Comment by jjav
17 hours ago
> It's likely that a fully-attentive human driver would have done worse.
We'd have to see video of the full scene to have a better judgement, but I wouldn't call it likely.
The car reacted quickly once it saw the child. Is that enough?
But most humans would have been aware of the big picture scenario much earlier. Are there muliple kids milling around on the sidewalk? Near a school? Is there a big truck/SUV parked there?
If that's the scenario, there is a real probability that a child might appear, so I'm going to be over-slowing way down pre-emptively even thought I haven't seen anyone, just in case.
The car only slows down after seeing someone. The car can react faster that I can after seeing someone, but as a human I can pre-react much earlier based on the big picture, which is much better.
As someone who lives on a residential street right by a primary school in the UK, the majority of drivers are going over 20mph even at the peak time when there are children everywhere.
While in theory human drivers should be situationally aware of the higher risks of children being around, the reality is that the majority will be in their own bubble of being late to drop their kid off and searching for the first free spot they can find.
the human driver would usually drive more closely to the centerline of such a residential road. If the road is clear ahead i'd drive almost over the centerline of the road having enough clearance between my path and the parked cars for any such "jumper" to be visible long enough for me to react. If there is an opposite traffic i get back strictly into my lane and slow down much more if the parked cars are close and they block sidewalk view, etc.
The autonomous cars have really got more aggressive recently as i mentioned before:
https://news.ycombinator.com/item?id=39298290
I'd really like to see the video of the incident.
I have a similar school drop-off, and can confirm that the cars are typically going around 17-20mph around the school when they're moving. Also that yes, human drivers usually do stay much closer to the centerline.
However, Waymo was recently cleared to operate in my city, and I actually saw one in the drop-off line about a week ago. I pulled out right in front of it after dropping my kid off. And it was following the line of cars near the centerline of the road. Honestly its behavior was basically indistinguishable from a human other than being slightly more polite and letting me pull out after I put my blinker on.
> the human driver would usually drive more closely to the centerline of such a residential road
I certainly do this. But asserting that most humans would usually do this? Have you ever actually seen humans drive cars? This is absolutely not what they do. On top of that, they run stop signs, routinely miss pedestrians in blind spots, respond to texts on their phone, or scroll around on their display to find the next song they want to put on.
I vividly recall a shot within a commercial, in which a driver was shown in slow motion, chucking his coffee into the passenger foot well in order to have two hands on the wheel for an emergency. I don’t remember what was about to happen to the car or the world around it. I’m pretty sure that a collision occurred.
Have you been in a waymo? It knows when there are pedestrians around (it can often see over the top of parked cars) and it is very cautious when there are people near the road and it frequently slows down.
I have no idea what happened here but in my experience of taking waymos in SF, they are very cautious and I'd struggle to imagine them speeding through an area with lots of pedestrians milling around. The fact that it was going 17mph at the time makes me think it was already in "caution mode". Sounds like this was something of a "worst case" scenario and another meter or 2 and it would have stopped in time.
I think with humans, even if the driver is 100% paying attention and eyes were looking in exactly the right place where the child emerged at the right time, there is still reaction times - both in cognition but also physically moving the leg to press the pedal. I suspect that a waymo will out-react a human basically 100% of the time, and apply full braking force within a few 10s of milliseconds and well before a human has even begun to move their leg.
You can watch the screen and see what it can detect, and it is impressive. On a dark road at night in Santa Monica it was able to identify that there were two pedestrians at the end of the next block on the sidewalk obscured by a row of parked cars and covered by a canopy of overgrown vegetation. There is absolutely no way any human would have been able to spot them at this distance in these conditions. You really can "feel" it paying 100% attention at all times in all directions.
Your opinion of "most humans" is vastly overinflated. The median human driver would be going 5 over the speed limit, on their cell phone, and paying fuck all attention. Humans never drive as slow as 17 mph, even in the context of being directly in front of schools with visible children.
You're describing the median driver in America or India. This is not universal.
True, but it seems fair to evaluate Waymo against the median American driver. If they expand to whatever other countries you're thinking of, then it will be fair recalibrate accordingly.
For context, Santa Monica is in America or India.
[dead]
> But most humans would have been aware of the big picture scenario much earlier.
I wouldn't call it likely. Sure, there are definitely human drivers who are better than Waymo, but IME they're few and far between. Much more common to be distracted or careless.
When walking along a busy street facing traffic, I like to play a game of "who's using a phone?" I sometimes score in excess of 50% of drivers texting or otherwise manipulating a phone instead of actually driving.
It's amazing how much nonsense we let slide with human drivers, and then get uptight about with anything else. You see the same attitude with bicycles. Cars run stop signs and red lights all day long and nobody bats an eye, but a cyclist does it and suddenly they're a menace.
I don't think it makes sense to lump some drivers better than waymo and worse than waymo. A human brain automatically thinks of all the scenarios, where Waymo has pre-programmed ones (and some NN based ones). So it's scenarios by scenario.
Consider this scenario:
5 kids are walking on the sidewalk while you're driving past them. But suddenly a large dumpster is blocking your view of them just as you pass. You saw them before the dumpster, but not after your car and the dumpster completely blocks the view.
Does a human brain carry some worry that they suddenly decide to run and try to cross the street after the dumpster? Does Waymo carry that worry or just continue to drive at the exact same speed.
Again, it's not like every driver will think about this, but many drivers will (even the bad ones).
> A human brain automatically thinks of all the scenarios
I don't think this is true. There are infinitely many scenarios in a complex situation like a road with traffic, cars parked, pedestrians about, weather, etc. My brain might be able to quickly assess a handful, but certainly not all.
12 replies →
You are vastly overestimating most drivers. Most drivers aren't even looking out the window the majority of their time driving.
> A human brain automatically thinks of all the scenarios, ...
Patently, obviously false. A human brain will automatically think of SOME scenarios. For instance, if a collision seems imminent, and the driver is holding a cup of coffee, these ideas are likely to occur to the driver:
IF I GRAB THE STEERING WHEEL AND BRAKE HARD, I MIGHT NOT HIT THAT PEDESTRIAN IN FRONT OF ME.
IF I DON'T CONTINUE HOLDING THE COFFEE CAREFULLY, I MIGHT GET SCALDED.
THIS SONG ON MY RADIO IS REALLY ROCKING!
IF I YANK MY WHEEL TO THE LEFT, I MIGHT HIT A CAR INSTEAD OF A HUMAN.
IF I BRAKE HARD OR SWERVE AT ANY TIME IN TRAFFIC, I CAN CAUSE AN ACCIDENT.
Experiments with callosal patients (who have damaged the connective bridge between the halves of their brains) demonstrate that this is a realistic picture of how the brain makes decisions. It offers up a set of possible actions, and attempts to choose the optimal one and discard all others.
A computer program would do likewise, EXCEPT it won't care about the coffee cup nor the radio (remove two bad choices from consideration).
It still has one bad choice (do nothing), but the SNR is much improved.
I'm not being hyperbolic; self-preservation (focusing on keeping that coffee in my hand) is a vital factor in decision-making for a human.
> ...where Waymo has pre-programmed ones (and some NN based ones).
Yes. And as time goes on, more and better-refined scenarios will be added to its programming. Eventually, it's reasonable to believe the car software will constantly reassess how many humans are within HUMAN_RUN_DISTANCE + CAR_TRAVEL_DISTANCE in the next block, and begin tracking any that in an unsafe margin. No human on Earth does that, continually, without fail.
> Does a human brain carry some worry that they suddenly decide to run and try to cross the street after the dumpster? Does Waymo carry that worry or just continue to drive at the exact same speed.
You continue to imply that Waymo cannot ever improve on its current programming. Does it currently consider this situation? Probably not. Will it? Probably.
1 reply →
For what it's worth, that kind of lumping of drivers is more-or-less one of the metrics Waymo is using to self-evaluate. Perfect safety when multi-ton vehicles share space with sub-300-pound humans is impossible. But they ultimately seek to do better than humans in all contexts.
According to the article the car was traveling at 17 miles an hour before it began braking. Presumably this was in a 25 mph school zone, so it seems the Waymo was already doing exactly what you describe - slowing down preemptively.
This is close to a particular peeve I have. Occasionally I see signs on the street that say "Slow Down". I'm not talking about the electronic ones connected to radar detectors. Just metal and paint.
Here's my problem. If you follow the instructions on the sign, it still says to slow down. There's no threshold for slow enough. No matter how slow you're going, the sign says "Slow Down". So once you become ensnared in the visual cone of this sign, you'll be forced to sit stationary for all eternity.
But maybe there's a loop-hole. It doesn't say how fast you must decelerate. So if you come into the zone going fast enough, and decelerate slowly enough, you can make it past the sign with some remaining non-zero momentum.
You know, I've never been diagnosed on the spectrum, but I have some of the tendencies. lol.
Obviously a static sign is not aware of your current state, so it's message can only be interpreted as relevant to your likely state... i.e. the posted speed limit.
5 replies →
A lot of clickbait headlines have the same problem. "You're using too much washing powder!"
Everyone's replying to you as if you truly don't understand the sign's intention but I'm sure you do. It's just annoying to be doing everything right and the signs and headlines are still telling you you're wrong.
There was a driving safety safety ad campaign here: "Drive to the conditions. If they change, reduce your speed." You can imagine how slow we'd all be going if the weather kept changing.
We might have OCPD.
6 replies →
Think of it like they're saying "my children play on this street and my neighbors walk here. Please think about that when you decide how fast to go here."
You learn how to put those signs into context during your driving lessons, and fail your test if you don't apply that correctly.
2 replies →
Think of the sign as a flag, not an instruction.
25 mph is an upper limit, not a minimum.
A 25mph school zone? That seems fast. 15mph would be more the norm, which is in line with the 17mph the car believed itself to be traveling.
FYI, unless you are a commerical truck, a cop, or a racer, your speedometer will read slightly fast, sometimes as much as 5 to 10%. This is normal practice for cars as it limits manufacturer liability. You can check this using independant gps, ie not an in-dash unit. (Just imagine the court cases if a speedo read slower than the actual speed and you can understand why this started.)
I mostly see 25 mph for school zones, though I'm in NC. Checking California, it sounds like 25 is standard there as well.[0] Some will drop to 15, but 25 is the norm as far as I can find.
[0] https://www.dmv.ca.gov/portal/handbook/california-driver-han...
2 replies →
Also, a different wheel diameter than the speedometer was calibrated with and you will have a larger difference between actual velocity and speedometer reading. The odometer will also not record actual distance traveled.
2 replies →
It does seem fast to me -- school zones are 20 mph in Seattle, at least when children are present. But Google suggests 25 is the norm in Santa Monica, where the incident occurred.
In Encinitas, California, that sign would have no more than 20 MPH. In adjacent Carlsbad, I believe 25 is normal.
In this situation, the car was already driving under the legal speed required for a school zone (25mph when children are present) [edit: some comments in the post suggest there is a 15mph sign, which is sometimes posted; to me, driving 17mph in a 15mph zone is acceptable).
I think any fair evaluation of this (once the data was available) would conclude that Waymo was taking reasonable precautions.
> was already driving under the legal speed
That's exactly part of the problem. If it is programmed to be over-cautious and go 17 in a 25 zone, that feels like it is safe. Is it?
It takes human judgment of the entire big picture to say meaningfully whether that is too slow or too fast. Taking the speed limit literally is too rigid, something a computer would do.
Need to take into account the flow of the kids (all walking in line vs. milling around going in all directions), their age (younger ones are a lot more likely to randomly run off in an unsafe direction), what are they doing (e.g. just walking, vs. maybe holding a ball that might bounce and make them run off after it), their clustering and so on.
Driving past a high school with groups of kids chatting on the sidewalk, sure 20mph is safe enough. Driving past an elementary school with a mass of kids with toys moving in different directions on the same sidewalk, 17mph is too fast.
And if I'm watching some smaller kids disappear behind a visual obstruction that makes me nervous they might pop up ahead of it on the street, I slow down to a crawl until I can clearly see that won't happen.
None of this context is encoded in the "25mph when children are present" sign, but for most humans it is quite normal context to consider.
But would be great to see video of the Waymo scene to see if any of these factors was present.
It was going 17 mph. That is rather slow.
To put it another way. If an autonomous vehicle has a reaction time of 0.3 seconds, the stopping distance from 17 mph is about the same as a fully alert human driver (1 second reaction time) driving 10.33 mph.
>It was going 17 mph. That is rather slow.
There's a case to be made that it wasn't slow enough.
I have a hard time believing that a human driver would be as slow as this Waymo, or even slower. I drive my kid to school where it's posted 20mph and there are cameras (with plenty of warnings about the presence of said cameras) and witness a constant string of flashes from the camera nailing people for speeding through there.
1 reply →
A small child jumped out in front of it, which is about the worst case scenario you can have... and the kid was fine. So it sounds like it was slow enough?
Considering the car hit the child at only 6 mph and the kid just got up and brushed themselves off, it was plenty slow enough.
Nobody was injured.
Maybe. That level of safetyism seems pretty unreasonable when humans are 100x worse and still allowed on the road.
Curiously enough Google could have access to how fast humans usually drive through that street.. if they record people's Google Maps trips, they can show the court that "Look, 80% of Google Maps users drive through here at 30 mph!".
Waymo itself has this as well. They record their drives after all which means they know the speed of the vehicles around them.
They even wrote a blog post about it:
https://waymo.com/blog/2023/07/past-the-limit-studying-how-o...
Google might even know how many drivers aren't obeying the speed limit or slow-rolling through stop signs. I wonder if they already have partnerships with law enforcement to detect areas where the traffic law is more ignored than others.
Two things:
I've read studies saying that most drivers don't brake at max effort, even to avoid a collision. This may be at least one of the reasons that Waymo predicted that an attentive human would likely have been going faster than their car at the moment of impact. I've got a good idea of my fun-car's braking performance, because I drive it hard sometimes, but after reading that I started practicing a bit with my wife's car on the school run, and... Yeah: it's got a lot more braking power than I realized. (Don't worry, I brake hard on a long straight exit ramp, when no one's behind me, a fast slow-down is perfectly safe, and the kiddo loves it.) I've now got an intuitive feel for where the ABS will kick in, and exactly what kind of stopping distance I have to work with, which makes me feel like a safer driver.
Second, going off my experience of hundreds and hundreds of ride-share rides, and maybe thirty Waymo journeys, I'd call the best 10-15% of humans better drivers than Waymo. Like, they're looking further up the road to predict which lane to be in, based on, say, that bus two blocks away. They also drive faster than Waymos do, without a perceptual decrease in safety. (I realize "perceptual" is doing some work in that sentence!) That's the type of defensive and anticipatory urban driver I try to be, so I notice when it's done well. Waymo, though, is flat-out better, in every way, than the vast majority of the ride-share drivers I see. I'm at the point where I'll choose a Waymo any time it'll go where I'm headed. This story reinforces that choice for me.
> I've read studies saying that most drivers don't brake at max effort, even to avoid a collision.
Ha! It is unbelievable how difficult it is to make someone brake hard. You'd think it's the easiest thing possible in the age of ABS - just press hard as you can.
I have a lot of experience on this, I used to teach car control both to teens and adults. One of the frequent exercises was seemingly very simple: Drive at Xmph until this spot, then brake at maximum power.
The vast majority of people can't do it on the first or second try, they'll just meekly press on the brake like they're coasting to a stop. After more coaching that hard means hard, they start to get it, but it takes many many tries.
If you brake too hard, you might fly over the handlebars.
1 reply →
The reason attentive humans don't equal the Waymo here is reaction time. When a thing happens the human takes a moment to process what it means, and choose a reaction. It's not, by our standards, a long time but it's way longer than it takes the Waymo.
Going early means you slow early, which means you also take longer to reach the child, but you're braking for all of that extra time, so you're slowing down even more.
It would be nice to see the video (although maybe there are some privacy issues, it is at a school after all).
Anyway, from the article,
> According to the NHTSA, the accident occurred “within two blocks” of the elementary school “during normal school drop off hours.” The safety regulator said “there were other children, a crossing guard, and several double-parked vehicles in the vicinity.”
So I mean, it is hard to speculate. Probably Waymo was being reasonably prudent. But we should note that this description isn’t incompatible with being literally in an area where the kids are leaving their parents’ cars (the presence of “several double parked cars brings this to mind). If that’s the case, it might make sense to consider an even-safer mode for active student unloading areas. This seems like the sort of social context that humans might have and cars might be missing.
But things speculation. It would be good to see a video.
> But most humans would have been aware of the big picture scenario much earlier. Are there muliple kids milling around on the sidewalk? Near a school? Is there a big truck/SUV parked there?
Waymos do this and have for years. They know where the people are around them and will take precautionary action based on that.
Here's a video from 2019 of one understanding that a car in the bike lane means the cyclists may dart out into the lane it's in and taking action based on that. https://waymo.com/blog/2019/05/safety-at-waymo-self-driving-...
That video is nearly 7 years old at this point and they've gotten much, much better since then.
If you think a fully-attentive human driver would have done better, I think you're kidding yourself.
I know you didn't make this point, but if anyone think the average LA driver would have done better than this I've got a bridge to sell you and that's really what matters more. (I say that as someone who used to live like half a mile from where this happened)
> The car only slows down after seeing someone.
How do you know that? The article says it slowed from 17 mph. That’s cautious progress speed, not cruising speed.
In principle, attentive drivers, who have either somehow come independently to the appropriate understanding or have been trained in how to react to hazards ahead...
https://www.gov.uk/theory-test/hazard-perception-test
... could in some circumstances know that there's a likelihood that a child will emerge suddenly and reduce their speed in anticipation where circumstances allow.
Note that: If you cut speed but other drivers can't see why they may overtake, even unsafely, because you are a nuisance to them. Slowing in anticipation that a child will run out from behind the SUV, only for a car behind you to accelerate around you and smack straight into the child at even higher speed, is not the desired outcome even though you didn't hurt anybody...
And yes, we'd need to see the video to know. It's like that Sully scenario. In a prepared test skilled pilots were indeed able to divert and land, but Sully wasn't prepared for a test. You're trained to expect engine failure in an aeroplane - it will happen sometimes so you must assume that, but for a jet liner you don't anticipate losing both engines, that doesn't happen. There's "Obviously that child is going in the road" and "Where the fuck did they come from?" and a lot in between and we're unlikely to ever know for sure.
I live in an area where there are pedestrians stepping into the street without looking, all over the place, and you can drive / cycle without hitting them but have to slow down appropriately if you have to go near something that you can't see behind. Like you say it would be interesting to see the video.
It was already moving slowly. 17MPH is pretty conservative. Most human drivers going past my local school are doing at least 30.
There's a bus stop right behind my house. I routinely hear the driver honking and yelling at people who ignore when the stop sign is extended (which is a misdemeanor in my state). So forgive me for not assuming a human would have done better.
The car was driving 17mph before braking. I don’t think I’ve ever seen a human drive at 17mph in a school zone or other area children congregate.
Meaning you’ve never seen a human drive that slowly in such an area, or you've never seen a human exceed the speed limit in a school zone?
Drive that slowly. I’m exaggerating obviously but the norm is to speed.
I drive like this too, but I think we’re a small minority. Especially here in LA.
Are they not using a ton of ML to take exactly this sort of context into account?
> But most humans would have been aware of the big picture scenario much earlier. Are there muliple kids milling around on the sidewalk? Near a school? Is there a big truck/SUV parked there?
Waymos constantly track pedestrians nearby, you can see it on the status screen if you ride in one. So it would be both better able to find pedestrians and react as soon as one was on a collision course. They have a bit more visibility than humans do due to the sensor placement, so they also can see things that aren't that visible to a person inside the car, not to mention being constantly aware of all 360 degrees.
While I suppose that in theory, a sufficiently paranoid human might outdo the robot, it looks to me like it's already well above the median here.
Do they speculate about things like “we’re near a school zone, kids are unloading, there might be a kid I’ve never seen behind that SUV?” (I’m legitimately asking I’ve never been in a Waymo).
It's not particularly meaningful to ponder the subjective experience of the waymo driving computer. Instead, focus on its externally visible behavior.
3 replies →
>The car can react faster that I can after seeing someone
and that can potentially allow internal planning algorithm to choose more risky and aggressive trajectories/behavior, etc. say to reach target destination faster and thus deliver higher satisfaction to the passengers.
Anecdote, but I live next to an elementary school and also on a route frequented by Waymos. Human drivers routinely cruise down the 25mph roads at 40+ and blow stop signs, even during school intake and release. Waymo vehicles always seem a lot more cautious.
When thinking about these things you have to factor in the prior probability that a driver is fully attentive, not just assume they are.
If you’ve ever been in a Waymo you quickly realize their field of view is pretty good. You often see the vehicle sensing small pets and children that are occluded to a passenger or driver. For this reason and my experience with humans near aforementioned school, I doubt a human would out perform the Waymo in this particular incident and it’s debatable they even have more context to inform their decisions.
All that said, despite having many hours in a Waymo, it’s not at all clear to me how they factor in sidewalk context. You get the sense that pedestrians movement vectors are accounted for near intersections, but I can’t say I’ve experienced something like a slow down when throngs of people are about.
Precisely. Environmental context is not considered in Waymo's "peer-reviewed model" (I encourage reflexive commenters to first read it: https://waymo.com/safety/collision-avoidance-benchmarking), only basic driver behavior and traffic signal timings.
Note the weaselly "immediately detected the individual as soon as they began to emerge" in the puff piece from Waymo Comms. No indication that they intend to account for environmental context going forward.
If they already do this, why isn't it factored in the model?
How is "immediately detected the individual as soon as they began to emerge" worded weaselly?
Not OP but I interpret that as they are focusing exclusively on what happened after the car saw the kid.
And I completely agree that from that instant forward, the car did everything correctly.
But if I was the accident investigator for this, I would be far more interested in what happened in the 30 seconds before the car saw the kid.
Was the kid visible earlier and then disappear behind an obstruction? Or did the kid arrive from the side and was never earlier visible? These are the more important questions.