autonomous cars - the future of motoring is driverless

[T & P]

Researchers in The Netherlands say self-driving cars and cyclists are not very compatible at the moment...

Bikes put spanner in works of Dutch driverless car schemes

 

[8ball]

Researchers in The Netherlands say self-driving cars and cyclists are not very compatible at the moment...

Bikes put spanner in works of Dutch driverless car schemes

Other incompatible things:

Self-driving cars and cars
Self-driving cars and drivers

 

[Chilli.s]

Self driving bikes... what a thrill, every time you chuck yer leg over it. Also incompatible with common sense.

When I was a kid everyone thought flying cars would be a thing. The completely self driving car that you just sit in the back as it drives you anywhere you want, is about the same kind of likely.

 

[weltweit]

There are self driving bikes, I saw a video about them.

 

[HAL9000]

Tesla Shifts the Goalposts For 'Full Self-Driving' Technology

Tesla has been selling "full self-driving" capability since 2016, promising that "you will be able to summon your Tesla from pretty much anywhere," and that "once it picks you up, you will be able to sleep, read or do anything else en route [sic] to your destination." Last week Tesla shifted the goalposts, redefining "full self-driving" as a number of Level 2 driver assistance features that were already available, and a few new tricks to be delivered later. All will require a qualified driver behind the wheel, paying attention at all times and ready to take over if the car can't handle the situation. Worse, owners who bought the previous full self-driving feature paid $8,000 for it. Tesla is now offering owners who bought their cars prior to the change the same package for $5,000. Owners who paid the $3,000 higher price are unsure if the previously promised technology has been abandoned and Level 2 is now the most they can expect.

Tesla Shifts the Goalposts For 'Full Self-Driving' Technology - Slashdot

 

[weltweit]

I drive through a complex junction on my way home.

I arrive on a fast main A road which goes on through a staggered junction. I want to turn right and just past that right turn is the left turn onto a smaller road. I always indicate plenty early to warn the cars behind that I might have to slow to a complete stop before being able to turn right.

In order to turn right safely I have to negotiate cars arriving at speed around a blind corner from the other direction on the A road, and cars turning right from the smaller road which joins from the left but close to the road I want to take on the right.

Sometimes vehicles arriving at the junction from the left stop and give way to me, even though they shouldn't. Sometimes they pull out onto the A road and then flash me to cross their path, which they also shouldn't.

Sometimes as I am slowing to make my turn, vehicles appear coming towards me on the A road, travelling at speed and I have to stop on the A road. Often this causes a queue behind me of other vehicles that also have to stop. Despite the junction sign, many of them are not expecting to stop.

There are regularly nasty accidents at this junction.

Personally I doubt, given their present levels of development, I would trust an automated car to navigate this junction safely.

 

[Crispy]

I expect automated cars to be deployed in "easy driving" areas to start with (eg. Waymo in Phoenix AZ). As they become popular and the benefits become obvious, road infrastructure will be adapted to suit them, just as it was the last time there was a disruptiive change in vehicle technology.

 

[weltweit]

I expect automated cars to be deployed in "easy driving" areas to start with (eg. Waymo in Phoenix AZ). As they become popular and the benefits become obvious, road infrastructure will be adapted to suit them, just as it was the last time there was a disruptiive change in vehicle technology.

Do you anticipate the auto-cars will warn their owners they are entering a danger zone when they stray from the easy roads?

 

[Almor]

Do you anticipate the auto-cars will warn their owners they are entering a danger zone when they stray from the easy roads?

 

[BigTom]

Do you anticipate the auto-cars will warn their owners they are entering a danger zone when they stray from the easy roads?

They will be geo locked to those easy roads and won't be able to take you outside them.

Also, whilst i don't think auto cars will be networked, it is possible and that would make the junction situation considerably easier for an auto car then for a human.
Even without a network, a range of sensors might allow an auto car to detect vehicles we can't see try, more accurately judge their speed/distance/direction etc. I don't think it's a given that they will struggle in situations we struggle in, and as crispy says, places will get redesigned around their problems if/when they become popular.

I think problems with pedestrians and cyclists are a much bigger issue than navigating around other cars.

 

[weltweit]

They will be geo locked to those easy roads and won't be able to take you outside them.

Not a popular idea as a means of transport, a vehicle that can't use most of the road network.

Also, whilst i don't think auto cars will be networked, it is possible and that would make the junction situation considerably easier for an auto car then for a human.
Even without a network, a range of sensors might allow an auto car to detect vehicles we can't see try, more accurately judge their speed/distance/direction etc. I don't think it's a given that they will struggle in situations we struggle in,

There is a whole humans being human to other road users which auto-cars just won't get, for example pushing your nose out into traffic so that you do actually get to join the busy slow moving road, watching for the lifting of a finger to indicate you can join a road in front of the driver whose finger lift has just given way to you, the many meanings behind a flashing of the lights, the flash of hazard lights to say thank you.

I don't think an auto-car will be able to either see or make sense of or use these small human indications.

and as crispy says, places will get redesigned around their problems if/when they become popular.

I think problems with pedestrians and cyclists are a much bigger issue than navigating around other cars.

That may be true, I haven't thought a great deal about it yet, but what is also true is that at the moment every drivers seat is presently occupied by a pedestrian / cyclist.

 

[kabbes]

Not a popular idea as a means of transport, a vehicle that can't use most of the road network.


There is a whole humans being human to other road users which auto-cars just won't get, for example pushing your nose out into traffic so that you do actually get to join the busy slow moving road, watching for the lifting of a finger to indicate you can join a road in front of the driver whose finger lift has just given way to you, the many meanings behind a flashing of the lights, the flash of hazard lights to say thank you.

I don't think an auto-car will be able to either see or make sense of or use these small human indications.




That may be true, I haven't thought a great deal about it yet, but what is also true is that at the moment every drivers seat is presently occupied by a pedestrian / cyclist.

An auto-car will be able to see and make sense of any indications you can and with more reliability too. That isn’t the problem. The problem is novelty not established mores.

 

[BigTom]

Not a popular idea as a means of transport, a vehicle that can't use most of the road network.

I was thinking about the kind of situation Crispy mentioned, which is a whole town. I can see this as being the start - auto taxi services operating within towns which are easy - it's what Uber were looking to do before they killed a pedestrian. That's the start, and more places get added over time as the companies spend the time mapping out towns and as towns redesign junctions which cause problems. Motorways should be the easiest situation to deal with so potentially it wouldn't be long before auto-cars could drive around cities and between them on motorways. Once again, rural areas will get left behind new transportation options but I think there's a plausible development path in this. Weather is a massive issue though, hence the start being in places like arizona and new mexico which are clear & sunny basically all the time.

There is a whole humans being human to other road users which auto-cars just won't get, for example pushing your nose out into traffic so that you do actually get to join the busy slow moving road, watching for the lifting of a finger to indicate you can join a road in front of the driver whose finger lift has just given way to you, the many meanings behind a flashing of the lights, the flash of hazard lights to say thank you.

I don't think an auto-car will be able to either see or make sense of or use these small human indications.

So here I was really thinking about the situation where all cars are automated and networked together, rather than there being a mix. I wonder how well a mix will work but can't see us skipping to all automated in one step.

Small movements like flicks of the finger maybe, but auto-cars will see lights and stuff and can be taught/learn what they mean. Flash of hazards is an interesting one to deal with but worst case scenario is the auto car brakes in reaction to the hazards then stops braking once the hazards don't flash again and/or it detects the car in front moving away.

That may be true, I haven't thought a great deal about it yet, but what is also true is that at the moment every drivers seat is presently occupied by a pedestrian / cyclist.

Most drivers don't cycle and most drivers behaviour around cyclists is not very good. I personally look forward to automated cars which will always give space whilst passing, will make better judgements about the value of passing (so many drivers go past you straight into a traffic jam or red light, auto cars will be better able to judge speed/distance and work out that there's no point in passing the cyclist), will be far, far, far, far, far less likely to pull out on a cyclist from a side road (the most common cause of serious injuries to cyclists), trucks won't turn left onto cyclists etc.

To do this they need to be able to reliably detect and identify cyclists (and pedestrians) and they are getting better and better at doing this.

The range and scope of sensors on auto cars is potentially so much better than what humans have that the potential to detect and react to risky/dangerous situations is far greater. We have two eyes, giving us vision in what, 120-150degrees? and a lot of that is side/peripheral vision and ears which are compromised by the car body and engine noise. Cars can have eyes that see 360degrees, that see in infra-red so see just as well at night as in the day. Then they can have radar, lidar, heat sensors, audio sensors on the outside of cars, not sure what else, but so much more than we can do, always paying full attention to the road and surroundings.
There's a lot of work to do but the potential is enormous.

 

[kabbes]

weltweit , can you explain your understanding of the way in which the cars have learnt how to drive?

 

[weltweit]

weltweit , can you explain your understanding of the way in which the cars have learnt how to drive?

My understanding is that many are using a combination of maps, sensors and cameras, none of which on their own would be sufficient for safety, but in combination can be relatively effective.

 

[BigTom]

My understanding is that many are using a combination of maps, sensors and cameras, none of which on their own would be sufficient for safety, but in combination can be relatively effective.

Those are the things they use to understand what their surroundings are - that's not the same thing as learning how to drive.

They are learning from experience - and that's the collective experience of all the cars in their network (even if they are not actively networked they will all be using the aggregated data from every car to learn, but I doubt google and uber and waymo and etc will be sharing data). They record what happens, how other drivers behave and build up a collective memory of situations. Kind of like we do, except we only have our own experiences. The eg: google cars probably have way more hours of combined on-road experience than me or you do. We get to benefit from past experience to some extent (it's what gives us the road laws and some collectively observed cultural behaviours), but nothing like the benefit the cars get from it.

Now there's a whole load of things about machine learning that I don't understand, and that may present problems that I can't foresee, but they gain experience of how other drivers have behaved in the past in a particular situation or similar situations, and use that to decide how to act.

 

[Poot]

Those are the things they use to understand what their surroundings are - that's not the same thing as learning how to drive.

They are learning from experience - and that's the collective experience of all the cars in their network (even if they are not actively networked they will all be using the aggregated data from every car to learn, but I doubt google and uber and waymo and etc will be sharing data). They record what happens, how other drivers behave and build up a collective memory of situations. Kind of like we do, except we only have our own experiences. The eg: google cars probably have way more hours of combined on-road experience than me or you do. We get to benefit from past experience to some extent (it's what gives us the road laws and some collectively observed cultural behaviours), but nothing like the benefit the cars get from it.

Now there's a whole load of things about machine learning that I don't understand, and that may present problems that I can't foresee, but they gain experience of how other drivers have behaved in the past in a particular situation or similar situations, and use that to decide how to act.

I think that the human element of relying on drivers' facial expressions and really very small movements (which we all do) will be a bit problematic for learning purposes when there is a mixture of cars and driverless cars. But the thing that's almost impossible to teach will be that often you notice a driver who has a specific style: aggressive or dithery or slow, and sometimes you notice them several cars away and react accordingly by either being careful not to drive into the back of them at roundabouts or when you can tell that they don't know where they're going (ditherers), or just keeping out of the way (aggressive drivers who tailgate and undertake) or overtaking in the fast lane as soon as possible (middle lane hoggers). I'm not sure a driverless car will ever be able to pick this up.

 

[weltweit]

Those are the things they use to understand what their surroundings are - that's not the same thing as learning how to drive.

Fair enough.

They are learning from experience - and that's the collective experience of all the cars in their network (even if they are not actively networked they will all be using the aggregated data from every car to learn, but I doubt google and uber and waymo and etc will be sharing data). They record what happens, how other drivers behave and build up a collective memory of situations. Kind of like we do, except we only have our own experiences. The eg: google cars probably have way more hours of combined on-road experience than me or you do. We get to benefit from past experience to some extent (it's what gives us the road laws and some collectively observed cultural behaviours), but nothing like the benefit the cars get from it.

It will be interesting to see, what the effect is on their driving?

Now there's a whole load of things about machine learning that I don't understand, and that may present problems that I can't foresee, but they gain experience of how other drivers have behaved in the past in a particular situation or similar situations, and use that to decide how to act.

I worked in the auto industry for an electronics company who, among more complex systems, developed software for window lift mechanisms. That may seem a simple mechanism, except that the mechanism had to work in icy conditions without failing, and also stop and back off the window if it detected a child's arm or worse neck trapped in the window. The code for the device was massive and with a startling amount of complexity, just for a single electric window.

We have seen lots of innovation in the automotive sector: seat belts, power steering, crumple zones, cruise control, ABS, ASR, automatic braking, sleeping driver detection, airbags, double glazed windows, heated windscreens, automatic wipers, ultrasonic reversing sensors, reversing cameras, lane departure warning, head up displays, run flat tyres, tyre pressure warning, fuel electric hybrids, hydrogen fuel cells, electric cars, GPS and satnav, catalytic converters, electronic stability control, diagnostics, smart contactless keys, start stop tech, Bluetooth, four wheel drive, etc etc

They have all been by and large incremental improvements. What we are talking about with fully autonomous vehicles is a level of complexity far and away greater than a safety critical ABS, or a cruise control, we are talking about a massive jump forward in engineering.

My argument is that the car industry has always progressed in increments, I see a lot of people like politicians, who usually don't know much about engineering, talking about autonomous vehicles as if they will be just one more step. I don't believe this. I think we should continue to progress in increments.

 

[kabbes]

Welweit, how did you learn to drive? How did you learn to respond to the things you say will be difficult for a machine to respond to?

 

[BigTom]

I think that the human element of relying on drivers' facial expressions and really very small movements (which we all do) will be a bit problematic for learning purposes when there is a mixture of cars and driverless cars. But the thing that's almost impossible to teach will be that often you notice a driver who has a specific style: aggressive or dithery or slow, and sometimes you notice them several cars away and react accordingly by either being careful not to drive into the back of them at roundabouts or when you can tell that they don't know where they're going (ditherers), or just keeping out of the way (aggressive drivers who tailgate and undertake) or overtaking in the fast lane as soon as possible (middle lane hoggers). I'm not sure a driverless car will ever be able to pick this up.

I agree with the first bit, but not the second, I think that will be teachable. Middle lane hoggers are easy - how do you spot a middle lane hogger? Because they haven't moved into the lane 1, despite not having a vehicle to overtake. That is surely easy for machines to learn. SImilarly, they should be able to spot tailgaters and undertakers easily - more easily than we can, since auto cars could be able to detect accurately the distance between two cars a number of cars ahead of itself, which we can see. Also it can see behind and beside accurately at the same time, which we can't. They will have the potential to track behaviour of more drivers simultaneously than we can, and access to a far bigger database of driver behaviour history. They could even remember licence plate numbers and expect similar behaviour to the last time they encountered that driver, perhaps sharing that information between cars so if your car encountered a driver behaving in a wayward way, then my car would know when I also came across them a little later.

 

[BigTom]

Fair enough.

It will be interesting to see, what the effect is on their driving?

I worked in the auto industry for an electronics company who, among more complex systems, developed software for window lift mechanisms. That may seem a simple mechanism, except that the mechanism had to work in icy conditions without failing, and also stop and back off the window if it detected a child's arm or worse neck trapped in the window. The code for the device was massive and with a startling amount of complexity, just for a single electric window.

We have seen lots of innovation in the automotive sector: seat belts, power steering, crumple zones, cruise control, ABS, ASR, automatic braking, sleeping driver detection, airbags, double glazed windows, heated windscreens, automatic wipers, ultrasonic reversing sensors, reversing cameras, lane departure warning, head up displays, run flat tyres, tyre pressure warning, fuel electric hybrids, hydrogen fuel cells, electric cars, GPS and satnav, catalytic converters, electronic stability control, diagnostics, smart contactless keys, start stop tech, Bluetooth, four wheel drive, etc etc

They have all been by and large incremental improvements. What we are talking about with fully autonomous vehicles is a level of complexity far and away greater than a safety critical ABS, or a cruise control, we are talking about a massive jump forward in engineering.

My argument is that the car industry has always progressed in increments, I see a lot of people like politicians, who usually don't know much about engineering, talking about autonomous vehicles as if they will be just one more step. I don't believe this. I think we should continue to progress in increments.

I agree, it will be interesting to see.
It is moving in increments - Tesla's auto-pilot, Volvo & other's self parking, Google, Waymo, Uber and others doing full auto trials in easy places with humans to intervene. All of these are incremental steps to fully autonomous vehicles aren't they?

 

[weltweit]

Welweit, how did you learn to drive? How did you learn to respond to the things you say will be difficult for a machine to respond to?

You are probably asking how machine learning will take place as well as how I learnt to drive so I will try to answer both.

I know there is such a thing as machine learning but I don't know how robust it is and how it will apply, without humungous levels of complexity in an autonomous car, or across a fleet of autonomous cars. The reason I mentioned the window lift above was because I wanted to illustrate the amazing level of software complexity in an item as simple as a electric window controller. And that software came into existence because at least one child was killed by an electric window. Such an item is tiny in complexity compared to actually driving a car on busy roads.

What are the levels of complexity involved in actually operating a car without human input and learning from the experience? What things will need be learnt? how will it be encoded? how will the system be designed in order to accept this on the go reprogramming? how will a fleet of vehicles gain from this learning? and who will take responsibility for the legal implications if one companies system does not learn something another makers vehicles learn which proves safety critical?

Me, I started to learn the skills required to drive while being driven by my parents, I began learning things like vehicle movements, I was interested in cars so I learnt to recognise the different makes, then, perhaps aged 12, I was cycling on the road so I continued learning road sense, probably aged 15 I started riding motorbikes on farms and learnt about things like gears, I drove tractors on farms. At 17 I started riding motorbikes on the road and passed my test, I then got a more powerful motorbike. Aged 18 I took car lessons and passed my test. I then had both a car and a motorbike on the road probably until the age of 45 when I stopped riding motorbikes. During that time I spent 10 years as a salesman driving 50,000 miles a year around the UK and Europe.

Aged 54 I continue to learn how to be a better driver, I pick up on minor mistakes and issues around the roads I use, and regularly tune my driving accordingly.

Now you kabbes might be going to say how much better it could be if a fleet of hundreds or thousands of vehicles could all learn from their collective experiences? They are going to drive many tens of thousands of miles more than an individual will, and this is true, except that there has never in the history of engineering thus far been a computerised system of self learning like that, in a safety critical application, it is completely new and largely unproven. Are we prepared to accept autonomous cars making mistakes and causing deaths because they haven't yet learnt enough? The Tesla which was not fully autonomous driving into a truck without seeing it, the other American autonomous car mowing down and killing a pedestrian because its sensors didn't see them (was that the Uber?). It turns out Google cars have been involved in a number of US accidents, most strenuously defended as not being the G car's fault, by definition though they were there where an accident happened, could we not expect more from our computer overlords?

My belief is that we need to advance incrementally, step by step, autonomous cars are coming, but we should proceed gradually towards them.

 

[weltweit]

I agree, it will be interesting to see.
It is moving in increments - Tesla's auto-pilot, Volvo & other's self parking, Google, Waymo, Uber and others doing full auto trials in easy places with humans to intervene. All of these are incremental steps to fully autonomous vehicles aren't they?

Yes, you are right, things like automatic parking are steps forward rather than giant leaps.

There just might be a continental thing going on here also. The Americans have long made their vehicles as easy to use as possible with things like automatic transmission so that younger drivers can use them, and they automate things in a big way which perhaps indicates a greater trust in technology than we have. This could be rubbish . The US does seem to have more autonomous car projects on the go than Europe though.

 

[BigTom]

Are we prepared to accept autonomous cars making mistakes and causing deaths because they haven't yet learnt enough? The Tesla which was not fully autonomous driving into a truck without seeing it, the other American autonomous car mowing down and killing a pedestrian because its sensors didn't see them (was that the Uber?). It turns out Google cars have been involved in a number of US accidents, most strenuously defended as not being the G car's fault, by definition though they were there where an accident happened, could we not expect more from our computer overlords?

To be proper about this you need to look at collisions, injuries and deaths per mile traveled, and compare that to humans. It was Uber that killed a pedestrian, yes - and that is something that needs to be learnt from and addressed. But to decide if you need to stop it completely, you need to make the comparison as a rate against human drivers. If human drivers would have killed two pedestrians in the miles that uber auto cars covered, then yes, I am prepared to accept those deaths, since they are actually lives saved.

 

[weltweit]

To be proper about this you need to look at collisions, injuries and deaths per mile traveled, and compare that to humans. It was Uber that killed a pedestrian, yes - and that is something that needs to be learnt from and addressed. But to decide if you need to stop it completely, you need to make the comparison as a rate against human drivers. If human drivers would have killed two pedestrians in the miles that uber auto cars covered, then yes, I am prepared to accept those deaths, since they are actually lives saved.

I would have thought we should be looking at a step change, a factor change, if the effort and expense and nannying of autonomous cars were to be deemed worthwhile and a success.

 

[Poot]

I agree with the first bit, but not the second, I think that will be teachable. Middle lane hoggers are easy - how do you spot a middle lane hogger? Because they haven't moved into the lane 1, despite not having a vehicle to overtake. That is surely easy for machines to learn. SImilarly, they should be able to spot tailgaters and undertakers easily - more easily than we can, since auto cars could be able to detect accurately the distance between two cars a number of cars ahead of itself, which we can see. Also it can see behind and beside accurately at the same time, which we can't. They will have the potential to track behaviour of more drivers simultaneously than we can, and access to a far bigger database of driver behaviour history. They could even remember licence plate numbers and expect similar behaviour to the last time they encountered that driver, perhaps sharing that information between cars so if your car encountered a driver behaving in a wayward way, then my car would know when I also came across them a little later.

I'm not sure that I agree. Don't you find that when you are driving, it's not just a case of identifying what sort of drivers are around you but also of reacting accordingly incredibly swiftly based - and I hate to say it - pretty much on instinct? Like you can figure out exactly who's going to cut you up on a roundabout because they're distracted just by seeing that their reactions are slightly slow or undisciplined but not enough to comment on, and that would be imperceptible to any sort of sensor. I'm surprised that you always find it easy to spot a middle lane hogger, too. I find it less simple because they are quite often going very slightly faster than the vehicles to the left, but not necessarily fast enough to actually overtake because that's not really their aim.

Of course, if it was all automated cars rather than a mix, this would all be much easier.

 

[kabbes]

To be clear, anything you do on “instinct” whilst driving is actually a conditioned response developed from repeated cases of experiencing a particular pattern and the outcome of that pattern. It isn’t innate. Replicating “instinct” is exactly what machine learning is designed to do and it does it better than any human because for every 1000 such patterns you experience and learn from, the machine can experience and learn from 1000 million. How robust is it? Incredibly robust. It’s being used in your everyday life in everything from search engines and sat navs to traffic lights and medical diagnostics. Next up — the nuclear industry is developing machine learning for predicting failures in its systems. It outperforms humans on any task that involves learning from experience to develop “instinct” and that’s where the use cases are going.

I think if you want to look for the problems with autonomous cars, it’s precisely the opposite to the instincts you’re proud to have developed. It’s the times you override that instinct or had to figure out something you don’t have instinct for. Novelty is a problem for machines because they don’t have intentionality, so they can’t take a top-down view of a system to spot where it is failing them. That infamous Uber death occurred when the light source confused the car’s sensors, but whereas a person with the sun in their eyes gets that a problem with their driving model itself has developed, the machine can’t do that kind of meta-analysis. Either the failsafe has been preprogrammed or learnt or it doesn’t exist.

 

[Chilli.s]

People are gonna hate auto drive cars. They'll be so slow, drive you into trafic jams, or areas you dont know, and when they break...gridlock!

 

[Chilli.s]

Only got to have one crash with you in it and you'll never want to go in one again.

 

[BigTom]

Only got to have one crash with you in it and you'll never want to go in one again.

ever had a crash in a human driven car? Fancy going in one again?
I don't see a problem here. Planes, trains, cars, bikes, shoes. People crash whilst using these and keep using them. Why would auto cars be any different?

Also why would auto-cars be slow? Why would they drive you into traffic jams, where you wouldn't? They have the potential to be networked and avoid traffic jams by working out together all the routes they are going to take and planning their individual routes as a group for maximum efficiency (I don't think this will happen but it's possible).

What's the issue with driving into areas you don't know - isn't that an advantage, that you don't need to know the places you are going?

And when they break, they will not cause any more gridlock than a manual car, since you will still be able to push them out of the way, or if they are still somewhat functional they will know they've had a problem or there will be a park/breakdown button so they will pull up on the side like we do.