> by the lower number of accidents that AVs are involved in per distance travelled compared to human-driven vehicles
Just a note that this number is hard to calculate accurately with an acceptable degree of certainty.
Anyone claiming that AVs are involved in fewer accidents per distance traveled than human drivers is either extrapolating from incomplete data or baking some unreliable assumptions into their statement.
Tesla just announced 500 million miles driven by FSD [1]. Per the video, were it fully autonomous they could have a 95% CI on "safer than human" at only 275 million miles [2], but obviously having human supervision ought to remove many of the worst incidents from the dataset. Does anyone know if they publish disengagement data?
This just shows how statistics can mislead. I own a Tesla with FSD and it's extremely unsafe for city driving. Just to quantify, I'd say at its absolute best, about 1 in 8 left turns result in a dangerous error that requires me to retake control of the car. There is no way it even comes close to approaching the safety of a human driver.
The caveats are missing the point that FSD is very obviously less safe than a human driver, unless you constrain the data to long stretches of interstate road during the day, with nice weather, clearly marked road lines and minimal construction. At that point, my "intuition" tells me human drivers probably still safer, but under typical driving conditions they very obviously are (at least with Tesla FSD, I don't know about Waymo)
The reason why I spent 3/4 of my post on caveats was because I didn't want people to read my post as claiming that FSD was safe, and instead focus on my real point that the unthinkable numbers from the video aren't actually unthinkable anymore because Tesla has a massive fleet. You're right, though, I could have spent 5/6 of my post on caveats instead. I apologize for my indiscretion.
Tesla's mileage numbers are meaningless because the human has to take over frequently. They claim credit for miles driven, but don't disclose disconnects and near misses.
California companies with real self driving have to count their disconnects and report all accidents, however minor, to DMV. You can read the disconnect reports online.
Do you think they lied about miles driven in the investor presentation?
Nah, that would be illegal. Their statement leaves plenty of room for dirty laundry though. I'm sure they won't disclose disengagement data unless forced, but they have plenty of legal battles that might force them to disclose. That's why I'm asking around. I'd love to rummage through. Or, better, to read an article from someone else who spent the time.
Your video and my response were talking about fatal crashes. Humans don't go 100 million miles between crashes.
Has FSD had a fatality? Autopilot (the lane-follower) has had a few, but I don't think I've heard about one on FSD, and if their presentations on occupancy networks are to be believed there is a pretty big distinction between the two.
Isn't "FSD" the thing they're no longer allowed to call self driving because it keeps killing cyclists? Google suggests lots of Tesla+cyclist+dead but with Tesla claiming it's all fine and not their fault, which isn't immediately persuasive.
> Google suggests lots of Tesla+cyclist+dead but with Tesla claiming it's all fine and not their fault, which isn't immediately persuasive.
With human drivers -- are we blaming Tesla for those too?
You do you, but I'm here to learn about FSD. It looks like there was a public incident where FSD lunged at a cyclist. See, that's what I'm interested in, and that's why I asked if anyone knew about disengagement stats.
Can someone summarize the video? That was my first thought as well: crash data for humans is clearly underreported. For example, police don't always write reports or human drivers agree to keep it off the books.
The probability of a human driver causing a fatality on any given mile driven is 0.00000109% (1.09 fatalities occur per 100 million miles driven).
Applying some basic statistics to show to a 95% confidence level that a self driving system causes fewer fatalities than a human you would have to drive 275 million autonomous miles flawlessly.
This would require a fleet of 100 vehicles to drive continuously for 12.56 years.
And in practice self-driving vehicles don't drive perfectly. Best estimates on the actual number of miles driven to validate their safety is around 5 billion autonomously driven miles and that's assuming that they actually are safer than a human driver.
Then you get into the comparison itself. In practice AVs don't drive on all the same roads, at the same times, as human drivers. A disproportionate number of accidents happen at night, in adverse weather, and on roads that AVs don't typically drive on.
Then you have to ask if comparing AVs to all drivers and vehicles is valid comparison. We know for instance that vehicles with automated breaking and lane assist are involved in fewer accidents.
Then of course if minimizing accidents is really what you care about there's something easy we could do right now: just mandate all vehicles must have an breathalyzer ignition. We do this for some people who have been convicted of DUI but doing it for everyone would eliminate a third of fatalities.
Then of course if minimizing accidents is really what you care about there's something easy we could do right now: just mandate all vehicles must have an breathalyzer ignition. We do this for some people who have been convicted of DUI but doing it for everyone would eliminate a third of fatalities.
In a similar vein, if we put geo-informed speed governors in cars that physically prevented you from exceeding, say, 20% of the speed limit, fatalities would also likely plummet.
I'm fine with it notifying the driver it thinks you might be speeding, but I don't like the idea of actually limiting the speed of the car. I've used several cars which had a lot of cases where it didn't track the speed right. Driving near but not in a construction zone. Driving in an express lane which has a faster posted speed than the main highway. School zones. I've seen cars routinely get these things wrong. A few months ago I was on an 85 MPH highway and Google Maps suddenly thought I was on the 45 MPH feeder. Add 20%, that's 54 MPH max speed. So what, my car would have quickly enforced the 30 MPH drop and slam on the brakes to get it into compliance?
I'd greatly prefer just automatic enforcement of speeding laws rather than doing things to try and prevent people from speeding.
Honestly I would think something like transponders on every freeway would work better than GPS. Regardless, I think everyone in the thread could think of 10 technological ways of making this work. I think the biggest barriers are political, not logistical, and definitely not engineering.
Yeah, because need to be able to use your vehicle to escape pursuers and also as a ramming weapon. I assume police would get an exception from this rule, but they don't actually have more of a legal right to use their vehicle as a weapon than anyone else, they are just less likely to be questioned on their judgement of the situation as an emergency by the DA. Probably also a second amendment violation, but the Supreme Court might be too originalist (and not textualist enough) to buy that argument, as cars did not exist in the decades surrounding the founding.
I sustained traumatic injuries a few months ago when a driver on a suspended learner's permit hit me. The lazy cop issued no tickets for the multiple traffic violations. He couldn't be bothered to show up for the trial and the lazy prosecutor who only notified me of the trial three days in advance went with a bare minimum wrist slap for the suspension. It's as if it officially never happened.
That's awful. The amount of egregious vehicular violence the US has tolerated is disgusting. Waymo seems like the best bet to making experiences like yours a thing of the past.
Odometers are pretty well regulated and insurance companies will often have a good record of the readings over long periods. I'm not sure how the org doing the data collection does it precisely, but pretty accurate data is out there.
It would be interesting if some kind of active/inductive measurement could be made.
As a human driver, I'm keenly aware of the "close calls" I've had where I've narrowly avoided a probable collision through good decisions and/or reaction times. I'm probably less aware of the times when I've just gotten lucky.
No doubt self-driving companies have internally available data on this stuff. Can't wait til superior performance to human drivers is a major selling point!
Although things are changing, the overwhelming majority of AV miles are generally-safer miles on protected highways. And yet their statistics are typically compared against vehicle miles on all road types.
Further, most AVs explicitly disengage when they encounter situations beyond their understanding. So again, we’re selecting AV statistics based on generally favorable conditions, but we don’t track human-operated miles the same way.
But also, there are explicitly times and areas where Cruises don’t operate due to being insufficiently able to operate. And times where they do just pull over and wait for human intervention when they don’t know what else to do. Both of which are safe and reasonable, but which also selects themselves into a safer set of miles actually traveled.
Human drivers "disengage" also, though we don't think of it that way. I have aging friends who refuse to drive at night. I'll stay home in bad weather. When I was younger, I'd sometimes ask a passenger to back out of tricky parking space for me.
It's not too useful to lump Tesla, Cruise, and Waymo together here. Tesla is years behind Cruise and cruise is years behind Waymo in terms of driving capability. Waymo doesn't even drive on highways, so we don't know how safe it would be (probably very safe).
>Further, most AVs explicitly disengage when they encounter situations beyond their understanding
the bigger problem here is that the machine may not realize it isn't fully understanding the situation; I think that's the more common situation : a computer situational model doesn't match reality, but the machine has the confidence in perception to proceed, producing dangerous effects.
Not sure that's fair - the number of accidents in human-driven vehicles varies significantly by the human driving it. Do you compare with the teen that just obtained their license and is more interested in their phone than in driving properly, or the 65-year old doctor who's been driving for work and other purposes every day for the past 40 years? According to https://injuryfacts.nsc.org/motor-vehicle/overview/age-of-dr... it's at least a 6x difference.
You compare what matters — mean accidents per mile driven. And like any actuary, you can compare distributions with finer granularity data (location, tickets, ages, occupations, collusion severity, etc). None of this is new or intractable. We can have objective standards with confident judgements of what’s safe for our roads.
As an aside, public acceptance of driverless cars is another story. A story filled with warped perspectives due to media outlets stoking the audience emotions which keep outlets in business — outrage, fear, and drama. For every story covering an extreme accident involving human drivers, there might be 100 stories covering a moderate driverless accident. No matter how superhuman autonomous cars perform, they’ll have a massive uphill battle for a positive image.
I think you compare it to the average as that's who you are encountering when on the road. You have no way of selecting the other drivers to bias toward the doctor.
Teens who’ve just gotten their license are often quite good drivers. It’s after that when they drive poorly. (This also applies to adults, to a lesser extent.)
Well, obviously the number of autonomous vehicles involved in accidents is going to be lower, but that's because barely any of them exist compared to the vast majority of people driving their cars. If you had statistics on proportions though, that might be a different story.
You missed "per distance travelled" - that partially normalises the results. You still have to adjust for situation and road type (which Tesla conveniently doesn't do in their advertised numbers) for a better result.
Just a note that this number is hard to calculate accurately with an acceptable degree of certainty.
Anyone claiming that AVs are involved in fewer accidents per distance traveled than human drivers is either extrapolating from incomplete data or baking some unreliable assumptions into their statement.
Welch Labs has a good introductory video to this topic: https://youtu.be/yaYER2M8dcs?si=XEB4aWlYf6gnnTqM