Self Driving Cars?

[ERD50]

...

Perhaps you have not seen the TED talk that I posted earlier. It shows how Waymo tries to address a much tougher problem than "lane following" on the freeway. I will repeat here that people have done freeway cruising long ago.

....

PS. OK I will repeat that link here for people who are interested and want to know what is involved. Skip to about 7:30 if you are impatient and want to see recorded data of what their computer saw. This is meant for people who want to understand more about self-driving car technology, not just some demo for ordinary people.

[see link in post]

Fascinating (and thanks for reposting, this is a loooong thread, stuff is getting buried!), but some of his comments seemed kind of twisted.

At 4:15-4:45 he described how the test driver let himself completly take his eyes off the road for a long time at 65 mph (grabbing the cable for his phone, plugging it in, etc). That was scary.

But his response was that we need the car to be good enough to completly take over. I don't agree - why not take measures to keep the driver involved? Then you have the 'eyes & ears' of the car, plus the eyes & ears of the human, working together.

He also seemed to be defining "driver assistance" as only being 'dumb' or something? That it can't develop into autonomous driving? Why not? Why can't the assistance get better and better - that seemed like a false view.

And I really didn't follow at 6:35 - 6:50. Did he say drivers make mistakes that lead to accidents once very 100,000 miles? That's about 6-7x the numbers we have from those papers, but OK, within an order of magnitude. But my bigger question is - doesn't he leap from accidents/100,000 miles to the number of decisions per mile that the autonomous system must make? What's the connection?

Again, I'm going to say that driver assistance plus a system that monitors the driver to keep them engaged and aware is more doable, and likely to save more lives, sooner, than waiting for a car that can function without a driver. It's the old 'perfect is the enemy of good' story.

And I still can't see why the assisted tech can't keep developing. When we get to the point that the data says "We kept the driver alert and engaged, but we never requested the driver to do anything for X million vehicle miles", then maybe it's time to say OK, we don't need the driver. Seems like a natural progression to me. And as I said earlier, by the time we get there, we may not even need it, other approaches will take over.

-ERD50

 

[NW-Bound]

Ah, finally someone who I can discuss this with. I kept trying to present some data, but was called a naysayer.

Chris Urmsom described how some beta users of his system back in 2013 had so much more confidence in the test cars than he himself had. There was that guy who fumbled around at 65 mph, reached in the back of the car for the laptop, etc..., despite the instruction to keep eyes on the road, hands on the steering wheels. Exactly the behavior shown in another video that I posted of the "Look ma no hands, I can even sleep" of some Tesla drivers.

So, Chris said that he had to go to Level 4 (full automation), because he could not count on the driver to take over when a lesser system needed help from a human. Many other car makers seem to agree. They are all trying to do Level 4. For example, see Cruise Automation demo video posted by Eroscott above.

Then, he lost me at 6:35 same as you.

He quoted the number of accidents at 1 per 100K miles. Using Midpack's number of 17,249 accidents/day (including fender benders), and the number on the FRED site of 8,500 million miles/day, I got close to 1 accident per 500K miles.

OK, perhaps Chris used urban driving accident number, while the FRED number includes highway miles, which tend to have less accidents.

Let's assume that the 5,000-mile number between manual intervention on his test cars is also urban driving, then Waymo system currently requires 20x more interventions than human-caused accidents. What would happen if the test drivers did not intervene? Big crash, or little fender bender, or just a close call? We will never know, but the frequency of human assistance is still high.

And that's why he said his system was not ready. This is a honest guy.

PS. By the way, the report from California from which I quoted the numbers shows that Waymo was the best of the pack. The TED talk shows a very good understanding of the problem that they are trying to tackle, and what they were achieving. Pretty darn good, but it needs to do better. And there's that rain and snow problem that impacts the lidar performance.

 

[Midpack]

There is a huge difference between self-driving/autonomous and driver-assistance. If you don't understand that, it might explain why you are so positive on the idea of self-driving/autonomous cars.

That's funny. I'm one of several people who posted the level definitions earlier. And I've referenced "level 5" in several posts instead of leaving it ambiguous, and noted others were mixing driver assisted and autonomous without noting the distinctions.

But they aren't mutually exclusive. Many of the driver assistance technologies are very similar to those used in level 5 cars. Some have been/are being developed for level 5, intermediate steps, providing a partial benefit today.

You are playing word games with 'potential' - you aren't talking just 'potential', you're talking about this tech becoming available in the foreseeable future, like our driving lifetimes.

In the same post, I included a chart with one authors theoretical timeline for self-driving cars. The technology still has a long way to go but many people believe we'll see level 4-5 cars in our lifetimes, and I hope it's true, but I don't pretend to know. You know they won't be a available? Beyond cars, you've predicted how quickly (or not) innovation has taken place in the last 30 years?

 

[Ronstar]

NW - thanks for reposting the TED video. Very interesting.

I believe that the self driving car technology will be implemented in stages. First - interstate highways. Less anomalies on the interstate. Then open up to highways and eventually city streets when safety is up to snuff. Hope it works and reduces traffic.

 

[JoeWras]

Ah, finally someone who I can discuss this with.

I wasn't good enough?

Seriously, last night I was presented with a scary driving situation that I would like to see the current crop of autonomous vehicles handle. Most of the videos we see are on nice days in sunny locals on good streets. Any videos from the Pittsburgh Uber experiment yet?

Anyway, that scenario: we had strong winds yesterday. A road is under construction and already has a mess of switch over lines, half ground down with ghosting, etc. The road is dynamically changing day to day. On top of that, the strong winds actually moved the cones and signs! They were pointing every which way. Humans were making sense of this complex mess. I honestly would love to see what the current technology would do with this today, just as a check point.

 

[NW-Bound]

Oops, sorry.

That's why Google (Waymo) said that they could not predict all the things that could happen, and had to go out to collect data and real-world situations.

Once we get past the "wow" factor of current experimental cars, and think about it, we will see that the human cognitive abilities are amazing, and not at all easy to transfer to a computer. Computers are fast and precise, but they are digital and the real world is analog.

Once we get the software to work well enough, before we deny the human driver the steering wheel, we have to worry about hardware failure. The current accident rate is 1 every 500K miles. What is the current hardware failure rate that Waymo is seeing? I will report on this later.

 

[Midpack]

I want to reduce accidents too, and as mentioned several times elsewhere, I do not care to drive and do not love my car like some people do. I do not even mind an ugly car with a Lidar on top.

But you bring up a good point, and make me go check some numbers. FRED site says that Americans drove a total of 3.1 trillion miles in 2015. Egads!

That works out to 8,493 million miles per day. Holy moly! But the USA is a big country. Let me double check the number by dividing it by the population of 319 million. That works out to 27 miles/day per capita. Ah, that's reasonable.

So, out of 8,494 million miles per day, Waymo cars would require human intervention 1.7 million times each day. Whoa! They need to be 100x better to match lousy humans.

PS. Now Waymo would not go out and do test drives when it rains and snows. That would bring the 5,000-mile/incidence number way down. Not good.

A fair illustration of how much further the technology has to come to match humans and cars today. Waymo isn't ready for prime time, no one has said it is.

As admittedly complex as driving is, I believe the technology will far surpass humans, others may believe otherwise. I believe we'll have level 5 capability in our lifetimes, and level 4 long before.

When I look at my iPhone today, it reminds me to be humble about my ability to foresee innovation. And I've seen this progression with technology way too many times to give too much weight to the naysayers predictions - All truth passes through three stages. First, it is ridiculed. Second, it is violently opposed. Third, it is accepted as being self-evident. Arthur Schopenhauer?

 

[NW-Bound]

And I've seen this progression with technology way too many times to give too much weight to the naysayers predictions...

If someone identifies the difficulties and remaining obstacles to conquer, is he a naysayer?

How does one solve problems and make progress if he does not understand what the problems are?

 

[Midpack]

Not recent, but some basic insights:

Why are self-driving cars said to be the answer?

Because as brilliant as your brain is, it has a many limitations. A computer can react almost instantly to a hazard, whereas you take precious milliseconds – and that's assuming you're even paying attention in the first place – and aren't affected by adverse weather and visibility conditions.

Moreover, autonomous cars can 'see' a lot more than we can. In Google's car, the LIDAR system on the roof (located there so as to offer an unimpeded view around the car) acts like a really advanced version of the Microsoft Kinect motion sensing camera, reading the surroundings up to 1.3 million times a second.

Autonomous cars can see around corners (if they're networked they always know where other vehicles may be), over blind crests and have 360 degrees of vision. Computers also don't get angry, fall asleep at the wheel, spill their coffee, get put off by whining kids in the back or drink alcohol.

Are there other benefits?

There are also the aforementioned safety benefits of a machine that can do its thing all day, every day without getting tired, so trucks never need to stop while the driver has a rest. Talk about improved productivity.

It removes the likelihood of human stupidity, too, which is said to be the culprit in 90 per cent of road accidents in the UK [USA too]. Elderly people, meanwhile, would be able to get around regardless of their health, removing the isolation associated with giving up driving for good.

So what are the negatives?

Self-driving cars have the potential to be safer, in theory. But unlike the autopilot on a plane, which only has to steer through the clouds and avoid other planes, an autonomous car has an insane number of variables to watch and react to.

Machine learning is getting there, as is software, but autonomous cars aren't infallible. Computers crash, GPS systems lose reception, software is hacked – we could go on.

The brain is an incredibly complex computer itself that processes huge amounts of data it sees, hears, feels and thinks while you go from A to B. It may react more slowly, but it can compute complicated events better than any machine, and emotional decisions are impossible for microchips.

Would a machine know, for instance, when a traffic light is out of action or how to get out of the way of an emergency vehicle? What if a mum and child step out in the road – can and should a computer decide whether to try and save the former or latter if that's the only choice?

https://recombu.com/cars/article/self-driving-cars-everything-you-need-to-know

 

[Midpack]

And I've seen this progression with technology way too many times to give too much weight to the naysayers predictions

If someone identifies the difficulties and remaining obstacles to conquer, is he a naysayer?

How does one solve problems and make progress if he does not understand what the problems are?

I said "too much weight," not none. They need to see and solve problems, and they are.

Probably just me, but while some of your posts on this thread have been insightful and thought provoking (despite our different POVs, I've enjoyed this thread), some seem to confidently depict some remaining problems as impossible to solve.

In retrospect, this may be a simple as I'm glass half full and you're glass half empty.

I don't have much more to contribute to this topic anyway...we can only wait and see how this plays out. It will be fascinating to watch.

 

[Amethyst]

But no driver is an island. If you are on the road with 100 other people, what are the chances that one of you is about to make that 100K-mile mistake? (I'm not arguing with anything you said - it just occurred to me).

Fascinating (and thanks for reposting, this is a loooong thread, stuff is get Did he say drivers make mistakes that lead to accidents once very 100,000 miles?

-ERD50

 

[NW-Bound]

I have no axe to grind, and do not care either way really, other than being alarmed at the cavalier attitude of Tesla vs. someone as responsible as Waymo. I will get back to this later.

As I said many times before, I don't think that a reliable Level 4 car will happen as soon as some optimistic predictions. Or if it does, it will not be as cheap as people hope, but I do not mind getting pleasantly surprised.

I do not get to work on this interesting problem, but as an engineer I am curious to see what they are after. The software required is quite complex. Anybody who has written software appreciates that. And as I understand their problems, particularly the intrinsic limitations of the sensors they are using, I will be more impressed when they can get around them.

 

[NW-Bound]

But no driver is an island. If you are on the road with 100 other people, what are the chances that one of you is about to make that 100K-mile mistake? (I'm not arguing with anything you said - it just occurred to me).

The 100K-mile number applies to innocent drivers too, as it is the average. A bad driver causing more accidents will have a much lower number. Good drivers have a slightly higher number. There are more good drivers than bad, and one bad driver brings down the number of 100 other drivers.

 

[JoeWras]

When I look at my iPhone today, it reminds me to be humble about my ability to foresee innovation. And I've seen this progression with technology way too many times to give too much weight to the naysayers predictions - All truth passes through three stages. First, it is ridiculed. Second, it is violently opposed. Third, it is accepted as being self-evident. Arthur Schopenhauer?

I think we agree pretty much this is our future.

I think we are not in agreement at the pace it will come about.

I think we are also discussing the complex issue of insertion of this technology into an existing infrastructure. I think I'm in near agreement with Midpack on the timeframe of reliable level 4 or 5 self-driving cars, but I'm in disagreement as to how they will insert into the existing system.

 

[Midpack]

I think we agree pretty much this is our future.

I think we are not in agreement at the pace it will come about.

I think we are also discussing the complex issue of insertion of this technology into an existing infrastructure. I think I'm in near agreement with Midpack on the timeframe of reliable level 4 or 5 self-driving cars, but I'm in disagreement as to how they will insert into the existing system.

Have you posted your thoughts on the latter? It seems driving assistance features so far are being assimilated easily, but level 4-5 may not be so easy. I have a vision of what the world may be like when level 5 cars have taken over (50 years?), but I'll be the first to admit the transition could be a tangled mess.

 

[NW-Bound]

... Probably just me, but while some of your posts on this thread have been insightful and thought provoking (despite our different POVs, I've enjoyed this thread), some seem to confidently depict some remaining problems as impossible to solve...

Not in an inexpensive way, using the technology available today. I only know about what is out there today.

Somebody may emerge from his garage with a superior and cheap sensor tomorrow. But that is conjecture.

For example, I mentioned the use of stereoscopic cameras to get distance to objects, and cameras are cheap. I am sure that occurs to all developers, but that takes huge amount of CPU power. Big cost!

Now, maybe someone discovers a brilliant algorithm. Bingo! But as a lot of people have been working this problem, as more time goes on, the chance of it getting discovered or rather invented becomes slimmer.

 

[meierlde]

Have you posted your thoughts on the latter? It seems driving assistance features so far are being assimilated easily, but level 4-5 may not be so easy. I have a vision of what the world may be like when level 5 cars have taken over (50 years?), but I'll be the first to admit the transition could be a tangled mess.

Actually its not so clear that the assistance features are that popular (all be it that they add $1000 to the price. If you look at say Honda dealers inventory only a small percent have the Honda Sensing feature (which is the Honda name for current assistance tech) Or note that Chevy regards the Driver Confidence package as a feature only on premium cars. I talked to some folks that don't want even standard features such as stability control. One guy did not like hill start assist either. Of course this is tied into folks that want a clutch pedal in their cars also. Of course all that stuff is more stuff to go bad, many perhaps want to go back to the simplicity of cars of the 1960s.

 

[Midpack]

Actually its not so clear that the assistance features are that popular (all be it that they add $1000 to the price. If you look at say Honda dealers inventory only a small percent have the Honda Sensing feature (which is the Honda name for current assistance tech) Or note that Chevy regards the Driver Confidence package as a feature only on premium cars. I talked to some folks that don't want even standard features such as stability control. One guy did not like hill start assist either. Of course this is tied into folks that want a clutch pedal in their cars also. Of course all that stuff is more stuff to go bad, many perhaps want to go back to the simplicity of cars of the 1960s.

I wasn't commenting on how popular they are, though that's relevant to frequency. My only point was, there are some cars on the road with all of us now using some driving assistance functions - without any of us needing to know or adapt. No idea how often, but I am sure I've driven alongside cars with lane keeping deployed, adaptive cruise control deployed, during auto park, maybe even collision avoidance or driver drowsiness detection.

Presumably level 5 vehicles won't be as easily assimilated, but that remains to be seen. I am sure the first times I see a car with no driver, I'll be startled first, then amazed.

 

[Amethyst]

We like: lane departure alarm, blind spot alarm
We LOVE in-car navigation (I know many people use their smartphone GPS, but we prefer the on-dash application)
Haven't cared for: Smart cruise control. We don't do that much long-distance driving to even need cruise control, and the "smart" (set a distance between you and the next car) feature has been a nuisance. The smart distance is impossible to maintain in mixed driving conditions...I mean, when I used to commute to the Pentagon, it was common practice on the route home to close up ranks till one car's nose was practically on the next car's bumper.

 

[ERD50]

....

I do not get to work on this interesting problem, but as an engineer I am curious to see what they are after. The software required is quite complex. Anybody who has written software appreciates that. ....

As an aside, does anyone have links to any basic explanations of how this kind of software is developed?

It's obviously not procedural based, you just can't write enough lines of codes that cover all the possibilities:

IF (BICYCLE approaches from LEFT) AND (>1 CAR IN_FRONT) AND (>1 CAR BEHIND) AND, AND< OR ad infinitum

As I (barely) understand, AI depends on weighting of a bunch of different conditions. So each identified object gets some sort of weighting of importance to the decision process. So those far away, slow moving pedestrians get a low weighting, a nearby pedestrian moving towards us get a very high weighting, as does a car moving towards us.

Then maybe they have thousands of 'scenarios', and pattern match to a scenario that is at least similar to the current condition? So driving on the highway isolates the decision process to one subset, while approaching an urban intersection is a different subset to analyse?

And I wonder how the programmers handle a situation that needed an intervention. Do they go in to the simulators, and tell the computer "watch what I do and learn from it"? And then maybe do dozens, hundreds of variations of that scenario so the computer can build up knowledge of similar situations?

It's somewhat mind-boggling. But then again, if you break down just about anything our computers do, it's mind-boggling. Right now, my computer is reading my keyboard, processing it on a display, spell checking, one more mouse click and a whole series of actions occur, formatting this all up, byte-by-byte sending it over wi-fi, getting feedback on that process, caching it to my hard drive, updating my display - all while handling hundreds (thousands?) of tasks in the background.

Their not AI based, but it does appear to us as an intelligence of sorts.

-ERD50

 

[NW-Bound]

AI software is not in my realm. I thought about learning about neural network a while back, but never got into it, and I am not going to do it now.

The Waymo (Google) guy talked about how they captured and stored all the sensor data collected in their drive tests for simulation afterwards. It makes a lot of sense. For example, if they improve their software to identify, track a bicyclist and project his trajectory, instead of going out to collect data all over again, they can just play back existing data set. And they keep adding and adding to their real-world experience.

Note that it is why a special test car is needed that has equipment for storing and uploading huge amount of data. You cannot get this kind of data from a bunch of ordinary cars. You would not know why something does not work in some cases, and may not be able to duplicate those conditions again. Google guys are pretty smart (as usual).

 

[NW-Bound]

I mentioned stereoscopic computer vision earlier and the difficulty in using it. However, I forgot to bring up something else. That is, a human can learn to do well with just one eye.

Yes. Think about it. People who are blind in one eye do not stumble all over the place.

So, how do they do it? Part of it is from memory when they navigate familiar surroundings. Part of it is from cognitive reasoning. They know from experience how high a step is, how wide and long an object is, etc... Part of it is that the brain and its memory can determine the distance to objects when the line of vision changes. It's the same as having multiple eyes, but the brain combines the vision at one instant to that a second later but at a different viewpoint when one moves. That's how one can get the depth perception with one eye.

Similar to the combining of vision from different viewpoints is the SAR (Synthetic Aperture Radar). An aircraft looking down with a radar while flying can combine the images later to have an extremely fine view of the ground. It takes special equipment for this.

Anyway, while we are in awe of the digital computer, we should not forget to pat ourselves on the back for the awesome cognitive power of our brain.

 

[marko]

I've not read through every single post here but with all the concern for safety and accidents one would think that the current intelligence behind the wheel (us) was foolproof!

 

[ERD50]

We like: lane departure alarm, blind spot alarm
We LOVE in-car navigation (I know many people use their smartphone GPS, but we prefer the on-dash application)
Haven't cared for: Smart cruise control. We don't do that much long-distance driving to even need cruise control, and the "smart" (set a distance between you and the next car) feature has been a nuisance. ...

I recently bought a new car, and was able to convince my old LBYM self to go several trim levels up to get some of the driver assistance features. I don't really drive much, so I'm still getting accustom to these features. But so far:

I absolutely love the blind spot detection. I'm not going to rely on it, but it's nice to have a system helping you out, providing a little more info. It's more like having a passenger with you, helping you out. I'm sort of mentally thinking, when I see the indicator "Yep, I'm aware of that car there, thank you.".

lane departure alarm: It's interesting. I don't think it has ever warned of me something I wasn't really aware of. Sometimes I will intentionally drift a bit, like to avoid something on the road, or maybe cut a corner in a bit more of a straight line, and it will let me know. But it does keep me a little more involved I think, like "hey, I'm watching you, so pay attention!".

Front Collision Detection: This one doesn't actually apply the brakes, it gives a green car symbol on the dash when a car ahead is in it's view, it turns yellow if you approach it too fast, then will beep and flash red light on the window if you get really close. I set it to the most sensitive setting, and was able to get it to flash a couple times when I intentionally 'pushed the envelope' to test it.

I like them all, the added info seems to keep me more engaged, rather than kicking back to let them do their thing (though of course, they aren't anywhere near that level anyhow).

Don't have adaptive cruise control, so can't comment.

And also, rear camera, with side alerts. I love this. I can see so much more when I back up than in any car I've ever owned. The side warning does a good job, almost too good, but that's OK. It will alert me to someone pushing a cart in the parking lot - they are moving slow, so I can back slowly w/o worry, but it's warning me. I kid my DW, that it takes me too long to back out of a parking space now - in the past, I'd move slow in my old car, and people would get out of the way - now I feel guilty backing up if this is beeping at me!

Front and Rear object detection - helps in parking a bit, you know how close you are coming to something. Seems pretty good.

These features are advancing pretty fast, I'll bet that I don't keep this car as long as I have kept cars in the past, I'm going to want the latest features, and won't wait 11 years for them (actually went 16 years in my old Volvo, which was safe for the time in terms of the air-bag/restraint systems).

-ERD50

 

[ERD50]

I've not read through every single post here but with all the concern for safety and accidents one would think that the current intelligence behind the wheel (us) was foolproof!

No, that's not it at all. We posted some stats on current drivers. As I posted recently (I'll just repeat here, this is a long thread)...

Since we are getting better and better driver assistance features on cars, the relevant question is not whether a self-driving/autonomous car will have fewer accidents per mile than a current car with an average driver. The relevant question is whether a self-driving/autonomous car will have fewer accidents per mile than a car with driver assistance features, and especially one that keeps the driver involved (which I don't think we have yet - monitoring the head/eye movements as I mentioned before).

If the autonomous car is going to notify a driver when it can't figure out what to do, and that driver hasn't been paying attention, it is very likely too late for that driver to get involved look around, analyze the situation, and respond.

-ERD50