Autonomous Cars

[Andres125sx]

Maybe the problem with the comments above is humans tend to confuse assistance functions with autonomous functions

That Tesla driver crashing with the white truck because Tesla AP didn´t see it due to the bright sky while he was reading a newspaper instead of keeping focus on the road as Tesla recommend is a good example. Take an assistance funciont as if it would be autonomous, then when it does not act as an autonomous function, blame it

Problem then is not the function itself, but humans assuming it should do more than it is designed to do

[AngusF1]

"Anyone care to posit how autonomous vehicles will be made to work in a snowstorm? road grit, snow and ice will be a significant impediment to cameras and wheel speed sensors are useless at best, a hindrance at worst. "

Yup. Before AVs become ubiquitous we will have to redesign our roads. They'll effectively be slot car tracks. There will be wireless sensors embedded along the road way and the cars will drive between them.

Given the current state of technology, yes, although putting electronic signals or sensors in the roads presents difficulties which make the concept impractical for use on all roads, hence limiting the usefulness of the concept. (Not to say it won't happen. Maybe they'll finally let us do 200 on the freeway?)

IMO the robust solution will be once AI vision is good enough to look through the normal windscreen and windows to interpret the car's surroundings without having to rely on easy-to-obscure external sensors. Amusingly the driver, set well back from all the windows and who can move his head to see around any dust, dirt or ice on them is much better placed to achieve this than any feasibly mounted cameras. (The idea here is that any piece of crud on the windscreen approximately camera/sensor-sized or larger becomes relatively small once the viewer is set back by a foot or two.) Perhaps they'll mount a pair of goggly-eyes on stalks inside the cabin, or hang them from the rear-view mirror?

Developing the AI vision to this point will require (typical Google / Facebook solution) putting spy-cameras and spy-sensors on all cars and car controls (wheel, brakes, throttle etc) on all cars everywhere to spy on everybody all the time to learn how real drivers drive in realistic edge-case scenarios where the sensory information is extremely difficult to interpret. (Is Tesla not doing this already?) Examples: on a moonless night in the driving rain/fog/snow with dirt-covered road and oncoming headlights, being blinded by glare coming over a crest, etc.

And that is the catch with computer-assisted driving; the circumstances where artificial assistance would be most useful are the circumstances under which artificial technologies are most limited, and in road vehicles the timeframe between sensor confusion and catastrophic collision is on the order of a second, which is too short for the driver to intervene. The AI vision people are doing their best, but it is very difficult to do better than the ruthless Darwinian selection behind the eyeball and visual cortex. For 350 million years, brains that guessed wrong under limited information simply died.

[henry]

Maybe the problem with the comments above is humans tend to confuse assistance functions with autonomous functions

That Tesla driver crashing with the white truck because Tesla AP didn´t see it due to the bright sky while he was reading a newspaper instead of keeping focus on the road as Tesla recommend is a good example. Take an assistance funciont as if it would be autonomous, then when it does not act as an autonomous function, blame it

Problem then is not the function itself, but humans assuming it should do more than it is designed to do

Well put.

This misconception isn’t helped when Elon Musk repeatedly tells the world that Tesla’s are ready to be Full Self Driving and all it needs is a bit more processing power and a few software tweaks. I really like the Tesla assistance implementation but it’s nowhere near autonomous.

Legislators need to get their act together. If we really are going to have autonomous vehicles the worlds’ governments need to set up a research group to understand the implications and how they should be regulated. The exact opposite is happening with regional and national governments vying to be as lax as possible to encourage investment. I guess it took a while, and quite a few incidents, before air safety got regulated so I’m guessing AV regulation will follow a similar path.

[henry]

"Anyone care to posit how autonomous vehicles will be made to work in a snowstorm? road grit, snow and ice will be a significant impediment to cameras and wheel speed sensors are useless at best, a hindrance at worst. "

Yup. Before AVs become ubiquitous we will have to redesign our roads. They'll effectively be slot car tracks. There will be wireless sensors embedded along the road way and the cars will drive between them.

Given the current state of technology, yes, although putting electronic signals or sensors in the roads presents difficulties which make the concept impractical for use on all roads, hence limiting the usefulness of the concept. (Not to say it won't happen. Maybe they'll finally let us do 200 on the freeway?)

IMO the robust solution will be once AI vision is good enough to look through the normal windscreen and windows to interpret the car's surroundings without having to rely on easy-to-obscure external sensors. Amusingly the driver, set well back from all the windows and who can move his head to see around any dust, dirt or ice on them is much better placed to achieve this than any feasibly mounted cameras. (The idea here is that any piece of crud on the windscreen approximately camera/sensor-sized or larger becomes relatively small once the viewer is set back by a foot or two.) Perhaps they'll mount a pair of goggly-eyes on stalks inside the cabin, or hang them from the rear-view mirror?

Developing the AI vision to this point will require (typical Google / Facebook solution) putting spy-cameras and spy-sensors on all cars and car controls (wheel, brakes, throttle etc) on all cars everywhere to spy on everybody all the time to learn how real drivers drive in realistic edge-case scenarios where the sensory information is extremely difficult to interpret. (Is Tesla not doing this already?) Examples: on a moonless night in the driving rain/fog/snow with dirt-covered road and oncoming headlights, being blinded by glare coming over a crest, etc.

And that is the catch with computer-assisted driving; the circumstances where artificial assistance would be most useful are the circumstances under which artificial technologies are most limited, and in road vehicles the timeframe between sensor confusion and catastrophic collision is on the order of a second, which is too short for the driver to intervene. The AI vision people are doing their best, but it is very difficult to do better than the ruthless Darwinian selection behind the eyeball and visual cortex. For 350 million years, brains that guessed wrong under limited information simply died.

Some thought provoking opinions. Thanks.

I’m not sure that vision, at least in the human visible spectrum, is going to sufficient. But of course it doesn’t have to be.

The eyeball and visual cortex are part of an even more complex system. Not only do we process images but we also interpret the data. This interpretation is based on whole life experience. When you first become a driver you already have knowledge of thousands of important objects and their behaviours and interactions. Dogs behave like dogs whether you’re driving a car or not. A human holding a phone to their face is different, behaviourally, than one with there arms by their sides. Differentiating between a bicycle and a motorcycle is only in part a visual problem. Etc.

I don’t know how much data Tesla collect, and in what circumstances. It’s not clear to me how any data they collect is transformed into training data with suitably classified inputs and outcomes. But then there’s lots I don’t know.

I think we might get a trade off between edge cases, where humans have the advantage of their life experiences, and run of the mill situations, where humans vary in their experience and hence the degree to which they can act automatically but AVs are less easily distracted. Who’s to say where the boundary should be drawn?

[Big Tea]

Surely the sensor its self would detect obstructed vision? Low range cars already have automatic wipers, should it not clean its own 'window' if it even works in the visible spectrum

[roon]

Apart from the conclusions on the functionality the article contains this, to my mind very sensible, prescription for how these technologies should be introduced.

“Before selling these systems, automakers should be required to give the public validated evidence of that system’s safety—backed by rigorous simulations, track testing, and the use of safety drivers in real-world conditions.”

Human drivers are released onto roadways without these barring the latter i.e. some dozens of hours with a licensed human driver and a few hours with a driving instructor. If the AI can drive an instructor around for a few hours it will have demonstrated entry level human driving competence.

[Nickel]

Given the current state of technology, yes, although putting electronic signals or sensors in the roads presents difficulties which make the concept impractical for use on all roads, hence limiting the usefulness of the concept. (Not to say it won't happen. Maybe they'll finally let us do 200 on the freeway?)

IMO the robust solution will be once AI vision is good enough to look through the normal windscreen and windows to interpret the car's surroundings without having to rely on easy-to-obscure external sensors. Amusingly the driver, set well back from all the windows and who can move his head to see around any dust, dirt or ice on them is much better placed to achieve this than any feasibly mounted cameras. (The idea here is that any piece of crud on the windscreen approximately camera/sensor-sized or larger becomes relatively small once the viewer is set back by a foot or two.) Perhaps they'll mount a pair of goggly-eyes on stalks inside the cabin, or hang them from the rear-view mirror?

Developing the AI vision to this point will require (typical Google / Facebook solution) putting spy-cameras and spy-sensors on all cars and car controls (wheel, brakes, throttle etc) on all cars everywhere to spy on everybody all the time to learn how real drivers drive in realistic edge-case scenarios where the sensory information is extremely difficult to interpret. (Is Tesla not doing this already?) Examples: on a moonless night in the driving rain/fog/snow with dirt-covered road and oncoming headlights, being blinded by glare coming over a crest, etc.

And that is the catch with computer-assisted driving; the circumstances where artificial assistance would be most useful are the circumstances under which artificial technologies are most limited, and in road vehicles the timeframe between sensor confusion and catastrophic collision is on the order of a second, which is too short for the driver to intervene. The AI vision people are doing their best, but it is very difficult to do better than the ruthless Darwinian selection behind the eyeball and visual cortex. For 350 million years, brains that guessed wrong under limited information simply died.

This is an important set of ideas to keep in mind.

I would actually really like to see an autonomous vehicle rally race series. I think it would be fascinating to watch it develop. either run it with a top level human forerunner in a car as closely comparable performance wise or run it as a sub category in an existing series like wrc.

Show me an autonomous car deliberately holding a four wheel slide while it's snowing >5cm an hour and I'll eat crow.

I am reasonably convinced by now that in normal driving conditions, AI can do a significantly better job at driving than your average human. People are terrible at it overall. fast and slow in all the wrong places and habits that exacerbate traffic and reduce fuel economy. But adverse weather is a whole other ball of wax (although many humans do very poorly in these situations too).

[AJI]

... But adverse weather is a whole other ball of wax (although many humans do very poorly in these situations too).

It could be argued that an AV would do better in adverse weather conditions as an AV wouldn't take a risk a human would take. There was an instance some time ago when an AV wouldn’t merge onto a highway because in order to do it safely it would have had to have broken the speed limit!

Humans often have their priorities backwards when taking risks, for example, when crossing the road. I've found that when the weather is great humans tend to cross the road in a very responsible manner. However, when it's raining we love to bolt across the road in the most idiotic way possible. Our fear of getting wet overrides our fear of potential death. Our fear of many seemingly inconsequential things often overrides our fear of potential death, like our fear of being late. Speeding, driving too fast for the conditions, being too tired, driving drunk, driving angry.., humans do all of those things all the time. An AV won’t.

[Phil]

Some thought provoking opinions. Thanks.

I’m not sure that vision, at least in the human visible spectrum, is going to sufficient. But of course it doesn’t have to be.

The eyeball and visual cortex are part of an even more complex system. Not only do we process images but we also interpret the data. This interpretation is based on whole life experience. When you first become a driver you already have knowledge of thousands of important objects and their behaviours and interactions. Dogs behave like dogs whether you’re driving a car or not. A human holding a phone to their face is different, behaviourally, than one with there arms by their sides. Differentiating between a bicycle and a motorcycle is only in part a visual problem. Etc.

I argued this very specific point half way through this topic and I'm glad you brought it up again. Categorizing and distinguishing a pedestrian from a cyclist and that from a vehicle on the car is easy, relatively speaking. However, distinguishing a young kid on a bicycle from an adult, or playing kids running along a road from a jogger are very different. Yesterday I was driving and the car ahead of me was swerving slightly. Within tolerable levels, but the subtle bouncing between the lanes led me to believe he was either driving under influence, tired or sidetracked. Or maybe it was an AV. Either way, I knew it was important to keep a safety distance. I doubt any software would ever be able to pick up that subtle nuance and interpret it accordingly.

Same applies to the video that was linked of the AV driving on potentially icy roads and saving a momentary loss of control. No doubt, a human could find itself in that very same situation, but many drivers also preemptively take action in such situations by driving with the necessary foresight and taking into account how the other cars are behaving in front to gauge the according grip levels and potential dangers to apply with additional care. I'm not sure I'd trust a piece of software relying on a dozen sensors to correctly assess the situation accordingly. "Seeing" and interpreting are two different things.

Maybe I'm more critical than most in this topic precisely because of my IT background and perhaps because I have a somewhat broad understanding on how one would try to program the algorithms for such software. I simply refuse to believe that the tech will somehow mature to the point it's possible. Just as an unsafe driver might statistically drive for thousands of kilometers before being involved in a preventable accident, I suspect even 'unsafe' AVs could be roaming the roads before being involved in accidents that would reveal the shortcomings of what technology can accomplish.

I see that the main arguments in favor of AVs have progressed to the point that it's assumed that AVs could be potentially safer than humans. No doubt, there are many avoidable accidents on roads and road safety is an important issue. Having said that, I think road safety is something that could be improved greatly without the necessity of replacing 'human drivers' with AVs. For one, better and more strict driving tests and make it required by law to do every 5 to 10 years. More stricter enforcement and punishments of driving under the influence or when breaking the law (e.g. speeding). This could achieve a lot already in order to make the roads a safer place.

[AJI]

...better and more strict driving tests and make it required by law to do every 5 to 10 years. More stricter enforcement and punishments of driving under the influence or when breaking the law (e.g. speeding). This could achieve a lot already in order to make the roads a safer place.

All of the above is already happening, making what was once considered the 'freedom' of driving less 'free' by the second. In Australia, an annual driving test now applies to the aged regardless of driving record and competency.

Disabled drivers are scrutinised at levels that are verging on discriminatory (fortunately common sense prevailed and the decision to revoke the licenses of one eyed drivers was reversed...)

Zero tolerance for a blood alcohol above 0.05 has now been applied to everyone (3 month suspension), regardless of driving record, sobriety, competency, or whether it's a 1st offence or not.

Travelling at speed in a supercar easily capable of safely doing +200km/h is limited to 110 no matter what the conditions.

Eventually AV's will be the saviour of that freedom. I only hope you guys can come up with a solution sharpish. I, for one, would be most appreciative.

[Phil]

For sure. I was only bringing it up because a few pages earlier, there was the famous 3000 deaths per hour number being thrown around. For sure there are that many, but there's also a direct correlation between the accident rate and enforcement of law.

Australia is on one end of the far extreme in how far authorities go to enforce that these laws are enforced, especially alcohol at the wheel. Many countries could take the standards enforced there as an example.

In my country [Switzerland], speeding can even put you into jail. I hate to admit it, but increasing the punishments for driving offenses does actually work.

Perhaps certain laws were sufficient 30 years ago, but in most areas, the density (and the number of distractions inside and outside the car) have increased to such a degree that updating the laws is a necessity many countries haven't followed up on yet.

[Just_a_fan]

The issue about ongoing checks/tests is interesting. Here in the UK, we have to renew our firearm licence on a 5 yearly basis. There are checks, an interview (police officer talks to you) and the licence can be withdrawn for non-firearms offences. E.g. convicted of drink driving - firearm licence will usually be revoked. The former being evidence that you're not a "fit person". Be tricky to get it back too.

One can pass a driving test at 17yo and then drive for the next 50+ years unfettered by ongoing checks.

I would argue that my 2.5 tonne, 130mph SUV is a more dangerous weapon than my shotgun in terms of the ability to take life. Walk down the street with a shotgun and someone will phone the police. Drive down the same street and no one cares. See that bus queue? 10 people could be dead in seconds with the SUV. A bit less likely with the shotgun as people would scatter and be out of range fairly quickly.

Why don't we police driving as strictly as we do firearms?

[Andres125sx]

Yesterday I was driving and the car ahead of me was swerving slightly. Within tolerable levels, but the subtle bouncing between the lanes led me to believe he was either driving under influence, tired or sidetracked. Or maybe it was an AV. Either way, I knew it was important to keep a safety distance. I doubt any software would ever be able to pick up that subtle nuance and interpret it accordingly.

True, AVs would act different, they don´t need to notice any weird behaviour in the car in front to keep correct safety distance, they´ll do it all the time

[Andres125sx]

For sure. I was only bringing it up because a few pages earlier, there was the famous 3000 deaths per hour number being thrown around. For sure there are that many, but there's also a direct correlation between the accident rate and enforcement of law.

Australia is on one end of the far extreme in how far authorities go to enforce that these laws are enforced, especially alcohol at the wheel. Many countries could take the standards enforced there as an example.

In my country [Switzerland], speeding can even put you into jail. I hate to admit it, but increasing the punishments for driving offenses does actually work.

Perhaps certain laws were sufficient 30 years ago, but in most areas, the density (and the number of distractions inside and outside the car) have increased to such a degree that updating the laws is a necessity many countries haven't followed up on yet.

So you agree human drivers are far from responsible and need to be forced to comply with the laws

That´s exactly the reason AVs will save many lives, even when responsible drivers and motor nerds like myself will hate it, reality is humans are far from reliable. Some are responsible and good drivers, some are good drivers buy not responsibles, some are responsible but poor drivers, and some are irresponsible and poor drivers. While we are humans, this will not change and AVs will be a lot more reliable (the average) than ourselves... and I´ll repeat this again, when bugs are solved obviously, because I agree at first they will need a lot of testing to become really safer than humans

[henry]

Apart from the conclusions on the functionality the article contains this, to my mind very sensible, prescription for how these technologies should be introduced.

“Before selling these systems, automakers should be required to give the public validated evidence of that system’s safety—backed by rigorous simulations, track testing, and the use of safety drivers in real-world conditions.”

Human drivers are released onto roadways without these barring the latter i.e. some dozens of hours with a licensed human driver and a few hours with a driving instructor. If the AI can drive an instructor around for a few hours it will have demonstrated entry level human driving competence.

When a human comes to learn to drive they have built up over many years a back catalogue of useful information that an AV has to learn from scratch. They also have a lifetime of the habits of learning.

But you’re right a human has to demonstrate a set of control and situational awareness skills, plus a catalogue of knowledge in the U.K., and if they do that they are free to drive. There is a tacit assumption that they have the capability and the aptitude to continue to learn.

Right now I know of no legislature that requires an AV to sit an equivalent test to humans. That would seem to be a bare minimum.