Interestingly, due to the required reporting by Waymo and Cruise, we effectively have a highly accurate record of the ability of normal drivers. Obviously, these records may be slightly biased by the driving style of a driverless car; however, but when stopped they are indistinguishible from cars with drivers. Waymo was hit 17 times while stationary, compared to hitting 2 stationary vehicles—this is a vast improvement in safety!
Nice post. The CMU professor Phillip Koopman brings more rigor to comparing AVs to humans and deciding when to launch. He wrote a book on it! If you’re looking to dive deeper, check out this 1hr talk on YouTube https://m.youtube.com/watch?v=UTdR_HE3DDw
I think that while you are definitely considering this, you still aren't seriously _enough_ considering the heterogeneity of human driver behavior. Waymo/Cruise vehicles are (a) not driving in rural settings (b) not driving in bad weather (mostly; SF and Phx don't have a lot of bad weather) (c) always driving expensive, late-model cars (d) are driven at fairly low speeds. Combined with the factors you do mention, the human crash rate and even more the human serious injury rate for that situation is most likely very low, but crucially we have no real idea what it is. Giving every driver in the US a Jaguar I-Pace would almost certainly reduce the accident rate a bunch too.
Of course, how relevant that is depends on what question we're trying to ask about self-driving (eg, is Waymo increasing the safety of ride-hail in SF?). The Waymo study of high-severity accidents in Phoenix is useful for that, but there are still a lot of confounders.
One consideration that I haven't seen discussed in safety articles about self-driving cars is induced demand. If self-driving cars are safer than human drivers but still substantially more dangerous than transit, and the availability of self-driving cars causes people to travel more passenger miles, the absolute number of car crash fatalities could still rise.
I agree with you that Waymo seems to be doing a better job than Cruise. From records for yearly miles driven with and without a test driver, reports filed, news articles, and responses from company leaders, the overall picture appears to be that Waymo has proceeded more gradually while Cruise has less regard for their impact on others.
On the other hand, you’re giving short shrift here to the difference in conditions under which these companies miles have been accumulated vs human drivers. Both companies have been heavily restricted in their operation with respect to location, time, speed, and weather while human drivers are not. We might find a far better safety record for human drivers in city conditions if we lowered speed limits and gave substantial fines for speed violations, as discussed in this article comparing driving in Finland with driving in the US. (https://heatmap.news/politics/helsinki-cars-pedestrians-bikes-finland ) With an increase in driverless vehicles we are likely to find new classes of safety issues, such as those that arise around autonomous vehicle confusion in construction or emergency scenes. While the cars in those situations may not actively hit someone, they can nonetheless be responsible for worse outcomes for those impacted - consider this report filed with SF regarding a delay to an ambulance which a Cruise vehicle caused this August. (https://journa.host/@cfarivar/110985056243205061 which links to https://www.forbes.com/sites/cyrusfarivar/2023/08/30/cruise-robotaxis-waymo-san-francisco-firefighters )
That last relates to two big issues I see with the entire discussion around vehicle safety. First, we talk about the safety of a car or of a driver without considering that safety is a system property. It is not the car or the driver that is “safe” but they are able to operate safely in some location under some set of conditions. None of us is truly safe driving in a blizzard; we are even less safe if the side of the road we’re driving in a blizzard is a cliff. We blame drivers for accidents when the road, signage, or vehicle have design flaws. And second, vehicle safety is considered only with respect to vehicle occupants, despite those vehicles having an impact on people who are not in them. I was surprised when I read the CPUC and DMV safety requirements for Cruise and Waymo because the focus was almost entirely on passengers. To me, that seems like dereliction of duty; our regulatory agencies should be considering safety for all, not just safety for those in vehicles. That’s a system-wide regulatory problem, not just one for driverless cars, but the issues arising in SF are pointing up some of the shortcomings.
Isn't this argument (that we can be relatively confident they're safer than humans already) pretty sensitive to the actual numbers around "how much safer/less safe is SF driving than the weighted average of human miles" and "what fraction of accidents are caused by drunk people/teenagers/road raging people/old people"? Even for Waymo, the expected number of "serious accidents" under the null of no-better-than-humans is below 10 (and since serious is so heavily correlated with speed, I'd guess that SF is better than average?)
A much more minor point, but I don't think "the read-ending driver is always at fault" works as a way of describing the world in big cities, even if that's how the law works
Thanks for sharing. I am not sure that one crash for every 60,000 miles on average sounds very reassuring. You do make a good point though on why the experiment should keep going.
Driverless cars are almost undoubtedly safer drivers than me, not due to poor skill but the opposite. I am a very skilled driver and MOST of the time- a defensive one.
But when someone drives dangerously and it sets me off and they react poorly to my warnings- I become an insanely dangerous, aggressive, all round horrible driver. I doubt you've ever seen someone get out of their vehicle at red lights more than me. I think I'm addicted to the feeling now. The road is far better off without me than with me (well that's true about everyone but this isn't an article on why you should take public transit and bike lol).
Self driving cars can't come soon enough. I'd volunteer for a program like that in a heartbeat. It's why I try to cycle whenever I can, not just for the environment but because I don't stay as angry for as long. Unfortunately driving on the road with a bike is quite dangerous, so I don't do it super often and when I see a cyclist on the road and I'm not in a rush (not frequently, thanks AutDHD) I'll throw my cruise control on behind them, throw them a thumbs up and just enjoy being a big meat shield for them. People bully other vehicles less than cyclists so I like to think that I'm keeping them safer.
If they don’t already, couldn’t these services utilize remote human intervention in some of these non destructive corner cases? For example, say in case of accident, obstruction or traffic jam causing scenario, an alert is triggered and some remote human takes control of the vehicle temporarily (to get out of intersection, or to side of the road, etc.)? And so you’d have support of a human backup driver for every X self-driving vehicles on the road?
Much promise imho for the significant number of A to B routes that are highly standardized, and less standardized routes could be geofence restricted or given more precautions…
See for example work on airbags by George Hoffer (Testing for Offsetting Behavior and Adverse Recruitment Among Drivers of Airbag-Equipped Vehicles): "Earlier studies reported that an insurance industry index of personal-injury claims rose after automobiles adopted driver's side airbags and that drivers of airbag-equipped vehicles were more likely to be at fault in fatal multivehicle accidents. These findings can be explained by the offsetting behavior hypothesis or by at-risk drivers systematically selecting vehicles with airbags (i.e., adverse recruitment). We test for offsetting behavior and adverse recruitment after airbag adoption using a database containing information on fatal accidents including information on drivers' previous records and drivers' actions that contributed to the occurrence of the accident. Further, we reexamine the personal injury claims index data for newly airbag-equipped vehicles and show that the rise in the index after airbag adoption may be attributable to moral hazard and a new vehicle ownership pattern. "
Who is considered at fault in these accidents , the only thing I could find "California, the law explicitly states that “an operator of an autonomous vehicle is the person who is seated in the driver’s seat, or if there is no person in the driver’s seat, causes the autonomous technology to engage.” ... Copywrited 2023 but I hope it is old. If true though the companies or third party could claim , the hiring passenger is liable.
Here's a (heuristic, but working from what we have) argument that this is probably underrating humans driving comparable miles: death rates per 100m miles vary a LOT (more than 3x!) between states:
In particular MA (which features an unusually high number of modern cars driving slowly in cities) is the absolute lowest. If SF driving is comparable to MA (and there's no reason to think of MA as a lower bound, since 1/3 of it by pop is outside the Boston metro), that could be a significantly different baseline
I’m puzzled by your conclusion that letting Cruise operate is a “close call.” It seems like it’s better than human drivers, and even if it was only even with them we benefit from the technology improving. This seems like a sop to the anti-progress forces, but it just makes their positions seem stronger than they are.
Thank you -- well written and clear; quite insightful. I’m in the Middle East; many different cultures and approaches to driving on display here every day. Absolutely no question AVs would make driving safer, just by dint of reducing variation (speed, optional use of turn signals, following distances). Probably also faster travel times if all the vehicles were networked to cooperate on flow and merging, etc.
Interestingly, due to the required reporting by Waymo and Cruise, we effectively have a highly accurate record of the ability of normal drivers. Obviously, these records may be slightly biased by the driving style of a driverless car; however, but when stopped they are indistinguishible from cars with drivers. Waymo was hit 17 times while stationary, compared to hitting 2 stationary vehicles—this is a vast improvement in safety!
Nice post. The CMU professor Phillip Koopman brings more rigor to comparing AVs to humans and deciding when to launch. He wrote a book on it! If you’re looking to dive deeper, check out this 1hr talk on YouTube https://m.youtube.com/watch?v=UTdR_HE3DDw
I think that while you are definitely considering this, you still aren't seriously _enough_ considering the heterogeneity of human driver behavior. Waymo/Cruise vehicles are (a) not driving in rural settings (b) not driving in bad weather (mostly; SF and Phx don't have a lot of bad weather) (c) always driving expensive, late-model cars (d) are driven at fairly low speeds. Combined with the factors you do mention, the human crash rate and even more the human serious injury rate for that situation is most likely very low, but crucially we have no real idea what it is. Giving every driver in the US a Jaguar I-Pace would almost certainly reduce the accident rate a bunch too.
Of course, how relevant that is depends on what question we're trying to ask about self-driving (eg, is Waymo increasing the safety of ride-hail in SF?). The Waymo study of high-severity accidents in Phoenix is useful for that, but there are still a lot of confounders.
One consideration that I haven't seen discussed in safety articles about self-driving cars is induced demand. If self-driving cars are safer than human drivers but still substantially more dangerous than transit, and the availability of self-driving cars causes people to travel more passenger miles, the absolute number of car crash fatalities could still rise.
I agree with you that Waymo seems to be doing a better job than Cruise. From records for yearly miles driven with and without a test driver, reports filed, news articles, and responses from company leaders, the overall picture appears to be that Waymo has proceeded more gradually while Cruise has less regard for their impact on others.
On the other hand, you’re giving short shrift here to the difference in conditions under which these companies miles have been accumulated vs human drivers. Both companies have been heavily restricted in their operation with respect to location, time, speed, and weather while human drivers are not. We might find a far better safety record for human drivers in city conditions if we lowered speed limits and gave substantial fines for speed violations, as discussed in this article comparing driving in Finland with driving in the US. (https://heatmap.news/politics/helsinki-cars-pedestrians-bikes-finland ) With an increase in driverless vehicles we are likely to find new classes of safety issues, such as those that arise around autonomous vehicle confusion in construction or emergency scenes. While the cars in those situations may not actively hit someone, they can nonetheless be responsible for worse outcomes for those impacted - consider this report filed with SF regarding a delay to an ambulance which a Cruise vehicle caused this August. (https://journa.host/@cfarivar/110985056243205061 which links to https://www.forbes.com/sites/cyrusfarivar/2023/08/30/cruise-robotaxis-waymo-san-francisco-firefighters )
That last relates to two big issues I see with the entire discussion around vehicle safety. First, we talk about the safety of a car or of a driver without considering that safety is a system property. It is not the car or the driver that is “safe” but they are able to operate safely in some location under some set of conditions. None of us is truly safe driving in a blizzard; we are even less safe if the side of the road we’re driving in a blizzard is a cliff. We blame drivers for accidents when the road, signage, or vehicle have design flaws. And second, vehicle safety is considered only with respect to vehicle occupants, despite those vehicles having an impact on people who are not in them. I was surprised when I read the CPUC and DMV safety requirements for Cruise and Waymo because the focus was almost entirely on passengers. To me, that seems like dereliction of duty; our regulatory agencies should be considering safety for all, not just safety for those in vehicles. That’s a system-wide regulatory problem, not just one for driverless cars, but the issues arising in SF are pointing up some of the shortcomings.
Isn't this argument (that we can be relatively confident they're safer than humans already) pretty sensitive to the actual numbers around "how much safer/less safe is SF driving than the weighted average of human miles" and "what fraction of accidents are caused by drunk people/teenagers/road raging people/old people"? Even for Waymo, the expected number of "serious accidents" under the null of no-better-than-humans is below 10 (and since serious is so heavily correlated with speed, I'd guess that SF is better than average?)
A much more minor point, but I don't think "the read-ending driver is always at fault" works as a way of describing the world in big cities, even if that's how the law works
Thanks for sharing. I am not sure that one crash for every 60,000 miles on average sounds very reassuring. You do make a good point though on why the experiment should keep going.
Driverless cars are almost undoubtedly safer drivers than me, not due to poor skill but the opposite. I am a very skilled driver and MOST of the time- a defensive one.
But when someone drives dangerously and it sets me off and they react poorly to my warnings- I become an insanely dangerous, aggressive, all round horrible driver. I doubt you've ever seen someone get out of their vehicle at red lights more than me. I think I'm addicted to the feeling now. The road is far better off without me than with me (well that's true about everyone but this isn't an article on why you should take public transit and bike lol).
Self driving cars can't come soon enough. I'd volunteer for a program like that in a heartbeat. It's why I try to cycle whenever I can, not just for the environment but because I don't stay as angry for as long. Unfortunately driving on the road with a bike is quite dangerous, so I don't do it super often and when I see a cyclist on the road and I'm not in a rush (not frequently, thanks AutDHD) I'll throw my cruise control on behind them, throw them a thumbs up and just enjoy being a big meat shield for them. People bully other vehicles less than cyclists so I like to think that I'm keeping them safer.
If they don’t already, couldn’t these services utilize remote human intervention in some of these non destructive corner cases? For example, say in case of accident, obstruction or traffic jam causing scenario, an alert is triggered and some remote human takes control of the vehicle temporarily (to get out of intersection, or to side of the road, etc.)? And so you’d have support of a human backup driver for every X self-driving vehicles on the road?
Much promise imho for the significant number of A to B routes that are highly standardized, and less standardized routes could be geofence restricted or given more precautions…
I first saw this on Ars. Very interesting. The one thing I noticed? NO Tesla.
See for example work on airbags by George Hoffer (Testing for Offsetting Behavior and Adverse Recruitment Among Drivers of Airbag-Equipped Vehicles): "Earlier studies reported that an insurance industry index of personal-injury claims rose after automobiles adopted driver's side airbags and that drivers of airbag-equipped vehicles were more likely to be at fault in fatal multivehicle accidents. These findings can be explained by the offsetting behavior hypothesis or by at-risk drivers systematically selecting vehicles with airbags (i.e., adverse recruitment). We test for offsetting behavior and adverse recruitment after airbag adoption using a database containing information on fatal accidents including information on drivers' previous records and drivers' actions that contributed to the occurrence of the accident. Further, we reexamine the personal injury claims index data for newly airbag-equipped vehicles and show that the rise in the index after airbag adoption may be attributable to moral hazard and a new vehicle ownership pattern. "
Who is considered at fault in these accidents , the only thing I could find "California, the law explicitly states that “an operator of an autonomous vehicle is the person who is seated in the driver’s seat, or if there is no person in the driver’s seat, causes the autonomous technology to engage.” ... Copywrited 2023 but I hope it is old. If true though the companies or third party could claim , the hiring passenger is liable.
Here's a (heuristic, but working from what we have) argument that this is probably underrating humans driving comparable miles: death rates per 100m miles vary a LOT (more than 3x!) between states:
https://worldpopulationreview.com/state-rankings/fatal-car-accidents-by-state
In particular MA (which features an unusually high number of modern cars driving slowly in cities) is the absolute lowest. If SF driving is comparable to MA (and there's no reason to think of MA as a lower bound, since 1/3 of it by pop is outside the Boston metro), that could be a significantly different baseline
I’m puzzled by your conclusion that letting Cruise operate is a “close call.” It seems like it’s better than human drivers, and even if it was only even with them we benefit from the technology improving. This seems like a sop to the anti-progress forces, but it just makes their positions seem stronger than they are.
Thank you -- well written and clear; quite insightful. I’m in the Middle East; many different cultures and approaches to driving on display here every day. Absolutely no question AVs would make driving safer, just by dint of reducing variation (speed, optional use of turn signals, following distances). Probably also faster travel times if all the vehicles were networked to cooperate on flow and merging, etc.
It would be interesting to see how well a driverless car might fare in a stock car race.😇