Hacker News Reader

Comment by athenot | Hacker News Reader

Perhaps San Francisco rain is solved, but I doubt the downpours we get in the South East are solved. Zero visibility is just one thing, how about hydroplaning when there is over 1" of standing water on the road because it's coming down faster than gravity can drain it away? Or spotting that your lane is fine but the opposite one is flooded so that car coming the opposite direction is going to have to borrow your lane for a bit?

These sound like edge cases but there are places where they occur very frequently and local human drivers, for all their faults, do have a way of dealing with them.

View on HN

> hydroplaning when there is over 1" of standing water on the road because it's coming down faster than gravity can drain it away?

I'd say that's pretty calculable, and has a solution, slow down to a speed that the tires can shed water.

Perhaps most telling about your qualification is the misplaced risk; the light precipitation with a little oil slick is likely more dangerous, mostly because you don't slow down for it.

Anecdotally, an even harder problem is when water exceeds 4", negotiating with traffic so you won't float away or get water in your intake.

Yes, it is trivial to calculate how to respond to known hazards. What is hard to calculate is what the hazards actually are in any given situation. Humans don't merely look at the road and compute how much standing water is there, they have prior experiences and a sense of what other drivers are doing on the road and why they are doing those things.

Self driving cars also have prior experiences and a sense of what other drivers are doing and why. That's sort of the point of machine learning.

In fact, they will have a lot more prior experience than any human driver in the world.

Unfortunately in the realm of machine learning, quantity does not equate quality.

The parent of any teenage driver can also attest to this.

Machine Learning still confuses Elephants for Cats.

While it's quite fun if you pin up a cat dressed up in an elephant costume and surely amuses a few colleagues, a self-driving car is not something that should confuse an Elephant for a Cat.

ML is IMO not reliable enough for use in self-driving cars, in complicated situations the driver should take over.

It's not reliable enough now, sure.

Similarly, my 6 year old is also not reliable enough to pilot a car.

Both will change as they mature.

Sure, multiple decades sounds about right (as opposed to current confusing statements like "has full self driving capabilities").

They have a lot more experience driving, but humans do not only use their driving experiences as priors.

humans regularly get killed misjudging how much water is in underpasses, I guess an autonomous car based on detailed mapping can measure water deep quite better comparing sensor result with its map (eventually).

Those conditions sound like you wouldn't be driving. So, solved. Pull over and stop

While probably true, "then don't do that" almost never works as an excuse to convince new adopters of your technology.

I'm not so sure that's true. Or, maybe, "that's user error" is what works because it seems like it's far too easy to convince users to blame themselves.

Well, when all your friends can drive in the snow but your fancy-pants car can't, that's not a very compelling value proposition.

This; arguing that a computer can't drive safely in conditions where a human can't either isn't an argument against computer drivers. Just because humans sometimes get away with it doesn't make it a good idea.

Well, I don't think there is any question about it. It can only be attributable to human error. This sort of thing has cropped up before, and it has always been due to human error.

It's not always possible to pull over. If it's snowing so heavily that driving has become dangerous there probably is no shoulder to pull over to, much less a shoulder you can safely pull over to. Source: I've been in these conditions. You do the best you can.

you should pull over before then. There is in my experience plenty of warning to find a hotel first.

You must never have visited Hemet, CA, or driven anywhere it rains a lot sometimes, or has a real winter.

And don't pick up your phone with your right hand.

You can drive in a torrential downpour. Short of actual flooding or the creation of large puddles, rain isn't all that bad of a driving condition. Be sure not to drive past the limits of visibility and increase follow distances from 2-3 to 5 seconds and you're good. Even motorcycle races are held in the rain: https://www.youtube.com/watch?v=tv4_425K8r4

I have been in a lot of monsoon level downpours where it is effectively impossible to drive as visibility goes to 0. There are different types of rain.

At some point "no one[0] should be driving" becomes a valid answer.

[0] I mean, there are going to be cases, but the number of people currently driving in horrific weather conditions because it would be inconvenient not to is too damn high.

It seems a lot of posters aren't willing to consider self-driving cars feasible until the cars are as foolhardy as the bottom-decile driver while still maintaining a perfect safety record. But most of accident prevention is in the planning more than the execution. It seems only reasonable to expect SDCs to avoid situations that some human drivers would brave if the risk is too high.

LOL. In Atlanta, highway traffic just slows down to 40 in torrential downpour conditions where you can’t see lane markings.

40 mph?

On I95 in the DE-NY area, it's more like 60 mph in heavy rain. Or even in moderate snow. Everyone knows that they could never stop. But they're counting on others' inability to stop ;) But occasionally, there are massive chain-reaction pileups :(

These are daily occurrences in parts of TX and FL in the summer, and people generally do drive through them.

It sounds like you've never lived in the southeast US? These conditions can occur on a daily basis in the summer, especially in Florida. "Pull over and stop" isn't realistic advice here.

Humans can and do drive safely in these conditions, however, and it's a valid concern to point out that Waymo's West Coast-centric models might not yet be trained to handle these cases.

(To be clear, I don't doubt that machines will eventually surpass humans in whatever weather conditions, if they haven't already.)

> safely

I've seen people do all sorts of foolish things. When the rain is so hard that I can't see more than 15 feet, I pull over and stop.

How does a car know it can't see well?

It uses its sensors. If the LIDAR shows only white noise it knows that it can't see well.

White noise certainly isn't the only indicator that one can't see well.

How does a human know it can't see well?

Well, that's a great question. It's also basically equivalent to the one that I asked. My point would be that I don't have an answer to that question, and given your response, I bet you don't either. Which leaves either of us at a loss when we try to create a safe and reliable autonomous vehicle.

It's certainly not an unanswerable question. It's just quite hard to verify that they answer you have is a sufficiently correct one.

> It's just quite hard to verify that they answer you have is a sufficiently correct one.

Why? It seems to me that estimating the confidence of a prediction by the quality of measurements is not an infrequent machine learning task. It'd be more complex for a driving algorithm than calculating the confidence interval of a linear regression, but it's in the same theme.

A computer can predict with 100% confidence if you give it few enough measurements. That doesn't mean it will make the correct decisions. You have to already know you're measuring the right things.

If you don't want it to do that, don't program it that way. Code it to assume some and to estimate the degree of mismeasurement.

Fair enough! You're absolutely right: I don't know the answer. I misinterpreted your post as suggesting/implying that there is no answer for a computer.

Nobody said anything about 15 feet but you. If you're pulling over in a normal heavy downpour with 500 ft of visibility, you're less safe than the people who continue to drive. Being pulled over to the side of any high-speed road is an unsafe condition that should be avoided.

Self-driving cars would be putting lives at risk if they just pulled over and shutdown on every summer thunderstorm.

With 500 ft of visibility, the auto-steering/auto-propelled car shouldn't have any trouble either.

"These conditions can occur on a daily basis in the summer, especially in Florida"

Can confirm, just spent 30 summer days in Florida. People don't pull over and stop, they just put on their "hazards" and slow down, sometimes to 25-30 MPH depending on the intensity of rain. This happens every single day.

I grew up with that kind of summer rain in Illinois, too. After about five minutes of it, yes, a lot of (sane) drivers pull over. But it comes on very suddenly, and for the first two or three minutes, traffic flows continue.

If you're in the center lane and traffic is heavy on both sides of you, while your windshield (or camera) is overwhelmed with water, you need to keep moving with traffic for a bit -- or you get rear-ended.

You have to pull over when it happens. Sometimes those southeast flash flood rains go from bad to opaque in an instant. Depending on the road you may not have enough of a shoulder to pull over, and any other driver will only be able to see you at 5-10 feet away maximum and (initially) going 45-60mph. Not ending up in a ditch can be difficult.

I’d recommend visiting Finland. In these parts, there’s basically 3 months of optimal visibility. The rest is fog, rain, snow, sleet and road conditions that would shut down most of US. Yet people drive and do so safely. “What people can do” is a bit different from “what an average us driver can be expected to do”

Doesn’t findland have a higher vehicle fatality rate than other nations?

Seems to be less than half the US rate.

https://en.wikipedia.org/wiki/List_of_countries_by_traffic-r...

Ah, I was thinking of the UK. Finland has 50% more fatalities per mile than the UK.

It may be that regions with better climates receive these vehicles first. It's relatively straightforward to detect the prevalent conditions and refuse to drive if the weather is outside the scope for which the vehicle is rated. As the technology matures, that scope should broaden.

So road trips are unaccounted for? Not a comforting idea that taking your car to another area will render it functionally unsafe on the (differently designed) roadways.

That's a price I would very very happily pay for a self-driving car that works in my city. If I want to take a road trip, i'll rent a traditional car.

We clearly have different lifestyles. Renting a car 3 times a week to venture into the mountains is beyond unaffordable, much less once a month. Do you stay that static in your life?

Essentially the same as "smartphones are about 95% useless without mobile broadband". Except for the tiny detail that lack of mobile coverage is unlikely to kill you...

And as a taxi driver once told me in Helsinki there are four seasons in Finland - Autumn, Winter, Spring and Construction. Even when the visibility is optimal they'll likely be tearing up the road anyway.

Reply on news.ycombinator.com