Friday, November 24, 2017

Autonomous Cars - Part 4

I don't know about you, but my conclusion about radars was that the problem of detecting things, how far away they are and relative speeds is doable.  A radar can detect a large signature like a car at any distance you expect to find in traffic.  Radars routinely find objects with a smaller cross section than a car over hundreds of miles; tracking a car "7 car lengths" ahead is trivial.  The accident I started this discussion with, a Tesla on autopilot decapitating its driver by going under a tractor trailer, was due to Tesla misusing the system by configuring it to "avoid false braking events" (by either the software response to strong returns or the position of the antenna in the car.)  Note to Tesla fanboys: I know NHTSA absolved Tesla of any fault, saying it's not their problem.  I disagree with NHTSA and think they were stupid. To quote EETimes
Did NHTSA let Tesla off the hook too easily? Absolutely.
Likewise, the problems raised in the comments here about cars interfering with each other or signals bouncing off of tunnel walls or bridges and other common examples are not fundamentally different from problems managed successfully in other systems.  Yes we have problems with cellphones and other commercial systems.  There's a fundamental difference between systems that were designed to be robust in the face of understandable problems and systems that aren't.  The high reliability world approaches things differently than the commercial "let's slap this together and ship it" world.  Your wired phone in your house will fail if everyone in the neighborhood tries to use their phone because the systems are designed for about 10% of the lines being used.  It doesn't have to be that way, it's just cheaper that way and has been proven to work well enough. 

As I said in the section on the image recognition piece, I think we have to conclude an image system can't do everything we need it to do, but could it "work well enough"?  I don't think so.  The problem with image rejection systems is that they take millions of training sessions to be reasonably accurate for a small set of things, and here we have a nearly limitless set it has to recognize and know what to do about.  Do we need nearly infinite training time?

In a larger sense, the problem with the whole idea of Advanced Driver Assistance Systems (ADAS) is that individually, all the pieces of hardware can be envisioned to do the job, but taken as a system, there isn't a chance that the software can work.

If we think of autonomous cars as "cruise control that includes the steering wheel" and simply stays in its lane, even that's getting dicey.  A fully autonomous car that you get into and say, "take me to work" or "take me to grandpa's house" and does so while you just sit as a passive cargo is probably more like 50 years away.  I did five hours of driving yesterday and I really would have liked to not have to deal with traffic in the rain and let JARVIS drive, but JARVIS is a comic book character.  Note to fans of the the "Singularity" when computers suddenly surpass the sum of all human intelligence: let's just call it vaporware.  People have been talking about this coming for close to 30 years.  It's much like nuclear fusion reactors that I first read were "20 years away" back in 1971.  If it happens before then, all bets are off, but as I've said many times and places (most recent), AI is over-hyped.  

Driving is a perfect example of the old saying, "hours and hours of boredom interrupted by moments of shear terror".  What do you do when you see a truck stopped on your street for a delivery?  You probably slow or even stop to look for other traffic and an opportunity to go around the truck.  Unless the ADAS car is programmed to do so, it's not going to know what to do.  What if there's a broken down car in the middle of your lane?  Or any one of a thousand oddities you see on the road in the course of a year.  Those are easy problems.  Without true intelligence, the software has to recognize it's a truck, understand the laws, understand what it can and can't do and choose the right option.  What about watching for kids darting between cars on the same street, or riding their bike on the shoulder of the road?


As I concluded last time, it's impossible to teach the computer "if a child runs out after that ball, slam on the brakes, and if you can't stop, hit something like a parked car".   If it was my kid in the intersection, I'd still prefer a real human mother to any computer driving a car because a real human mother is very likely going to care more about any child than damaging her own car.  A computer isn't going to understand the concept of "child" or "person".  They need to be much more sophisticated AI systems than we have now.  I'm not talking about the actor-voiced "Watson" on the commercials; I'm talking about a really advanced Turing test where you could talk to the computer for a long time and not know it's artificial.  Good luck with concepts like "do I hit the adult on the bike or injure my passengers by hitting the parked bus?" 

The truth of the matter is that driving has difficult moments and not all accidents are due to someone drunk, texting, or on the phone.  Some accidents come down to not having good options left.  

As many commenters have pointed out, the biggest risk in these systems is the combined mad dash to put them into place on the part of industry and the Fed.gov itself.  The public is less sanguine about autonomous cars, with a AAA survey I reported on in September showing
...three-quarters of Americans reported feeling afraid to ride in a self-driving car. One year later, a new AAA survey found that fear is unchanged. While the majority are afraid to ride in a fully self-driving vehicle, the latest survey also found that the majority (59%) of Americans are keen to have autonomous features in their next vehicle. This marked contrast suggests that American drivers are ready embrace autonomous technology, but they are not yet ready to give up full control.
Commenter Dan said on the second post,
I suspect that sooner rather than later we are going to see one of these autonomous vehicles murder a family of children or a school bus full of kindergarden kids and the legal fallout will end this experimentation for a very long time.
and I've frankly been waiting to hear that news with dread.  I'm sure the ambulance chasers are chomping at the bit for it to happen.  They like to go after the deepest pockets possible and among the deepest pockets possible are in play here.  I can see them going to court even though the "expert" NHTSA ruled in Tesla's favor in the decapitation accident.  Juries are under no requirement to follow the  NHTSA, in my understanding.  If a jury awards a big punitive award in a case like this or like Dan describes, it could well put the development of these systems on ice. 
Now what do I do?


14 comments:

  1. Replies
    1. It's either the only thing they allow or they break it into four moves and fine separately for each one.

      Delete
  2. It is not affordable. The closest technology to self driving is the aircraft industry moving passengers. Primary systems are expensive and backup secondary systems add more cost. When using backup systems and deferring maintenance there are additional rules put in place. An aircraft can spread that cost out over a couple hundred people but an auto will only have a couple of people to spread this cost. Maintenance costs and reliability go hand in hand. Lawyers love to get involved in air crashes...and they will love self driving vehicles. The Teslas of the world will be in bankruptcy in mere months. indyjonesouthere

    ReplyDelete
    Replies
    1. Yeah, but Tesla may "be in bankruptcy in mere months" from now!

      Delete
  3. Before the advent of the Automatic Information System (which broadcasts position, course and speed) being seen by commercial vessels was a real concern of recreational sailors.

    There are a variety of radar reflectors you can hang in the rigging. None of them work all that well, or only in extremely limited circumstances. Very little metal in a fiberglass sailboat, and most of it is at or below the waterline.

    So while the radar in a Tesla may find a pickup truck, what will it miss? Pedestrians? People lying in the road? Random junk that could cause an accident? And sailboats travel at about 10 to 15MPH (mostly) and commercial vessels 35MPH or so. How much harder is that going to be at 70mph?

    ReplyDelete
  4. "It's impossible to teach ethics to a computer. It's impossible to teach the computer "if a child runs out after that ball, slam on the brakes, and if you can't stop, hit something like a parked car". A computer isn't going to understand the concept of "child" or "person". Good luck with concepts like "do I hit the adult on the bike or injure my passengers by hitting the parked bus".

    You reach this sweeping conclusion based on exactly what? An article in a newspaper quoting some self-serving study by MIT students trying to make a name for themselves? AFAIK neither you nor "Borepatch" have any background in modern AI. And by modern I mean the last few years since the field is moving so fast. In an article you quote your teacher in a physical optics class. Well, optics is involved in the design of the camera but not at all in what you do with the images it produces. Neither you nor your teacher really know what you are talking about regarding the limits of AI.

    You quote Commenter Dan; "I suspect that sooner rather than later we are going to see one of these autonomous vehicles murder a family of children or a school bus full of kindergarden kids and the legal fallout will end this experimentation for a very long time."

    Well how about this quote: "In 2010, there were an estimated 5,419,000 crashes (30,296 fatal crashes), killing 32,999 and injuring 2,239,000,[2] and around 2,000 children under 16 years old die every year due to traffic collisions.[3] Records indicate that there has been a total of 3,613,732 motor vehicle fatalities in the United States from 1899 to 2013."
    https://en.wikipedia.org/wiki/List_of_motor_vehicle_deaths_in_U.S._by_year#See_also

    So it's not like the current situation is acceptable. We have gotten used to this slaughter of people on our streets and think that it is the natural order of things--but it is not. According to AAA, "Car drivers were to blame in 75 percent of the accidents." We may all think we are great drivers but the evidence is we are terrible at it. The bar for AI driving systems is much lower than you portray.

    How about the use of resources. Truck driver is the largest single occupation in the US. Driving uses up a huge fraction of everybody else's time. As an engineer, you can easily do a back if the envelope calculation to see that if the average person drives 12,000 miles per year at an average speed of 30 miles per hour that is 400 hours spent driving. Compare that to the total hours at work, 40x50=2000. That is a huge amount of productivity wasted.

    Sebastian Thrun of the Google self-driving car project points out that on a heavily crowded freeway that is still moving at the speed limit, cars only occupy about 10% of the road. If, with self-driving technology, we could increase that to 20% we would double the carrying capacity. The savings would be huge in a more efficient use of this extremely expensive and disruptive to our environment resource--the roads.

    I am also an engineer graybeard, 69 years old with a PhD in engineering from Stanford and worked all my life in techology research and my conclusion is way different than yours. There is a confluence of a terribly destructive and expensive current situation and a rapidly expanding state of the art so I think driverless cars are inevitable and I am really looking forward to it.

    Bob the Engineer

    ReplyDelete
  5. Re:
    Did NHTSA let Tesla off the hook too easily?

    When you own the right politicians, you get the decisions you want. But then surely Elon Musk would NEVER stoop to that level...

    ReplyDelete
  6. "Sometimes it come down to having no good options left"
    Sounds like a new branch of System Safety Assessment;The Options cone.... probably too many random/unquantifiable variables, but an interesting thought exercise nonetheless.
    Differ

    ReplyDelete
  7. Autonomous cars are a Hindenburg type of idea, and in due course, they'll achieve a Hindenburg outcome. Mind the flaming bodies as they fall, and try not to book passage on that ride.

    I also note that in general, the less technically literate a person is, the more they like the idea, and conversely, the more savvy someone is, and the better they understand the actual problems, the less they want to ever be within a country mile of an autonomous car, ever, in their lifetime.
    {cf. Harry Potter/Magical thinking/Underpants Gnome}

    ReplyDelete
  8. While it is certainly possible that a particularly gruesome f**k up by an autonomous vehicle could result in them being if not banned then at least curtailed. It is also quite possible that the people and companies pursuing autonomous vehicles could get legislatures and perhaps even congress to pass laws exempting them from damages their vehicles cause. With enough money and ownership of the right politician just about any law is possible.

    And there has to be a motive behind the massive push to create driverless vehicles. It's not as if the technology is cheap, it's not. It's a very expensive bit of engineering to design, create and implement. It makes one ask why....why are they trying so hard to get this technology into the real world. The realist in me leads me to conclude that such technology will assist those in power with what they enjoy most. CONTROL. If left unfettered eventually autonomous vehicles will become practical ( a term that is subject to debate of course). Eventually the technology will become widespread.....and once it does the power mongers will do what they always do....legislate. They will seek to make it illegal to use a vehicle that is NOT autonomous....because "do it for the children" etc. Once they succeed in banning vehicles that humans can control they will essentially have TOTAL control over all transit and travel in America. And THAT is worth the cost....both in $$$ and in lives...at least to a politician.

    ReplyDelete
    Replies
    1. I'm not really going to go there. I think it's a possible justification, but that's really playing the long game.

      My explanation is a bit too long for this field, so I think it will be another post.

      Delete
  9. I am all for my vehicle making information available to me. Right now it tells me how fast I'm going, the coolant temperature, the oil pressure, how much fuel there is in the tank and so on. I recently drove a 2010 Ford Mondeo that a friend of mine wons. He's an avionics engineer and he's modified his car to give a heads-up-display (HUD). This gives speed, warnings of speed limits ahead and an indication of any warning light coming on on the dash display. It's very useful and conducive to good driving.
    I think that this is the way to go, giving the DRIVER pertinent information in a timely way. Allowing a silicon chip to control the car is NOT the way.

    ReplyDelete
  10. This comment has been removed by the author.

    ReplyDelete