesn't have a way of participating in.
[1] https://blog.waymo.com/2022/02/utilizing-key-point-and-pose-...
(I appreciate 2019 feels like 7 years ago)
Suppose the collision avoidance is perfect. Now put it on the roads of NYC, or New Delhi.
There are a lot of people who will just walk in front of a car going 40mph, if they know for sure it will brake hard and stop.
The problem isn't technology, it's humanity.
The solution is to change the rules of the road, have protected lanes for self-driving buses and taxis and cars, and enforcement.
Let vehicles that can take full advantage of communicating with each other and the road go fast and use the infrastructure to maximum theoretical capacity, without having to worry about dumb human drivers.
We reached this conclusion about 150 years ago and came up with rails. In addition, you get cheap electricity so reliably that modern trains don't even bother having batteries.
Yes, rail lines as they are deployed now might not be the ideal future proof solution, but something similar which allows 'cars' to go off track for the last mile but otherwise not incur wear and tear on your own tires and engine/transmission for the long haul might be a practical idea.
I think the solution here is to issue tickets to those people. You could probably ticket them already under some statute like "endangering road users" or something.
With self driving cars having always on cameras, you only need to ticket each idiot once or twice, and they'll stop doing it.
We already punish people who run around on the runway of airports - seems no different.
It would suck to be in a world where the only way to do self-driving is indistinguishable from the Uber or taxi service we already have (and likely wouldn’t even be cheaper if it’s proprietary to one or two mega-companies who can extract nearly all the productivity surplus from this as monopoly rents).
i may be biased since i use public transit or bike for everything
In general I think the trend of personal car ownership is something that will become somewhat of a hobby rather than a daily necessity, even outside of cities as long as Waymo (and others) are able to actually achieve their ambitious goals. The only way I see that reversing is if people are forced to live out of their cars due to absurd home costs, which is a very very bleak future.
There’s options with fixed parking spots, and services that allow pick up and drop off anywhere.
You tend to structure your life a little different once you don’t own your own car anymore, you start to think twice about little trips you would’ve done otherwise. On the flip side I now have access to 5 different types of cars ranging from small to big (vans) from my phone. It doesn’t even require that much more planning considering it’s reached critical mass around here and there’s a ton of cars available.
The biggest player around is profitable too, so it’s not going away any time too. It’s saved me thousands and spared me from so much hassle surrounding car ownership. I consider myself an enthousiast but I just got a motorcycle for the weekend instead, pennies on the dollar compared to a car.
All in all I notice I’m just happier not being in a car all the time anymore, you might consider it your safe area but it might as well be a golden cage at times.
I understand it’s different once your throw small children in the mix so it might make sense there, but the reality is that a lot of people could do with a lot less car at most points in their life.
If you are doing longer travel, using the train is actually awesome. Those trains have actual places for children to play in.
If you cities and that goes for small cities as well are properly designed its very possible. Its just that in the US cities are literally designed so as to make it impossible.
It is good to recognize this. A very large portion of the population does not live somewhere that makes good sense for pervasive public transit, walking, or biking for regular transportation needs. And many of those people actively don't want to live somewhere like that. Personal vehicles have a use case, and that does not become invalid just because it does not match your own preferences.
That's just a different way of saying that you never have any cargo (children, groceries, etc) to move.
You should also bear in mind that not everyone wants to live in such high density that everything you could ever need is 5 minutes away on foot.
We have plenty of cargo (children and groceries) and do not own a car.
I also expect long term self driving cars will be safer than humans, and as a person that primarily walks around instead of driving that's a benefit to me even if I'm not in the car.
In 2018 this would be a really good argument. What does Tesla do better now, compared to another modern purpose built EV, for example a Ford Mustang Mach E, or a Hyundai Ioniq 5, Kia EV6, etc?
I struggle to identify any particular feature I would say they are better at, much less something that would make it the best mass produced EV. I say this as a two-time Model 3 owner, having just bought the most recent one two weeks ago. I don't quite have buyers remorse yet, but it's nagging at me that I may have just made a foolish choice for the wrong reasons.
As someone in the EV market I almost went for a Tesla for this reason.
They are not obvious winner among EV cars currently. They were first to do actual high end EV car and that vision changed the market back then.
Tesla also used to have reputation of being well build cars few years back. I don't know whether their quality went down due to mass production or whether it was just not known about them.
Some other comments on this thread suggest that self-driving is only perceived to be unsafe, but is statistically much more safe. Unless the stats improved significantly since then (and what specifically achieved that?), and an independent analysis can agree, I'm not trusting corporate PR.
The true number is somewhere between this worst case and the numbers Waymo presents.
Most driver interventions that I've seen on video were not narrowly missed accidents; they're the car being confused by road construction, a double parked driver, pedestrians spilling onto a street etc.
I have also seen (for Tesla at least), videos of driver interventions that definitely would have been accidents if the driver hadn't stepped in.
I definitely agree with your point that I'd love to see more in-depth figures and the CA DMV might release those more detailed figures? I'm not sure.
EDIT: Sure enough, the CA DMV disengagement reports list the exact cause (https://www.dmv.ca.gov/portal/vehicle-industry-services/auto...)
I don't think it's a realistic expectation that drivers are always fully attentive and able to respond in time to a crash situation when the self-driving mode is active. Such a light would give me the heads up that I should stay on my toes.
"Will this car suddenly decide to change lanes?", "Not likely, it's self driving and there aren't exits or major changes in traffic in front of it".
There will also be quarks in any automatic system. Learning and then predicting these will make the streets safer and more comfortable. For example, perhaps self-driving cars are overly cautions around some local crosswalks. If I'm behind one of these things in the winter I might be aware I should leave even more extra room for the sudden slow downs that other cars will be less likely to do. If I'm smart this wont be the difference between an accident or not, but it will make for a smoother ride.
My larger question is, will proliferation of AI cars increase or decrease net traffic flow. People seem to be driving larger and larger cars slower and slower anyway, so maybe this target is achievable, but this all worries me.
Free highways being a huge issue. Not having car registration be based on weight. Far higher requirements to be allowed to drive. There is a long list.
If you reason is safety, cars are the issues, not bigger cars.
That's a lovely straw man you built. I've not met anyone who bought an SUV for that reason. 99 times out of 100 it's for the utility, especially the third row. Which, if you ask me, isn't as useful as people think when they buy it, but still, it's a big factor when you expect to be regularly driving the kids around along with their friends.
https://www.theguardian.com/cities/2019/oct/07/a-deadly-prob...
Want a minivan? Good luck.
Station wagon? You need a time machine.
I suspect the cyclist in the video is not a $500k/year ML engineer, it's a $50K/year veteran trying to stay out of the welfare line.
For people it is asking for them to give up control either as drivers or as pedestrians in a city. Most people like to have a sense of control of their lives. (My take anyways. I know lots of people like having tracks in their career and progress but I would argue they still feel in control.)
For the whole truck shipping lobby on the other hand it will be: save money. (I know that is a dystopian take, but I find me more likely for trucking or something similar to be the first fortress to "fall")
Take people passing a bike in a bike lane for example. Some will go way under the speed limit, and not just pass the bike even though it's in it's own lane. Or others will swing WAY wide to pass them, even though again, they are in their own line.
Will AI itself learn that indeed, bikes do suddenly veer into your lane, or will AI in fact learn that no, they don't do that?
These tests are nice and all but what is going to happen when a situation invariably arises where someone is going to get hurt? What's going to happen then? Is the car going to opt to protect the passenger (because customer) or the pedestrian (much more likely to be hurt/killed).
If the car is, for instance, tracking likely pedestrians that will enter the road and slows down for safety when they’re near, someone would have to irresponsibly run into the road. If the choice is to hit them or turn over a cliff, I’d expect the car to hit the careless pedestrian.
Long after self-driving systems are superior to human drivers on average, the headlines will still scream about humans being killed by self-driving tech. The sensationalism will still sell and people will still be very outraged about it.
The expectation will be no mistakes. Anything short of that will always draw a hyper emotional negative response, which will lure in political/regulatory responses.
For starters, that's not the correct metric. Self driving systems have to at least surpass the median driver, not the average (mean). Auto-related fatality stats are heavily skewed by a small subset of drivers who engage in very risky habits.
> Auto-related fatality stats are heavily skewed by a small subset of drivers who engage in very risky habits.
Right and it would be great if those people used self driving cars instead.
Because you have to convince people like me to buy a self-driving car, and as long as that car is more likely to get me killed than I am, my family will remain in a car that I drive. I do not drive drunk, I avoid driving in inclement weather when not required, or at night, or when I'm really tired. I don't race, I don't road rage, I am a very defensive driver. I have not had an at-fault accident ever (in 30 years and counting since I got my license) and the only accidents I've ever been in at all were minor fender-benders.
So convince me why I should endanger myself so that you can have an unsafe computer driven auto on the road?
> Right and it would be great if those people used self driving cars instead.
So make a self-driving car for them. You will need to subsidize it, since these types of drivers are more likely than not unable to afford a fancy new toy. When the technology can finally cross the median point, then we can talk again about regular, good drivers hanging up their keys.
Will they?
I mean for example Tobacco companies lied and the truth that we know today is that smoking is very very detrimental to their health. It's also detrimental non-smokers in society via second hand smoke, and secondary effects like cigarette butt litter. It doesn't even provide any solid utility like transportation does, it just feels good.
Not only do people still smoke today, people _start_ smoking today given all the information we have.
So when I see behavior like that, I'm not confident that people won't want FSD just because it's 'dangerous'.
FSD will be incredibly convenient, which means humans will always be motivated to come up a reason, valid or otherwise, that justifies their own use of the tech while allowing themselves to condemn others for mishaps incurred doing the exact same thing.
"They didn't maintain it correctly." "They didn't listen to the warnings." "They bought the wrong brand." "They weren't current on software updates."
Someone posted upthread that current fatalities on something stupid like 1.5 per 100000 miles.
Humans are currently ahead in the safety stats game.
How to crash an airplane – Nickolas Means
https://www.youtube.com/watch?v=099cHWSbAL8
Reflecting on the video, they don't need to save everyone, but it certainly should be the first time when not everyone dies.
While I appreciate the spirit you're being overly glib, and we should all be wary of overly simplistic answers since reality can have surprising emergent effects. Like, your proposal as written would give behind manufacturers who cut corners and were bad an effective veto over companies doing well. If say Tesla finds they have to pretty much start from scratch, that they're 5 years behind Waymo now because they took a shoddy haphazard approach, well they might as well create a bunch of impossible tests that Waymo can't pass. It doesn't matter that they can't pass them, because they couldn't pass the ones Waymo proposes either, so this way they could throw a spanner in the works and pull back on the leaders. You could alter your suggestion to be that manufacturers can only propose tests that they themselves can pass with ready-to-ship vehicles, so everyone has to meet each other's standards. But what if there really are tests that everyone should pass but that no one can yet? Or what if there is explicit or tacit collusion in the other direction, where everyone low balls the tests because ultimately it's more profitable to get stuff shipping?
Basically I don't see any reason to not just have the government continue to be involved here and come up with independent road safety standards as advised by their own experts, with public comment and rationale. Ultimately it's the public interest at stake and the rules are about use of public infrastructure. Why not just have an aggressive federal standard course and set of tests that everyone must meet?
I also think in terms of incentives that FSD car manufacturers should be fully liable for any accidents caused by the car while FSD is active, simple as that.
See the millions of examples of trolley problems, for example.
I'm willing to negotiate. If companies don't want to release their code, they can put up $50M in a trust, and $1M bond per car, and automatically accept at-fault for any fatalities involving one of their self-driving units, with a minimum of $1M paid out for each deceased to their survivors.
"NIEON is defined by (1) gaze being directed through the windshield toward the forward path during the conflict and (2) a lack of sleepiness and intoxication-related impairment."
You can use this for comparison, as in "what would NIEON do in this situation"?
The goal should be less cars overall specially in cities. Technology that will cause more cars and more traffic is just terrible.
Cities should ban cars, make it pedestrian and bicycle friendly, build 'self driving' metros and S-Bahn style train systems.
Large scale pedestrianization of cities is a gigantic for for improvement of the world. Both in terms of less accidents, cleaner air, more attractive city, more efficient transit, increase health from bicycle use and walking. I could go on.
So the program should be ban cars from anything that could even remotely be consider a core part of the city, massively slow cars down around the city, remove all on-street parking on all public land, eliminate all minimum parking requirements. The list of policies goes on.
> self-driving tech could make our cities far more "people" friendly
Having cars with 1 person use up the most valuable space in the world is bad enough, doing it with cars that have 0 people in them is worse.
Self driving at best should be for a very small % of trips, maybe less then 1% in very few special cases, maybe for old people, people with disabilities and stuff like that.
Can you give an example of a such a city (other than the obvious examples of having sections of the inner core of a city where car usage is restricted, which is pretty much the norm in many older European cities)? But it doesn't make any sense to build (or even maintain) a city with roads designed for cars, then not allow them to be used for that purpose. If the roads and other car-based infrastructure were re-purposed in a way that made them clearly unsuitable for cars, again, as I said, you don't need to ban them - it just wouldn't be feasible or practical to use them for the vast majority of people.
> a city with roads designed for cars, then not allow them to be used for that purpose
Actually, walking, bikes, buses and commercial vehicles can use them just fine without personal cars being there.
You can also put stuff there, like restaurants can expand. You can make use the space in lots of different ways.
Yes eventually you want to move away from asphalt streets towards something more reasonable (and more environmentally friendly) but there is lots of things you can do on asphalt.
> it just wouldn't be feasible or practical to use them for the vast majority of people.
Actually removing cars and giving priority for walking, bicycles and public transport increases the transport capacity of the streets. Its also better for business along those streets as has been proven in study after study done on pedestrianization.
I think you missed the point. People drive cars on roads. Stop building, and start removing, roads and people no longer drive on them. There is no need to go through the process of banning cars if you just change how and what infrastructure gets built.
You can't just remove roads in the middle of cities. Like what does that even mean?
You just block of cars and let people retake those spaces.
https://en.wikipedia.org/wiki/Freeway_removal
Plenty of other info online describing various successful transformation projects that basically involved getting rid of car-based (or worse, car-exclusive) infrastructure. My main point is that it doesn't need to be replaced with something where cars aren't allowed at all, just with infrastructure that gives equal priority to all modes of getting around. And it's just not removal, I'd hope to see that as a guiding policy for the building of new urban developments too (sadly we're still someway off that in Australia - most of that occurs on the city fringes and further entrenches car-dependency).
If Tesla gave all the data to Waymo, Waymo would reach L5 in no time.
The self driving AI can gain 100 years experience in just 1 day using simulation: https://blog.waymo.com/2020/04/off-road-but-not-offline--sim...
If simulated data were enough, why _hasn't_ Waymo launched in placed like NYC or Boston?
Waymo just went live with full driverless for the public in Phoenix. Tesla can't do full driverless so I think that evidence settles the debate of simulated vs. data collection, unless you have evidence otherwise?
But in fact that isn't actually why 10 billion $ were invested.
You still didn't answer the question - where's the evidence?
Elon love/hate is a powerful force.
> One of the scariest parts about Tesla is that they don't even seem to know what they don't know.
I'm curious what the reference here is? Again, are you taking coverage of Waymo's test environment as evience of its absence at its competitors?
Waymo describes in detail how they test their algorithm against specific scenarios and makes statements about those experiments.
One company is trying to sell you cars and telling you what you want to hear, the other company is extremely careful with their statements.
How? The only problem with this page[0] is that they haven’t released 2022 stats. Otherwise:
> To ensure our statistics are conservative, we count any crash in which Autopilot was deactivated within 5 seconds before impact, and we count all crashes in which the incident alert indicated an airbag or other active restraint deployed.
This is a press release, not a safety report.
And those are the mistakes that they were willing to show! To be clear, I mean situations where the driver needed to take over but either didn't, or didn't in sufficient time to avoid an illegal or dangerous maneuver.
Having said that, he's also been really clear that the video doesn't always make it obvious just how many aspects of FSD are just plain weird. Even when its not actively failing, it moves in odd ways that are uncomfortable.
He's also said that earlier versions resulted in curbed wheels.
Tesla's cameras have huge blind spots near the car and aren't stereoscopic in all directions. They had ultrasonic sensors at one time, but those had blind spots and major limitations too.
GM has Milford and Yuma + others, Waymo has Castle, Zoox uses Altamont Raceway, Nuro has a track at the Vegas speedway. I can't imagine Tesla's small Fremont loop and their winter track in Fairbanks are sufficient.
Do we actually know that this is true? October of last year "internal QA" found regressions on an already public release and they rolled out an update in less than 24 hours after that was published. Both the fact that it was already public when QA found an issue and that they were able to push new public version in such a short time seems to me like they don't necessarily have a release gauntlet for each version, or at least not a very robust one.
https://www.cnbc.com/2021/10/25/tesla-rolled-back-fsd-beta-v...
Tesla cars are selling faster than they can make them and adding a Radar will slow down the manufacturing, not make it faster.
Tesla told the FCC that it plans to market a new radar starting next month.
https://electrek.co/2022/12/06/tesla-radar-car-next-month-se...
If this is true, why are they reportedly cutting production and offering incentives? If they were booked solid, no incentives would be needed.
https://www.reuters.com/business/autos-transportation/tesla-...
> Radar will slow down the manufacturing, not make it faster.
? They are adding radar, though: https://electrek.co/2022/12/06/tesla-radar-car-next-month-se...
I just don't believe that you can infer the dimensions of objects using stereoscopic images with a reliability that you need to make FSD work.
And also having a computer behind them that deeply understands what an object is, the forms it should take and what its expected behaviour should be. We don’t ever confuse billboards for real people.
Fair but to compensate the car has many more cameras than we do.
On top of that these cameras can't move, can't be re-aimed, and have generally far worse dynamic range than our eyes.
Also, when you're in the car, you probably move your head more than you realize. Moving your head around and looking around also gives you better understandings of distance. Far more than a couple of cameras feeding a few megapixel images into a ML model.
People can scream bloody murder till the cows come home, but Tesla for all its faults is king of the hill - literally more valuable the rest of the car industry combined. Musk is the invincible and irrespective of any shortcomings he's literally the richest person on Earth.
Tesla's approach is demonstrably the best approach in the court of the customer, irrespective of what we (others on the road) think or are put at risk.
Not anymore https://www.bbc.com/news/business-63963239
Tesla's market share is eroding every year, and is expected to decline to less than 20% by 2025.
https://www.cnbc.com/2022/11/29/teslas-dominance-of-evs-is-e....
There are plenty of people like me, who are choosing non-Teslas when they buy EVs.
This utter disaster of a situation leads to 10000s of deaths that has well known solutions.
But instead of preventing these deaths with well known low tech solution, a super expensive technological holy grail is gone somehow fix those problems.
[1]https://twitter.com/TaylorOgan/status/1602774335244177408?re...
Consensus from informed industry observers who are sounding the alarm bells about how dangerous Tesla's approach is from a first-principles standpoint
The "data" Tesla released is a meaningless attempt to pull the wool over your eyes.
The same Tesla who argues to the California DMV that FSD is a level 2 system to dodge reporting requirements? The Tesla who has a major EV outlet demanding that they release meaningful disengagement data? [1]
Perhaps I missed something but I really have no idea what data you could be talking about, unless by "data" you mean access to the FSD beta.
[1] https://electrek.co/2022/12/14/tesla-full-self-driving-data-...
Find anything on Waymo. The NHTSA released autopilot and AV vehicle crash data in raw number of crashes, and I wanted to normalize both Tesla and Waymo somehow, and I couldn't find anything from Waymo other than some repeated claim of "over 20 million miles" which dates back years. If it is only 20 million, then Waymo is insanely dangerous, and no one should be defending them, but I have a feeling they're into the hundreds of millions at this point. Also, look up similar YouTube experiences. Waymo has almost nothing--a few people following their vehicles
The reports they have released have been pretty rudimentary and therefore misleading. I'm glad they released the data when they did and they should keep doing it and expand it to FSD beta with disengagements: https://twitter.com/Tweetermeyer/status/1488673180403191808
> you can find endless videos of user experiences. I said thousands of hours before, but it very well could be hundreds of thousands of hours.
I don't really give Tesla all that much credit for allowing their users to publish video of them using the software and hardware that they paid for, especially when they have asked users to consider Tesla's reputation when choosing what to publish. Further, while videos can reveal interesting qualitative information, they are a really bad substitute for the quantitative data that Tesla is already sitting on and choosing not to release.
Reports for every AV company that tests in CA (except Tesla): https://www.dmv.ca.gov/portal/vehicle-industry-services/auto...
With newer algorithms, you can stitch multiple cameras to reconstruct a scene and semantically identify important objects. I do like what Tesla is doing with its new voxel neural net.
Yeah it will take a few more years, but the hard problems are doing what our visual cortex does with two cameras (our eyes). Also keep in mind a car has multiple cameras around to give full 360 view. They also sense in IR range that our eyes do not.
LIDAR may be expensive now, thats not to say it won't get cheeper, with fewer parts.
But the main issue is a social one, I don't think the public will except any fatalities as a result of a vision based system making a mistake. Even if the deal rate from accetends is lower than with a human driver.
A self driving car can't be 50% better, or even 200% better it needs to be thousands of times better than a human driver for it to succeed. I think the certainty of a LIDAR baed system is the only way to do that.
An AI based vision system is Tesla saying that can make a "brain" better than a human at understanding vision.
LIDAR is saying, humans can't measure how far away things are, we can do better than humans by doing something they can't.
Newer lidar units are orders of magnitude cheaper, and solid state. If I could only choose between a camera and lidar, I'm going with the latter.
I often wonder if these companies are truly trying to revolutionize driving, or just trying to put a couple of "Johnny Cabs" in the southwest and call it a day. Their strategy really does seem geared towards the latter outcome.
This is precisely the type of presumptions I'd like to see tested. I mean, if you're going to go to the trouble of all this and have a 118 acre road course, it seems the height of hubris to just say "well, the sensors will probably be better than cameras at night."
It's amazing to me that on Hacker News people people have this puritanical embarrassment over obvious technical questions.
Night is very concerning and is a problem for computer vision systems.
During the day illumination is fairly consistent. Sure the sun moves around, but it's not a spotlight. At night, illumination varies a lot. A detector that works during the day may not work at all at night. This is no joke, there are papers that show dramatic performance losses for tasks like pedestrian detection at night.
Cameras are worse at night. They need to be more sensitive which dramatically increases their noise. They may need to have longer exposures leading to blur, which is of course made worse if you are moving the camera.
Headlights also don't provide the same visibility so your reactions must be faster. Reflections are a big problem too. Oncoming headlights are also an issue.
Nighttime testing will be critical.
This is solved by having multiple sensors. Nighttime is really not a problem, especially given the broad spectrum and LIDAR.
I think it's worth putting them into different classes, don't you/
It's like criticizing the Wright Brothers, saying stuff like "meh, they barely were off the ground, who cares"
I suspect we’ll start seeing more training and testing in places with less nice weather as time goes on.
One example: https://arstechnica.com/cars/2017/10/waymo-starts-testing-in...