Live data from Hacker News

Replaying real life: how the Waymo Driver avoids fatal human crashes

blog.waymo.com

21–30 of 182 posts

Re: Replaying real life: how the Waymo Driver avoids fatal human crashes

#21
post #14
post #7

Why limit simulations to fatal crashes instead of all reported crashes?

How could they recreate every crash that's happened? It seems the police only keep detailed records of the fatal crashes that happened there.

"The simulations were not conducted independently of the company, nor were they reviewed by any third-party for verification prior to the company making them public." from https://www.theverge.com/2021/3/8/22315361/waymo-autonomous-...

Re: Replaying real life: how the Waymo Driver avoids fatal human crashes

#22
post #14
post #7

Why limit simulations to fatal crashes instead of all reported crashes?

How could they recreate every crash that's happened? It seems the police only keep detailed records of the fatal crashes that happened there.

This gap in data is a bugbear for cycling/walking advocates such as myself who feel that non-vehicles do an unseen and unfair portion of the work of preventing crashes which might otherwise involve them. That they basically paper over a myriad of bad driver behaviours and ghastly infrastructure designs by always being on high alert and generally getting really good at "jumping out of the way." You can partly tell how effective this system is because when there are pedestrians killed, commentators express regret that the victim didn't just jump out of the way like a normal pedestrian, or imply that it was their fault for other reasons such as wearing the wrong clothing, looking at their device, or choosing an inappropriate time of day to go for a walk. All of this leads to attitude among drivers and politicians that current solutions are fine because the only numbers they have are literally deaths.

Anyway, one would think with SDVs that there'd be an explosion in data for this kind of thing, since the cars will witness and participate in many more close calls than could ever be reported or tracked by anyone, much less the police. I wonder how much Google is able to process their car data for these kinds of things? Potentially it could be a huge aid, not even just for law enforcement, but also for urban planners looking for the close call hotspots.

Re: Replaying real life: how the Waymo Driver avoids fatal human crashes

#23
post #18

You can read the California DMV's set of autonomous vehicle accident reports.[1] Almost all the Waymo reports are "vehicle was entering intersection, detected cross traffic, stopped, was rear-ended by human driver". There's one Waymo report where someone ran a stop sign and hit them. The rear-ending problem will be solved as automatic emergency braking becomes standard on cars. Already, it's shipping on almost all hi…

>The rear-ending problem will be solved as automatic emergency braking becomes standard on cars. Already, it's shipping on almost all high-end cars and 70-80% of midrange cars. There's a US auto industry goal of it being standard by 2022. That's probably the main feature needed for self-driving cars to coexist with human-driven cars. I really wish we could just enforce proper education of being observant. Most times…

I support tackling this problem with both your solutions. But I certainly wouldn't bet my money on yours.

Re: Replaying real life: how the Waymo Driver avoids fatal human crashes

#24
post #16

Figure 3 has the responder slowing unusually soon. Makes me wonder if their model is overly fit to the sample. Still, if it's genuinely that good at responding and mitigating then the future looks bright.

Figure 3 almost looks like the responder is lagging. Feels like it slows down and speeds up in several places. Assuming that’s not a gif/video rendering bug, it’s a great way to get rear-ended.

Here's the semi-annual reminder to everyone that you are probably following too closely, and this is exactly why you shouldn't.

The rule of thumb is that you want 3-4 seconds between cars, which at highway speeds is 80-100 meters or yards.

Yes, almost no one leaves that much room. That's part of why so many accidents happen. It's easy for you to do better.

Re: Replaying real life: how the Waymo Driver avoids fatal human crashes

#25
post #18

You can read the California DMV's set of autonomous vehicle accident reports.[1] Almost all the Waymo reports are "vehicle was entering intersection, detected cross traffic, stopped, was rear-ended by human driver". There's one Waymo report where someone ran a stop sign and hit them. The rear-ending problem will be solved as automatic emergency braking becomes standard on cars. Already, it's shipping on almost all hi…

Solving the automated driving problem will require these systems to behave in ways that regular human drivers consider acceptable - at least for a transition period of time where both live on the road together. Sudden hard breaking in situations that humans see absolutely no need for it is an unacceptable behavior for human drivers to have. Just because it's not your fault doesn't mean you can't do things to prevent it. I have prevented many crashes throughout my life due to my mitigations around others' driving mistakes. We should expect self driving cars to do this at least as well as - not worse than - humans.

Re: Replaying real life: how the Waymo Driver avoids fatal human crashes

#26
post #18

You can read the California DMV's set of autonomous vehicle accident reports.[1] Almost all the Waymo reports are "vehicle was entering intersection, detected cross traffic, stopped, was rear-ended by human driver". There's one Waymo report where someone ran a stop sign and hit them. The rear-ending problem will be solved as automatic emergency braking becomes standard on cars. Already, it's shipping on almost all hi…

>The rear-ending problem will be solved as automatic emergency braking becomes standard on cars. Already, it's shipping on almost all high-end cars and 70-80% of midrange cars. There's a US auto industry goal of it being standard by 2022. That's probably the main feature needed for self-driving cars to coexist with human-driven cars. I really wish we could just enforce proper education of being observant. Most times…

Unfortunately, current cars have to be controlled by humans, and the human brain has certain characteristics (selected for by evolution) that makes it near impossible for it to be able to react to 100% of the driving situations. There are built-in efficiency circuits (similar to CPU branch prediction) that takes immediate action based on what it typically sees.

So if 99.9999% of the time, when a car in front starts to accelerate, and doesn't slam on the breaks immediately after, the brain makes the assumption that the care in front will never slam on the breaks and devotes less brain-CPU time to "Oh No, that care in front just slammed on the breaks" processing.

In general, human drivers don't do this unless they are very inexperienced. You don't mash on the gas and immediately mash on the breaks, instead you make sure there is nothing coming. If you can't see well enough that there is something coming, then you ease up slowly and only mash on the gas when you have enough information that you won't get side swiped going through the intersection.

And yes, intersections that have a "stop line" far enough back that you can't see cross traffic are poorly designed. Or if cars are really expected to stop every time before a crosswalk, then advance 2 feet to stop again, this needs better enforcement so that activity becomes normalized to the point where the car behind would expect that (one way of doing this is to put two stop signs, one before the sidewalk, and one at the intersection).

Re: Replaying real life: how the Waymo Driver avoids fatal human crashes

#27
post #14

Earlier quoted context omitted.

How could they recreate every crash that's happened? It seems the police only keep detailed records of the fatal crashes that happened there.

"The simulations were not conducted independently of the company, nor were they reviewed by any third-party for verification prior to the company making them public." from https://www.theverge.com/2021/3/8/22315361/waymo-autonomous-...

Why is that relevant to my point? How can they do all crashes without knowing the specifics of each crash?

"This research was made possible because of the thorough crash reports the Chandler Police Department and Arizona Department of Transportation (ADOT) prepare following fatal crashes, and we’re grateful to ADOT, Chandler Police Department, and Arizona Department of Public Safety for making these reports available to the public upon request."

Re: Replaying real life: how the Waymo Driver avoids fatal human crashes

#28
post #2

What do you call that when you subconsciously suppress any tests or scenarios you know will break your code? Like you notice somethings kinda racey but it involves other components you don't control and so you don't do the parallel tests you probably should, that kinda thing. Because that is exactly the bad feeling I get with Waymo never actually testing anywhere that doesn't have mm-precise maps and the all importan…

> There seems to be this implicit understanding that when you drop the thing in light snow it will just careen off the side and nobody is particularly incentivized to find that yep, that is exactly what happens.

Most people from Southern states will careen off the side when dropped into light snow--I'm exaggerating, but not by a lot.

If the Waymo car does nothing else short of "maintain appropriate following distance" it is already ahead of 90% of all drivers in all weather.

Re: Replaying real life: how the Waymo Driver avoids fatal human crashes

#29
post #18

You can read the California DMV's set of autonomous vehicle accident reports.[1] Almost all the Waymo reports are "vehicle was entering intersection, detected cross traffic, stopped, was rear-ended by human driver". There's one Waymo report where someone ran a stop sign and hit them. The rear-ending problem will be solved as automatic emergency braking becomes standard on cars. Already, it's shipping on almost all hi…

This does not, by the way, imply that the machines are working correctly and human drivers are the problem. These cars are rear-ended more often than human drivers, which suggests they stop more suddenly or at times humans don't expect.

Slamming on the brakes is a flaw. Even if every car on the road had the ability to respond instantly to the car ahead, being forced to brake hard still risks loss of control especially in bad conditions.

Also, having automatic emergency braking on all vehicles is at least decades out. It will never work on motorcycles or scooters or bicycles, period. It cannot be retrofitted onto existing cars. Requiring it for all new cars would require pretty hamfisted regulations. I expect human drivers to be widespread for at least the rest of my life.

Re: Replaying real life: how the Waymo Driver avoids fatal human crashes

#30

This is really cool. I suspect that when we have self driving cars, it's going to take off very quickly due to insurance. In particular, if you are in a non self driving car and in an accident with a self driving one, it'll be hard to not be shown as at fault. So insurance I would guess go up based on the percentage of self driving cars, and likely be really cheap for self driving cars. I wonder if there's a period t…

On the other hand, if self-driving cars are as good at avoiding accidents as the article says that may cause the price of insurance for non-self-driving cars to fall.
Post reply on HN