Hacker News Reader

Comment by heywire | Hacker News Reader

On days like today, when there are several inches of snow on the ground and more to come, I always wonder how autonomous vehicles will handle such situations. What happens when snow (or mud, etc) accumulates on the camera or sensors? I don't doubt that a computer can react quicker and with more precision in an "event", but what about when it simply can't "see"? I don't see a steering wheel in these videos, so I guess there is no manual failsafe?

View on HN

Ford has been quietly testing in snow in Ann Arbor, MI.[1] Volvo tests in snow a lot. Volvo discovered that ice on the radars could be a problem. (So, de-icing heaters on the front of the sensors, like aircraft.) The big auto companies have tested their cars in bad conditions for many decades. They all have test tracks in awful locations and performance standards for bad weather. Even Google tests in Tahoe occasionally.

What's becoming clear is that the self-driving car industry will be the car industry. Autonomous driving equipment may come from suppliers, but they'll just be suppliers to the automakers. That's not a great place to be; the margins are low and suppliers are totally under the thumb of the automakers.

Volvo, which will deploy 100 self driving cars with actual customers in Gothenberg, Sweden in 2017, is way ahead on the user interface.[2] Volvo takes the firm position that, in autonomous mode, the driver is not required to pay attention at all, and if something goes wrong, it's Volvo's problem. They have redundant sensors, actuators, and computers.

Volvo is also way ahead on ads for self-driving.[3]

[1] http://www.volvocars.com/intl/about/our-innovation-brands/in...

[2] http://www.dailymotion.com/video/x4di8ds

[3] https://www.youtube.com/watch?v=uDB6fFflTVA

Not just Ford; most auto OEM's are involved with MCity: http://www.mtc.umich.edu/partners/industry

These cars will likely have many sensors of different types at overlapping locations, its not likely to keep going if all are completely unusable.

But I wonder about the snow though. How does the car know where to go when there are no identifying lane features? Maybe its not as hard of a problem as I believe it to be?

I really don't get these questions. "How will the car drive in whiteout conditions where the road is frozen over and not at all visible?" Same way a (smart) human drives: it will pull over.

For substantially less crazy conditions, you can infer where the lanes would be by (a) looking at the spots of the road you can see, (b) prior knowledge of the road from experience, and (c) looking at oncoming cars and the flow of traffic, and driving in a way that doesn't surprise them. Though it's likely the case that you shouldn't be driving in these conditions anyway.

You must not live somewhere that gets a lot of snow :) For those of us that do, sometimes driving conditions are much less than ideal, even unsafe. Several times per winter I have to drive home from work on unplowed roads. As you put it, I have prior knowledge of the road (maybe precise GPS will help here?), and I can figure out where it is safe to drive based on tire tracks. I guess these will just be places that you probably won't have one of these vehicles (just as you wouldn't ride your motorcycle in such weather).

> Same way a (smart) human drives: it will pull over.

You mean carefully get to the nearest settlement as soon as possible? Some of these storms can last for days at a time. Stopping to wait it out in the middle of nowhere, especially if you don't have a full tank of fuel to keep the car warm for extended periods, is a pretty scary situation.

Hmmm. If the car decided, before your trip started, that it's too dangerous for anyone to drive where you want to go because of possible weather and environmental conditions mid-route, would that be useful or enraging?

I imagine a bit of both. Getting caught in a snow storm is a pretty horrible experience, and I would gladly stay home knowing I would have otherwise ended up in one. But at the same time the forecasting isn't all that great, so often the storms don't amount to anything, and you would have been perfectly fine to be out there. Sometimes the storms even come up without any warning (most of the snow we get – especially that which causes driving issues – is from the lake, not the clouds). Having to essentially hibernate in the winter because the car is always erring on the side of caution could become pretty enraging.

As haywire said, if you live where it snows, you drive regardless of whether or not you can see the road. I got a job to do, computers don't fix themselves, and my job is necessary even if the weather sucks outside.

While I think it's plausible to teach a self-driving car to drive in snow (like I do, generally following in the tracks of the cars ahead of me), I honestly think this is one of the best justifications that cars should continue to have steering wheels and pedals: Because automated snow driving is going to take a lot longer to become a solved problem than fair weather driving.

How does a human do it? It's extremely hard but recognizing landmarks like a guard rail or plow markers would be possible. Or maybe self driving cars will apply a sane approach and stay off the streets in bad weather unlike human piloted vehicles. (There is a whole other rant about letting employees work from on bad snow days)

Presumably the cars won't continue driving if their systems are inoperable or unsure how to proceed. One could also argue that riding around in several inches of snow is a bad idea regardless of who's driving, human or computer.

That's a little concerning. Especially to the lee of the lake, snow storms can pop up out of nowhere and can last for days at a time. Usually, with care, a human driver can traverse to the nearest shelter to weather the storm. But if the car just stops out of the middle of nowhere, with limited energy to keep the cabin warm in the harsh cold, that seems like a recipe for disaster.

We're talking about mobility-as-a-service, so presumably the service would use other means to get you to your destination, like having the car towed, or sending out an SUV with a professional driver.

These kinds of extreme situations are scary, but that's what insurance is for.

Everyone is worried about how we get autonomous driving that's perfect in every situation, but that's not how automation works. You just have to automate the common case so a small number of humans can focus on the corner cases.

> We're talking about mobility-as-a-service

We are? I went up the comment chain again and still did not get that impression the second time around. What statement in particular gave you that idea?

> These kinds of extreme situations are scary, but that's what insurance is for.

Like... life insurance?

> You just have to automate the common case so a small number of humans can focus on the corner cases.

If I'm going to be sitting in the car anyway, I can drive it in those conditions, assuming the vehicle allows me to. However, Google has stated before that it is unlikely that they will release a car that needs to be taken over by human drivers. It's all or nothing for them.

It's a fair question, but it doesn't seem unsolvable. And humans have huge issues driving in snow themselves, so by comparison even sightly impaired self driving will be preferable.

If it can't see, you can't see either.

The sensors are primarily radar based. Mud/snow won't really affect them.

False -- https://en.wikipedia.org/wiki/Clutter_(radar)

Anyways, if this is true, it's a grim future then. Radar is incredibly trivial to interfere with.

From the "How it works" Google self driving car page: "Sensors Lasers, radars and cameras detect objects in all directions"

It uses many kinds of sensors. Yes I know about clutter, I've spent quite a lot of time in the radar industry. But by combining data from radars of multiple wavelengths, it becomes pretty feasible. Though yes, difficult.

https://www.google.com/selfdrivingcar/how/

It strikes me as curious that someone who spent a lot of time in the radar industry would make the original statement you made, at least in that fashion. Also, radar is subject to interference as well. In addition, there are many things that act as "natural" corner reflectors that would pose more challenges, I can only think of some mountainous snowy / icey nothern US/Canadian routes and just thinking No way Light sensors and Radar sensors will conquer this.

We need to fundamentally redesign our road systems to accommodate self driving cars. That might happen... in 50-100 years...after we address the already crumbling infrastructure we have. For perfect, sunny conditions, like the roads in Nevada all these self driving startups are doing their testing, with straight flat landscapes, I am sure the tech can work fine. Rest of the country, maybe not so much.

The original comment was hasty, I admit. But there are a lot of ways to tackle these issues.

Radar clutter, attenuation, penetration, etc are all affected by wavelengths. By using multiple frequencies, you can get a better idea of what is and isn't really there. As far as interference, there are ways to filter out noise. The car presumably knows how fast it is going. Therefor, it can do doppler filtering on the received waves. IE... car knows it sent waves at 50 GHz and is traveling at 60 mph. It knows to expect a response at ~54 GHz from the front and 50 GHz from the side.

For tracking objects, you can use a pulse-doppler radar [0] to get both range and rate information.

[0] https://en.wikipedia.org/wiki/Pulse-Doppler_radar

And there's a dual risk of interference, too.

One is the basic "environmental" interference with competing radar signals from other devices. Imagine how much more complicated this gets when a majority of competing traffic is likewise equipped.

But the other risk, which seems to be mostly ignored thus far, is sabotage. Can you design a system that is hardened not only against stray competing EMR but also against attempts to hack or vandalize your system into misbehavior or malfunction? For example, consider a terrorist plot. Or even a smart neighborhood crazy who wants the neighborhood kids off his lawn and neighborhood traffic to stop during his midday nap. Or, a British organized crime gang using stopped traffic to cover an escape[0].

[0] http://www.imdb.com/title/tt0064505/

When all the cars on the road are self driving, I don't see why they would need to have all their radars on at once. I imagine they would be able to communicate with each other and form a complete picture of the entire roadway without all of them having their radars on (or perhaps running at a lower power?)

This would be very difficult to train with ML, however. I'm sure there are other downsides I'm not thinking of

Reply on news.ycombinator.com