Sure, it was able to handle this trip once during ideal conditions. What happens when road conditions become tricky? Will the AI be able to navigate a lane closure that forces it to drive on the shoulder due to an accident? How about adverse weather conditions?
How does that work on ice or snow? With a bus behind you? Or on a curving downgrade? With a trailer? With your unbuckled luggage or dog in the back seat?
When the driving is easy, the AI _can_ do it. We are by default discussing the unusual, difficult situations.
Another edge case I've wondered about is whether there will be driving circumstances during inclement weather in which a sensibly-programmed autonomous car will simply pull over and refuse to operate, even as human drivers happily drive forth. Autonomous cars will be capable of producing reasonably objective measurements of driving condition safety... I wouldn't be surprised winter storms are more routinely capable of inducing conditions that are worse than we humans really realize. Or, alternatively, that we realize are dangerous but choose to drive in anyhow, but no autonomous care programmer would ever choose to accept.
The one thing that sticks with me in all of these automated driving demonstrations is the lack of operation in adverse environmental conditions such Rain, Fog, Snow/Ice. I would expect that any one of those conditions would post extraordinary challenges to an automated driving system.
People also do really poorly in edge cases. Don't forget self driving cars are going to quickly have billions of road miles worth of data.
There are plenty of videos of autonomous cars is highly chaotic road conditions, but that's all old hat. Weather is an issue, but weather is also forecastable. Even if you only get rid of truckers in areas stay above freezing that's still a massive change.
One thing that is going to be awkward to navigate in the transition to self-driving cars is how they handle driving in bad weather. Obviously that's a technical problem but I'm not sure it's actually solvable because it's not actually solvable for human drivers. We just kind of do it anyway. But when a self-driving car has some quantified sense that "driving in these conditions is unsafe" will we get to the point where we just say f it and let them keep going? Or, maybe more likely, we remain in the gray zone where the car is fully self-driving but the human driver has to be ready to take control at any time which I think limits some of the larger impacts.
Does self-driving have to handle all weather conditions right away? A sensible implementation needs to take the current conditions into account, such as the weather and the status of the road and car. If those are bad, it would refuse to active itself, similarly to how a responsible human would choose to not drive in bad conditions.
People are bad drivers. Every winter at the first snowfall there are tonnes of accidents because folks forgot to adjust following distances.
The promise of an AI driver is that there will never be a second accident of the same type. The only question is how fast the learning rate can ramp up.
Also we then get to some more interesting challenges. Certain snow conditions can have rather bad traction like very muddy roads do. And self-driving should be able to navigate these better than regular drivers.
The real world is full of 'edge cases' that humans can deal with through their intuition and understanding of context. Humans are unlikely to mistake a yellow-ish moon for a yellow traffic light unless very intoxicated.
I'd like to see a Tesla on full autonomous autopilot take a drive in Calgary or Edmonton a few days after a heavy snow, when it's been mostly plowed but the lines are missing from the roads, all the lines at the side of the road are covered in snow, etc.
All good points, but just a side note on one of my reservations about self-driving cars:
> and to top it all off, if the AI feels that it's in over its head than the car can just stop whenever and everyone can climb out of the vehicle
Just stopping in a car is always less dangerous than just stopping a plane, but there are cases where just stopping is very dangerous. Desert driving in the absence of sandstorms is probably unlikely to overwhelm an AI. However, in Alaska or the upper Midwest, you'll get snowstorms where you've got 6 inches or more of snow on top of the road, IR and the visible spectrum are limited to tens of feet due to blowing snow, exposed flesh will begin to freeze in under 10 minutes, and you might get hypothermia if you're in your car overnight. Getting out of the car and trying to flag down a passerby is likely to get you struck by an out-of-control car, so your best bet is to call a tow truck and hope they can get to you before the inside of your car gets too cold. In major blizzards, there are so many cars needing rescue, that some people need to wait in their cars until well into the following morning. The local newscast will warn people that driving is potentially life-threatening, but that doesn't mean that essential workers won't be driving. Having half of your trauma surgeons stuck in ditches in self-driving cars during a blizzard isn't a great situation to be in.
People keep candles, matches, and candy/energy bars in old coffee cans, plus water and blankets in their trunks (boots, if you're British) for these circumstances in order to be able to shelter-in-place for a day. Even humans with years of driving experience under icy conditions are pretty bad at driving in these conditions.
Having grown up in the upper Midwest, I was amazed how terrible drivers in the D.C. metro area get with just a dusting of snow or slightly icy conditions. I'm a bit worried that California-based automated driving companies will miss lots of common corner-cases associated with Winter driving in Alaska / the upper Midwest. I expect that if the road is uniformly icy, an AI would adapt fine with the assumption that traction levels here are the same as those up ahead. However, due to lower thermal mass, bridges undergo larger temperature swings and are likely to be icy. There are a bunch of gotchas associated with driving near active snowplows. Sometimes it's best to drive in the ruts created by other cars, but sometimes that's where the road is iciest and it's better to drive offset from the center of your lane. Under normal conditions, your car has a much shorter stopping distance than a loaded tractor-trailer, but in some icy conditions, the much higher tire loading will allow the tractor-trailer to stop in a shorter distance than a passenger car. Many people put sand bags in their trunks (boots) to increase their tire loading during winter.
There are just tons of little rules like this that you're either going to need to hard-code, or get good simulations plus a lot of driving time in International Falls, and hoping that a closed track still gets you conditions that teach the AI when to drive in the center of the the lane and when to drive off-center. Hopefully the AI also learns to visually identify different kinds of icy or likely icy patches.
I'm sure it'll all get worked out, but I wouldn't be surprised if the first Winter or two with self-driving cars ends up with a bunch of cars stuck in ditches in Alaska and the upper Midwest. It might look a bit like the newscasts from when they get icy roads in Texas or Georgia for the first time in decades.
Getting back to aircraft, even for everything that's in the emergency procedures manuals, writing good realistic simulations in order to get enough training data for the AI is going to be very difficult. For situations that aren't in the emergency procedures manuals, both most humans and most AIs are going to be in deep trouble.
They're allowed to have limits like "local driverless taxis don't operate outside SF city limits or below 35 degrees with precip in the forecast" etc. at level 4, but to meet level 5 (per the bet) it has to be able to "drive everywhere and in all conditions," [0] which adds a lot of really difficult edge cases.
Situations that come immediately to mind:
- Driving in the hurricane lane on the shoulder during an evacuation
- Reversible lanes and streets
- Sizing up an icy hill and figuring out whether it's safe to keep going
- Ferries
- Knowing a baseball entering the road from behind a parked car will probably be followed by a child
- Understanding traffic police, sign turners, "follow me" trucks, etc.
I have every confidence that once self-driving cars have done a pretty good job of mastering city/fair-weather driving conditions, that ML will be able to handle poor weather just as good or better than humans. For example, humans are susceptible to panic when they start to lose control of the vehicle and do unproductive things like lock up the brakes. Even experienced drivers do this. Self-driving autos won't.
But yes, in the short term, self-driving is a non-starter for many regions during winter months or inclement weather.
Their head of AI talked about this during their AI Investor Day and he said that even when the road is entirely covered, there is still enough subtle clues for the neural net to pick up on that it will be possible to have a self-driving car in the snow. Just like how humans are able to observe the edge of the road (enough) to slowly drive to their destination, the car will pick up and learn from those clues too eventually.
At least Scenario 2 should be much easier for self driving cars to manage than humans since they can (at least in theory) monitor the dampness, temperature and traction of the road they are on in real time. There are camera systems today that can detect black ice with combination of IR and image recognition, and in the future there might be systems like that along most roads that can send real time updates about upcoming hazards directly to the car.
Scenario 3 is functionally no different than a truck driving slowly and weaving back and forth erratically, something that should be part of the standard scenarios any self driving car should handle.
Scenario 1 is almost certainly the hardest one and the one that will take longest to solve.
An open problem that Google has admitted with their car automation project is a certain problem endemic to the northern part of this country: Snow and Ice.
Admittedly, ice is rather difficult in any condition. However, many human drivers drive safely on snow cover with minimal problems. Accidents are higher during those times, but traffic still keeps flowing. Google-car-AI cannot handle these events at all (at this time).
>The one sensor that is guaranteed to have sufficient information to drive in all conditions is vision.
Minor nit but humans can't drive--certainly not safely--in all conditions. You can certainly get to a point in fog, blizzards, and even very heavy rain where you really would like to get off the road if possible. (Not always possible of course and in snow particularly, pulling off to the side of a highway isn't a great option.)
The scenarios you are describing are incredibly common, and trivial to develop for in comparison to real edge scenarios (not to say they are trivial to develop for). More challenging are things like, "How do you respond to a homeless guy standing on the street directing traffic into an alley", "How do you respond to a flooding on the road resulting in 6" of standing water obscuring all traffic lane markings", "How do you respond to a flooding on the road resulting in 18" of standing water? 24" of standing water?" How do you respond to a road covered in snow (12", 18", 24")? How do you respond to a road covered in snow with tire prints that seem to go into the oncoming lane? How do you respond to a road covered in snow with what appears to be a small 24" high snow drift in your lane?
Etc, Etc..
I think the answer to all these questions is, "Initial versions of Automated vehicles will only perform under ideal road scenarios. During Snow, Flooding, or other inclement weather, they will not be dispatched. Wait for future iterations in 2030, 2040, 2050, etc...
I think there are better potential solutions for a self driving car to orient itself in those kinds of conditions than for a person. I've driven in miserable weather, and it is very disconcerting - you just guess and go. A self driving car could take measurements of the road + gps coords + look its map database to know how many lanes there should be, then estimated where the lanes and limit lines are. Once my eyes fail me, I don't have a fallback.
The problem with the mountains though is that these conditions can manifest themselves in a matter of minutes. Last time I drove through the mountains, it was sunny on the way up, but as we were leaving it was already overcast and within 10 minutes we were in thick fog and a near blizzard. Also, roads completely iced over 1/2 way through the 30 minute drive back down the mountain. I'm not sure any AI would have done better than I did in this case.
The edge cases are where the real challenges are.
reply